Modern RF and Microwave Measurement Techniques (Valeria Teppati, Andrea Ferrero and Mohamed Sayed, 2013) Livro - 476 Páginas

more information - www.cambridge.
org/9781107036413
Modern RF and Microwave Measurement Techniques
This comprehensive, hands-on review of the most up-to-date techniques in RF and

microwave measurement combines microwave circuit theory and metrology, in-depth
analysis of advanced modern instrumentation, methods and systems, and practical advice
for professional RF and microwave engineers and researchers.
Topics covered include microwave instrumentation, such as network analyzers, real-
time spectrum analyzers, and microwave synthesizers; linear measurements, such as
VNA calibrations, noise figure measurements, time domain reflectometry and multiport
measurements; and nonlinear measurements, such as load- and source-pull techniques,
broadband signal measurements, and nonlinear NVNAs.
Each technique is discussed in detail, and accompanied by state-of-the-art solutions
to the unique technical challenges associated with its deployment. With each chapter
delivered by internationally recognized experts in the field, this is an invaluable resource
for researchers and professionals involved with microwave measurements.
Valeria Teppati is a Researcher in the Millimeter Wave Electronics Group of the Depart-
ment of Information Technology and Electrical Engineering at ETH Zürich, developing
innovative solutions to aspects of linear and nonlinear measurement techniques.
Andrea Ferrero is a Professor in the RF, Microwave and Computational Electronics group
of the Department of Electronics and Telecommunications at Politecnico di Torino. He
is a Distinguished Microwave Lecturer of the IEEE Microwave Theory and Techniques
Society, and a Fellow of the IEEE.
Mohamed Sayed is the Principal Consultant for Microwave and Millimeter Wave Solu-
tions, and has nearly thirty years’ experience of developing microwave and millimeter
wave systems for Hewlett-Packard Co. and Agilent Technologies Inc.
The Cambridge RF and Microwave Engineering Series
Series Editor
Steve C. Cripps, Distinguished Research Professor, Cardiff University
Peter Aaen, Jaime Plá, and John Wood, Modeling and Characterization of RF and
Microwave Power FETs
Dominique Schreurs, Máirtín O’Droma, Anthony A. Goacher, and Michael Gadringer
(Eds), RF Amplifier Behavioral Modeling
Fan Yang and Yahya Rahmat-Samii, Electromagnetic Band Gap Structures in Antenna
Engineering
Enrico Rubiola, Phase Noise and Frequency Stability in Oscillators
Earl McCune, Practical Digital Wireless Signals
Stepan Lucyszyn (Ed.), Advanced RF MEMS
Patrick Roblin, Nonlinear FR Circuits and the Large-Signal Network Analyzer
Matthias Rudolph, Christian Fager, and David E. Root (Eds), Nonlinear Transistor
Model Parameter Extraction Techniques
John L. B. Walker (Ed.), Handbook of RF and Microwave Solid-State Power Amplifiers
Anh-Vu H. Pham, Morgan J. Chen, and Kunia Aihara, LCP for Microwave Packages
and Modules
Sorin Voinigescu, High-Frequency Integrated Circuits
Richard Collier, Transmission Lines
Valeria Teppati, Andrea Ferrero, and Mohamed Sayed (Eds), Modern RF and
Microwave Measurement Techniques
Forthcoming
David E. Root, Jason Horn, Jan Verspecht, and Mihai Marcu, X-Parameters
Richard Carter, Theory and Design of Microwave Tubes
Nuno Borges Carvalho and Dominique Schreurs, Microwave and Wireless
Measurement Techniques
Modern RF and Microwave
Measurement Techniques
Edited by
VAL ERIA T E PPAT I

ETH Zürich
ANDREA FE RRE RO
Politecnico di Torino
MOHAME D SAYE D
Microwave and Millimeter Wave Solutions
cambridge university press
Cambridge, New York, Melbourne, Madrid, Cape Town,
Singapore, São Paulo, Delhi, Mexico City
Cambridge University Press
The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9781107036413
© Cambridge University Press 2013
This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2013
Printed and bound in the United Kingdom by the MPG Books Group
A catalogue record for this publication is available from the British Library
Library of Congress Cataloguing in Publication data

Modern RF and microwave measurement techniques / [edited by] Valeria Teppati,
Andrea Ferrero, Mohamed Sayed.
pages cm. – (The cambridge RF and microwave engineering series)
Includes bibliographical references and index.
ISBN 978-1-107-03641-3 (hardback)
1. Radio measurements. 2. Microwave measurements. 3. Radio circuits.
I. Teppati, Valeria, 1974– editor of compilation.
TK6552.5.M63 2013
621.382028 7–dc23 2013000790
ISBN 978-1-107-03641-3 Hardback
Cambridge University Press has no responsibility for the persistence or

accuracy of URLs for external or third-party internet websites referred to in
this publication, and does not guarantee that any content on such websites is,
or will remain, accurate or appropriate.
This book is dedicated to the memory of our colleague Dr. Roger D. Pollard,
innovator, educator, contributor and friend.
Contents
Preface page xvii

List of contributors xix
List of abbreviations xxi
Part I General concepts 1
1 Transmission lines and scattering parameters 3

Roger Pollard and Mohamed Sayed
1.1 Introduction 3
1.2 Fundamentals of transmission lines, models and equations 3
1.2.1 Introduction 3
1.2.2 Propagation and characteristic impedance 4
1.2.3 Terminations, reflection coefficient, SWR, return loss 7
1.2.4 Power transfer to load 8
1.3 Scattering parameters 8
1.4 Microwave directional coupler 11
1.4.1 General concepts 11
1.4.2 The reflectometer 12
1.5 Smith Chart 13
1.6 Conclusions 16
References 20
Appendix A Signal flow graphs 16
Appendix B Transmission lines types 18
2 Microwave interconnections, probing, and fixturing 21

Leonard Hayden
2.1 Introduction 21
2.2 Device boundaries and measurement reference planes 21
2.2.1 Devices 22
2.2.2 Transmission lines 22
2.2.3 Circuits 23
viii Contents
2.3 Signal-path fixture performance measures 24

2.3.1 Delay 24
2.3.2 Loss 24
2.3.3 Mismatch 25
2.3.4 Crosstalk 27
2.3.5 Multiple-modes 28
2.3.6 Electromagnetic discontinuity 29
2.4 Power-ground fixture performance measures 30
2.4.1 Non-ideal power 30
2.4.2 Non-ideal ground 32
2.5 Fixture loss performance and measurement accuracy 33
2.6 Microwave probing 34
2.6.1 Probing system elements 35
2.6.2 VNA calibration of a probing system 36
2.6.3 Probing applications – in situ test 37
2.6.4 Probing applications – transistor characterization 37
2.7 Conclusion 38
References 38
Part II Microwave instrumentation 39
3 Microwave synthesizers 41
Alexander Chenakin
3.1 Introduction 41
3.2 Synthesizer characteristics 41
3.2.1 Frequency and timing 42
3.2.2 Spectral purity 43
3.2.3 Output power 47
3.3 Synthesizer architectures 47
3.3.1 Direct analog synthesizers 47
3.3.2 Direct digital synthesizers 50
3.3.3 Indirect synthesizers 52
3.3.4 Hybrid architectures 54
3.4 Signal generators 55
3.4.1 Power calibration and control 55
3.4.2 Frequency and power sweep 57
3.4.3 Modulation 58
3.5 Conclusions 62
References 62
Contents ix
4 Real-time spectrum analysis and time-correlated measurements applied to

nonlinear system characterization 64
Marcus Da Silva
4.1 Introduction 64
4.1.1 Types of spectrum analyzers 65
4.2 Spectrum analysis in real-time 68
4.2.1 Real-time criteria 69
4.2.2 Theoretical background 69
4.3 Spectrum analysis using discrete Fourier transforms 70
4.3.1 The Fourier transform for discrete-time signals 70
4.3.2 Regularly spaced sequential DFTs 71
4.4 Windowing and resolution bandwidth (RBW) 72
4.4.1 Windowing considerations 74
4.4.2 Resolution bandwidth (RBW) 75
4.5 Real-time specifications 76
4.5.1 Real-time criteria 76
4.5.2 Minimum event duration for 100% probability of intercept at the
specified accuracy 76
4.5.3 Comparison with swept analyzers 78
4.5.4 Processing all information within a signal with
no loss of information 80
4.5.5 Windowing and overlap 81
4.5.6 Sequential DFTs as a parallel bank of filters 83
4.5.7 Relating frame rate, frame overlap, and RBW 85
4.5.8 Criteria for processing all signals in the input waveform with no
loss of information 85
4.6 Applications of real-time spectrum analysis 85
4.6.1 Displaying real-time spectrum analysis data 85
4.6.2 Digital persistence displays 86
4.6.3 The DPX spectrum display engine 86
4.7 Triggering in the frequency domain 88
4.7.1 Digital triggering 88
4.7.2 Triggering in systems with digital acquisition 89
4.7.3 RTSA trigger sources 90
4.7.4 Frequency mask trigger (FMT) 90
4.7.5 Frequency mask trigger time resolution and time alignment 91
4.7.6 Other real-time triggers 92
4.8 Application examples: using real-time technologies to solve
nonlinear challenges 92
4.8.1 Discovering transient signals 92
4.8.2 Adjacent channel power (ACP) violation caused by power supply
fluctuations 93
4.8.3 Software errors affecting RF performance 93
x Contents
4.8.4 Memory effects in digitally pre-distorted (DPD) amplifiers 95

4.9 Conclusions 96
End Notes 96
References 97
5 Vector network analyzers 98

Mohamed Sayed and Jon Martens
5.1 Introduction 98
5.2 History of vector network analyzers 98
5.2.1 Pre-HP-8510 VNA – 1950–1984 98
5.2.2 HP-8510 VNA System – 1984–2001 99
5.2.3 Evolution of VNA to the Present – 2001–2012 101
5.3 Authors’ remarks and comments 101
5.4 RF and microwave VNA technology 101
5.4.1 Sources 103
5.4.2 Switches 107
5.4.3 Directional devices 109
5.4.4 Down-converters (RF portion of the receivers) 113
5.4.5 IF sections 117
5.4.6 System performance considerations 119
5.5 Measurement types in the VNA 121
5.5.1 Gain, attenuation, and distortion 121
5.5.2 Phase and group delay 121
5.5.3 Noise figure measurements 121
5.5.4 Pulsed RF measurements 121
5.5.5 Nonlinear measurements of active and passive devices 122
5.5.6 Multi-port and differential measurements 122
5.5.7 Load-pull and harmonic load-pull 122
5.5.8 Antenna measurements 122
5.5.9 Materials measurements 122
5.6 Device types for VNA measurements 123
5.6.1 Passive devices such as cables, connectors, adaptors, attenuators,
and filters 123
5.6.2 Low power active devices such as low noise amplifiers, linear
amplifiers, and buffer amplifiers 123
5.6.3 High power active devices such as base station amplifiers and
narrow-band amplifiers 123
5.6.4 Frequency translation devices such as mixers, multipliers,
up/down-converters and dividers 123
5.6.5 On-wafer measurements of the above devices 124
5.7 Improving VNA measurement range 125
5.7.1 Using a switch matrix box 125
5.7.2 Using multiple sources 125
Contents xi
5.7.3 Using reversing couplers 126

5.7.4 Using an external amplifier/attenuator 126
5.7.5 All-in-one VNA box 126
5.8 Practical tips for using VNAs 127
5.8.1 User training 127
5.8.2 Connector care 127
5.8.3 Temperature environment and stability 128
5.8.4 Measurement locations: production, development or research 128
5.9 Calibration and calibration kits 128
5.10 Conclusions 128
References 129
6 Microwave power measurements 130

Ronald Ginley
6.1 Introduction 130
6.1.1 Why power and not voltage and current? 131
6.2 Power basics, definitions, and terminology 131
6.2.1 Basic definitions 132
6.2.2 Different types of power measurements 132
6.3 Power detectors and instrumentation 136
6.3.1 Bolometric detectors 137
6.3.2 Thermoelectric detectors 138
6.3.3 Diode detectors 139
6.3.4 Power meters 142
6.3.5 Power measurements and frequency ranges 142
6.3.6 Power levels and detectors 143
6.4 Primary power standards 143
6.4.1 The microcalorimeter 145
6.4.2 The dry load calorimeter 146
6.4.3 Voltage and impedance technique 147
6.5 Basic power measurement techniques 148
6.5.1 Mismatch factor 149
6.5.2 Measuring power through an adapter 150
6.5.3 Power meter reference 151
6.6 Uncertainty considerations 151
6.6.1 Power meter uncertainty – uncertainty in Psub 152
6.6.2 ηDet uncertainty 152
6.6.3 Mismatch uncertainty 152
6.6.4 Adapter uncertainty 153
6.6.5 Device repeatability 153
6.7 Examples 154
6.8 Conclusions 157
References 157
xii Contents
7 Modular systems for RF and microwave measurements 160

Jin Bains
7.1.1 Virtual instrumentation 161
7.1.2 Instrumentation standards for modular instruments 163
7.1.3 PXI architecture 165
7.1.4 The role of graphical system design software 167
7.1.5 Architecture of RF modular instruments 169
7.2 Understanding software-designed systems 171
7.2.1 Measurement speed 171
7.3 Multi-channel measurement systems 175
7.3.1 Phase coherence and synchronization 176
7.3.2 MIMO 179
7.3.3 Direction finding 180
7.3.4 Phase array 183
7.4 Highly customized measurement systems 184
7.4.1 IQ data conditioning (flatness calibration) 184
7.4.2 Streaming 184
7.4.3 Integrating FPGA technology 186
7.5 Evolution of graphical system design 189
7.6 Summary 190
References 191
Part III Linear measurements 193
8 Two-port network analyzer calibration 195

Andrea Ferrero
8.2 Error model 195
8.3 One-port calibration 198
8.4 Two-port VNA error model 201
8.4.1 Eight-term error model 202
8.4.2 Forward reverse error model 204
8.5 Calibration procedures 207
8.5.1 TSD/TRL procedure 208
8.5.2 SOLR procedure 210
8.5.3 LRM procedure 211
8.5.4 SOLT procedure 215
8.6 Recent developments 216
8.7 Conclusion 217
References 217
Contents xiii
9 Multiport and differential S-parameter measurements 219

Valeria Teppati and Andrea Ferrero
9.2 Multiport S-parameters measurement methods 220
9.2.1 Calibration of a complete reflectometer multiport VNA 221
9.2.2 Calibration of a partial reflectometer multiport VNA 225
9.2.3 Multiport measurement example 229
9.3 Mixed-mode S-parameter measurements 230
9.3.1 Mixed-mode multiport measurement example 235
References 237
10 Noise figure characterization 240

Nerea Otegi, Juan-Mari Collantes, and Mohamed Sayed
10.2 Noise figure fundamentals 241
10.2.1 Basic definitions and concepts 241
10.2.2 Two noise figure characterization concepts:
Y-factor and cold-source 246
10.3 Y-factor technique 247
10.4 Cold-source technique 249
10.5 Common sources of error 251
10.5.1 Mismatch 252
10.5.2 Temperature effects 258
10.5.3 Measurement setup 260
10.6 Noise figure characterization of mixers 265
10.6.1 Noise figure definitions for frequency translating devices 266
10.6.2 Obtaining the SSB noise figure from Y-factor and cold-source 270
10.7 Conclusion 274
References 275
11 TDR-based S-parameters 279

Peter J. Pupalaikis and Kaviyesh Doshi
11.2 TDR pulser/sampler architecture 279
11.3 TDR timebase architecture 282
11.4 TDR methods for determining wave direction 286
11.5 Basic method for TDR-based S-parameter measurement 290
11.6 Summary of key distinctions between TDR and VNA 293
11.7 Dynamic range calculations 294
11.8 Dynamic range implications 298
11.9 Systematic errors and uncertainty due to measurement noise
in a network analyzer 300
11.9.1 Error propagation for a one-port DUT 300
xiv Contents

References 304
Part IV Nonlinear measurements 307
12 Vector network analysis for nonlinear systems 309

Yves Rolain, Gerd Vandersteen, and Maarten Schoukens
12.2 Is there a need for nonlinear analysis? 309
12.2.1 The plain-vanilla linear time-invariant world 309
12.2.2 Departure from LTI 310
12.2.3 Measuring a non-LTI system 310
12.2.4 Figures of merit to characterize the nonlinearity 311
12.3 The basic assumptions 312
12.3.1 Restricting the class of systems: PISPO systems 313
12.3.2 Influence of the excitation signal 315
12.3.3 The definition of the nonlinear operating point 320
12.4 Principle of operation of an NVNA 320
12.4.1 Introduction 320
12.4.2 Basic requirements for nonlinear characterization 321
12.4.3 A calibration for nonlinear measurements 323
12.5 Translation to instrumentation 327
12.5.1 Oscilloscope-based receiver setups 328
12.5.2 Sampler-based receiver setups 331
12.5.3 VNA-based setups 338
12.5.4 IQ-modulator based setups 341
12.6 Conclusion, problems, and future perspectives 342
References 343
13 Load- and source-pull techniques 345

Valeria Teppati, Andrea Ferrero, and Gian Luigi Madonna
13.2 Setting the load conditions: passive techniques 347
13.2.1 Basics 348
13.2.2 Harmonic load-pull with passive tuners 349
13.3 Setting the load conditions: active, open-loop
techniques 350
13.4 Setting the load conditions: active-loop techniques 352
13.4.1 Active loop: basics 352
13.4.2 Stability analysis of the active loop 353
13.4.3 Practical active-loop implementations 355
13.4.4 Wideband load-pull 356
13.4.5 Combining passive tuners and active techniques 357
Contents xv
13.5 Measuring the DUT single-frequency characteristics 359

13.5.1 Real-time vs. non-real-time load-pull measurements 360
13.5.2 Calibration of real-time systems 361
13.5.3 Mixed-mode, harmonic load-pull systems 365
13.6 Measuring the DUT time domain waveforms 368
13.6.1 Load-pull waveform techniques in the time domain 368
13.6.2 Load-pull waveform techniques in the frequency domain 370
13.6.3 Other calibration approaches 373
13.6.4 Measurement examples 374
13.7 Real-time source-pull techniques 375
References 379
14 Broadband large-signal measurements for linearity optimization 384

Marco Spirito and Mauro Marchetti
14.2 Electrical delay in load-pull systems 385
14.3 Broadband load-pull architectures 386
14.3.1 Detection scheme 386
14.3.2 RF front-end 388
14.3.3 System calibration 392
14.4 Broadband loads 392
14.4.1 Closed-loop active loads 392
14.4.2 Mixed-signal active loads 395
14.5 System operating frequency and bandwidth 400
14.6 Injection power and load amplifer linearity 400
14.7 Baseband impedance control 403
14.8 Broadband large-signal measurement examples 405
14.8.1 IMD asymmetries measurements 405
14.8.2 Phase delay cancellation 407
14.8.3 High power measurements with modulated signals 408
References 411
15 Pulse and RF measurement 414

Anthony Parker
15.2 Dynamic characteristics 414
15.3 Large-signal isodynamic measurements 417
15.3.1 Measurement outside safe-operating areas 418
15.3.2 Pulsed-RF characteristics 418
15.4 Dynamic processes 418
15.4.1 Temperature and self-heating 419
15.4.2 Charge trapping 420
xvi Contents
15.4.3 Impact ionization 422

15.5 Transient measurements 423
15.5.1 Measurement of gate lag 423
15.5.2 Time evolution characteristics 426
15.6 Pulsed measurement equipment 427
15.6.1 System architecture 427
15.6.2 Timing 430
15.7 Broadband RF linearity measurements 431
15.7.1 Weakly nonlinear intermodulation 432
15.7.2 Intermodulation from self-heating 433
15.7.3 Measuring heating response 435
15.7.4 Measuring charge trapping response 436
15.7.5 Measurement of impact ionization 437
15.8 Further investigation 438
References 439
Index 442
Preface
In the last few years, the field of microwave testing has been evolving rapidly with the
development and introduction of digital techniques and microprocessor based instru-
ments, and reaching higher and higher frequencies. Nevertheless, the basic underlying
concepts, such as frequency synthesis, network analysis and calibration, and spectrum
analysis, still constrain even the more modern equipment.
In recent years, microwave instrumentation has had to meet new testing requirements,
from 3G and now LTE wireless networks, for millimeter wave and THz applications. Thus
instrumentation and measurement techniques have evolved from traditional instruments,
such as vector network analyzers (VNAs), to increasingly more complex multifunction
platforms, managing time and frequency domains in a unified, extensive approach.
We can identify two main directions of evolution:
• linear measurements, essentially S-parameter techniques;
• nonlinear measurements, for high power and nonlinear device characterization.
S-parameter measurements have been moving towards the multiport and millimeter wave
fields. The first to characterize multi-channel transmission structures such as digital
buses, and the latter for space or short-range radio communication or security scanner
applications. New calibrations and instrument architectures have been introduced to
improve accuracy, versatility and speed.
Nonlinear applications have also evolved. Traditional high power transistor charac-
terization by load-pull techniques now also typically includes time domain waveform
measurements under nonlinear conditions. These techniques can nowadays also handle
the broadband signals used in most communication links, or pulsed signals. Moreover,
even nonlinear measurements had to evolve to multiport, with differential and com-
mon mode impedance tuning, due to the spreading of amplifiers and devices exploiting
differential configuration.
The idea of a comprehensive book on microwave measurement was born when we
noticed that the knowledge of these modern instrumentation and measurement techniques
was scattered inside different books or papers, sometimes dealing more with design or
modeling than with the measurement itself or the metrological aspects, and there was no
recent book covering these topics extensively.
xviii Preface
We thus tried to make an effort to produce a book that could:

• give an overview of modern techniques for measurements at microwave frequencies;
• be as complete and comprehensive as possible, giving general concepts in a
unitary way;
• treat modern techniques, i.e. the state of the art and all the most recent developments.
As editors of the book, we have been honored to work with several international experts
in the field, who contributed their invaluable experience to the various chapters of this
book. This multi-author approach should guarantee the reader a deep understanding of
such a complex and sophisticated matter as microwave measurements.
The book is structured in four main sections:
1. general concepts
2. microwave instrumentation
3. linear measurement techniques
4. nonlinear measurement techniques.
An already expert reader may directly jump to a specific topic, to read about innova-
tive instruments or techniques, such as synthesizers, modular RF instruments, multiport
VNAs or broadband load-pull techniques, or follow the book’s organization that will
guide him/her through the development of the instruments and their applications.
Fifteen chapters form the body of the four book sections. Two of them describe funda-
mentals, from the theory behind the S-parameters to the interconnections; five chapters
are then devoted to microwave instrumentation: synthesizers, network and spectrum
analyzers, power meters, up to modern microwave modular instrumentation. The third
section on linear measurements covers traditional two-port S-parameter calibration,
multiport S-parameter techniques, noise measurements and time domain reflectome-
try techniques. Finally the last section on nonlinear measurements describes nonlinear
VNAs, load-pull, broadband load-pull, and concludes with pulsed measurements.
All the content is correlated with details on metrological aspects whenever possible,
and with some examples of typical use, though we have tried to be as independent as
possible of a specific device under test and to concentrate on the measurement technique
rather than the particular application.
Contributors
Jin Bains
National Instruments Corp., USA
Alexander Chenakin
Phase Matrix, Inc., USA
Juan-Mari Collantes
University of the Basque Country (UPV/EHU), Spain
Kaviyesh Doshi
Teledyne LeCroy, USA
Andrea Ferrero
Politecnico di Torino, Italy
Ronald Ginley
NIST, USA
Leonard Hayden
Gian Luigi Madonna
ABB Corporate Research, Baden, Switzerland
Mauro Marchetti
Anteverta Microwave B.V., the Netherlands
Jon Martens
Anritsu Company, USA
Nerea Otegi
University of the Basque Country (UPV/EHU), Spain
Anthony Parker
Macquarie University, Sydney, Australia
Roger Pollard
Agilent Technologies, USA and University of Leeds, United Kingdom
Peter J. Pupalaikis
xx List of contributors
Yves Rolain
Vrije Universiteit Brussel, Belgium
Mohamed Sayed
Microwave and Millimeter Wave Solutions, USA
Maarten Schoukens
Marcus Da Silva
Tektronix Inc., USA
Marco Spirito
Delft University of Technology, the Netherlands
Valeria Teppati
ETH Zürich, Switzerland
Gerd Vandersteen
Abbreviations
ACLR Adjacent Channel Leakage Ratio

ACPR Adjacent Channel Power Ratio
AD Analog-to-Digital
ADC Analog-to-Digital Converter
ADS Advanced Design System
ALC Automatic Level Control
AM Amplitude Modulation
ASB All-Side-Band
AWG Arbitrary Waveform Generator
BER Bit Error Rate
BWO Backward Wave Oscillators
CDMA Code Division Multiple Access
CIS Coherent Interleaved Sampling
CMOS Complementary Metal-Oxide Semiconductor
CPU Central Processing Unit
CPW CoPlanar Waveguide
CRT Cathode Ray Tube
CW Continuous Wave
CZT Chirp Zeta Transform
DAC Digital-to-Analog Converter
DDC Digital Down Converter
DDS Direct Digital Synthesizer
DFT Discrete Fourier Transform
DMM Digital Multimeter
DPD Digital Pre-Distortion – used to linearize power amplifiers
DPX Digital Phosphor Processing – Tektronix implementation of a variable
persistence spectrum display.
DSA Discrete-time Spectrum Analyzer
DSB Double Side Band
DSP Digital Signal Processing
DTC Digital to Time Converter
DTFT Discrete-Time Fourier Transform
DUT Device Under Test
DVM Digital Volt Meter
xxii List of abbreviations
DWT Discrete Wavelet Transform

EDA Electronic Design & Automation
EDGE Enhanced Data rates for GSM Evolution
eLRRMTM enhanced Line-Reflect-Reflect Match (Cascade Microtech)
ENR Excess Noise Ratio
EVM Error Vector Magnitude
FEM Finite Element Method
FET Field Effect Transistor
FEXT Far End CrossTalk
FFT Fast Fourier transform
FM Frequency Modulation
FMT Frequency Mask Trigger
FOM Figure of Merit
FPGA Field-Programmable Gate Arrays
GMSK Gaussian Minimum Shift Keyed
GPIB General Purpose Interface Bus
GPU Graphics Processing Unit
GSG Ground-Signal-Ground
GSM Global System for Mobile Communications
GUI Graphical User Interface
HBT Heterojunction Bipolar Transistor
HEMT High Electron Mobility Transistor
HW HardWare
I In-phase component of vector modulation
IC Integrated Circuit
IDE Integrated Drive Electronics
IDFT Inverse Discrete Fourier transform
IDWT Discrete Wavelet Transform
IEEE Institute of Electrical and Electronics Engineers
IF Intermediate Frequency
IFFT Inverse Fast Fourier Transform
IL Insertion Loss
IM3 Third-order Inter Modulation distortion
IM5 Fifth-order Inter Modulation distortion
IMD Inter Modulation Distortion
IP Intellectual Property
IQ Cartesian vector modulation format – In-phase and Quadrature.
IVI Interchangeable Virtual Instruments
JTFA Joint Time-Frequency Analysis
KCL Kirchhoff’s Current Law
KVL Kirchhoff’s Voltage Law
LAN Local Area Network
LCD Liquid Crystal Display
LDMOS Laterally Diffused Metal Oxide Semiconductor
List of abbreviations xxiii
LO Local Oscillator
LPF Low Pass Filter
LRL Line Reflect Line
LRM Line Reflect Match
LSNA Large Signal Network Analyzer
LTE Long Term Evolution
LTI Linear Time Invariant
LUT Look Up Table
LXI LAN eXtensions for Instrumentation
MESFET MEtal-Semiconductor Field Effect Transistor
MIMO Multiple Input Multiple Output
MMIC Monolithic Microwave Integrated Circuit
MTA Microwave Transition Analyzer
NCO Numerically Controlled Oscillator
NEXT Near End CrossTalk
NF Noise Figure
NFA Noise Figure Analyzer
NIST National Institute of Standards and Technology
NVNA Nonlinear Vector Network Analyzer
OFDM Orthogonal Frequency-Division Multiplexing
OIP2 Output Second Order Intercept Point
OIP3 Output Third Order Intercept Point
ORFS Output RF Spectrum
OS Operating System
P2P Peer-to-Peer
PA Power Amplifier
PAE Power Added Efficiency
PC Personal Computer
PCB Printed Circuit Board
PCI Peripheral Component Interconnect
PCMCIA Personal Computer Memory Card International Association
PDF Probability Density Function
PFER Phase and Frequency Error
pHEMT pseudomorphic High Electron Mobility Transistor
PICMG PCI Industrial Computer Manufacturers Group
PISPO Periodic In, Same Period Out
PLL Phase Locked Loop
PM Phase Modulation
PMC PCI Mezzanine Card
PVT Power Versus Time
PXI PCI eXtensions for Instrumentation
PXISA PXI Systems Alliance
Q Quadrature component of vector modulation
QMF Quadrature Mirror Filter
xxiv List of abbreviations
QSOLT Quick Short Open Load Thru

RAID Redundant Array of Inexpensive Disks
RAM Random Access Memory
RBW Resolution BandWidth- The minimum bandwidth that can be resolved
in a spectrum analyzer display.
RF Radio Frequency
RL Return Loss
RMS Root Mean Square
ROM Read Only Memory
RSA Tektronix nomenclature for its real-time signal analyzer family
RSS Root Sum of Squares
RTSA Real-Time Signal Analyzer
SA Spectrum Analyzer
SATA Serial Advanced Technology Attachment
SI Signal Integrity
SMVR R&S nomenclature for its real-time signal analyzer family
SNR Signal to Noise Ratio
SOL Short Open Load
SOLT Short Open Load Thru
SSB Single-Side Band
SW SoftWare
SWR Standing Wave Ratio
TDEMI Gauss Instruments nomenclature for its real-time spectrum analyzers
targeted at emissions measurement
TDMA Time Division Multiple Access
TDR Time Domain Reflectometry
TDT Time Domain Transmission
TE Transverse Electric
TEM Transverse ElectroMagnetic
TL Transmission Line
TM Transverse Magnetic
TOI Third Order Intercept
TRL Thru Reflect Line
TSD Thru Short Delay
TXP Transmit Power
UHF Ultra High Frequency
UML Universal Modeling Language
USB Universal Serial Bus
VCO Voltage Controlled Oscillator
VISA Virtual Instrument Software Architecture
VNA Vector Network Analyzer
VPN Virtual Private Network
VSA Vector Signal Analyzer
VSG Vector Signal Generator
List of abbreviations xxv
VSWR Voltage Standing Wave Ratio

VXI VMEbus eXtensions for Instrumentation
W-CDMA Wideband Code Division Multiple Access
WLAN Wireless Local Area Network
YIG Yttrium-Iron-Garnet
Part I
General concepts
1 Transmission lines and scattering
parameters
Roger Pollard and Mohamed Sayed
1.1 Introduction
This chapter introduces the reader to the topics presented in the rest of the book,
and serves as a quick guide to the basic concepts of wave propagation and scattering
parameters.
Understanding these concepts becomes very important when dealing with RF and
microwave frequencies, as is shown in Section 1.2, where a simplified formulation for
the transmission line theory is given.
Section 1.3 provides the definition of the scattering matrix or S-matrix, the key element
to describe networks at RF, microwaves, and higher frequencies.
Section 1.4 deals with the most important component in microwave measurements,
the directional coupler, while Section 1.5 revises a common way to represent quantities
in the RF domain, the Smith Chart.
Finally, in Appendix A signal flow graphs, a typical way to represent simple linear
algebra operations, are presented, while Appendix B summarizes the various types of
transmission lines cited in this book.
1.2 Fundamentals of transmission lines, models and equations
1.2.1 Introduction
Electromagnetic waves travel at about the speed of light (c = 299 792 458 m/s) in air.
Using the relationship
ν = f λ, (1.1)
where ν is velocity (= c in air), f is frequency and λ is wavelength, the wavelength of

a 100 GHz wave is about 3 mm. If a simple connection on a circuit is of the order of
magnitude of a wavelength, it is then necessary to consider its behavior as distributed
and regard it as a transmission line. In fact, propagation phenomena already appear for
lengths of 1/10th of a wavelength.
Let’s clarify this concept by a simple example. When a source of electrical power is
connected to a load, as shown in Figure 1.1, the voltage appears at the load instantaneously
over a short distance.
4 Roger Pollard and Mohamed Sayed
Fig. 1.1 Connection of a light bulb close to the source of electrical power.
150 million kilometers
Fig. 1.2 Connection of a light bulb at 150 million km from the source of electrical power.
I(0,t) I(z,t)
v(0,t ) v (z,t)
Fig. 1.3 Two-wire line.
However, if the connection wiring is very long, as shown in Figure 1.2, it takes time
for the signal to propagate to the load. In this example, using the approximate distance
from the sun, the bulb would light some 8 minutes after the switch is closed.
This means that the connection cannot be modeled with a short circuit anymore, since
the voltage and current (or electric and magnetic fields) are now functions of both time
and position.
Let us consider a two-wire line, as shown in Figure 1.3.
Here both the voltage and current are functions of position and time. Now, if we model
the line as an infinite number of very short sections, each element can be considered as a
series inductance and shunt capacitance with associated losses, as shown in Figure 1.4.
This model can actually be applied to any kind of transmission line (waveguide, coaxial,
microstrip, etc.; see Appendix B for a brief description of the most common types of
transmission lines referred to in this book).
1.2.2 Propagation and characteristic impedance

In a two-conductor line, the model may be explained physically. The wire properties
and skin effect generate the inductance, the two conductors the capacitance, and leakage
Transmission lines and scattering parameters 5
I I + ΔI
LΔz RΔz
V CΔz GΔz V + ΔV
Fig. 1.4 Lumped-element model of a section of the two-wire line of Fig. 1.3.
and losses produce the parasitic resistances. These model elements are also functions of
frequency.
Solving the model circuit for the voltage and current, yields
V (z, t) = (Rz + j ωLz)I (z, t) (1.2)
and
I (z, t) = (Gz + j ωCz)V (z, t). (1.3)
Taking z as infinitely short, the partial derivatives of voltage and current with respect
to the z coordinate appear as:
∂V (z, t)
= − (R + j ωL) I (z, t) (1.4)
∂z
∂I (z, t)
= − (G + j ωC) V (z, t). (1.5)
∂z
Then, by differentiating (1.4) again with respect to z and substituting (1.5) in the obtained
equation (and vice versa) one gets:
∂ 2 V (z, t) ∂ 2 I (z, t)
= γ 2 V (z, t) and = γ 2 (z, t)I (z, t), (1.6)
dz2 dz2
√
where γ = (R + j ωL)(G + j ωC) = α + jβ is the propagation constant.
The equations have exponential solutions of the form
V = V1 e−j γ z + V2 e+j γ z , (1.7)
where the first part of the solution (V + = V1 e−j γ z ) is referred to as an incident wave,
and the second part (V − = V2 e+j γ z ) as a reflected wave.
In the same way, one can write the solution for the current as
I = I1 e−j γ z + I2 e+j γ z . (1.8)
By substituting (1.7) and (1.8) inside (1.4) and (1.5) one can find the relationship between
I1 -V1 and I2 -V2 , which are:
V1 = Z0 I1 (1.9)
and
V2 = −Z0 I2 (1.10)
with √
R + j ωL
Z0 = √ , (1.11)
G + j ωC
where Z0 is referred to as the characteristic impedance of the transmission line. Note that
the wave number β can be expressed as a function of vp , the so-called phase velocity,
or of the wavelength (λ):
ω 2π
β= = . (1.12)
vp λ
The time dependence of the voltage and current can be made explicit in this way
V (z, t) = V (z)ej ωt I (z, t) = I (z)ej ωt (1.13)
and the circuit equations rewritten as

∂V ∂I ∂I ∂V
= − RI + L and = − GV + C . (1.14)
∂z ∂t ∂z ∂t
Again, differentiating gives

∂ 2V ∂V ∂V ∂ 2V
= R GV + C +L G +C 2 (1.15)
∂z2 ∂t ∂t ∂t
or
∂ 2V ∂V ∂ 2V
= −(RC + LG) − LC − RGV = 0. (1.16)
dz2 ∂t dt 2
Note that the current I satisfies an identical equation.
In the case of lossless transmissions lines with R = G = 0, the propagation constant
and the characteristic impedance simplify to the trivial

√ L
γ = jβ = j ω LC and Z0 = . (1.17)
C
For most practical purposes, however, especially in a hollow pipe waveguide, the low-loss
case (R = ωL, G = ωC) provides accurate values:

√ 1√ R G
γ ≈ α + jβ = j ω LC + LC + (1.18)
2 L C
with
1√ R G 1
α= LC + = (RY0 + GZ0 ) (1.19)
2 L C 2
1.2.3 Terminations, reflection coefficient, SWR, return loss

We have seen how the total voltage on a transmission line is the vector sum of the
incident and reflected voltages and the phase relationship between the waves depends on
the position along the line. The nature of a discontinuity determines the phase relationship
of the incident and reflected waves at that point on the line and that phase relationship is
repeated at points that are multiples of a half-wavelength (180◦ ).
The classical example is when the line is terminated with a load impedance ZL that is
not the characteristic impedance. Some of the incident energy may be absorbed by the
load and the rest is reflected. The maximum and minimum values of the standing wave
voltage and the positions of these maxima and minima are related to ZL . The maximum
occurs where the incident and reflected voltages are in phase, the minimum where they
are 180◦ out of phase.

Emax = |Vincident | + Vreflected and Emin = |Vincident | − Vreflected (1.20)
with Vincident a constant and Vreflected a function of ZL , EEmax

min
is the Voltage Standing
Wave Ratio, abbreviated VSWR or SWR and is a way of describing the discontinuity at
the plane of the load. The SWR is 1 when the load termination is equal to the characteristic
impedance of the line, since Vreflected = 0, and infinite when a lossless reflective termina-
tion (short circuit, open circuit, capacitance, etc.) is connected, since Vreflected = Vincident
in that case. SWR is commonly used as a specification for components, most commonly
loads and attenuators.
For a finite ZL , the magnitude and phase of the reflected signal depends on the ratio of
ZL /Z0 . Since the total voltage (and current) across ZL is the vector sum of the incident
and reflected voltages (and currents) we have
VL Vincident + Vreflected
ZL = = . (1.21)
IL Iincident + Ireflected
The voltage and current in each of the waves on the transmission line are related by the
characteristic impedance, as already shown in (1.9) and (1.10)
Vincident Vreflected
= Z0 and = −Z0 (1.22)
Iincident Ireflected
so
Vreflected
Vincident + Vreflected 1+ Vincident 1+
ZL = = Z0 = Z0 (1.23)
Vincident
−
Vref elected
1−
Vreflected 1−
Z0 Z0 Vincident
where is the reflection coefficient, a complex value with magnitude and phase. The
magnitude of is usually denoted by the symbol ρ and its phase by θ . The values of ρ
vary from zero to one. It is common practice to refer to the magnitude of the reflection
coefficient as the return loss (20log10 ρ).
Note that ρ, the magnitude of , remains constant as the observation point is moved
along a lossless transmission line. In this case, the phase θ changes and thus the complex
value of rotates around a circle on a polar plot. Since, at the plane of the load
ZL − Z0
= (1.24)
ZL + Z0
the value of the impedance seen looking into the transmission line at any point is readily
calculated by rotating by the electrical length (a function of the signal frequency,
360◦ = λ/2) between the plane of the load and the point of observation. Thus, for
example, at a quarter-wavelength distance (180◦ electrical length) from the plane of a
short circuit, the impedance appears as an open circuit. The same impedance repeats at
multiples of a half-wavelength.
1.2.4 Power transfer to load

The maximum power transfer from sources with source impedance of Rs to load
impedance of RL occurs at the value of Rs equal to RL . For complex impedances,
the maximum power transfer occurs when ZL = RL + jXL , Zs = Rs − jXs and Rs = RL ,
and XL = Xs , otherwise there will be a mismatch and standing wave ratio.
1.3 Scattering parameters
A key assumption when making measurements is that networks can be completely charac-
terized by quantities measured at the network terminals (ports) regardless of the contents
of the networks. Once the parameters of a (linear) n-port network have been determined,
its behavior in any external environment can be predicted.
At low frequencies, typical choices of network parameters to be measured and handled
are Z-parameters or Y-parameters, i.e. the impedance or admittance matrix, respectively.
In microwave design, S-parameters are the natural choice because they are easier to
measure and work with at high frequencies than other kinds of parameters. They are
conceptually simple, analytically convenient, and capable of providing a great insight
into a measurement or design problem.
Similarly to when light interacts with a lens, and a part of the light incident is
reflected while the rest is transmitted, scattering parameters are measures of reflection
and transmission of voltage waves through an electrical network.
Let us now focus on the generic n-port network, shown in Figure 1.5
To characterize the performance of such a network, as we said, any of several para-
meter sets can be used, each of which has certain advantages. Each parameter set is
related to a set of 2n variables associated with the n-port model. Of these variables,
n represents the excitation of the network (independent variables), and the remaining
n represents the response of the network to the excitation (dependent variables). The
network of Figure 1.5, assuming it has a linear behavior, can be represented by its
I1 Ii
V1 Vi
I2
V2
In
Vn
Fig. 1.5 Generic n-port network.
Z-matrix (impedance matrix):

⎡ ⎤ ⎡ ⎤⎡ ⎤
V1 Z11 Z12 ··· Z1n I1
⎢ V2 ⎥ ⎢ Z21 Z22 ··· Z2n ⎥⎢ I2 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ .. ⎥=⎢ .. .. .. .. ⎥⎢ .. ⎥, (1.25)
⎣ . ⎦ ⎣ . . . . ⎦⎣ . ⎦
Vn Zn1 Zn2 ··· Znn In
where V1 -Vn are the node voltages and I1 -In are the node currents. Alternatively, one
can use the dual representation:
⎡ ⎤ ⎡ ⎤⎡ ⎤
I1 Y11 Y12 ··· Y1n V1
⎢ I2 ⎥ ⎢ Y21 Y22 ··· Y2n ⎥⎢ V2 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ .. ⎥=⎢ .. .. .. .. ⎥⎢ .. ⎥. (1.26)
⎣ . ⎦ ⎣ . . . . ⎦⎣ . ⎦
In Yn1 Yn2 ··· Ynn Vn
Here, port voltages are the independent variables and port currents are the depen-
dent variables; the relating parameters are the short-circuit admittance parameters, or
Y-parameters. In the absence of additional information, n2 measurements are required to
determine the n2 Y-parameter. Each measurement is made with one port of the network
excited by a voltage source while all the other ports are short-circuited. For example,
Y21 , the forward trans-admittance, is the ratio of the current at port 2 to the voltage at
port 1, when all other ports are short-circuited:

I2
Y21 = . (1.27)
V1 V2 =...=Vn =0
If other independent and dependent variables had been chosen, the network would have
been described, as before, by n linear equations similar to (1.24), except that the vari-
ables and the parameters describing their relationships would be different. However, all
parameter sets contain the same information about a network, and it is always possible
to calculate any set in terms of any other set [1,2].
“Scattering parameters,” which are commonly referred to as S-parameters, are a param-

eter set that relates to the traveling waves that are scattered or reflected when an n-port
network is inserted into a transmission line.
Scattering parameters were first defined by Kurokawa [3], where the assumption was
to have real and positive reference impedances Zi . For complex reference impedances,
Marks and Williams [4] addressed the general case in 1992 and gave a comprehensive
solution to it. They describe the interrelationships of a new set of variables, the pseudo-
waves ai , bi , which are the normalized complex voltage waves incident on and reflected
from the ith port of the network, defined as:
Vi + Zi Ii
ai = α {Zi } .
2|Zi |
Vi − Zi Ii
bi = α {Zi } . (1.28)
2|Zi |
where voltage Vi and Ii are the terminal voltages and currents, Zi are arbitrary (complex)
reference impedances and α is a free parameter whose only constraint is to have unitary
modulus, from now on assumed to be 1.
The linear equations describing the n-port network are therefore:
⎡ ⎤ ⎡ ⎤⎡ ⎤
b1 S11 S12 · · · S1n a1
⎢ b2 ⎥ ⎢ S21 S22 · · · S2n ⎥ ⎢ a2 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ .. ⎥ = ⎢ .. .. .. .. ⎥ ⎢ .. ⎥ (1.29)
⎣ . ⎦ ⎣ . . . . ⎦⎣ . ⎦
bn Sn1 Sn2 ··· Snn an
where by definition

bj
Sij = . (1.30)
ai a2 =...=an =0
Note that in principle each port can use a different reference Zi , and they need not be
related to any physical characteristic impedance.
The ease with which scattering parameters can be measured makes them especially
well suited for describing transistors and other active devices. Measuring most other
parameters calls for the input and output of the device to be successively opened and short-
circuited. This can be hard to do, especially at RF frequencies where lead inductance and
capacitance make short and open circuits difficult to obtain. At higher frequencies these
measurements typically require tuning stubs, separately adjusted at each measurement
frequency, to reflect short or open circuit conditions to the device terminals. Not only is
this inconvenient and tedious, but a tuning stub shunting the input or output may cause
a transistor to oscillate, making the measurement invalid.
S-parameters, on the other hand, are usually measured with the device embedded
between a matched load and source, and there is very little chance for oscillations to
occur. Another important advantage of S-parameters stems from the fact that traveling
waves, unlike terminal voltages and currents, do not vary in magnitude at points along
a lossless transmission line. This means that scattering parameters can be measured on
a device located at some distance from the measurement transducers, provided that the
measuring device and the transducers are connected by low-loss transmission lines.
The relationship between some of the most commonly used parameters can be found in
[1], which is valid for real reference impedances. When dealing with complex reference
impedances, then the corrections of [2] should be taken into account.
1.4 Microwave directional coupler
1.4.1 General concepts

Probably the most important passive component in the microwave measurement field,
the directional coupler [5], [6], is a device that can separate the incident and reflected
waves, which were described in Section 1.2.2 of this chapter.
A sketch of a generic directional coupler is shown in Figure 1.6(a). Independent of the
typology and the coupling strategy, any directional coupler is made of two transmission
lines, respectively the main line (line 1–2 in Figure 1.6(a)) and the coupled line (line
3–4 in Figure 1.6(a)).
The directional coupler properties can be understood by inspecting its S-matrix:
⎡ ⎤ ⎡ ⎤
S11 S12 S13 S14 ρ1 l1 k ι
⎢ S21 S22 S23 S24 ⎥ ⎢ l1 ρ1 ι k ⎥
⎢ ⎥ ⎢ ⎥.
⎣ S31 S32 S33 S34 ⎦ = ⎣ k ι ρ2 l2 ⎦
(1.31)
S41 S42 S43 S44 ι k l2 ρ2
Here we assume that the device is passive (reciprocal) and perfectly symmetrical, thus the
return losses of the main line and the coupled line are S11 = S22 = ρ1 and S33 = S44 = ρ2 .
If the device is well designed (well matched to the reference impedance) then ρ1 ≈ ρ2 ≈ 0.
The terms l1 and l2 represent the losses of the two transmission lines.
The term k in the S-matrix is the coupling factor (expressed in linear units). When a
signal source is connected to port 1 of the coupler and port 2 is connected to non-matched
load L , if port 3 and 4 are well matched (i.e. a3 ≈ a4 ≈0) we have
b3 = ka1 + ιa2 (1.32)
b4 = ιa1 + ka2 (1.33)
r1 r2
3 4
am bm L
s
a
1 2 b
(a) (b)
Fig. 1.6 Sketch of a generic directional coupler (a) and a directional coupler used as a reflectometer (b).
where ai and bi are, respectively, the incident and reflected waves at each i-port of the
coupler. As long as ι, the isolation factor, is kept small, b3 is proportional to a1 through
the coefficient k and b4 is proportional to a2 , through the same factor.
Coupling and isolation factors are typically expressed in dB. The directivity expresses
the ratio ι/k, in other words how much the coupler is capable of separating the incident
and reflected waves.
The different typologies of directional couplers available depend on the type of the
transmission line (see Appendix B for a quick overview) used for the main and the
coupled lines [7].
For example, in microstrip technology, the coupling between the two lines can be
realized by progressively reducing the distance between the main and coupled line or in
a “branch line” configuration (with two parallel microstrips physically coupled together
with two or more branch lines between them, placed at proper distance).
In a waveguide, coupling is typically realized with single or multiple holes along one
side or other of the guide. Coaxial couplers can be realized by manufacturing holes in
the external shields of the coaxial lines. Mixed-technology couplers are also possible,
such as waveguide-coaxial, coaxial-microstrip, etc.
In all cases the design involves finding the proper physical dimensions in order to
achieve the desired performances, in terms of:
• coupling factor,
• directivity,
• insertion losses,
• frequency bandwidth.
Depending on the application, it is also common to find 3-port directional couplers, where
one of the coupled ports (3 or 4) is typically physically terminated with a matched load.
1.4.2 The reflectometer

The directional coupler can be used as a reflectometer, as shown in Figure 1.6(b). From
(1.31), a1 and b2 are related through
b2 = l1 a1 + ρ1 a2 . (1.34)
Under the assumption that ρ1 ≈ ρ2 ≈ 0 and ι = 0, then
k
b3 = b2 (1.35)
l1
b4 = ka2 (1.36)
which means that the reflected waves at ports 3 and 4 are proportional to the reflected
wave at port 2 and the incident wave at port 2, respectively.
Thus the directional coupler can be used to physically separate the incident and
reflected waves at a certain port. There are more details on this topic in Chapter 8,
which also considers the case of a non-ideal directional coupler.
1.5 Smith Chart
The Smith Chart, shown in Figure 1.7, is a graph of the reflection coefficient in the
polar plane. Phillip H. Smith invented this chart in the 1930s [8]. Using the Smith Chart
it’s very easy to convert impedances to reflection coefficients and vice versa.
The Smith Chart represents the bilinear conformal transformation = (z − 1)/(z + 1)
where z is the normalized impedance (Z/Z0 ). In other words, the Smith Chart is the
transformation of the right part of the Z complex plane (only positive real parts of Z are
considered) into a circle, where the infinite values for the real and imaginary parts of Z
converge to the point (1, 0) on the transformed plane.
The normalization to Z0 of the Smith Chart implies, for example, that an impedance
of (30 + j10) will be plotted as (0.6 + j 0.2) on the Smith Chart, normalized to 50 .
Fig. 1.7 The Smith Chart.

The lines of the Smith chart define the loci of the constant real part of the impedance
or constant imaginary part, as shown in Figure 1.8. Constant resistance maps to circles
and constant reactance maps to arcs. In the Z plane, these would simply be vertical and
horizontal lines, respectively. Note that since the transformation is conformal, the 90◦
angles formed between these lines are also maintained in the transformed plane.
A perfect load (equal to Z0 ) occurs when equals zero, which is the center of the
Smith Chart. An open load will have a of unity and 0o (point (1,0) on the Smith Chart)
and a short load will have a of unity and 180o (point (−1,0) on the Smith Chart).
Figure 1.8 shows a Smith Chart with the constant VSWR contours. A constant VSWR
corresponds to a constant ||.
The constant impedance magnitude and phase can also be plotted on the Smith Chart
as shown in Figures 1.9(a) and (b).
Re{Γ}
Im{Ζ}=k
Z=∞
Re{Z}=k
Z=0Ω |Γ|=k
Im{Γ}
Fig. 1.8 Constant resistance and reactance lines on the Smith Chart.
Re{Γ}
|z| = 1
|z| = 0.5 |z| = 2
Im{Γ}
(a)
Re{Γ}
Pha{z} = 45º
Im{Γ}
Pha{z} = 315º
(b)
Fig. 1.9 Constant impedance magnitude (a) and constant impedance phase (b) represented on the Smith
Chart.
The Smith Chart is typically used to map the (and the impedance) across the length
of the transmission line. In the absence of losses, (l) = 0 ej 2βl , where 2βl = 4π l/λ. So
while moving along a transmission line, the moves on a circle with constant || = 0 ,
changing only its phase. Clockwise direction represents moving towards the generator
and counter clockwise represents moving towards the load. Moving from short circuit
to open circuit represents a quarter of a wavelength.
1.6 Conclusions
In this chapter the concepts of wave propagation along a transmission line, which are
important when the excitation signal frequency increases to the RF and microwave
regions, were revised.
The representation of a generic linear n-port network in terms of scattering parameters
was presented and the most important passive component for microwave measurements,
the directional coupler, was described.
There follow two Appendices, one on signal flow graphs and the other summarizing
the types of transmission lines cited in this book.
Appendix A – Signal flow graphs
Microwave networks can be analyzed using signal flow graphs and scattering parameters.
Each variable becomes a node, and each parameter becomes a branch. A branch enters
a dependent variable node and emanates from an independent variable node. Each node
is equal to the sum of the branches entering the node.
A two-port network can be presented as two parts a1 , b1 , and a2 , b2 as shown in
Figures 1.A.1. and 1.A.2.
a1
S11
b1 S12 a2
Fig. 1.A.1 Signal flow graph describing the scattering equation: b1 = S11 a1 + S12 a2 .
a1 S21 b2
S22
a2
Fig. 1.A.2 Signal flow graph describing the scattering equation: b2 = S 21 a1 + S22 a2
Thus, the complete two-port flow graph is shown in Figure 1.A.3.
a1 S21 b2
S11 S22
b1 S12 a2
Fig. 1.A.3 Signal flow graph describing a set of two-port scattering equations.
The generator and load add more nodes and branches as shown in Figure 1.A.4.
bS bG aL
1 bG = a1 a L = b2
ΓS ΓL
a G = b1 aL = a2 bL
aG
(a) (b)
Fig. 1.A.4 Generator (a) and load (b) representation with signal flow graphs.
Thus the overall flow graph can be combined as shown in Figure 1.A.5.
bS a1 S21 b2
ΓS S11 S22 ΓL
b1 S12 a2
Fig. 1.A.5 Full representation of a microwave source and load connected to an S-matrix.
Finally, some basic rules for the nodes are described in Figures 1.A.6–1.A.9.
a1 S21 a2 S32 a3 a1 S21 S32 a3
Fig. 1.A.6 Series rule.
SA
a1 a2 a1 SA + SB a2
SB
Fig. 1.A.7 Parallel rule.

S21
a1 S21 a2 S32 a2 a1 1–S21 a2 S32 a3
S22
Fig. 1.A.8 Self-loop rule.
a1 S21 a2 S32 a3
a3 = S21S32S1
S42 a4 a4 = S21S42a1
Fig. 1.A.9 Splitting rule.
Appendix B – Transmission lines types
Complete information and description of transmission line typologies is far from the
purpose of this book. The detailed formalism and modal analysis can be found in many
other books, as for example [6]. Here we summarize the basic properties of some of the
most used.
Figure 1.B.1 shows the cross sections of the most common typologies of transmission
lines, some of which cited in this book. The metal conductors are depicted in black,
while dielectric material is indicated with a dashed filling.
The coaxial transmission line (Figure 1.B.1(a)) has an inner conductor and an outer
ground shield. It supports TEM modes if the dielectric is homogenous and a DC current
can flow through such a transmission line. Closed formulas for the computation of the
characteristic impedance from the physical dimensions are available.
The circular (Figure 1.B.1(b)) or rectangular (Figure 1.B.1(d)) waveguides do not have
an internal conductor and support only TE and TM modes; DC current can flow. The
typical medium within the metal shield is air; this keeps the losses of a waveguide very
low, typically much lower than those of coaxial cable of the same length.
The stripline (Figure 1.B.1(c)) is the natural evolution of a coaxial cable when a
transmission line must be realized on a planar circuit board, or in an integrated circuit. The
inner conductor has a rectangular shape and is surrounded by a homogenous dielectric.
Like the coaxial cable, this TL supports TEM modes and DC current. Only approximated
formulas are available for the computation of the characteristic impedance; nevertheless
modern simulators (e.g. FEM) can perform such computations.
The microstrip (Figure 1.B.1(e)) is also typical of integrated circuits or PCBs; it’s very
easy to fabricate since the strip does not need to be embedded in the circuit but can be
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 1.B.1 Cross section of common use transmission lines: coaxial (a), circular waveguide (b), stripline (c),
rectangular waveguide (d), microstrip (e), coplanar waveguide (f). Dashed lines represent a
dielectric material.
fabricated by the classical exposure-etch methods. As the medium is non-homogenous,

the supported modes are quasi-TEM.
Finally, the coplanar waveguide (Figure 1.B.1(f)), also typical of PCBs and ICs, has
two ground shields placed at the side of the main central line. The ground potential is
achieved by drilling some via-holes, to reach the bottom ground conductor. CPWs reduce
the crosstalk between different lines on the same circuit. The drawback is that the top
ground shields must be kept at the same zero potential along all the line length: via-holes
must be then placed with proper spacing all along the line length.
Table 1.B.1 summarizes the basic properties of these TLs.
Table 1.B.1 Basic properties of the most common transmission line types
Closed
Formulas
Name Figure Type DC supported Available
Coaxial 1.B.1 (a) TEM Yes Yes

Waveguide 1.B.1 (b–d) non-TEM No Yes [6]
Microstrip 1.B.1 (e) quasi-TEM Yes No [9], [10]
Stripline 1.B.1 (c) TEM Yes No [11]
Coplanar waveguide 1.B.1 (f) quasi-TEM Yes Yes [6], [10], [12]
References
[1] D. A. Frickey, “Conversions between S, Z, Y, h, ABCD, and T parameters which are valid for
complex source and load impedances,” IEEE Trans. Microw. Theory and Tech., vol. MTT-42,
no. 2, February 1994.
[2] R. B. Marks and D. F. Williams, “Comments on ‘Conversions between S, Z, Y, h, ABCD,
and T parameters which are valid for complex source and load impedances’,” IEEE Trans.
Microw. Theory and Tech., vol. 43, no. 4, April 1995.
[3] K. Kurokawa, “Power waves and the scattering matrix,” IEEE Trans. Microw. Theory and
Tech., vol. MTT-13, no. 2, March 1965.
[4] R. Marks and D. Williams, “A general waveguide circuit theory,” Journal of Research of the
National Institute of Standards and Technology, vol. 97, no. 5, September–October 1992,
pp. 533–562.
[5] H. A. Wheeler, “Directional Coupler,” U.S. Patent 2 606 974, issued 12 August 1952.
[6] R. E. Collin, Foundations for Microwave Engineering. 2nd Edition, NewYork: McGraw-Hill,
1992.
[7] L. Young, Parallel Coupled Lines and Directional Couplers. Dedham, MA: Artech House,
1972.
[8] P. H. Smith, “Transmission line calculator,” Electronics, vol. 12, no. 1, pp. 20–31, January
1939.
[9] H. A. Wheeler, “Transmission-line properties of a strip on a dielectric sheet on a plane,” IEEE
Trans. Microw. Theory Tech., vol. MTT-25, pp. 631–647, Aug. 1977.
[10] K. C. Gupta, R. Garg, I. J. Bahl, and P. Bhartia, Microstrip Lines and Slotlines, 2nd Edition.
Dedham, MA: Artech House, 1996.
[11] H. Howe, Stripline Circuit Design, Dedham, MA: Artech House, 1974.
[12] T. Q. Deng, M. S. Leong, and P. S. Kooi, “Accurate and simple closed-form formulas for
coplanar waveguide synthesis,” Electronics Letters, vol. 31, is. 23, pp. 2017–2019, November
1995.
2 Microwave interconnections,
probing, and fixturing
Leonard Hayden
2.1 Introduction
In this chapter concepts related to connecting test equipment to a device-under-test are

explored. Application-specific definitions of device boundaries and measures for signal
path and power-ground performance are introduced. Practical measurement system accu-
racy implications of fixture losses are examined, with the surprising result that sometimes
more fixture loss can be beneficial to measurement precision. An introduction to the basic
elements of microwave probing and probing applications concludes the discussion.
2.2 Device boundaries and measurement reference planes
It is necessary to clearly define the boundary of the target of a measurement (known

generally as the Device Under Test, or DUT), to distinguish it from the test system –
fixture, probes, or other interconnections. The DUT can take many forms. It could be a
functional block in a housing with connectors or waveguides for the inputs and outputs.
Or a circuit DUT could be an embeddable semiconductor functional design element
with a standard interface point such as a microstripline or other transmission lines. At
the other extreme, the DUT could be a circuit component such as a transistor, inductor,
capacitor, or resistor with no interface elements other than the constituent electrical
contacts. Somewhere in between is the fully distributed circuit element. In all cases, the
measurement reference planes define the boundaries of the DUT; see Figure 2.1.
For the purposes of discussion let us broadly and perhaps arbitrarily assign DUTs into
three categories: Devices, Transmission Lines, and Circuits.
Test
Equipment
Fixture
Test System
Measurement
Reference Plane
DUT
Fig. 2.1 The measurement reference plane is the dividing line between the test system, including test
equipment and fixturing, and the Device Under Test.
22 Leonard Hayden
2.2.1 Devices
This category of DUT includes the components making up, for example, an integrated cir-
cuit design. When electrically small, these devices may have essentially lumped element
behavior, changing to lumped element with parasitics, and to fully distributed behavior
as the electrical size grows with frequency. Often, it is desirable to consider devices with
distributed behavior in the transmission line category of DUT.
Device models describe the behavior of the circuit component either with a functional
black-box with a network parameter description, through an equivalent circuit made
up of a topology of ideal components arranged to mimic the device behavior, or some
hybrid of the two. A so-called “compact” circuit model of a CMOS transistor can easily
exceed 100 parameters defining the functioning behavior and combines lumped element
topology modeling with special mathematical expressions.
The goal of a device model is to predict the performance of a circuit from a theoretical
array (in a circuit simulator) of the devices and topology of the design. The measure of
success of a device model is the degree of its success for the required application. In
modeling applications, the generality of a model is compromised for efficiency and the
model is always created with a context or range of applicability in mind. Device modeling
measurements, likewise, are scaled and evaluated based upon the application needs.
2.2.2 Transmission lines

Transmission lines, distributed circuits, and other interconnection elements, such as
adapters, pose a particular challenge for a measurement system. The desired electrical
behavior can approach the ideal with close to no insertion loss and no reflections. Mea-
surement attempts are often limited by the residual uncertainty of the system itself. For
example the mismatch measured at the input of an ideal transmission line is the mismatch
of the termination; or, in calibrated systems, the residual error in the characterization of
the termination mismatch.
The location of the measurement reference planes is critical for transmission lines. In
modeling a transmission line, one of the most fundamental properties is the propagation
constant – the attenuation and delay/phase per-unit-length. For accurate calculation of the
per-unit-length normalization, the physical distance between the measurement reference
planes must be known accurately.
For this class of measurement, the reference planes must be in the middle of uni-
form sections of the transmission line: see the later discussion of Thru-Reflect-Line
(TRL) and Line-Reflect-Line (LRL) network analyzer calibration methods. These meth-
ods support a DUT embedded in a well-defined transmission line environment that
effectively becomes part of the DUT as far as the measurement is concerned. But
using such line-surrounded definitions is not well suited to device modeling. Moving
the reference planes to the physical boundaries of a device can be a challenging de-
embedding or modeling exercise. The embedding transmission line behavior can change
as it nears a discontinuous transition to the device due to the complex electromagnetic
interaction.
Microwave interconnections, probing, and fixturing 23
2.2.3 Circuits
The simplest conceptual measurement case is the basic amplifier (or passive two-port
device) with a coaxial connector at the input and output; see Figure 2.2. The reason these
circuits seem simple is that they directly connect to instrumentation and, in the simplest
cases, the only measurement concern is transmission gain or attenuation in a 50 ohm
environment, either as a frequency response or pulse response in the time domain. This
measurement can be made either with a pulse source and an oscilloscope or with a swept
frequency signal generator and a power meter. A direct connection of the input to the
output provides the input reference excitation that can be removed/normalized from the
measured response to isolate the DUT behavior.
Adding one level of complexity raises concern about the input and output match, per-
haps characterized as a standing wave ratio (SWR); see Figure 2.3. When a circuit is
always tested and used in an environment supplying perfect 50 ohm terminations, then
DUT reflection behavior is simply a contributor to the transmission response. A scalar
value of the mismatches allows an estimation of the uncertainty bounds on the transmis-
sion behavior associated with the DUT and system or test fixture mismatches interacting
with each other. Adding a reflectometer or SWR bridge to the scalar measurement
transmission test system facilitates scalar mismatch measurement.
t t
DUT
f f
Fig. 2.2 Simple circuit transmission behavior.
t t
DUT
f f
Fig. 2.3 Circuit behavior with reflections.

24 Leonard Hayden
Adding yet another level of complexity, we can consider the signal distortion in wide
bandwidth signals caused by frequency-dependent delay (dispersion/deviation from lin-
ear phase). Often the absolute delay is not a concern, but dispersion can significantly
change a pulse shape through increased rise-time and the ringing of a step response.
For this case, adding a vector voltmeter to our previously scalar measurement system
would work, but the more general answer is to use a Vector Network Analyzer (VNA)
with two or more port signal switching, directional couplers (or bridges), wide dynamic
range receivers, and “calibration” software. These features allow a VNA to clearly define
the measurement reference planes and computationally remove the fixture behavior from
the measurement.
For well-behaved linear systems, the VNA measurement is capable of fully charac-
terizing the DUT behavior at the VNA calibration reference planes, independent of the
non-idealities of the measurement system. While it is a very powerful instrument, the
VNA has limitations and there are measurement situations that add further complexity:
multi-mode excitations, nonlinear DUT gain sensitivity to port terminations, frequency
conversion and intermodulation products, etc.
2.3 Signal-path fixture performance measures
At the most fundamental level, a fixture would be most ideal if it could be electrically
represented as a node, i.e. a connection with zero loss, delay, or other signal or impedance
impacts. This is a useful concept in the regime where physical dimensions are small
compared to the electrical wavelengths and lumped element approximations are usable
for circuit work.
2.3.1 Delay
The roughly meter-long cables used in bench-top network analysis equipment fail to
behave as nodes for all but the lowest of frequencies. Delay or phase-shifts become
a behavior that must be considered. Linearly increasing phase-shift with frequency, or
constant-delay, has the least impact, as signals can propagate without distortion – a
complex waveform will retain its shape from input to output of the interconnection.
The alternative dispersive propagation, frequency-dependent propagation delay, causes
changes in the wave-shape as the various frequency components change relative position
in time due to unequal delays.
2.3.2 Loss
Loss in the form of a uniform attenuation with frequency preserves the relative signal
wave-shape (it is modified only by a scale factor), but it can limit the measurement
system dynamic range as attenuated signals grow closer to the floor of uncertainty, due
to various types of noise or correlated or uncorrelated interfering signals.
Frequency-dependent loss is present in interconnections when the skin-depth, i.e. the

conductivity and frequency-dependent dimension of the electromagnetic penetration of
current distribution in a metal, becomes smaller than the cross-sectional dimensions of
the interconnection. For the lower range of frequencies where the skin-depth exceeds the
conductor dimensions, the series resistance and commensurate attenuation are essentially
constant as the entire cross-sectional area carries current. At increasing frequencies, the
depth of penetration effectively limits the cross-sectional conductor area carrying current,
resulting in series resistance that is proportional to the square root of the frequency. This
behavior is known as skin-effect loss.
The skin-effect loss behavior occurs for frequencies above which the conductors are
no longer fully penetrated by the current distribution. The transition to skin-effect loss
behavior may be observed in the MHz frequency range for larger dimension structures
like coaxial cables used in instrumentation. However, for small dimension interconnec-
tions such as thin-film traces on ceramic or semiconductor substrates, the conductors are
very thin and the transition to partial penetration may occur as high as several GHz.
Dielectric loss is also a component of interconnection loss, particularly for lower
quality materials such as common FR-4 circuit boards. Measurement systems tend to
avoid using material with significant dielectric loss contribution, frequency-dependence,
or resonant behavior over the frequencies of use.
2.3.3 Mismatch
Even a lossless, constant-delay transmission line can contribute to signal attenuation and
distortion. In any distributed system, impedance mismatch will cause signal reflections,
and pairs of mismatches work to cause a multiply reflected signal path to combine with
the direct signal path resulting in a frequency dependence of loss (as the path-length
varies between constructive and destructive interference) and a phase-shift (when the
path-length combines to cause signal leading or lagging). Example reflection magnitude
values are shown in Table 2.1.
The single mismatch section example suggests that a severe mismatch is required to
cause a significant non-ideality. Indeed, for a significantly mismatched 40 ohm section in
a 50 ohm environment, the attenuation ripple amplitude is only about 0.2 dB, the deviation
from the linear phase is less than 1 degree, and the peaks in return loss exceed 30 dB.
However, a real interconnection system may have several regions of mismatch each
contributing such a response. With the phase coherence of constructive interference,
the reflections add as voltages, so the worst-case combination of two 30 dB return
loss non-idealities is 6 dB higher or a 24 dB return loss. As the number of transitions
increases, the performance can rapidly degrade with each contributor of non-ideal match
(all combinations of impedance discontinuity interactions).
Estimation of lumped parasitic values – a practical tool

Transmission line theory can be helpful in estimating the behavior of electrically
short sections of interconnections such as bond-wires, trace-overlap regions, or socket
elements. While transmission line equations represent l and c on a per-unit-length basis,
26 Leonard Hayden
Table 2.1 Impedance vs. reflection coefficient, SWR, and return

loss relative to a 50 ohm environment.
Impedance (ohms) SWR Return Loss (dB)
200 0.60 4.00 −4.4

150 0.50 3.00 −6.0
120 0.41 2.40 −7.7
100 0.33 2.00 −9.5
80 0.23 1.60 −12.7
75 0.20 1.50 −14.0
70 0.17 1.40 −15.6
65 0.13 1.30 −17.7
60 0.09 1.20 −20.8
55 0.05 1.10 −26.4
54 0.04 1.08 −28.3
53 0.03 1.06 −30.7
52 0.02 1.04 −34.2
51 0.01 1.02 −40.1
50 0.00 1.00 –
49 −0.01 1.02 −39.9
48 −0.02 1.04 −33.8
47 −0.03 1.06 −30.2
46 −0.04 1.09 −30.2
45 −0.05 1.11 −25.6
40 −0.11 1.25 −19.1
35 −0.18 1.43 −15.1
30 −0.25 1.67 −12.0
25 −0.33 2.00 −9.5
20 −0.43 2.50 −7.4
10 −0.67 5.00 −3.5
5 −0.82 10.00 −1.7
with a constraint of electrically short regions we can use these same equations for total
inductance L and total capacitance C. The characteristic impedance and total delay of
the electrically short section are then given by:

L
Z0 = (2.1)
C
√
T= L·C (2.2)
and the total inductance and capacitance in terms of impedance and total delay are then:
L = T · Z0 (2.3)
T
C= (2.4)
Z0
For any particular trace region, an estimate of the physical length and approximate
dielectric coefficient is enough to determine a surprisingly accurate estimate of the total
delay T . Transmission lines created using coax, stripline, or microstrip require extremely
small or large physical dimensions to realize particularly high or low impedances. Going
much above 100 ohms or much below 10 ohms becomes difficult and the impedance
extreme is readily apparent from the extreme dimensions. It becomes possible to look
at structures and have a good idea of the impedance value to better than an order of
magnitude – or even as close as a factor of 2 or 3. For example, a very wide trace over
a ground plane is likely to be close to an estimate of 10 ohms, while a thin wire in
air far from ground might be approximated by 100 ohms. Using estimated delay and
impedance values, the total inductance and capacitance are computed using (2.3) and
(2.4) with reasonable accuracy.
For example, consider a narrow bond wire 200 μm above a conductive plane and
1.2 mm long in air. Air dielectric has a propagation velocity of 3 × 108 m/s, or 300 μm/ps
for convenience at this scale. The total delay of the 1200 μm length is then 4 ps to high
precision. Characteristic impedance equations or a cross-sectional simulator could be
used for best accuracy, but a safe guess for the impedance of the bond wire is of the order
of 100 ohms. This estimate is expected to be well within a factor of two of the actual
value. The impedance is safely above 50 ohms, since a typical 25 μm diameter bond wire
in air would have to be much closer than 25 μm to realize 50 ohms. And, as a bond wire
becomes very far from a ground plane, the impedance curve flattens out making 200 ohms
a likely maximum to achieve. The total inductance estimate is then 400 pH and the less
significant capacitance is 40 fF. These values should be within about a factor of two of
the actual values, providing an easy to obtain and often very useful estimate.
2.3.4 Crosstalk
Ideally, multiple interconnections do not electrically interact and closed transmission
line structures such as coaxial cables exhibit very low crosstalk. However, open inter-
connection structures such as parallel microstriplines on circuit boards, integrated circuit
packaging, or other interconnections will exhibit crosstalk. Crosstalk effects are cumula-
tive and grow with the complexity of the circuit (e.g. number of inputs and outputs), but
even a single input, single output circuit such as a transistor test fixture can be sensitive
to input/output port coupling when used for device modeling or critical performance
metrics such as Fmax derived from Mason’s gain.
Crosstalk is particularly difficult to remove from a measurement because crosstalk
mechanisms in test fixtures can have electrical behavior dependent on a DUT. The
thought experiment of a parasitic capacitive coupling between input and output in a two-
port measurement system easily demonstrates this. The current through the capacitor
depends on the dynamic voltage. If the DUT measured is low-impedance, low voltages
are present and capacitive crosstalk is small. A high-impedance DUT maximizes the
capacitive crosstalk.
Particularly subtle and often difficult to identify is the measurement “suckout” which
is often related to crosstalk. A suckout is a frequency response magnitude aberration
28 Leonard Hayden
with a characteristically narrow and small dip in transmission – often of the order of
0.1–1 dB in depth. One cause for a suckout is when an otherwise clean transmission line
is lightly coupled to an adjacent transmission line without terminations. The adjacent
open-ended line acts as a high-Q, half-wavelength resonator that is only lightly loaded
by the coupling to the primary signal path. Energy is sucked away from the signal path
at the peak of the resonance.
2.3.5 Multiple-modes
Measurement systems often expect interconnections to only allow a single propagating
mode at a network port. Network analysis and S-parameter theory depend upon this
and when a physical interconnection has significant energy in another mode, then this
mode must be mathematically separated and considered as effectively an additional
network port.
The problem with multiple-modes at a reference plane is with the transmission line
definition of the port. Non-degenerate modes propagate at different velocities, creating
an interaction pattern with distance that is not compatible with a propagation constant
description. Over electrically short transitions this cannot happen and the effect does not
have an impact.
The consistent summation of all mode behavior at a location may be used to instead
define a port voltage and current allowing network modeling where this non-distributed
behavior is appropriate. This is the situation for an abrupt transition – there is no single
propagation behavior; multiple-modes or even higher-order electromagnetic coupling
exists, but the region with this behavior is electrically short allowing voltage-current
based circuit modeling to effectively describe the transition non-idealities.
The conductor-backed finite ground coplanar waveguide is an example of such a multi-
mode path; see Figure 2.4. The three conductors along with the backside ground plane
allow three modes of propagation: the desired coplanar waveguide mode with outbound
Coplanar Microstrip
Mode Mode
–0.1 +0.9 –0.1 +0.5 +0.5 +0.5
Slotline
Mode
+0.5 0 –0.5
Fig. 2.4 Propagating modes in conductor-backed coplanar waveguide with finite ground conductor width.
current in the signal trace and equally split return currents in the ground traces; microstrip
mode where all traces carry outbound symmetric currents and the return current is in the
ground plane; and slot-line mode where the signal trace is current-free and the ground
traces carry equal currents in opposing directions.
2.3.6 Electromagnetic discontinuity

In an electrical transition between two different transmission lines, such as between a
coaxial cable and a microstripline, the discussion above shows that maintaining a constant
impedance is necessary to avoid mismatch losses. But even if every cross section along the
transition has the same equivalent impedance, does this guarantee optimal performance?
When approaching this problem from the electrical circuit or even distributed circuit
approach the answer would seem to be yes, but these are only approximations to the
physical world modeled by Maxwell’s electromagnetic equations. Examining the electric
field patterns for coaxial cable and microstrip shows two radically different shapes. An
abrupt transition can be made between the two but it will incur a mismatch associated with
the mismatched fields. The simple transmission line approximation fails to predict this.
Eisenhart [1] used the concept of an electromagnetic transition as a way to optimize
behavior and minimize mismatch effects. A continuously varying cross section is used
that progressively shifts the electrical field patterns from the radial pattern of the coax
case to the vertical with fringing microstrip field. When done over a distance that is
electrically long at frequencies where the electrical discontinuity would otherwise be of
significance, an optimal transition is created; see Figure 2.5.
Fig. 2.5 The Eisenhart launcher creates a continuous field transition between coaxial cable and
microstripline, minimizing mismatches.
30 Leonard Hayden
2.4 Power-ground fixture performance measures
Power and ground path performance often has greater impact on a measurement of circuit
performance than non-ideal signal paths. Circuits expect a “stiff” or low-impedance
supply with a constant voltage independent of the dynamic current draw. Ground is
expected to be an equipotential reference everywhere it is used.
De-embedding or compensation of the impact of a non-ideal power or ground is not
something that has been demonstrated. Unlike the signal path where correction may be
possible, it is essential that the power and ground paths are optimized as much as practical
in our measurement system. In some cases the best test is obtained when the power and
ground connections are identical to what would be used in application. Often this means
using the original application circuit with most connections retained, but modified to
allow microwave measurement launches to the signal input and output.
2.4.1 Non-ideal power

If circuits could function nominally when supplied DC power from a 50 ohm source,
then this would be a very short topic. But many active circuits will suffer degraded
gain, dynamic range, linearity, and increased tendency to oscillate with a high-resistance
power supply. The latter concern requires not just a low and controlled-impedance power
supply over the frequencies where there is signal energy, but the impedance must be well
controlled anywhere there is gain and the possibility of oscillation.
Bypass capacitors are the key components for reducing power bounce, as they have
the tendency to stabilize voltage by acting as a reservoir of charge to supply surges in
current. A perfect bypass would have a very large capacitance immediately connected
across the DUT power leads with sufficient charge storage to hold a constant voltage
across it for the expected dynamics of current. But in practice, large value capacitors are
physically large and must be located further away from, for example, the tiny pads of an
integrated circuit. This distance can pose a problem.
Consider the impedance looking back into that poor 50 ohm line used as a power
source. For reasonable construction and to a reasonable tolerance, 50 ohms will be seen,
and it will be constant over low to high frequencies. This is not good enough, so the first
thought is to add a large shunt capacitor along our 50 ohm line so large that the reactance
is very close to zero even at low frequencies. But the impedance seen by the circuit
looking back into the power supply is only small at frequencies where the capacitor is
electrically close to the circuit. When this delay approaches one-quarter wavelength, the
impedance goes through a maximum, becoming an open circuit in the absence of loss; see
Figure 2.6. For these frequencies this condition is far worse than the matched 50 ohm line.
One approach to fixing the problem of standing waves on a power interconnect would
be to use the matched line approach, but with a much lower impedance line. Feasibility
can be a concern though, since to reach, for example, a 10 ohm impedance can require an
impractically wide trace. The use of power planes is an approach that does benefit from
this concept, but isn’t available when considering sockets, integrated circuit packages,
or wafer probing measurement cases.
50
+ Z = 50 DUT
– o
|Zps|
50
+ c DUT
– Zo = 50
Fig. 2.6 When electrically far from a bypass capacitor undesired extremes of impedance occur.
Z Z
L C L1
C1 L2
C2
L
L2C1
log |Z| log |Z|
C C2 L2 C1 L1
L2C2 L1C1
Log freq
log freq
Fig. 2.7 A distributed power system makes use of progressively larger capacitors at progressively greater
distances from the DUT.
A distributed power system is often the answer to situations with significant regions
of non-zero impedance. Instead of using a single, very large capacitor (value and size) a
set of progressively larger valued capacitors is used – with a very small value very close
to the DUT, a moderately valued capacitor at a moderate distance from the DUT, and a
large valued capacitor at a large distance from the DUT until finally the low-frequency
behavior is controlled by the feedback circuits in the power supply that maintain an
arbitrarily low impedance.
Careful study of Figure 2.7 leads to insight into the behavior of the power system. In
each section the transmission line behavior is modeled well as an inductor, since they are
electrically short for the frequency components that can get past the capacitors closer to
the DUT.
At the highest frequencies, C1 behaves as a short circuit and the equivalent behavior
is simply that of L1 . As the frequency is lowered, a series resonance of L1 and C1 occurs
creating an impedance minimum. Below this minimum the impedance grows with the
behavior dominated by C1 . Looking past C1 we see L2 terminated by C2 . C2 is large for
this frequency range and acts as a short.
32 Leonard Hayden
An undesired parallel resonance between L2 and C1 creates an impedance maximum

limited only by losses. Adding resistance in series with C1 will load the parallel resonance
and reduce the magnitude of the maximum impedance. Below the parallel resonance, L2
will dominate the response to a series resonance with C2 causing a second impedance
minimum and dominating the response at even lower frequencies.
It is interesting to note that with this kind of distributed power structure, loss is required
in the transmission lines or the capacitors to avoid peaks of impedance due to high-Q
parallel resonance. Bypass capacitor manufacturers often feature their extremely low
series resistance, when this is not always an advantage.
2.4.2 Non-ideal ground

Grounding in circuits has been one of the black arts of analog RF and microwave circuit
design [2]. Conceptually a ground is a common potential reference, independent of the
current distribution present in it. In practice a ground plane must be of sufficiently low
impedance that the return currents imposed upon it do not impact circuit performance.
If this condition cannot be met, then the equipotential requirement of a ground reference
is not met.
In many applications a ground plane is sufficient as a ground reference. However,
high power amplifiers running low-voltage swing and high currents can easily exhibit
the effects of non-ideal grounding where different spots on the plane have different
potentials.
Conversely, a fully differential circuit design works on the principle of a virtual ground
where net ground currents sum to zero. For differential circuits the behavior is indepen-
dent of the ground impedance. Zero current means zero voltage drop and the ground has
a single potential level.
Evaluating the impedance of a ground path is particularly troublesome. For a socket,
integrated circuit package, or probing environment a reasonable approach can be taken
by using a transmission measurement where a signal transmission path is shorted by the
imperfect ground connection; see Figure 2.8.
i 50 Ω
Lg
V 50 Ω
di
V = Lg
dt
Signal For small Lg
Ground
Signal
Fig. 2.8 Direct measurement of ground impedance is a difficult problem; the shunt imperfect short
method shown allows a reasonable inference of the equivalent ground inductance.
2.5 Fixture loss performance and measurement accuracy
As a general rule it is not bad to assume that higher performance fixtures make for more
accurate measurement results. This is true when the measurement and DUT reference
planes don’t coincide due to limitations in our ability to calibrate our system. In these
cases de-embedding or modeling techniques are used to identify and remove the inter-
vening interconnection behavior. Usually, the more ideal this element, the easier it is to
identify.
In the ideal case the losses and mismatches of the fixture are small enough that they
may be ignored, but in all cases interconnection delay (even if distortion free) is present
and may need to be identified. Practical interconnections have losses. Loss is usually split
between conductor and dielectric loss. Other loss mechanisms are possible, but effects
like radiating loss are considered something to avoid in a measurement system, since the
energy may be going to unpredictable locations.
Conductor losses are often dominant in test fixtures, particularly when good qual-
ity dielectrics are used (e.g. air, alumina, semi-insulating GaAs, SiO2 , etc) and when
conductor cross-sectional dimensions are small and resistance is high. Both propagation
constant and characteristic impedance may vary with extreme loss values, but in low-loss
cases the mismatch effect may be small enough to ignore. Fundamental transmission line
theory [3] tells us how the line impedance is determined from per-unit-length r, l, g, and
c (resistance, inductance, admittance, and capacitance, respectively).

r + j ωl
Zo = , (2.5)
g + j ωc
where ω is the radian frequency given by 2πf .

When g is small compared to ω. c, the case for low dielectric loss to surprisingly low
frequencies, then at higher frequencies where r is small compared to ω. l we find that Zo
simplifies to the real and often constant high-frequency approximation:

l
Zo = . (2.6)
c
In this region, the loss of the interconnection no longer contributes to mismatch, as the
inductive reactance and capacitive susceptance dominate the behavior.
In some measurement systems loss can be a stabilizing factor. In vector network
analysis, the measurement system dynamic range is often sufficient to tolerate some
signal loss without significant impact, and calibration effectively corrects for the fixture
loss in the signal path. When an imperfect measurement system has electrically separated
elements with high reflections, the interaction resonance errors may be noticeable in the
frequency response. Similarly, a suckout caused by coupling to an adjacent unterminated
interconnection can create a sharp resonance response.
Corrections need repeatability in the measurement system to be applicable. A small
temperature change will cause physical and electrical length changes in the metal
interconnect conductors, moving the resonances and invalidating the calibration; see
34 Leonard Hayden
Calibration New Corrected

measurement Correction measurement measurement
Immediately
after cal: + =
After drift or
system change + =
occurs:
(Drift) (Wrong)
Fig. 2.9 A small change in fixture electrical length can cause any resonances to move in frequency,
invalidating the calibration.
Figure 2.9. The presence of loss between the interacting reflections or in the coupled
adjacent line will load the resonance causing the effect to be shallower and broader. The
calibration will be imperfect but still helpful over a broader temperature range.
Mechanisms other than temperature change will contribute to frequency shifts in
the small bumps inherently present in a non-ideal measurement system. Sensitivity to
changes from cable bending or twisting is also mitigated by the presence of loss.
Adding loss to a VNA port has a significant drawback, however. The port directivity,
the ability to distinguish between the incident and reflected waves, is reduced by twice
the attenuation added to the port. For any specific measurement system the solution to
the trade-off for optimal loss will differ. A well-matched, suck out-free system with high-
quality phase stable cables with minimal bending and twisting during use will work best
with minimal fixture loss; while, as reported in [4], a broadband probing measurement
system using a poorly matched combiner (between a coaxial low-band and a waveguide
high-band), benefits from some degree of loss to stabilize the response from the probing
discontinuity and combiner interaction. Loss improves the raw source and load match of
the system, and the impact on directivity is not important since the waveguide couplers
have very high natural directivity and some degradation can be tolerated.
2.6 Microwave probing
Before the concept of a monolithic microwave integrated circuit became viable hybrid
circuits were common practice. These used passive components patterned on a ceramic
substrate and individual transistors connected via bond wires. Larger value resistors
and capacitors were attached through bonding or direct attach (epoxy or solder). Test
fixtures were often of a similar concept using coax-to-microstrip launches and bond
wires connecting the DUT to the microstrips; see Figure 2.10. For best performance, the
electromagnetically tapered launch developed by Eisenhart was used.
Microwave probes [5] were significant to the development of microwave integrated
circuits. Probes enabled much more accurate transistor models due to smaller and more
consistent launch structures and the small tip geometry facilitates precise calibration. The
Thru DUT
Fig. 2.10 Use of microstrip launchers to characterize a microwave component. Hybrid circuits bonded
transistors and capacitors to ceramic substrates and connected them using thin-film interconnects
and resistors.
Coax-Connector
Mount
EM Absorber
EM Absorber
EM Absorber
Coax cable
Contact fingers
Fig. 2.11 Probe cross section showing typical microwave probe features.
ability to probe devices, test patterns, and circuits while on an undiced, unpackaged wafer
reduces the design, fabricate, test cycle time speeding up iterations for improvement.
2.6.1 Probing system elements

A typical microwave probe consists of a coaxial connector, a coaxial cable, and a tip
made up of multiple contact fingers; see Figure 2.11. The probe design is optimized to
make the size transition from the 1–3 mm diameter coax connector to the 100–500 μm
tip width while maintaining a constant 50 ohm impedance.
For best electrical repeatability, the probe tip is designed with contacts that maintain a
fixed pitch (center-to-center spacing); see Figure 2.12. Signal path shielding is optimized
when ground contacts are immediately adjacent on both sides. Electromagnetic fields
are most confined in this configuration and crosstalk to nearby traces or probes is also
minimized.
Unlike simple needles, a microwave probe tip is an array of points making contact
with multiple pads simultaneously. These pads can be as small as 50 μm square so
a precision positioning system is required. A typical microwave probe station uses a
high power microscope for viewing and the probe is driven in x-y-z by a three-axis
36 Leonard Hayden
Poor
Better
Best
Fig. 2.12 High-performance microwave probes used precisely fabricated fixed spacing contacts in a
ground-signal-ground (GSG) configuration. The double ground contact provides better
electromagnetic shielding of the signal path and the shorter fingers minimize the impedance and
field discontinuity.
micro-positioner. Often the positioner arm provides a roll axis adjustment so that the
plane of the probe contact fingers can be oriented to be parallel to the plane of the probe
pads – so that all contacts touch with equal force (planarizing).
Balanced and controlled tip forces and the proper choice of tip material are necessary to
ensure constant and repeatable contact resistance when probing pads with an aluminum
surface. A thin barrier layer of aluminum oxide, known as native oxide, forms over
exposed aluminum pads. Probe contacts must penetrate this insulating layer to make
electrical contact with the aluminum underneath. Special probe versions optimizing
performance on aluminum pads are available from microwave probe vendors. The extra
cost of these premium probes is avoided if gold-plated pads are available in the process
(common in GaAs, but rare in silicon foundries).
2.6.2 VNA calibration of a probing system

Calibration standards for probes can be relatively simple to fabricate. A nearly ideal low-
inductance short-circuit is made by a metal shorting bar for the contacts. A probe elevated
in air or landed on open pads behaves like it is capacitively terminated (with negative
and positive value, respectively). Constant value lumped models (inductor or capacitor)
are often used; these are much simpler than the polynomial of frequency dependence
combined with delay offset used to model coaxial short and open calibration standards.
A match termination for the GSG configuration uses equal 100 ohm thin-film resistors
from the signal to each ground in parallel to provide a well-defined 50 ohm load. These
resistors are symmetrically trimmed to tight tolerance with a laser or other mechanism.
The electrical model of these standards is simply 50 ohms in series with a lumped,
frequency-independent inductor. Coaxial standards normally assume perfectly matched
loads, but a probe load requires an extra parameter describing the inductance of the load.
Unlike the case for complementary gender coaxial connectors, no direct thru connec-
tion of probes is possible. Instead, a transmission line of known electrical behavior is
fabricated on the calibration substrate. By keeping the length of this structure very short
electrically (e.g. 1 ps) uncertainty in the impedance or accuracy of the delay parameter
has minimal impact on the measured S-parameters of a DUT.
Actual standard electrical behavior varies with specific positional placement of the
probes on the standards. A change in probe to pad overlap of even a few microns is
discernible in the calibration result. Alignment marks are fabricated and help the probing
operator to dock probes into a precise separation. With the aid of the reference and high-
power optics, better than 5 μm placement repeatability can usually be achieved. Using a
probe station separation stroke the probes are lifted together and landed on simultaneous
shorts, loads, or a thru standard. Visible symmetry helps to ensure that both probes
overlap equally.
The calibration method plays a part in the sensitivity of the calibration to probe
placement. Traditional Short-Open-Load-Thru (SOLT) calibration requires fully known
electrical behavior of the standards and is most sensitive to probe placement error caused
by variation from the definitions.
The enhanced Line-Reflect-Reflect-Match (eLRRM) with automatic determination of
load inductance calibration method [6] uses the same physical standards as SOLT, but
does not require the shorts and opens to be anything other than symmetric on the two
ports. The inductance of the load/match standard (only one port’s match measurement
is required) is determined with redundancy in the calibration data. The load need only
be modeled well by an R-L equivalent circuit (where R is the known low-frequency
value, usually 50 ohms). The eLRRM calibration is far less sensitive to imprecise probe
placement than the SOLT calibration method.
In probing two-port standards, such as thru or line structures, the structure orientation
and geometry must match the probes. Having a straight thru standard is of no help if the
probes are oriented orthogonally or on the same side of a DUT.
This requirement of geometric compatibility between two-port standards and ports is
not a requirement in coaxial setups, where the cables may be reoriented as needed to
complete the standard measurement. Conversely, the coaxial setup does have to worry
about connector gender. These differences in important issues, along with the desire to
automate calibration steps, has created a need for specialized calibration software for
microwave probing [7]. This software goes beyond the mainly coaxial and rectangular
waveguide calibration support provided in most commercial vector network analyzers.
2.6.3 Probing applications – in situ test

While microwave probes were developed for the specific needs of testing devices and
circuits while still in wafer form, their advantages can also be applied in areas more
familiar to the days of the hybrid circuit. Small ceramic elements containing a GSG
probe interface with a transition to microstrip are commercially available [8]. These
probe launches can be bonded to the input and output of a compatible circuit that is
otherwise in its native environment (power, ground, and other connections).
2.6.4 Probing applications – transistor characterization

A measurement of a transistor (or other small device) has, as its goal, the behavior of
the intrinsic device (the device without pads). Measurement after probe tip calibration
38 Leonard Hayden
provides the extrinsic device behavior (the device with pads). Separate measurement and
modeling of the device pads allows their mathematical removal. The most commonly
used version of this process uses two-steps and is known as Y-Z de-embedding.
In the first step of Y-Z de-embedding, the lumped, frequency-dependent Pi-model
Y-parameters of the parasitic shunt parasitics are determined from a measurement of
the open pads (device removed). In the second step, the frequency-dependent T-model
Z-parameters of the series parasitics are determined from a measurement of the shorted
pads (device replace by a short of both signals to ground). Simple subtractions of Y and
Z parameters facilitate the de-embedding [9].
2.7 Conclusion
This chapter has provided some key concepts and tools related to device fixturing and
performance and boundary determination. There are many more topics that could be
explored in much greater depth – particularly as the unique needs of specific applications
are considered.
References
[1] R. L. Eisenhart, “A Better Microstrip Connector,” Microwave Symposium Digest, 1978 IEEE-
MTT-S International, pp. 318–320, 27–29 June 1978.
[2] E. Holzman, Essentials of RF and Microwave Grounding. Norwood, MA: Artech House, 2006.
[3] Philip C. Magnusson, et al., Transmission Lines and Wave Propagation, 4th ed., Boca Raton,
FL: CRC Press, 2001.
[4] Agilent Technologies application note 5989–1941 [Online]. Available: www.agilent.com.
[5] Eric Strid, “A History of Microwave Wafer Probing,” ARFTG Conference Digest-Fall, 50th,
vol. 32, pp. 27–34, Dec. 1997.
[6] L. Hayden, “An enhanced Line-Reflect-Reflect-Match calibration,” ARFTG Conference, 2006
67th, pp. 143–149, 16 June 2006.
[7] WinCal T M Calibration Software [Online]. Available: www.cascademicrotech.com.
[8] ProbePoint T M Adapter Substrates [Online]. Available: www.jmicrotechnology.com.
[9] M. C.A. M. Koolen, et al., “An improved de-embedding technique for on-wafer high-frequency
characterization,” Bipolar Circuits and Technology Meeting, 1991, Proceedings of the 1991,
pp. 188–191, 9–10 Sep. 1991.
Part II
Microwave instrumentation
3 Microwave synthesizers
Alexander Chenakin
3.1 Introduction
A frequency synthesizer is the most versatile piece of microwave equipment. Synthesizers

come in a variety of forms ranging from tiny chips to complex instruments. Single-chip
synthesizers are available in a die form or as surface-mount ICs. They include the
key elements (such as RF and reference dividers, phase detector, and lock indicator)
required to build a simple single-loop synthesizer. Such ICs are mounted on a PCB
with additional circuitry. The PCB-based modules range from small, surface-mount,
“oscillator-like” designs to more complex connectorized assemblies. Such PCB assem-
blies can be packaged into a metal housing and are presented as stand-alone, complete
synthesizer modules. Connectorized synthesizer modules can be used to build larger
instruments such as signal generators for test-and-measurement applications.
Not surprisingly, frequency synthesizers are among the most challenging of
high-frequency designs. Many approaches have been developed to generate clean output
signals [1–17]. This chapter presents a brief overview of today’s microwave synthesizer
technologies. It starts with general synthesizer characteristics followed by a review of
the main architectures. Direct analog, direct digital, and indirect techniques are com-
pared in terms of performance, circuit complexity, and cost. Synthesizer parameters can
be further improved in hybrid designs by combining these main technologies and tak-
ing advantage of the best aspects of each. Finally, sophisticated test-and-measurement
signal generator solutions are reviewed. The signal generators come with high-end tech-
nical characteristics and extended functionality including output power calibration and
control, frequency and power sweep, various modulation modes, built-in modulation
sources, and many other functions.
3.2 Synthesizer characteristics
A frequency synthesizer can be treated as a black box that translates one (or more)
input frequency (usually called reference) to a number of output frequencies as shown
in Fig. 3.1. This black box contains various components such as VCOs, frequency mul-
tipliers, dividers, mixers, and phase detectors, which being properly connected, perform
the desired translation function. Although all synthesizers exhibit significant differences
as a result of specific applications, they share basic core characteristics depicting their
42 Alexander Chenakin
frequency and timing, spectral purity, and output power parameters. These core charac-
teristics are reviewed below. Other specifications (not listed here) may include AC or DC
power consumption, control interface, mechanical and environmental characteristics as
well as some special functions such as modulation and output power control.
3.2.1 Frequency and timing

Operating frequency range denotes the range of frequencies that can be generated by
the synthesizer. It is specified in the units of Hz (MHz, GHz) by indicating the minimum
and maximum frequencies generated by the synthesizer.
Frequency resolution or step size is the maximum frequency difference between two
successive output frequencies, indicated in Fig. 3.1 as f = fn+1 − fn . The operating
frequency range and frequency resolution are fundamental synthesizer specifications
set by a particular application. Some applications (e.g. test-and-measurement) require
very wide coverage and fine frequency resolution while others need a relatively narrow
bandwidth with a rough step size or just a single fixed frequency.
Frequency accuracy indicates the maximum deviation between the synthesizer’s set
output frequency and its actual output. Frequency accuracy is normally determined by
the reference signal, which can be internal or external to the synthesizer. Frequency
synthesizers usually employ a crystal oscillator as an internal reference. The crystal
oscillator’s temperature stability and aging are important characteristics that define the
synthesizer’s frequency accuracy. Temperature stability denotes the maximum frequency
drift over the operating temperature range and is usually expressed in ppm (the term ppm
is an acronym for parts-per-million – a dimensionless coefficient equal to 10−6 ). Aging is
a change in frequency over time that occurs because of changes in the resonator material
or a buildup of foreign material on the crystal. It is also specified in ppm over a certain
period of time. Aging leads to a permanent frequency error; thus, it is good practice to
Frequency resolution Switching time

Δf t
In Out
Synthesizer
fREF fMIN fMAX

Frequency range
Output power
Phase noise Spur
Fig. 3.1 Frequency synthesizer core characteristics.

Microwave synthesizers 43
use mechanical or electronic frequency adjustment means to compensate for internal

reference aging.
Switching or tuning speed determines how fast the synthesizer transitions from one
desired frequency to another and is defined as time spent by the synthesizer between these
two states. Thus, the switching time is a more proper term (indicated as t in Figure 3.1).
The switching time is calculated from the time when the synthesizer receives a command
to the time it approaches the desired frequency with a specified accuracy.
3.2.2 Spectral purity

Harmonics appear in the synthesizer spectrum as integer multiples of the output fre-
quency because of signal distortion in nonlinear components. For example, if the
fundamental frequency is represented by f , the frequencies of the harmonics would
be represented by 2f , 3f , etc. Harmonics are expressed in dBc (decibels relative to the
carrier) and represent the power ratio of a harmonic to a carrier signal, as depicted in
Figure 3.2. For narrow-band synthesizer harmonics can be easily controlled by adding a
simple low-pass filter. However, in test-and-measurements applications, a tunable filter
or a switched filter bank is required since the synthesizer bandwidth normally exceeds
an octave.
Sub-harmonics appear at frequencies that are “sub-harmonically” related to the main
signal such as f/2, f/3, etc. Propagating through nonlinear components, these signals
exhibit their own harmonics. Thus, in a more general case, the sub-harmonics are consid-
ered as products appearing at N/K of the output frequency, where Nand K are integers.
Sub-harmonics can be created in some nonlinear devices such as a frequency doubler.
The doubler generates a number of harmonics of the incoming signal. It usually employs
a balanced scheme that intends to suppress odd products, as shown in Figure 3.3. Since
the second harmonic now becomes the main signal, all the odd products (which are not
completely suppressed) do not meet the harmonic relationship with respect to the desired
output and are, therefore, treated as sub-harmonics.
Spurious signals or spurs are undesired artifacts created by the synthesizer at some
discrete frequencies that are not harmonically related to the output signal. As a typical
example, Figure 3.4 shows reference spurs often created in PLL synthesizers. Other
spurs can come from many other sources such as mixer intermodulation products,
P (dBm)
ΔP (dBc)
f 2f 3f
Fig. 3.2 Harmonics appear at integer multiples of the output frequency.

Main tone
Sub-harmonics
f 2f 3f 4f
f Frequency
doubler
Fig. 3.3 Sub-harmonic products.
fREF ΔP (dBc)
Fig. 3.4 Reference spurs of a PLL synthesizer.
local oscillator leakages, and external signals coming through the bias or control inter-
face. In contrast to harmonics, the spurs are much more troublesome products that can
limit the ability of receiving systems to resolve and process a desired signal. Spurs
can sit very close to the main tone and in many cases cannot be filtered. Thus, the
spurious level has to be minimized, typically to −60 dBc relative to the main sig-
nal, although many applications require bringing this level even lower. This presents
a certain design challenge, especially if a small step size is required. A different con-
cern is mechanically induced spurs, usually referred to as “microphonics.” These spurs
appear due to the sensitivity of certain synthesizer components to external mechanical
perturbations.
Phase noise is one of the major parameters that ultimately limits the performance of
microwave systems. In general, phase noise is a measure of the synthesizer’s short-
term frequency instability, which manifests itself as random frequency fluctuations
around the desired tone. The output of an ideal synthesizer is a pure sine-wave signal
VOU T = A0 sin ω0 t with amplitude A0 and frequency ω0 = 2π f0 . However, in reality
the output signal demonstrates amplitude and phase variations (Figure 3.5), which can
be represented as follows
VOU T = (A0 + a(t)) sin(ω0 t + ϕ(t)), (3.1)

Amplitude fluctuations
Phase fluctuations
Fig. 3.5 Amplitude and phase fluctuations.
where a(t) and ϕ(t) are the amplitude and phase fluctuations, respectively. Amplitude
noise is rarely as critical as phase noise. The amplitude variations can be easily reduced
by balanced mixers, amplifiers in compression, diode limiters, or an automatic level
control circuit. Hence, the phase effects generally dominate, reducing (3.1) to VOUT =
A0 sin(ω0 t + ϕ(t)).
These phase fluctuations result in uncertainty in the signal zero-crossing, which in the
time domain is referred as jitter. Assuming that the phase fluctuations ϕ(t) are caused
by an unwanted fixed-frequency signal ωm = 2π fm that modulates the synthesizer out-
put frequency and is expressed as ϕ(t) = Am sin ωm t, then the output signal can be
described by
VOUT = A0 sin(ω0 t + Am sin ωm t)
and using the well-known trigonometric identity sin(α + β) = sin α cos β + cos α sin β
is further transformed to
VOUT = A0 [sin ω0 t cos(Am sin ωm t) + cos ω0 t sin(Am sin ωm t)]. (3.2)
Assuming that the amplitude of the modulating signal Am is small, we can simplify the
corresponding terms of (3.2) to
cos(Am sin ωm t) ≈ 1
sin(Am sin ωm t) ≈ Am sin ωm t
reducing (3.2) to
VOUT ≈ A0 (sin ω0 t + Am cos ω0 t sin ωm t). (3.3)
1
Using another elementary trigonometric formula, sin α cos β = [sin(α + β) +
2
sin(α − β)], (3.3) is further modified to
Am
VOUT ≈ A0 {sin ω0 t + [sin(ωm t + ω0 t) + sin(ωm t − ω0 t)]}
2
and finally
A0 Am A0 Am
VOUT ≈ A0 sin ω0 t + sin(ω0 + ωm )t − sin(ω0 − ωm )t. (3.4)
2 2
Note that (3.4) has three sinusoidal terms related to ω0 , ω0 − ωm , and ω0 + ωm . Thus,
in the frequency domain, the output signal is no longer a single spectral line but adds two
spurious sidebands equally spaced by fm (below and above the main signal). Obviously,
if fm is not a fixed frequency but changes randomly, the sidebands also spread randomly
over frequencies both above and below the nominal signal frequency. Phase noise can be
quantified by measuring the output power at many frequencies away from the nominal
frequency and comparing it to the power at the nominal frequency, as illustrated in
Figure 3.6. This leads to a quantitative definition of the phase noise as the ratio of the
noise power found in a 1-Hz bandwidth at a certain frequency offset f to the total
power at the carrier frequency f0 , which can be written as

Pf10H z
+f
L = 10 log (3.5)
Pf0
This ratio is normally taken in the logarithmic scale; hence, the phase noise is expressed
in units of dBc/Hz (dBc per hertz) at various offsets from the carrier frequency and is
usually specified by a table or as a graphic representation.
ΔP
1 Hz
Δf
f0 f
Fig. 3.6 Phase noise is quantified by measuring the output power at many frequencies away from the
nominal frequency and comparing it to the power at the nominal frequency.
3.2.3 Output power

Output power is a measure of the synthesizer output signal strength specified in units
of watt or, more frequently, in dBm. The term dBm refers to the ratio in decibels of the
measured power referenced to one milliwatt. The relationship between these two units
is expressed by
PdBm = 10 log PmWatt . (3.6)
A simple synthesizer usually delivers a fixed power level that cannot be changed. More
complex designs provide an ability to control the output power in a specified range. In
the latter case, the output power control range (i.e. the minimum and maximum values
between which power can be set) and the power step size (i.e. the minimum change
between two consecutive power settings) are specified as well. Note that output power
can differ from its set value. This discrepancy is described by the output power accu-
racy that defines the absolute maximum variance between programmed and actual (i.e.
measured) power values.
3.3 Synthesizer architectures
Synthesizer characteristics depend heavily on a particular architecture. Synthesizer archi-

tectures can be classified into a few main groups, as shown in Figure 3.7. The direct
architectures are intended to create the output signal directly from the available input
frequency signals either by manipulating and combining them in the frequency domain
(direct analog synthesis) or by constructing the output waveform in the time domain
(direct digital synthesis). The indirect methods assume that the output signal is regen-
erated inside the synthesizer in such a manner that the output frequency relates (e.g. is
phase-locked) to the input reference signal. A practical synthesizer, however, is usually
a hybrid design that combines various techniques to achieve specific design goals.
3.3.1 Direct analog synthesizers

The direct analog synthesizer is today’s most advanced technique, offering exceptional
tuning speed and phase-noise characteristics. The output signal is obtained by mixing
input frequencies followed by switched filters, as shown in Figure 3.8. These input
frequencies can be created by mixing, dividing, and multiplying the output of low-
noise fixed-frequency oscillators [1–5]. The output frequency change is accomplished by
Direct analog Direct digital Indirect (PLL)

synthesizers synthesizers synthesizers
Hybrid architectures
Fig. 3.7 Synthesizer architectures.

f1
f2
f3
f4
f5
f6
Fig. 3.8 Basic block diagram of a direct analog synthesizer.
f01
f02
f03
f0N
f11 f12 f13 f1N f21 f22 f23 f2N
Fig. 3.9 The number of output frequencies is increased by cascading individual mixer stages.
switching appropriate input frequencies; thus, tuning speed is only limited by propagation
delays inserted by the switches and their control circuits. Phase noise mainly depends on
the noise of the available fixed-frequency sources and can potentially be very low. The
main disadvantage of this simple scheme is the limited frequency coverage and step size.
In our example, only eighteen output frequencies can be generated (even by utilizing
both mixer sidebands). The number of output frequencies can be increased by cascading
individual mixer stages, as shown in Figure 3.9. However, this rapidly increases the
design complexity and overall component count.
The frequency resolution can also be improved by repeatedly mixing and dividing
the input frequencies, as conceptually shown in Figure 3.10. The synthesizer contains a
chain of frequency mixer-divider cells that generate a signal at
fi f1 f2 fi
fOUT = = f0 + + + ··· + i , (3.7)
Ni N N2 N
i=0
where fi is a frequency driving the corresponding mixer and N is the division coefficient
of the utilized frequency dividers. Using proper fixed frequencies and a sufficient number
of individual cells, an arbitrarily small step size can be achieved. In general, the frequency
fN
÷N
fN
÷N
f1
fOUT
÷N
f0
Fig. 3.10 Mixing and dividing technique.
division coefficients can also be arbitrary; however, N = 10 is the most commonly used
scenario that leads to
f1 f2 fi
fOUT = f0 + + + ··· + i . (3.8)
10 100 10
The frequencies fi are usually generated from a common reference F0 by utilizing its
harmonics, i.e. fi = Ai F0 , where Ai is an integer between 1 and 9. This allows rewriting
the synthesizer tuning formula to

A1 A2 Ai
fOU T = F0 A0 + + + ··· + i , (3.9)
10 100 10
where the decimal coefficients Ai simply show which harmonic is chosen. Therefore, the
output frequency is conveniently represented in a decimal form by setting corresponding
digits. Similarly, the synthesizer can be constructed using different frequency division
coefficients to represent its output frequency in a binary or any other desired form, or a
combination thereof.
The main disadvantage of the direct analog synthesizers is the large number of mixing
products that have to be filtered. These include the undesired mixer sideband, LO leakage,
and intermodulation products. Depending on a particular frequency plan, filtering close-
in spurs can be a challenging task. Another serious issue is cross-coupling between
individual filter channels and whole stages. Although a large variety of mixing and
filtering schemes are possible, they tend to be hardware intensive if a small frequency step
and wide coverage are required. Therefore, while direct analog synthesis offers excellent
fCLOCK
FDDS
Phase Look-up
DAC LPF
accumulator table
W
Fig. 3.11 Direct digital synthesizer concept.
tuning speed and phase-noise characteristics, its usage is limited to applications where
a fairly high cost can be tolerated.
3.3.2 Direct digital synthesizers

Direct digital synthesizers utilize digital signal processing to construct an output signal
waveform in the time domain piece-by-piece from an input (called clock) signal [1–6].
The direct digital synthesizer includes a phase accumulator, digital look-up table, DAC,
and LPF as depicted in Figure 3.11. The phase accumulator allows the entering and
storing of a digital word W called the phase increment. At each clock pulse, the phase
accumulator adds (i.e. accumulates) the phase increment to the previously stored digital
value that represents an instantaneous digital phase of the generated signal. This digital
phase is continually updated until it reaches the capacity of the accumulator.
For an N-bit accumulator and the smallest increment of one least significant bit, it
will take 2N clock cycles to fill up the accumulator. Then the accumulator resets and the
process starts over again. Hence, the lowest generated frequency is given by
fCLK
fMI N = (3.10)
2N
that also equals the smallest frequency step. With a larger phase increment W , the phase
accumulator obviously fills up faster and the DDS output frequency increases to
W
fDDS = fCLK . (3.11)
2N
Therefore, frequency tuning is accomplished by changing the phase increment word.
This word defines the DDS output frequency and can be loaded into the accumulator
through either a serial or parallel interface. The tuning process has essentially no settling
time delays other than what is inserted by the digital interface. The frequency can be
changed in very small steps determined by the length of the phase accumulator. For
example, assuming that fCLK is 100 MHz and N equals 32, we can calculate fMIN to
approximately 0.023 Hz. The length of the phase accumulator can be further increased;
thus, millihertz or even microhertz steps are easily achievable.
The next step is to convert the digital phase value into a digital representation of the
signal waveform. This is accomplished with a look-up table. It uses a ROM to store a
digital code that sets a proper address on the DAC’s bus, and consequently, its output
voltage. In general, any desired waveform can be created; however, the sine wave is
most commonly used. The waveform construction process completes with a low-pass
10
–10
–20
–30
–40
–50
–60
–70
–80
–90
Fig. 3.12 DDS output contains a number of spurious signals.
filter required to remove some unwanted spurious components because of the imperfect
approximation of the desired waveform.
Practical realization of this concept brings further modifications. For example, the
length of the phase accumulator, required to achieve the necessary resolution, can exceed
practical limits for ROM and DAC devices. Due to the sine function’s symmetry, only
one-fourth of the cycle needs to be stored, thus, greatly reducing the required memory
capacity. Furthermore, the DAC usually utilizes a smaller number of bits available from
the phase accumulator. This reduction in DAC resolution is called phase truncation and
leads to increased spurious levels.
The DDS output contains a number of spurious signals (Figure 3.12) as a result of
truncation, amplitude quantization, and DAC nonlinearities. However, the most signifi-
cant ones are aliased images of the output signal that appear on either side of the clock
frequency and its multiples because of the sampling nature of digital signal synthesis.
From this point of view, the DDS works as a frequency mixer producing spurs at
fSPUR = ±n fCLK ± m fDDS , (3.12)
where n and m are integers. Similar to mixer intermodulation products, these spurs
require careful frequency planning, since they can be very close to the output signal and,
therefore, cannot be filtered. While spur location in the frequency domain can be easily
determined, its amplitude is much less predictable. As a general rule, lower-order spurs
are the strongest; although, fairly high-order spurs can still be harmful and must be taken
into account. Typical DDS spurious levels are −50 to −60 dBc for output signal ranges
between a few tens to a few hundreds of megahertz.
The DDS also provides reasonably low phase-noise levels, even showing an improve-
ment over the phase noise of the clock source itself. From the phase-noise point of view,
the DDS works as a fractional frequency divider with a very fine, variable, frequency divi-
sion coefficient. The phase-noise improvement is described by the 20 log (fCLK /fDDS )
function and is limited by the residual noise floor.
The DDS is currently available as a tiny, yet highly integrated, surface-mount IC that
includes the phase accumulator, look-up table, and DAC in a single chip. It needs only
a few external components (LPF and bias circuitry) to build a powerful and versatile
module. The most valuable DDS feature is its exceptionally fine frequency resolution
and fast switching speed that is comparable to direct analog schemes. The main disad-
vantages are limited usable bandwidth and relatively poor spurious performance. While
a DDS starts working from nearly DC, its highest frequency is limited within one half
of the clock frequency according to the sampling theory. It is theoretically possible to
use DDS aliased images above the one half of the clock limit; however, the spurious
content is further degraded. As a rule of thumb, the usable DDS bandwidth is limited
to about 40% of the clock signal by practical LPF design considerations. Typical clock
speeds for today’s commercial DDS ICs are in the range of a few hundred megahertz
to a few gigahertz. The DDS technique is rarely used alone at microwave frequencies
because of the previously mentioned bandwidth and spurious limitations. Rather DDS is
used as a fine-frequency-resolution block in conjunction with direct analog and indirect
architectures.
3.3.3 Indirect synthesizers

Indirect frequency synthesizers utilize phase-lock loop techniques offering a smaller
step size and lower complexity in comparison with direct analog schemes [9–17]. A
typical single-loop PLL synthesizer (Figure 3.13) includes a tunable voltage-controlled
oscillator that generates a signal in a desired frequency range. This signal is fed back to
a phase detector through a frequency divider with a variable frequency division ratio N .
The other input of the phase detector is a reference signal equal to a desirable step size.
The phase detector compares the signals at both inputs and generates an error voltage,
which following filtering and optional amplification, slews the VCO until it acquires the
lock frequency given by fOUT = N fPD , where fPD is the comparison frequency at the
phase detector inputs.
fREF fPD
÷R fOUT
PD VCO
fPD = fOUT/N
÷N
Fig. 3.13 Single-loop PLL synthesizer.

The frequency tuning is achieved in discrete frequency steps equal to fP D by changing

the division coefficient N. The available reference frequency can be divided down by
another divider to reduce the step size. If the division coefficient of the reference divider
is R, then the output frequency is set by
N
fOUT = fREF . (3.13)
R
Since the output signal of the PLL synthesizer is generated at microwave frequencies,
all spurs associated with the direct architectures are generally absent. The only source of
the spurs in the PLL block diagram shown in Fig. 3.13 is the reference signal itself. The
reference signal and its harmonics modulate the VCO tuning port and create sidebands
both above and below the main signal. The loop filter bandwidth has to be significantly
lower than fP D (usually ten times or more) to keep the reference spurs at a reasonable
level. However, the loop bandwidth is inversely proportional to the settling time. Thus,
achieving fine frequency resolution, low spurs, and fast switching is an arduous task as
it means balancing mutually exclusive terms.
Another important consideration and design tradeoff is phase noise. The noise outside
the PLL filter bandwidth is mainly determined by the VCO’s free-running noise. The
phase noise within the loop filter bandwidth is given by
LPLL = L PD + 20 log N, (3.14)
where L P D is the cumulative phase noise of the reference signal, reference and feed-
back dividers, phase detector, LPF, and loop amplifier recalculated to the phase detector
input. The phase noise generated by PLL components is degraded by the large division
ratios required to provide a high-frequency output with a fine resolution. Moreover, pro-
grammable dividers are usually not available at high frequencies; thus, an additional,
fixed-division-coefficient divider (called a prescaler) is required. In this case, the total
division ratio increases by the prescaler division coefficient resulting in further phase
noise degradation. At high frequency offsets, the VCO’s free-running noise can be (and
normally is) better than the multiplied PLL noise. The optimal phase-noise profile is
achieved by choosing the loop bandwidth at the cross point of the multiplied PLL noise
and VCO free-running noise curves, as depicted in Figure 3.14. Clearly, by utilizing a
low noise VCO and narrower loop bandwidth, it is possible to mask some excessive
PLL noise at high-frequency offsets. However, this results in a slower switching speed.
Alternatively, a good PLL design can suppress VCO noise at higher-frequency offsets
and also provide faster tuning.
Overall, the major advantages of the PLL schemes are reduced levels of spurious
signals resulting from the low-pass filter action of the loop and a much less complex
compared to the direct analog architectures. In fact, all key PLL components can be
integrated into a single chip that leads to low-cost, miniature designs. The main dis-
advantages are slower tuning, limited step size, and considerably higher phase noise
compared to direct analog and direct digital architectures.
Phase noise
20logN
fOPT Frequency offset
Fig. 3.14 Phase noise of a single-loop PLL synthesizer.
fREF
÷ R1 fOUT
PRD1 VCO1
÷ N1
÷ R2
PRD2 VCO2
÷ N2
Fig. 3.15 Multiloop synthesizer concept.
3.3.4 Hybrid architectures

A practical synthesizer is usually a hybrid design that combines various techniques to
achieve specific design goals.A good example is a multiloop synthesizer that is essentially
a combination of direct analog (frequency mixing) and indirect (PLL) techniques. The
idea is to convert the VCO output to a much lower frequency with the aid of a mixer and
an offset frequency source, as illustrated in Figure 3.15. The second PLL generates an
auxiliary signal, which is used as an offset signal for the first loop. Splitting the design
in two loops allows for the reduction division coefficients in both loops, thus, improving
the overall phase-noise performance.
fCLOCK
DDS fOUT
PD VCO
M1 M2 MN
xK1 xK2 xKN
÷ N1 ÷ N2
fOFFSET
Fig. 3.16 Hybrid design combines three frequency synthesis techniques.
Another interesting design employs a chain of mixers converting the VCO signal to a
lower frequency, as depicted in Figure 3.16. This scheme allows for the minimization,
or even complete removal, of the frequency divider from the PLL feedback path that
results in very low phase noise and low spurious performance [18]. Note that the local
oscillator offset signals are created from a common source utilizing frequency mixing,
division, and multiplication. In other words, these offset signals are created by direct
analog synthesis means. The fine frequency step is achieved by adding a DDS. Thus, this
design combines all three main frequency synthesis techniques (i.e. direct analog, direct
digital, and indirect) to achieve high performance and extended functionality [19].
3.4 Signal generators
Although synthesizers can be found in virtually any microwave system, test-and-

measurement is probably the most challenging application that calls for advanced
synthesizer solutions. Broadband operation, very fine frequency resolution, low spurs,
and low phase noise are the key specifications for signal generator instruments. Another
important parameter that impacts overall test-and-measurement system throughput is the
switching speed. The time spent by the synthesizer transitioning between frequencies
becomes increasingly valuable since it cannot be used for measurement and data process-
ing. Besides these key parameters, signal generators feature extended functionality that
includes output power calibration and control, frequency and power sweep, list mode,
and various modulation functions [20].
3.4.1 Power calibration and control

For a simple synthesizer design, the output power can vary across the operating frequency
range because of individual component gain variations. More sophisticated designs bring
fIN fOUT
Control
DAC
Fig. 3.17 Open-loop power control.
fIN fOUT
–
Control
DAC
Fig. 3.18 Better performance is achieved with a closed-loop control.
the ability to equalize the output power response and also change the power level as
needed. A synthesizer’s output power can be controlled in many different ways, for
example, using an open-loop technique illustrated in Figure 3.17. The amplitude con-
trol circuit includes an RF attenuator and a DAC. The DAC generates a proper voltage
for any given frequency to ensure a flat output response across the entire operating
frequency range. The DAC values are generated during a calibration routine and are
stored in a look-up table. The output power can be changed within certain limits (set by
the available attenuator dynamic range), adding one more dimension to the calibration
table. Furthermore, the synthesizer output circuit may include many devices that exhibit
temperature variations. Thus, the synthesizer may also include a temperature sensor to
provide further correction if required. By employing a sophisticated interpolation rou-
tine, this technique provides reasonably flat and repeatable output power characteristics
across operating frequency and temperature ranges. Note that the output power is set
almost instantaneously. Therefore, the open-loop method is well suited for fast switch-
ing applications. The main disadvantage of this method is limited accuracy caused by
component temperature variations. The output power delivered to an external load also
depends on how well the synthesizer and load impedances are matched.
Better performance can be achieved with a closed-loop ALC method. The output
power is sampled with a directional coupler and routed to an RF detector, as depicted
in Figure 3.18. The detector generates a voltage proportional to the output power. This
voltage is compared to a reference voltage generated by a DAC. An error signal controls
the attenuator, thus, closing the loop. In other words, the RF detector continuously
measures and adjusts the output power to a value set by the DAC. This configuration
ensures a precise output power level regardless of the load mismatch. Furthermore,
temperature variations of the synthesizer components are also taken into account. The
only significant source of temperature instability is the RF detector itself (and – to a
smaller degree – the directional coupler). Temperature variations of the detector are
further reduced by controlling (i.e. stabilizing) its temperature. The power control range
can be further extended by adding an electromechanical step attenuator.
3.4.2 Frequency and power sweep

It is often desired to linearly change the output frequency within certain limits. This func-
tion is called frequency sweep and is widely used in test-and-measurement applications
for characterizing various devices across their operating frequency range. The frequency
sweep is defined by setting start frequency, stop frequency, and sweep time. There are
two basic modes – continuous and stepped sweeps. The continuous (analog) sweep is
realized by changing the VCO tuning voltage directly (i.e. breaking the phase-lock-loop)
with a sawtooth generator or a DAC. This mode requires linear and repeatable tuning
characteristics; however, a frequency error is always present since the oscillator remains
unlocked during the sweep. The stepped (also called discrete or digital) frequency sweep
is realized by changing the synthesizer output frequency in discrete frequency increments
(steps) as illustrated in Figure 3.19. It assumes that the synthesizer is locked at every
discrete point across the sweep range. Hence, this mode provides significantly better
frequency accuracy compared to the analog sweep.
Similarly, synthesizer output power can be swept between desired power levels. This
function (called power sweep) is used in the characterization of output power and linearity
characteristics of various devices such as transistors, amplifiers, mixers, and many others.
The synthesizer output power can be swept continuously or in steps.
The frequency (or power) sweep mode normally assumes that the synthesizer steps lin-
early in equal increments. However, frequency and power can be set to any desired value.
A list mode provides better flexibility. The idea is to create a table (list) of frequencies and
Continuous sweep
f
fSTOP
Stepped sweep
fSTART
Fig. 3.19 Frequency sweep methods.

store it in the synthesizer’s memory. The list is executed by sending a proper command
or by applying a trigger signal (which is a voltage pulse) to a corresponding control line.
Once the synthesizer’s control circuit detects a trigger pulse, it commands the synthesizer
to move from one frequency to another according to the programmed list. Alternatively,
the synthesizer can go to the next frequency, stop there and wait for the next trigger pulse;
then, the process repeats. One of the advantages of the list mode is a significant through-
put improvement compared to normal programming, since it is possible to precalculate
and memorize all necessary parameters required to control individual components of the
synthesizer.
3.4.3 Modulation
Signal generators utilize various modulation forms ranging from simple pulse, ampli-
tude, frequency, and phase modulation to complex digital modulation formats. The most
commonly used modulation modes are reviewed below. Further details on modulation
theory and implementation techniques can be found in [1], [9], [20], [21].
Pulse modulation
Pulse modulation is achieved by switching the output signal on and off in accordance
with the applied modulating pulses. The result is a sequence of RF pulses that replicate
(or tend to replicate) the input modulating signal as shown in Figure 3.20. The minimum
RF pulse width, rise time, fall time, and overshoot are important characteristics that
define how well the modulating signal is replicated. Typical rise time and fall time
numbers required are in the order of ten nanoseconds. The pulse modulation on/off ratio
is another critical parameter. A typical specification is 80 dB or higher. The modulating
signal frequency (also called rate) can be between DC and several megahertz.
Amplitude modulation
Amplitude modulation historically has been one of the most popular methods for carrying
information via RF frequencies. It is realized by varying the output signal amplitude in
accordance with an applied modulating signal, as indicated in Figure 3.21. The simplest
Fig. 3.20 Pulse modulated signal.

Fig. 3.21 Amplitude-modulated signal.
Fig. 3.22 Frequency-modulated signal.
way to implement AM is to control the insertion loss of an attenuator inserted into the
synthesizer output circuit. The maximum power variation (which can also be expressed
in terms of modulation index or depth) is achieved by setting the output power level
in the middle of its control range. Another important requirement is linearity, since the
modulator must translate the modulating signal with minimal distortion. This may further
limit a realizable modulation depth. Various linearization techniques can be applied to
minimize AM signal distortion.
Alternatively, amplitude modulation can be implemented by summing the modulating
signal into the ALC loop. In general, the ALC-based amplitude modulation offers better
linearity and repeatability characteristics. However, the modulation depth may be limited
by the available ALC dynamic range which, in turn, depends on the detector that is used.
The maximum modulating signal rate is also lower compared to the open-loop alternative
because of the settling time of the closed-loop ALC system.
Frequency and phase modulation

Frequency modulation is another popular form of analog modulation that offers bet-
ter signal immunity than AM. The process of producing a frequency-modulated signal
involves the variation of the synthesizer output frequency in accordance with the modu-
lating signal, as shown in Figure 3.22. The frequency bandwidth where the synthesized
signal fluctuates is proportional to the peak amplitude of the modulating signal and is
called frequency deviation. FM can also be described by the modulation index, which is
the ratio of the maximum frequency deviation to the frequency of the modulating signal.
Note that we can vary not only the frequency but also the phase of the synthesized
signal, thus, producing PM. Both processes are quite similar since in both cases we vary
the argument (the angle) of the same sine function. Hence, the angular modulation is a
more general case that represents both FM and PM. FM and PM modulated signals can be
produced in many different ways. For example, it is possible to modulate the synthesizer’s
VCO tuning voltage around the value where it is settled. The problem, however, is that the
PLL will tend to correct any voltage change. For proper operation, the modulating signal
rate has to be well above the PLL filter bandwidth. Typical achievable modulating rates
range from a few kilohertz to tens of megahertz. An alternative solution is to modulate
not the VCO but the reference oscillator. Furthermore, a higher deviation can be achieved
by changing not the reference frequency itself but rather its phase by inserting a variable
phase-shifter in the reference signal path. If the modulating signal rate is sufficiently
low, the PLL will track the reference frequency (or phase) change and, hence, translate
the modulation to the VCO output. The loop filter bandwidth should be set as wide as
possible to allow higher modulating rates. Typical rates start from nearly DC to a few
tens of kilohertz. Thus, these two modes complement each other and can extend the
overall modulating frequency range.
Digital modulation
Modern communication systems migrate from simple analog modulation to more sophis-
ticated digital modulation techniques. Note that more effective modulation forms are
possible by simultaneously varying both amplitude and phase. The simplest way to
visualize such a complex signal is to draw it as a vector on a polar diagram. The
amplitude and phase are represented as the length and the angle of the vector, as
shown in Figure 3.23. In digital communication systems, such a signal is expressed
in I (in-phase) and Q (quadrature) terms, which are projections of the signal vector on a
corresponding orthogonal axis. Therefore, the amplitude and phase modulation assumes
the change of the signal vector, which can be conveniently accomplished by varying two
Amplitude
Phase
Q
Fig. 3.23 More effective modulation forms are possible by simultaneously varying both amplitude and
phase.
IIN
0º
LOIN RFOUT
Hybrid Σ
90º
QIN
Fig. 3.24 Basic diagram of an IQ-modulator.
independent IQ-components. Hence, such a complex modulation is also called vector or

IQ-modulation.
Vector modulation can be realized with an IQ-modulator. It consists of two identical
mixers driven with a 90-degree phase-shift at their LO ports, as shown in Figure 3.24.
The base-band data signals are applied directly to the IF ports, upconverted and summed
together with no phase-shift between them. The resulting output is an IQ-modulated
signal at the same carrier frequency as the LO. The quality of the synthesized signal
can be tested by applying two base-band signals of the same frequency and amplitude
with a 90-degree phase-shift with respect to each other. For a perfect modulator, only
one sideband should be present. However, in reality, the output signal contains another
sideband because of imperfect amplitude and phase balance. Furthermore, an LO leakage
also takes place. The undesired sideband can be suppressed (cancelled) by adjusting the
amplitude and phase of the applied IQ signals. The LO leakage can be controlled by
adjusting DC offset voltages for the diodes used in the balanced mixers. Therefore, it
is generally possible to calibrate the modulator characteristics to a degree where it can
be practically used. The difficulty is that this calibration has to be implemented at many
frequencies across the entire operating range. Moreover, the calibration has to survive
over time and temperature changes. Thus, achieving a good image and LO leakage
suppression for a broadband, high-frequency, direct IQ modulator is a challenging task.
An alternative solution is to create a desired IQ-modulated signal at a lower, fixed
frequency and upconvert it to microwave frequencies. Obviously, it is much easier to
achieve better cancellation of undesired products at a single (and lower-frequency) point.
However, the difficulty is now to remove undesired up-conversion products in a wide
frequency range. This can be accomplished with a YIG-tuned filter. The disadvantage of
the YIG filter is slow tuning speed and relatively narrow passband that can be insufficient
in certain applications. A switched filter bank offers better tuning characteristics. How-
ever, it usually requires a large number of channels and, hence, is hardware extensive.
Nevertheless, IQ-modulation is a very desirable function in modern signal generator
instruments.
3.5 Conclusions
Frequency synthesizers are among the most challenging of high-frequency devices. The
industry feels persistent pressure to deliver higher-performance designs. Broadband oper-
ation, fine frequency resolution, low spurs, low phase noise, and fast switching speed are
desirable characteristics. Another challenge is size and cost reduction. In the past, com-
plex microwave synthesizers were often built using individual connectorized modules
connected with coaxial cables. The designer could easily isolate and refine individual
blocks to make them perfect. These days, such complex assemblies have to be made on
a common PCB using tiny surface-mount parts. A great effort is required to minimize
interactions between individual components sitting on the same board. Furthermore,
many parts are reused to accomplish different functions, which are distributed through
the whole assembly. The net result is a significant increase in “design density,” meaning
both component count and functionality per square inch.All these factors drastically com-
plicate the design process. Nevertheless, this seems to be a “must” approach these days.
References
[1] V. Manassewitsch, Frequency Synthesizers: Theory and Design, 3rd ed. NJ: Wiley, 2005.
[2] V. F. Kroupa, Frequency Synthesis: Theory, Design and Applications. NewYork: Wiley, 1973.
[3] V. Reinhardt, et al., “A Short Survey of Frequency Synthesizer Techniques,” Proc. 40th
Annual Symposium on Frequency Control, May 1986, pp. 355–365.
[4] R. R. Stone Jr., “Frequency Synthesizers,” Proc. 21st Annual Symposium on Frequency
Control, April 1967, pp. 294–307.
[5] Z. Galani, and R. A. Campbell, “An overview of frequency synthesizers for radars,” IEEE
Trans. Microw. Theory and Tech., vol. 39, no. 5, May 1991, pp. 782–790.
[6] V. F. Kroupa (ed.), Direct Digital Frequency Synthesizers. New York: IEEE Press, 1999.
[7] A. Chenakin, “Frequency synthesis: Current solutions and new trends,” Microwave Journal,
May 2007, pp. 256–266.
[8] A. Chenakin, Frequency Synthesizers. Concept to Product. Norwood, MA: Artech House,
2010.
[9] J. A. Crawford, Advanced Phase-Lock Techniques. MA: Artech House, 2008.
[10] R. E. Best, Phase-Locked Loops: Theory, Design and Applications. NewYork: McGraw-Hill,
1984.
[11] W. F. Egan, Phase-Lock Basics, 2nd ed. NJ: Wiley, 2007.
[12] W. F. Egan, Frequency Synthesis by Phase Lock, 2nd ed. New York: Wiley, 1999.
[13] F. M. Gardner, Phaselock Techniques, 3rd ed. NJ: Wiley, 2005.
[14] J. Klapper and J. T. Frankle, Phased-Locked and Frequency-Feedback Systems. New York:
Academic Press, 1972.
[15] U. L. Rohde, Digital PLL Frequency Synthesizers: Theory and Design. NJ: Prentice-Hall,
1983.
[16] V. F. Kroupa, Phase Lock Loops and Frequency Synthesis. NJ: Wiley, 2003.
[17] S. J. Goldman, Phase-Locked Loop Engineering Handbook for Integrated Circuits. MA:
Artech House, 2007.
[18] A. Chenakin, “Low phase noise PLL synthesizer,” US Patent No. 7 701 299, April 2010.
[19] A. Chenakin, “A compact synthesizer module offers instrument-grade performance and

functionality,” Microwave Journal, February 2011, pp. 34–38.
[20] C. F. Coombs, Jr., (ed.), Electronic Instrument Handbook, 3rd ed. New York: McGraw-Hill,
1999.
[21] F. E. Terman, Electronic and Radio Engineering. New York: McGraw-Hill, 1955.
4 Real-time spectrum analysis and
time-correlated measurements
applied to nonlinear system
characterization
Marcus Da Silva
Nonlinear performance characterization and measurements have become an important

consideration for today’s modern technologies as digital clock rates increase and wireless
communication systems attempt to keep up with the dynamic delivery of voice, video,
and data over a finite wireless spectrum. Transient anomalies caused by the interaction
of various digital and analog signals within a system, nonlinear behavior and device
memory effects, can all degrade system performance, causing EMI/EMC regulatory
violations, lost calls, packet errors, and system inefficiency.
This chapter explores the techniques used in real-time spectrum and signal analysis
and presents applications where real-time technologies can be applied to offer a modern
approach to the detection, discovery, and analysis of transient behavior, including those
caused by nonlinear memory effects and interactions between the digital and analog
parts of modern embedded systems.
4.1 Introduction
The class of instruments called spectrum analyzers has evolved with the uses of the
electromagnetic spectrum and with the available technology. Early instruments, then
called Wave Analyzers, were manually tuned receivers that measured the signal level
at the frequency to which they were tuned. The addition of sweep tuning and a CRT
enabled a two-dimensional display of amplitude versus frequency and engendered the
Swept Tuned Spectrum Analyzer (SA). The advent of digital modulation techniques and
the availability of precision Analog-to-Digital Converters (ADCs), coupled with enough
computing power for Digital Signal Processing (DSP), brought forth the Vector Signal
Analyzer (VSA). The explosion of digital communications and the need to maximize the
ever-increasing amount of information that must be transferred across a limited spectrum
created a need for techniques that separate signals in the time domain as well as the
frequency domain and the ability to observe and measure signals that happen far too fast
for traditional analyzers. This need for speed and the need to correlate time and frequency
in the analysis of RF signals led to the creation of the Real-Time Signal Analyzer (RTSA).
This chapter describes the architecture of real-time signal and spectrum analyzers;
explores some of the theoretical implications of the techniques used; and provides some
examples of RTSA applications. It also covers methods of sequentially applying Discrete
Real-time spectrum analysis and time-correlated measurements 65
Resolution Envelope Video

Low-pass Display
bandwidth detector bandwidth
RF Down-converter filter filter
Attenuator
X–Y
Input
YIG
Swept-tuned
Pre-selector
local
oscillator
Sweep
generator
Fig. 4.1 Swept-tuned spectrum analyzer (SA) architecture.
Fourier Transforms to a continuous stream of time domain samples in order to perform

real-time spectrum analysis. Methods of analyzing and displaying the fast-changing
spectral events captured by real-time spectrum analysis are explored as well.Applications
of the technique to identify and measure transient events caused by interactions between
digital modulation and nonlinear circuit effects are also presented.
4.1.1 Types of spectrum analyzers

Swept-tuned spectrum analyzer
The swept-tuned, super-heterodyne spectrum analyzer shown in Figure 4.1 is the tradi-
tional architecture that first enabled engineers to make frequency domain measurements
several decades ago. The swept SA has since evolved along with the applications that
it serves. Current generation swept SAs include digital elements such as ADCs, DSPs,
and microprocessors. The basic swept approach, however, remains largely the same and
is best suited for observing signals that change slowly relative to the sweep speed.
The swept SA makes power vs. frequency measurements by down-converting the
signal of interest and sweeping it through the pass-band of a resolution bandwidth (RBW)
filter. The RBW filter is followed by a detector that calculates the amplitude at each
frequency point in the selected span. This approach of effectively sweeping a filter across
the span of interest is based on the assumption that the signal being measured does not
change during the time it takes the analyzer to complete one sweep. Measurements
are valid for relatively stable, unchanging input signals. The assumption that signals
are stable for the duration of a sweep is certainly valid in many cases, but becomes
an impediment when analyzing the more dynamic signals prevalent in modern digital
communications, imaging, and radar applications.
Vector signal analyzer

Analyzing signals carrying digital modulation requires vector measurements that pro-
vide both magnitude and phase information. The capability of performing vector
measurements gave the Vector Signal Analyzer, shown in Figure 4.2, its name.
A VSA digitizes all the RF signals within the pass-band of the instrument and puts the
digitized waveform into memory. The waveform in memory contains both the magnitude
66 Marcus Da Silva
Low-pass Display
Decimation X-Y
RF Down-converter IF filter filter PC
Attenuator
Micro-
Input ADC Memory
processor
Band-pass Acquisition Post capture

Local
oscillator bandwidth processing
Fig. 4.2 Vector signal analyzer (VSA) architecture.
and phase information which can be used by DSP for demodulation, measurements, or
display processing. Transformation from the time domain to the frequency domain is
done using Discrete Fourier Transform (DFT) algorithms. Digital signal processing is
also used to provide a variety of other functions including the measurement of modulation
parameters, channel power, power versus time, frequency versus time, phase versus time,
and others.
While the VSA uses DSP to greatly expand its signal analysis capability, it is limited in
its ability to analyze transient events. Signals that are acquired must be stored in memory
before being processed. The serial nature of this batch processing means that the instru-
ment is effectively blind to events that occur between acquisitions. Single events or events
with low repetition rates cannot be reliably captured into memory unless a trigger is avail-
able to isolate the event in time. The dynamic nature of modern RF signals is not always
accurately portrayed due to the relatively slow cycle time for acquisition and analysis.
Real-time signal analyzer

The RTSA shown in Figure 4.3 is designed with dynamic and transient signals in mind.
Like the VSA, a wide pass-band of interest is down-converted to an IF and digitized.
The time domain samples are continuously digitally converted to a baseband data stream
composed of a sequence of I (in-phase) and Q (quadrature) samples. This digital down-
conversion is done in real-time, allowing no gaps in the time record. The same IQ samples
that are fed to the real-time engine1,2 can also be simultaneously stored in memory for
subsequent analysis using batch mode digital signal processing.
The real-time processing Engine3 is a combination of hardware and software optimized
to perform computations at a rate that keeps up with the incoming data stream. DFTs
are sequentially performed on segments of the IQ record, generating a mathematical
representation of frequency occupancy over time. RTSAs generate spectrum data at rates
that are much too fast for the human senses. Visual data compression techniques must
be used to generate meaningful displays. Persistence Spectrum4 and DPX®Spectrum5
(Digital Phosphor Spectrum) are techniques that provide an intuitive “live” view of
complex and dynamic spectrum activity. The real-time processing engine can also be
used to generate a trigger signal based on specific occurrences within the input signal.
These occurrences can be frequency domain patterns, time domain events, or modulation
events. The trigger signal can be used to store specific segments of the IQ time record
for further analysis using batch-mode DSP.
RF Down-converter Real-Time digital processing Post capture processing
Low-pass
Down-convert
Attenuator IF filter & filter Capture Displays
Amp./Phase DDC/
Input ADC Memory
corrections Decimation
X-Y
Ext
Band-pass Local
ADC clock
oscillator
Free Micro-
run processor Live signal processing
Acquisition bandwidth Trigger
DPX™
Real-Time engine
Display
X-Y
trigger
analysis
Real-Time
I-Q out
Real-Time bandwidth Display processing
Fig. 4.3 Real-time signal analyzer architecture.

68 Marcus Da Silva
Table 4.1 Swept-tuned spectrum analyzer (SA) architecture
Real-time spectrum and signal analyzers
Manufacturer Gauss instruments Rhode & Schwarz Tektronix
Product family TDEMI series FSVR series RSA 5000 series

RSA 6000 series
Purpose EMI Measurements Spectrum and signal Spectrum and signal
analysis analysis
Frequency range 0.03, 1, 3, 6, 18, 26.5, 7, 13.6, 30, 40 GHz 3, 6.2, 14, 20 GHz
40 GHz Extendable to 110 GHz
Real-time BW 162.5 MHz 40 MHz 25, 40, 85, 110 MHz
Real-time displays Weighted Spectrogram DPX
spectrogram Persistence spectrum Zero-span
Spectrogram
AM
FM
Phase
Real-time triggers N/A Amplitude Runt
Frequency mask Power (Amplitude)
Frequency edge (FM)
Frequency mask
DPX Density trigger
In addition to the traditional spectrum analysis, RTSAs can perform multiple time
domain, frequency domain, modulation-domain, and code-domain measurements on RF
and microwave signals and can display these measurements in a way that is correlated
in both time and frequency.
Table 4.1 shows the key performance parameters for real-time RF/uW spectrum and
signal analyzers from three manufacturers that are available at the time of writing. The
list is not exhaustive and will undoubtedly grow as the real-time processing technology
advances and as the need for making measurements in real-time becomes critical. The
performance quoted will also change as the technology advances.
The information presented in this chapter is generally applicable to all real-time
spectrum and signal analyzers.
4.2 Spectrum analysis in real-time
The term “Real-Time” has its origins in early work on digital simulations of mechanical
systems. A digital system simulation was said to operate in real-time if its operating
speed matched that of the real system which it was simulating.
To analyze signals in real-time means that the analysis operations must be performed
fast enough to account for all the relevant signal components in the frequency band
of interest. This definition, when applied to Fourier analysis, implies that a real-time
spectrum analyzer that is based on sequential DFTs, as shown in Figure 4.3, must sample
the analog IF signal fast enough to satisfy the Nyquist criteria. This means that the
sampling frequency must exceed twice the bandwidth of interest. It must also perform
all computations continuously at a fast enough rate that the output analysis keeps up
with the changes in the input signal. At the time of publication, RTSAs can process
bandwidths exceeding 160 MHz and generate in the order of 300 K spectrums per
second.
4.2.1 Real-time criteria

The statement that a real-time system must keep up with changes in the incoming signal
requires a more precise definition. There are two useful criteria based on scenarios in
which RTSAs are used to perform spectrum analysis.
1. Transient detection and measurement: An RTSA is often used to observe a section

of spectrum and to detect and measure short-duration events that happen with
unknown timing. A useful measure of performance is the minimum single-event
duration that can be detected and measured with 100% probability, at the specified
accuracy.
2. Spectrum monitoring with no loss of information: An RTSA is often used to
analyze unknown signals in spectrum management and surveillance applications. It
is desirable that all information contained in the signal of interest be included in the
analysis, with no gaps or lost content.
An elaboration of the two definitions requires us to explore the basic RTSA operation of
performing sequential DFTs on a continuous input stream of time domain samples.
4.2.2 Theoretical background

Sequential discrete Fourier transforms
The heart of an RTSA is a DFT engine capable of analyzing frequency behavior over
time. Figure 4.4 shows a simplified schematic of a real-time processing engine sequen-
tially computing DFTs. The frequency domain behavior over time can be visualized
as a spectrogram, shown in Figure 4.5, where frequency is plotted horizontally, time
is plotted vertically, and the amplitude is represented as a color, intensity, or shade. A
swept spectrum analyzer, in contrast, is tuned to a single frequency at any given point
in time. The diagonal line in Figure 4.5 traces the time-frequency trajectory taken by a
sweep. The slope of the line becomes steeper as the sweep slows, so that the function
of a spectrum analyzer in zero-span can be represented as a vertical line indicating that
the instrument is tuned to a single frequency as time advances. Figure 4.5 also shows
how a sweep can miss transient events such as the single frequency hop depicted in the
spectrogram.
70 Marcus Da Silva
DFT*-based spectrum analysis

Input signal x (t ) Memory contents
A/D DFT engine
Time
Time samples
Successive DFTs produce a

series of spectrums over time.
* The Fast Fourier Transform (FFT) is a common

implementation of a Discrete Fourier Transform (DFT).
Fig. 4.4 A DFT-based spectrum analyzer computing a series of transforms over time.
Spectrum
Color scale
Frequency
Sw
ee
p
Sequential DFT results

plotted over time
Spectrogram
Fig. 4.5 Spectrum, Spectrogram and their relation to sweep.
4.3 Spectrum analysis using discrete Fourier transforms
4.3.1 The Fourier transform for discrete-time signals

Consider an input signal x (t) as shown in Figure 4.4 that is digitized with a sampling
period TS resulting in the sampled representation of the input signal given by (4.1).
x[n] = x(nTs ), for n = 0, 1, 2, 3, 4, 5, . . . (4.1)

The Fourier integral, computed for N samples starting at sample nm becomes the sum
+N−1
nm
X(ω) = x (nTs ) e−j ωnTs Ts . (4.2)
n=nm
Normalizing the frequency variable and scaling the magnitude by TS , leads us to the
Discrete Time Fourier Transform (DTFT) for a finite time interval6 .
+N −1
nm
X () = x [n] e−j n . (4.3)
n=nm
The DTFT has N time domain samples at its input and generates a continuous function
of frequency, X (). X () is periodic in with a period of 2π . The time interval
represented by the N samples covered in the summation, starting at sample nm and
ending at nm + N − 1, is called a frame.
Many DTFT algorithms provide an output that is sampled in the frequency domain.
Assuming regularly spaced frequency sampling, the output of each DTFT can be
denoted as
+N −1
nm
x [n] e−j k K n ,
2π
X [k] = (4.4)
n=nm
where 2π
K is the spacing between frequency domain samples.
The formula in (4.4) can be recognized as the Discrete Fourier Transform or DFT7
performed at an arbitrary starting point. It should be noted that, although many algorithms
have an equal number of input and output points (Cooley-Tukey FFT for example),
the number of samples in the frequency domain output does not need to be equal to
the number of time domain samples in the input. (4.4) simply samples the continuous
frequency function X ().
4.3.2 Regularly spaced sequential DFTs

One can extend the above formula for a sequence of transforms by using the index m
to denote the transform number in a sequence. Consider now the continuous flow of
regularly spaced DFTs shown in Figure 4.4. We have a new frequency domain output
being generated every L samples. The time-frequency output can be expressed as a
sampled function of time and frequency:
+N −1
nm
x [n] e−jk K n ,
2π
X [k, m] = (4.5)
n=nm
where K represents the number of frequency points in the output of each transform, k
represents the frequency sample index, L represents the number of time samples between
the start of each successive DFT frames, and N represents the number of time domain
samples in each DFT frame. The indices for frequency, time, and frame number are k, n,
72 Marcus Da Silva
DFT sequence
10
dB
21
DFT frame index
(m)
41
–90
0 50 100
Frequency index
(k)
Fig. 4.6 A 3-dimensional view of a sequence of DFTs taken over time.
and m, respectively. Each value of m represents a new DFT frame with N time domain
samples and produces a frequency spectrum with K frequency domain samples as shown
in Figure 4.6.
There are three relevant cases to consider:
1. L > N : The spacing between frames is greater than the frame duration. There is a
gap between frames. The portion of the input that lies in the gap is ignored. Data is
lost.
2. L = N: The first sample of a frame is the sample immediately following the last
sample of the previous frame. There is no gap. The frames are back-to back. Every
sample of the input is included.
3. L < N : The spacing between frames is smaller than the frame length. The frames
overlap. Not only are all input samples included but a given frame shares some of the
samples with frames that precede it and with frames that follow it.
4.4 Windowing and resolution bandwidth (RBW)
The mathematics of DFTs assumes that the data to be processed is a single period of
a periodically repeating signal. The upper graph in Figure 4.7 depicts a series of time
domain samples. When DFT processing is applied to the 64 samples starting at sample
32 in Figure 4.7, the periodic extension is made to the signal as shown in the lower
graph. The resulting discontinuities can generate spectral artifacts that are not present in
the original signal. This effect produces an inaccurate representation of the signal and
is called spectral leakage8 . Spectral leakage not only creates signals in the output that
Sampled input waveform

4
3
2
1
Voltage
0
–1 0 32 64 96 128
–2
–3
–4
Samples
Periodic extension assumed by DFT

2.5
2
1.5
1
Voltage
0.5
0
–0.5 0 32 64 96 128
–1
–1.5
–2
–2.5
Samples
Fig. 4.7 Time domain samples and the discontinuities caused by periodic extension of samples in a single
frame.
were not present in the input, but also reduces dynamic range, the ability to observe
small signals in the presence of nearby large ones.
Windowing is a technique that is commonly used to reduce the effects of spectral
leakage and to improve the resulting dynamic range. Before performing the DFT, the DFT
frame is multiplied by a window function with the same length sample by sample. The
window functions usually have a bell shape, reducing or eliminating the discontinuities
at the ends of the DFT frame. Figure 4.8 shows an example of a Kaiser window and its
Fourier transform.
Figure 4.9 shows the effects of spectral leakage on dynamic range and how windowing
can be used to reduce its effects. The input signal in Figure 4.9 contains two pure
sinusoids, one at full amplitude with a frequency of 1/13th of the sampling rate; the
second signal is at 1/7th of the sample rate and has an amplitude 1000 times lower
(−60 dB). The trace in black shows the magnitude in dB of a 1024-point DFT with a
rectangular window (un-windowed). Spectral leakage reduces the dynamic range so that
the weaker of the two signals is not visible. The application of a window similar to the
one shown in Figure 4.9 increases the dynamic range so that the weaker signal is easily
visible, as shown in the lower trace.
The choice of window function depends on its frequency response characteristics such
as side-lobe level, equivalent noise bandwidth, and amplitude error. The window shape
74 Marcus Da Silva
Kaiser window 0
1.2 –20
1 –40
–60
0.8 –80
Weight
0.6 –100
–120
0.4 –140
0.2 –160
–180
0 –200
0 200 400 600 800 1000 –8 –6 –4 –2 0 2 4 6 8
Sample number
Fig. 4.8 Kaiser window (beta=16.7) and its Fourier transform.
Rectangular and Kaiser window

0
–20 Kaiser window (a=16.7)
–40 Rectangular window
–60
–80
dB
–100
–120
–140
–160
–180
–200
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Frequency (f/fs)
Fig. 4.9 Dynamic range improvement with windowing.
also determines the effective resolution bandwidth (RBW)9 . Figure 4.9 shows a widening
of the line-width in the spectrum. Many DFT-based spectrum analyzers vary window
parameters as a means of allowing a selectable resolution bandwidth or RBW. RTSAs
offer either user selectable RBW or a choice of several popular window functions.
4.4.1 Windowing considerations

Consider a window function as shown in Figure 4.8. The mth frame of the input x(n)
signal is multiplied by the window to generate the windowed signal, xW [n, m]. The DFT
of the mth frame is then taken as

mL+N−1
mL+N−1
XW [k, m] = x [n] W [n − mL]e−j kS n = xW [n, m] e−j kS n . (4.6)
n=mL n=mL
The windowed function, xW [n, m], has a value of zero at n = mL and at n = mL +

N − 1, the beginning and end of each summation. Discontinuities, like those shown in
Un-windowed signal
0 200 400 600 700 1000
Windowed signal
0 200 400 600 700 1000
Fig. 4.10 Un-windowed and windowed signal x[n] and xW [n, m].
Figure 4.7, are thus attenuated. Figure 4.10 shows a frame of a sinusoidal input signal
and the same frame after windowing.
4.4.2 Resolution bandwidth (RBW)

The term “Resolution bandwidth” (RBW) was coined in the days when spectrum ana-
lyzers actually swept a signal through a physical filter as shown in Figure 4.1. Resolution
bandwidth is defined as the smallest bandwidth that can be resolved in a spectrum ana-
lyzer. A swept analyzer, when presented with a pure sinusoidal tone at its input, traces
the RBW filter shape on its display.
RTSAs use windowed DFTs to generate spectrum displays. The windowing operation
involves multiplying the window shape with the incoming signal. The multiplication of
the window function W (t) and the signal to be analyzed x(t) implies that the DFT of
the time domain product is the convolution of the frequency domain functions.
In commercially available spectrum analyzers, the RBW filter defines the spectrum
shape traced by a perfect sinusoid presented at the input. Therefore, the RBW of a DFT-
based spectrum analyzer is defined as the convolution of a spectral impulse with the
Fourier transform of the window function. The effective RBW filter shape is the Fourier
transform of the window function.
The 3 dB bandwidth of the RBW filter is given by
KW
RBW = , (4.7)
DF rame
76 Marcus Da Silva
where KW is a coefficient that is related to a particular window and DFrame is the frame
duration in seconds.
The right side of Figure 4.8 Figure 4.1 shows the Fourier transform of the Kaiser
window shown on its left. KW is approximately 2.23. The shape factor, defined as the
ratio of the 3dB bandwidth to the 60 dB bandwidth, is approximately 4:1.
The RBW filter shape is the same as the frequency domain shape of the window
function. Performing spectrum analysis with a particular RBW requires choosing a win-
dow whose transform yields the required RBW shape and applying the window to each
DFT frame.
4.5 Real-time specifications
4.5.1 Real-time criteria

The introductory sections of this chapter described two useful criteria for real-time
spectrum analysis. We now elaborate further on these criteria.
1. Transient detection and measurement: A useful measure of performance is the
minimum single-event duration that can be detected and measured with 100% prob-
ability, at the specified accuracy. For simplicity, we define this minimum event as the
narrowest rectangular RF burst that can be detected with 100% probability and have
its underlying RF signal measured without degradation of accuracy.
2. Spectrum monitoring with no loss of information: It is desirable that all information
contained in the signal of interest be included in the analysis, with no gaps or lost
content. A useful test for no information loss in a sequence of DFTs is that input time
domain data can be recovered from the frequency domain output.
An exploration of the two criteria requires us to look at the sequence of DFTs and the
relationship between the DFT length, the spacing between successive DFTs, and the
effects of windowing.
4.5.2 Minimum event duration for 100% probability of intercept at the

specified accuracy
Consider a rectangular burst of RF that is down-converted and sampled. The stream of
samples can be expressed as

x[n] = a[n]ej ω0 nTs u(n − n0 ) − u(n − n0 − np ) , (4.8)
where a(n) is the sampled complex envelope of the RF signal within the burst, ωo is the
RF carrier frequency after down-conversion, n0 is the starting sample of the pulse, and
nP is the number of samples contained within the pulse. A representation of such an RF
burst is shown in Figure 4.11.
The DFT for any particular frame will be the same as that for a continuous signal as
long as that frame is completely contained within the pulse. The spectrum of the RF
t0 t0 + TP
RF Pulse
Time
DFT DFT Frame DFT Frame DFT Frame DFT Frame DFT Frame
mR mF
Fig. 4.11 A DFT frame must be contained within the pulse for an accurate representation of the RF signal
contained in the pulse.
signal within the pulse will be accurately represented as long as the pulse contains at
least one entire DFT frame. Frames that contain either a rising edge or a falling edge
will reflect the presence of those edges in their spectrums and will not accurately depict
the underlying RF signal.
Consider now performing successive DFTs on the single RF burst shown in
Figure 4.11. The starting point of the burst is unknown and not synchronous with the
sequential DFT operation.

mL+N−1
XW [k, m] = W [n − mL]a(n)ej ω0 nTs [u(n − n0 ) − u(n − n0 − np )]e−j ks n .
n=mL
(4.9)
If the rising edge is contained in frame mR and the falling edge in frame mF then
the above expression can be expanded to include all frames that contain some of the RF
burst.
The frames that occur before the beginning and after the end of the burst contain no
signal. The frames that contain the rising and falling edges have a truncated summation.
Their DFT will show the spectral effects of the rise and fall.

mR L+N−1
XW [k, mR ] = W [n − mR L]a(n)ej ω0 nTs e−j kS n
n=n0
n0 +np

XW [k, mF ] = W [n − mF L]a(n)ej ω0 nTs e−j kS n . (4.10)
n=mF L
The DFT for frames that fall completely within the burst will be indistinguishable
from those of a continuous signal. These frames do not contain the rising and falling
edges and have a complete summation. Their spectrum shows a faithful representation
of the signal inside the burst. Any modulation, distortion, or other spectral effects present
78 Marcus Da Silva
in the signal will be accurately represented.
(mR +1)L+N−1

XW [k, mR + 1] = W [n − (mR + 1)L]a(n)ej ω0 nTs e−j kS n
n=(mR +1)L
(mR +2)L+N−1

XW [k, mR + 2] = W [n − (mR + 2)L]a(n)ej ω0 nTs e−j kS n
n=(mR +2)L
..
. (4.11)
(mF −1)L+N−1

XW [k, mF − 1] = W [n − (mF − 1)L]a(n)ej ω0 nTs e−j kS n .
n=(mF −1)L
Pulses and DFT frames are, in the general case, asynchronous. Figure 4.12 illustrates
the case where the DFT frames overlap and the case where there is a gap between frames.
Figure 4.12 also shows that a burst must be at least as wide as the time it takes for two
consecutive frames to be acquired in order to have a 100% probability of containing a
complete frame, considering the arbitrary timing for the RF burst.
Let TP min be the minimum pulse duration required for a pulse to be captured with
a 100% probability with full accuracy. Let TF rame and TGap be the frame duration and
gap duration, respectively. Then
TP min = 2TFrame + TGap . (4.12)
The requirements for discovering, capturing, and analyzing transients are:

• Enough capture bandwidth to support the signal of interest.
• A high enough ADC clock rate to exceed the Nyquist criteria for the capture
bandwidth.
• A frame duration long enough to support the narrowest resolution bandwidth (RBW)
of interest as defined by the Fourier transform of the window function.
• A fast enough DFT transform rate to support the minimum event duration as given
by (4.12). It should be noted that overlapping frames reduce the minimum time by
making TGap negative.
4.5.3 Comparison with swept analyzers

The definitions for TP min can apply to any spectrum analyzer, including the traditional
swept spectrum analyzers. If we replace TF rame by the sweep time, TSW , and TGap by
the retrace time, TRetrace , then (4.12) becomes
TP min = 2TSW + TRetrace . (4.13)
The minimum time for 100% probability of capture and measurement with full accu-
racy for a swept analyzer is twice the sweep time plus the retrace time. TP min for a
Overlapped DFT frames Gap between DFT frames
Minimum pulse duration must Minimum pulse duration must
contain at least one full acquisition contain at least one full acquisition
Acquire data Compute DFT
Gap
Overlap
Time Time
Fig. 4.12 Overlapped DFT frames and frames with a time gap.
80 Marcus Da Silva
swept analyzer observing a 100 MHz span with a 1 MHz RBW is in the order of many
milliseconds. An RTSA can achieve TP min < 6 μs for the same settings.
4.5.4 Processing all information within a signal with no loss of information

It is sometimes required to process all the information contained within a signal, making
sure that no information is lost. As mentioned in the introduction to this section, a useful
test for no loss of information is that the input data can be recovered from the output.
Consider performing consecutive DFTs on a sampled time domain signal x[n] that
has been windowed on a frame-by-frame basis. The DFT of the mth frame is related to
its time samples by
−1
mL+N
2π
X[k, m] = x[n]e−jk K n . (4.14)
n=mL
The original time domain samples for the mth frame can, in general, be reconstructed
from the frequency domain samples with the application of an inverse DFT and divi-
sion by the window function. It must be noted that the value of a time domain sample
becomes indeterminate if the window function has a weight of zero at a particular sam-
ple. Numerical resolution can also affect the ability to accurately recover time domain
data, especially for the samples where the window has small values.
Let yW [n, m] be the output of an inverse DFT (IDFT) for the mth frame. The windowed
time domain samples can be recovered by
2 −1
K
1 2π
yW [n, m] = XW [k, m]ej k K n , mL ≤ n ≤ N − 1. (4.15)
N K
k=− 2
The effects of the window can be removed by dividing yW [n, m] by the window
function. Applying the equation for the mth DFT and exchanging the index n for p, we
have
2 −1 mL+N
K
1 −1 2π
yW [n, m] = xW [p, m]e−j kS p ej k K n . (4.16)
N K
k=− 2 p=mL
Inverting the order of summation and combining the exponentials, we get
2 −1
K
1
mL+N−1 2π
yW [n, m] = xW [p, m] ej k K (n−p) . (4.17)
N
p=mL k=− K2
The second summation above has a value of zero except for the cases where the
argument of the complex exponential is either zero or a multiple of 2π , where its value
is K.
K
mL+N−1
yW [n, m] = xW [p, m]∂[p − n + iK], where i is an integer. (4.18)
N
p=mL
We must now remember that xW [p, m] is non-zero only for the N samples in the
window, mL ≤ p < mL + N. We must also remember that the DFT is periodic with
period K.
Consider the following cases:
1. K N: Data recovery is possible. The number of frequency domain points is greater
than or equal to the number of time domain points. The only value that falls inside
the summation limits is i = 0. The result is an exact reproduction of the original
windowed function for the mth frame.
K
mL+N−1
yW [n, m] = xW [p, m]∂[p − n]
N
p=mL
K
yW [n, m] = xW [n, m]. (4.19)
N
2. K < N: Aliasing occurs. The number of frequency domain points is smaller than
the number of time domain points. There are more than K points in the summation.
Non-zero values of i fall within the summation limits. The recovered output contains
contributions from multiple periods. The frequency domain signal is said to be under-
sampled. The time domain samples cannot be uniquely recovered from the DFT
output. ⎧ ⎫

⎪
⎪
N
int K ⎪
⎪
K⎨ ⎬
yW [n, m] = xW [n, m] + xW [n − iK, m] . (4.20)
N⎪⎪ ⎪
⎪
⎩ i=1 ⎭
4.5.5 Windowing and overlap

The derivations in (4.18) – (4.20) show that the original time domain data for a particular
frame can be recovered from its DFT provided that the number of frequency samples is
at least as great as the number of time samples and that the window has non-zero values
across the DFT frame. The recovery of all data can still be achieved with DFT frame
overlap even if the window function goes to zero at the beginning and end of each frame.
Consider now a sequence of windowed DFTs taken at regular intervals. Taking inverse
DFTs can recover the windowed time domain samples for each frame. There are three
cases to consider.
1. L > N: The spacing between each DFT is greater than the DFT length. There is a gap
between consecutive DFT frames. Any variations in the signal that happen inside the
gap are lost.
2. L = N: Consecutive DFTs begin with the sample immediately following the last sam-
ple of the previous one. There are no gaps between successive DFTs. Every time
domain sample is considered. The application of inverse DFTs can, in theory, recover
each and every sample of the input. There are, however, practical considerations:
82 Marcus Da Silva
L>N: Gap between frames
–2
–6
–10
–14
–18
0 32 64 96 128
Fig. 4.13 L > N – Gap between DFT frames.
L=N: End-to-between frames
–2
–6
–10
–14
–18
0 32 64 96 128
Fig. 4.14 L = N- Frame arranged end-to-end, with no gap.
• The window cannot be zero at any point in the frame.

• Numerical precision may cause an unrecoverable degradation of data at points
where the window value is small.
Figure 4.14 shows the same input signal, now with an end-to-end frame arrangement.
Limited numerical precision makes the recovery of samples near the ends of the frames
impractical. Figure 4.14 illustrates this point since the anomaly around sample 32 is
attenuated to a value very near zero by the window.
3. L<N. Consecutive DFTs overlap. All DFT frames share some samples with both their
successors and predecessors. Successive inverse DFTs can uniquely reconstruct the
input signal from the output as long as the windowing function is known and the
zeroes of overlapping windows don’t occur at the same place. Bell-shaped windows,
such as the ones typically used in spectrum analysis, allow reconstruction when the
amount of overlap is enough to overcome any numerical precision issues. Information
is not lost.
Figure 4.15 shows the same input signal as in the previous two examples, this time
processed with 50% frame overlap. Note that the effects of the anomaly near sample
50% Frame overlap
–2
–6
–10
–14
–18
0 32 64 96 128
Fig. 4.15 L = N – DFT frames with 50% overlap.
32 are included in the first and second frames. An inverse DFT (IDFT) provides two
estimates of the input signal for samples lying within the gap. The input signal can be
recovered with high confidence by choosing the best of the two estimates or by optimally
processing the two results.
It must be noted here that the proceeding discussion assumes that both magnitude and
phase information are available from the DFT outputs. Although this is inherently true
for DFT computations, most spectrum analyzers typically only display the magnitude.
The test that the time domain data be recoverable is still useful for us to determine if all
information present in a signal is included in the analysis and reflected in its results.
4.5.6 Sequential DFTs as a parallel bank of filters

Figure 4.16 illustrates an intuitive interpretation of the process used in real-time spectrum
analysis. Taking sequential equally spaced DFTs over time, as shown in the upper part
of Figure 4.16, is conceptually equivalent to the filtering and sampling system shown in
the lower part of the figure.
Consider passing the input signal though a bank of K band-pass filters whose frequency
spacing is the same as the spacing between DFT bins and whose bandwidth and frequency
response is the same as the RBW shape described in Figure 4.8. The resulting K analog
waveforms are then sampled and digitized in magnitude and phase (or I and Q) at the rate
and timing with which the DFTs are computed. The resulting output is indistinguishable
from the result of the computations in the upper part of Figure 4.12 Overlapped DFT
frames and frames with a time gap.10
Consider one of the band-pass filter paths. Slow signal variations, those with frequency
components contained within one RBW, result in changes in the level and phase of that
path over time. A faithful representation of all signals present within the RBW requires
that the magnitude and phase be sampled at a fast enough rate to meet the Nyquist criteria,
at least one complex sample (a complex sample contains two samples) for every Hz of
bandwidth. The required time between frames is related to the RBW by
1
TFrame ≤ . (4.21)
RBW
DFT*-based spectrum analysis
Input signal
Memory contents
AID Memory
Time Time samples
N-point FFT
Tim
Equivalent bank of filters e
Bank of N Bandpass
filters with centers Complex
separated by one FFT Envelope
frequency bin width detection
M/θ
Input signal
M/θ
M/θ Sampled at the
same rate that
transforms are
Time computed
M/θ
* The Fast Fourier Transform (FFT) is a common implementation

of a Discrete Fourier Transform (DFT).
Fig. 4.16 DFTs can be interpreted as a parallel bank of band-pass filters.

Signal variations that are too fast to be contained within one RBW affect the magnitudes
and phases of adjacent paths and are resolved in the spectrum graph.
4.5.7 Relating frame rate, frame overlap, and RBW

Equation (4.7) relates the frame duration to RBW. Combining this with (4.21) gives us
DFrame
≥ KW . (4.22)
TFrame
The duration of the frame must exceed the time between consecutive frames. Consec-
utive frames must overlap in time. Each contains some samples in common with both the
previous and the next frames. For the Kaiser window shown in Figure 4.8, KW = 2.23.
This window requires an overlap of at least 55% to ensure that all data contained in
the time domain input signal is included in the three-dimensional output of spectrum
versus time.
4.5.8 Criteria for processing all signals in the input waveform with no loss of information
In order to take all the information contained in a time domain waveform and transform it
into a frequency domain representation with no loss of information in real-time requires
several important signal processing requirements:
• Enough capture bandwidth to support analysis of the signal of interest.

• A high enough ADC clock rate to exceed the Nyquist criteria for the capture
bandwidth.
• A long enough frame duration to support the narrowest resolution bandwidth (RBW)
of interest as defined by the Fourier transform of the window function
• Overlapping DFT frames. The amount of overlap depends on the window or BRW
used (4.22).
4.6 Applications of real-time spectrum analysis
4.6.1 Displaying real-time spectrum analysis data

RTSAs generate spectrum data at rates that are far too fast for the human eye to see.
A means must be employed to compress the available data so that the user can observe
dynamic, time-varying signals in a live manner. Several methods are used in the industry
to compress the data into a form that is observable. These include three-dimensional
displays (level, frequency, and time) like the Weighted Spectrogram11 offered by Gauss
industries, as well as ways to display the statistics of frequency occupancy over time such
as Persistence Spectrum12 and DPX®13 , offered by R&S and Tektronix, respectively.
This text focuses on the Tektronix DPX method.
86 Marcus Da Silva
4.6.2 Digital persistence displays

The names “Digital Phosphor” and “Persistence Spectrum” are derived from the phos-
phor coating on the inside of cathode ray tubes (CRTs) used in older oscilloscopes and
spectrum analyzers, where the electron beam is directly controlled by the waveform to
be displayed. When the phosphor is excited by an electron beam, it fluoresces, light-
ing up the path drawn by the stream of electrons. The phosphors had the properties of
persistence and proportionality.
Persistence is the property that the phosphor continues to glow even after the electron
beam has passed by. Persistence allows the human eye to see events that would otherwise
occur too fast to be seen. Proportionality means that the brightness of a phosphor is
proportional to the number of electrons that hit a target point in the phosphor. The
brightness of the spot increases as the electron beam hits it more frequently.
Digital Phosphor (DPX)14 technology was developed by Tektronix to bring the analog
benefits of a variable persistence CRT to modern digital instruments. DPX includes digital
enhancements such as intensity grading, selectable color schemes, and statistical traces
that communicate more information in less time. In the DPX display, both color and
brightness provide z-axis emphasis.
4.6.3 The DPX spectrum display engine

A real-time signal analyzer computes hundreds of thousands of spectrums every second.
This high transform rate is the key to detecting infrequent events, but it is far too fast
for the liquid-crystal display to keep up with, and well beyond what human eyes can
perceive. The DPX engine writes the incoming spectrums into a bitmap database at full
speed then transfers the resulting image to the screen at a viewable rate. The bitmap
database can be viewed as a dense grid created by dividing a spectrum graph into rows
representing trace amplitude and columns for points on the frequency axis. Each cell in
this grid contains the count of how many times it was occupied by an incoming spectrum.
The result is a three-dimensional database where each point in the grid is represented by
its x-axis (frequency), y-axis (amplitude) and z-axis (number of occurrences).
Persistence
Persistence is obtained by accumulating the contents of many DFTs and storing the
results in a bitmap where the x and y axes correspond to frequency and amplitude,
respectively and the z-axis, usually represented as a color, is an indication of how often
a particular point is occupied.
The 11 × 10 matrix shown in Figure 4.17 illustrates the concept. The bitmap is com-
puted by adding the contents of consecutive grids, each grid corresponding to a DFT
frame. The picture on the left of Figure 4.17 shows what the database cells might contain
after a single spectrum is mapped into it. Blank cells contain the value zero, meaning that
no points from a spectrum have fallen into them yet. The grid on the right shows values
that our simplified database might contain after an additional eight spectral transforms
have been performed and their results stored in the cells. One of the nine spectrums
1
1
1 2
Amplitude
Amplitude
1 4
2 1
1 1 1 5 7
1 1 7 8
1 1 1 1 1 1 9 9 9 1 1 1 1 1 9 9 9
Frequency Frequency
Fig. 4.17 Example 3-D bitmap database after 1 (left) and 9 (right) updates. Note that each column contains
the same total number of “hits.”
happened to be computed at a time during which the signal was absent, as indicated by
the string of “1” values at the noise floor.
The DPX engine receives spectrum information at the full DFT rate and then accu-
mulates the bitmap for a large number of DFTs, passing the bitmap to the display at a
much slower rate. Each accumulated bitmap that is passed to the display is called a dis-
play frame. The display frame is computed by accumulating the contents of the bitmap.
The Tektronix RSA6100A series of RTSAs, for example, performs nearly 292 000 DFT
operations per second and updates the display frame at 20 times per second.
Variable persistence occurs when only a fraction of each count is carried over to the
next display. Adjusting the fraction changes the length of time it takes for a signal event
to decay from the database, and thus fade from the display. Raising the fraction to unity
provides infinite persistence, where each point in the bitmap contains a histogram of the
number of times it was hit since the process was started. A value of zero means that there
is no persistence and it is freshly updated each display frame.
Color mapping and persistence

The display processing system then maps the number of occurrences to a color scale.
Warmer colors (red, orange, yellow) indicate more occurrences. Cooler colors (blues and
purples) indicate fewer occurrences. Black indicates no occurrences. Other intensity-
grading schemes can also be used.
Imagine a 20 dB step change in RF level happening at the end of a display frame.
Assume that the initial level was present for all 14,600 of the spectrum updates in one
display frame and that the variable persistence factor causes 25% attenuation after each
frame (α = 0.75). After the initial frame, the cells affected by the initial level would
start out with a value of 14,600 and be displayed at full force (Red). One display frame
later, the values of those cells would drop by 25% (10,950), while those occupied by the
new level would be at 25% (3650). After the next frame, the cells occupied by the initial
level decrease by another 25%, while the value for the cells occupied by the new level
88 Marcus Da Silva
Fig. 4.18 With variable persistence, a brief CW signal captured by DPX remains in the display for an
adjustable period of time before fading away.
increases. Eventually all the cells occupied by the original level vanish and only those
cells associated with the new level are visible. On the screen, you would initially see a
bright trace with a spike at the signal frequency. The part of the trace where the original
signal occurred fades away. During this time, the pixels start to brighten at the new level
below the fading signal. In the end, there is only the new trace in the display, as shown in
Figure 4.18. Infinite persistence can catch even a single occurrence of a spectrum event.
Variable persistence can provide an insight into dynamic signal behavior as it happens.
DPX line traces

The DPX engine can also produce line traces. The +Peak and −Peak traces show signal
maxima and minima instantly and clearly. The average trace finds the mean level for the
signal at each frequency point.
Using Hold on the DPX + Peak trace is almost exactly the same as the Max Hold trace
on a typical spectrum analyzer, with the important difference that the DPX trace’s update
rate is orders of magnitude faster. The +Peak and −Peak traces show signal maxima and
minima over arbitrarily long time periods. The +Peak trace displays a single occurrence
of an RF burst lasting a few microseconds in an observation period of many hours or
even days. Similarly, the −Peak trace shows a momentary gap in a signal lasting as little
as a few microseconds.
4.7 Triggering in the frequency domain
Triggering is critical to capturing time domain information. Triggering allows a user

to concentrate attention and analysis on a window of time where an event of interest
happens. RTSAs offer unique trigger functionality, providing the ability to trigger on a
frequency mask, modulation and signal statistics as well as power in a bandwidth. Other
trigger functions including external triggering and gated triggers are included as well.
This section explores the frequency mask trigger (FMT), beginning with an overview of
triggering in general.
4.7.1 Digital triggering

Triggering was originally implemented as a way to stabilize oscilloscope displays. In
traditional analog oscilloscopes, the signal to be observed is fed to one input while the
Most recent sample

Oldest sample
First In first out (FIFO) memory
1 1 1 N NN
A/D 0 1 2 3 4 5 6 7 8 9 0 1 2 – – – N
3 2 1
Memory control
Sampling Freeze
clock
TD = 0 Pre-trigger samples only
TD = Memory length Post-trigger samples only
TD = 50% Memory length Half Pre-trigger, Half Post-trigger
Trigger TD
Variable delay
0 to N clocks Signal
Trigger
Trace
Fig. 4.19 Triggering in digital acquisition systems.
trigger is fed to another. The trigger event causes the start of a horizontal sweep while the
amplitude of the signal is shown as a vertical displacement superimposed on a calibrated
grid. In its simplest form, analog triggering allows events that happen after the trigger
to be observed.
4.7.2 Triggering in systems with digital acquisition

The ability to represent and process signals digitally, when coupled with large memory
capacity, allows the capture of events that happen before the trigger as well as after
it. Digital acquisition systems of the type used in Tektronix RTSAs use an analog-
to-digital converter (ADC) to fill a deep memory with time samples of the received
signal. Conceptually, new samples are continuously fed to the memory, while the oldest
samples fall off. The example shown in Figure 4.19 shows a memory configured to store
N samples. The arrival of a trigger stops the acquisition, freezing the contents of the
memory. The addition of a variable delay in the path of the trigger signal allows events
that happen before a trigger, as well as those that come after it, to be captured.
Consider a case in which there is no delay. The trigger event causes the memory to
freeze immediately after a sample concurrent with the trigger is stored. The memory
then contains the sample at the time of the trigger as well as N − 1 samples that occurred
before the trigger. Only pre-trigger events are stored.
Consider now the case in which the delay is set to match exactly the length of the
memory. N samples are then allowed to come into the memory after the trigger occurrence
before the memory is frozen. The memory then contains N samples of signal activity
after the trigger. Only post-trigger events are stored.
90 Marcus Da Silva
Digital down
conversion (DDC)
I
X Filter and decimate
+ Captured
A/D Sample memory IQ data
phase/amplitude
Q
X corrections
Level/
power
External triggers
trigger
Quadrature
Frequency
NCO
mask
Real trigger
time
FFT Digital
DPX trace data
Phosphor
Fig. 4.20 Real-time signal analyzer trigger system.
Both post- and pre-trigger events can be captured if the delay is set to a fraction of the
memory length. If the delay is set to half the memory depth, then half the stored samples
are those that preceded the trigger and half are the stored samples that followed it.
Once data is stored in memory, it is available for further analysis using a DSP. The
stored signal data can subsequently be analyzed in the frequency, time, and modulation
domains. This is a powerful tool for applications such as signal monitoring and device
troubleshooting.
4.7.3 RTSA trigger sources

RTSAs provide several methods of internal and external triggering, as shown in
Figure 4.17. Triggers can come from an external source, from a computation of sig-
nal power, from a real-time demodulation of the input signal or from the frequency
domain content of the input signal in a frequency mask trigger (FMT).
4.7.4 Frequency mask trigger (FMT)

Frequency mask triggering compares the spectrum shape to a user-defined mask. This
powerful technique allows changes in a spectrum shape to trigger an acquisition. Fre-
quency mask triggers can reliably detect signals far below full-scale even in the presence
of other signals at much higher levels. This ability to trigger on weak signals in the
presence of strong ones is critical for detecting intermittent signals, the presence of
intermodulation products, transient spectrum containment violations, and much more.
Trigger events are determined in the frequency domain using a dedicated hardware DFT
processor, as shown in the block diagram in Figure 4.20. A full DFT is required to
compare a signal to a mask, requiring a complete frame.
Consider a sequence of log-magnitude traces generated by consecutive DFTs. Each
of these spectrums is compared with a stored mask. A trigger is generated whenever one
of the incoming spectrums violates the mask conditions.
Fig. 4.21 Frequency mask trigger operation.
Figure 4.21 shows frequency mask triggering applied to a frequency hop. The fre-
quency mask, shown as a shaded area, is drawn to exclude the normal hop behavior
and to catch a frequency overshoot. The spectrogram in the upper part of Figure 4.21
shows that there is a momentary transient as the frequency hops from its original setting
just to the right of center to a new one about 6 MHz lower. The overshoot violated the
frequency mask and generated the trigger for the acquisition. Trigger parameters were
set to display events that preceded the trigger as well as events that happened after the
trigger, as shown in the spectrogram in the lower part of Figure 4.21.
Markers can be used to display the spectrum corresponding to each horizontal line
in the spectrogram, pinpointing both the time and frequency of a spectrum feature. The
full time record represented by the spectrogram is also stored and can be subjected to
analysis using the many other analytic functions of the RTSA.
4.7.5 Frequency mask trigger time resolution and time alignment

Triggering in the time domain as illustrated in Figure 4.19 can uniquely locate an event
to a resolution of a single time sample (some oscilloscope trigger systems use interpo-
lation to achieve sub-sample time resolution). This is the case for external, level, and
power triggers in RTSAs. Frequency mask triggering, however, compares the output of
a sequence of DFTs with a frequency domain mask. Each spectrum is computed for an
92 Marcus Da Silva
entire DFT frame containing a large number of samples. The location of an event within
a DFT frame cannot be known. The center of the frame, corresponding to the sample for
which the window has the highest value, is chosen by convention. The narrowest burst
of RF that can be captured at full accuracy with a frequency mask triggering follows the
principles outlined in the development of (4.13).
4.7.6 Other real-time triggers

The ability to digitally process signals in real-time provides a myriad ways to detect
events within a signal and isolate those events in time. This allows the user to analyze
the events and their causes by applying the power of digital signal processing to a short
time period near the occurrence of interest. A list of these triggers includes:
• Frequency-edge trigger (FM trigger): A trigger is generated when the frequency
trajectory of a signal crosses a threshold.
• Amplitude (power) trigger: A trigger is generated when the amplitude of a signal
crosses a user-defined threshold.
• DPX density trigger: The DPX processing engine essentially keeps statistics of the
signal occupancy or density on a frequency-amplitude grid. A DPX density trigger
allows the user to select a two-dimensional region or box on this grid and trigger when
the average density over the box exceeds a user-selectable level. Alternatively, the
system can learn the average density and trigger when the density changes from the
“normal” value.
4.8 Application examples: using real-time technologies to solve

nonlinear challenges
4.8.1 Discovering transient signals

Behaviors that create transient nonlinear effects have traditionally been difficult to iden-
tify and to troubleshoot. These events often cause non-stationary behavior that is causally
linked to other parts of a complex system. Some examples of nonlinear behavior caused
by coupling between independent parts of complex systems include:
• Systems that combine digital and analog circuitry can exhibit power supply voltage
dips caused by variations in current draw as computations are performed that create
momentary clipping in the analog signal path.
• Systems involving DSP can have software errors that cause incorrect filter values to
be momentarily applied.
• Phase hits can be caused by the physical effects of component heating.
• Systems that use RF bursts can exhibit nonlinear effects that change in time due to
thermal transients in amplifying devices.
Discovering the root cause of transient events often involves mixed domain analysis and
multiple instruments. The tools that are used to correlate events require precise triggering
Oscilloscope RTSA
TDS6000 RSA6100A
Trigger
VDC
IDC
Data
Modulator DAC
In
Digital signal processing Transmitter

(Microprocessor/FPGA)
Fig. 4.22 Frequency mask triggering enables precise event capture.
and time alignment across multiple measurement domains. The real-time process of
frequency mask triggering provides a unique method of transient event isolation.
4.8.2 Adjacent channel power (ACP) violation caused by power supply fluctuations
Consider a case where a power supply current spike is causing a spectrum anomaly in a
transmitter system. Figure 4.22 demonstrates an approach to resolving the root cause of
transient events.
A frequency mask trigger (FMT) is used to trigger on spectrum anomalies such as
excessive adjacent channel power (ACP). The trigger is then sent to the oscilloscope
to reveal the simultaneous behavior of the power supply drain current. The real-time
triggering functionality on the RTSA can not only enable a trigger output to cross-trigger
other instruments: it can also capture the IQ representation of the RF signal containing
the transient event into internal memory. Once captured into memory, the RF signal can
be analyzed for power fluctuations, modulation errors, phase stability, or any other RF
parameter of interest.
4.8.3 Software errors affecting RF performance

Software errors are often the root cause of a transient spectrum anomaly. Tracking down
a software problem often requires the simultaneous observation of signals in many parts
of a complex system, often crossing the RF, analog baseband, digital, and software
domains.
Figure 4.23 illustrates what might be needed to track an RF signal integrity problem
across the various building blocks of a modern RF transmitter. The RTSA is used to trigger
the oscilloscope and logic analyzer when a transient event appears in the spectrum.
94 Marcus Da Silva
Antenna
Baseband UP Converter
processing DAC
IQ PA
(ASC, FPGA, Modulator
DSP,..) DAC
Data Trigger
transfer
Logic analyzer Oscilloscope Real-time signal

analyzer
Fig. 4.23 Multi-domain event analysis enable by real-time triggering.
Fig. 4.24 Frequency mask trigger used to troubleshoot analog and digital errors.
The external trigger output from the RTSA is then fed to the oscilloscope and the logic
analyzer. The time-correlated baseband analog signal and the time-correlated digital
signal can be simultaneously captured and displayed on an integrated view. The screen
shot on the left of Figure 4.24 shows a spectrum mask violation in a BPSK signal that
was used to trigger the scope and logic analyzer shown in Figure 4.23. The IQ baseband
signal from the oscilloscope and the digital bus that time correlates to the trigger event
are shown in the screen shot on the right of Figure 4.24. With common source code
debug tools, a trace can be put on the real-time hardware and real-time instructions
being executed, so the line of code being executed at the time of the spectrum event can
be isolated.
Power
Digital I amplifier
PAR
limiter + DAC
Digital Q linearizer
DSP
Output
signal
Local sample
oscillator
ADC
Feedback
Fig. 4.25 Digital pre-distortion with feedback.
4.8.4 Memory effects in digitally pre-distorted (DPD) amplifiers

RF power amplifiers used in digital communications systems need to be linear to min-
imize out-of-band spurious emissions and distortion. They also need to be efficient to
minimize heat dissipation and extend battery life. Various adaptive digital pre-distortion
(DPD) techniques have emerged as promising linearization methods and are being
applied to applications ranging from cell phones to radars. DPD employs digital base-
band pre-distortion ahead of the amplifier to compensate for the nonlinear distortion
contributed by PAs as they operate in high-efficiency yet nonlinear regions. Compen-
sation levels are monitored and adapted based on a sensing feedback loop, as shown in
Figure 4.25.
Memory effects have been identified as a major source of degradation of DPD perfor-
mance in PA systems primarily due to the finite delay in the sensing and adapting of the
feedback system. Memory effects can be defined as the frequency-dependent distortion
in RF power amplifiers caused by circuit elements with memory. Memory effects in
PAs15 are generally categorized into two groups: electrical effects and thermal effects.
Electrical effects are usually caused by the source and/or load impedance variations
over the modulation bandwidth. Thermal memory effects are caused by electro-thermal
coupling. Since PAs always exhibit some sort of thermal inertia and most of the elec-
trical parameters of the transistors are functions of temperature, thermal effects are
unavoidable.
The left side of Figure 4.26 shows an example of a traditional spectrum display of a
power amplifier output after the DPD correction has been applied. The right side shows
a DPX display of the same signal. The DPX display shows the effects of infrequent
spectral regrowth. These infrequent events go undetected on the traditional spectrum
analyzer. Adjacent channel power (ACP) measurements and other conventional ways
96 Marcus Da Silva
Fig. 4.26 Spectrum of a power amplifier after DPD correction has been applied.
of measuring and characterizing power amplifiers are based on averaging signal power
over a relatively long time period and are incapable of showing the transients caused by
memory effects.
4.9 Conclusions
Spectrum and signal analyzers have evolved along with the many novel ways we’ve
learned to use the electromagnetic spectrum and with the changes in technology. Modern
digital communications formats have created a need to simultaneously observe events
in the time as well as the frequency domain and to measure dynamic signals that change
far too fast for traditional analyzers.
Real-time spectrum analysis provides a methodology to observe events that are far
too brief to be detected with traditional equipment. Advanced trigger functions allow the
pinpoint capture of these rare RF events and a way to time-correlate them to their causes.
Nonlinear effects in RF devices such as power amplifiers often display anomalous
spectrum behaviors that coincide with occurrences in the digital portions of a system or
as a result of modulation anomalies (occasional power peaks for example) that are carried
on the RF. Errors in programming, noise from switching power supplies, and RF energy
radiated from nearby digital circuitry can also cause spectrum abnormalities that are rare
and of short duration. Real-time spectrum analysis, with its ability to perform gapless
analysis and to detect single events lasting as little as a few microseconds, provides a
tool to isolate these rare events in an RF signal and track down their causes.
End notes
1. R&S White Paper “Implementation of Real-Time Spectrum Analysis,” page 6.

2. Tektronix, “Fundamentals of real-time Spectrum Analysis.”
3. Implementations differ among various equipment manufacturers. The discussion in
this book focuses on the Tektronix implementation.
4. R&S FSVR Data Sheet, page 16.
6. Stremler [5], p. 135.
7. Oppenheim [6], p. 520.

8. We have effectively used a “rectangular window [6].”
9. Oppenheim [7], p. 717.
10. Rabiner and Gold [7], p. 386.
11. Gauss Instruments TDEMI 40G Data Sheet.
12. R&S White Paper “Implementation of Real-Time Spectrum Analysis”, p. 19.
14. US Patent # 7,216,046.
15. Y. He, D. McCarthy and M. daSilva.
References
[1] R & S FSVR Data Sheet. Münich, Germany: Rohde & Schwarz GmbH KG, 2010.
[2] Implementation of Real-Time Spectrum Analysis. Munich, Germany: Rohde & Schwarz
GmbH KG, 2011.
[3] Gauss Instruments TDEMI 40G Data Sheet. Münich, Germany: Gauss Instruments GmbH,
2012.
[4] Fundamentals of Real-time Spectrum Analysis. Beaverton, OR: Tektronix Inc., 2008.
[5] F. G. Stremler, Introduction to Communications Systems. Boston, MA: Addison-Wesley,
1990.
[6] A. V. Oppenehim and R. W. Schafer, Discrete Time Signal Processing. Englewood Cliffs,
NJ: Prentice-Hall, 1989.
[7] L. R. Rabiner and B. Gold, Theory and Application of Digital Signal Processing. Englewood
Cliffs, NJ: Prentice-Hall, 1975.
[8] A. V. Oppenheim, A. S. Willsky, and I. T. Young, Signals and Systems. Englewood Cliffs,
NJ: Prentice-Hall, 1983.
[9] Y. He, D. McCarthy, and M. daSilva, “Different measurement methods for characterizing and
detecting memory effects in nonlinear RF power amplifiers,” ARFTG Conference, Phoenix,
December 2007.
[10] Measurement of the nonlinearities of RF amplifiers using signal generators and a spectrum
analyzer. Münich, Germany: Rohde & Schwarz GmbH KG, 2006.
[11] Fundamentals of Digital Phosphor Technology in Real-time Spectrum Analyzers, Beaverton,
OR: Tektronix Inc., 2008.
[12] DPX turns a light on in a dark room, Beaverton, OR: Tektronix Inc., 2006.
[13] K. Bernard and E. Gee, “Real time power mask trigger,” US Patent 7 251 577, July 31, 2007.
[14] K. Bernard, “Time-arbitrary signal power statistics measurement device and method,” US
Patent 7 298 129, November 20, 2007.
[15] K. Bernard and E. Gee, “Real time power mask trigger,” US Patent 7 418 357, August 26,
2007.
[16] M.Agoston, W. B. Harrington, and S. L. Halsted, “Method of generating a variable persistence
waveform database,” US Patent 7 216 046, May 8, 2007.
[17] S. R. Morton and J. C. Demogalla, “Method and apparatus for identifying, saving, and
analyzing continuous frequency domain data in a spectrum analyzer,” US Patent 5 103 402,
April 7, 1992.
[18] J. D. Earls and A. K. Hillman, “Multichannel simultaneous real-time spectrum analysis with
offset frequency trigger,” US Patent 7 352 827, April 1, 2008.
5 Vector network analyzers
Mohamed Sayed and Jon Martens
5.1 Introduction
The VNA is the instrument that measures the S-parameters (and related quantities) of pas-
sive and active devices and components. The phase and magnitude of these S-parameters
are displayed in different formats in accordance with the user’s application. Scalar net-
work analyzers measure only the magnitude of the device’s performance and that is not
the focus of this chapter.
VNA measurements can be done using one or many ports, over swept frequency or
swept power and with a variety of receiver configurations, depending on the measure-
ment requirement. This chapter explores the history of this instrument, some aspects
of its structure and performance, and a brief introduction on how specific measurement
applications are affected by the VNA attributes. Many microwave measurement concepts
and instruments are based on the VNA and some are discussed later in this book. As such,
this chapter serves as something of an introduction to many subtopics.
5.2 History of vector network analyzers
5.2.1 Pre-HP-8510 VNA – 1950–1984

Rohde and Schwarz introduced the first impedance measuring device that could warrant
the term “network analyzer” in 1950. Wiltron introduced the 310 VNA in 1965. This
was followed by the HP VNA in 1966, 1968, and 1970.
Table 5.1 shows the VNA model numbers and the years for this time period. In [1],
Doug Rytting describes in detail these early VNAs.
During this period, sweepers were used as narrow-band sources (2–4 GHz, 4–8 GHz,
etc.). The displays shown on the monitor were rectangular or Smith Charts. External
computers were used to control the measurements and displays. To cover the 2–20
GHz frequency range, multiple sources were used. Calibration was done manually with
external calibration kits.
Four full racks of instruments were needed to perform as an “Automatic Network
Analyzer” or ANA. BWOs were used as the source. Multiple plug-ins were used to cover
different types of displays and there was one rack for computing and displaying results.
All these instruments were located in measurement rooms staffed by engineers and
Vector network analyzers 99
Table 5.1 Pre-HP 8510 Vector Network Analyzers (VNA)
Date Company Country Model number Frequency range
1950 Rohde & Schwarz Germany DuZ-g Diagraph 30–300 MHz

1952 Rohde & Schwarz Germany ZD-9D Diagraph 300–2400 MHz
1965 Wiltron USA W-310/311 1-2 / 2-4 / 4-8 GHz
1966 HP USA HP-8405 18 GHz
1967 HP USA HP-8410 12.4 GHz
1968 HP USA HP-8540 18 GHZ
1970 HP USA HP-8542 18 GHZ
1972 HP USA HP-8409 18 GHZ
technicians. Lab engineers needed to reserve specific times to take their measurements
and make calibrations.
Microwave samplers were used for down-converting the input signal to a fixed IF.
Harmonics of the low-frequency oscillator were used to mix with the input microwave
frequency. Thus, the system performance degraded as frequency increased. The noise
floor and stability were not as good as at the present time.
Waveguide frequency bands were used to define different frequency ranges, e.g.
L-Band of 1–2 GHz, S-Band of 2–4 GHz, C-Band of 4–8 GHz, X-Band of 8–12 GHz,
Ku-Band of 12–18 GHz, K-Band of 18–26.5 GHz, and Ka-Band of 26.5–40 GHz. The
Type N connector was the typical one used up to 18 GHz. Precision connectors were
developed to go higher in frequency, e.g. APC–7 will go up to 20 GHz and APC–3.5 will
go up to 26.5 GHz.
5.2.2 HP-8510 VNA System – 1984–2001

The 26.5 GHz HP-8510 VNA system was introduced in 1984. This VNA included a
synthesized source (HP-8340A), error correction, time domain and pulse measurements.
The analyzer system bus was used to choose the frequency range and number of points
(51, 101, 201, or 401). The IF frequency was 20 MHz and the lowest frequency was 45
MHz. The HP-IB data bus was used for automatic operation of the VNA.
Several test sets were introduced to cover different frequency ranges: 45 MHz–26.5
GHz, 2–18 GHz, 2–20 GHz, and 45 MHz–20 GHz. The HP–8510 rack system consisted
of one test set, one source, and an HP–8510 display / user interface. The display was
black and white and a magnetic tape was used to collect test results. This system became
known as the HP-8510A.
In 1987, Wiltron introduced the 40 GHz 360 VNA system. This system had a color
display and extended the lower frequency to 10 MHz. Users enjoyed the competition
between HP and Wiltron and benefitted by getting the best performance per dollars
or dB/$.
The upper frequencies of these VNAs started to increase to 50 and 65 GHz in coaxial
and to 110, 220, and 325 GHz in waveguide.
100 Mohamed Sayed and Jon Martens
Table 5.2 HP 8510 Vector Network Analyzers (VNA) systems
Date Company Country Model number Frequency range
1984 HP USA HP-8510 26.5 GHz

1987 Wiltron USA 360 40 GHz
HP USA HP-8510 50 GHz
Wiltron USA W-360 65 GHz
Wiltron USA W-360 110 GHz
HP USA HP-8753 3/6 GHz
HP USA HP-8720 20/40 GHz
1994 Wiltron USA 37XXX 20/40 GHz
1998 Wiltron USA MS462XX 9 GHz
Rohde & Schwarz Germany ZVT8 8 GHz
To extend the HP-8510 system to 40 GHz, a doubler was installed into the test set
using a 20 GHz HP-8340 source. The 50 GHz HP-8510 system was introduced along
with the APC-2.4 connector. A set of calibration kits, cables, and verification kits was
also introduced for each system and each different frequency range.
Many applications in material, measurements, antenna measurements, and radar
measurements were shown and used by customers during the period 1984–2001.
A few years later, a one-box VNA was introduced for RF in the HP-8753 and
for microwave (HP-8720). Wiltron introduced the Scorpion Network Analyzer which
included two sources and noise figure measurements up to 9 GHz. Rohde and Schwarz
introduced an 8 GHz ZVT8 which included 8 measurement ports. Table 5.2 shows VNA
model numbers and years for this period.
Several technologies were quickly developing: 1) low-cost solutions for low-frequency
and production environments; 2) compact sources with fast tuning times and high reso-
lution; 3) calibration routines and kits for higher accuracy; 4) multiple-port applications
for production systems and solutions; 5) wideband components such as couplers, cables,
and mixers to extend the VNA bandwidth range.
Pulsed measurements for the on-wafer application of high power devices were intro-
duced by HP and Wiltron. Load-pull measurements under pulsed bias and pulsed RF
were developed by both companies to test high power devices on wafer. Wafer probes
were developed by Cascade Microtech and other vendors.
The customer’s need to extend the frequency range and dynamic range of the VNA
motivated vendors to develop high-resolution sources, receivers, and calibration rou-
tines. During 1984–2000, customer seminars were developed by different vendors and
were presented all over the world. Engineers and scientists attended annual Interna-
tional Microwave Symposiums (IMS) to view the latest VNA systems, solutions, and
applications.
The automated production of devices using either HP’s VNA or Wiltron’s VNA made
great progress during the period 1984–2001. More details about this history can be found
in Rytting’s paper [1].
On wafer
Performance and
System All-in-one box
application
Automated fast
Lower cost
measurement
Fig. 5.1 VNA growth and applications.
5.2.3 Evolution of VNA to the Present – 2001–2012

During 2000–2001, all-in-one VNA systems were introduced by Anritsu/Wiltron
(Lightning) with a 65 GHz range and HP/Agilent (PNA) with a 50 GHz range.
The need for wider frequency ranges and higher output power has sped up the intro-
duction of the Vector Star from Anritsu and the PNA-X from Agilent during the last
few years.
Rohde and Schwarz later introduced the ZVA with four internal sources up to 67 GHz
and extended the frequency range up to 220 GHz with external millimeter wave modules.
VNA growth and applications over the last generation are shown in Figure 5.1. The
evolution of the VNA is shown in Figure 5.2.
5.3 Authors’ remarks and comments
VNA in this chapter refers to Linear VNA. The NVNA and related LSNA are discussed
in Chapter 12. In addition, waveform engineering, which is currently being pursued by
a number of researchers, is not addressed in this chapter.
There are several factors which will be important in the near future: cost vs. perfor-
mance, digital designs to 40 GHz and higher, modeling and verification for high power
devices, time to market of new technologies, and the role of microwave measurements
technology for future devices and mobile technology.
5.4 RF and microwave VNA technology
The most basic objective of the VNA (and many related instruments) is to measure
S-parameters or the constituent wave quantities (e.g. a1 , b1 …). In this most basic form,
one must acquire incident and reflected waves at each port of interest while providing
input signals at the different ports. Carrying this simplistic picture forward, one then
requires a signal source, some receivers, and some way of separating the incident and
reflected energy. The purpose of this section is to explore some of the elemental blocks
Pulsed
(Chapter 15)
Load-pull Multi-ports
(Chapter 13) (Chapter 9)
NVNA E-Cal
VNA
TDR Power
Noise
(Chapter 10)
Fig. 5.2 Evolution of VNA.
of a VNA, the performance considerations of relevance, technological evolution over the

more recent history of VNAs, and the effect the blocks have on a variety of different
measurements. Some of the historical concepts discussed earlier are used by inference
and more details can be found elsewhere (e.g. [1–5]).
The concepts above lead to an elemental block diagram (for a 2-port case) like
that shown in Figure 5.3, although many variations are possible. Some of the basic
elements are:
• One or more signal sources having at least controllable CW/swept frequencies with
sufficient spectral purity that measurements can be made. It is also preferred for the
power to be controllable.
• Some directional devices (see Chapter 1) for separating incident and reflected waves at
the ports. In some cases these devices need not be physically directional, but they could
be generalized splitting devices of sufficient stability that they can be computationally
directional.
• If there are fewer sources than ports or if there are more or fewer receivers than ports,
there must be some means of switching signals.
• One or more receivers, usually incorporating down-converters, to take the incident
and reflected waves down to some convenient IF for processing.
a1 a2
Receiver Receiver
IF
processing/digitizing
Receiver Receiver
b1 b2
Port 1 Port 2
Fig. 5.3 One possible VNA block diagram (for a 2-port case) is shown here that illustrates the key blocks
and the flow paths to be discussed.
• An IF section and digitizer to process the converted wave amplitudes into a form useful
for computation and display.
Among the possible variations of Figure 5.3, one could use a source per port instead of
switching one between two ports. The coupling devices could also be repositioned and,
of course, there could be many more ports or just one. The point of the diagram is that
the functions listed above are generally present in one form or another.
5.4.1 Sources
Historically, the source in a VNA has taken many forms ranging from simple analog
sweepers in the earliest implementations to complex synchronized synthesizers in more
modern instruments. Sweepers can be quite fast (and before about 2000, they were
generally faster than synthesizers), but the spectral purity is not as good and there are
potentially synchronization issues since the LO and IF systems must be semi-coherent
with the source system (resulting in sometimes substantial frequency errors). As a result,
more of the recent VNAs are synthesizer-based. While there are an infinite number of
variations possible, a core block diagram of a synthesized source is shown in Figure 5.4.
Considerably more detail on synthesizer structure is presented in Chapter 3 and, increas-
ingly, more of these design concepts are migrating on to VNA platforms. The usual
difference is the higher importance that is assigned to point-to-point tuning speed in
VNA applications than in classical synthesizer applications, but this line has also been
blurring over time.
VCO/YTO or other
controlled oscillator
Reference Phase-locked loop
Feedback circuitry
Mixing, multiplication,
Control signals
division, modulation,…
Power leveling
To signal distribution and

separation
Fig. 5.4 A very generic source block diagram to illustrate some of the choices to be made.
Among the issues to be considered:

• How fast can one get from one frequency to another? While this may not be the
dominant part of the measurement speed, it can be relevant.
• How clean is the signal (in terms of phase noise, harmonics, and spurs)? In doing only
S-parameter measurements, some of these parameters may not be that important since
we know what frequency we are measuring (and there is only one per point) and a
narrow IF bandwidth can be used to reduce noise effects. As applications proliferate
(IMD, mixer, and nonlinear measurements) and speed needs increase, however, one
must pay more attention to the spectral purity.
• How is the frequency plan organized (multipliers, dividers, and mixers operating on a
base range)?
• How is power control done? How much power range and accuracy is available? This
becomes more important for nonlinear and quasi-linear measurements.
The block diagram in Figure 5.4 is extremely general, since there are many possible
structures and large portions of the source may be generated digitally. The reference can
come from a crystal oscillator or from some other synthesizer. Feedback for the loop is
shown (by the dashed line) to be coming from the VCO, but it could also come from
elsewhere in the system (e.g. from a receiver). The oscillator output may be frequency-
converted or may be modulated. In some cases, the source may not even be locked
(although there are accuracy penalties for that as discussed in the sweeper case).
As discussed before, the source need not actually be locked as Figure 5.4 might
suggest. The sweeper type of structure can be quite fast and devoid of some spectral
artifacts. The downside is that controlling the timing relationships between source, LO,
and acquisition can be challenging. Doing this over temperature and aging can be more
difficult and normally requires sophisticated internal calibration structures. Integrating
more complicated applications (such as mixer measurements) requiring external sources

or otherwise changing the sweep dynamics adds additional challenges.
A fully synthesized approach avoids most of these problems at the expense of some
different types of control complexity. In order to make a fully synthesized version fast, but
still with good spectral purity, more careful loop design and perhaps more sophisticated
control electronics are required. It is not the intent of this chapter to cover details of
phase locked loop and synthesizer design (for details, see for example [6]) but there are
some key points that may aid the discussion:
• Generally, the wider the loop bandwidth, the faster the settling time. Concomitantly,
wider loop bandwidths generally lead to more phase noise in an integrated sense.
• From a measurement point of view, it is the settling of the final receiver IF that is
somewhat more important than the independent settling of the source and LO. If the
source and LO can settle together, faster net measurement times are possible (there
are limits; one must be on frequency to within a certain tolerance).
• Generally the loops will settle faster (to within a fixed tolerance) for smaller frequency
steps. For larger steps, dynamic loop response changes can help.
• Increasing levels of (and frequency ranges of) fully digital synthesis can greatly speed
the sweep at the cost of some spur control complexity.
At some point in the process, locking is usually required. The next question is where the
locking is performed. The simplest approach may be treating both the source and LO as
separate synthesizers with their own integrated phase-locked loops with shared reference
frequencies at some level. Other possibilities include locking through the receiver, essen-
tially locking the LO to the IF or locking the source to the IF (sometimes termed follower
mode and source-locking, respectively). This can reduce the individual loop complex-
ities somewhat and can lead to a very clean received signal since the IF becomes the
locking reference. This approach does complicate the application space somewhat since
one receiver must be made the locking parent and hence no longer has meaningful phase
information. Also the source and LO must have a fixed offset to enable the loop to close
which makes mixer, IMD, and other translating measurements quite difficult.
There are many technological decisions to be made on the individual PLLs and most
of these are beyond the scope of this chapter, but a few key comments can be made.
Historically, YIG oscillators have often been used for the source with a source-locking
architecture. While the phase noise of such oscillators is quite good (particularly far from
the carrier), they tend to be slow (at least in broadband configurations). More recently,
VNAs where all sources are based on varactor-tuned VCOs, which can tune much faster,
have been introduced. The trade-off is degraded phase noise at offsets much larger than
the loop bandwidth, but even these differences have been shrinking as VCO technology
improves.
The fine-tuning structure of these loops has also changed in recent history. Fractional-N
structures (e.g. [6]) are very popular and can offer fine-tuning resolution with decent
spurious and noise performance. Increasingly wide bandwidth direct digital synthesizers
have become more common and have had ever-improving spurious performance. The
fine-tuning capabilities of such structures are needed since the VNA tuning resolution
Loop amp+ DAC

–
Detector
Variable
attenuator/
modulator Test port
Fig. 5.5 A very simplified ALC loop.
must typically be of the order of 1 Hz (or better if high-order multipliers are part of the
system).
Another important aspect of the source side of the system is power control. Aside from
having a vague idea of what the DUT is being driven with, swept power measurements
are increasingly important to the VNA user for measurements such as gain compression,
IMD vs. power, harmonics vs. power, etc. Thus a reasonably accurate and wide range
ALC is of importance. Complicating things, like so much else in the system, is that this
leveling system must be fast enough to keep up with the measurement.
Leveling subsystems are used in many applications and are conceptually quite simple.
They use a power detector of some kind and, in the context of a negative feedback loop,
compare the detected output to some desired reference voltage (usually from a DAC)
and feed the result to a power modulator of some kind (see Figure 5.5).
For the purposes of illustration, a number of assumptions were built into this diagram
that are not mandatory:
• A coupled detector is shown for power detection.
◦ Sometimes non-directional splitters are used instead of a coupler. This is much

simpler and can lead to a more fixed delivered power instead of incident power at
the potential expense of stability.
◦ Arrays of detectors are sometimes used for improved control range and even a
thermal sensor could be used, although there may be a speed penalty.
• A variable attenuator is shown for power modulation.
◦ Amplifier bias is sometimes used for this kind of control. Harmonic generation in
that case could be a concern as the requested power is reduced.
◦ Cold FET and PIN diode attenuators are both popular for variable attenuators. PIN
diode structures often have an advantage in power handling and FET structures
often perform better at low frequencies (although there are exceptions to these
generalizations and the technologies are constantly evolving). Hybrids are possible.
• A simple loop amp (often an integrator) is shown.
◦ Multi-stage and distributed loop amps are often used for more control of loop gain.
◦ Variable poles are often used for stability in different operating modes.
Example variable attenuator response

Attenuation (dB) 0
–10
–20
–30
–40
–5 –4 –3 –2 –1 0
Control voltage (V)
Fig. 5.6 An example response curve of a commercial voltage-variable attenuator.
Level- Loop amp+ DAC

dependent gain –
Linearizer
Detector
Variable
attenuator/
modulator
Test port
Fig. 5.7 A more complete ALC block diagram.
The issue of loop bandwidth is an important one to consider. Since a VNA has to operate
over wide frequency ranges and, often, wide power ranges, the overall loop gain will not
be flat. To see this, consider the attenuation curve of a commercially available voltage-
variable attenuator (Figure 5.6).
The slope variations in this curve represent changes in loop gain. If this was uncom-
pensated, the bandwidth of the loop could become very small at some states (making the
measurement slow at low attenuation levels) and very large at other states (potentially
leading to oscillation at higher attenuation levels). In addition to simple level-dependent
gain changes, there may be other frequency-dependent gain changes such as when one
moves from a fundamental source band to a multiplied source band that may use a differ-
ent variable attenuator. Since detectors have nonlinear responses over wide power ranges
as well, some linearization may be desirable again to keep loop gain relatively flat. From
all of these complications, one may end up with a leveling system that looks more like
Figure 5.7.
5.4.2 Switches
RF switches are needed in VNAs for a number of reasons including the desire to allow
one source to drive two or more ports (thus saving the expense of multiple sources) or
to selectively route to multiple receivers (e.g. in a multi-port scenario or to different

application-specific receivers).
Very often, the demands on the switches can be extreme in terms of isolation, inser-
tion loss, bandwidth, and, perhaps, power handling/linearity. Using a 2-port VNA as an
example, there is usually a main switch (normally called a transfer switch even if it is
SPDT) allowing one source to drive port 1 or port 2. The isolation of this switch directly
translates to the raw isolation of the VNA. The insertion loss and linearity directly affect
what the maximum available port power can be and its bandwidth can limit that of the
VNA. For a high-performance microwave VNA, this can be a challenging combination.
In the distant past, electromechanical switches were sometimes used due to their
favorable insertion loss/isolation ratio. The repeatability of these switches, typically
no better than a few hundredths to a tenth of a dB at microwave frequencies, led to
some measurement errors. Also, the lifetime of many mechanical switches does not
exceed 10 million cycles. Even at a slow sweep rate of one sweep per second (which
may be relevant depending on how the switch is used), this switch would last less than
3000 hours.
Thus electronic switches are normally used, typically either a PIN diode (e.g. [7]) or
cold FET circuit (e.g. [8]), or some combination thereof. It is beyond the scope of this
discussion to analyze the device physics in detail but a quick summary is provided below
(see, for example, [7–9] for more information).
A PIN diode consists of heavily doped P and N layers surrounding a relatively thick
intrinsic layer (hence the acronym). Because of this thickness, the reverse biased capac-
itance of the diode is quite low compared to many other diode types. This leads to better
isolation when used in a series construction and less insertion loss in a shunt topology.
When forward biased, carriers are injected into the intrinsic layer but do not recom-
bine immediately. This leads to some complications at lower frequencies since the RF
frequency can be on the same scale as the recombination rate and distortion occurs.
A typical cold FET switch is just that: a MESFET or similar structure setup with no
drain bias. When the gate is biased strongly negative, no carriers are available in the
channel and the device provides reasonable isolation in a series sense. Like the PIN
diode, the off capacitance (drain to source) is fairly low due to the geometry, so shunt-
topology insertion losses can be low as well (although typically worse than with a PIN
structure). The elevated capacitance can be mitigated by embedding the switch in a
transmission line structure. When the gate is near ground potential, carriers are available
in the channel and a relatively low series resistance is available. Unlike the PIN diode,
the recombination time remains fast so there are few low-frequency effects. Since one
is usually operating against a 0-bias limit, there can be linearity issues, although these
have been overcome at least in part with more novel topologies.
Whichever technology, or combination of technologies, is chosen, the issue of switch
topology is critical. For simple applications requiring low isolation, a single series-shunt
element per arm may be appropriate. In some cases, even a single element can be used
but there may be severe match implications on a multi-throw switch. When high levels
of isolation are required, more elements are often used per path. There are many choices
one can make about the combinations of elements, but a few items can help:
• Series elements generally become less effective at higher microwave frequencies and
more shunt elements will be used in that frequency range.
• Sometimes series-shunt pairs are available as a single die or cell and they are often
convenient to bias that way.
• Proper allocation for biasing inductors must be made (for PIN switches primarily) and
their layout is critical since above 50 GHz or so, bias circuit parasitics may contribute
as much to insertion loss as the switch itself.
• Isolation may end up being limited by radiative effects thus making housing design
and layout quite important.
• As has been pointed out in the literature, the switch spacing in higher isolation
structures is quite important due to the standing waves that will appear between
switches.
• Terminating switches are often required, which usually means an additional branch to
a load is needed at the output ports although there are other approaches.
As an example of some of these behaviors, an empirical curve of isolation and insertion

loss for a SPDT switch using a particular PIN diode technology operating at up to 70
GHz is shown in Figure 5.8. The return loss of the output is roughly constant in on-
and off-states due to the presence of a terminating load (with phasing differences and
multiple reflection interactions causing differences). The isolation in this measurement
is largely noise floor limited but one can see that the insertion loss to isolation ratio
supports a reasonable dynamic range.
5.4.3 Directional devices

Key to the concept of reflectometry is the directional device used to collect the incident
and reflected wave energy (see also Chapter 1). While there are a number of ways to do
this, broader band microwave VNAs tend to rely on directional couplers, while RF VNAs
may use a directional bridge (or the above concepts may be combined). It is beyond the
scope of this chapter to delve into the theory of the directional devices (see, for example,
[9]) but we need the following definitions (see Figure 5.9 for port assignment, reverse
coupling assumed for the drawing):
Coupling: S24 (sometimes coupling is defined to include the insertion loss, much like
S41 with a perfect reflect connected)
Insertion loss: S21
Isolation: S41
Directivity is often defined as |S24 /S41 |, although variations on this definition may
include insertion loss as part of the numerator (in the sense of reducing the numerator).
Coupling is usually dictated by the signal levels needed by the rest of the system
subject to the constraint that directivity usually worsens (for a broadband coupler) if the
coupling gets too tight (i.e. |S24 | gets larger). With these constraints in mind, coupling
factors usually end up in the 10–20 dB range although there are exceptions. Of course,
one wants a minimal forward insertion loss (to maximize the available port power) and
Example switch match vs. state

0
Reflection cofficient (dB) on off
–10
–20
–30
–40
0 20 40 60
Frequency (GHz)
(a)
Examples switch insertion loss and isolation
0
–20
IL Isolation
Transmission (dB)
–40
–60
–80
–100
–120
–140
0 20 40 60
Frequency (GHz)
(b)
Fig. 5.8 Example insertion loss and isolation of a broadband, high-isolation switch construction.
3 4
1 2
Fig. 5.9 An example coupler block is shown here to help with the definitions. Assume the path 2-4 is the
desired coupling direction.
reasonable match (since this may dictate the raw port match and is usually connected to
directivity).
The wildcard, which is principally a function of the construction techniques and level
of assembly tuning, is directivity or isolation. In view of the power of VNA calibrations,
one may wonder how important these raw parameters are to overall system performance.
In an instantaneous sense, the answer is usually not significantly. In the longer term (in
the sense of calibration stability), it can be considerably more important.
Before exploring these comments, we must revisit the concepts of residual vs. raw
parameters (such as directivity and source match). The raw parameters describe the
physical performance of the components involved such as the directivity described above
for the directional device. The residual directivity is that left after the calibration and
also describes the quality of the calibration components, the calibration algorithm, and
the calibration process. This concept is discussed in more detail in Chapter 8. It is the
residuals, at the time of DUT measurement, that describe the measurement uncertainty
to a great degree, not the raw parameters. Now an individual DUT may be sensitive to
the raw parameters (e.g. an amplifier may or may not be stable for a given raw port match
on the VNA) but the measurement itself can be largely invariant to them.
To see this, consider two calibrations performed on a VNA. The first is with the VNA
as it is normally configured, with a raw directivity of about +10 to +15 dB across the
band 70 kHz−70 GHz. Also, perform a calibration with a 10 dB pad on the test port
so the raw directivity in that case (ignoring pad mismatch so this is an upper bound)
is −5 to −10 dB. With the two calibrations, measure the return loss of the same delay
line. The results are shown in Figure 5.10 and indicate agreement (to within connector
repeatability limits) even in the deepest notches. This indicates the residual directivities
are nearly identical.
In a practical sense, however, it is important since the raw parameters have an impact
on the stability of the calibration. Consider the directivity correction. In a reflection
measurement, the directivity error adds to the DUT’s reflected wave to produce the net
Match of a delay line measured with positive and

negative raw directivity systems
0
–10
|S11| (dB)
Positive
–20
Negative
–30
–40
0 20 40 60
Frequency (GHz)
Fig. 5.10 The impact of positive and negative raw directivities on a calibrated measurement. As long as the
environment is stable, both calibrations are roughly equivalent.
Actual DUT
Good raw directivity case
Raw measurement
Raw directivity Correction
Poor raw directivity case Actual DUT Raw measurement
Raw directivity Correction
Fig. 5.11 The mathematics of directivity correction for two different raw directivities.
measurement. In the correction, the directivity is subtracted out (as well as other tasks
being performed). If that subtraction is small in magnitude, a small drift in the actual
amount of directivity does not affect the end result very much. If the subtraction is large,
however, a fairly minor drift in that directivity vector can result in a substantial change
in the final result (see Figure 5.11).
Thus one often strives for the best directivity possible within the boundaries of the
other constraints. In the example of Figure 5.10, both measurements were done shortly
after the calibrations. Had the delay line been measured several hours after the calibration
in a thermally dynamic environment, the results might have been quite different.
One of the other constraints on the directional element is bandwidth. While the upper
end is relatively easy to understand with the collapse of directivity under the wavelength
limits, the low end is often misunderstood. Obviously as the coupling section becomes
electrically short, the coupling factor must typically fall and often at a 6 dB/octave rate.
Thus the available signal level decays rapidly and signal to noise becomes a problem.
Directivity usually also suffers at this end but more for reasons of match problems,
although this is not a correct generalization for some of the more exotic coupling
topologies.
Bridges are a slightly different structure and do bear some resemblance to the classical
Wheatstone bridge idea. The difficulty from an RF point of view is how to generate the
non-ground referenced nodes. Typically this is done with a transmission line balun,
although there are other possibilities (including entirely with active elements). This
in turn helps set the bandwidth along with the parasitics of the lumped components
being used.
Reasonable directivity can be maintained over large frequency ranges through proper
balun design. The example shown in Figure 5.12 could be further optimized by use of a
more elaborate balun structure at the expense of some insertion loss.
Example bridge directivity

40
35
30
Directivity (dB)
25
20
15
10
0
0.01 0.1 1 10 100 1000 10000
Frequency (MHz)
Fig. 5.12 Directivity of an RF bridge structure. Reasonable performance down to very low frequencies is
possible with the right balun structure.
Non-directional splitters (including simple three resistor designs) are also sometimes
used, particularly for the reference waves. In the reference wave case, the argument can
be made that it is far enough from the test port that load-impedance-induced effects are
small enough to not impact stability significantly. These non-directional structures have
also been used on test ports where the directivity is entirely computational. Stability may
be a prime consideration and sometimes extensive thermal stabilization efforts are made
in these contexts. Cost, simplicity, and size are the obvious advantages.
5.4.4 Down-converters (RF portion of the receivers)

An entire book could be generated on receiver design, so this discussion is limited to some
VNA-specific topics/decisions and some general analyses. In the context of a broadband
microwave VNA, one currently has to perform some means of down-conversion. There
are many choices to make:
How many down-conversion stages and with what frequency plan?

Classically, measuring receivers have used multiple up and down-conversions to pro-
vide better image rejection and to allow a more flexible frequency plan for the purpose
of avoiding spurs. In a VNA, the image is usually less of a concern since there is one
known signal present (and perhaps its harmonics) that one wants to measure. As the
application space changes to include IMD and other more spectrally rich measurements,
the image behavior takes on added performance, but there are often ways around the
issue. Sources are increasingly clean and converters increasingly linear so that spurs
have been a declining problem. Couple that with cost, complexity, and again with the
situation of a single known signal and many of the reasons for multiple conversions
are mitigated. Depending on how the IF is implemented and for other signal processing
reasons, two conversions are sometimes desired but it is less common now to go beyond
that. For reasons of stability, homodyne receivers have been avoided in recent years but
that may change as the adaptive conversion circuitry used in non-measurement receivers
improves.
In an ideal scenario, one would be able to fundamentally mix over the entire frequency
range of the instrument. This would have the lowest spurious possibilities and best
receiver noise figure (and probably best linearity). For most middle microwave and
lower-frequency systems, this is the choice made.
For broadband microwave systems, this can get very expensive since the isolation
chains (see Figure 5.13) have to run over this full frequency range, somehow provide
enough LO power for the converter (10–20 dBm typically), and provide 100–120 dB
of round trip isolation. One could imagine having a separate LO for each of the four or
more converters (Figure 5.14) but this gets even more expensive and maintaining phase
LO
Fig. 5.13 Another four-channel receiver architecture is shown here. Now a single LO is shared between the
four down-converters and amplifier chains are used to ensure channel-to-channel isolation. More
or fewer amplifiers could be used and, in some circumstances, isolators or filters can be used
instead.
Fig. 5.14 Concept schematic of a four-channel receiver where each down-converter has its own LO. This
can be expensive at higher frequencies and it can be challenging to ensure adequate phase
synchronization between channels.
synchronization can be challenging. Thus typically, some kind of harmonic conversion

process is used to limit the required range of the LO. The harmonic conversion may
occur in the final converting device itself or in some pre-multiplier but the important
concept is that often, most of the distribution is done at lower frequencies.
Converter type?
There are many possible configurations of converters and the distinctions can be subtle.
As indicated above, fundamental mixing (e.g. [10]) is generally used at RF and up
into the middle microwave frequency ranges and that usage may continue to march up
in frequency. The early VNAs used relatively low LO frequencies (tens to hundreds
of MHz) and higher-order harmonic conversion via samplers, in part because of the
technologies available (e.g. [11]). The latter structure has many image responses and
relatively low conversion efficiency. Depending on how the device is implemented, the
linearity can be quite good. The sampler method used involved some form of edge
sharpening (originally a step-recovery diode, SRD and more recently using nonlinear
transmission lines) followed by a passive differentiator to create a sharp pulse. This pulse
turned on the sampling diodes and captured a small window of RF energy at the period
of the LO. After filtering, this created an equivalent IF for later processing. The concept
is shown in Figure 5.15.
More recently when harmonic conversion has been desired (higher microwave and into
the mm-wave range), the trend has been toward harmonic mixers and high LO samplers.
Both of these use relatively low harmonic orders and higher LOs (into tens of GHz) and
obviously have fewer image responses and tend to have better conversion efficiencies if
for no other reason than less energy redistribution but this can vary with implementation.
The distinction between a harmonic mixer and sampler of this type can be very subtle
(e.g. [12]) and may come down to the degree of LO waveform modification performed
prior to the physical converting device. Even this may be a distinction without value
as the physical converting device can be engineered to do the waveform modification
itself. Instead, we consider some differences with greater or lesser degrees of waveform
modification at relatively low orders of harmonic conversion, as is commonly seen in
higher frequency VNAs.
SRD
Edge sharpener Differentiator
Edge
sharpener+SRD
Optional bias
Tee
Fig. 5.15 Block diagram of a sampler construction as used in some VNAs.

0 Conversion vs. topology

Conversion efficiency (dB)
–10
–20 Unshaped (LO<20 GHz)

Unshaped (LO<10 GHz)
–30 Shaped (LO<10 GHz)
–40
0 20 40 60 80
Frequency (GHz)
Fig. 5.16 Example plot of conversion efficiency versus frequency and converter LO structure.
Conversion efficiency is of interest since this plays a role in determining the VNA’s
dynamic range. If the final converting device sees an unmodified sinusoid of some lim-
ited frequency range, at higher frequencies less RF energy is captured per cycle and the
effective conversion efficiency decreases. The higher the implied harmonic order, the
more the decrease. If the LO waveform is highly shaped, the conversion efficiency tends
to be flatter but starts off at a lower level, since energy is being somewhat dispersed to the
images in all cases. This idea is illustrated in Figure 5.16 where the LO is constrained into
two different ranges for the unshaped case and a doubly-balanced structure is assumed
that favors odd conversion products. Thus for RF frequencies up to the LO limit, funda-
mental conversion is used, then 3x the LO is used, then 5x and so on. In the plot, one can
see the transitions clearly for the unshaped topology for the first few and then it becomes
somewhat muddier as the interaction of many mixing products may occlude the picture.
The higher the LO range one can use, the less roll-off one sees in the unshaped cases
and the higher the baseline conversion efficiency in the shaped cases. The downside is
that the LO distribution becomes increasingly expensive and complicated the higher the
LO frequency goes.
The image responses are also of concern, particularly in the non-S-parameter applica-
tions. As might be expected from Figure 5.16, the shaped case image responses will all
be of roughly equal conversion efficiency and there may be a fair number of them. In the
unshaped case there is a roll-off, so if one is operating on a higher harmonic, lower-order
images (which are undesired in this operating mode) are higher in conversion efficiency.
The details of the frequency plan will determine the relative merits of these two cases.
Linearity is also of interest since it impacts the compression point referenced to the
VNA port (as modified by couplers, pre-amplifiers, and other networks). If harmonic
mixing is used where the conversion device itself (in the classical diode sense) is respon-
sible for generating LO harmonics, higher drive levels and incomplete saturation are a
common effect. As a result, linearity tends to be lower for equivalent device technolo-
gies than for a fundamental mix or one where the LO waveform was altered prior to the
final conversion device. This concept is shown in Figure 5.17 where the input-referred
Intercept point vs. converter topology

example
30
Fund. mix
IIP3 (dBm)
20 N=5 harmonic mix

N=5 harmonic mix
10 Shaped
0
0 20 40 60
Frequency (GHz)
Fig. 5.17 Plot of converter input-referred intercept point vs. frequency and LO structure.
Very low IF: high self-

conversion noise possible
Very high IF: low self-
conversion noise but complex
A/D possible
Increasing IF
Fig. 5.18 Comparison of the noise effects of high and low IF frequencies.
intercept point for a collection of converters is plotted versus frequency where it was
attempted to hold device technology and equivalent drive levels constant. The absolute
values of the intercept point and the frequency dependencies will, of course, be a strong
function of the technology employed and may not hold for more exotic topologies.
5.4.5 IF sections
The IF section of any receiver often gets less attention from a technological point of view
but it is a critical component of instrument performance. The possible floor for speed,
dynamic range, and trace noise can be set in this section, although systems are usually
designed such that noise performance is not IF-limited. One of the first questions is what
IF frequency (or range) should be used. If it is very low, then the A/D circuitry can be
simple, but converted LO phase noise becomes more of a problem (depending on the
conversion structure) as suggested by Figure 5.18. A very high IF frequency requires a
more complex A/D structure and potentially more noise injection at the IF level, but the
Classical oversampling
An undersampling
example
Fig. 5.19 Pictorial diagrams of undersampling and oversampling.
DC: in-phase
component
Incoming IF
at f0 cos(2πf0)
sin(2πf0)
DC: quadrature
component
Fig. 5.20 Concept of synchronous detection, where the final IF is down-converted to DC for A/D sampling.
noise and spur contributions from the RF section are usually lower. Certain applications
may demand certain ranges of IF frequency (larger bandwidths needed, for example).
Once the IF frequency is selected, the frequency plan for the A/D system usually comes
next. Classically, an oversampled structure would be used to allow the extraction of maxi-
mal spectral information. This requires a faster A/D clock and places more of a constraint
on cleanliness and on the A/D converter. Returning again to the concept of knowing the
signal that is being measured, one could move to an undersampled structure which can
improve noise and simplicity. The downside is an increase in spurious responses that
may require more analog filtering. The classical difference between undersampling and
oversampling is illustrated in Figure 5.19.
Another method of detection, termed synchronous detection, works by performing a
final down-conversion to DC in an in-phase and in a quadrature sense. Since the A/D
converters are operating at DC, the clocking structure can be simpler. Like homodyne
systems, however, there are DC defects such as offsets and channel skews that must
be minimized and/or corrected for. The concept of synchronous detection is shown in
Figure 5.20. This approach was used in many of the earlier VNAs but has been largely
supplanted by some of the previously discussed techniques as ADC technology has
advanced.
Digital
hardware:
IF from
receivers
Digital filtering
ADC
for noise
control;
Simple filtering for decimation and
aliases, images, other processing
known spurs
Fig. 5.21 Common IF filtering scheme where some simple filtering is analog but most of the variable
filtering (and narrow bandwidth filtering) is done digitally.
Implementation of filtering is another major topic. Usually some analog filtering is

required to handle aliases, images, and other known large interferers. This might be
simply to avoid overloading any IF amplifiers or theA/D converters. Beyond that, filtering
is required for noise reduction (and sometimes reduction of close interferers). This is
usually termed an IF bandwidth in VNAs, although an analogous concept of resolution
bandwidth applies to spectrum analysis. Historically, this was done with a collection of
analog filters but these had issues of stability and measurement accuracy when changing
the setting between calibration and measurement. More recent instruments implement all
of this filtering digitally. Thus a common filtering implementation may look something
like that in Figure 5.21.
5.4.6 System performance considerations

It may be useful to examine how the performance of some of the blocks just discussed
impact the overall system performance parameters.
• Frequency range
◦ Obviously set by source, receiver, and signal separation.

◦ Applications requiring only simple S-parameter measurements have clear selection
criteria except sometimes extreme out-of-band results are needed for characteri-
zation or regulatory reasons. Other applications requiring harmonic measurements
have their own demands.
• Dynamic range
◦ Two parts to this: noise floor and maximum power. The latter is addressed below
(see compression or port power).
◦ Noise floor is impacted by front-end loss (couplers and attenuators), conversion
loss/gain and initial IF gain stages. An RF pre-amplifier can help at the potential
expense of compression and stability. Noise floor and dynamic range are often
specified in a 10 Hz bandwidth. Scaling rules help at other bandwidths over at least
a limited range.
◦ Access loops can be used to improve the reference plane noise floor at the expense
of compression limits (skipping coupler loss).
• Trace noise
◦ Usually measured far away from the noise floor so that is not an impact.
◦ LO/source phase noise folds over and converts to the IF.
◦ IF system noise.
• Port power
◦ Source power and loss between source and port are determining factors. Maximum
power levels of +10 dBm and higher are increasingly common. Step attenuators
are often used to reach −90 dBm or lower.
◦ Compression limits of switches and other test set components may play a role.
• Power accuracy
◦ Generally limited by the structure of the ALC loop, the temperature compensation
methods employed, the calibration procedure, and the power ranges involved.
• Harmonics
◦ Usually from the source and related components.

• Compression
◦ Usually the converter linearity sets this limit although any front-end RF amplifiers
can sometimes contribute.
◦ RF attenuators can help in some applications.
◦ IF systems can sometimes contribute.
• Raw port parameters
◦ The front-end components (couplers, attenuators, transfer switches, etc.) tend to set
these parameters.
• Residual port parameters
◦ Generally the calibration kit and calibration algorithms set these limits. The first
instrument parameters to affect the residuals are usually linearity-related.
• Stability
◦ Many factors, some of which are hard to measure (measurement dynamics).

◦ Temperature stability of couplers, switches, and cables.
◦ LO power stability and sensitivity of the converter to its changes.
◦ Linearity of system and stability of port power.
• Measurement speed
◦ A number that is dependent on many variables including frequency step size, num-
ber of points, display setups, power levels, and external data transfer setups. It is
therefore very hard to compare amongst applications or vendors.
◦ Source tuning speed, receiver acquisition time, digital hardware processing time,
and software overhead can all play a significant role.
All of the blocks discussed play a role in how the instrument performs and how these spec-
ifications are created. In terms of non-S-parameter measurements, some of the impacts
will be discussed in later chapters. While the above are usually still important, the
source purity aspects (harmonics, spurs, and phase noise) take on added importance in
quasi-linear and nonlinear measurements as does receiver compression. In time domain
contexts (when transformed), stability and repeatability play a larger role. Residual port
parameters are often dominated by calibration considerations and those are covered in
Chapter 8.
5.5 Measurement types in the VNA
5.5.1 Gain, attenuation, and distortion

Basic VNA measurements are gain, attenuation, and distortions. This is similar to mea-
surements with scalar network analyzers. However, the VNA is more accurate, based
on correction availability and the greater dynamic range (tuned receiver vs. a broadband
receiver). Minor variations in these measurements (in the quasi-linear realm) include
DUT harmonic output and intermodulation distortion.
5.5.2 Phase and group delay

The change of phase between the input and output to and from the DUT determines the
phase differences going through the DUT. The display of phase on the VNA is usually
between +180o and −180o although unwrapping of phase (or absolute phase displays)
is often available. Group delay is calculated from the derivative of phase with respect to
frequency. This derivative is normally calculated numerically, so the interval over which
it is calculated can be important (termed the “aperture”): a larger aperture effectively
applies smoothing to the data but reduces the resolution (of group delay distortions) in
frequency.
5.5.3 Noise figure measurements

More recent VNAs can measure noise figures and noise parameters. Hot/cold and cold
only techniques are used for this application. Menus can guide users to connect, calibrate,
and measure accurately the DUT noise parameters and figures.
More details are in Chapter 10.
5.5.4 Pulsed RF measurements

Some applications for pulsed RF measurements using VNA are:
• Eliminating thermal effects when testing high power devices on wafer;

• Troubleshooting high power devices to pinpoint trapping effects;
• Testing DUT under similar environments for real-life pulsed applications;

• Optimizing DC pulsed bias to minimize over- and under-shooting;
• Monitoring phase variations through the DUT during RF pulsed signals.
5.5.5 Nonlinear measurements of active and passive devices

Under high-input signals to the DUT, the output signal will contain the fundamentals and
harmonics. Thus a nonlinear measurement technique needs to be defined to characterize
and test the DUT.
X-parameters are an example of this type of measurement using the Agilent PNA-
X [13].
5.5.6 Multi-port and differential measurements

The latest devices include multiple ports and differential ports. Some devices are intrin-
sically multiport (e.g. couplers) and others are driven differentially. In signal integrity
applications (e.g. a backplane) there can be many ports where common measurement
needs are insertion loss and crosstalk.
5.5.7 Load-pull and harmonic load-pull

The VNA is used to determine the impedance required to deliver a specific output power.
The locus of the output impedance for different powers is drawn on a Smith Chart.
Harmonics of the input power are also used to locate the harmonic load pull impedances.
This information is used to optimize the output load impedance to cover a specific
frequency ranges over a specific power ranges.
More details are in Chapters 13 and 14.
5.5.8 Antenna measurements

A variety of antenna measurements are also coordinated with a VNA including antenna
patterns, antenna match, and radar cross section. These measurement systems often
include positioners, additional amplifiers (due to remote location of the antennas), and
distribution networks.
5.5.9 Materials measurements

The permittivity and permeability of dielectric materials can be measured by the VNA
using specific measurement techniques including resonator structures, transmission
(free-space or in media), and open reflection probes. This is critical since environmen-
tal conditions may be difficult, e.g. very high temperatures or very low temperatures.
Material shapes may differ, e.g. biological materials, hot liquids, or large flat surfaces.
5.6 Device types for VNA measurements
5.6.1 Passive devices such as cables, connectors, adaptors, attenuators, and filters
Insertion losses and input and output return losses are the main vector measurements
needed for passive devices. Calibration needs may be relatively simple except in cases
of very low insertion loss. Dynamic range needs can vary widely, becoming the highest
for low crosstalk measurements and filter stopband measurements, for example. Mea-
surement speed, especially for production environments, is essential for this application.
Special care needs to taken for millimeter wave measurements of passive devices.
5.6.2 Low power active devices such as low noise amplifiers, linear amplifiers, and
buffer amplifiers
Linear amplifiers can also be measured using a VNA. Reverse isolation may be needed
in this case, especially for buffer amplifiers. For high-gain amplifiers a dynamic range of
more than 100 dB may be needed. A trade-off between speed, accuracy, and the number
of points needs to be decided before doing measurements.
Compression can also be measured at 1, 2, or 3 dB, and even the DC current variation
may be of interest. Noise figure, intermodulation, distortion, and return loss may also be
of interest.
5.6.3 High power active devices such as base station amplifiers and narrow-band
amplifiers
High power active devices can be considered as two parts. The linear part is already
covered in Section 5.6.2. The nonlinear part is covered in Chapter 12. Attenuators need to
be used to protect the VNA from being damaged. Many of these applications are relatively
narrowband (10%) and have power higher than watt-scale. The thermal management of
devices and device measurement is critical. Pulsing the DC bias can be used to reduce
the thermal effects. Within the pulsed bias, the RF measurement can be performed using
pulsed RF VNA. This pulsed measurement is covered in Chapter 15.
5.6.4 Frequency translation devices such as mixers, multipliers, up/down-converters

and dividers
Newer VNAs are capable of measuring frequency translation devices which have dif-
ferent input and output frequencies. Typical devices are: 1) multipliers where the output
frequency is a multiple of the input frequency. The step recovery diode is capable of
mixer
fin fIF = fin –nfLO
fLO
Fig. 5.22 Down converter block diagram.
delivering a comb generator (n = 1 to 100) for the input frequency. Common diode-
based multipliers are used for lower orders and high-speed digital circuitry is also often
used for harmonic generation. 2) Dividers devices where the input frequency is a multiple
of the output frequency. 3) Mixer devices are more general where the output frequency
is related to the input frequency by the local oscillator frequency as: fIF = fin + / − fLO
where the fI F is the intermediate frequency, which is often lower than the input fre-
quency, as shown in Figure 5.22. 4) Harmonic mixers are the same as mixers except
that the fLO is a multiple of a lower frequency LO. 5) Up/down-converters are the most
general frequency translation devices. They include amplifiers, filters, mixers, or multi-
pliers. Some applications require that the up/down-converter is in the same package and
a high isolation switch is used to choose the operation mode. Receivers for consumer
electronics often have this or higher levels of integration.
5.6.5 On-wafer measurements of the above devices

On-wafer measurements represent another type of challenge. The probe interface from
devices on-wafer to the VNA is an important and challenging aspect of this measurement
class as there is additional loss between the couplers and the DUT and the probe contact
repeatability can affect uncertainties. A variety of different calibration and de-embedding
techniques have been developed for on-wafer measurements that are suitable for the stan-
dards involved and are discussed in Chapter 8. For high power on-wafer measurements,
a pulsed bias and pulsed RF can be used, as explained in Chapter 15.
At the present time VNAs are used for many applications with the following features:
• High frequency up to 1100 GHz or higher
• Low frequency down to 70 KHz or lower
• New calibration techniques including electronic calibration
• High-speed measurements down to 2.5 microseconds per point
• Multiple ports up to 16 ports or more
• Differential and balanced measurements
• Time domain for signal integrity analysis
• Pulsed RF measurements
• Error-corrected power measurements
• Error-corrected mixer measurements for conversion loss and phase
• Error-corrected noise measurements.
Input
fin
Pin Switch matrix box
φin
Output 1 Output 2 ………… Output n
Fig. 5.23 Switch matrix box.
5.7 Improving VNA measurement range
5.7.1 Using a switch matrix box

The basic switch matrix box is used to switch microwave signals between the output of
a 2-port VNA and input to one or more input devices as shown in Figure 5.23. Another
switch matrix box (or another set of switches in the same box) is used to connect from
the output of multiple devices to the input of the 2-port VNA.
Mechanical switches can be used for reasons of linearity and lower insertion losses.
This has the disadvantages of slower measurement time, repeatability, and, potentially,
switch lifetime (which may be in the tens of millions of cycles). Solid-state switch
matrices (using technologies discussed earlier in this chapter) are also used that have
excellent repeatability, switching time, and lifetime, but have higher insertion losses and,
potentially, worse linearity.
Passive-DUT switch boxes are typically used to test multiple or single devices.
However, active-DUT switch boxes are also used to configure the input to devices
for specific power, frequency range, noise or multiple signals for intermodulation
distortion.
Depending on the software implementation, the switch box may be de-embedded
using measured S-parameters or, more commonly, multiport calibrations are applied at
the test ports of the switch box.
VNAs used now include many of the switch structures to accomplish the same purpose
previously done by the switch matrix box.
5.7.2 Using multiple sources

Testing mixers requires two sources – one for the input signal and the other for the
local oscillator. Also testing for intermodulation distortion requires two sources with a
frequency offset.
Some VNAs now used include more than one source and a high isolation combiner.
The frequency ranges of these sources are important, and there is the need to control
each source separately.
The phase noise of these multiple sources can be important for intermodulation
measurements where the offset frequency is low.
Measurement
6
Receiver 1
Bias tee Test coupler TP1
Reference coupler 3
Reference
1
Receiver 1
Fig. 5.24 Reverse coupler.
5.7.3 Using reversing couplers

Incident and reflected signals are usually measured using 10–20 dB directional cou-
plers. Thus, reversing the coupler reduces the noise floor by the coupling factor of
approximately 10–20 dB. However, this comes at the expense of reducing the maximum
input power by 10–20 dB.
Using the switch box to do the signal switching in portions of the measurement routine
is typically used for testing low noise devices or systems.
Using reversing couplers does not change the VNA’s dynamic range. However, it
offsets the dynamic range by 20 dB, as shown in Figure 5.24. Instead of connecting
point-1 to point-2, point-3 to point-4, and point-5 to point-6, the reverse coupler technique
connects point-1 to point-2, point-4 to point-6, and point-3 to point-5.
5.7.4 Using an external amplifier/attenuator

To characterize a power amplifier for full saturation, higher power than is available from
the VNA is needed. Thus, an external amplifier can be connected between point-3 and
point-4 as shown in Figure 5.25.
To characterize an amplifier with higher gain which produces higher power than is
specified by the VNA input, an external attenuator is needed. It can be connected between
point-5 and point-6 as shown in Figure 5.24.
5.7.5 All-in-one VNA box

Adding more couplers, attenuators, receivers, and bias-tees can expand the VNA to a
four (or more) port analyzer.
Measurement
6
Receiver 2
Bias tee Test coupler TP2
Reference coupler 3
Reference
1
Receiver 2
Fig. 5.25 High power amplifier.
Adding more than one source with switches and high isolation combiners can expand
the VNA to systems for testing mixers and intermodulation distortion.
Users can then reconfigure the “all-in-one” VNA box into a custom system to meet
the test requirements for specific devices or specific applications.
Recently, up to eight sources were introduced by Rohde and Schwarz for the 8 GHz
box (ZVT8).
5.8 Practical tips for using VNAs
5.8.1 User training

The user’s experience with VNAs can impact measurement results. Thus a basic under-
standing of operatingVNAs and calibration are essential to obtain accurate and repeatable
data. The cable type, connector type, operating frequency, and power will influence the
levels of user training. When waveguide is used at millimeter wave frequencies, then
user training is different from that at the microwave frequencies.
5.8.2 Connector care

Connectors need to be regularly inspected to assure that the connector mating is cor-
rect, clean, and precise. At millimeter wave frequencies this is more critical than at RF
frequencies. At higher microwave power, connectors need to be checked to see if they
can handle the power going through them without degradation. The specific application
and operating frequency determine, what type of connector is needed.
5.8.3 Temperature environment and stability

The stability of the temperature of the environment where the VNA is located impacts
the stability and repeatability of measurements. This is especially true when comparing
devices with mdB variations in their S-parameter values. Therefore, a stable environment
is required to ensure accurate measurements. The VNA data is specified for operating in
temperatures in a limited range and any deviation from the calibration temperature can
introduce other changes in data.
5.8.4 Measurement locations: production, development or research

Production environments require accurate calibration and rapid methods of taking mea-
surements. VNA calibration should be performed on a regular basis to ensure high
yields.
Development environments demand a highly reconfigurable VNA system. Measure-
ments need to be performed to prove the development concept. The objective is to increase
yield and decrease test time. Automated test systems may be required to measure a large
number of devices.
Research environments require much wider frequency range systems and are associ-
ated with very different calibration techniques.
5.9 Calibration and calibration kits
Due to imperfections of the VNA and of any networks between the VNA and the DUT,
calibrations need to be performed on the VNA to calibrate the measurement system to
the DUT reference planes.
Chapter 8 presents different methods of calibration and calibration kits.
5.10 Conclusions
In this chapter, the history and background behind the modern VNA have been discussed
as have been many of the fundamental building blocks of that class of instrument. The
objective has been to show how the attributes of those building blocks affect measurement
performance and how various architectures can optimize or enable certain applications.
The core technologies and structures presented here are themselves building blocks for
instruments and measurement classes presented in later chapters.
References
[1] D. Rytting, “ARFTG 50 year network analyzer history,” 67 th ARFTG Conf. Dig., pp. 1–8,
June 2006.
[2] J. A. C. Kinnear, M.A., A.M.I.E.E. “An automatic swept frequency meter,” British
Communications & Electronics, p. 359, May 1958.
[3] “An advanced new network analyzer for sweep-measuring amplitude and phase from 0.1 to
12.4 GHz,” HP Journal, Feb. 1967.
[4] R. A. Hackborn, “An automatic network analyzer system,” Microwave Journal, Vol. 11,
pp. 45–52, 1968.
[5] The essentials of vector network analysis: from α to Z 0 , Anritsu Company, 2008.
[6] J. A. Crawford, Frequency Synthesizer Design Handbook, Artech House, 1994.
[7] Microsemi, The PIN Diode Circuit Designer’s Handbook, 1992.
[8] R. S. Pengelly, Microwave Field Effect Transistors – Theory, Design and Applications,
Research Studies Press, 1986.
[9] G. D. Vendelin, A. M. Pavio, U. L. Rohde, Microwave Circuit Design Using Linear and
Nonlinear Techniques, Wiley, 2005, chp. 12.
[10] S. A. Maas, Microwave Mixers, Artech House, 1993.
[11] M. Kahrs, “50 years of RF and microwave sampling,” IEEE Trans. Microw. Theory and Tech.,
vol. 51, pp. 1787–1805, June 2003.
[12] J. Martens, “Multiband mm-wave transceiver analysis and modeling,” 2012 WAMICON
Dig., Apr. 2012.
[13] Company web sites:
• Rohde & Schwarz, http://www2.rohde-schwarz.com/en

• Anritsu Corporation, http://www.anritsu.com
• Agilent Technologies, http://www.agilent.com
• Maury Microwave, http://www.maurymw.com
• Focus Microwaves, http://www.focus-microwaves.com
• NMDG, http://www.nmdg.be
6 Microwave power measurements
Ronald Ginley
6.1 Introduction
In physics, power is the rate at which energy is transferred, used, or transformed. For
example, the rate at which a light bulb transforms electrical energy into heat and light is
measured in watts – the more wattage, the more power, or equivalently the more electrical
energy is used per unit time [1]. Energy transfer can be used to do work, so power is also
the rate at which this work is performed [2].
For systems or circuits that operate at microwave frequencies, the output power is
usually the critical factor in the design and performance of that circuit or system. Mea-
surement of the power (signal level) is critical in understanding everything from the basic
circuit element up to the overall system performance. The large number of signal mea-
surements that can be made and their importance to system performance means that the
power-measurement equipment and techniques must be accurate, repeatable, traceable,
and convenient.
In a system, each component in a signal chain must receive the proper signal level
from the previous component and pass the proper signal level on to the succeeding
component. If the output signal level becomes too low, the signal becomes obscured in
noise. If the signal level becomes too high, though, the performance becomes nonlinear
and distortion can result. The uncertainties associated with the measurement of power
also play a very important role in the development and application of microwave circuits.
For example, a 10 W transmitter costs more than a 5 W transmitter. Twice the power
output means twice the geographical area is covered or 40% more radial range for a
communication system. Yet, if the overall measurement uncertainty of the final product
test is of the order of ±0.5 dB, the unit actually shipped could have output power as
much as 10% higher or lower than the customer expects, with resulting lower operating
margins [3].
At low frequencies, the concepts of voltage, impedance, and current can be used to
describe how energy is transported through a circuit. At microwave frequencies, voltage
and current lose significance and are replaced by “power.” The question of how much
signal is present is answered by a power measurement. The importance of the mea-
surement of microwave power in microwave circuits is easily seen, as it is the power
that “does the work” or in the case of a communication system, it is the power that
carries the information [4]. It is at the higher operating power levels that each decibel
increase in power level becomes more costly in terms of complexity of design, expense
Microwave power measurements 131
of active devices, skill in manufacture, difficulty of testing, and degree of reliability.

“The increased cost per dB of power level is especially true at microwave and higher
frequencies, where the high power solid state devices are inherently more costly and the
guard-bands designed into the circuits to avoid maximum device stress are also quite
costly. Many systems are continuously monitored for output power during ordinary
operation” [3].
This chapter covers the fundamentals of making microwave power measurements;
beginning with basic concepts, definitions, and terminology. Next, there is a brief dis-
cussion on different types of power measurements and how they are applied. To make
power measurements, there must be some form of a power detector. The main types
of power detectors are covered. The power detectors that are used are not perfect and
the results of their measurements must be corrected for different error mechanisms. The
first of these, effective efficiency, is discussed in terms of its definition and how it is
measured at the highest level of accuracy. The concept of a basic power measurement is
explored and the effect of effective efficiency and different error mechanisms and other
adaptations are discussed. This leads into a brief discussion of the uncertainty of basic
power measurements and a look at the larger contributors to the overall uncertainty of a
power measurement. Finally, a few examples of different aspects of power measurements
are given.
6.1.1 Why power and not voltage and current?

The first question many ask with regards to measuring microwave power is: why not just
use measurements of voltage and current? From a DC and low-frequency perspective,
power is defined in terms of the voltage (V), current (I), and resistance (R) as:
V2
P = IV = = I 2 R. (6.1)
R
Voltage and current measurements are straightforward and easy to make. However,
as the frequency approaches 1 GHz, it becomes necessary to directly measure power
because voltage and current measurements become impractical. One of the main reasons
for this is that voltage and current can vary with position along a lossless transmission
line, whereas power maintains a constant value with position. Another example of the
decreased usefulness is in waveguide transmission structures, where voltage and current
are even more difficult to define due to the structure of the electric and magnetic fields
inside the guiding structure. For these reasons, at radio and microwave frequencies,
power is more easily measured, easier to understand, and more useful than voltage or
current as a fundamental quantity.
6.2 Power basics, definitions, and terminology
Just what do we mean when we talk about microwave power and the measurement of
microwave power? First, the unit of power is the Watt. The International System of
132 Ronald Ginley
b a
Termination
Fig. 6.1 Simple termination to help define power.
Units has established one Watt to be one joule per second. Note that there are no other
electrical units used in this definition. We can talk about power in terms of the complex
electromagnetic wave amplitudes that are travelling along or incident on a microwave
structure. Take, for example, the simple termination shown in Figure 6.1, where “a” is
the electromagnetic wave incident on the termination and “b” is the wave that is reflected
from the termination. The reflection coefficient is defined as: S11 = ab . Different “powers”
are defined in terms of the electromagnetic waves as:
incident power: Pinc = a 2 ,

reflected power: Pref = b2 ,
and net power into the termination: Pnet = |a|2 − |b|2 .
It is difficult to directly measure the complex waveforms “a” and “b”. Instead, the
measurement of microwave power is performed by transforming the waves into some-
thing more easily measured such as a temperature change or rectified energy. There are
efforts underway to directly measure the electric or magnetic fields; these techniques
will, hopefully, allow us to make much more accurate power measurements [5].
6.2.1 Basic definitions

To be able to understand microwave power measurements, there are a few useful basic
definitions:
dB (decibel): the ratio of two powers is often used instead of absolute power. The ratio
is dimensionless
and is commonly
expressed as decibels. The dB is defined as:
dB = 10log 10 Power Level 2 where Power Levels 1 and 2 are arbitrary power levels
Power Level1
dBm: another method of expressing a power level is to reference it toa known level. In
the case of dBm, the reference level is 1 mW. Thus, dBm = 10 log 10 Power Level1
1 mW .
dBW: power expressed in dB with a reference level of 1 W.
A comparison of the different terms is given in Table 6.1.
6.2.2 Different types of power measurements

When making a power measurement, what “type” of power measurement must be defined
to avoid confusion and incompatible results. There are many different ways to define
“power” when looking at sinusoidal or other complex, periodic waveforms. The most
Table 6.1 Comparison of different power definitions
dBm dBW Watts
+60 +30 1,000 (1 kilowatt)

+50 +20
+40 +10
+30 0 1 (1 Watt)
+20 −10
+10 −20
0 −30 1 milliwatt
−10 −40
−20 −50
−30 −60 1 × 10−6 (1 microwatt)
DC
Component
P
i
e
Amplitude
e R i
Time
(a) (b)
Fig. 6.2 The product P of voltage e and current i varies during the sinusoidal cycle (figure courtesy of
Agilent Technologies).
common are average power, pulse power, and peak or peak envelope power. Modern
wireless system designs use different complex schemes for combining many channels
into broadband complex signal formats. A typical signal, like the EDGE system, requires
peak, average, and peak-to-average characterization of power signals.
The term “average power” is very popular and is used in specifying almost all RF
and microwave systems. In elementary theory, power is said to be the product of voltage
and current. But for an AC voltage cycle, this product V × I varies during the cycle,
as shown by curve P in Figure 6.2, according to a 2 × frequency relationship. Using
this example, a sinusoidal generator produces a sinusoidal current as expected, but the
product of voltage and current has a DC term as well as a component at twice the generator
frequency. The word “power” as most commonly used, refers to that DC component of
the power product. All the methods of measuring power to be discussed in this chapter
use power sensors which, by averaging, respond to the DC component.
The definition of power is energy per unit time. The important question to resolve is
over what time is the energy transfer rate to be averaged when measuring or computing
power? From Figure 6.2, we clearly see that if too narrow a time interval is used (say
close to one cycle) varying answers for energy transfer rate are found. But at microwave
134 Ronald Ginley
P
1
Tr =
fr
Ppulse
Pavg
Fig. 6.3 Pulse power Ppulse is averaged over the pulse width (figure courtesy of Agilent Technologies).
frequencies, such microscopic views of the voltage-current product are not common. For
this discussion, power is defined as the energy transfer per unit time averaged over many
periods of the carrier frequency involved.
In a more mathematical sense, average power can be written as [3]:
nT
1
Pavg = e(t) i(t) dt, (6.2)
nT 0
where T is the period of the lowest frequency component of e(t) and i(t) (e(t) and i(t)
are defined in Figure 6.2). The averaging time for average power sensors and meters
is typically from several hundredths of a second to several seconds and, therefore, this
process obtains the average of most common forms of amplitude modulation [3].
For pulse power, the energy transfer rate is averaged over the pulse width τ (Figure 6.3).
Pulse width τ is generally considered to be the time between the 50% rise-time/fall-time
amplitude points.
By its very definition, pulse power averages out any aberrations in the pulse envelope
such as overshoot or ringing. For this reason it is called pulse power and not peak power
or peak pulse power as is done in many radar references. The terms peak power and peak
pulse power are not used here for that reason. Peak power refers to the highest power point
of the pulse top, usually the risetime overshoot. For certain more sophisticated microwave
applications and because of the need for greater accuracy, the concept of pulse power is
not totally satisfactory. Difficulties arise when the pulse is intentionally non-rectangular
or when aberrations do not allow an accurate determination of pulse width; this is when
the peak power method can be used for more accurate measurements [3].
Measurements of modulated signals

Digital modulation provides more information capacity, compatibility with digital
data services, higher data security, better-quality communications, and quicker system
availability. Developers of communication systems face constraints such as available
bandwidth, permissible power, and the inherent noise level of the system. Over the past
few years, a major transition has occurred from simple analog amplitude modulation
and frequency/phase modulation to new digital modulation techniques. Another layer of
complexity in many new systems is multiplexing. Two principal types of multiplexing (or
“multiple access”) are TDMA and CDMA. These are two different ways to add diversity
to signals, allowing different signals to be separated from one another [6].
Although many RF and microwave measurements can be made with CW signals,
there are many other signal schemes that require sampling a signal at a certain point in
time, or applying non-CW excitation to a circuit under test. Pulses, on-off transitions,
power control steps, and some digital modulation schemes are not CW signals, and their
measurement requires more advanced techniques [7]. Depending on the application, the
accuracy of the power meter solution could have a significant impact on the overall
performance. For example, the output power transmitted at a cellular base station affects
the coverage area. When the base station is installed, the output power is measured and
verified. System designers try to optimize the coverage area while balancing trade-offs.
More output power leads to a greater coverage area but it can also create interference. If
the power output is below a minimum limit, the coverage area is reduced and this could
eventually lead to dropped calls and dissatisfied customers.
There are many different ways of looking at a digitally modulated signal. To exam-
ine how transmitters turn on and off, a power-versus-time measurement is very useful.
In addition, peak and average power levels must be well understood, since asking for
excessive power from an amplifier can lead to compression or clipping. These effects
distort the modulated signal and usually lead to spectral regrowth as well. The power
within one or more cycles of the signal is of interest when developing or troubleshooting
mobile radio systems.
When looking at these complex signals, the most appropriate power sensor needs to
be selected. Conventional thermal power sensors such as bolometric or thermoelectric
detectors (see Section 6.3 for a discussion of these types of sensors) cannot adequately
measure complex signal characteristics since they cannot delimit specific areas of power
contribution in a timeslot. This is because thermal sensors average the RF power occur-
ring over the entire time period. Sampling the power envelope over time is feasible with
diode sensors. However, diode sensors always include signal details such as overshoots,
interference pulses, and glitches as well as signal edges of a pulsed RF signal in pro-
portion to their power [8]. Peak power measuring instruments and sensors have time
constants in the sub-microsecond region which allow for measurement of pulsed power
modulation envelopes [3].
Diode-based power sensors can be used to display power versus time in the same way
that an oscilloscope does. This means that you do not miss a single detail of the signal you
want to investigate. Furthermore, you can add time-slot and gate structures to your pulsed
RF signals and configure them in the manner desired (see Figure 6.4 for an example). By
graphically editing the gates added to the “scope” window, you can selectively suppress
unwanted components at the beginning and end, which occur, for example, in the transi-
tion between two timeslots. Wideband power sensors can quickly and accurately measure
peak, average, peak-to-average ratio power measurement, rise/fall time, pulse width, and
complementary cumulative distribution function statistical data for wideband signals.
136 Ronald Ginley
Pulsed RF signal Illustrating gated measurements

10
Peak Peak
gate 1 gate 2
0
–10 Time-gated
mesurements
Power (dBm)
–20
–30
–40
–50
–60
0 1 2 3 4 5 6 7 8
Time (seconds)
Fig. 6.4 Time domain shot of a wireless signal format, in this case, an EDGE signal in a GSM system. It
is an ideal candidate for peak, average, and peak-to-average ratio measurements for time-gated
wireless formats (figure courtesy of Agilent Technologies).
The measurement of complex signals is indeed a complex topic. For further

information on the topic of modulated signal measurements, see [6, 7, 8, 9, 10].
For the purposes of the discussion in the rest of this chapter, we will focus on average
power measurements and CW signals.
6.3 Power detectors and instrumentation
As with power measurements, there are different types of power detectors. It is necessary
to have a basic understanding of how power detectors work in order to be able to choose
the most appropriate one for the measurement at hand. In addition to learning about the
detectors, it is important to understand how the electronic packages associated with the
detectors work (the “power meters”).
Due to the difficulty of measuring waveforms and power directly at higher frequencies,
the techniques used to measure power modify the microwave signal in some manner to
allow it to be measured more easily. The three main types of detectors are bolometric,
thermoelectric, and diode. Bolometric detectors work by substituting DC power for the
RF power; thermoelectric detectors work by substituting a thermally generated voltage
for the RF power; and the diode-type sensors work by rectifying the RF signal.
Each type of sensor has its strengths and weaknesses. The bolometric sensors are
typically very stable, linear, and have easily modeled behavior: however, they work only
in a narrow dynamic range and have limited power capabilities; they also react slowly.
Thermoelectric sensors are also linear, have better sensitivity, good dynamic and power
ranges; and for general use, they need a support set of electronics and require a reference
point at a known frequency to fix their operation. Diodes are very nonlinear and fast, and
the newer generation of diode-based detectors have good dynamic and power ranges;
they also need a reference point to fix their operation.
6.3.1 Bolometric detectors

Bolometric detectors use a temperature-sensitive resistor to measure the microwave
power. The most common form of bolometric detector is the thermistor detector. Simply
put, when microwave power is applied to a thermistor sensor, the resistive element heats
up and as a result changes its resistance. By measuring the change in resistance, you
can determine the amount of microwave power that was applied. Thus, the microwave
power level of the signal being measured is ultimately determined by a DC resistance
measurement.
Of course, it isn’t really that simple. There are other steps in the process that must
be considered. Figure 6.5 shows the basic structure of a generalized bolometric detector
and power meter. The DC blocking capacitor, if used, rejects any DC signal coming into
the detector from the microwave connector. This is important, as the sensing element
reacts to any signal and a DC signal would give a false microwave power level. The
sensing element is some form of temperature-sensing resistor. There are many resources
that discuss the different types of sensing elements; see [11] for a good analysis of
the different types. For this discussion, we consider a thermistor bead sensing element.
Thermistor sensors are small beads of metallic oxides with two very small wire leads.
The common type of bead is a negative-temperature-coefficient bead which refers to the
effect that as the temperature of the bead goes up the resistance of the bead goes down.
The resistance of the thermistor is monitored by a power meter circuit. To make sure
that there is minimal leakage of the microwave signal beyond the bead, there are filters
designed to block any microwave signal, in the operating range of the detector, from
leaking out of the DC leads to the power meter circuit.
The power meter circuit in its most basic form is simply a Wheatstone bridge
(Figure 6.6). In more complicated schemes, it is a self-balancing bridge circuit with
a digital back end. An example of a self-balancing, bridge circuit power meter is the
DC Sensing
blocking element Conditioning
capacitor electronics
RF
input Power meter
Fig. 6.5 Basic bolometric power sensor design.

138 Ronald Ginley
DC Bias
Voltmeter
Reference
resistance Ro
PRF Power meter
Thermistor
Bolometer
Fig. 6.6 Functional diagram of thermistor and power meter pair.
NIST Type IV power meter [12]. The interaction of the thermistor and the power meter
is straightforward. Initially, with no microwave power applied, the power meter supplies
a DC current to the thermistor bead. This current heats up the bead and brings its resis-
tance to a point where it is in balance with the internal reference resistor inside the power
meter (the balancing arm of the bridge circuit). Once the DC equilibrium is established,
the microwave signal is applied to the thermistor. The microwave signal heats up the
thermistor bead and, thus, drives the resistance of the bead down. This causes the bridge
to go out of balance and it responds by removing enough DC bias from the bead for it to
again be balanced against the reference resistor in the power meter. By measuring how
much DC power was removed, the amount of microwave power applied is determined.
This is termed the DC substituted power. The power for this type of power meter can be
calculated as:
2
VDC−off − VDC−on
2
Psub = , (6.3)
R0
where VDC−off is the DC bias voltage with no microwave power, VDC−on is the DC bias
voltage with microwave power applied and R0 is the power-meter reference resistance
value. Thermistor detectors have a workable dynamic range of −10 to +10 dBm.
6.3.2 Thermoelectric detectors

Thermoelectric detectors use a different method for detecting a microwave signal. How-
ever, like the bolometric detector, temperature is the medium of the method. The basic
circuit for a thermoelectric detector is shown in Figure 6.7(a). For this detector, the
microwave signal is applied to a terminating resistor. This is a fixed value resistor. The
microwave signal causes the resistor to heat up. In close proximity to the resistor, on
a thermally isolated “island” is the hot junction of a thermocouple. Thermocouples are
based on the fact that dissimilar metals generate a voltage due to temperature differences
at a hot and a cold junction of the two metals, as seen in Figure 6.7(b).
Thermal isolation Thermocouple

Hot Cold
Hot junction junction junction
Cold junction Conductor a
RF in T
Conductor b Conductor b
Termination
(a) (b)
Fig. 6.7 (a) Basic thermocouple detector design and (b) basic thermocouple design (figure courtesy of
Agilent Technologies).
When the terminating resistor heats up, it transfers heat energy to the thermocou-
ple, which generates a voltage that can be used to determine the amount of microwave
power applied. Since thermocouples and thermistors are heat-based detectors, they are
true averaging detectors. In most thermoelectric detectors, the voltage generated by the
thermocouple is small and needs to be amplified and conditioned. These electronics are
usually housed in the same shell as the terminating resistor and the thermocouple. This
overall device is connected to a power meter that converts the thermocouple signal into
a power value (a detailed discussion of thermocouples and power meters can be found
in [9] and [11]). These power meters have a design that takes advantage of the increased
power sensitivity of the thermocouple sensors and is still able to deal with the very
low output signals of the detectors. Thermocouple detectors have a dynamic range of
approximately −35 to +20 dBm.
With thermocouples, where there is no direct power substitution, sensitivity differences
between sensors or drift in the sensitivity due to aging or temperature can result in
a different DC output voltage for the same RF power. Because there is no feedback to
correct for different sensitivities, measurements with thermocouple sensors are said to be
open-loop. Thermocouple power meters have solved this need for sensitivity calibration
by incorporating a 50 MHz, 1 mW power-reference oscillator whose output power is
controlled with great precision (as low as ±0.4%) [9]. To verify the accuracy of the
system, or adjust for a sensor of different sensitivity, the user connects the thermocouple
sensor to the power reference output of the power meter and, by the use of a calibration
adjustment, sets the meter to read 1.00 mW. By applying the 1 mW reference oscillator
to the sensor’s input port, just like when an unknown signal is to be measured, the same
capacitors, conductors, and amplifier chain are used for measurement in the same way as
for the reference calibration. This feature provides confidence in the power results that
the detector/power meter pair is producing.
6.3.3 Diode detectors

Diode detectors use a method for determining power that is very different from the
thermistor and thermocouple detectors. Diodes convert AC signals to DC by way of
140 Ronald Ginley
15
10
V
–30 –20 –10 10 20 30 mV
–5
–10
–15
Fig. 6.8 Basic response curve for a semiconductor diode.
their rectification properties. These arise from the nonlinear current-voltage (I-V) char-
acteristics of a semiconductor diode. Metal-semiconductor junctions, exemplified by
point-contact technology, exhibit a low potential barrier across their junction, with a
forward voltage of about 0.3 V. They have superior RF and microwave performance,
and were popular in earlier decades. Low-barrier Schottky diodes, which are metal-
semiconductor junctions, succeeded point-contacts and vastly improved the repeatability
and reliability [9]. Figure 6.8 shows a typical diode I-V characteristic of a low-barrier
Schottky junction, expanded around the zero-axis to show the square-law region.
Mathematically, a detecting diode obeys the diode equation

i = IS eαV − 1 , (6.4)
where α = q/nKT , i is the diode current, V is the net voltage across the diode, and IS is
the saturation current and is constant at a given temperature. K is Boltzmann’s constant,
T is absolute temperature, q is the charge of an electron, and n is a correction constant to
fit experimental data (n equals approximately 1.1 for the devices used here for sensing
power). The value of α is typically a little under 40 v −1 [9].
For a typical diode, the square-law detection region exists for power levels Pin from
the noise level up to approximately −20 dBm. There is a transition region that ranges
from approximately −20 to 0 dBm input power, and there is a linear detection region
that extends above approximately 0 dBm. For wide-dynamic-range power sensors, it
is crucial to have a well-characterized expression of the transition and linear detection
ranges. If you are operating in these regions, then it is necessary to apply some form of
correction to the diode curve that is not necessary in the square law region.
The simplified circuit of Figure 6.9(a) represents an unbiased diode device for detecting
low-level RF signals. The matching resistor (approximately 50 ohms) is the termination
Linear
region
Square law
region of diode
sensor
Rs Vo
Vs
Rmatching Cb Vo
Noise
Floor
Pin
–70 dBm –20 dBm
(a) (b)
Fig. 6.9 (a) Diagram of a source and a diode detector with matching resistor and (b) power versus voltage
curve for diode sensor (figures courtesy of Agilent Technologies).
RF input
Low power path
High power path
Fig. 6.10 Example of new diode detector schemes (figure courtesy of Agilent Technologies).
for the RF signal. RF voltage is converted to a DC voltage at the diode, and the bypass
capacitor is used as a low-pass filter to remove any RF signal that leaks through the diode.
This example would work for signals in the square-law region (Figure 6.9(b)) of the
diode and would deviate appreciably from that as the power approaches approximately
−20 dBm [9].
Digital signal-processing and microwave semiconductor technology have now
advanced to the point where dramatically improved performance and capabilities are
available for diode power sensing and metering. New diode power sensors are now capa-
ble of measuring over a dynamic power range of −70 to +20 dBm (as a reference,
thermocouple sensors have a range of approximately −35 to +20 dBm and thermistor
sensors have a range of −10 to +10 dBm). This broad range permits the new sensors to
be used for CW applications that previously required two separate sensors. A simplified
diagram of one of these sensors is shown in Figure 6.10. Here, different diode chains are
used for different power levels.
142 Ronald Ginley
In detecting low power levels of about 100 pW, the diode detector output is about 50 nV.
This low-signal level requires a power meter with sophisticated amplifier and chopper
circuit design to prevent leakage signals and thermocouple effects from dominating
the desired signal. Earlier diode power sensors required additional size and weight to
achieve controlled thermal isolation of the diode. A dual-diode configuration balances
many of the temperature effects of those highly sensitive circuits and achieves superior
drift characteristics [9].
6.3.4 Power meters

The power meter plays an important role in the measurement chain of acquiring a signal,
reacting to the signal, and finally producing an output that is understandable. The power
meter’s function is the last part: producing the results of a measurement in a form that we
understand, meaning, in terms of power, a number in dB, dBm, Watts, or something that
can be used to signify power. For the bolometric-type sensors, the power meter is the unit
that maintains the resistive balance of the thermistor element and uses that process for the
calculation of power. For diodes and thermocouples, the power meter takes the voltage
from the sensing element and the calibration from the power reference port and calculates
power. Note that there are sensors that connect directly to a computer through a USB
connector. For these devices, the sensing element, conditioning electronics, and power
meter are all built into one housing, and a reference port calibration is not necessary, as
the sensor response data is calibrated at the factory and stored in the unit.
6.3.5 Power measurements and frequency ranges

The detector-type concepts described so far can be applied to almost all frequency
ranges. Practical application is another matter. At low frequencies, 100 kHz to 1 MHz,
bolometric-type detectors have issues with the RF energy leaking through the detector
and influencing the power meter and other associated electronics. Thermoelectric-style
detectors do not have this problem, as the RF signal terminates in a resistor with no
electrical connection to the power meter side. The diode detector circuit also limits
the leakage signal. At very high frequencies, above 110 GHz, the waveguide structures
become so small that the use of classical thermistor beads and terminating resistors
becomes impractical. There are several different techniques that are being used in the
110 GHz to approximately 1 THz range for power measurements. The first technique
uses a photo-acoustic method to measure power. “A thin film absorbs incoming THz
photons, creating heat. The heat causes expansion of a closed air-cell which is then
measured acoustically” [17]. Another technique uses a dry-load, WR-10 (75–110 GHZ)
calorimeter operating in an over-moded condition (for more information on dry-load
calorimeters, see Section 6.4.2). The detector can be used to measure power in higher-
frequency waveguide bands through the use of waveguide tapers or simply attaching the
WR-10 port directly to the other waveguide port (if they are compatible) [18]. Finally,
there are waveguide-based diode detectors that allow banded power measurements up to
1.7 THz.
While detectors used in the 100 kHz to 50 MHz range are traceable to the voltage and
impedance technique, and detectors used in the 50 MHz to 110 GHz range are trace-
able to calorimetric methods (see Section 6.4), above 110 GHz there are no established
methods for tracing power measurements to the SI. There have been efforts to extend
the calorimetric methods to higher frequencies. Also, the photo-acoustic and dry-load
calorimetric solutions described above try to establish calibration capability through
DC-heating substitution-type techniques. These techniques remain unverified, and the
problem of establishing traceability in the sub-millimeter ranges still remains.
6.3.6 Power levels and detectors

We have seen that bolometric-style detectors have a limited dynamic range, approxi-
mately −10 to +10 dBm, and the modern thermoelectric and diode sensors can have a
range of about −70 to +20 dBm. What can you use if you need to measure outside of these
levels? For higher power levels there are various forms of microwave wattmeters. These
devices measure the higher output directly and generally use rectification, or applying
the power to a load and measuring the load’s temperature change. They can be either
terminating or feed-through devices. Another technique cascades microwave couplers
together and after calibration allows standard detectors to measure the coupler-scaled
power signal [19]. Due to heating effects caused by the higher levels of power and other
error mechanisms, measurements with these detectors tend to have higher uncertainties
than those with standard power detectors.
For power levels below −70 dBm, heterodyne detectors/receivers are commonly used.
For these devices, the high-frequency signal is mixed with a lower-frequency signal. The
resultant signal accurately reproduces the amplitude and phase of the original signal, but
at a lower frequency, and simpler low-frequency techniques can be applied to determine
the power-signal values providing a larger dynamic range [20].
6.4 Primary power standards
No power detector is perfect. That is to say that power detectors do not indicate in their
electronics the exact amount of microwave power being applied to the units. To have an
accurate representation of the power incident on the detector, it is necessary to calibrate
the detector. While power detectors give readings in terms of the absolute power that is
applied to the unit, when we refer to “calibrating” the detector, we do not calibrate it in
terms of absolute power; instead, we use the concept of effective efficiency. Figure 6.11
shows a generalized power detector and power meter. Microwave power is applied to
the detector. This is represented as Pinc , the power incident on the detector. Not all of
the incident power reaches the sensing element of the detector. Power is lost through
absorption, imperfect conductors, reflections, and other loss mechanisms. Pnet is the
power that is dissipated in the sensing element. Note that while Pnet takes into account
the loss mechanisms of the input section, it does not account for the microwave energy
that leaks out of the detector through the DC connections. Finally, after signal processing
144 Ronald Ginley
Detector Conditioning
element electronics
RF input
Power meter
Detector
Pinc Pnet Psub
Fig. 6.11 Generalized power detector showing different powers levels that can be determined.
occurs in the power meter electronics, the power meter shows a resultant power level.
This is Psub , which is the substituted power (the terms Pinc , Pnet , and Psub were originally
defined in terms of bolometric power sensors, but they can be applied to the broader scope
of diode and thermoelectric detectors as well).
In an ideal detector, the three powers Pinc , Pnet , and Psub would all be equal. In the
real world they are all different. There are two terms that are used to describe the fact that
all of the incident power does not reach the sensing element and is not indicated in the
substituted power determined by the power meter. The first term is called the effective
efficiency (referred to as η or ηe ). Effective efficiency is defined as:
Psub
ηe = . (6.5)
Pnet
Thus, effective efficiency ηe for a detector is the ratio of the power determined by the
electronics of the power meter Psub to the power dissipated in the sensing element Pnet .
The other term is called the “calibration factor” (also called cal factor or Cf ). This is
defined as:
Psub
Cf = = ηe 1 − ||2 . (6.6)
Pinc
The calibration factor is the ratio of the substituted power to the power incident on the
detector. This can also be expressed in terms of the effective efficiency and the reflection
coefficient of the detector () as seen in (6.6). Another way of looking at the cal factor is
that it relates the substituted power in the detector to the power incident on the detector
and takes into account the fact that there is a difference between the reflection coefficient
of the detector and the reflection coefficient of the port where the detector is connected.
The concept of measurement error due to differences in reflection coefficients is termed
the mismatch factor or mismatch correction and is covered in more detail later in this
chapter. In the definition above of cal factor, it is assumed that the port to which the
detector is connected is non-reflecting.
We see that to make an accurate power measurement, we need to know the effective
efficiency (or CF) of the detector as well as its reflection coefficient. So how is ηe for
a detector determined? In general, ηe is determined through a transfer system that is
ultimately traceable to a primary standard for microwave power. The most widely used
primary standard is the calorimeter. The calorimeter essentially works by measuring both
the bolometric and calorimetric powers (electrical and thermal, respectively) simultane-
ously and comparing the results. There are several different forms of calorimeters. The
“microcalorimeter” and the “dry load calorimeter” are described here, as they account
for a large majority of systems used. Good references for information about calorimeters
can be found in [11, 13, 14]. An alternative approach for low frequencies, below 100
MHz, using AC voltage and impedance techniques, is also described.
6.4.1 The microcalorimeter

The microcalorimeter (Figure 6.12) measures the temperature rise of the bolometer
detector connected to it under different conditions of applied microwave power and DC
bias. In the coaxial microcalorimeter, the temperature increase of the detector is measured
with a thermopile. The thermopile is connected between the detector being measured and
a thermal reference; this is usually a mass with known thermal properties and has roughly
Type N connector
dc Bias lead
Thermopile
output lead
Hanger
Cable conduit
RF input semi-
rigid coax Cover
Water level
Base plate Thermopile

assembly
(a) (b)
Fig. 6.12 The microcalorimeter. (a) overall descriptive diagram and (b) photograph of Type-N
microcalorimeter and a primary transfer standard.
146 Ronald Ginley
the same thermal mass as the detector. During the measurement, the microcalorimeter
is kept in a thermally stable environment to minimize the effect of external temperature
changes. A commonly used technique is to immerse the calorimeter in a water-tight
housing, in a stable temperature-controlled water bath. To determine the ηe of a detector,
measurements are made at each frequency of interest, of the power meter and thermopile
output voltages (V1 and e1 ) with only DC applied to the detector, and then again (V2
and e2 ) with both RF and DC applied. The effective efficiency ηe is calculated at each
frequency using (6.7) [13]
2
V2
1−
V1
ηe = g 2 . (6.7)
e2 V2
−
e1 V1
The term g is a frequency-dependent correction factor for the microcalorimeter-
bolometer detector combination, which is also known as the calorimetric equivalence
correction. The uncertainty in the measurement is determined primarily by the uncer-
tainty in g. The determination of g is a major effort that is described in [14]. A real
advantage of the microcalorimeter is that the power reflected by the detector being mea-
sured and the power lost in the transmission lines leading to the calorimeter have a
minimal effect on the results.
6.4.2 The dry load calorimeter

The dry load calorimeter is similar to the microcalorimeter except that the thermal ref-
erence is replaced by a “dummy” detector. The essential design concept of the dry load
calorimeter is shown in Figure 6.13. Dry load calorimeters use a dual input to identical
loads where the DC power biasing one load is used to balance the RF power absorbed
in the other load. A thermopile is used to detect any temperature difference between the
two loads.
Calorimeter
Active load
Thermopile
RF input
Digital
voltmeter
Reference
load
Amplifier
(a) (b)
Fig. 6.13 Dry load calorimeters. (a) conceptual diagram and (b) photograph of a system that can be used as
a dry load calorimeter with two power detectors attached.
The two loads should see the same external environmental variations and a these
should cancel out. Thus, there is no need for as extreme environmental controls as
with the microcalorimeter. In microcalorimeters, the effective efficiency of an inserted
bolometer mount, which functions as the calorimeter load, is measured. After calibration,
this bolometer mount is used as a secondary standard for power measurements. In the case
of dry load calorimeters, the calorimeter itself functions as the calorimetric load, and its
effective efficiency is determined by measurements and theoretical analysis. Secondary
power standards are calibrated using a stable RF generator system by comparing their
response with that of the dry load calorimeter.
6.4.3 Voltage and impedance technique

Calorimeter techniques are generally used above 50 MHz. For calibrating power detectors
below 50 MHz, the voltage and impedance technique can be used.
If a thermistor detector with an associated power meter and a thermal voltage converter
is connected to two sides of a tee, and DC and RF power is supplied through a switch to
the third side of the tee (Figure 6.14), the effective efficiency of the thermistor mount can
be determined from the electrical parameters of the tee, the thermal voltage converter
standard, and the reflection coefficient of the thermistor mount [15, 16]. From before,
the effective efficiency of a thermistor mount is defined as
Psub
ηe = , (6.8)
PRF
Power
meter
Thermistor
DC detector
source
Characterized
Switch
tee
RF
source Thermal voltage
convertor
Voltmeter
Fig. 6.14 Equipment setup for determining the effective efficiency of a thermistor detector with a thermal
voltage convertor.
148 Ronald Ginley
where Psub is the DC substituted power determined from the electronics of the detector
power meter and PRF is the net RF power delivered to the thermistor mount. For the
voltage and impedance technique, the RF power delivered to the mount is determined
from the following equation:

Mu 2 1
PRF = Vdc ∗ Cf ∗ ∗ Re ∗
. (6.9)
MS Zpar
Vdc is the average of the absolute values of the plus and minus DC voltage supplied
when the thermal voltage converter output is identical to the output when RF voltage is
supplied to the system.
Cf is the correction factor for the thermal voltage converter standard.
Mu and Ms are the mismatch factors for the two sides of the tee.
∗ is the complex conjugate of the impedance at the reference plane of the thermistor
Zpar
mount.
Zpar can be calculated from the reflection coefficient of the detector using
1 + Det
Zpar = Zo , (6.10)
1 − Det
where Zo is the characteristic impedance of the transmission medium (in most cases
50 ) and Det is the complex reflection coefficient of the thermistor detector.
The DC substituted power of the thermistor mount is determined from the power
meter as
2 −V2
Voff on
Psub = . (6.11)
Rth
Voff is the voltage output of the power meter measured when the RF power is turned off.
Von is the voltage output of the power meter measured when the RF power is turned on.
Rth is the DC resistance that the power meter establishes for the thermistor mount.
By combining (6.8)–(6.11) the effective efficiency of the thermistor detector can be

determined.
6.5 Basic power measurement techniques
Power measurements can be as simple or as complex as you want them to be. There
is a direct trade-off between accuracy and simplicity. When making microwave power
measurements, there are several factors that must be considered when looking at the
accuracy of the measurement. The simplest form of measuring power is to connect the
detector to the output port of whatever and see what power is there (Figure 6.15). This
assumes that you have a power detector that has the same connector type as the output
port type you are measuring. This situation can be described by:
Po = Psub , (6.12)
Signal
Power
to be
meter
measured
Output Power
port detector
Fig. 6.15 Basic power measurement setup.
where Po is the power at the output port. If you do this; however, you will have accuracy
limitations due to not correcting for the effective efficiency of the detector, not correct-
ing for the mismatch, and not considering the errors in the power meter reference and
electronics. There are other additional errors that must be considered. These can include:
(a) if the connector of your detector does not match the output port of what you are
measuring; (b) uncertainties in the evaluation of the effective efficiency of the power
detector; and (c) repeatability.
A more detailed power measurement can be described by:
Psub
Po = , (6.13)
ηDet Mgl
where Po is the available power at the port you are measuring that would be delivered
to a load, Psub is the substituted power determined from the power meter, Mgl is the
mismatch factor, and ηDet is the effective efficiency of the power detector that is being
used for the measurement.
The effective efficiency of the detector can be determined through the use of a
calorimetric process or through a transfer process such as a direct comparison power
measurement system [21, 22, 23]. The mismatch factor is due to the difference in the
reflection coefficient of the output port and the reflection coefficient of the detector.
6.5.1 Mismatch factor

A uniform section of a microwave transmission line is “matched” when it is terminated
in such a way that no net reflection of energy occurs. A device that causes a net reflec-
tion of energy when connected to this transmission line is termed a “mismatch.” The
effects of mismatches have long been recognized and evaluated [4, 24, 25, 26, 27].
Mismatch corrections are often neglected when using microwave power detectors. This
neglect is often not justified, and large errors result. Returning to the setup to measure
power, and now taking reflection coefficients () into account (Figure 6.16), we can
determine the value of the mismatch factor which gives the maximum power transfer as:
150 Ronald Ginley
g l
Signal
Power
to be
meter
measured
Output Power
port detector
Fig. 6.16 Basic measurement setup with reflection coefficients identified.
Signal Sij
1 2
source
Adapter Detector
Fig. 6.17 Power measurements using an adapter.
2
1 − g 1 − |l |2
Mgl = ; (6.14)
1 − g l 2
g is the reflection coefficient looking into the port where the detector is
connected.
6.5.2 Measuring power through an adapter

Sometimes you end up trying to measure power with a power detector that has a
microwave connector type different to the connector type of the port you are trying
to measure. The easiest way to make this measurement is simply to place the correct
adapter between the output port and the detector. This works well, but the efficiency loss
introduced by the adapter must be accounted for to achieve the highest measurement
accuracy possible. This situation is seen in Figure 6.17.
The cascaded effective efficiency of the adapter and the detector is
η = η1 η2 , (6.15)
where η1 is the efficiency of the adapter and η2 is the efficiency of the detector. It can be
shown that [28]:

|S12 |2 1 − |L |2
η1 = , (6.16)
|1 − S11 L |2 − |(S12 S21 − S11 S22 ) L + S22 |2
where Sij are the scattering parameters of the adapter and L is the reflection coefficient
looking into the detector. Both the scattering parameters of the adapter and L can be
determined through measurements. These are most commonly made through the use of
a vector network analyzer.
6.5.3 Power meter reference

An inherent characteristic of both diode and thermocouple power measurements is that
these measurements are considered open-loop. Thermistor power measurements are
inherently more accurate because of their DC-substitution, closed-loop process. The
bridge feedback of substituted DC power compensates for differences between ther-
mistor mounts and for drift in the thermistor resistance-power characteristic without
recalibration. With diodes and thermocouples, where there is no direct power substitu-
tion, sensitivity differences between sensors or drift in the sensitivity due to aging or
temperature can result in a different DC output voltage for the same RF power.
All thermocouple and diode power sensors require a power reference to some absolute
power level that is traceable to the manufacturer or a national standard to compensate
for the open-loop nature of the measurements. Power meters accomplish this power
traceability by using a highly stable, internal 50 MHz, 1 mW power reference oscillator.
The 1 mW reference power output is near the center of the dynamic range of thermo-
couple power detectors, but above the range of the sensitive older-style diode detectors.
For these detectors, a special 30 dB calibration attenuator, designed for excellent preci-
sion at 50 MHz, is generally supplied with the diode detector. When that attenuator is
attached to the power reference output on the power meter, the emerging power is 1 μW
(−30 dBm) [9].
6.6 Uncertainty considerations
It is very important when making microwave power measurements to understand the

accuracy of the measurements that you are making.Also, a detailed look at the uncertainty
components for a measurement can help to point out problems in a measurement as well
as ways to improve the measurements. Returning to the equation for a general power
measurement:
Psub
Po = , (6.17)
ηDet Mgl
we can use the propagation of uncertainties method and RSS combination of terms to
describe the uncertainty in Po (uP o ) [29, 30]. Assuming normal distributions for all
uncertainty components (if other than normal distributions are assumed, see [29] and
[30] for the proper correction factor to be applied to the component):
"
2 2 2 #
∂P ∂P ∂P
uP o = !
o o o
dηDet + dP sub + dMgl (6.18)
∂ηDet ∂Psub ∂Mgl
152 Ronald Ginley
where the partial derivatives of (6.17) are determined to be:
∂Po −Psub
= 2 ,
∂ηDet ηDet Mgl
∂Po 1
= ,
∂Psub ηDet Mgl
∂Po −Psub
= 2
.
∂Mgl ηDet Mgl
In (6.18), dPsub , dηDet , and dMgl are the uncertainty estimates for the measurements of the
substituted power, the effective efficiency of the detector, and the mismatch correction,
respectively. These are determined through a thorough examination of the sources of
uncertainty for each respective term. Please note that there are other ways of determining
the overall uncertainty and the individual components. The method being shown here is
primarily for illustration, although it can be used for a complete analysis if desired.
6.6.1 Power meter uncertainty – uncertainty in Psub

Use of thermistors, thermocouples, or diodes requires some form of power meter to be
used. The external power meter introduces additional errors that must be accounted for
when determining the substituted power. The uncertainty component for the power meter
needs to include instrumentation contributions, RF/DC conversion contributions, and
contributions from the power meter reference, if it is used. To get the lowest uncertainty
using the power reference, the mismatch correction for the reference port/detector pair
should be made. There are several good methods for obtaining the reflection coefficient
(g ) of the reference port; a good method is described in [31]. Other contributions to
the uncertainty, when determining Psub , include general instrumentation errors, power
meter zero-setting errors, noise, drift, and power linearity.
6.6.2 ηDet uncertainty

The uncertainty in the effective efficiency of the detector comes from the measurement
process that was used to determine the value of the effective efficiency. If your detector
is a primary transfer standard, then this is the uncertainty that comes from the calorime-
ter system. If it has been calibrated through a lower-level transfer standard, then this
uncertainty component comes from the transfer system used to evaluate the effective
efficiency of your detector.
6.6.3 Mismatch uncertainty

References [23] and [26] contain a detailed analysis of the derivation of the uncertainty
component due to the mismatch factor. In summation, from the definition of Mgl , as
shown in (6.14) and repeated here,

2
1 − g 1 − |l |2
Mgl = . (6.19)
1 − g l 2
The RSS mismatch uncertainty can be found from the separate partial derivatives of
(6.19) as:
⎡ ⎤
2 2 2
∂Mgl ∂Mgl ∂Mgl
uMgl = !⎣ d g + d |e | + dφgl ⎦, (6.20)
∂ g ∂ |l | ∂φgl

φgl = φg + φl (φg and φl are the phase angles of g and |l |, respectively).
where

d g , d |l |, and dφgl are the uncertainty estimates for the magnitude of the output
port reflection coefficient, the magnitude of the detector reflection coefficient, and the
uncertainty in the reflection coefficient arguments. Explicit expressions for the partial
differentials can be determined [23]. The important point here is that the mismatch
uncertainty is related to the uncertainties in the determination of the complex reflection
coefficients. These uncertainties come from the analysis of the network analyzer system
that was used to determine the reflection coefficients.
6.6.4 Adapter uncertainty

If use of an adapter is necessary for a power measurement, then the uncertainty for the
cascading of efficiency must be determined. This follows from (6.15) and (6.16). The
exact formula for the adapter uncertainty can be determined with the techniques used in
the previous sections. It can be seen that the uncertainty is composed of contributions from
the determination of the scattering parameters of the adapter. The process of determining
these scattering parameters, especially for adapters with different connector types on each
end, has fairly high uncertainties and is generally one of the larger contributors to the
overall uncertainty of a power measurement.
6.6.5 Device repeatability

The connection repeatability for the detector to the output port must also be considered.
The easiest way to estimate this term is to use multiple connections of the detector to
the output port. Multiple measurement passes need to be made and the detector should
be disconnected and reconnected between each measurement pass. The average of the
results for the measurement passes is the final result, and the standard deviation of all of
the connections is the uncertainty term as defined by:
" $N 2 #1/2
j =1 ηj − ηa
urepeat = , (6.21)
N −1
154 Ronald Ginley
where N is the number of connections/measurements, ηj is the j -th measurement of the

effective efficiency, and ηa is the mean of the N measurements.
6.7 Examples
Example 1: The direct comparison power system
While calorimeters are a very good way to evaluate the effective efficiency of a
substitution-type detector, they are slow, up to 30 minutes per frequency point, costly,
and very hard to evaluate. A much easier approach to determine the effective efficiency of
a detector is to use a transfer system. The direct comparison power measurement system
is a good example of a simple transfer system [21–23]. The mathematics for the direct
comparison system will be developed as an example of the power measurement process.
Figure 6.18 shows a block diagram of a direct comparison system. Overall the system
is very simple.A signal generator sends an RF signal into a power splitter/divider that then
splits the signal between a monitor detector and either the calibration standard detector
or the DUT. The detectors are connected to power meters whose output is connected
to a DVM, if needed (depending on the type of power meter), or whose output is read
directly by a connected computer through an instrument interface bus.
To calibrate the system, a detector with a known effective efficiency (ηstd ) is connected
to the test port of the power splitter. From the known ηstd of the standard, the reflection
coefficient of the standard, the reflection coefficient looking back into the splitter, and the
substituted power measured in both the standard and the monitor detector, a value can be
determined (Ka ) for each measurement frequency that relates the power available at the
test port to the power measured in the monitor detector. Mathematically, the calibration
is represented by
Pdc−std
Ka = , (6.22)
ηstd PM−std Mgl−std
Test port
DVM
Standard
Power
or DUT
Resistive meter
detector
power
splitter
Computer
Signal
generator
DVM
Monitor
Power
power
meter
detector
Computer
Fig. 6.18 Block diagram of a direct comparison system.

where Ka is the calibration factor for the splitter, Pdc−std is the substituted power read
from the calibration standard; ηstd is the known effective efficiency of the calibration
standard; PM−std is the substituted power read from the monitor detector during the
calibration; and Mgl−std is the mismatch factor from the reflection coefficients of the
standard and the splitter.
With Ka known, the effective efficiency (ηe ) of the DUT can be determined. The
process for measuring the ηe of the DUT is the reverse of the calibration process. From the
substituted power readings of the monitor detector and the DUT, the reflection coefficient
of the DUT, the reflection coefficient looking into the test port, and Ka , the ηe of the DUT
can be determined. Note that all of the power readings are used in ratios (monitor detector
to the standard or DUT) and are never used as absolute power values. By doing this, any
drift of the signal amplitude is negated. The measurement process is represented by
PDC−DUT
ηDUT = , (6.23)
Ka PM−DUT Mgl−DUT
where ηDUT is the effective efficiency of the DUT; PDC−DUT is the substituted power
read from the DUT; PM−DUT is the substituted power read from the monitor detector
during the DUT measurement; and Mgl−DUT is the mismatch factor from the reflection
coefficients of the DUT and the splitter.
Now, taking (6.23) for ηDUT and substituting for Ka from (6.22), we get:
ηstd P DC−DUT PM−std Mgl−std

ηDUT = . (6.24)
PDC−std PM−DUT Mgl−DUT
The terms Mgl−std and Mgl−DUT in (6.24) are expanded as:
2
1 − g 1 − |std |2
Mgl−std = (6.25)
1 − g std 2
2
1 − g 1 − |DUT |2
Mgl−DUT = , (6.26)
1 − g DUT 2
with std being the reflection coefficient of the detector with known effective efficiency,
DUT the reflection coefficient of the unknown detector, and g the equivalent source
mismatch looking into the splitter.
The reflection coefficient of the standard and the DUT are measured directly with a
VNA. There are several techniques for determining the equivalent source mismatch term
g looking into the test port of the splitter. g is not a true reflection coefficient, as it is
determined in such a manner that it is invariant with respect to what is connected to the
156 Ronald Ginley
other ports of the splitter. In terms of the scattering parameters of the splitter (Sij ), g is
defined as:
S12 S23
g = S22 − . (6.27)
S13
Normally the true reflection coefficient of a port of a power splitter is dependent on
what is connected to the other ports of the splitter. Good references for the evaluation of
g are [32, 33].
Example 2: Available power calculation

For this example, assume that the power detector you are using for measurement has an
effective efficiency of ηDet = 0.985 and this detector is connected to the output of a signal
generator. The signal generator’s output level is set so that the detector reads 100 mW.
The magnitude and angle of the reflection coefficients for the detector and the output
port of the generator have been measured to be Det = 0.065 at 30◦ and g = 0.055 at
67◦ , respectively.
What is the available power at the output port of the signal generator?
Returning to (6.13) for a general power measurement:
Psub
Po = . (6.28)
ηDet Mgl
The substituted power reading from the detector needs to be corrected for the effective
efficiency of the detector and the mismatch factor, from the signal generator output port
and the detector, to determine the actual power at the output port (Po ). Calculating the
mismatch factor for maximum power transfer:
2
1 − g 1 − |l |2 1 − 0.0552 1 − 0.0652
Mgl = = 2 = 0.992. (6.29)
1 − g l 2 ◦ ◦
1 − 0.05567 0.06530
Thus,
100 mW
Po = = 102.34 mW. (6.30)
0.985 0.992
Example 3: Power measurement uncertainty calculation

A measurement is made using a Type-N thermoelectric power detector. Calculate both
the worst-case uncertainty and the RSS uncertainty for the measurement. For the RSS
calculation, use a coverage factor of 2 and assume that all the individual components
have a normal distribution.
Assume that the principal error sources are the uncertainty in the calibration of the
effective efficiency of the detector, the mismatch uncertainty, the power meter uncertainty
(instrumentation and power reference), and the repeatability of the device.
Through evaluation of your system, you have estimated the individual uncertainty
terms (with a coverage factor of 2) to be:
us = ±1.8% (effective efficiency uncertainty)

uMgl = ±2.1% (mismatch uncertainty)
upm = ±3.1% (power meter uncertainty)
ur = ±0.8% (repeatability uncertainty).
To calculate the worst-case uncertainty, the uncertainty components are linearly summed,
giving:
UW C = 1.8 + 2.1 + 3.1 + 0.8 = ±7.8%. (6.31)
To calculate the RSS uncertainty with a coverage factor of two and normal distributions:

2 2 2 2
1.8 2.1 3.1 0.8
URSS = 2 × + + + = ± 4.23%. (6.32)
2 2 2 2
The RSS method is a more realistic method for calculating the uncertainty, as the worst
case method assumes that all the components have their maximum value and are in such
a direction as to add together constructively.
6.8 Conclusions
Power measurements are very important in describing how microwave circuits work
and how information is transferred within and through the circuits. To be able to make
a good microwave power measurement, several questions need to be considered: how
the data is going to be used, what level of accuracy is necessary, and how the signal is
to be measured. The proper choice of microwave power detector type, understanding
the way these detectors work, different considerations for making measurements, and
the uncertainty components related to the measurements have been described to help
answer the above questions. A firm understanding of the basics of power measurements
will allow you to more easily make the measurements that support research and product
development, and evaluation.
References
[1] D. Halliday and R. Resnick, Fundamentals of Physics. Wiley Illustrated, 1974, Chapter 6,
Section 7.
[2] R. P. Feynman, R. B. Leighton, and M. Sands, The Feynman Lectures on Physics. Pearson
Education, Volume I, 1963, pp. 13–23.
[3] “Fundamental of RF and Microwave Power Measurements (Part 1),” Agilent Technologies,
Application Note 64-1, 5988-9213EN, April 2003.
158 Ronald Ginley
[4] G. F. Engen, Microwave Circuit Theory and Foundations of Microwave Metrology. London,
UK: Peter Peregrinus Ltd., 1992, pp. 103–128.
[5] M. Kinishita, “Atomic microwave power standard based on the Rabi frequency,” IEEE Trans.
Instrum. and Meas., vol. 60, issue 7, July 2001, pp. 2696–2701.
[6] “Digital Modulation in Communications Systems – An Introduction,” Agilent Technologies,
Application Note 1298, 5965–7160E, 2001.
[7] G. Breed, “Fundamentals of pulsed and time-gated measurements,” High Frequency
Electronics, Nov. 2010, pp. 52–56.
[8] “Gated Measurements Made Easy,” News from Rohde & Schwarz, No. 185, 2005,
pp. 15–17.
[9] “Fundamentals of RF and Microwave Power Measurements (Part 2),” Agilent Technologies,
Application Note 64-2, 5988-9214EN, July 2006.
[10] “Using Error Vector Magnitude Measurements to Analyze and Troubleshoot Vector-
Modulated Signals,” Agilent Technologies Product Note 89400-14, 2000.
[11] A. Fantom, Radio Frequency and Microwave Power Measurement. London, UK: Peter
Peregrinus Ltd., 1990, chapters 2–8.
[12] N. T. Larsen, “A new self-balancing DC-substitution RF power meter,” IEEE Trans. Instrum.
and Meas., vol. IM-25, no. 4, Dec. 1976, pp. 343–347.
[13] F. R. Clague and P. G. Voris, “Coaxial Reference Standard for Microwave Power,” NIST
Technical Note 1357, April 1993.
[14] F. R. Clague, “A Calibration Service for Coaxial Reference Standards for Microwave Power,”
NIST Technical Note 1374, May 1995.
[15] P. S. Filipski, R. F. Clark, and D. C. Paulusse, “Calibration of HF thermal voltage converters
using an asymmetrical tee,” IEEE Trans. Instrum. and Meas., vol. 50, no. 2, April 2001,
pp. 345–348.
[16] M. Halawa and N. Al-Rashid, “Performance of the single junction thermal voltage con-
verter at 1 MHz via equivalent circuit simulation,” Cal Labs Magazine, Apr./May/Jun. 2009,
pp. 40–45.
[17] R. J. Wylde, “Installation and Operating Instructions for the TK TeraHertz Absolute Power
Meter System,” product manual, Nov. 2002.
[18] “Power Measurement above 110 GHz,” VDI Application Note, Oct. 2007.
[19] K. E. Bramall, “Accurate microwave high power measurements using a cascaded coupler
method,” Journal of Research of the National Bureau of Standards, vol. 75c, no. 3 and 4,
July–Dec. 1971, pp. 185–192.
[20] “Fundamental of RF and Microwave Power Measurements (Part 4),” Agilent Technologies,
Application Note 64-2, 5988-9214EN, Sept. 2008.
[21] R. A. Ginley, “A direct comparison system for measuring radio frequency power (100 kHz
to 18 GHz),” Proceedings of the Measurement Science Conference, 2006.
[22] J. R. Juroshek, “NIST 0.05–50 GHz direct comparison power calibration system,”
Proceedings of the Conference on Precesion Electromagnetic Measurements, 2000,
pp. 166–167.
[23] M. P. Weidman, “Direct comparison transfer of microwave power sensor calibrations,” NIST
Technical Note 1379, Jan. 1996.
[24] R. W. Beatty and A. C. McPherson, “Mismatch errors in microwave power measurements,”
Proceedings I.R.E., 1953, vol. 41, Sept. 1953, pp. 1112–1119.
[25] A. Y. Rumsfelt and L. B. Elwell, “Radio frequency power measurements,” Proc. IEEE,
vol. 55, no. 6, June 1967, pp. 837–850.
[26] “Fundamental of RF and Microwave Power Measurements (Part 3), Agilent Technologies,
Application Note 64-32, 5988-9215EN, April 2011.
[27] G. F. Engen, “A method of determining the mismatch correction in microwave power
measurements,” IEEE Trans. Instrum. Meas., vol. IM-17, no. 4, Dec. 1968, pp. 392–395.
[28] D. M. Kerns and R. W. Beatty, Basic Theory of Waveguide Junctions and Introductory
Microwave Network Analysis. London, UK: Pergamon Press, pp. 42–50.
[29] “Evaluation of Measurement data – Guide to the Expression of Uncertainty in Measurement,”
BIPM – JCGGM 100:2008, 1995.
[30] B. N. Taylor and C. E. Kuyatt, “Guidelines for Evaluating and Expressing the Uncertainty
of NIST Measurement Results", NIST, Technical Note 1297, 1993.
[31] K. Shimaoka, M. Shida, and K. Komiyama, “Source reflection coefficient measurements
of the power reference of power meters,” Proceedings of the Conference on Precision
Electromagnetic Measurements, July 2006,
[32] J. R. Juroshek, “A direct calibration method for measuring equivalent source mismatch,”
Microwave Journal, Oct. 1997, pp. 106–118.
[33] H. Jager, “Measurement Method for Determining the Equivalent Reflection Coeffi-
cient of Directional Couplers and Power Splitters,” Rohde & Schwarz Application Note
08.02-1EZ51_1E.
7 Modular systems for RF and
microwave measurements
Jin Bains
7.1 Introduction
One of the major progressions in RF, microwave, and wireless testing is the ability to make
fast, flexible, and accurate measurements using software-designed, modular instruments.
As RF applications have become increasingly more complex and challenging, legacy test,
validation, and design systems, which are generally expensive and rigid, have become
increasingly less competitive, and are being replaced by modular, software-designed
instruments that are more flexible, extensible, and designed to keep up with the rapid
pace of change in the RF and wireless industry. There has been an inflection point in the
industry, and the momentum behind software-designed modular instruments is expected
to continue accelerating.
Combining Moore’s Law with advances in RF technologies and processes has enabled
the development of smaller form-factor, lower-cost modular products to match the per-
formance and features of more traditional test products. Modular systems can take full
advantage of multi-core processors and make use of the latest FPGA technologies to
allow for the greatest measurement flexibility and timing control. These advances have
resulted in measurement devices whose core functionality is designed, at least par-
tially, by software written by the system designer(s). Software-designed instruments
are mainstream in today’s test systems. They allow scientists and engineers to use
software to specify pass/fail criteria, test execution flow, signal processing and math-
ematics, data/logging, and other required elements of test and measurement systems.
Software-designed instruments may have vendor-defined elements as part of the sys-
tem, but unlike purely vendor-defined solutions, software-designed instruments empower
engineers with the ability to design their test systems and instruments specifically for
their needs.
The expandability of modular systems allows for synchronized, phase-coherent mea-
surements on systems comprising multiple sources or receivers. The increased RF
performance of modern modular products has enabled highly accurate measurements
with greatly reduced time, space, and cost.
There is an increasing importance in the role of modular measurement systems in
radio frequency (RF) and microwave applications. This chapter discusses the fundamen-
tals of modular instruments and reviews some of the salient features of these systems
which allow them to be highly effective for many RF and microwave measurement
applications.
Modular systems for RF and microwave measurements 161
Traditional instrument Virtual instrument
Proprietary
PC processor, OS
processor, OS
Firmware User software
Bus
User-
Fixed user
defined
Timing and Measurement interface Timing and Modular
interface
control subsystem control hardware
Power supply Shared power supply
GPIB, LAN, USB Connectivity GPIB, LAN, USB Connectivity
Fig. 7.1 Traditional and virtual instrumentation.
7.1.1 Virtual instrumentation

The rapid adoption of the PC in the last 20 years catalyzed a revolution in instrumenta-
tion for test, measurement, and automation. One major development resulting from the
omnipresence of the PC is the concept of virtual instrumentation. Virtual instruments
offer several benefits to engineers and scientists who require increased productivity,
accuracy, and performance [1]. Figure 7.1 illustrates a high-level comparison of a virtual
instrument versus a traditional instrument. Note that the virtual instrument is always
composed of modular hardware.
“Virtual instruments” is a term that has a long history and it is a broad term used in
various industries and many applications. We will briefly explain the various definitions
and history of “virtual instruments” or “virtual instrumentation” and then continue the
rest of the chapter referring only to “modular instruments,” which is a more common
term to use in the current era to represent measurement systems composed of mod-
ular hardware products. There are several definitions associated with the term virtual
instruments [2]. In a general sense, a virtual instrument consists of an industry-standard
computer or workstation equipped with powerful application software, cost-effective
hardware such as plug-in boards, connected by a high-speed bus, utilizing driver software,
which together perform the functions of traditional instruments. Virtual instruments rep-
resent a fundamental shift from traditional hardware-centered instrumentation systems
to software-centered systems that exploit the computing power, productivity, display,
and connectivity capabilities of popular desktop computers and workstations. Although
the PC and integrated circuit technology have experienced significant advances in the
last two decades, it is software that truly provides the wherewithal to build on this pow-
erful hardware foundation to create virtual instruments, thus providing better ways to
innovate and significantly reduce cost. With virtual instruments, engineers and scientists
can build measurement and automation systems that suit their needs exactly, instead of
being limited by traditional fixed-function instruments.
162 Jin Bains
Brief history of virtual instruments and the introduction

of graphical programming
In the early days of instrumentation, prior to and for a decade or two after the Second
World War, measurement products were mostly analog devices, such as oscilloscopes
and voltage meters. The key technology driving these products was the vacuum tube,
and a variety of other purely analog components. In the 1950s, there was a gradual
shift towards digitizing the measured signals, so that digital processing of data became
possible.
As computer technology became available in the 1960s, Hewlett-Packard developed
the general purpose instrument bus (GPIB). This was originally called the HPIB bus (for
Hewlett-Packard) and provided an interface between the measuring instrument and the
computer. Through the 1970s, virtual instruments were controlled via a GPIB bus, and
the instrument control programs were generally written in BASIC. During this period
there was a massive expansion in measurement capability provided by test instruments.
This expansion was enabled to a large extent by significant progress in integrated cir-
cuit technology, which was following a pace of evolution described very accurately by
Moore’s Law [3].
With further rapid advances in computer technology, and in particular with the advent
of the Macintosh, the instrumentation world was prepared for a new paradigm. This came
with the emergence of LabVIEW (Laboratory Virtual Instrument Engineering Work-
bench), developed by National Instruments in 1986 [4]. LabVIEW greatly enhanced
the ease of use of instrument products through a graphical user interface, resulting in
a friendly, powerful method for control, measurement, and analysis. LabVIEW also
allowed customers to extend or add functionality to previously fixed, closed instruments
and as such, it started a trend of software being used to design the functions of test sys-
tems. In addition to the growth of graphical programming, the performance capabilities
of virtual instruments were also enhanced by the rapid evolution of high-speed com-
puter buses, enabling measurement data to be transferred to a computer processor with
increasing bandwidth and decreasing latency. All of these factors suggest that the most
significant future advances in measurement products are likely to be driven by software
technologies.
Virtual instruments versus traditional instruments

Stand-alone traditional instruments, such as oscilloscopes and waveform generators,
can be very powerful and expensive, and are generally designed to perform one or more
specific tasks defined by the vendor. However, it is generally not possible to extend or
customize them. The knobs and buttons on the instrument, the built-in circuitry, and the
functions available, are all specific to the nature of the instrument. In addition, special
technology and expensive components must be developed to build these instruments,
making them very expensive and slow to adapt.
Virtual instruments, by virtue of being PC-based, inherently take advantage of the
benefits of the latest technology incorporated into off-the-shelf PCs. For example, vir-
tual instruments are generally significantly faster than boxed instruments because users
can always upgrade to the latest desktop PC processors rather than relying on the older
processors built inside the boxes. Traditional instruments also frequently lack portabil-
ity, whereas virtual instruments running on notebooks automatically incorporate their
portable nature. Engineers and scientists whose needs, applications, and requirements
change very quickly, need flexibility to create their own solutions.You can adapt a virtual
instrument to your particular needs without having to replace the entire device because
of the application software installed on the PC and the wide range of available plug-in
hardware.
7.1.2 Instrumentation standards for modular instruments

Modular instrumentation concepts have evolved over time to deal with increased demand
for lower cost and greater lifetime of instruments. Figure 7.2 shows a basic modular
instrumentation system, comprising the chassis, embedded controller, high-speed data
bus on the chassis backplane, the modular instruments themselves (cards inserted in the
chassis), and the graphical user interface.
Various instrumentation standards have been developed to meet the growing require-
ment for flexible test systems. These instrumentation standards have generally made
use of an existing PC bus technology. Some of the modular instrument standards are
discussed here.
VXI
VMEbus eXtensions for Instrumentation (VXI) is a standard platform for instrumentation
systems based on the VMEbus standard. Besides using the VME bus in the backplane,
VXI also implemented timing and synchronization features that were required for instru-
ments. The VXIbus Consortium was formed in 1987 with the intention of defining a
multivendor instrument-on-a-card standard. This consortium has defined system-level
components required for hardware interoperability. The IEEE officially adopted the VXI
specification, IEEE 1155, in March 1993. The VXIplug&play Systems Alliance, founded
in September 1993, sought a higher level of system standardization to cover all VXI sys-
tem components. By focusing on software standardization, the alliance defined standards
to make VXI systems easy to integrate and use while maintaining multivendor software
interoperability. The success of VXI as an open, multivendor platform is a testament to
the value of multivendor standards, and for a period of time, made VXI the platform of
choice for open instrumentation systems.
Fig. 7.2 Modular instrumentation system, with graphical user interface.

164 Jin Bains
The demand for an industry-standard instrument-on-a-card architecture was driven by

the need for a reduction in the size of rack-and-stack instrumentation systems, tighter
timing and synchronization between multiple instruments, and transfer rates faster than
the 1 MB/s rate of the 8-bit General Purpose Interface Bus (GPIB). The modular form-
factor, high bandwidth, low-latency, and commercial success of the VMEbus made it
particularly attractive as an instrumentation platform.
LXI
LAN eXtensions for Instrumentation (LXI) is an instrument control standard based on
Local Area Network (LAN) and Ethernet technologies, web interfaces, and IEEE 1588
[5]. LXI offers three levels of synchronization that vendors can choose to implement in
their boxes. The LXI Consortium was founded in 2004 and the LXI 1.0 specification
was released in September 2005. The LXI Consortium’s goals were to increase the inter-
operability and functionality of Ethernet-based instruments by standardizing common
operations and interfaces and to develop, support, and promote the LXI standard.
The need for the LXI standard arose owing to the widespread use of Virtual Private
Networks (VPN) and an increase in the number of instruments available on the Internet.
PXI/PXI Express
PCI eXtensions for Instrumentation (PXI) is a rugged, PC-based platform [6]. PXI
combines the Peripheral Component Interconnect (PCI) electrical bus with the rugged,
modular Eurocard mechanical packaging of CompactPCI and adds specialized synchro-
nization buses and key software features. PXI also adds mechanical, electrical, and
software features that define complete systems for test and measurement, data acquisi-
tion, and manufacturing applications. These systems are used for applications such as
manufacturing test, military and aerospace, machine monitoring, automotive, and indus-
trial test. PXI is currently the most popular and fastest-growing modular instruments
form factor.
National Instruments developed and announced the PXI specification in 1997 and
launched it in 1998 as an open industry specification to meet the increasing demand
for complex instrumentation systems. Currently, PXI is governed by the PXI Systems
Alliance (PXISA), a group of more than 70 companies that are chartered to promote
the standard, ensure interoperability, and maintain the PXI specification. Because PXI
is an open specification, any vendor can build PXI products. CompactPCI, the standard
regulated by the PCI Industrial Computer Manufacturers Group (PICMG), and PXI
modules can reside in the same PXI system without any conflict, because interoperability
between CompactPCI and PXI is a key feature of the PXI specification.
The demand for a high-performance, low-cost deployment solution for measurement
and automation systems paved the way for developing this specification.
PCI Express, the next-generation of the PCI bus, was introduced in 2004 to increase
the measurement throughput of PXI. Today, most PCs ship with a combination of PCI
and PCI Express slots. It will not be long before the PCI bus is completely phased out.
The integration of PCI Express signaling into the PXI standard increases the backplane
bandwidth from 132 MB/s to 6 GB/s, an improvement of 45 times. The PXIe specification
also enhances PXI timing and synchronization features by incorporating a 100 MHz
differential reference clock and differential trigger lines. The PXI Express specification
adds these features to PXI while maintaining backwards compatibility.
7.1.3 PXI architecture

There is a growing emphasis on increasing measurement speed and improving the flex-
ibility and performance of RF instruments. For these reasons, PXI and PXIe (used
interchangeably here, since the basic architecture is identical) are considered the solution
for next generation RF and microwave test systems. The following section discusses the
PXI hardware and software architectures.
Hardware architecture
PXI systems are composed of three basic components – chassis, system controller, and
peripheral modules, as shown in Figure 7.3.
PXI chassis
The chassis provides rugged and modular packaging for the system. Chassis are generally
available in 4-, 6-, 8-, 14-, and 18-slot 3U and 6U sizes. Some chassis include AC and
DC power supplies and integrated signal conditioning.
PXI controllers
Most PXI chassis contain a system controller slot as the leftmost slot of the chassis.
There are a few options when determining the best system controller for an application,
including remote controllers from a desktop, workstation, server, or laptop computer
and high-performance embedded controllers with either a general purpose (OS) such as
Windows or a real-time OS.
Chassis
Controller
Modules
Fig. 7.3 Components of a PXI system.

166 Jin Bains
PXI peripheral modules

Currently, PXI provides the industry’s highest-bandwidth and lowest-latency bus with
modular I/O for applications ranging from high-resolution DC to 26.5 GHz RF. While
the smaller, 3U height instruments are by far the most widely used, because PXI is
compatible with CompactPCI, there is an option to use 3U or 6U CompactPCI modules
in a PXI or PXI Express system. Additionally, CardBus/PCMCIA and PCI Mezzanine
Card (PMC) cards, among others, can be installed in PXI systems using carrier modules.
Software architecture
The development and operation of Windows-based PXI systems is no different from
that of a standard Windows-based PC. Additionally, because the PXI backplane uses the
industry-standard PCI bus, writing software to communicate with PXI modules is, in
most cases, identical to that of PCI boards. Therefore, there is no need to rewrite existing
application software, example code, and programming techniques when moving software
between PC-based and PXI-based systems.
PXI Express systems also provide software compatibility to help preserve the invest-
ment in existing software. Because PCI Express uses the same driver and OS model as
PCI, the specification guarantees that there is complete software compatibility among
PCI-based systems. As a result, neither vendors nor customers need to change driver or
application software for PCI Express-based systems.
As an alternative to Windows-based systems, a real-time software architecture can
be used for time-critical applications requiring deterministic loop rates and headless
operation (no keyboard, mouse, or monitor). Real-time operating systems help prioritize
tasks so that the most critical task always takes control of the processor when needed.
With this feature, an application can be programmed with predictable results and reduced
jitter. The PXI specification presents software frameworks for PXI systems based on
Microsoft Windows operating systems.
Understanding PC technologies
To address the growing appetite for bandwidth, the PCI Express bus was introduced by
Intel in 2004. Originally designed to enable high-speed audio and video streaming, PCI
Express is used to improve the data rate from measurement devices to PC memory by
up to 30 times more than the traditional PCI bus.
PCI and PCIe bandwidth versus latency

For considering the technical merits of alternative buses, bandwidth and latency are two
of the most important bus characteristics [8]. Bandwidth measures the rate at which
data is sent across the bus, often represented in bits or bytes per second, while latency
measures the inherent delay in data transmission across the bus. By analogy, if we were
to compare an instrumentation bus to a road, bandwidth would correlate to the width of
the road and the speed of travel, while latency would correlate to the number of stoplights
in the road.
Table 7.1 and Figure 7.4 compare the bandwidth and latency performance of various
instrument buses. An ideal instrument would have a very high bandwidth with very low
Table 7.1 Comparison of instrument control buses
Maximum Dedicated Latency (μs) Distance (m)

Bandwidth (MB/s) Bandwidth (with no
repeaters)
GPIB 1.8 (488.1) No 30 20

8(HS488)
USB 60 (USB HS) No 1000 (USB 1.1) 5
125 (USB 2.0)
LAN/LXI 12.5 (Fast) No 1000 (Fast) 100
125 (Gigabit) 1000 (Gigabit)
PXI/PXIe 132 (PXI) No (PXI) 0.7 Internal PC bus
4,000 (PXIe) Yes (PXIe)
Increasing (Improving) Bandwidth
10000
Good Better Best
Max Bandwidth (MB/s)
PCI Express (x4)

1000
Gigabit Ethernet PCI/PXI (32/33)

100 USB 2.0
IEEE 1394a
Fast Ethernet VME/VXI
10 GPIB(HS 488)
USB 1.1 GPIB(488.1)
1
10000 1000 100 10 1 0.1
Decreasing (improving) latency
Fig. 7.4 Bandwidth versus latency.
latency. A bus with high bandwidth can transmit more data in a given period than a
bus with low bandwidth [7]. A bus with low latency introduces less of a delay between
the time data was transmitted from one end and processed at the other end. Bandwidth
is important because it determines whether data can be sent as fast as it is acquired
and how much onboard memory instruments will need. Latency, while less observable,
has a direct impact on applications such as Digital Multimeter (DMM) measurements,
switching, and instrument configuration, because it affects how quickly a command sent
from one node on the bus, such as the PC controller, arrives at and is processed at another
node, such as the instrument.
7.1.4 The role of graphical system design software

Graphical system design, using an open platform of productive software and reconfig-
urable hardware, shortens the integration cycle for new technology and functionality.
It allows engineers to visualize and implement systems faster because the platform
168 Jin Bains
Fig. 7.5 Example of graphical system design software.
makes technology easier to access through intuitive interfaces to accelerate the design,
prototyping, and deployment of the system.
The productivity benefits of graphical system design span every industry in which
engineers create systems that need measurement and control. When using graphical
system design, one can make use of the work of other engineers in the platform ecosys-
tem by accessing thousands of software and hardware components to efficiently build
an application. Figure 7.5 shows an example of a graphical system design software
environment.
Instrument drivers
Instrument drivers are an integral component in modern automated test systems. They
perform the actual communication and control of the instrument hardware in the system,
and provide a high-level and easy-to-use programming model that turns complex instru-
ment measurement capabilities into simple software function calls. Instrument drivers
are used to simplify instrument control and reduce test program development.
Interchangeable Virtual Instruments (IVI) is a standard for instrument driver soft-
ware technology. IVI builds on the VXI plug&play specifications and incorporates new
features that address issues such as system performance, development flexibility, and
instrument interchangeability. IVI drivers also take advantage of the power of the VISA
I/O library defined by VXI plug&play to seamlessly communicate with instruments
across different I/O buses such as GPIB, VXI, PXI, Serial, Ethernet, and USB.
Analysis routines
Graphical System Design software allows for integrated analysis routines. In general, raw
data must be processed before it can be used for collecting information. Signal processing
involves analysis, interpretation, and manipulation of signals. Analysis routines give a
proper procedure for the analysis process. A very well-tested and correct analysis routine
helps to save development time. Analysis routines help to improve efficiency and iterative
correlation input variables to process. Therefore, the availability of well-defined and
tested analysis routines saves time in creating them.
7.1.5 Architecture of RF modular instruments

RF modular instruments provide all the capabilities of traditional stand-alone instru-
ments, but go well beyond the limitations inherent in the traditional paradigm. The
modular instrument breaks down the RF instrument into its key measurement blocks,
disaggregating the overall system and thereby opening the way for a lower-cost, more
flexible, highly expandable measurement system. As part of the modular instrumentation
system, each RF modular product also can take full advantage of the high-speed data
buses as well as the triggering and clocking functions available.
To demonstrate the distinction between a traditional instrument and a modular instru-
ment, we will take a closer look at both a vector signal analyzer (VSA) and a vector
signal generator (VSG), although we could have just as easily looked at a vector network
analyzer (VNA) or any other RF instrument.
Vector signal analyzer

A vector signal analyzer is a measurement device that can measure signals at RF and
microwave frequencies. The VSA converts the signal from a time domain representation
to a frequency domain representation. The time domain representation gives the ampli-
tudes of the signal at the instants of time during which it was sampled. However, in many
cases you need to know the frequency content of a signal rather than the amplitudes of
the individual samples. In the frequency domain, you can separate conceptually the sine
waves that add to form the complex time domain signal, as shown in Figure 7.6.
Magnitude
Freq
n uen
ai cy d
m om
ain
do
e
m
Ti
Freq
uen
cy
e
m
Ti
Fig. 7.6 Frequency and time domain measurements.

170 Jin Bains
Down-converter Digitizer
ADC
Local oscillator
Fig. 7.7 Modular vector signal analyzer block diagram.
An example of a modular VSA block diagram is given in Figure 7.7.

In many ways the block diagram of a modular VSA is identical to that of a traditional
stand-alone VSA. The signal to be analyzed passes through the attenuator. The atten-
uators reduce the gain of the signal. The mixer converts the incoming RF signal to a
low-frequency IF signal from the attenuator and local oscillator and passes to the IF
amplifier. The signal is amplified by the IF amplifier tuned to the frequency of the down-
converted signal, and is then filtered by the IF filter. The IF band pass filter removes the
unwanted signals and retains the desired IF signal only. The digitizer then converts the
signal to digital form. Finally, the signal data is displayed on the host computer.
However, the key distinction with the modular VSA is that the VSA is implemented
in three different modules, one for the down-converter, one for the local oscillator (LO),
and one for the digitizer. This is where the real power of the modular approach comes
into play, since the door is now open for a far more flexible, expandable, and upgradable
measurement solution.
Vector signal generator

VSGs are used to generate continuous or arbitrary waveforms which can be used as
an input for devices under test. A VSG uses direct RF up-conversion from differential
baseband I and Q signals. A digital-to-analog converter (DAC) generates baseband I
and Q signals. The signal passes through the filter and is modulated. The baseband
modulated signal undergoes analog up-conversion to frequency-translate the signal to
the RF frequency at which the signal is transmitted.
An example of a modular VSG block diagram is given in Figure 7.8.
Similar to the VSA, the modular VSG block diagram is distributed amongst multiple
core modules. There are three parts: the arbitrary waveform generator (AWG) module,
the local oscillator (LO), and the RF up-converter or IQ-modulator module. The VSG
instrument driver software operates all three hardware modules as a single instrument
by handling all module programming and interaction. This implementation allows the
Arbitrary waveform Up-Converter or IQ

generator Modulator
DAC
90° +
DAC
Local oscillator
Fig. 7.8 Modular vector signal generator block diagram.
user the flexibility to upgrade any module within the overall VSG system to change or
improve the performance of the system.
7.2 Understanding software-designed systems
The term “software-designed systems” means not only using software to measure data,
analyze it, and generate results, but also using software to design, prototype, and deploy
measurement systems. Using software for measurement systems enables the user to
design systems that generate and analyze RF signal measurements four times faster
than other modular instrumentation solutions and more than ten times faster than tradi-
tional box instruments. Because the solution is software-designed, engineers can easily
configure the same measurements used in hardware to fulfill multiple functionalities,
which in turn reduces the cost of testing. Another advantage of using software is that
the hardware has fixed functionality and vendor defined measurements, while soft-
ware enables the user to design the measurements required and also has an integrated
software GUI.
It is clear that in the modern world of increased abstraction and the need for
increasingly flexible systems, a software-designed measurement system has significant
advantages over a traditional hardware-centric system.
7.2.1 Measurement speed

To get clear benefits such as productivity and faster time-to-market, measurements have
to be performed faster. PXI modular instruments already perform RF measurements
significantly faster (around 3 to 10 times) than traditional instruments. This section
explains the factors that affect RF measurement time using PXI and how to make the
system work faster.
172 Jin Bains
RF in
PCle Interface
RF front end ADC DDC
PXle-5663
RF Signal
Analyzer Onboard
memory
Multicore
Hard drive Host memory
CPU
PCle Interface
PXle-8108
Embedded
Controller
Operating +
Motherboard
system
Fig. 7.9 Software-designed PXI instrument using a multicore CPU.
From RF input signal to measurement result

Typically, a software-designed RF measurement system consists of a PXI RF signal ana-
lyzer along with a PXI chassis and controller. While the simplest configuration consists
of only the analyzer, many PXI RF measurement systems contain additional modules
for RF signal generation and mixed signal or DC input and output. Figure 7.9 illustrates
the block diagram of a basic PXI RF signal analyzer, such as the NI PXIe-5663.
Measurement begins when an RF vector signal analyzer (VSA) starts to collect an
IF signal from the RF front end. The IF signal is processed in the analyzer’s digital
down-converter (DDC), and IQ samples are generated, which are stored in the onboard
memory. A PXI controller then fetches the IQ samples from the VSA’s onboard memory
through a PCI or PCIe data bus. Once the IQ samples are in the host controller’s memory,
a software-designed measurement algorithm produces the measurement result. By using
different measurement algorithms, PC-based measurement systems can compute a wide
range of time and frequency domain measurements including: power, frequency, spectral
mask margin, error vector magnitude (EVM), and many others.
The different stages of the measurement process can be evaluated and documented as
follows: Step 1: Acquires IQ samples; Step 2: Transfer IQ samples to host PC; Step 3:
Apply measurement algorithm.
During each step of the measurement process, various factors affect the overall mea-
surement time. Let us understand and evaluate how each factor affects the measurement
time of a typical spectral measurement. Assuming a frequency span of 50 MHz and a
resolution bandwidth (RBW) of 30 kHz, we can observe which step takes the longest
time. The full analysis is shown in Figure 7.10.
During step 1, factors such as software latency and the signal analyzer’s internal
acquisition engine are the biggest sources of delay. Because these are minor contributors
Anatomy of measurement time
PXI Express Vector Signal Analyzer (< 4 ms)

Total time = 4 ms*
Acquire IQ Transfer data Processing *Spans greater than 50 MHz
samples to CPU time (FFT) pay a 0.5 ms tuning penalty
< 4 ms
1 1
Acquisition time = = =
RBW 30 kHz
1 1
Number of samples = × 1.25 × Span = × 1.25 × 50 MHz = 2083
RBW 30 kHz
Data size (bytes) 8332 bytes

Transfer time = + Latency (sec) =
Data rate (bytes/sec) 600 MB/s
Fig. 7.10 Measurement time for a modular VSA.
to measurement time, a typical RF VSA can produce IQ samples within 30–40 µs of the
time that the acquisition was initiated in the software.
In step 2, the data bus bandwidth is the biggest contributor to measurement delays.
Step 3, which involves execution of the measurement algorithm, is fundamentally the
largest contributor to overall measurement time. In fact, one way to evaluate the influ-
ence of signal processing time is to compare the results from steps 1 and 2 to the
overall measurement time. Using a PXIe controller, the 50 MHz spectral mask mea-
surement (30 kHz RBW) can typically be performed in 2.8 ms. Given that steps 1
and 2 in the measurement process add up to a maximum of 90 µs, we note that for
this example, signal processing accounts for 97% (or more than 2.71 ms) of the total
measurement time.
The following section deals with the factors that affect the speed of measurement.
Role of the CPU

By observing the effect of the CPU on the overall measurement time, we can determine
that signal processing time is the bottleneck for measurement time. If all the calculations
are correct, the measurement time should improve with the use of a more powerful host
PC (CPU). Additionally, more intensive measurements take longer to perform.
While many factors, from RBW to number of symbols, affect the RF measurement
time, the easiest method to reduce measurement time without affecting measurement
quality is to use the fastest CPU available. The availability of a high-performance CPU
on a PXI measurement system is the main contributor to the speed of PXI measurements
over traditional instrumentation.
Signal processing and parallelism

The need to reduce the cost of wireless handsets continues to increase as wireless
devices increase in complexity and volume. Fortunately, multicore processors provide
174 Jin Bains
today’s software-designed instruments with a high-performance test solution. Multi-

core processors significantly improve test times in single device-under test (DUT)
testing. Processing time can be reduced further with the implementation of parallel DUT
configurations.
There are several software packages that give you ready-to-run, stand-alone signal
processing capabilities with high-level digital signal processing (DSP) tools and utilities
[9]. These software packages have functions that are designed for performing advanced
DSP and designing digital filters interactively. The following section explains some of
the advanced signal processing capabilities:
1 Joint time-frequency analysis

Unlike conventional analysis technologies, joint time-frequency analysis (JTFA) rou-
tines examine signals in both the time and frequency domains simultaneously. You
can apply JTFA in almost all applications, such as biomedical signals, radar image
processing, vibration analysis, machine testing, and dynamic signal analysis.
2 Wavelet analysis
Wavelets are a relatively new signal processing method. A wavelet transform is almost
always implemented as a bank of filters that decomposes a signal into multiple signal
bands. Wavelet transform separates and retains the signal features in one or more
of these sub-bands. Thus, you can easily extract signal features. In many cases, a
wavelet transform outperforms the conventional FFT in feature extraction and noise
reduction. Because the wavelet transform can extract signal features, it is used for
data compression, echo detection, pattern recognition, edge detection, cancellation,
speech recognition, texture analysis, and image compression.
3 Super-resolution spectral analysis
FFT is the primary tool for spectral analysis. For high-resolution spectra, FFT-based
methods need a large number of samples. However, in many cases, the data set is
limited because of a genuine lack of data or because users need to ensure that the
spectral characteristics of the signal do not change over the duration of the data record.
For cases where the number of data samples is limited, a model-based analysis can
be used to determine spectral characteristics. By using spectral analysis, a suitable
signal model is assumed and the coefficients of the model are determined. Based on
this model, the application can predict the missing points in the given finite data set
to achieve high-resolution spectra. In addition, model-based methods can be used to
estimate the amplitude, phase, damping factor, and frequency of damped sinusoids.
Additionally, super-resolution spectral analysis can be used in diverse applications
including biomedical research, economics, geophysics, noise, vibration, and speech
analysis.
Parallel programming techniques

The parallel execution structure of programming languages and the test executive intro-
duces a variety of potential techniques to reduce overall test time. Some test executive
software actually contains built-in functionality to allow engineers to more easily con-
figure parallel device testing. Programming techniques such as data and task parallelism
Multiple processing cores PFER
TXP
Raw IQ Data Parallel
PXI Vector measurement
Signal Analyzer results
PVT
ORFS
Fig. 7.11 Multicore processors enable parallel measurement algorithms.
allow measurement algorithms to operate more efficiently on multicore processors. When

using task parallelism, a programmer will structure the algorithm such that multiple
operations are performed simultaneously on a single data set.
In cellular measurement systems, task parallelism enables measurements such as
PFER, TxP, PvT, and ORFS to be performed simultaneously using a single set of IQ
data. This is illustrated in Figure 7.11.
Some typical GSM measurements are described below:
Phase and Frequency Error (PFER) – Phase and frequency error is a measurement
performed on signals that use the Gaussian Minimum Shift Keyed (GMSK) modulation
scheme. It is a comprehensive measurement of modulation quality that can identify a
wide variety of signal impairments.
Transmit Power (TXP) – The transmit power measurement describes the average power
of a GSM burst.
Power Versus Time (PVT) – A PVT measurement is actually a time domain power mea-
surement which compares the power of a GSM burst to a power mask. This measurement
is used to ensure that the ramp-up power of the transmitting device does continue for the
duration of the burst.
Output RF spectrum (ORFS) – ORFS is used to characterize the power output of the
transmit signal at a series of offsets from the carrier. Nonlinear components such as
power amplifiers can contribute to poor ORFS performance.
These results indicate that parallel programming techniques enable efficient processor
utilization on multicore CPUs. Because measurement speed is processor-limited, virtual
instrumentation systems can be upgraded to faster processors to reduce measurement
time further.
7.3 Multi-channel measurement systems
Modular instruments have many significant advantages when an application requires

multiple channel measurements. Figure 7.12 compares the architecture of a modular
176 Jin Bains
Fig. 7.12 Comparison of virtual instruments and traditional instruments for multi-channel systems.
Fig. 7.13 Typical PXI phase-coherent RF measurement system.
system versus a traditional system for multi-channel systems, clearly showing the ben-
efits of the modular system, in addition to the configurability possible through the
built-in FPGA.
7.3.1 Phase coherence and synchronization

The modular architectures of PXI RF instruments lend themselves to the phase-
coherent RF measurements required for multiple-input multiple-output (MIMO) and
beam-forming applications. Figure 7.13 illustrates a modular measurement system with
four synchronized RF analyzers (left side of chassis) and two synchronized RF signal
generators (right side of chassis).
Phase-coherent RF signal generation

The configuration of any phase-coherent RF system requires synchronization of every
clock signal present on the devices. There are different degrees of phase coherency. The

Generator Modulator
DAC
90° +
DAC

Generator Modulator
DAC
90° +
DAC
Local oscillator
Fig. 7.14 Synchronization of two RF generation channels.
phase in phase coherency refers to the relative phases of the different clock signals used in
the instruments. The best situation is when all channels share the same clock signals. The
next degree of phase coherency is when the different channels share a common reference
clock (10 MHz clock is common). Each channel uses the same reference clock as an
input to a PLL that synthesizes an LO for mixing, but each PLL introduces unique phase
noise to that channel’s LO. Finally, there is no phase coherency when no synchronization
or timing signals are shared between channels, making correlation of timing/phase data
between those channels difficult or impossible.
Some RF VSGs use direct up-conversion to translate baseband waveforms into RF
signals. Figure 7.14 illustrates the basic architecture of a two-channel RF VSG. Note that
both baseband sample clocks and the local oscillators (LOs) are shared between both
channels.
In Figure 7.14, observe that the VSG again consists of three modules: the local oscil-
lator (CW synthesizer), the arbitrary waveform generator, and the RF up-converter or
IQ modulator. These modules are used together as a single-channel RF vector signal
generator or they can be combined (as illustrated) with additional arbitrary waveform
generators and RF IQ modulators for multichannel signal generation applications. Addi-
tional cables are necessary for daisy-chaining the LO signal from the first IQ modulator
to the second as well as for daisy-chaining the reference clock from the first baseband
AWG to the second.
178 Jin Bains
It is clear that this approach can be used to extend the modular system to more than
two channels. The limitation on the number of channels is really only driven by the
requirements on phase coherence and the ability to preserve the integrity of the shared
LO signal.
Phase-coherent RF signal acquisition

Similarly, an RF/microwave VSA can be configured for multichannel applications. When
configuring multiple modules for phase-coherent RF signal acquisition, sharing the LO
between the RF down-converters, and sharing the reference clock between the IF digitiz-
ers ensures best performance. One way to build a VSA is by implementing signal stage
down-conversion to IF and digital down-conversion to baseband. This architecture is
one of the simplest to configure for phase-coherent applications because, unlike a three-
stage super-heterodyne VSA, only a single LO must be shared between each channel.
While synchronizing multiple VSAs, to achieve best performance, distribute a shared
baseband reference clock and LO between each analyzer to ensure that each channel is
configured in a phase-coherent manner. An example of a two-channel system is shown
in Figure 7.15.
The RF/microwave VSA consists of a local oscillator (CW synthesizer), an RF down-
converter, and an IF digitizer. When the VSA is combined with an additional down-
converter and digitizer, and cables are included for sharing the LO and the digitizer
reference clock, a complete two-channel RF acquisition system has been created. Again,
this approach allows the system to expand to well beyond two channels. The number of
channels is only really limited by the requirements on phase coherence and the ability
to preserve the integrity of the shared LO signal.
To understand the method of synchronization between multiple phase-coherent
RF VSAs, consider the block diagram of a typical RF signal analyzer as given in
Figure 7.16. Observe that even though a single LO is used to down-convert from RF
to IF, each analyzer must share three clocks. The analyzers share two signals – the
Vector signal analyzer Local oscillator
Shared LO Shared
ADC
Sample
Downconverter Digitizer Clock
Rx0
ADC DSP
Downconverter Digitizer
Rx1
ADC DSP
Downconverter + Digitizer
Fig. 7.15 Synchronization of a two-channel phase-coherent RF VSA system.

Downconverter Digitizer I
ADC
Q
NCO
CW LO 0°
VCO 90°
CIK10
LO out CIK10 out
Fig. 7.16 Block diagram of typical VSA, showing sharing of clocks.
LO and the reference clock. The digitizer synthesizes the ADC sample clock and
the NCO from the reference clock. An additional channel of phase-coherent acqui-
sition could be added by adding another analyzer that shares the LO and reference
clocks.
The local oscillator and 10 MHz digitizer clock are being shared between each RF
channel. While sharing a reference clock between each digitizer introduces uncorrelated
channel-to-channel phase jitter on the ADC sample clocks, the level of phase noise
introduced at IF is negligible compared to the phase noise of the rest of the system.
While emerging technologies such as MIMO and beam-forming produce new chal-
lenges for test engineers, modular RF instrumentation provides a cost-effective and
high-performance measurement solution to meet these challenges.
7.3.2 MIMO
As the prevalence of wireless communications continues to grow, there is an increasing
demand for more effective use of channel bandwidth. One of the most recent inno-
vations that helps achieve this is the development of multiple-input multiple-output
(MIMO) technology. MIMO uses multiple transmitters and receivers to increase the
effective signal-to-noise ratio (SNR). MIMO exploits a radio-wave phenomenon called
multipath: transmitted information bounces off walls, doors, and other objects, reaching
the receiving antenna multiple times through different routes and at slightly different
times. MIMO harnesses multipath with a technique known as spatial multiplexing. Spa-
tial multiplexing is a process by which a single data stream is multiplexed into multiple
data streams, within the same channel. In a physical channel with sufficient multipath
reflections, the maximum theoretical improvement in data rates scales linearly with the
number of spatial streams. You can refer to Figure 7.17 for a block diagram of spatial
multiplexing. MIMO is thus significantly different from the traditional method, whereby
the data rate is increased by using more of the limited bandwidth resources. MIMO is
the foundational technology used in the WLAN standard 802.11n, which can transmit
as high as 140 Mbits per second. The more recent 802.11ac VHT specification allows
for use of up to 8x8 MIMO, which allows for double the maximum spatial streams of
4x4 MIMO offered in 802.11n.
180 Jin Bains
011
Spatial Spatial
011010 Combiner
parser demux
010
Fig. 7.17 Block diagram of spatial multiplexing.
Spatial multiplexing requires multiple antennas at both the transmitter and receiver.
Accurate testing of MIMO transceivers presents significant challenges to existing test
instrumentation architectures. New architectures require advanced signal processing
algorithms to multiplex and de-multiplex various spatial streams, and tight synchroniza-
tion between each transmit and receive antenna. Efficient MIMO transceivers cannot
be designed by merely imitating or replicating the architecture of the traditional test
equipment. A more versatile and efficient solution to the challenge of designing MIMO
equipment is the use of software-designed modular instruments.
A MIMO system requires that the phase between each transmitter antenna remain
constant. When synchronizing baseband I and Q signals, skew must be minimal to
prevent distortion of the RF signal. When synchronizing multiple RF signals, phase
skew between each of the RF signals is tolerable but must be minimized as much as
possible. One of the challenges that MIMO poses for RF instrumentation is the need for
increased levels of phase coherence between instruments.
The existing architecture of traditional RF instruments produces uncorrelated phase
noise. Each instrument has a 10 MHz reference input and output, which is used for
synthesizing the required local oscillator and baseband clock signals. Each instrument has
an independent frequency synthesizer, which produces phase noise. The phase noise acts
as an additional source of EVM to the system. Sharing a common clock causes significant
phase-shifts, even for relatively small thermal changes and generates a need for frequent
calibration. Testing some of the advanced MIMO operating modes requires knowledge
about the relative phase difference between channels, which becomes difficult due to the
phase noise. One method of overcoming this problem is to distribute the local oscillator
from a single RF frequency source to the various RF up-converters and down-converters,
as explained in section 7.3.1. This modular architecture with shared LOs leads to a better
error vector magnitude (EVM) performance and requires less frequent calibration.
The modular PXI platform can be used to design and deploy MIMO systems quickly,
while ensuring both baseband and RF synchronization.
7.3.3 Direction finding

Direction finding is one of many applications that benefits from phase-coherent
analysis and generation. With phase-coherent RF vector signal analyzers (VSAs), a
phase-comparison direction-finding system can be built relatively easily.
Remember jumping on a trampoline as a child and “stealing the bounce” of a friend?

A perfectly timed jump would create the destructive interference necessary to bring the
unfortunate jumper to their knees. Sometimes, you would try “giving a bounce,” using
constructive interference to send your friend much higher than they could have reached
on their own. This behavior was observed in waves long before the introduction of the
modern trampoline and has found its way into many applications, such as direction
finding.
Figure 7.18 illustrates two transmitters and two possible scenarios of a signal source
creating constructive and destructive interference at the receiver with the phase shown
in the simplified polar plots. In beam-forming applications, a delay in the transmission
(phase change) from one of the sources will steer the direction of highest RF intensity,
controlling the direction of transmission.
In Figure 7.19, two phase-coherent receivers are used to measure the difference in
phase of a signal received by two different paths. Using this relative phase comparison,
the direction from which the transmission originated can be determined.
Even though this may not be practical on a trampoline, detecting the direction of an
RF transmission in this way is possible using the right tools and techniques.
Beamforming
Tx1
Tx2 Rx
Constructive interference Destructive interference
ΦTx1 = ΦTx2 ΦTx1 = ΦTx2 = 180°
Fig. 7.18 Controlling direction of optimal transmission by adjusting the phase difference between two
transmitters.
Direction finding
Tx
Rx1
Rx2
Phase comparison
ΦRx1
ΦRx1-ΦRx2 ΦRx2
Fig. 7.19 Determining a signal’s direction of arrival.

182 Jin Bains
Constructing a direction finder

To construct a basic phase-comparison direction finder, multiple receivers are needed,
as well as the ability to measure the phase difference between the received signals, and
some math. The first requirement is satisfied by adding more analyzers to the system, but
accurately measuring the difference in phase between two signals is more challenging. To
compare the phase difference between two measurements, the phase differences between
each oscillator used along the down-conversion path from the RF must be known pre-
cisely, as well as the time difference between multiple records from the analog-to-digital
converters (ADCs).
Figure 7.20 shows a solution using two VSAs sharing a common local oscillator (LO)
for down-conversion from RF and a 10 MHz reference clock.
With two phase-coherent signal analyzers, any phase difference between the two
RF channels can be easily measured and applied to direction-finding applications. For
example, a two-way family radio is used as a transmitter at 462.56 MHz with a pair
of general-purpose ultra-high-frequency (UHF) telescoping antennas connected to two
VSAs. By positioning the antennas 32.3 cm (one-half wavelength) apart, the expected
phase difference is 180 degrees when the antennas share a line of sight to the receivers
Phase-Synchronized NI PXIe-5663E
Vector Signal Analyzers
NI PXIe-5601 NI PXIe-5622
1
Rx1
ADC DOC 0
NI PXIe-5662
10 MHz Shared
CLK
Clock
NI PXIe-5601 NI PXIe-5622
1
Rx2
ADC DOC 0
Fig. 7.20 Sharing common LO and sample clock between multiple analyzers for tight phase
synchronization.
Tx
For R>> d:
360˚
r1
R.sin(θ)
R
r2
360˚
Rx1 Rx2
θ
d R.cos(θ)–01/2
R.cos(θ)+90
Fig. 7.21 Phase-comparison direction finding using two receivers.
and the expected phase difference is zero degrees when the transmitter is equidistant
from both antennas.
By tuning the VSAs to the carrier frequency of 462.56 MHz, I and Q samples are
continuously acquired to extract the phase. Verify the zero and 180 degree cases by
observing the difference between the phase measurements of the VSAs. The last step is
to solve for the intermediate cases.
As shown in Figure 7.21, the goal of a direction finder is to solve for θ . This math is
greatly simplified if R is assumed to be much larger than d, which is a valid approximation
for most signals of interest.
Knowing the frequency of interest, the distance between the antennas, and the differ-
ence in the measured phase, it is possible to solve for the corresponding values for θ .
Measuring a phase difference between two analyzers of 58 degrees would translate to a
θ of 71.2 degrees, whereas a phase difference of −121 degrees would yield a θ of 132.2
degrees.
7.3.4 Phase array

A phase array is an array of antennas in which the relative phases of the respective sig-
nals feeding the antennas are varied in such a way that the effective radiation pattern of the
array is reinforced in a desired direction and suppressed in undesired directions. A phased
array antenna is composed of multiple radiating elements each with a phase-shifter.
Beams are formed by shifting the phase of the signal emitted from each radiating ele-
ment, to provide constructive or destructive interference and steer the beams in the desired
direction. Similar to the examples given above with direction finding and multiple-
channel systems, a modular measurement system can give tremendous advantages over
traditional measurement systems.
184 Jin Bains
7.4 Highly customized measurement systems
Open software and off-the-shelf hardware can be used to customize a measurement

system, and the following are a few examples of concepts used in highly customized
measurement systems:
7.4.1 IQ data conditioning (flatness calibration)

IQ data conditioning or flatness calibration is the process of adjusting the strength of
certain frequencies within a signal to make the real signal more ideal. IQ data conditioning
creates a flatter frequency response, reduces spurs and images in the signal, and increases
the linear phase response.
For fixed instruments, a spectrum is returned after IQ data conditioning. For modular
instruments, the raw I/Q samples are returned and there is then the ability to apply various
manners of correction, before translating it to the typical spectral display, with improved
performance.
7.4.2 Streaming
Streaming is the process of transferring data to or from an instrument at a rate high
enough to sustain continuous acquisition or generation. Streaming involves direct data
transfer to or from memory. This memory can be the onboard memory of the instrument,
the RAM of the controller, or the hard drive of the controller. The rate at which data
is transferred to these various types of memory is limited by several factors, from the
system’s bus bandwidth to the read/write speed of the memory media.
The following are the advantages of using streaming for measurement systems:
• Higher and sustained acquisition and generation rates

• Reduction in measurement times due to the elimination of delays associated with
downloading or uploading waveforms to or from instruments
• Enables new applications.
A variety of media can be used for streaming applications, including IDE (Integrated
Drive Electronics) drives, SATA (Serial Advanced Technology Attachment) drives, and
RAID (Redundant Array of Inexpensive Disks) drive systems. For example, a chunk of
real-world RF spectrum can be recorded to test a device, and played back in the lab for a
virtual field test. If the device doesn’t behave as expected, the exact same scenario which
caused it to fail is replayed, and allows the problem to be debugged.
Streaming is also very useful in spectrum monitoring and signal intelligence systems,
where a large amount of spectral content needs to be analyzed.
Peer-to-peer (P2P) streaming

P2P streaming technology uses PCI Express to enable direct, point-to-point transfers
between multiple instruments without sending data through the host processor or memory
Fig. 7.22 NI FlexRIO peer-to-peer architecture.
[10]. This enables devices in a system to share information without burdening other
system resources.
With peer-to-peer technology, data streaming rates of more than 800 MB/s are possible
in a single direction. Maximum throughput is dependent on the streaming modules,
chassis, and, if the configuration warrants it, the controller. Generally, the lowest of
these rates is the maximum possible P2P bandwidth. Peer-to-peer transfers are designed
to have a very low latency, but it varies depending on the system configuration. The
main advantage of peer-to-peer streaming is that the data need not travel through the
host, reducing latency, increasing reliability, and increasing total system bandwidth.
Figure 7.22 shows the NI FlexRIO peer-to-peer architecture.
Device-to-host or host-to-device streaming

Device-to-host or host-to-device streaming is the process of streaming data from an
instrument (e.g. FlexRIO) to the host controller, and vice versa, for either subsequent
processing, or writing to disk.
With this technology, data streaming rates for a PCI Express Gen1 ×4 device of more
than 800 MB/s are possible in a single direction. The data transfer rate depends on the
bus technology, such as Gen1 or Gen2 PCI Express with a ×4 or ×8 link.
Integrating hard drives for record and playback

Record and playback is the process of using platforms that use high-bandwidth buses
(e.g. PCI and PCIe) to enable instruments to stream data to and from RAID 0 (striped)
hard drive arrays at high sustained rates [11]. RF record and playback systems combine
PXI RF signal analyzers and RF signal generators with RAID arrays for high-speed, long-
duration recording and playback. These systems reduce the cost for memory expansion by
using PC memory for contiguous acquisitions. You can use record and playback systems
to perform host-to-device or device-to-host streaming. For low-rate applications, where
186 Jin Bains
Fig. 7.23 Record and playback system.
the host CPU can keep up with the data streaming rates, you can use record and playback
systems for inline or real-time processing.
The ability to generate or acquire terabytes of continuous data can help you implement
applications previously possible only with custom hardware, such as the following:
• spectrum monitoring,
• packet sniffing,
• wireless receiver design, validation, and verification,
• digital video broadcasting Bit Error Rate (BER) tests.
Figure 7.23 shows a typical RF record and playback system.
7.4.3 Integrating FPGA technology

Field-programmable gate arrays (FPGAs) are reprogrammable silicon chips. Unlike
multi-core PC processors, which run software applications to implement functionalities,
programming an FPGA rewires the chip itself to implement the functionalities.
As shown in Figure 7.24, FPGA chip specifications include the amount of configurable
logic blocks, number of fixed function logic blocks, such as multipliers, and size of
memory resources such as embedded block RAM. There are many other parts to an
FPGA chip, but these are typically the most important when selecting and comparing
FPGAs for a particular application.
At the lowest level, configurable blocks of logic, such as slices or logic cells, are
made up of two basic things: flip-flops and look-up tables (LUTs). This is important
to note because the various FPGA families differ in the way flip-flops and LUTs are
packaged together. Virtex-II FPGAs for example, have slices with two LUTs and two
flip-flops, whereas Virtex-5 FPGAs have slices with four LUTs and four flip-flops. The
LUT architecture itself may also differ (4-input versus 6-input).
Every FPGA chip is made up of a finite number of predefined resources with pro-
grammable interconnects to implement a reconfigurable digital circuit. For any given
P R OGRAMMABLE I/O BLOCKS

INTERCONNECT
LOGIC BLOCKS
Fig. 7.24 Parts of an FPGA.
piece of synthesizable code, either graphical or textual, there is a corresponding cir-

cuit schematic that describes how logic blocks should be wired together. Synthesis
is the process of translating high-level programming languages into true hardware
implementations.
Beyond being user-programmable, FPGAs offer hardware-timed execution speed as
well as high determinism and reliability. They are truly parallel, so different processing
operations do not have to compete for the same resources. Each independent processing
task has its own dedicated section of the chip, and each task can function autonomously
without any influence from other logic blocks. As a result, adding more processing does
not affect the performance of another part of the application.
To increase performance even further, FPGAs offer the computational performance
to provide real-time measurements that occur faster than the time it takes to acquire the
data [12].
Figure 7.25 shows the difference in computational performance between a host-based
implementation and a FPGA-based implementation of an adjacent channel leakage ratio
(ACLR) calculation.
While the host-based implementation takes advantage of multiple high-performance
CPU cores and the high-bandwidth PXI Express data bus, the FPGA implementation
reduces measurement time even further by using dedicated, real-time processing and
eliminating unnecessary host data transfers. Furthermore, the peer-to-peer FIFO is con-
figured only once regardless of the number of averages, so measurement time scales are
based on the amount of time you need to acquire the RF data necessary to perform the
measurement.
To be useful in a software-designed instrumentation context, FPGAs must be repro-
grammable by the test engineer in software; in other words, they should be used to push
software programmability down into the hardware itself. In the past, FPGA technology
was available only to engineers with a deep understanding of digital hardware design
188 Jin Bains
Host FFT Implementation
ACLR
Host Configure FFT + Sum FFT + Sum 310
Calculation
Transfer Transfer
Instrument Configure Acquire Acquire Acquire Acquire Acquire Acquire Acquire Acquire Acquire Acquire
29 ms
FPGA FFT Implementation
ACLR
Host Configure
Calculation
Instrument Configure Acquire
Peer-to-Peer
FFT + Sum Transfer

FPGA
10 averages
4.6 ms
Fig. 7.25 Comparing performance of host and FPGA implementations.
software, such as hardware description languages like Verilog or VHDL, which use low-
level syntax to describe hardware behavior. Most test engineers do not have expertise in
these tools. However, the rise of high-level design tools is changing the rules of FPGA
programming, with new technologies that convert graphical block diagrams or even C
code into digital hardware circuitry. These system-level tools that abstract the details of
FPGA programming can bridge this gap.
The following are the advantages of using FPGA-based test methods:
• Real-time, continuous measurements

With their processing throughput, FPGAs can perform measurements faster than the
I/O hardware can acquire data, so it is now possible to test the DUT continuously.
Instead of acquiring, transferring data, then processing, which has a limited duty
cycle, we can now acquire and measure continuously.
• Custom triggering and acquisition
With an FPGA always acquiring data, we can define the measurement back-end by
adding custom triggering and data recording. We may want to add a complex digital
trigger, OR-ing and AND-ing several digital lines to detect a trigger condition.
• Closed-loop and dynamic test
As modern devices are increasingly integrated into the world around them, testing them
without incorporating feedback into the test system may not provide adequate test
coverage. For instance, modern communication schemes often incorporate acknowl-
edgement packets or bits. If the test system does not correctly interpret these and
respond appropriately and in a timely manner, then it is not obeying the protocol and
the DUT may not be accurately tested. Often, only an FPGA can provide these kinds
of low latency responses.
• Protocol emulation
Instead of constantly using software to go back and forth between protocol-level and
signal-level information, which can be tedious and slow, we can actually implement
protocols on the FPGA, allowing the test system to interact with general test hardware
at a protocol level.
7.5 Evolution of graphical system design
Graphical system design is an approach to designing an entire system, using more

intuitive graphical software and off-the-shelf (non-custom) hardware devices.
Graphical system design is a valuable technique for creating completely user-designed
instrumentation, in which the instrument is completely user-programmable, down to the
pin, allowing customization of every aspect of its behavior. A default behavior may be
used which resembles that of a “traditional” modular instrument, incorporating record-
based data movement paradigms, or a custom protocol may be implemented on the
instrument, allowing it to interact with the DUT as if the DUT is in a real-world operating
environment, not just a test mode.
The following are a few important milestones in the evolution of graphical system
design:
• Portable measurement algorithms
If the microprocessor initiated the virtual instrumentation revolution, then the FPGA
is ushering in its next phase. FPGAs have been used in instruments for many years. For
instance, today’s high-bandwidth oscilloscopes collect so much data that it is impossi-
ble for users to quickly analyze all of it. Hardware-defined algorithms on these devices,
often implemented on FPGAs, perform data analysis and reduction (averaging, wave-
form math, and triggering), compute statistics (mean, standard deviation, maximum,
and minimum), and process the data for display, all to present the results to the user in a
meaningful way. While these capabilities present obvious value, there is lost potential
in the closed nature of these FPGAs. In most cases, users cannot deploy their own
custom measurement algorithms to this powerful processing hardware. New software
tools strive to open these FPGAs to a variety of test and measurement algorithms,
effectively and automatically porting them from the CPU to the higher-performance
capabilities of the FPGA.
• Reconfigurable instruments
Test systems are reconfigured for endless reasons – from adapting to new test
requirements to accommodating instrument substitutions during calibration and repair
cycles. Software-designed instrumentation is based on a modular architecture that
enables a high degree of reconfigurability. Software-designed instruments consist of
modular acquisition/generation hardware whose functionality is characterized through
user-defined software running on a host multicore processor.
190 Jin Bains
The new software-designed architecture can meet application challenges that are
impossible to solve with traditional methods that require real-time decision making
by the host to properly test the device. Instead, engineers can fully deploy the intel-
ligence to the FPGA embedded on the instrument for pass/fail guidance. For some
applications, engineers also perform the communication over a protocol – wireless or
wired – which requires a significant layer of coding and decoding before making a
decision.
• Heterogeneous computing
Automated test systems have always comprised multiple types of instruments; each
best suited to different measurement tasks. An oscilloscope, for example, can make
a single DC voltage-level measurement, but a DMM provides better accuracy and
resolution. It is this mix of different instrumentation that enables tests to be conducted
in the most efficient and cost-effective manner possible.
A heterogeneous computing architecture is a system that distributes data, processing,
and program execution among different computing nodes that are each best suited to
specific computational tasks. For example, an RF test system that uses heterogeneous
computing may have a CPU controlling program execution with an FPGA performing
inline demodulation and a graphics processing unit (GPU) performing pattern match-
ing before storing all the results on a remote server. Test engineers need to determine
how to best use these computing nodes and architect systems to optimize processing
and data transfer.
• IP to the pin
The next phase in integrating design and test is the ability for engineers to deploy design
building blocks, known as intellectual property (IP) cores, to both the device under
test (DUT) and the reconfigurable instrument. This capability is called “IP to the pin”
because it drives user-defined software IP as close to the I/O pins of next-generation
reconfigurable instruments as possible. The software IP includes functions/algorithms
such as control logic, data acquisition, generation, digital protocols, encryption, math,
RF, and signal processing.
Graphical programming languages allow the creation of virtual instruments that can be
easily tested before embedding as a subroutine into a larger program. The graphical
approach also allows non-programmers to build programs by dragging and dropping
virtual representations of lab equipment with which they are already familiar.
7.6 Summary
There has been a rapid increase in the importance of the role of modular measurement
systems in RF and microwave applications. This chapter discussed the fundamentals of
modular systems and reviewed some of the salient features of these systems which allow
them to be highly effective for many RF and microwave measurement applications.
There is no doubt that the advent of modular instruments has been one of the major
progressions in RF and microwave testing recently. This approach has resulted in the
ability to make fast, flexible, and accurate measurements using SW-designed modular
test products. This is a trend that has gained momentum and will continue to accelerate.
For instance, it can be very difficult to solve the ever-changing needs of the wireless
industry with traditional test products that are often expensive, fairly large, and usually
rigid. The expandability of modular systems allows for synchronized, phase-coherent
measurements on systems comprising multiple sources or receivers. The increased RF
performance of modern modular products has enabled highly accurate measurements in
a fraction of the time, space, and cost.
Further advances in RF technologies and processes have enabled the development
of smaller form-factor, lower-cost modular products to match the performance and fea-
tures of more traditional test products. Modular systems, wrapped in graphical design
software, can take full advantage of multi-core processors and make use of the latest
FPGA technologies to allow for the greatest measurement flexibility and timing control.
This capability extends from the early design phase all the way through deployment of
a measurement system, and is commonly referred to as Graphical System Design.
References
[1] National Instruments. (2009, May). Virtual Instrumentation. [Online]. Available:

http://zone.ni.com/devzone/cda/tut/p/id/4752
[2] C. F. Coombs, “Virtual Instruments,” in Electronic Instrument Handbook, 2nd ed. New York:
McGraw-Hill, 1995.
[3] G. E. Moore, “Cramming more components onto integrated circuits,” Electronics, vol. 38,
no. 8, pp. 114–117, Apr. 1965.
[4] M. Santori, “An Instrument that isn’t really,” IEEE Spectrum, vol. 27, no. 8, pp. 36–39, Aug.
1990.
[5] National Instruments. (2011, March). What is LXI?. [Online]. Available: http://zone.ni.com/
devzone/cda/tut/p/id/7255
[6] National Instruments. (2011, Dec.). What is PXI?. [Online]. Available: http://zone.ni.com/
[7] National Instruments. (2010, Oct.). PXI Express FAQ. [Online].Available: http://zone.ni.com/
[8] M. Friedman and J. Schwartz, “Techniques for Architecting High-Performance Hybrid Test
Systems,” in IEEE Autotestcon 2008, Salt Lake City, UT: 2008, pp. 282–285.
[9] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing. New Jersey: Prentice-
Hall, Inc, 1999.
[10] National Instruments. (2011, Dec.). An Introduction to Peer to Peer Streaming. [Online].
Available: http://zone.ni.com/devzone/cda/tut/p/id/10801
[11] National Instruments. (2011, Aug.). Introduction to Record and Playback. [Online].
Available: http://zone.ni.com/devzone/cda/tut/p/id/7209
[12] National Instruments. (2012, Feb.). Make Your Measurements Faster With FPGA Technology.
[Online]. Available: http://zone.ni.com/devzone/cda/pub/p/id/1513
Part III
Linear measurements
8 Two-port network analyzer
calibration
Andrea Ferrero
8.1 Introduction
Although VNAs are probably the most advanced microwave systems, with broadband
sources, high-speed and high-dynamic-range receivers, the intrinsic property of distribute
components makes a calibration procedure mandatory to obtain reasonable results due
to an enormous systematic error. To stress this fundamental problem, imagine weighing
300 g of ham with a one-ton plate scale! This is more or less the same influence as
systematic phase error introduced by a 1 meter teflon cable in front of a VNA port at
10 GHz if we are trying to measure 1 degree of phase-shifting on a DUT S11 parameter.
Not only the phase, but also the magnitude as well is affected, due to different attenuation
paths in various system sections. Clearly without a proper correction the measurement
quality would be unacceptable.
In the early development of VNA, hardware compensation with line stretchers and
variable gain amplifiers was attempted, but it’s only with the introduction of computer-
controlled digital VNAs, that specific signal processing techniques allow a real-time
correction of the most important errors. During the last forty years several algorithms
have been proposed especially for one- or two-port VNAs; some of them like TRL, SOLT,
LRM, and SOLR, became a de-facto standard in all modern VNA firmware; however,
many others have been proposed to solve particular problems [1– 4]. This chapter
presents a review of the error models and the main VNA calibrations, by focusing the
attention on their commonalities and by pointing out their different fields of application.
The development follows the system approach born in the early 1990s, rather than the
traditional one based on the analysis of all possible sources of error [3– 5].
8.2 Error model
Let’s consider a one-port VNA essential block scheme, as shown in Figure 8.1; from the
microwave ports to the digital data we have the following significant parts:
• the source,
• the microwave test set,
• the down-converter,
• the IF Digitizer.
196 Andrea Ferrero
The source provides the required microwave signals, while the microwave test set
includes all the microwave components such as couplers, cables, and adapters from the
source reference plane to the DUT one. This block is modeled as a 4-port network,
where two ports are loaded with the mixers/samplers. The basic hypothesis to develop
a general error model is the overall system LINEARITY, i.e. it’s assumed that every
component from the cables to the A/D is linear. Thus a set of linear equations link the
digital outputs of the IF A/D converters with the a and b waves at the DUT reference
plane. The validity of such an approach is mainly constrained by the linear region of
the mixer/sampler, while the nonlinearity of the A/D converter is negligible. The system
works in the frequency domain, i.e., the source is supposed to be sinusoidal and the
receiver is strictly narrowband, typically 100 Hz, for accurate measurements. At each
frequency we have a pair of complex numbers at the two IF digitizer outputs usually
called the measured waves, am and bm . The linearity assumption sets a C2 =⇒ C2 linear
application which obviously can be written as:
% & " #% &

b1 d11 d12 am1
= , (8.1)
a1 d21 d22 bm1
where the dij are four error coefficients. However, it’s worth deriving (8.1) from classical
network theory. Let S be the scattering matrix of the four-port test set with the wave
convention of Figure 8.1 as:
⎡ ⎤ ⎡ ⎤⎡ ⎤
b0 S11 S12 S13 S14 a0
⎢a1 ⎥ ⎢S21 S22 S23 S24 ⎥ ⎢b1 ⎥
⎥ ⎢
⎢ ⎥=⎢ ⎥. (8.2)
⎣b3 ⎦ ⎣S31 S32 S33 S34 ⎦ ⎣a3 ⎦
b4 S41 S42 S43 S44 a4
am1 bm1
IF Digitizer
Vm1 Vm2
Down-conversion
b3 a3 b4 a4
b0 a1 D
U
Microwave components b1 T
a0
Fig. 8.1 One-port VNA essential block scheme.

Two-port network analyzer calibration 197
The two mixers load ports 3 and 4 with two generic reflection coefficients 3 and 4 as:
a3 = 3 b3 (8.3)
a4 = 4 b4 . (8.4)
The IF output voltages Vm1 and Vm2 are the low-frequency images of the total voltage
at the RF input ports, i.e.
Vm1 = α3 (1 + 3 )b3 (8.5)

Vm2 = α4 (1 + 4 )b4 , (8.6)
where αi are the proper mixer conversion factors. Finally the readings are linked with
the IF voltages by the A/D coefficients:
am1 = β1 Vm1 (8.7)

bm1 = β2 Vm2 (8.8)
We have:
% & % &% &
β1 α3 (1 + 3 )b3 = am1 b ξ 0 am1
⇒ 3 = 1 (8.9)
β2 α4 (1 + 4 )b4 = bm1 b4 0 ξ2 bm1
' () *

while from (8.2)–(8.4):

⎡ ⎤⎡ ⎤ ⎡ ⎤
−S11 1 −S12 0 a0 S13 3 S14 4
% &
⎢−S21 0 −S22 1⎥ ⎢ ⎥ ⎢ S24 4 ⎥
⎢ ⎥ ⎢b0 ⎥ = ⎢ S23 3 ⎥ b3 . (8.10)
⎣−S31 0 −S32 0⎦ ⎣b1 ⎦ ⎣(S33 3 − 1) S34 4 ⎦ b4
−S41 0 −S42 0 a1 S43 3 (S44 4 − 1)
' () * ' () *
W Q
Finally:
⎡ ⎤
a0
% & % &
⎢b 0 ⎥
⎢ ⎥ = W−1 Q am1 = D am1 . (8.11)
⎣b 1 ⎦ bm1 bm1
a1
By taking the last two rows of (8.11), the linear system of (8.1) is obtained. It’s worth
noting that the elements of the matrix D are independent of the loading conditions at the
source port, i.e. the error coefficients of a full reflectometer-based VNA are independent of
the source, which means the source can be changed after the calibration without affecting
its validity. This important result is not obvious and fundamental for many applications.
Traditionally the error coefficients are organized as S parameters of a fictitious network,
called the error box, which is interposed between the DUT and an ideal VNA, as shown
198 Andrea Ferrero
am1 a1
Error
Ideal
box DUT
VNA
E
bm1 b1
Fig. 8.2 One-port error box.
in Figure 8.2; thus the usual form of the error coefficient becomes:
% & " #% &
bm1 e11 e12 am1
= . (8.12)
a1 e21 e22 b1
' () *
E
It’s straighforward to obtain the eij from the dij as:

% &% & % &% &
−d32 0 bm1 d −1 am1
= 31
−d42 1 a1 d41 0 b1
⇓
" # % &−1 % &
e11 e12 −d32 0 d31 −1
=
e21 e22 −d42 1 d41 0
% &
1 −d31 1
= . (8.13)
d32 d41 d32 − d31 d42 d42
However the elements of E do not behave in any way as a scattering matrix, i.e.
they do not have any particular properties of physical networks, but they are only four
complex numbers for each frequency, which include all the systematic, i.e. time-invariant,
characteristics of the whole system. Since no assumption has been made about the nature
of the error terms, it follows that the calibrated values of the a and b waves are a function
of how the error coefficients are computed, or:
the reference impedance of a VNA is set by the calibration and not by the hardware
The VNA readings have no physical meaning until the calibration is performed. To
calibrate a VNA means to determine the required set of error coefficients which define
the error model.
8.3 One-port calibration
Let’s proceed toward the complete solution in the elementary case of one port; from
(8.12) we have:

) *' (
bm1 e11 − (e11 e22 − e12 e21 )
= m = , (8.14)
am1 1 − e22
where = ba11 is the desired reflection coefficient while m is defined as the measured
one. From (8.14) we can see that only three error coefficients are needed to compute the
corrected reflection coefficient: e11 , e22 , and , and the de-embedding equation, i.e. the
formula which gives the corrected value from the measured one, follows from (8.14) as:
m − e11
= . (8.15)
e22 m −
To solve the calibration problem it’s useful to write (8.15) as:
e22 m − = m − e11
⇓ (8.16)
e11 + e22 m − = m .
This equation shows a simple and effective way to compute the error terms by mea-
suring three different standards i and stacking the corresponding equations as in (8.16)
for each mi measurement to form the linear system:
⎡ ⎤⎡ ⎤ ⎡ ⎤
1 1 m1 −1 e11 m1
⎣1 2 m2 −2 ⎦ ⎣e22 ⎦ = ⎣m2 ⎦ . (8.17)
1 3 m3 −3 m3
The first one-port technique was called SOL because the three standards were a Short,
an Open, and a Load [5]. The difficulties of making precise microwave standards were
immediately obvious, especially for the open and the load ones; furthermore the fre-
quency behaviour of these devices was not ideal nor constant, so a set of electrical
models were developed and included in the VNA firmware to describe the response vs.
frequency of the standards. Figure 8.3 shows the adopted standard models. They are
simple networks where a parameter is obtained as a polynomial fitting of the frequency
response. These models became a de facto standard and every standard manufacturer
publishes the parameters of its devices in this way.
Zc
Open Delay jBC = 2π f (C0 + C1f + C2f 2 + C3f 3)
loss R
Zc jXL = 2π f (L0 + L1f + L2f 2 + L3f 3)

Short Delay
loss R
Zc L
Load Delay
loss R
Fig. 8.3 One-port standard models.

200 Andrea Ferrero
Since all the measurements are functions of the calibration standards, to obtain their
parameters with high accuracy is a must and cannot be done through experiments. For a
coaxial environment the parameters were, in the past, obtained from a scale model of the
most critical one (the open) measured at low frequency. Nowadays FEM simulators are
used, which poses the metrological question of how accurate they are. Among the three
standards, the easiest one to manufacture is the short one; thus a simple solution takes
three offset shorts as the three standards. This technique is mathematically identical to
the SOL, but the three standards are of the same type and have only a different delay.
The problem arises when the line lengths resonate, thus the linear system becomes
undeterminate. For this reason the offset short technique is narrowband and has its main
application in waveguides. To complete the one-port case, it’s worth noting that if we
make the following assumptions:
• the dual directional coupler is well matched, balanced, but has finite isolation, i.e. its
S-matrix becomes:
⎡ ⎤
0 α β γ
⎢α 0 γ β⎥
S=⎢
⎣β
⎥.
γ 0 α⎦
γ β α 0
• the two mixers are identical and perfectly matched, i.e.:
ξ = ξ1 = ξ2 = 1
3 = 4 = 0.
From (8.10) and (8.11) it follows:
⎡ ⎤
−β γ
1 ⎢ αγ −αβ ⎥
D= 2 ⎢ ⎥ (8.18)
γ − β2 ⎣ γ −β ⎦
−αβ αγ
⇓
γ −αγ
e11 = e22 = =α (8.19)
β β
⇓
βm − γ
= . (8.20)
−αγ m − βα
Thus in this particular case, e11 represents the directivity of the directional coupler. In
the ideal case of infinite isolation and no insertion loss we have:
e11 = 0 e22 = 0 =1 (8.21)

Finally let’s consider (8.14). If = 0, i.e. we are connecting a perfect load, we have:
m = e11 . Thus the direct measurement of a standard gives the directivity† . Unfortunately
an ideal matched load does not exist, but only an approximation ( ≈ −40dB) can be
manufactured, which means limiting the VNA accuracy to that level. This was the reason
to develop the so called sliding load calibration [6]. This calibration uses a sliding load,
i.e. a transmission line with a load that can slide along it. From (8.14) note that if
≈ 0 ⇒ m ≈ e11 − , and by measuring this device for several load positions, i.e. for
different phases, we obtain a small circle on the complex plane whose center is e11 . This
techinique coupled with a short and an open was widely used on one-port VNAs.
8.4 Two-port VNA error model
Two-portVNAs generally have two different architectures which are modeled by different
error models:
• a reflectometer on each port, as shown in Figure 8.4,
• one reference coupler and a single coupler on each port, as shown in Figure 8.5.
The first case has more complex hardware, but, as demostrated above, does not suf-
fer from the switch imperfection or repeatability due to the independence of the error
terms from the source termination. The latter fewer components, but it requires a highly
repeatable switch and its error model does not allow the use of more modern calibration
techniques.
am1 bm1 am2 bm2
IF Digitizer
a1 D a2
U
b1 T b2
Fig. 8.4 Two-port VNA with a complete reflectometer on each port.

† However, this is true only if the S matrix of the microwave part was referred to the same impedance as the
load standard
202 Andrea Ferrero
bm1 bm2
a1 D a2
U
b1 T b2
amR
Fig. 8.5 Two-port VNA with a single reference channel.
8.4.1 Eight-term error model

The dual reflectometers VNA of Figure 8.4 is the generalization of the one port. Its error
model is straightforward to obtain with the introduction of a second error box as shown
in Figure 8.6. Here the four measured quantities am1 , bm1 , am2 and bm2 are linked to the
corresponding waves as:
% & " A #
A % &
bm1 e11 e12 am1
= (8.22)
a1 eA eA b1
' 21 () 22 *
EA
% & " B B %
# &
bm2 e11 e12 am2
= . (8.23)
a2 B
e21 B
e22 b2
' () *
EB
It’s more convenient to use a cascade matrix representation, where:

% & "A #
A % &
bm1 t11 t12 b1
= (8.24)
am1 t A t A a1
' 21 () 22 *
TA
% & " DU T #
DU T % &
b1 t11 t12 a2
= (8.25)
a1 DU T
t21 DU T
t22 b2
' () *
TDUT
am1 a1 a2 am2
DUT
EA S EB
bm1 b1 b2 bm2
IDEA LVNA
Fig. 8.6 Error model for 2-port VNA with two reflectometers.
" B
#% & %
B &
t11 t12
a2 a
= m2 (8.26)
t B t B b2 bm2
' 21 () 22 *
TB
and the relationships among the parameters become:

" DU T DU T
# % DU T &
S11 S12 1 t12 DU T
= DU T T (8.27)
DU T
S21 DU T
S22 t22 1 −t21
DU T
"A A
# % &
t11 t12 1 −A A
e11
= A E (8.28)
A
t21 A
t22 e21 −e22
A 1
' () * ' () *
TA XA
" A A
# % &
e11 e12 1 A
t12 A
= T (8.29)
A
e21 A
e22 A
t22 1 −t21
A
' () *
EA
"B B
# % &
t11 t12 1 1 −e22B
= (8.30)
B
t21 B
t22 B eB
e21 −B
' () * ' 11 () E *
TB XB
" B B
# % &
e11 e12 1 B
t21 B
= T . (8.31)
B
e21 B
e22 B
t11 1 −t12
B
' () *
EB
The fundamental calibration equation, i.e. the relationship among the measured and
desired quantities, can be now written in terms of the T matrix† as:
% & % &
bm1 a
= TA TDUT TB −1 m2
am1 ' () * bm2
TM
⇓
TM = TA TDUT TB −1 (8.32)
† For historical reasons we adopt the convention where the T matrix of port 2 is used inverted.
204 Andrea Ferrero
To obtain TM , two sets of different measurements are required, which are normally given
by switching the source between port one and port two. By combining the eight readings
we have: % & % &
bm1 bm1 am2 am2
= TM
a a b b
' m1 () m1 * ' m2 () m2 *
M1 M2 (8.33)
⇓
TM = M1 M1 −1
where am1 ,b ,a ,b
m1 m2 m2 are with the source at port 1 (forward measurements) while
are with the source at port 2 (reversed measurements). (8.32) can also
am1 ,bm1 ,am2 ,bm2
be written as:
TM = αXA TDUT XB −1 . (8.34)
B
e21
where α = A
e21
and
1 −1
TDUT = XA TM XB . (8.35)
α
which shows that the number of error coefficients required to obtain the corrected S
matrix in the two-port case is seven and not eight. However the common name for this
model is the eight-term error model [3].
8.4.2 Forward reverse error model

The second possibility for a two-port VNA is to adopt only one reference coupler and
a single coupler on each port, as shown in Figure 8.5. Here the error model discussed
above must be changed by assuming that each port has two different states, as shown in
Figures 8.7 and 8.8.
1. When the source is connected to the port, the reference and the test coupler form a
complete reflectometer and the model is identical to the one discussed above, but here
the slightly different notation of (8.36) and (8.37) is introduced.
2. However, when the port is not connected to the source, it is loaded with the internal
switch termination and a different model must be adopted where the linearity implies
that the two waves will be linked with the single reading + bmi as in (8.38) and (8.39).
amR bmi
ai
ai = −hi amR + li bmi (8.36)
bi
bi = −mi amR + ki bmi (8.37)
Fig. 8.7 Two-state error model with the port i connected to the reference channel.
bmi
ai
ai = gi +
bmi (8.38)
bi
bi = f i +
bmi (8.39)
Fig. 8.8 Two-state error model with the port i teminated
To derive a single equation for calibration and de-embedding, let’s note that the total
possible measurements obtainable by switching the source on the two ports are six, and
let’s organize them as follows:
% & % & % &
, a 0 , bm1 0 + 0 +
bm1
Am = R B = B = + . (8.40)
0 aR
m m
0 bm2 bm2 0
At the same time let’s write the eight DOT waves, two sets of four waves for each
source position, as:
% & % & % &
a1 a1 , + a1 0 0 a1
A= = A+A = + (8.41)
a2 a2 0 a a 0
' () 2 * ' 2 () *
,
A +
A
% & % & % &
b1 b1 b10 0 b1
B= =,
B ++
B= + (8.42)
b2 b2 0 b b 0
' () 2 * ' 2 () *
,
B +
B
and the error coefficients as well as:
% & % &
h1 0 k1 0
H= , K= ,
0 h2 0 k2
% & % &
l 0 m1 0
L= 1 , M= , (8.43)
0 l2 0 m2
% & % &
f 0 g 0
F= 1 , G= 1 .
0 f2 0 g2
From (8.36)–(8.43) we obtain the calibration equation (8.46) as:
,
A = L,Bm − H ,Am
,
B = K,Bm − M ,Am ,
+ (8.44)
A = G+Bm
+B = F+
Bm ,
A=, A++ A = L,Bm − H ,Am + G +
Bm
, +
B = B + B = K Bm − M Am + F +
, , Bm , (8.45)
S = BA−1 ,
206 Andrea Ferrero
and finally
Bm + F+
−SG+ Bm − SL,
Bm + K,
Bm + SH,
Am − M,
Am = 0. (8.46)
Equation (8.46) is also valid for multiport VNAs as explained in the following chapter;
however here we write it in scalar form for 2-port VNAs [7]. In the forward case, i.e.
with the source at port 1, it gives:
-
−S12 g2+ + (k − S l )b + (S h − m )a = 0
bm2 1 11 1 m1 11 1 1 R
(8.47)
(f2 − S22 g2 )+ − S l b + S h a = 0
bm2 21 1 m1 21 1 R
which can be normalized by k1 and if we define:

bm1
Sm11 = aR
+
(8.48)
bm2
Sm21 = aR
as a measured S-parameter we obtain:

⎧
⎨−S12 gk2 Sm21 + (1 − S11 kl1 )Sm11 + (S11 hk 1 − mk 1 ) = 0
1 1 1 1
(8.49)
⎩ ( fk12 − S22 gk12 )Sm21 − S21 kl11 Sm11 + S21 hk11 = 0.
By doing the same procedure for the reverse case, i.e. with the source on port 2, we
obtain: -
−S21 g1+ + (k − S l )b + (S h − m )a = 0
bm1 2 22 2 m2 22 2 2 R
(8.50)
(f1 − S11 g1 )+ − S l b + S h a = 0
bm1 12 2 m2 12 2 R
which can be normalized by k2 and if we define:

bm2
Sm22 = aR
+
(8.51)
bm1
Sm12 = aR
as a measured S-parameter we obtain:

-
−S21 gk21 Sm12 + (1 − S22 kl22 )Sm22 + (S22 hk22 − mk22 ) = 0
(8.52)
( fk21 − S11 gk21 )Sm12 − S12 kl22 Sm22 + S12 hk22 = 0.
Equations (8.49) and (8.52) form a linear system in the four unknown DUT S-parameters
that can be easily solved once the error coefficients have been computed. Note that the
same system can be applied to calibrate the VNA by using a set of proper standards, i.e.
by knowing Sij , and by solving it for the error coefficients. There are ten unknowns so this
model is known as the ten-term error model. Apparently, this model should have eleven
unknowns (six error coefficient times 2 ports minus 1 for the normalization); however
the two-port case is the degenerative one and the coefficients reduce to ten, leaving
Table 8.1 Equivalence between different error coefficients

notations
VNA This Book
EdF m1 /k1
EsF l1 /k1
ErF (m1 /k1 )(l1 /k1 ) − h1 /k1
EtF f2 /k1
g2 /k1
ElF
f2 /k1
EdR m2 /k2
EsR l2 /k2
ErR (m2 /k2 )(l2 /k2 ) − h2 /k2
EtR f1 /k2
g1 /k2
ElR
f1 /k2
the forward and reverse case equations completely independent. With the formulation
here presented the eight-term model can be considered as a subcase of the more general
one, but the link between the two is possible only for those VNA architectures with
a reflectometer on each port. In this case the eight-term model and the ten-term one
are both applicable and interchangeable. However, when there is only one reference
coupler this is not possible and only the ten-term model can be used. Table 8.1 reports
the formulas for the conversion of the error coefficients notation introduced here with
the more common one which is typically included in the VNA firmware.
8.5 Calibration procedures
The two error models can be identified by means of particular procedures which require
the measurement of known or even partially known standards. There are many techniques
which differ by the kind of standards and the math adopted. Here the more common ones
are presented and in particular those called:
• Thru-Short-Delay,
• Thru-Reflect-Line,
• Short-Open-Load-Reciprocal,
• Line-Reflect-Match,
• Short-Open-Load-Thru.
All the above methods but the last are usable ONLY with the eight-term error model,
while the SOLT is usable also on the forward/reverse model. This is due to the need for the
knowledge of eight readings during the two-port standard measurements, while the SOLT
is the only one which does not have this requirement and where the six measurements
of the forward/reverse model are enough.
208 Andrea Ferrero
8.5.1 TSD/TRL procedure

This technique uses a Thru which is typically either a direct connection between the ports
or a short straight line, a longer Line and a known or even unknown reflection standard.
Let’s consider the two-port calibration equation as (8.32) and let’s write it for the Thru,
TmT and Line TmL measurements:
TmT = TA TT T−1
B (8.53)
TmL = TA TL T−1
B , (8.54)
where TT and TL are the transmission matrix of a fully known standard Thru and of a
Line. Let’s compute:
Rm = TmL T−1 −1 −1 −1
mT = TA TL TB (TA TT TB ) =
= TA TL T−1 −1 −1
T TA = TA m TA (8.55)
Rn = T−1 −1 −1 −1
mT TmL = TB TT TA TA TL TB =
= TB T−1 −1 −1
T TL TB = TB n TB . (8.56)
If the thru and the line have the same characteristic impedance and their transmission
matrices are referenced to an impedance equal to the characteristic one, then:
% &
eγ (lL −lT ) 0
m = n = = TL T−1 −1
T = TT TL = . (8.57)
0 e−γ (lL −lT )
where lL and lT are the line and thru electrical lenghts and γ is the propagation con-
stant. (8.55) and (8.56) are eigenvalue equations where is the eigenvalues matrix. The
corresponding eigenvector matrices are the desired error coefficient ones:
" #
k
pa b
TA = p k = pXA (8.58)
p 1
% u &
1
TB = w w = wXB . (8.59)
f wu g
From the solution of the eigenvalues/vector problem f, g, a, and b are known, but pk ,
p
w and α = w (see (8.34)) are still unknown. Let’s first consider the measurement of a
u
fully known reflective standard SA , as an ideal short at port 1. From (8.14), (8.28), and
(8.58), we have:
A − A A
e11 b + pk aSA
A
mS = E S
= (8.60)
1 − e22
A A
S 1 + pk SA
⇓
k b − mS
A
= a (8.61)
p S (mS − 1)
A A
k
Once p is known, XA in known. From the thru measurement (8.53), we have:
−1
TmT = αXA TT XB (8.62)
⇓
−1
Y = αXB = T−1 −1
T XA TmT (8.63)
Given Y, the calibration problem is solved since the de-embedding equation (8.35) can
be now written as:
−1
TDU T = XA TM Y−1 . (8.64)
Since the used standards are: a Thru, a Short and a Delay line the TSD acronym was
used. The evolution of this technique was the so-called Thru-Reflect-Line which assumes
that the reflective standard X is not known, but is measured at the two ports, so:
A − A
e11 b + pk aX
E X
A
mX = = (8.65)
1 − e22
A
X 1 + pk X
B − B
e11 E X f + wu gX
B
mX = = . (8.66)
1 − e22
B
X 1 + wu X
Let’s compute the symbolic form of the thru measurement matrix:

" #
e 2γ lT ag k u − bf e 2γ lT a k u − b
−1 α
TmT = αXA TT XB = γl u pw pw
. (8.67)
e T w (g − f ) e2γ lT g pk wu − f e2γ lT pk wu − 1
If we consider the element TmT (1, 2)/TmT (2, 2) we obtain:
TmT (1, 2) e T a p w − b
2γ l k u
SmT (1, 1) = = (8.68)
TmT (2, 2) e2γ lT pk wu − 1
and solving the nonlinear system formed by (8.65), (8.66), and (8.68), X is obtained as:

(f − mX
B )(b − A )(a − S
mT (1, 1))
X = ±e γ lT mX
. (8.69)
(g − mX
B )(a − A )(b − S
mX mT (1, 1))
k u
Once X is known the procedure either follows the TSD algorithm or p and w are
given by (8.65) and (8.66), while α can be computed from TmT (2, 2) as:
e2γ lT k u
−1
w (g − f )
u
pw eγ lT
TmT (2, 2) = α ⇒ α = TmT (2, 2) . (8.70)
w (g − f )
u
eγ lT e2γ lT pk wu − 1
The sign of X must be known and typically is given by a rough knowledge of the
reflection type (a short or an open). Finally note that X cannot be a match load, i.e
210 Andrea Ferrero
X = 0, because we would simply obtain mX A = b and B = f which are already

mX
known.
The main characteristics of the TRL calibration are:
1. The propagation constant of the line is obtained as a by-product of the calibration.

2. The characteristic impedance of the line sets the reference impedance of the VNA.
3. If a zero length thru, i.e. a direct port connection, is used, only the line must be known.
4. The length difference between the line and the thru does not have to be a multiple of
the wavelength.
The S21 parameter of the line is given by the solution of the eigenvalues problem, while
the propagation constant is obtainable if the length is known without the calibration.
For this reason the line can be a partially known standard. This property of TRL was
successfully used to characterize different structures [8]. However the characteristic
impedance of the line automatically becomes the reference impedance of the VNA since
the diagonal property of the eigenvalues matrix is obtainable only by assuming that the
reference impedance is equal to the characteristic one. Furthermore its value MUST be
known a priori and not from the measurement. The TRL procedure is the only one which
sets the reference impedance based on a distributed component, the LINE, while all the
other calibration methodologies use a lumped component to set the reference impedance.
Thus it’s questionable if waves are really measured using calibration procedures other
than TRL [9] and all the national metrology labs in the world use a set of lines as their
primary microwave coaxial and waveguide standards. If the THRU has zero length,
i.e. a unitary transmission matrix, the characteristic impedance of the line remains the
ONLY parameter required to obtain a successful calibration and this property makes
TRL easily traceable to the mechanical dimension of the standard. The main drawback
of the TRL technique is the relatively small bandwidth because of the line resonance. At
the frequency where eγ (lL −lT ) = 1, the eigenvalue matrix becomes unitary and obviously
the problem becomes undeterminate. The calibraton fails and typically, a glitch appears
in the measurement date. To avoid this problem, for a broadband calibration several lines
are mandatory and a Multiline TRL is used [10].
8.5.2 SOLR procedure

The Short-Open-Load-Reciprocal calibration was introduced in 1992 and it avoids the
use of a fully known two-port device, generally the THRU, which is common in all
the other techniques[4]. If a linear system based on (8.65) and (8.66) is formed with
three different standards on each port, as done in the one-port case, the matrices XA
and XB are easily obtained, i.e six out of seven error coefficients are given by one-port
measurements. Let’s now take the measurement of a reciprocal device TmR , i.e with
S12 = S21 , from (8.34):
−1
TmR = αXA TR XB . (8.71)
If the determinant on both sides on (8.71) is taken and by noting that the reciprocity
condition implies det (TR ) = 1, the last term α is easily obtained as:
=1
) *' (
det (X A ) det (T R ) det (XB ) det (TmR )
det (TmR ) = α 2 ⇒α=± . (8.72)
det (XB ) det (XA )
The characteristics of this technique are
1. Easily applicable with ordinary one-port standards.

2. Any reciprocal device can be used as THRU.
3. Different connection problems can be easily solved.
4. The one-port standards must be fully known.
The SOLR uses the same one-port standard set as the old SOLT and this means a
straighforward applicability to all the full reflectometer VNAs. However the freedom
from the THRU device, the main characteristic of this technique, allows a much easier
solution of the calibration problem in many situations. As an example, if the two ports
are far apart a fully known THRU may be difficult to obtain while a simple cable, used
as reciprocal, is a very easy replacement. Another typical example is on-wafer mea-
surements with right-angle probes, where a bended THRU is far from being an ideal
line while the reciprocity condition is easily achieved with non-giromagnetic struc-
tures. The main constraint given by the SOLR is the need for perfectly known one-port
standards.
8.5.3 LRM procedure

One of the more interesting two-port techniques is the Line-Reflect-Match. Here a
different notation is used instead of the original one [3], but this one is also useful
for the multiport case in the next paragraph. Let’s consider (8.36) and (8.37), in the full
reflectometer case:
a1 = −h1 am1 + l1 bm1 (8.73)

b1 = −m1 am1 + k1 bm1 (8.74)
a2 = −h2 am2 + l2 bm2 (8.75)
b2 = −m2 am2 + k2 bm2 . (8.76)
As done before for the ten-term model, we can organize the eight readings, four for each
source position, as:
% & % &
a
am1 b
bm1
Am = m1
Bm = m1
, (8.77)
am2 am2 bm2 bm2
212 Andrea Ferrero
the eight DUT waves as:

% & % &
a1 a1 b1 b1
A= B= (8.78)
a2 a2 b2 b2
and also the error coefficients as (8.43). From (8.77), (8.78), and (8.43) we obtain the
calibration equation (8.80) as:
A = L Bm − H Am
(8.79)
B = K Bm − M Am
−SLBm + KBm + SHAm − MAm = 0. (8.80)
This equation is the scattering version of (8.32). Let’s introduce the line measurement
matrices
%
& %
&
a aLm1 bLm1 bLm1
ALm = Lm1 B Lm = (8.81)
aLm2 aLm2 bLm2 bLm2
and the line S-parameter matrix
% &
SL11 SL12
SL = . (8.82)
SL11 SL22
Equation (8.80) becomes:
−SL LBLm + KBLm + SL HALm − MALm = 0, (8.83)
i.e. in scalar form:

% &% &%
&
SL11 SL12 l1 0 bLm1 bLm1
− +
SL11 SL22 0 l2 bLm2 bLm2
% &%
&
k 0 bLm1 bLm1
+ 1 +
0 k2 bLm2 bLm2
% &% &%
& (8.84)
SL11 SL12 h1 0 aLm1 aLm1
+ +
SL11 SL22 0 h2 aLm2 aLm2
% &%
& % &
m 0 aLm1 aLm1 0 0
− 1 = .
0 m2 aLm2 aLm2 0 0
Equation (8.84) provides four independent equations, one for each ij element. By doing
the same procedure for the load measurement matrices:
% & % &
a 0 b 0
Am = m1 Bm = m1 (8.85)
0 am2 0 bm2
and by remembering that the reflection coefficient for the load is null, we have:
% &% &% &

0 0 l1 0 bm1 0
− +
0 0 0 l2 0 bm2
% &% &
k 0 bm1 0
+ 1 +
0 k2 0 b
% &% & % m2

& (8.86)
0 0 h1 0 am1 0
+ +
0 0 0 h2 0 a
% &% & m2 % &
m 0 am1 0 0 0
− 1 = ,
0 m2 0 am2 0 0
i.e. other two independent equations are given by (8.86), because only the ii elements
are not null. These six equations form a linear calibration system as:
⎡ ⎤
m1
⎡ ⎤
−aLm1 0 −bLm1 SL11 −bLm2 SL12 aLm1 SL11 aLm2 SL12 bLm1 0 ⎢ m2 ⎥
⎢ ⎥
⎢−a 0 S
−bLm1 S
−bLm2 S
aLm1 S
aLm2
bLm1 0 ⎥ ⎢ l1 ⎥
⎢ Lm1
L11

L12

L11

L12

⎥ ⎢ ⎥
⎢ 0 −aLm2 −bLm1 SL21 −bLm2 SL22 aLm1 SL21 aLm2 SL22 0 bLm2 ⎥⎢ l2 ⎥
⎢ ⎥⎢ ⎥=0
⎢ 0
−aLm2 S
−bLm1 S
−bLm2 S
aLm1 S
aLm2 0 bLm2 ⎥
⎢ h1 ⎥
⎢ L21 L22 L21 L22 ⎥⎢ ⎥
⎣−a
m1 0 0 0 0 0
bm1 0 ⎦⎢⎢ h2 ⎥
⎥
0
−am2 0 0 0 0 0
bm2 ⎣ k1 ⎦
' () * k2
N ' () *
u
⇓
⎡ ⎤⎡ ⎤
−aLm1 0 −bLm1 SL11 −bLm2 SL12 aLm1 SL11 aLm2 SL12 m1
⎢−a 0 S
−bLm1 S
−bLm2 S
aLm1
aLm2 SL12 ⎥ ⎢m2 ⎥
⎢ Lm1
L11

L12

L11

⎥⎢ ⎥
⎢ 0 −aLm2 −bLm1 SL21 −bLm2 SL22 aLm1 SL21 aLm2 SL22 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ l1 ⎥ +
⎢ 0
⎢

−aLm2 S
−bLm1 L21
S
−bLm2 L22
S
aLm1 L21 aLm2 SL22 ⎥ ⎢
⎥ ⎥
⎢ l2 ⎥
⎣−a 0 0 0 0 0 ⎦ ⎣ h1⎦
m1
0
−am2 0 0 0 0 h2
' () * ' () *
, ,
u
⎡ N ⎤
bLm1 0
⎢b 0 ⎥
⎢ Lm1
⎥% &
⎢ 0 b ⎥
+⎢ Lm2 ⎥ k1 = 0
⎢ 0 bLm2 ⎥

⎢ ⎥ k2
⎣b
m1 0 ⎦ ' () *
+
u
0 b
' () m2 *
+
N
⇓
,u++
N, N+u=0
⇓
u = −,
, N−1 +u = W+
N+ u. (8.87)
The matrix W is fully known from the measurements and the definitions of the LINE
and MATCH standards. (8.87) defines a subset of the normalized error coefficients ,
u as
214 Andrea Ferrero
k2
a linear combination of the ratio k1 as:
m1
k1 = w11 + w12 kk21
m2
k1 = w21 + w22 kk21
l1
k1 = w31 + w32 kk21
(8.88)
l2
k1 = w41 + w42 kk21
h1
k1 = w51 + w52 kk21
h2
k1 = w61 + w62 kk21 .
The unknown reflection measured at the two ports gives:

⎡ ⎤
m1
⎢m ⎥
%
&⎢
⎢ ⎥
2⎥
−aRm1 0 −bRm1 R 0 aRm1 R 0 ⎢ l1 ⎥
⎢ ⎥+
0 −aRm2 0 −bRm2 R 0 aRm2 R ⎢ l2 ⎥
' () *⎢ ⎥
⎣ h1 ⎦
,
NR
h2
' () *
% &% & ,
u
b 0 k1
+ Rm1 = 0.
0 bRm2 k2
' () * '()*
+
NR +
u
(8.89)
bRm1 bRm2
Let Rm1 = aRm1 and Rm2 = aRm2 ; from (8.87) and (8.89) we have:
, u++
NR W+ NR+u=0
⇓ (8.90)
NR W + +
(, u = 0,
NR )+
which has a non-null solution only if the determinant of
NR W + +
, NR =
% &
−w31 Rm1 R + w51 − w11 + Rm1 −w32 Rm1 R + w52 − w12
(8.91)
−w41 Rm2 R + w61 − w21 −w42 Rm2 R + w62 − w22 + Rm2
is null, i.e:
(−w31 Rm1 R + w51 − w11 + Rm1 )(−w42 Rm2 R + w62 − w22 + Rm2 )
(8.92)
−(−w41 Rm2 R + w61 − w21 )(−w32 Rm1 R + w52 − w12 ) = 0.
Equation (8.92) is a second-order equation in R which can be easily solved. Once R

is known the ratio kk21 is easily computed from (8.90) and all the other error coefficients
from (8.88). The main charateristics of LRM are:
• Broadband performances.
• Suitable for on-wafer measurement where the use of a load does not require a probe
shifting.
• High accuracy as the TRL, if good broadband loads are used.
• The VNA reference impedance is set by the load.
8.5.4 SOLT procedure

The Short-Open-Load-Thru is the oldest, most used and widely adopted calibration tech-
nique. This is due to the applicability of the technique to both a complete or a partial
reflectometer VNA architecture, as shown in the following. The SOLT was first imple-
mented inside the VNA firmware due to its simplicity and applicability on both models;
furthermore the commercial availability of coaxial standards made it easy to manage.
Based on the mathematics presented in the previous paragraphs, we use (8.80) to
write the corresponding calibration system for the seven measured standards (Opens:
O1 , O2 ; Shorts: S1 , S2 ; Loads: L1 = L2 = 0 and Thru ST ) as [11]:
⎡ ⎤
−aT m1 0 −bT m1 ST 11 −bT m2 ST 12 aT m1 ST 11 aT m2 ST 12 bT m1 0
⎢ ⎥
⎢ −aT m1 0 −bT m1 ST 11 −bT m2 ST 12 aT m1 ST 11 aT m2 ST 12 bT m1 0 ⎥
⎢ ⎥
⎢ 0 −aT m2 −bT m1 ST 21 −bT m2 ST 22 aT m1 ST 21 aT m2 ST 22 bT m2 ⎥
⎢ 0 ⎥
⎢ ⎥
⎢ 0 −aT m2 −bT m1 ST 21 −bT m2 ST 22 aT m1 ST 21 aT m2 ST 22 bT m2 ⎥
⎢ 0 ⎥
⎢ ⎥
⎢ −a 0 0 0 0 0
bLm1 0 ⎥
⎢ ⎥
N = ⎢ Lm1 ⎥
⎢ 0 −aLm2 0 0 0 0 0 bLm2 ⎥
⎢ ⎥
⎢ ⎥
⎢ −aSm1 0 −bSm1 S1 0 aSm1 S1 0 bSm1 0 ⎥
⎢ ⎥
⎢ ⎥
⎢ 0 −aSm2 0 −bSm2 S2 0 aSm2 S2 bSm2 0 ⎥
⎢ ⎥
⎢−a
−bOm1 0 ⎥
⎣ Om1 0 S1 0 aOm1 O1 0 bOm1 ⎦
0
−aOm2 0
−bOm2 O2 0
aOm2 O2
bOm2 0
T
u = m1 m2 l1 l2 h1 h2 k1 k2
Nu = 0. (8.93)
The linear system (8.93) is overdetermined in this case† and it can be either solved
in the least squared sense or better, three equations are non-necessary. This case was
called QSOLT (Quick SOLT) because it may avoid the use of one-port standards at
port 2 [12, 13]. However for the forward/reverse model, since we have more unknowns,
all the standard measurements are required. Let’s take (8.49) and specialize it for the
corresponding standard measurements with the source at port 1, i.e. the forward case, as:
⎡ ⎤⎡ ⎤ ⎡ ⎤
1 ST 11 ST m11 −ST 11 0 ST 12 Sm21 m1 /k1 ST m11
⎢0 S S ⎥⎢ ⎥ ⎢ 0 ⎥
⎢ T 21 T m11 −ST 21 −ST m21 ST 22 ST m21 ⎥ ⎢ l1 /k1 ⎥ ⎢ ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢1 0 0 0 0 ⎥ ⎢ h1 /k1 ⎥ = ⎢ SLm1 ⎥ (8.94)
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣1 O1 SOm11 −O1 0 0 ⎦ ⎣ f2 /k1 ⎦ ⎣ SOm1 ⎦
1 S1 SSm11 −S1 0 0 g2 /k1 SSm1
† We have 7 unknowns and ten equations.

216 Andrea Ferrero
While in the reverse case, i.e source at port 2, we use (8.52) and obtain:
⎡ ⎤⎡ ⎤ ⎡ ⎤
1 ST 22 ST m22 −ST 22 0 ST 21 Sm12 m2 /k2 ST m22
⎢0 S S −ST 12 −ST m12 ST 11 ST m12 ⎥ ⎢ ⎥ ⎢ ⎥
⎢ T 12 T m22 ⎥ ⎢ l2 /k2 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢1 0 0 0 0 ⎥ ⎢ h2 /k2 ⎥ = ⎢ SLm2 ⎥ . (8.95)
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣1 O2 SOm22 −O2 0 0 ⎦ ⎣ f1 /k2 ⎦ ⎣ SOm2 ⎦
1 S2 SSm22 −S2 0 0 g1 /k2 SSm2
The solution of (8.94) and (8.95) directly gives the error coefficients.
8.6 Recent developments
The VNA error models here presented were all based on the Non Leakage hypothesis,
i.e it’s assumed that port 1 and port 2 are isolated and no signal appears on the other side
channel unless a 2-port DUT is connected. This assumption limits the applicability of
the model where the crosstalk signals are significantly lower than the DUT ones, which
is the typical case with coaxial or waveguide devices measurements. However for high
attenuation testing or for on-wafer critical applications the crosstalk cannot be negleted.
The leakage error models have been introduced since the 1970s, but a successful standard
sequence was only invented in the 1990s [14], [15]. Following the linearity principle it’s
easy to develop a leakage model by considering the eight-port network formed by the
two sides of the VNA, as shown in Figure 8.9. As done before there is a C4 =⇒ C4 linear
am1 bm1 am2 bm2
IF Digitizer
Leakage block
a1 D a2
U
b1 T b2
Fig. 8.9 2-port VNA with crosstalk part shaded.

application between the measured and DUT quantities as:

⎡ ⎤ ⎡ ⎤⎡ ⎤
b1 d11 d12 d13 d14 am1
⎢a1 ⎥ ⎢d21 d22 d23 d24 ⎥ ⎢bm1 ⎥
⎢ ⎥=⎢ ⎥⎢ ⎥ (8.96)
⎣b2 ⎦ ⎣d31 d32 d33 d34 ⎦ ⎣am2 ⎦ .
a2 d41 d42 d43 d44 bm2
Equation (8.96) contains 16 error terms and a particular solution of the identification
problem was given in [14]. Although the leakage calibration has been formally and
experimentally tested, the leakage terms are often position-dependent, as in the on-
wafer environment where the probes distance dramatically affects the crosstalk. In this
case, the correction given by the 16-term calibration may fail because the error terms
may change after the calibration. Finally, the author wishes to point out that the simple
solution of the leakage problem that can be found in many VNA firmware, which adds
two error terms to the ten-term model, may lead to incorrect measurements.
Since the two-port VNAs are the most widely used and the calibration is a must to
obtain reliable measurements, in the 1990s automatic calibrator devices were introduced
and called Electronic Calibrators. They are typically PIN diode-based networks that con-
tain different loads and are precharacterized with a metrological grade VNA, calibrated
with TRL. The broadband measurement of the electronic calibrator for all the possible
states is stored in a file shipped with each unit and used by the VNA firmware to solve the
calibration, typically with an SOLT or SOLR algorithm. Since these devices substitute
the traditional mechanical standards they are called Transfer Standards because their
electrical behavior is measured and not computed from mechanical dimensions and EM
theory[16]. By using an Electronic Calibrator the VNA calibration is greatly simplified
to a single connection, but it’s always better to verify the obtained accuracy by measuring
at least one traditional mechanical device on both ports.
8.7 Conclusion
The calibration of two-port VNAs has greatly enhanced the measurement accuracy of
microwave devices. During the last 40 years many different algorithms have been pro-
posed, but the error models and the techniques shown here are now established and
implemented in the majority of modern VNAs.
References
[1] B. Donecker, Determining the measurement accuracy of the HP8510 microwave network
analyzer. Santa Rosa, CA: HP, 1985.
[2] R. A. Franzen, N. R. Speciale, “New procedure for system calibration and error removal in
automated s-parameter measurements,” in 5th European Microwave Conference, Sept. 1975.
[3] H. Eul and B. Schieck, “A generalized theory and new calibration procedures for network
analyzer self-calibration,” IEEE Trans. Microw. Theory Tech., vol. MTT-39, pp. 724–731,
Apr. 1991.
218 Andrea Ferrero
[4] A. Ferrero and U. Pisani, “Two-port network analyzer calibration using an unknown thru,”
IEEE Microw. Guid. Wave Lett., vol. 2, pp. 505–507, Dec. 1992.
[5] R. A. Hackborn, “An automatic network analyzer system,” Microwave J., vol. 11, pp. 45–52,
May 1968.
[6] I. Kasa, “A circle fitting procedure and its error analysis,” IEEE Trans. Instrum. Meas., vol.
IM-25, p. 8, Mar. 176.
[7] A. Ferrero, V. Teppati, M. Garelli, and A. Neri, “A novel calibration algorithm for a special
class of multiport vector network analyzers,” IEEE Trans. Microw. Theory Tech., vol. 56,
no. 3, pp. 693–699, Mar. 2008.
[8] D. Williams and R. Marks, “Accurate transmission line characterization,” IEEE Microw. and
Guid. Wave Lett., vol. 3, pp. 247–249, Aug. 1993.
[9] R. Marks and D. Williams, “A general waveguide circuit theory,” J. Res. NIST, vol. 97,
pp. 533–561, Sept. 1992.
[10] R. Marks, “A multiline method of network analyzer calibration,” IEEE Trans. Microw. Theory
Tech., vol. 39, no. 7, pp. 1205–1215, July 1991.
[11] K. Silvonen, “A general approach to network analyzer calibration,” IEEE Trans. Microw.
Theory Tech., vol. 40, no. 4, pp. 754–759, Apr. 1992.
[12] A. Ferrero and U. Pisani, “Qsolt: a new fast calibration algorithm for two-port S-parameter
measurements,” in 38th ARFTG Conf. Dig., San Diego, CA, Dec. 1991, pp. 15–24.
[13] H. Eul and B. Schieck, “Reducing the number of calibration standards for network analyzer
calibration,” IEEE Trans. Instrum. Meas., vol. IM-40, pp. 732–735, Aug. 1991.
[14] J. V. Butler et al., “16-term error model and calibration procedure for on-wafer network
analysis measurements,” IEEE Trans. Microw. Theory Tech., vol. 39, no. 12, pp. 2211–2217,
Dec. 1991.
[15] K. Silvonen, “A 16-term error model based on linear equations of voltage and current
variables,” IEEE Trans. Microw. Theory Tech., vol. 54, no. 4, pp. 1464–1469, June 2006.
[16] V. Adamian, “Simplified automatic calibration of a vector network analyzer,” in ARFTG
Conference, Nov. 1994, pp. 1–9.
9 Multiport and differential
S-parameter measurements
Valeria Teppati and Andrea Ferrero
9.1 Introduction
The last ten years have witnessed an increasing interest in multiport S-parameter
measurements, i.e. S-parameter measurements of devices with more than two ports,
for two main reasons: the first one is the increasing complexity of modern microwave
devices and circuits and the use of more complex MMICs.
But the main reason is definitely the shift toward microwave frequencies of the personal
computer’s processors speed, which implies that such digital applications must now face
typical microwave challenges. These topics have recently been addressed in [1]. Preserv-
ing the signal integrity of a microwave signal through the packages, sockets, connectors,
and PCB traces, commonly found in today’s computer systems, is one of the main issues.
System architectures with hundreds of parallel channels, operating at higher and higher
data rates, involve microwave multiport measurements for the characterization, design,
and analysis of the structures and their effects on the signals. Microwave designers
and engineers are thus facing new challenges in multiport measurement hardware and
calibrations.
The first challenge comes from the typical media of digital interconnections: the PCB.
It can include both planar and three-dimensional (3-D) DUTs, as found, for example,
in memory modules. So, on one hand many data lines must be connected and measured
simultaneously, and they do not necessarily lie on a single plane. On the other hand,
these connections from the boards to the typically coaxial test ports of the VNA must
have good performances at microwaves, i.e. be “transparent” for the measurements.
If the structure is not three-dimensional but planar, the best choice for contacts is
connecting directly to the PCB surface with high-performance microwave probes. The
alternative for 3-D structures, when probing is not possible, is to use coaxial to PCB
launchers, but this solution is not the best in terms of insertion losses. Microwave
probes provide better high-frequency transitions to the boards, compared with coaxial
launchers, both in terms of connection repeatability and of electrical transparency of the
transition. Thus, probing typically offers improved calibration and better measurement
accuracy.
For multiport devices, as the number of ports increases, traditional single-port probes
are unsuitable, since it could be very difficult to mechanically put in place all the probes,
even if specific probe stations could be designed for the purpose. The typical solution
is to use multiport probes, such as the ones shown in Figure 9.1. With these probes it is
220 Valeria Teppati and Andrea Ferrero
(a) (b)
Fig. 9.1 Multiport GSG probe tips on the same probe head [3]. Courtesy of GGB Industries.
possible to measure various channels simultaneously, but the probe pitch and the patterns
on the board must be designed to match.
Probe manufacturers provide different types of multiport probes, depending on the
customer needs. In particular, one of the aspects to be taken into account for accu-
rate measurements is the presence of crosstalk between probe fingers. Two multiport
probe configurations are possible; the first one is obtained simply by tiling standard
ground-signal-ground (GSG) probes (GSGGSG), as shown in Figure 9.1. The second
configuration allows narrower patterns to be tested as it does not have ground fingers
between the various signal lines (GSSG). In this case, the crosstalk between two adjacent
fingers might not be negligible, and to achieve accurate calibrations and measurements
it should be taken into account. A possible solution to this problem is revised in this
chapter [2].
In any case, for both coaxial launchers and probes, the VNA calibration for on-board
testing can be performed following two approaches. Either the VNA test ports are cal-
ibrated, so that the transition performances are included as part of the DUT, and then
a separate de-embedding of the transition is performed. Or, to achieve more accurate
results, the reference planes are moved on-board through an on-board calibration [3–5].
In the next sections the problems with the measurement and calibration of multiport
systems are described in detail.
9.2 Multiport S-parameters measurement methods
Multiport measurements find the error-corrected S-parameters of a DUT having more

than two ports.
Multiport and differential S-parameter measurements 221
There are two approaches to this problem. The first one, available since the early 1980s,
consists of performing multiple (round robin) measurements, with two-port VNAs and
matched loads on the unused ports [6–9]. This method is still used nowadays, by taking
into account the non-idealities of the matched terminations [10, 11]. The procedure is
quite cumbersome, since it requires n(n−1)/2 different two-port measurements, for each
n-port measurement. Besides, the overall accuracy is affected by the multiple connections
required and by the accuracy of the terminations.
The alternative and more modern approach is the multiport VNA, i.e. a measurement
system able to perform straightforward calibrated multiport measurements, with a single
DUT connection. The calibration and the measurement problem are the two main aspects
to be considered when dealing with a multiport measurement architecture.
Various system architectures are currently available with different numbers of sources
and measurement receivers per port, from a maximum of one source and two receivers for
each port, as shown in Figure 9.2(a), to a minimum of a single source and two receivers
with a proper switch matrix; see Figure 9.1(b). Of course, the solution of Figure 9.2(a)
is very expensive, but fast, while the solution of Figure 9.1(b), at the cost of speed, can
be used to extend any two-port VNA to multiport.
A number of possible intermediate solutions lies between the two configurations
depicted in Figure 9.2, e.g. one source for each couple of ports, etc.
To more clearly delineate why certain choices in the architecture may be made, the
following constraints need to be taken into account [1].
(i) Extra sources and receivers are more expensive than extra couplers.
(ii) Cable and connector losses are very high due to the frequencies involved. It is
particularly important to minimize losses after the test couplers as this affects the
raw directivity.
(iii) Switch isolation becomes worse at higher frequencies.
(iv) At higher frequencies, single-pole double-throw (SPDT) switches perform much
better (in terms of isolation) than single-pole triple-throw (SP3T) or single-pole
quadruple-throw (SP4T) switches. But of course, substituting SP4T switches with
STDT ones complicates the switching matrix.
(v) For error model simplicity, it helps if the load match presented by a port is
independent of the driving port.
As the frequency rises, the number of sources, receivers, and directional couplers
should be reduced. A fairly good compromise is the partial reflectometer architecture
(see Figure 9.7), which is analyzed in Section 9.2.2.
9.2.1 Calibration of a complete reflectometer multiport VNA

We now present the generalized and simple error model formulation, that was provided
by [12], for a complete reflectometer VNA architecture, i.e. based on the use of two
directional couplers for each port. In this case, it is always possible to measure the
incident and the reflected waves at each port, wherever the source excitation is. The error
Microwave
sources
a1 a2 a3 a4
Mixer
b1 b2 b3 b4
Port 1 Port 2 Port 3 Port 4
(a)
Test receiver Microwave

(sampler or mixer) sources
Reference
receiver
a1 a2 a3 a4
b1 b2 b3 b4
(b)
Fig. 9.2 The most expensive (a) and minimum (b) multiport architecture, for a four-port example with
complete reflectometers.
model, in the most simple case of no leakage between ports, has 4n − 1 unknowns, where
n is the number of ports.
The incident and reflected waves at each port can be organized in the following
matrices:
⎡ ⎤ ⎡ ⎤
am11 am12 ··· am1n bm11 bm12 ··· bm1n
⎢am21
⎢ am22 ··· am2n ⎥
⎥
⎢bm21
⎢ bm22 ··· bm2n ⎥
⎥
Am = ⎢ . .. .. .. ⎥ , Bm = ⎢ . .. .. .. ⎥
⎣ .. . . . ⎦ ⎣ .. . . . ⎦
amn1 amn2 ··· amnn bmn1 bmn2 ··· bmnn
⎡ ⎤ ⎡ ⎤
a11 a12 ··· a1n b11 b12 ··· b1n
⎢a21
⎢ a22 ··· a2n ⎥
⎥
⎢b21
⎢ b22 ··· b2n ⎥
⎥
A=⎢ . .. .. .. ⎥ , B=⎢ . .. .. .. ⎥ ,
⎣ .. . . . ⎦ ⎣ .. . . . ⎦
an1 an2 ··· ann bn1 bn2 ··· bnn
where amij and bmij represent, respectively, the measured incident and reflected waves
at port i, when the source excitation is at port j , while aij and bij are the actual incident
and reflected waves at the port i reference plane, when the source excitation is at port j .
Calling S the scattering matrix of the multiport DUT, we can write
B = SA. (9.1)
The error model is defined by the following:
A = L B m − H Am
(9.2)
B = K Bm − M A m ,
where L, M, H, and K contain the error coefficients, and can be full matrices (in this
case the error model is full leaky), diagonal matrix (for a non-leaky model), or block
diagonal (for a partially leaky model [2]).
By combining (9.1) and (9.2), the equation for the error coefficient computation is
SLBm − SHAm − KBm + MAm = 0 (9.3)
thus for de-embedding we have:
S = (KBm − MAm ) (LBm − HAm )−1 . (9.4)
Note that (9.3) is written in terms of measured waves rather than measured
S-parameters, as in [13]. In other words, no switch correction technique has been applied
here to obtain the measured scattering matrix. The calibration equations are written
directly in terms of the measured quantities, with computational advantages.
It is useful to express (9.3) in the following iterative form (written here for simplicity
for a non-leaky error model),

n
n
Sip Lpp bmpj − Sip Hpp ampj − Kii bmij + Mii amij = 0
p=1 p=1
(9.5)
(i = 1, . . . , n)
(j = 1, . . . , n)
because in this form it can be easily used for one- or two-port standards, by simply
eliminating the proper rows.
For example, a one-port standard connected at port 1 (i = 1, j = 1) gives:
L11 m − H11 + M11 = K11 m , (9.6)

Ideal multiport vector network analyzer

P1 P3
P2 P4
1 3
2x2-leaky 2x2-leaky
error 2 DUT 4 error
model model
(a) (b)
Fig. 9.3 Four-port partially leaky error model (a) and coupled lines, measured with two GSGGSG probes
(b). © 2005 IEEE. Reprinted, with permission, from [2].
where m and are the measured and the defined reflection coefficient of the standard,
respectively.
A typical calibration consists of finding the error coefficients by solving a system in
the form (9.5), obtained by measuring a proper sequence of one- and two-port standards.
Since the coefficients are 4n − 1, there must be 4n − 1 linearly independent equations
in order to find all the unknowns. The rules to grant the independence of the multiport
calibration equations are given in [13].
Calibration of a partially leaky multiport VNA

For some measurement problems, the non-leaky model could be not accurate enough.
An example is the multi-finger probes, of Figure 9.1, where crosstalk between fingers
belonging to the same probe cannot be neglected. In these cases a more complete error
model is required, and is shown in Figure 9.3(a). The measurement system is split in two
halves, and the leakage error terms are present in each of the halves, but not between the
two parts of the model. This error model is mathematically described by (9.2), where L,
M, H, and K are block diagonal [2].
Considering the case of a four-port VNA, with two double signal probes, the partially
leaky model takes into account the leakage between the fingers on the same probe, while
the side-by-side crosstalk is neglected. It has been proved that, in such cases, this model
has better accuracy than both the non-leaky model and the fully leaky one. The crosstalk
terms of the half-leaky case are fixed and constant, due to the fixed position of each probe
finger, while the side-by-side crosstalk, which is usually minimal, variable, and difficult
to model, is neglected. This is better than trying a correction with wrong error terms, as
the full-leaky calibration would do [2].
An optimized standard sequence, especially useful for on-wafer measurements, is sim-
ilar to a classical LRM calibration, where the number of probe touchdowns is minimized:
only three on-wafer probe placements are required, as shown in Figure 9.4. The three
standard combinations are: 1) thru ports 13 and shorts ports 2 and 4; 2) thru ports 24 and
shorts ports 1 and 3; and 3) thru ports 1–4 and loads ports 2 and 3.
3
Match
1 3 1 3 1 4
Thru S S Thru
2 4 2 4 2
S S Thru Match
1) 2) 3)
Fig. 9.4 Calibration sequence for a partially leaky four-port VNA, optimized for on-wafer touchdowns. ©
2005 IEEE. Reprinted, with permission, from [2].
Figures 9.5 and 9.6 report measurements and simulations of loosely coupled coplanar
lines, similar to the ones depicted in Figure 9.3(b). All the simulations were performed
with a commercial simulator, implementing a simple circuital model. Figure 9.5(a) refers
to 0.58 mm long coupled lines. Scattering parameter S12 , i.e. the near-end coupling
between the two structures, is clearly overestimated by the non-leaky model, since it does
not include the correction for leakage between ports 1 and 2, while the proposed half-
leaky and full-leaky calibrations demonstrate a very good agreement with simulations.
But if we consider longer (6.6 mm) coupled lines, as shown in Figure 9.5(b), the 10
GHz resonance predicted by simulations is found only with the half-leaky model. The
effect of the wrong correction of the full-leaky model is evident since this device is much
longer than the thrus used during calibration. Also the non-leaky model does not provide
the right value of the resonance frequency.
Finally, in Figure 9.6, the far-end crosstalk of the same (6.6 mm) coupled lines is
shown. Also in this case, the half-leaky model agrees better with the simulation than the
other two models.
9.2.2 Calibration of a partial reflectometer multiport VNA

Available multiport VNA architectures do not always have complete reflectometers at
each port. Some ports might have a partial reflectometer, as the one shown in the example
of Figure 9.7. This architecture is typically used, especially for a high number of ports
(e.g. twelve or sixteen) to reduce the total number of directional couplers. This has a
clear cost advantage, and also improves measurement speed. In principle, the number of
directional couplers can be reduced from 2n to n + 1, thus measuring only one incident
wave at a time. This, anyway, reduces the calibration possibilities, so it is preferable to
have the multiport system split at least in two halves, as shown in Figure 9.7, and the
number of directional couplers equal to n + 2.
For the partial reflectometer multiport VNA, a different error model must be intro-
duced. Due to the presence of switches in the measurement paths, each port is modeled
in two different states:
• state A: complete reflectometer; see Figure 9.8 (a);

• state B: partial reflectometer, i.e. only the reflected wave can be measured; see
Figure 9.8 (b).
−20
−30
−40
−50
|S12| (dB)
−60
−70
Half−leaky model
Non−leaky model
−80
Full−leaky model
Simulation
−90
−100
0 2 4 6 8 10 12 14 16 18
Frequency (GHz)
(a)
−20
−30
−40
−50
|S12| (dB)
−60
−70
Half−leaky model
−80 Non−leaky model
Full−leaky model
Simulation
−90
−100
0 2 4 6 8 10 12 14 16 18
Frequency (GHz)
(b)
Fig. 9.5 Near-end crosstalk of the two loose coupled coplanar lines of Figure 9.3(b), compared with
simulations. (a): 0.58 mm lines, as are the thrus used during calibration. (b) 6.6 mm lines, i.e.
more than ten times the thrus used during calibration. © 2005 IEEE. Reprinted, with permission,
from [2].
In state A, a full reflectometer is present, thus the state is described by (9.2), that we
rewrite here for a non-leaky case:
aii = li bmii − hi amii

(9.7)
bii = ki bmii − mi amii .
−20
−30
−40
−50
|S14| (dB)
−60
−70
−80 Half−leaky model

Non−leaky model
−90 Full−leaky model
Simulation
−100
0 2 4 6 8 10 12 14 16 18
Frequency (GHz)
Fig. 9.6 Far-end crosstalk of the two loose coupled coplanar lines of Figure 9.3(b), compared with
simulations (6.6 mm long lines). © 2005 IEEE. Reprinted, with permission, from [2].
a1/2 a3/4
b1 b2 b3 b4
Fig. 9.7 An example of partial reflectometer architecture for a four-port VNA.
ami bmi bmi
Γi
bi bi
Source or
termination ai ai
(a) (b)
Fig. 9.8 State A and B configurations, (a) and (b), respectively. © 2008 IEEE. Reprinted, with
permission, from [14].
Instead, for state B, the equations:

aij = gi +
bmij
+
bij = fi bmij (9.8)
i = j
have been recently introduced [14].
The error-model extraction becomes easier if the matrices containing the measure-
ments are organized as follows:
⎡ ⎤ ⎡ ⎤
bm11 0 ··· 0 am11 0 ··· 0
⎢ 0
⎢ bm22 · · · 0 ⎥ ⎥ ,
⎢ 0
⎢ am22 · · · 0 ⎥ ⎥
,
Bm ≡ ⎢ . . . . ⎥ , Am ≡ ⎢ . . . .. ⎥ (9.9)
⎣ . . .
. . . .
. ⎦ ⎣ . . .
. . . . ⎦
0 0 · · · bmnn 0 0 · · · amnn
and
⎡ ⎤
0 bm12 +
+ bm13 ··· +
bm1n
⎢+ 0 + ··· +
bm2n ⎥
⎢bm21 bm23 ⎥
+ ⎢+
Bm ≡ ⎢bm31
+
bm32 0 ··· bm3n ⎥
+
⎥. (9.10)
⎢ . .. .. .. .. ⎥
⎣ .. . . . . ⎦
+
bmn1 +
bmn2 · · · +
bmnn−1 0
We make the assumption that all i ports are always in the B state while the source is
at port j , because it greatly simplifies the theory. The actual wave matrices at the DUT
ports can be seen as
A =,
A ++
A (9.11)
B=,
B ++
B. (9.12)
where
⎡ ⎤ ⎡ ⎤
a11 0 ··· 0 b11 0 ··· 0
⎢ 0 a22 ··· 0 ⎥ ⎢ 0 b22 ··· 0 ⎥
, ⎢ ⎥ , ⎢ ⎥
A≡⎢ . .. .. .. ⎥, B ≡ ⎢ . .. .. .. ⎥,
⎣ .. . . . ⎦ ⎣ .. . . . ⎦
0 0 ··· ann 0 0 ··· bnn
⎡ ⎤ ⎡ ⎤
0 a12 a13 ··· a1n 0 b12 b13 ··· b1n
⎢a21
⎢ 0 a23 ··· a2n ⎥
⎥
⎢b21
⎢ 0 b23 ··· b2n ⎥
⎥
+ ⎢
A ≡ ⎢a31 a32 0 ··· a3n ⎥
⎥, +
⎢
B ≡ ⎢b31 b32 0 ··· b3n ⎥
⎥.
⎢ . .. .. .. .. ⎥ ⎢ . .. .. .. .. ⎥
⎣ .. . . . . ⎦ ⎣ .. . . . . ⎦
an1 an2 ··· ann−1 0 bn1 bn2 ··· bnn−1 0
We can then write (9.7) and (9.8) in matrix form
,
A = L,
Bm − H ,Am
, Bm − M ,
B = K, Am
+ + (9.13)
A = G Bm
+
B = F+
Bm ,
In this formulation leakage is neglected, thus L, M, H, K, F, and G matrices are all

diagonal.
By substituting (9.13) in (9.11) and (9.12) we have
A =,
A ++
A = L,
Bm − H ,
Am + G +
Bm
(9.14)
B=,B ++
B = K,
Bm − M ,
Am + F +
Bm ,
By substituting (9.14) in (9.1), we find the new matrix equation for the error coefficient
computation
−SG+
Bm + F+
Bm − SL,
Bm + K,
Bm + SH,
Am − M,
Am = 0. (9.15)
Like (9.3), (9.15) is written in terms of measured waves rather than measured S-
parameters.
The generalized system (9.15) can be used to compute the error coefficients from the
standard measurements and definitions. As before, it is useful to write the n2 equations
as follows:

n
− (1 − δpj )Sip gp+
bmpj + (1 − δij )fi+
bmij − Sij lj bmjj +
p=1
+δij ki bmij + Sij hj amjj − δij mi amij = 0 (9.16)

(i = 1, . . . , n)
(j = 1, . . . , n)
where δij is the Kronecker delta. These equations can be easily used for one- and two-
port standard connections. A system of equations is obtained by putting together all
the equations coming from the different standard measurements. The solutions of this
system are the error coefficients.
Since this system is homogeneous, in order to avoid the trivial zero solution it is
normalized to one of the unknown coefficients; thus, the total number of unknown error
coefficients is 6n − 1, instead of 4n − 1 as in the complete reflectometer model [15].
From (9.15), the de-embedding equation is the following:
−1
Bm − M,
S = K, Am + F+
Bm L,Bm − H,
Am + G+
Bm . (9.17)
9.2.3 Multiport measurement example

Figure 9.9 shows a typical critical DUT. Port 1 and port 2 are connectorized through
APC7, and have no gender. Instead, ports 3 and 4 have SMA female connectors. This
means that a direct connection, i.e. a “thru” standard, can be inserted only between ports 1
and 2.
P_2
P_1
P_4
P_3
Fig. 9.9 An example of multiport DUT: a directional coupler [16]. © 2008 IEEE. Reprinted, with
One way to solve the calibration problem in this case, without resorting to adapter
removal, is to split it into simpler sub-problems. For example, it is possible to perform
a classical TRL between ports 1 and 2, while an “unknown thru” calibration can be
performed between ports 3 and 4 [18]. At this point, these two sets of error coefficients
must be merged into one, by means of another “unknown thru”, between ports 1 and 4,
for example.
The standard sequence is then:
• APC7 thru ports 1 and 2,

• APC7 line ports 1 and 2,
• APC7 reflect (e.g. a short) at ports 1 and 2,
• SMA female short, open and load at ports 3 and 4,
• SMA female–female adapter at ports 3 and 4,
• SMA female–APC7 adapter at ports 1 and 4.
The measurement results, after this calibration, are shown in Figure 9.10. This multi-
port calibration involving a TRL between ports 1 and 2 is capable of resolving very low
values of insertion loss.
9.3 Mixed-mode S-parameter measurements
Mixed-mode S-parameters have been introduced for the analysis of lines, circuits and sys-
tems in differential configuration at microwaves. In the following we revise the original
definition and provide a generalized method to compute the mixed-mode S-parameters
from single-ended ones.
The original definition of differential and common mode S-parameters is due to
Bockelman and Eisenstadt in 1995 [19]. They introduced the so called mixed-mode scat-
tering matrix, a linear transformation from the single-ended S-matrix to this new matrix,
0.1
Main line insertion 0
loss (dB)
–0.1
–0.2
–0.3
–0.4
0 2 4 6 8 10 12 14 16 18
50
Directivity (dB)
40
30
20
10
0
0 2 4 6 8 10 12 14 16 18
Frequency (GHz)
(a)
–24
–26
–28
–30
Coupling (dB)
–32
–34
–36
–38
–40
–42
0 2 4 6 8 10 12 14 16 18
Frequency (GHz)
(b)
Fig. 9.10 Calibrated measurements of coupler isolation and directivity (a) and coupling factor (b) © 2008
IEEE. Reprinted, with permission, from [17].
and designed and implemented an instrument able to measure directly the mixed-mode
S-matrix, i.e. the pure-mode VNA [20]. In [19] Bockelman and Eisenstadt showed, for
a four-port case, that if the differential and common mode voltages and currents are
defined as
Vd12 ≡ V1 − V2
Id12 ≡ (I1 − I2 )/2
Vc12 ≡ (V1 + V2 )/2
Ic12 ≡ I1 + I2
(9.18)
Vd34 ≡ V3 − V4
Id34 ≡ (I3 − I4 )/2
Vc34 ≡ (V3 + V4 )/2
Ic34 ≡ I3 + I4
then it is possible to define the differential and common mode waves similarly to the
single-ended ones
ad12 ≡ √1 (Vd12 + Rd Id12 ) =

2 Rd
√1 (a1 − a2 )
2
bd12 ≡ 2√1R (Vd12 − Rd Id12 ) = √1 (b1 − b2 )
d 2
ac12 ≡ 2√1R (Vc12 + Rc Ic12 ) = √1 (a1 + a2 )
c 2
bc12 ≡ 2√1R (Vc12 − Rc Ic12 ) = √1 (b1 + b2 )
c 2
(9.19)
ad34 ≡ 2√1R (Vd34 + Rd Id34 ) = √1 (a3 − a4 )
d 2
bd34 ≡ 2√1R (Vd34 − Rd Id34 ) = √1 (b3 − b4 )
d 2
ac34 ≡ 2√1R (Vc34 + Rc Ic34 ) = √1 (a3 + a4 )
c 2
bc34 ≡ 2√1R (Vc34 − Rc Ic34 ) = √1 (b3 + b4 )
c 2
where Rd and Rc are purely real reference impedances (typically Rd = 100 and
Rc = 25 ). Consequently, the mixed-mode S-matrix SMM was defined as,
⎛ ⎞ ⎛ ⎞
bd12 ad12
⎜bd34 ⎟ ⎜ad34 ⎟
⎜ ⎟ ⎜ ⎟
⎝bc12 ⎠ ≡ SMM ⎝ac12 ⎠ . (9.20)
bc34 ac34
This matrix can also be computed directly from the single-ended S-matrix as:
SMM = MSM−1 . (9.21)
where ⎛ ⎞
1 −1 0 0
1 ⎜ 0 0 1 −1⎟
M= √ ⎜ ⎟. (9.22)
2 ⎝1 1 0 0⎠
0 0 1 1
The measurement of the mixed-mode S-matrix SMM can then be performed by mea-
suring a single-ended S-matrix with a single-ended multiport VNA, and then applying
(9.21). This is the easiest and more common approach. An alternative is a modified VNA,
the pure-mode VNA, which is able to separately excite the differential and the common
mode, by using a 180◦ hybrid coupler. Two possible implementations are shown in
Figure 9.11.
Microwave
source
Σ 180° Δ
Hybrid
0° 180° MW sources
(variable phases)
0° 180° 0° 180°
a1 a2 a3 a4 a1 a2 a3 a4
b1 b2 b3 b4 b2 b2 b3 b4
Mixed-mode Mixed-mode Mixed-mode Mixed-mode

(a) (b)
Fig. 9.11 Possible implementations of a pure mode VNA, with hybrid couplers (a) [19], or independent
sources (b). © 2008 IEEE. Reprinted, with permission, from [21].
The Bockelman and Eisenstadt formulation is very simple and intuitive, but has the
following drawbacks:
• the waves are referred to a real reference impedance, so if a TRL-like calibration is

applied, the notation does not take into account the complex characteristic impedance
of the reference line, and will be inaccurate [22]
• it cannot be applied to a DUT having one (or more than one) single-ended ports.
These drawbacks have been overcome by a recently introduced formulation [23].

It applies to the measurement of an n-port DUT, with p mixed-mode ports and n − p
single-ended ports. For the latter n − p ports, the single-ended pseudo-waves are defined,
according to [22]
Vj + Ij Zj
aj ≡ α R j
2|Zj |
Vj − Ij Zj
bj ≡ α Rj ,
2|Zj |
where Zj is the complex reference impedances, Rj = {Zj }, and α is a unitary magnitude

complex normalization factor.
For the mixed-mode p ports, considering the port-pair j, k of the mixed-mode port
set, with reference impedances Zcj k for the common mode, and Zdj k for the differential
mode, the mixed-mode waves are
Vdj k + Idj k Zdj k

adj k ≡ Rdj k
2|Zdj k |
Vdj k − Idj k Zdj k
bdj k ≡ Rdj k
2|Zdj k |
Vcj k + Icj k Zcj k
acj k ≡ Rcj k
2|Zcj k |
Vcj k − Icj k Zcj k
acj k ≡ Rcj k ,
2|Zcj k |

where Rdj k = Zdj k and Rcj k = Zcj k .
All the pseudo-waves are re-ordered in the following vectors:
⎛ ⎞ ⎛ ⎞
ad12 bd12
⎜ ⎟
ad34 ⎜ ⎟ bd34
⎜ ⎟ ⎜ ⎟
⎜ ⎟
· ⎜ ⎟ ·
⎜ ⎟ ⎜ ⎟
⎜ ⎟
· ⎜ ⎟ ·
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎜ad(p−1)p ⎟ ⎜bd(p−1)p ⎟
⎜ ⎟ ⎜ ⎟
⎜ ac12 ⎟ ⎜ bc12 ⎟
⎜ ⎟ ⎜ ⎟
⎜ ac34 ⎟ ⎜ ⎟
◦ ⎜ ⎟ ◦ ⎜ bc34 ⎟
⎜
a≡ ⎜ ⎟ ⎜ ⎟
· ⎟ b≡ ⎜ · ⎟
⎜ · ⎟ ⎜ · ⎟
⎜ ⎟ ⎜ ⎟
⎜a ⎟ ⎜b ⎟
⎜ c(p−1)p ⎟ ⎜ c(p−1)p ⎟
⎜ ⎟ ⎜ ⎟
⎜ ap+1 ⎟ ⎜ bp+1 ⎟
⎜ ⎟ ⎜ ⎟
⎜ · ⎟ ⎜ · ⎟
⎜ ⎟ ⎜ ⎟
⎜ · ⎟ ⎜ · ⎟
⎜ ⎟ ⎜ ⎟
⎝ an−1 ⎠ ⎝ bn−1 ⎠
an bn
and a generalized mixed-mode matrix can now be defined
◦ ◦◦
b ≡ Sa. (9.23)
◦
Starting from these definitions, it is possible to find that the relationship between S
and S is a bilinear transformation,
◦
S ˜ 21 +
= ( ˜ 22 S)(
˜ 11 +
˜ 12 S)−1 (9.24)
bm1 bm5
1 5
bm2 bm6
am1-4 am5-8
2 6
bm3 bm7
3 7
bm4 bm8
4 8
Fig. 9.12 Probe setup and measurement architecture. © 2008 IEEE. Reprinted, with permission,
from [21].
where ˜ ij are transformation matrices containing all the single-ended and mixed-mode
reference impedances, computed in [23].
9.3.1 Mixed-mode multiport measurement example

The example presented here, courtesy of Intel Corporation, appeared in [24]. The problem
is the on-wafer measurements of transmission lines in differential configuration, with an
eight-port VNA system.
The probe setup is shown in Figure 9.12, where the measurement architecture is also
shown. The multiport system was a partial reflectometer system, divided in two halves
(ports 1–4 and 5–8), so that a two-port standard, connected for example between ports
1 and 5, can be measured as if the reflectometer architecture was complete. Thus, any
two-port calibration algorithm can be applied to these ports. In particular, a multiline
TRL was chosen for ports 1 and 5, while the rest of the ports were calibrated with
the thru connections summarized in Table 9.1, using QSOLT-type algorithms [25]. The
probe touchdowns for this calibration are shown in Figure 9.13. The total number of
Table 9.1 Calibration matrix
Port 1 Port 2 Port 3 Port 4 Port 5 Port 6 Port 7 Port 8
Port 1 X X X X MTRL X X X
Port 2 X X X X thru thru X X
Port 3 X X X X X thru thru X
Port 4 X X X X X X thru thru
Port 5 MTRL thru X X X X X X
Port 6 X thru thru X X X X X
Port 7 X X thru thru X X X X
Port 8 X X X thru X X X X
N/A
N/A
(a) (b)
N/A N/A
(c) (d)
Fig. 9.13 Calibration touchdowns: straight thrus (a), shifted thrus (b), are obtained with the same
calibration standard. Offset opens (c) are measured only at ports 1 and 5, as the set of three
different length lines (d). © 2008 IEEE. Reprinted, with permission, from [21].
touchdowns for the calibration of this eight-port system is only six, which is rather low,
considering that a multiline TRL with three lines would require five touchdowns for a
two-port VNA.
The measurement results of a pair of transmission lines in differential configuration
are shown in Figure 9.14. The differential parameters of interest, the insertion loss (IL)
and the return loss (RL) were computed from the single-ended S-parameter, applying
the transformations (9.24). It is interesting to note that the analysis of the mixed-mode
performances leads to different conclusions than the analysis of the single-ended ones.
If we consider a fixed (target) level of RL, e.g. -12 dB, this performance is achieved for
higher frequency if the differential RL is considered instead of the single-ended RL. The
same happens for the insertion loss, as shown in Figure 9.14(b).
In conclusion, this example has shown that choosing the proper error model and cali-
bration algorithm can considerably reduce the measurement time and costs. Moreover,
the mixed-mode S-parameter matrix can give important information for the design and
performance analysis of differential devices and circuits.
0
Single-ended RL (dB) –5
–10
–15
–20
–25
–30
–35
0
Differential RL (dB)
–5
–10
–15
–20
–25
–30
–35
0 2 4 6 8 10 12 14 16 18 20
Freq (GHz)
(a)
0
Single-ended IL (dB)
–2
–4
–6
–8
–10
0
Differential IL (dB)
–2
–4
–6
–8
–10
0 2 4 6 8 10 12 14 16 18 20
Freq (GHz)
(b)
Fig. 9.14 Single-ended and differential return loss (a) and insertion loss (b) of two lines in differential
configuration. © 2008 IEEE. Reprinted, with permission, from [21].
References
[1] T. G. Ruttan, B. Grossman, A. Ferrero, V. Teppati, and J. Martens, “Multiport VNA

measurements,” IEEE Microwave, pp. 56–69, June 2008.
[2] V. Teppati and A. Ferrero, “On-wafer calibration algorithm for partially leaky multiport vector
network analyzers,” IEEE Trans. Microw. Theory Tech., MTT-53, no. 11, pp. 3665–3671, Nov.
2005.
[3] B. Grossman, T. Ruttan, and E. Fledell, “Architectural considerations for multiport vector
network analyzers,” Proc. IEC DesignCon 2007, 13, Feb. 2007.
[4] B. Grossman and T. Ruttan, “Why multi-port VNAs?” 70th ARFTG Conf. Signal Integrity
Workshop Dig., pp. 120–139, Dec. 2007.
[5] B. Grossman, T. Ruttan, and E. Fledell, “Comparison of multiport VNA architectures
measured results,” Proc. 66th ARFTG Conf., Dec. 2005.
[6] J.-C. Tippet and R.-A. Speciale, “A rigorous technique for measuring the scattering matrix
of a multiport device with a 2-port network analyzer,” IEEE Trans. Microw. Theory Tech.,
MTT-30, no. 5, pp. 661–666, May 1982.
[7] U. Lott, W. Baumberger, and U. Gisiger, “Three-port RF characterization of foundry dual
gate FETs using two-port test structures with on-chip loading resistors,” Proc. IEEE Int.
Conference on Microelectronics Test Structures, pp. 167–180, Mar. 1995.
[8] C. S. Hartmann and R. T. Hartmann, “Software for multi-port RF network analysis with a
large number of frequency samples and application to 5-port SAW device measurement,”
Ultrasonics Symposium Proceedings, 1, pp. 117–122, Dec. 1990.
[9] H.-C. Lu and T.-H. Chu, “Multiport scattering matrix measurement using a reduced-port
network analyzer,” IEEE Trans. Microw. Theory Tech., MTT-51, no. 5, pp. 1525–1533, May
2003.
[10] J.-C. Rautio, “Techniques for correcting scattering parameter data of an imperfectly termi-
nated multiport when measured with a two-port network analyzer,” IEEE Trans. Microw.
Theory Tech., MTT-31, no. 5, pp. 407–412, May 1983.
[11] M. Davidovits, “Reconstruction of the S-matrix for a 3-port using measurements at only two
ports,” IEEE Trans. Microw. Theory Tech., MTT-5, no. 10, pp. 349–350, Oct. 1995.
[12] A. Ferrero and F. Sanpietro, “A simplified algorithm for leaky network analyzer calibration,”
IEEE Microw. Guid. Wave Lett., 5, no. 4, pp. 119–121, Apr. 1995.
[13] A. Ferrero, F. Sampietro, and U. Pisani, “Multiport vector network analyzer calibration: a
general formulation,” IEEE Trans. Microw. Theory Tech., MTT-42, no. 12, pp. 2455–2461,
Dec. 1994.
[14] A. Ferrero, V. Teppati, M. Garelli, and A. Neri, “A novel calibration algorithm for a special
class of multiport vector network analyzers,” IEEE Trans. Microw. Theory Tech., MTT-56,
pp. 693–699, Mar. 2008.
[15] A. Ferrero, U. Pisani, and K. Kerwin, “A new implementation of a multiport auto-
matic network analyzer,” IEEE Trans. Microw. Theory Tech., MTT-40, pp. 2078–2085,
Nov. 1992.
[16] V. Teppati and A. Ferrero, “A new class of non-uniform, broadband, non-symmetrical rect-
angular coaxial-to-microstrip directional couplers for high power applications,” IEEE Trans.
Microw. Wireless Compon. Lett., 13, no. 4, pp. 152–154, Apr. 2003.
[17] V. Teppati, A. Ferrero, and U. Pisani, “Recent advances in real-time load-pull systems,” IEEE
Trans. Instrum. Meas., 57, no. 11, pp. 2640–2646, Nov. 2008.
[18] A. Ferrero and U. Pisani, “Two-port network analyzer calibration using an unknown ’thru’,”
IEEE Microw. Guid. Wave Lett., MGWL-2, pp. 505–507, Dec. 1992.
[19] D. Bockelman and W. Eisenstadt, “Combined differential and common-mode scattering
parameters: theory and simulation,” IEEE Trans. Microw. Theory Tech., MTT-43, no. 7,
pp. 1530–1539, July 1995.
[20] D. Bockelman and W. Eisenstadt, “Pure-mode network analyzer for on-wafer measurements
of mixed-mode S-parameters of differential circuits,” IEEE Trans. Microw. Theory Tech.,
MTT-45, no. 7, pp. 1071–1077, July 1997.
[21] A. Ferrero and V. Teppati, “Multiport and mixed mixed-mode measurements,” in Proc. 72nd
ARFTG Conf., Signal Integrity Workshop, Portland, OR, Dec. 2008.
[22] R. Marks and D. Williams, “A general waveguide circuit theory,” J. Res. NIST, 97,
pp. 533–561, Sept. 1992.
[23] A. Ferrero and M. Pirola, “Generalized mixed-mode S-parameters,” IEEE Trans. Microw.
Theory Tech., MTT-54, pp. 458–463, Jan. 2006.
[24] E. Fledell and T. Ruttan, “Digital backplane interconnections and bus multi-port differential
characterization,” in IEEE MTT-S Intl. Microwave Symp. Dig., June 2006.
[25] A. Ferrero and U. Pisani, “Qsolt: a new fast calibration algorithm for two-port S-parameter
measurements,” in 38th ARFTG Conf. Dig., San Diego, CA, Dec. 1991, pp. 15–24.
10 Noise figure characterization
Nerea Otegi, Juan-Mari Collantes, and Mohamed Sayed
10.1 Introduction
Noise is one of the most critical issues in wireless systems because it is a fundamental
limiting factor for the performance of microwave receivers. Industry requirements for
increasingly higher performing communication systems require tighter noise specifica-
tions that make the noise figure measurement a critical step in the characterization of
modern microwave circuits and systems.
Noise figure measurements of circuits and sub-systems have been traditionally per-
formed with noise figure meters specifically developed for that purpose. A paradigmatic
example is the HP8970 (and associated family) that was considered for years as the ref-
erence meter for noise figure characterization. This instrument, as well as other modern
equipment, uses the popularY-factor technique to compute the noise figure from the ratio
of two power measurements (“cold” and “hot”). The scalar nature of the measurements
allows an easy and straightforward characterization process. This simplicity is undoubt-
edly part of its large success. However, its accuracy is limited by the match properties
of the device under test and measurement setup.
There are two factors that have been driving an evolution in the noise figure charac-
terization schemes. One factor is a growing tendency in microwave instrumentation to
integrate different types of measurements into a single instrument box. As a result, noise-
figure characterization is now available as an option in modern vector network analyzers
(VNA) from different manufacturers. The other factor is that the accuracy require-
ments in environments that are not perfectly matched (millimeter wave and beyond,
on-wafer setups, etc.) demand a noise figure characterization that takes advantage of
vector measurements to improve scalar results.
In this context, solutions have been proposed to enhance the original scalar Y-factor
technique with vector correction terms that account for systematic errors such as mis-
match. Moreover, techniques other thanY-factor are also proposed in modern equipment.
This is the case for some new VNAs that use the cold-source technique. Here, the device
is measured at a single “cold” state. The cold-source technique was mainly used in the
past to characterize the noise parameters of single transistors, which very often present
poor match characteristics. That is why, in its classical form, cold-source includes correc-
tions for mismatch errors and requires vector measurements. It is indeed a more complex
characterization approach than the scalar Y-factor technique.
Noise figure characterization 241
Noise figure characterization approaches with the ability to correct for a variety of
systematic errors are not exclusive to new VNAs. They can also be applied to other
microwave instruments with noise figure capabilities as most modern spectrum analyzers
(SA), although additional equipment for vector measurements is required in this case.
In this chapter we provide a detailed description of both Y-factor and cold-source
techniques for noise figure characterization. As starting point, the fundamentals of noise
figure are briefly summarized in Section 10.2. Section 10.3 is devoted to the classical
Y-factor technique, while cold-source is treated in Section 10.4. In Section 10.5 the main
sources of systematic errors in a noise figure measurement are analyzed (mismatch,
receiver bandwidth and linearity, etc.). Whenever it is relevant, their impact on each
technique is comparatively discussed and measurement examples are provided. Finally,
Section 10.6 is dedicated to the noise figure characterization of mixers. This is usually a
challenging measurement because, in addition to the frequency translation, nearly every
drawback affecting ordinary two-ports is magnified in mixers.
10.2 Noise figure fundamentals
10.2.1 Basic definitions and concepts

Noise
Noise is a random process associated with several sources. Thermal noise (Johnson [1]
or Nyquist [2] noise) is one of the principal noise mechanisms in RF and microwave
systems and it is caused by the random motion of thermally excited charge carriers in
any passive circuit element that contains losses. There are several other noise sources
such as shot noise, generation-recombination (G-R) noise, flicker noise, quantum noise
etc. The analysis of the properties of noise as a random process or the particular nature
of these several noise sources is beyond the scope of this chapter. Fundamental readings
on noise can be found, amongst others, in [3–6].
Any resistive element at a temperature T different from zero Kelvin generates thermal
noise that can be expressed as the available thermal noise power [2]:
N = kT B. (10.1)
where k is the Boltzmann constant (1.38e-23 Joule/Kelvin) and B is the considered

bandwidth (Hertz). Therefore, the available noise power generated by a passive element
is proportional to temperature and bandwidth, but does not depend on resistance or
frequency. A noise mechanism that fulfils this last property is referred to as white noise.
Actually, the spectral density of the available thermal noise power has a slight frequency
dependence that can be neglected up to TeraHertz frequencies [4]. Other types of noise,
such as shot and G-R noise, also behave as white noise in the RF and microwave frequency
ranges, so in a noise characterization the overall effect of these sources is usually treated
as equivalent to thermal noise [7].
242 Nerea Otegi, Juan-Mari Collantes, and Mohamed Sayed
Noise figure
Any two-port device, in addition to amplifying or attenuating both the signal and the noise
present at its input, adds extra noise generated by its own components, thus degrading
the signal-to-noise ratio (SNR). The noise figure is a figure of merit that characterizes
this degradation and it is defined as the ratio of the SNR at the input and the SNR at the
output when the input noise is thermal noise generated by a passive load at a reference
temperature of T0 = 290 K [8]:

Si /Ni
F= . (10.2)
So /No T =T0
Si and So are, respectively, the signal powers available at the input and output of the
two-port, while Ni and No are the available noise powers. This definition of noise figure
can be extended to multiport devices [9]. It is worthwhile to note that, because of its
definition, the relevance of the noise figure is limited to low input signals and low noise
levels. This is why the noise figure has little significance for a power amplifier, where
the added noise has a negligible contribution to the degradation of the SNR because of
the large levels involved.
The most basic concepts associated with noise figure are graphically shown in
Figure 10.1. In Figure 10.1(a), a generic block diagram of a two-port device with a
noise source connected at its input can be seen. The noise power available at the output
of the device as a function of the noise source temperature, the “noise line,” is plotted in
Figure 10.1(b).
According to the definition of noise figure, the available noise at the input of the
two-port device is thermal noise generated by a passive load:
Ni = kBT0 . (10.3)
In addition, the noise power available at the output port can be expressed as
No = Gav Ni + Nadd = kBGav T0 + Nadd (10.4)
Ni = kBT0 No = GavNi +Nadd No

Si So = GavSi kBGav
Gav , Nadd N0
kBGavT0
T0 Nadd
Γs Γout
T0 T
(a) (b)
Fig. 10.1 (a) Generic two-port device with noise source connected at input. (b) Graphical representation of
the noise at the output of the device as a function of the noise source temperature.
where, Nadd is the noise added by the two-port device and Gav is the two-port available
gain [7], defined as
1 − |s |2 1
Gav = |s21 |2 , (10.5)
|1 − s11 s | 2
1 − |out |2
where sij are the S-parameters of the two-port device, out is its output reflection coef-
ficient, and s is the reflection coefficient of the passive load connected at the input of
the two-port device.
The noise figure definition in (10.2) can be rewritten as:
kBGav T0 + Nadd
F= , (10.6)
kBGav T0
which is the ratio of the noise power available at the output port to the contribution to
the output of the input termination, when this termination is at the reference temperature
of 290 K. (10.6) represents the formal definition of noise figure adopted by the IRE
[10, 11].
It is seen from (10.6) that the noise figure characterizes the noise added by the device
and, thus, this added noise can be expressed as a function of the noise figure as
Nadd = kBGav T0 (F − 1) . (10.7)
The IRE introduced the equivalent denomination noise factor for the noise figure
(10.6), [12], sometimes called noise figure in linear terms. It is nowadays broadly
accepted to use noise figure NF for the quantity (10.8), expressed in dB, while noise
factor is used for the linear quantity F . From now on, this convention is followed in this
chapter.
NF = 10 log10 (F ) . (10.8)
Noise temperature
Sometimes, especially for low noise devices, the effective input noise temperature, Te , is
used instead of the noise factor to characterize the noise generated by a device. Accord-
ing to [11] the effective input noise temperature is the temperature at which a source
termination connected to a noise-free equivalent of the two-port device would lead to the
same output noise power of the real two-port device with a noise-free source termination.
Figure 10.2 illustrates the meaning of effective input noise temperature.
According to its definition, the noise at the output of the two-port device can be written
in terms of this effective input noise temperature as:
No = kBGav (T0 + Te ) . (10.9)

No = kBGavT0 + Nadd No = kBGav (T0 + Te)
Gav , Nadd + Gav
Noise-free
T0 DUT T0 Te equivalent
of DUT
Fig. 10.2 Graphical interpretation of effective input noise temperature.
i1 i2 i1 νn i2
Noisy Noiseless
ν1 ν2 ν1 in ν2
two-port two-port
(a) (b)
Fig. 10.3 (a) Two-port device with internal noise sources; (b) equivalent circuit with noise voltage source
vn and noise current source in at noise-free equivalent input.
It is seen from its definition that the effective input noise temperature is a translation
to the input port of the noise added by the device.
Nadd
Te = . (10.10)
kBGav
If (10.9) is brought into the definition of noise factor, a simple relationship between
the noise factor and the effective input temperature is obtained:
Te
F= + 1. (10.11)
T0
This relationship between the effective input noise temperature and the noise factor
is limited to two-port transducers with a single input frequency and a single output
frequency, as explicitly stated in [12]. This is a non-trivial assessment that is further
analyzed in Section 10.6, that is devoted to mixer noise figure characterization.
Noise parameters
The noise behavior of a linear two-port device can be fully modeled by two noise sources
added to a noise-free equivalent of the original two-port device [13], shown in the clas-
sical representation of Figure 10.3. As there are several internal noise processes that are
complex, these equivalent noise sources are generally not independent. Several represen-
tations in terms of different equivalent noise sources [13, 14], or parameterizations based
on noise-waves [14] can be found in the literature. A compilation of diverse represen-
tations is given in [16]. Whatever the noise representation, the noise sources associated
with the two-port device are correlated in a general case. Thus, four independent param-
eters, noise parameters, are required to fully characterize the internal noise of a linear
two-port device in terms of its source impedance (leading to the so-called noise correla-
tion matrix): two real parameters, one for each of the sources, and the real and imaginary
parts of a complex parameter that takes into account the correlation between the sources.
In the model of Figure 10.3, these four parameters are the mean square fluctuations of
the noise sources, vn 2 , and in 2 , and a complex correlation parameter vn in ∗ .
Accordingly, the noise factor of a two-port device depends
4 on the source
termination
5
through a set of four independent noise parameters [14]. Fmin , Rn , Re opt , I m opt
is the most common set of noise parameters for microwave two-port devices. These noise
parameters are derived from the noise fluctuations (vn 2 , in 2 , vn in ∗ ) and completely char-
acterize the noise response of the two-port device [14]. Fmin is the minimum noise factor
of the device and opt is the optimum source reflection coefficient, which provides the
minimum noise factor. The “noise resistance” Rn is a parameter that characterizes how
rapidly the noise factor diverges from Fmin as the source reflection coefficient varies
from the optimum case. In this representation, the noise factor is given as a function of
the reflection coefficient as:
Rn | − opt |2
F = Fmin + 4 , (10.12)
Z0 |1 + opt |2 1 − ||2
where Z0 is the characteristic impedance.

A three-dimensional representation of (10.12) shows a paraboloid with its minimum
(Fmin ) at opt (see Figure 10.4). If constant noise factor values are depicted as a function
of on the Smith-Chart, the well-known noise circles [17] are obtained. These circles
represent the projection of the paraboloid on the Smith-Chart.
Fmin
Γopt
Fig. 10.4 Three-dimensional representation of noise factor versus source reflection coefficient and noise
circles on a Smith-Chart.
Obtaining the noise parameters of a two-port device is not an easy task [18]. The
first methods for characterizing the noise parameters of a two-port device were based on
the actual experimental searching for the minimum noise factor and its corresponding
optimum source impedance [10]. However, computer-aided data fitting techniques were
soon proposed to extract the noise parameters from measured data [19], leading to faster
and more accurate noise parameter characterization techniques (source pull techniques),
as [19–23]. For that at least four noise figures corresponding to four source reflection
coefficients are required, although more than four terminations are normally used to
obtain the parameters from an overdetermined system and minimize errors. The noise
parameter extraction was further simplified on the basis of directly measuring noise
power values instead of noise figures [24–26]. Impedance tuners are used to synthesize
the required source reflection coefficients, which have to be adequately distributed on the
Smith-Chart in order to avoid ill-conditioning problems [27–30]. To facilitate the data
fitting, linearized versions of the classical noise parameter representation, sometimes
referred to as noise pseudoparameters, are used, as in [19]. Linearized parameterizations
based on noise-wave descriptions can also be found in [31, 32].
Noise figure of cascaded devices

Let us consider a system formed by several cascaded two-port devices, each of them
represented by its noise factor Fj and available gain Gavj , like the one shown in
Figure 10.5.
As shown by Friis [8], the overall noise factor of a cascaded system is given by
F2 (out1 ) − 1 F3 (out2 ) − 1
F = F1 (s ) + + + ··· (10.13)
Gav1 Gav1 Gav2
This expression reflects a well-known fact: the importance of the first stage in a system
receiver, since the contribution to the SNR degradation of later stages is reduced by the
product of gains of the preceding ones.
10.2.2 Two noise figure characterization concepts:

Y-factor and cold-source
In essence, characterizing the noise figure involves the knowledge of the “noise line”
in Figure 10.1(b). There are two fundamental methods for characterizing a line: obtain-
ing two points of the line or obtaining a point and the slope. These two concepts are
F1, Gaν1 F2, Gaν2 F3, Gaν3
Γs Γout1 Γout 2
Fig. 10.5 Block diagram of a cascaded system formed by several two-port devices.
directly related to the two basic noise figure measurement methodologies that are further
discussed in this chapter: Y-factor and cold-source.
The Y-factor technique obtains the noise factor from two noise powers (Nc , Nh )
corresponding to two different input temperatures (Tc , Th ), known as cold and hot temper-
6
atures, respectively [11]. The ratio of these two quantities is called Y-factor, Y = Nh Nc ,
and thus the name of the technique. (10.14) and (10.15) represent the noise powers avail-
able at the output of the two-port device for the two input temperatures. The noise factor
can be expressed as a function of the Y-factor and both temperatures as shown in (10.16).
Nc = kBGav [Tc + T0 (F − 1)] (10.14)

Nh = kBGav [Th + T0 (F − 1)] (10.15)
6 6
Th T0 − 1 − Y Tc T0 − 1
F= . (10.16)
Y −1
Assuming that the cold temperature Tc is very close to the reference temperature T0 ,
the noise factor can be approximated to:
6
Th T0 − 1
F≈ . (10.17)
Y −1
The cold-source approach obtains the noise factor from a single noise power corre-
sponding to the cold temperature [24]. The noise factor is directly derived from (10.14)
and does not require a hot noise power measurement. The “noise slope” kBGav has to
be previously determined for that. The noise factor is given by (10.18), which can be
approximated to (10.19) considering again that the cold temperature Tc is close to the
reference temperature T0 .

Nc Tc
F= − −1 (10.18)
kBGav T0 T0
Nc
F≈ . (10.19)
kBGav T0
10.3 Y-factor technique
The Y-factor technique is the most popular noise figure measurement methodology, used
by the majority of the commercially available noise figure meters from the classical
HP-8970 [33], to recent versions of the NFA series N897X-A [34]. Modern noise figure
measurement implementations included in spectrum analyzers are also usually based
on the Y-factor technique [35, 36]. A very detailed description of this technique can be
found in [37]. The basic diagram of a Y-factor measurement is shown in Figure 10.6.
As schematically shown in the previous section, in the Y-factor technique the noise
factor is obtained through two noise power measurements for two different input
temperatures. In order to physically generate the cold and hot input “temperatures,”
Measurement Calibration
Nc, Nh Nc_rec , Nh_rec

DUT Receiver Receiver
Th Th
Γout
Tc Γs Tc Γs
Noise Noise
source source
(a) (b)
Fig. 10.6 Basic block diagrams for Y-factor characterization technique: (a) measurement step; (b)
calibration step.
i.e. available noise powers corresponding to these temperatures, commercial noise

sources are normally used. These noise sources are avalanche diodes that provide two
different broadband noise levels, depending on whether they are biased or not [38]. When
the diode is not biased (“off” or “cold” state), it is basically a resistor that generates a
noise power proportional to the ambient temperature. However, when the diode is biased
in avalanche (“on” or “hot” state) it provides extra noise, significantly higher than the
thermal noise corresponding to the off state, which can be related to a hot temperature.
This hot temperature
is quantified
6 by the Excess Noise Ratio (ENR) of the noise source:
ENR = 10 log10 (Th − Tc ) T0 .
With the noise source connected to the DUT, two noise powers are measured
(Figure 10.6(a)) so that the noise factor can be calculated as shown in (10.16). However,
the noise factor obtained in this way is not the device noise factor, FDUT , but the noise
factor Fsys of the cascaded system composed by the DUT and the noise receiver. Thus,
from the measured noise powers (Nc , Nh ) the system noise factor is calculated as:
6 6
Th T0 − 1 − Y Tc T0 − 1 Nh
Fsys = , Y= . (10.20)
Y −1 Nc
A second stage correction is required to eliminate the noise contribution of the receiver.
To this end, a calibration step is needed to characterize the receiver noise factor. With the
noise source directly connected to the receiver (Figure 10.6(b)) two noise powers (Nc_rec
and Nh_rec ) are measured and the receiver noise factor Frec is again calculated as:
6
6
Th T0 − 1 − Yrec Tc T0 − 1 Nh_rec
Frec = , Yrec = . (10.21)
Yrec − 1 Nc_rec
Modern noise figure measurement equipment includes the possibility of correcting the
differences between the cold temperature Tc and the reference one T0 [39]. If the cold
temperature can be approximated to the reference, (10.20) and (10.21) can be simplified
according to (10.17) as:

6
Th T0 − 1 Nh
Fsys ≈ , Y= (10.22)
Y −1 Nc
6
Th T0 − 1 Nh_rec
Frec ≈ , Yrec = . (10.23)
Yrec − 1 Nc_rec
The calibration step is also used to obtain the DUT gain, necessary for the second-
stage correction. From the four measured scalar noise powers the insertion gain of the
device is computed by means of (10.24).
Nh − Nc
Gins ≡ . (10.24)
Nh_rec − Nc_rec
Finally, the device noise figure is calculated as:

Frec (s ) − 1
NF YF ≡ 10 log10 Fsys − . (10.25)
Gins
It is important to remark that all the measurements involved in the determination of

(10.25) are exclusively scalar measurements.
10.4 Cold-source technique
Following [24], the cold-source technique is a usual noise-measurement methodology

for noise parameter extraction. A basic diagram of the cold-source procedure is given in
Figure 10.7. In the cold-source technique the noise figure is characterized from a single
noise-power measurement of the device, (Nc ), with a source termination at ambient
Measurement Noise calibration

Nh_rec (Γsh)
Receiver
Th
Nc Γsh
DUT Receiver
Nc_rec (Γscj)
50 Ω Tc Γout Receiver
Γs
Tc
Γsc1, Γsc2, Γsc3, Γsc4, ...
(a) (b)
Fig. 10.7 Basic block diagrams for cold-source characterization technique: (a) measurement step; (b)
calibration step.
temperature connected at its input (see Figure 10.7(a)). To this end, the device available
gain Gav and the gain-bandwidth product of the receiver kB|s21rec |2 have to be previously
determined. To obtain an accurate measurement of the kB|s21rec |2 term requires the
use of a noise source in its cold and hot states, but it should be noted that the hot
noise source is only necessary in the calibration step. Cold-source implementations that
characterize the receiver gain-bandwidth product without a noise source can also be found
[40, 41]. The kB|s21rec |2 term can be estimated by obtaining the gain and bandwidth
responses of the receiver. For that, a narrowband frequency sweep is performed and
the effective noise bandwidth is computed by integrating this response [10]. A second
stage correction is again needed to properly characterize the device noise figure and,
thus, the noise contribution of the receiver has to be characterized in the calibration
step, as well. Generally devoted to noise parameter extraction, treated in Section 10.2.1,
the cold-source technique is normally a fully corrected procedure, including a complete
receiver noise calibration to get its four noise parameters. The DUT noise figure is
obtained from

Nc
NF CS ≡ 10 log10
kB|s21rec |2 G
av MM (out ) T0

Frec (out ) − 1 Tc
− − −1 , (10.26)
Gav T0
where Frec (out ) is the receiver noise factor corresponding to the output reflection coef-
ficient of the DUT, Gav is the device available gain given in (10.5), and MM (out ) is a
term accounting for the mismatch between the device and the receiver (with s11rec the
input reflection coefficient of the receiver).
1 − |out |2
MM (out ) = , (10.27)
|1 − s11rec out |2
In order to characterize the four noise parameters required to obtain Frec (out ) and
the gain-bandwidth product of the receiver kB|s21rec |2 a calibration step is needed
(Figure 10.7(b)). For that, at least four noise power measurements with four passive
loads (10.28) and one hot measurement (10.29) are required.

Nc_rec scj = kB|s21rec |2 MM scj Tc + T0 Frec scj − 1 ,
j = 1, 2, 3, 4 . . . (10.28)
Nh_rec (sh ) = kB|s21rec |2 MM (sh ) [Th + T0 (Frec (sh ) − 1)] , (10.29)
where
Rn_rec | − opt_rec |2
Frec () = Fmin_rec + 4 . (10.30)
Z0 |1 + opt_rec |2 1 − ||2
4
The
five unknowns (kB|s21rec |2 and the receiver noise parameters Fmin_rec , Rn_rec ,
5
Re opt_rec , Im opt_rec , or an equivalent set) are extracted from these measured
noise powers. For that, linearized versions of (10.28) (10.29), in terms of equivalent
noise pseudoparameters, are generally used. Diverse approaches to the noise description,
based on different parameterizations, can be found, as for instance in [25, 32, 42].
A simplified estimation of kB|s21rec |2 can be obtained from (10.31) assuming that the
receiver noise factor does not vary between one of the cold measurements and the hot
one. This estimation is sometimes used as the starting point of an iteration process [24].
Nh_rec (sh ) /MM (sh ) − Nc_rec (sc1 ) /MM (sc1 )

kB|s21rec |2 ≈ , (10.31)
Th − Tc
where Nc_rec (sc1 ) and Nh_rec (sh ) are, respectively, the cold and hot noise powers
measured by the receiver with the noise source connected to it (typically sc1 ≈ 0,
sh ≈ 0). Note however, that the actual kB|s21rec |2 would be given by:
Nh_rec (sh ) /MM (sh ) − Nc_rec (sc1 ) /MM (sc1 )

kB|s21rec |2 = . (10.32)
Th − Tc + T0 (Frec (sh ) − Frec (sc1 ))
As explained in Section 10.2.1, to minimize errors, more than four cold terminations
are normally used in the noise calibration. To this end, impedance tuners that synthesize
the required impedance states are generally used and the noise parameters are then
extracted by means of fitting methods [25, 26, 40]. The terminations have to be adequately
distributed on the Smith-Chart in order to provide a well-conditioned set of equations
that allow the accurate computation of the four noise parameters [27–30].
The described cold-source technique requires a more complex measurement bench
than the scalar Y-factor method [43]: vector measurements, impedance tuning, switch-
ing circuitry, etc. Thus, the cold-source method was not the usual option for a standard
noise figure measurement of circuits or subsystems, but was mainly focused on noise
parameter extraction. Only recently, with the appearance of modern VNAs with noise-
figure measurement capabilities [40, 41], has attention been brought to this technique in
the context of circuit noise figure characterization. A different noise figure characteriza-
tion approach implemented in a VNA and based on digital data processing techniques
can also be found [44].
10.5 Common sources of error
The accuracy of a noise figure measurement depends on a wide variety of factors includ-
ing the characteristics of the DUT, the measurement setup and, obviously, the degree of
approximations included in the methodology. For instance, when coming to analyze a
scalar noise figure measurement technique as the classical Y-factor, mismatches in the
measurement path will impact the accuracy of the final result. There are other effects
that can have an influence on noise figure accuracy, such as measurement temperature,
receiver linearity, and bandwidth. There is also uncertainty associated with the limited
accuracy of the measurement instruments; the imperfect knowledge of the noise tem-
peratures; the incomplete knowledge of the correction terms used to remove systematic
errors; and random effects such as connector variability, jitter, etc. Evaluating the overall
uncertainty of a noise figure measurement is not an easy task because of the complex for-
mulae involved, especially in those methodologies that include a variety of corrections,
and it is not treated in this chapter. The fundamentals on measurement uncertainty and
guidance on numerical methods for its evaluation are given in [45] and [46], respectively,
while works particularly focused on noise measurement uncertainty can be found, for
example, in [47–50]. Accuracy and uncertainty issues have been increasingly treated by
noise figure measurement equipment manufacturers, as for example in [39, 40, 51–53].
Nowadays, specific uncertainty calculators for noise figure measurements are offered by
different manufacturers [54–56].
In this section, some basic sources of error in noise figure characterization are dis-
cussed. The analysis is divided into three main categories (mismatch, temperature, and
measurement setup). Each systematic effect is treated separately to extract unambiguous
conclusions. Where it is pertinent, the accuracy implications forY-factor and cold-source
are compared. Numerical examples are used to help visualize the main results of the
analyses. In addition, measurement examples are provided to confirm and illustrate the
basic conclusions. Only ordinary two-port devices are considered in this section, since
frequency translating devices are specifically addressed in Section 10.6.
10.5.1 Mismatch
In this section, only errors coming from mismatch are considered. Any other source of
error in the measurement is neglected. Obviously, mismatch has a significant impact on a
scalar methodology such as the classical implementation of Y-factor. Therefore this part
is mainly focused on the Y-factor technique. The effect of including correction terms for
mismatch systematic errors in Y-factor is also studied.
Y-factor and the second stage correction

The Y-factor technique is based on a scalar approximation (10.25) of the actual Friis
formula [8] for the DUT-receiver system, given in (10.33).

Frec (out ) − 1
NF DUT = 10 log10 Fsys − . (10.33)
Gav
It should be noted that (10.25) neglects any mismatch effect in the measurement path
and includes two main approximations to the Friis formula (10.33).
The first approximation concerns the available gain. The device available gain, whose
accurate characterization requires vector measurements, is substituted by the insertion
gain, directly obtained through scalar noise power measurements. The insertion gain is
defined as:
|1 − s11rec s |2 1
Gins = |s21 |2 . (10.34)
|1 − s11 s |2 |1 − s11rec out |2
If the noise source and the receiver are properly matched, which is usually the case, the
insertion gain tends to |s21 |2 . If, in addition, the device is properly matched the available
gain also tends to |s21 |2 . Hence, if the device is adequately matched, the insertion gain
will be a good approximation of the available gain. When this is not the case, both gains
can diverge significantly.
The second approximation concerns the receiver noise factor. During the calibration
step, the source termination presented to the receiver corresponds to the reflection coef-
ficient of the noise source, s . Thus, the measured receiver noise factor is Frec (s )
instead of Frec (out ), as required by the Friis formula. The noise source can be con-
sidered to be a fairly matched device. As a consequence, if the DUT has a poor output
match, the receiver noise factor measured in the calibration step, Frec (s ), may not be
a good approximation of the noise factor that the receiver actually has during the DUT
measurement step, Frec (out ).
The expression resultant from the classical scalar Y-factor implementation (10.25) is
given by (10.35). This expression shows the error associated with the calculation of the
noise figure with respect to the true noise figure of the DUT, NF DUT = 10 log10 (FDUT ).
It can be observed how (10.35) converges to the true value, NF DUT , for perfect match
conditions (out = s = 0).

Frec (out ) − 1 Frec (s ) − 1
NF YF = 10 log10 FDUT (s ) + − . (10.35)
Gav Gins
Therefore, although the Y-factor technique is a simple technique from the implemen-
tation point of view, the lack of vector measurements and the neglecting of the receiver
noise factor dependence on the source termination can significantly degrade the accu-
racy of the measured noise figure, especially when measuring poorly matched, low-gain
devices.
In [57], a Y-factor technique complemented with vector measurements is proposed.
From these additional vector measurements, the available gain of the device is computed
and included in the second stage correction. It is important to note that the receiver
noise factor required for this second stage correction is still obtained as in the classical
scalar Y-factor technique. Therefore, any dependence of the receiver noise on the source
termination is still neglected. This implementation can be formulated as

Frec (s ) − 1
NF YF_Gav_CORR ≡ 10 log10 Fsys − , (10.36)
Gav
in contrast to the fully scalar technique given in (10.25), where the device insertion gain is
used in the second stage correction. The expression that results from this implementation,
under the assumptions made, is:

Frec (out ) Frec (s )
NF YF_Gav_CORR = 10 log10 FDUT (s ) + − . (10.37)
Gav Gav
The possible benefits of the partial correction included in (10.36) are not guaranteed
and have to be carefully analyzed. As shown in [58], for low gain mismatched devices,
the second stage correction requires the knowledge of receiver noise parameters to be
efficient and rigorous. Applying vector corrections without an accurate knowledge of
the receiver noise factor may end up in poor accuracy, even worse than the basic scalar
approach.
Let us illustrate this analysis with the help of a numerical example.
Numerical example 1. The following setup is considered:

Noise source: ENR = 15 dB, sc = sh = 0.05∠−20º.
DUT: s11 = 0.08∠45º, s22 = variable ∠180º, s21 = 5 dB, s12 = −50 dB; NF = 3 dB.
Receiver: s11rec = 0.06∠170º; NF min_rec = 6 dB, Rn_rec = 50 , opt_rec = 0.07∠60º.
Let us define the error associated with each methodology as the difference between the
resultant noise figure and the true noise figure (eYF = NFYF − NFDUT , eYF_Gav_CORR =
NFYF_Gav_CORR − NFDUT ). In Figure 10.8(a) the errors associated with both techniques
are plotted as a function of device output match. As can be seen, both methodologies
become considerably inaccurate as this output match worsens. The error given by the
corrected Y-factor version is slightly lower than the error of the classical approach for
good DUT output match. However, as the output match degrades, eYF_Gav_CORR increases
more rapidly than eYF . This rapid increase is due to the combination of an available gain
that tends to infinity with a noise factor obtained at the calibration step that remains
constant.
It is clear from the nature of the second stage correction that the effect of any mismatch
at the output stage will decrease with increasing gain. A new analysis can be performed
as a function of the device gain to see its influence on the resultant errors. For this
second analysis a fixed output match of s22 = 0.5 is taken, while the magnitude of the
s21 parameter is varied. The errors computed from the classical and corrected Y-factor
versions are shown in Figure 10.8(b). As expected, the error provided by both techniques
tends to zero as the device gain increases.
Finally, it is important to highlight that the accuracy associated with both techniques
depends highly on the characteristics of the DUT and receiver.
Noise source match variations

The above analysis dealt with mismatch effects associated with the second stage correc-
tion. However, mismatch effects at the input stage of the DUT are also important in the
Y-factor technique, since they affect the very principle of the method itself. Indeed, the
formulation of the Y-factor methodology relies on the basis that there is a single device
noise factor during the measurement process. However, if the variation of the reflection
coefficient of the noise source from the cold to the hot state is not negligible, the device
noise factor will actually vary [59], [60].
Let us come back to the basic Y-factor concept presented in Section 10.2.2, where no
consideration of the source reflection coefficient has been made. The cold and hot noise
powers given in (10.14) and (10.15) depend on the reflection coefficient of the source
1
eYF
0.8 eYF_Gav_CORR
Error (dB)
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
|s22|
(a)
2
eYF
eYF_Gav_CORR
1.5
Error (dB)
0.5
0
0 5 10 15 20 25 30
|s21| (dB)
(b)
Fig. 10.8 Errors associated with classical and corrected Y-factor methodologies as a function of: (a) DUT
output match (s21 = 5 dB); (b) DUT gain for a given s22 of 0.5.
termination at each temperature, as explicitly shown in (10.38) and (10.39)
Nc (sc ) = kBGav (sc ) [Tc + T0 (F (sc ) − 1)] (10.38)

Nh (sh ) = kBGav (sh ) [Th + T0 (F (sh ) − 1)] , (10.39)
where sc and sh are the reflection coefficients of the noise source in its cold and
hot states.
Hence, the noise factor computed from these noise powers will accordingly depend on
the device noise factors corresponding to both source reflection coefficients. Neglecting
any other sources of error, the noise figure calculated from theY-factor expression (10.17)
will be actually given by (10.40), which is a function of the device gains and noise factors
at the two noise source states.
NFYF = 10 log10
6
Gav (sc ) Th T0 − 1 F (sc )
6 . (10.40)
(Gav (sh ) F (sh ) − Gav (sc ) F (sc )) + Gav (sh ) Th T0 − 1
Only when changes in the reflection coefficient of the noise source are negligible
(sh sc = s ), does the noise figure characterized by (10.40) converge to NF (s ):
NF YF (sh sc = s ) 10 log10 (F (s )) = NF (s ) (10.41)
It should be noted that in (10.38) and (10.39) both the device available gain and noise
factor change with the source reflection coefficient. Therefore, inaccuracies associated
with discrepancies between sc and sh come from both the match and noise variations
of the DUT. Obviously, the magnitude of the error depends on the amount of change in
the noise source and on the intrinsic properties of the DUT, given by its gain and noise
characteristics.
A correction factor can be applied to the Y-factor to deal with the variations in the
noise source reflection coefficient [57]:
2
1 − |sc |2
|1 − s
11 sh |2
1 − out_h

Nh
YCORR = , (10.42)
Nc 1 − | |2 |1 − s |2 1 −
sh 11 sc out_c
2
where out_c and out_h are, respectively, the output reflection coefficients of the DUT
in the cold and hot measurements.
Nevertheless, it is clear from (10.38) and (10.39) that this factor can only correct for
match variations. With this correction applied, the resultant noise figure leads to (10.43),
which is still a function of the DUT noise factors for the two reflection coefficients
F (sc ) and F (sh ).
6
Th T0 − 1 F (sc )
NFYF_Y _CORR = 10 log10 6 . (10.43)
(F (sh ) − F (sc )) + Th T0 − 1
In (10.43), match variations from cold to hot measurement have been eliminated and
this will in general lead to an improvement in the overall accuracy. Nonetheless, the
variations in the device noise factor itself due to source termination changes cannot be
taken into account unless the four noise parameters of the device are fully determined.
Numerical example 2. Let us illustrate the above discussion with an example. For the example,
the following characteristics have been considered:
Noise source: ENR = 8 dB, sc = 0.22, sh = 0.19∠−30º.
DUT: s11 = variable ∠ − 60º, s22 = 0, s21 = 40 dB, s12 = −50 dB; NF min = 1 dB, Rn = 20 ,
opt = 0.45∠200º.
These match and ENR values could be realistically assigned to a Q347B millimeter wave
2
eYF
1.5 eYF_Y_CORR
Error (dB)
0.5
0
0 0.2 0.4 0.6 0.8 1
|s11|
Fig. 10.9 Errors resultant from NFYF and NFYF_Y _CORR versus DUT input match.
noise source from Agilent. Second-stage noise is not included. We take NF (sc ), the noise
figure corresponding to the cold state, as the true value. Then, the errors associated with NFYF
and NF YF_Y _CORR have been computed as the difference between the values obtained with
(10.40) and (10.43), respectively, and the true value. The results are plotted in Figure 10.9 as
a function of the DUT input match.
As can be seen in Figure 10.9, NFYF presents an increasing error as the DUT input
match worsens. In contrast, eYF_Y _CORR does not depend on DUT s11 , but a constant non-
negligible error remains. This is due to the variation of the device noise figure between
cold and hot measurements, as shown by (10.43). For the considered DUT, the errors
associated with the typical 346 noise source family from Agilent would be significantly
lower (in particular, eYF_Y _CORR will be negligible).
A more involved measurement strategy that takes into account noise source variations
in the Y-factor technique can be found [61]. For that, a complete noise characterization
of the device that obtains its four noise parameters is required.
In order to minimize mismatch-related effects, the use of isolators or attenuator pads
is recommended, although this solution has its own drawbacks [39]. The inclusion of iso-
lators limits the frequency range and several isolators may be required to cover the entire
band in a wideband measurement. In addition, a rigorous characterization of the influ-
ence of the isolator requires a vector characterization. If several isolators are required,
this process will accordingly enlarge. In contrast, attenuators have broadband response.
However, an attenuator reduces the ENR presented to the DUT by its insertion loss,
requiring a vector correction for its accurate characterization.
Cold-source
The cold-source technique presents a significant advantage over the Y-factor when deal-
ing with mismatch-related errors, since in this technique the device noise figure is
measured for a single-source impedance state. As a consequence, any inaccuracy related
to noise source reflection coefficient variations can be avoided in the cold-source proce-
dure. It should be noted that such variations affect obtaining the gain-bandwidth product
of the receiver, because this term is generally characterized from cold/hot measurements
[60]. However, if a fully corrected procedure is considered, as the one described by
(10.26), these variations can be properly accounted for. Furthermore, in this fully cor-
rected procedure any mismatch error associated with the second stage is eliminated at
the cost of a substantial increase in measurement complexity compared with the classical
Y-factor technique.
Measurement example
Let us illustrate the effect of mismatch on noise figure characterization by means of a
measurement example. To this end, measurements of a mismatched passive device are
provided. This kind of device represents a challenging test in this context, because the
lack of gain magnifies the mismatch effects in the output stage. Also, note that the true
noise factor of a passive device can be calculated analytically from its S-parameters as the
inverse of the available gain. The DUT has been built up by combining an attenuator with
a mismatch block. Figure 10.10(a) shows the output return loss of the DUT, while the
available gain is depicted in Figure 10.10(b). The measurements were carried out in an in-
home setup specifically conceived to implement different characterization approaches
(Y-factor and cold-source with different levels of corrections). The setup includes a
PNA E8358A, a low noise preamplifier, and a commercial 346B noise source.
Figure 10.11 compares the noise figures measured in a 1–2 GHz frequency range
through four different characterization approaches. NFYF is the classical scalar Y-factor
technique (10.25). NFCS is the cold-source method (10.26), with vector corrections and a
full noise receiver calibration. NFYF_Gav_CORR represents the partially corrected Y-factor
approach (10.36), which makes use of the DUT available gain instead of the insertion
gain. Finally, NFYF_CORR is a fully corrected version of the Y-factor technique, including
a correction for variations in the noise source match (10.42) and a noise calibration of
the receiver. The true DUT noise figure NFDUT computed from S-parameters is also
depicted in Figure 10.11. As can be seen, the scalar Y-factor technique NFYF cannot
equal the accuracy provided by the cold-source NFCS for this demanding DUT, because
of the lack of vector corrections and receiver noise calibration. In contrast, the fully cor-
rected Y-factor NFYF_CORR presents accuracy comparable to the cold-source technique,
as expected from the comparable level of corrections included in the methodology. The
residual effect of noise source reflection coefficient variations is negligible with the 346B
noise source in this case. Finally, it should be noted that the highest error corresponds to
the partially corrected Y-factor version NFYF_Gav_CORR .
10.5.2 Temperature effects

A usual approximation when making noise figure measurements is that of considering
the cold temperature Tc to be equal to the reference temperature T0 (290 K). However, Tc
is normally the ambient temperature and does not in general agree with the reference T0 .
As previously mentioned, a correction can be included in modern noise figure measure-
ment instruments to account for this difference. Moreover, current noise sources include
–2
DUT output return loss (dB)
–3
–4
–5
–6
–7
–8
1 1.2 1.4 1.6 1.8 2
Frequency (GHz)
(a)
–1
–1.5
DUT Gav (dB)
–2
–2.5
–3
1 1.2 1.4 1.6 1.8 2
Frequency (GHz)
(b)
Fig. 10.10 (a) DUT output return loss; (b) DUT available gain.
temperature sensors that can measure their own temperature [62] and automatically per-
form the temperature compensation. If differences between Tc and T0 exist and are not
corrected, an error is introduced in the measurement performed with both the Y-factor
and cold-source techniques. If no other systematic effect is present and the second stage
contribution is negligible, the noise figure computed by both techniques is

Tc
NFYF = NFCS = 10 log10 F+ −1 (10.44)
T0
where F is the true noise factor.
Numerical example 3. Figure 10.12 shows the error given by (10.44) for a range of Tc
between 280 K and 300 K as a function of the noise figure. It is clear from Figure 10.12 that
approximating Tc to T0 is acceptable for DUTs with NF > 5 dB.
8
NFDUT=-Gav (dB)
7 NFYF
NFCS
6
Noise figure (dB)
NFYF_Gav_CORR
NFYF_CORR
5
1
1 1.2 1.4 1.6 1.8 2
Frequency (GHz)
Fig. 10.11 True DUT noise figure and noise figures characterized with Y-factor (scalar, partially corrected,
and fully corrected) and cold-source in 1–2 GHz frequency range.
0.3
0.2
0.1
Error (dB)
–0.1 Tc = 280 K
Tc = 285 K
Tc = 290 K
–0.2 Tc = 295 K
Tc = 300 K
–0.3
0 5 10 15
NF (dB)
Fig. 10.12 Errors due to Tc = T0 as a function of actual noise figure for Y-factor, according to [39], and
cold-source.
Measurement example
Figure 10.13 shows the measurement results obtained from Y-factor and cold-source
techniques with and without cold temperature correction. The same measurement setup
as in Section 10.5.1 is used. The cold temperature Tc is the ambient temperature, 298 K
(25◦ C) in this case. The DUT is a low noise amplifier with an approximately 30–25 dB
gain in the 1–2 GHz frequency range. As expected from the analysis, the non-corrected
approaches are approximately 0.1 dB over the corrected ones.
10.5.3 Measurement setup

Until this point mismatch and temperature-related errors have been analyzed. We dis-
cuss here three important points related to the measurement setup characteristics that
1.5
Noise figure (dB)
1
NFYF (Tc corrected)
0.5 NFCS (Tc corrected)
NFYF
NFCS
0
1 1.2 1.4 1.6 1.8 2
Frequency (GHz)
Fig. 10.13 Y-factor and cold-source noise figure measurements of an LNA with and without Tc correction
(Tc = 298 K).
can have a non-negligible influence on the final accuracy of the measurement results:
selection of the ENR, receiver bandwidth, and receiver linearity. Moreover, their impact
on measurement accuracy is different depending on the technique we use for the noise
figure calculation: Y-factor or cold-source.
Noise source ENR selection

Selecting an adequate ENR can be important for an accurate noise figure measurement.
This selection depends on the measurement technique. For theY-factor technique, several
reasons recommend the use of a low ENR [39], unless high noise figures are to be
measured. First, a low ENR reduces the possibility of driving the receiver beyond its
linear region. Also, even if the receiver maintains its linear response, higher ENR values
may require internal attenuation that increases the receiver noise factor, thus reducing
the measurement accuracy. In addition, reflection coefficient variations from the cold
to the hot state are in general smaller in a lower ENR noise source, because of built-in
attenuators. Therefore, inaccuracies due to source reflection coefficient variations are
reduced, recalling (10.40) and (10.43).
In contrast, higher ENR values are better suited to the cold-source technique. The use
of the noise source, necessary for accurately obtaining the kB|s21rec |2 term, is limited
to the calibration step in this technique; thus, receiver compression should not be a
problem. When measuring the DUT, the noise power involved is generally much higher
than that of a calibration performed with a low ENR noise source. The kB|s21rec |2
characterized in that case magnifies the overall uncertainty of the calculated noise figure.
However, a higher ENR allows a calibration in a larger dynamic range, closer to the DUT
measurement level and thus providing lower uncertainty. In addition, greater reflection
coefficient variations associated with a high ENR are not as critical as in the Y-factor
technique since they only affect the characterization of the noise receiver and, thus, can
be corrected as mentioned earlier.
0.8
|s21|=5 dB
|s21|=20 dB
0.6
Uncertainty (dB)
0.4
0.2
0
0 5 10 15
ENR (dB)
Fig. 10.14 Uncertainty given by cold-source as a function of ENR for a 2% uncertainty in measured noise
powers and 0.2 dB uncertainty in ENR for two different DUT gains.
Numerical example 4. The uncertainty increase associated with a low ENR is illustrated in
this example. Let us consider the following setup:
Noise source: ENR = variable dB, s = 0.
DUT: s11 = 0, s22 = 0, s21 = 5/20 dB, s12 = −50 dB; NF = 3 dB.
Receiver: s11rec = 0; NFrec = 6 dB.
For the analysis a 2% uncertainty due to jitter is assigned to the measured noise powers
and a typical 0.2 dB uncertainty is assigned to the noise source ENR. Figure 10.14 shows
the uncertainty associated with the noise factor characterized by means of the cold-source
technique as a function of the ENR of the noise source for the two DUT gains considered. As
can be seen, the resultant uncertainty (standard deviation of the result) increases for decreasing
ENR values. This uncertainty increase is slightly magnified by the DUT gain.
Receiver bandwidth
The internal bandwidth of a classical noise figure instrument such as the HP8970 is about
4 MHz. Current noise figure analyzers have variable bandwidths that can be reduced much
further [34]. Selecting an adequate receiver bandwidth is fundamental for measurement
accuracy. During the calibration step the total noise power within the bandwidth of
the receiver is measured. In contrast, if the DUT (or the combination DUT plus noise
receiver) has a bandwidth narrower than the receiver itself, the noise bandwidth will
be restricted by the presence of the DUT during the measurement and thus, errors can
arise (see Figure 10.15). This is a situation that can typically happen when measuring at
the passband edge of a very frequency-selective DUT. For the sake of clarity, let us call
Bcal and Bmeas , respectively, the noise bandwidths during calibration and during DUT
measurement.
Neglecting any systematic error other than the bandwidth variation, the noise figure
obtained from a Y-factor technique can be approximated to (10.45), where FDUT is the
true DUT noise factor. In this technique, the error becomes insignificant if the gain of
the DUT is significantly larger than the ratio Bcal /Bmeas [39].

Bcal 1
NFYF ≈ 10 log10 FDUT + −1 . (10.45)
Bmeas Gav
BRec BRec
N N
BDUT BDUT
f f
Bcal Bcal
Bmeas Bmeas
(a) (b)
Fig. 10.15 Two possible error sources: (a) DUT bandwidth narrower than receiver bandwidth; (b)
measurement in DUT passband edge.
As in the Y-factor case, the fact of having a Bmeas < Bcal is also a source of error in
the cold-source technique. Nevertheless, there is a significant difference in the impact
of this error from one technique to the other. In the cold-source case, any bandwidth
difference between calibration and measurement affects the resultant noise figure, no
matter the value of the DUT gain. The noise figure measured with cold-source technique,
neglecting again any other source of systematic error, can be approximated to (10.46).

Bmeas Bmeas 1
NFCS ≈ 10 log10 FDUT + 1− . (10.46)
Bcal Bcal Gav
Numerical example 5. This example serves to visualize the differences between the errors
associated with both techniques. Characteristics of the setup are:
Noise source: ENR = 15 dB, s = 0.
DUT: s11 = 0, s22 = 0, s21 = 10 dB, s12 = -50 dB; NF = 3 dB.
Receiver: s11rec = 0; NFrec = 6 dB. Bandwidths: Bmeas /Bcal = variable/0.5.
For each technique (Y-factor and cold-source) an error function is calculated as the difference
between the computed noise figure and the true one. The error is calculated as a function of
the bandwidth ratio with a fixed s21 of 10 dB. The errors obtained are plotted in Figure 10.16.
As shown in this figure, the error associated with the cold-source technique is significantly
larger. This is because in the Y-factor case the error is attenuated by the DUT gain, while in
the cold-source case the error tends to the bandwidth ratio Bmeas /Bcal (in dB) as the DUT
gain increases. Obviously, both errors disappear if Bmeas = Bcal .
Receiver linearity
A noise figure measurement relies on the linearity of the whole measurement system, as
is clear from Figure 10.1(b). If the noise powers involved are high enough to drive the
receiver into compression, the computed noise figure will not be accurate. As previously
stated, the use of a low ENR is good practice to avoid the nonlinear behavior of the
receiver in the Y-factor technique. However, if the DUT gain is high, in-line attenuation
after the DUT may also be required. If this is the case, a correction has to be applied to
0
Error (dB)
–1
–2 eYF
eCS
–3
0.5 0.6 0.7 0.8 0.9 1
Bmeas/Bcal
Fig. 10.16 Errors arising from a measurement bandwidth narrower than calibration bandwidth as a function
of bandwidth ratio Bmeas /Bcal .
the measurement to eliminate the contribution of the attenuator. It should be noted that,
as long as it is not included in the calibration step, an accurate characterization of the
attenuator requires vector measurements [39].
Deviations from linear behavior affect the noise figure calculation for the Y-factor as
well as for the cold-source technique. Again, errors vary depending on the technique.
These errors depend on the DUT characteristics, the ENR value, and the compression
curve of the receiver. In contrast to the bandwidth discussion, here no general conclusion
can be easily extracted about which technique becomes less accurate when a linearity
deviation is taking place during the measurement process. However, the cold-source
technique is less susceptible to driving the receiver into its nonlinear range because of
the lack of a hot noise power measurement.
(10.47) and (10.48) are, respectively, the approximated noise figures computed from
theY-factor and cold-source techniques when there is a linearity deviation in the receiver.
In these expressions Cc and Ch are compression factors (typically 0 < Ch ≤ Cc ≤ 1)
so that the actual measured cold and hot noise powers are Cc Nc and Ch Nh , instead of
the ideals Nc and Nh (see Figure 10.17). Any systematic effect that is different from
the compression of the receiver has been neglected in these expressions. In the Y-factor
case, the noise figure will be overestimated because of the reduction in the denominator
of (10.47) due to Ch < Cc . Note that if both measurements presented equal compression
factors (Ch = Cc ) there would be no error in the Y-factor case. However, in the cold-
source technique, the noise figure will be underestimated if the measured noise power is
compressed.
6
Th T0 − 1
NFYF ≈ 10 log10 6 , (10.47)
Ch N h C c N c − 1

Cc Nc
NFCS ≈ 10 log10 . (10.48)
kB|s21rec |2 Gav T0
Nh
ChNh
Nc
CcNc
Tc Th T
Fig. 10.17 Basic diagram of compression in receiver: ideal noise powers (Nc , Nh ) and compressed noise
powers (Cc Nc , Ch Nh ) with Ch < Cc .
Measurement example
This example illustrates the effect of compression in the noise receiver. Again, the
same measurement setup described in Section 10.5.1 is used. Figure 10.18(a) shows
the noise response of this receiver as a function of its input noise. Compression for high
noise-power levels is clearly noticeable. Y-factor and cold-source measurements of a
variable-gain amplifier are performed. High gain values of the DUT bring the receiver
into compression. The measurement results are plotted in Figure 10.18(b). To compare
with a valid reference, the noise figure of the DUT, NFDUT , was measured with a linear
receiver that avoids compression through the use of input attenuators (superimposed in
Figure 10.18(b)). As previously analyzed, the Y-factor technique tends to overestimate
the noise figure because of a larger compression in the hot measurement. Indeed, the
computed noise figure is 5 dB over the reference one for a 36 dB DUT gain. In contrast,
for such gain the noise figure provided by the cold-source technique is approximately
0.2 dB below the reference NFDUT , due to a compressed cold measurement.
10.6 Noise figure characterization of mixers
Mixers have some particular characteristics that complicate obtaining accurate noise
figure measurements. They often present a poor output match (generally worse than
amplifiers) and can have losses instead of gain (diode-based and cold-FET mixers). In
addition, other effects specific to frequency translation appear. Besides, although mixer
noise theory was developed early [63]–[71], some degree of confusion has accompanied
mixer noise figure formulation from the very beginning, as was already pointed out in
[68]. In this section the noise figure definition specifically provided by the IEEE for
frequency translating devices [11] is analyzed. The definition and significance of the
single-sideband (SSB) noise figure of a mixer are revisited. Obtaining the SSB noise
figure through the Y-factor and cold-source techniques is comparatively discussed [72].
–10
Measured
–20 Linear
No (dBm)
–30
–40
–50
–60
–160 –150 –140 –130 –120
kBT (dBm)
(a)
10
NFDUT
9 NFYF
Noise figure (dB)
8 NFCS
4
10 15 20 25 30 35 40
|GDUT| (dB)
(b)
Fig. 10.18 (a) Compression curve of noise receiver. (b) Y-factor, cold-source, and DUT noise figures
versus gain.
10.6.1 Noise figure definitions for frequency translating devices

According to [11], the noise factor of a frequency translating device can be expressed
mathematically as:
No
F= . (10.49)
kBT0 Gav
where, No is the total noise power available at the output port at the output frequency
when the noise temperature of its input termination is T0 = 290 K at all frequencies.
kBT0 Gav is the portion of No that is engendered by the input termination at temperature
T0 at the input frequency/frequencies. It is important to note that in the denominator
of (10.49) only the contribution via signal-frequency transformation(s) is included. All
Ni No
G1av G3av Nadd

GavIF
G2av G4av
IF
kT0
IF LO 3LO f IF f
RF
Fig. 10.19 Basic diagram of noise behavior in a mixer. No includes noise converted from principal, image,
and idler frequencies, as well as noise added by the mixer.
other contributions, i.e. contributions from frequency conversions where the signal is not
present in operating conditions, are excluded [11].
Single-sideband, double-sideband, and all-sideband noise figure

Considering that, in the normal operation of a heterodyne system, the signal is only
present at a single frequency, Gav in (10.49) is simply Gav = G1av , with G1av being the
available conversion gain relating this single input frequency to the output one. Therefore,
according to the IEEE definition, the SSB noise factor of a mixer is given by (10.50).
No
FSSB = . (10.50)
kBG1av T0
It should be noted that this definition does not exclude from No any noise generated at
image or idler frequencies. On the contrary, it excludes from the denominator any gain
different from G1av , i.e. any gain not corresponding to a signal-frequency transformation.
Thus, No , given in (10.51), includes contributions from every possible conversion
from input to output, as schematically shown in Figure 10.19.
No = kB (G1av + G2av + · · · + Gnav ) T0 + Nadd . (10.51)
In (10.51) the Gj av terms represent each possible available conversion gain from input
to output: G1av , as already defined, is the principal available conversion gain, which
relates the input RF frequency to the output one; G2av is the image available conversion
gain; G3av ,…, Gnav represent the available gains associated with idler conversions.
Finally, Nadd is the noise added by the mixer (including white noise coming from LO
port). Analogous to the available gain of an ordinary two-port device (10.5), the available
conversion gains Gj av can be defined as:
2
1 − s fj 2 1
Gjav = 2 c21 fj , fIF , (10.52)
1 − s fj s11 fj 1 − |out (fIF )|2

where s fj and s11 fj are, respectively, the source and input reflection coefficients
at input frequency fj . out (fIF ) is the output reflection coefficient of the mixer at IF

frequency. Finally, c21 fj , fIF is a conversion parameter from input to output frequency,

analogous to the standard s21 S-parameter. Both s11 fj and out (fIF ) must be obtained
under operating conditions, i.e. with the LO power at its operating level, to correctly
describe the mixer behavior.
The definition of the SSB noise factor (10.50) is completely consistent with considering
the noise factor as a figure of merit that characterizes the degradation of the signal-
to-noise ratio from the input to the output of the device when operating in the SSB
heterodyne mode.
When the denominator of (10.49) includes noise contributions from every possible
transformation, an all-sideband (ASB) [73] noise factor is obtained, given in (10.53).
No
FASB = . (10.53)
kB (G1av + G2av + · · · + Gnav ) T0
It is usually assumed that idler contributions are negligible compared to the principal
and image contributions. If this is the case, the all-sideband noise factor equals the
double-sideband (DSB) noise factor, as defined in (10.54), where only the available
gains corresponding to RF and image conversions are considered in the denominator.
Obviously, when the system operates in DSB (as in receivers for radiometry applications
or in zero-IF receivers), the figure of merit that characterizes the degradation of the SNR
is FDSB .
No
FDSB = . (10.54)
kB (G1av + G2av ) T0
It can be directly deduced from (10.50) and (10.53) that the SSB noise factor is equal
to the ASB noise factor magnified by the quotient of the sum of all available conversion
gains contributing to the output over the principal available conversion gain, as shown by
(G1av + G2av + · · · + Gnav )

FSSB = FASB . (10.55)
G1av
Let us now consider a noise figure measurement Nc at a temperature Tc different from

the reference temperature T0 . Then, the SSB noise factor is given by:
Nc (Tc − T0 ) (G1av + G2av + · · · + Gnav )

FSSB = − . (10.56)
kBG1av T0 T0 G1av
It is clear from (10.56) that if Tc = T0 a correction factor that includes all the available
gains, is required. Note that in this case the temperature difference is magnified by the
gain ratio, in contrast to ordinary two-ports devices.
Noise temperature in mixers

Sometimes the noise temperature is used instead of the noise figure for characterizing the
noise behavior of a mixer [6]. Let us consider a mixer from a system point of view. For the
sake of simplicity, only the principal and image conversions are included. Figure 10.20
RF port RF port
G1av IF port G1av IF port
+ +
T0 TSSB No T0 TDSB No
+
G2av G2av
T0 T0 TDSB
Image port Image port
Fig. 10.20 Basic diagram of SSB and DSB noise temperature concepts.
shows a typical representation of SSB and DSB input noise temperature concepts. In
such a diagram, the input RF frequency and the image frequency are treated as separate
input ports.
In the SSB noise temperature, TSSB , all the noise generated by the mixer is translated to
the RF input port. In doing so, the noise available at the output can be given as a function
of TSSB by means of (10.57). In contrast, when referring to the DSB noise temperature,
TDSB , the total noise generated in the mixer is translated to both input ports, i.e. to RF
and image ports. In this case, the output noise can be written as in (10.58).
No = kB [G1av (T0 + TSSB ) + G2av T0 ] = kB [G1av TSSB + (G1av + G2av ) T0 ] (10.57)

No = kB [G1av (T0 + TDSB ) + G2av (T0 + TDSB )]
= kB [(G1av + G2av ) TDSB + (G1av + G2av ) T0 ] . (10.58)
Both concepts are simply a translation to the input (to RF port for SSB and to RF and
image ports for DSB) of the noise added by the mixer:
Nadd
TSSB = (10.59)
kBG1av
Nadd
TDSB = . (10.60)
kB (G1av + G2av )
As a consequence of (10.59) and (10.60), the SSB input noise temperature is equal
to the DSB input noise temperature multiplied by the ratio of the sum of principal and
image conversion gains over the principal conversion gain, as in the noise factor case.
(G1av + G2av )
TSSB = TDSB . (10.61)
G1av
We can also express the noise factor in terms of these input noise temperatures. If the
mixer added noise is rewritten in terms of TSSB , the relationship between the SSB noise
factor and the SSB input noise temperature can be achieved:
TSSB (G1av + G2av )

FSSB = + . (10.62)
T0 G1av
Equation (10.62) is the direct result of the application of the IEEE noise factor def-
inition, although other interpretations can be found [74]. If the principal and image
conversion gains can be considered to be equal, (10.62) simplifies to:
TSSB
FSSB ≈ + 2. (10.63)
T0
In a similar way, the relationship between the DSB noise factor and the DSB input
noise temperature can be found.
TDSB
FDSB = + 1. (10.64)
T0
Finally, let us generalize the previous analysis to include all mixer responses and let
us consider an effective input noise temperature Te common to all these responses [11].
In doing so, the noise available at the output is:
No = kB [G1av (T0 + Te ) + G2av (T0 + Te ) + · · · + Gnav (T0 + Te )] . (10.65)
Then the SSB noise factor can be written in terms of Te as (10.66), which simplifies
to (10.67) when idler conversions are negligible.

Te (G1av + G2av + · · · + Gnav )
FSSB = +1 (10.66)
T0 G1av

Te (G1av + G2av )
FSSB ≈ +1 . (10.67)
T0 G1av
Equation (10.68) relates the ASB noise factor to the effective input noise temperature
Te . It should be noted that this relationship applies to the DSB noise factor of a mixer
with negligible idler conversions, as shown by (10.64).
Te
FASB = + 1. (10.68)
T0
10.6.2 Obtaining the SSB noise figure from Y-factor and cold-source
As previously mentioned, typical noise figure meters such as the classical HP8970 and
derived implementations (including spectrum analyzers with noise measurement capa-
bilities) use the Y-factor technique to characterize the noise figure of circuits, including
mixers. In the Y-factor technique, the noise figure is characterized from two noise power
measurements. If the DUT is a mixer, both noise powers include noise contributions
from the image and idler frequencies, because the noise source is a broadband device
that provides extra noise in a wide frequency range.
Considering Tc = T0 for simplicity, and analogous to (10.14) and (10.15), the cold
and hot noise powers corresponding to a frequency converter are, in that order, (10.69)
and (10.70).
Nc = kB (G1av + G2av + · · · + Gnav ) T0 + Nadd (10.69)

Nh = kB (G1av + G2av + · · · + Gnav ) Th + Nadd . (10.70)
Then, applying the Y-factor expression (10.17), the obtained noise factor is an ASB
noise factor:
6
Th T0 − 1 (Th − T0 ) Nc
F= 6 =
Nh Nc − 1 T0 (Nh − Nc )
Nc
= = FASB . (10.71)
kB (G1av + G2av + · · · + Gnav ) T0
In (10.69) and (10.70) no noise contribution of the receiver has been considered. In
fact, in an actual measurement this contribution has to be eliminated, as usual, applying
the second-stage correction:
6
Th T0 − 1 Frec − 1
NFYF = 10 log10 6 − , (10.72)
Nh N c − 1 Gins
where Frec and Gins are computed from the standard calibration step, (10.23) and (10.24),
respectively.
In order to obtain the SSB noise figure from aY-factor technique, the following assump-
tions are often made [6]: image conversion is equal to principal conversion (G1av = G2av )
and all idler conversions are negligible (G3av = … = Gnav = 0). Then, the SSB noise
figure is considered to be simply 3 dB higher than the measured one:
NFSSB_YF = NFYF + 3 dB. (10.73)
Obviously, the above assumptions are not always satisfied. A common approach for
obtaining a “true” SSB noise figure measurement through a Y-factor technique includes
a filter at the input of the device that filters out image and idler frequencies. However,
impedance terminations of the mixer input port at the image and idler frequencies can
have a non-negligible influence on the device noise performances [73]. Therefore, if the
filter is not required for regular operation of the device, some amount of error should be
expected in the noise figure characterization.
In [75], which is a noise figure measurement implementation on a spectrum or signal
analyzer, the possibility of including a correction factor to the noise figure measurement
as a function of image rejection, instead of a fixed value of 3 dB, is provided. However,
this correction factor is only an estimate, because the actual gain ratio of the mixer, i.e.
the ratio of the sum of all available conversion gains over the principal conversion gain,
is not characterized.
Inaccuracies are not only related to the ASB to SSB translation. As in ordinary two-
port devices, there may be other sources of systematic error in the measurement. Indeed,
mixers can present worse measurement conditions for noise characterization than two-
port devices. On the one hand, most mixers are passive devices with usually poorer
matching conditions than two-port devices [76]. On the other hand, since conversions
from many frequencies might be involved, restrictions to usual approximations (Tc ≈ T0 ,
c ≈ h , etc.) are tighter than for the standard two-port devices.
In contrast to the Y-factor technique, cold-source is a straight implementation of the
SSB formulation. For that, a cold noise power measurement with a matched load con-
nected at the input of the mixer is performed, in the standard manner of the cold-source
approach. In addition, and according to (10.50), the principal available conversion gain
G1av has to be characterized. To obtain it, conversion characteristics as well as the input
and output matches of the device must be characterized (recall (10.52)). The contribution
of the noise receiver must be eliminated by applying the second-stage correction.
The total noise power measured by the noise receiver is:
Nc = kB|s21rec |2 MM (out )
× [T0 G1av FSSB + (Tc − T0 ) (G1av + G2av + · · · + Gnav )]
+ kB|s21rec |2 MM (out ) T0 (Frec (out ) − 1) , (10.74)
where out is the output reflection coefficient of the mixer and MM (out ) is the mismatch
between the mixer and the receiver, as given by (10.27). Note that these quantities are
obtained at fIF .
Therefore, the SSB noise figure of the mixer can be obtained from

Nc Frec (out ) − 1
NFCS ≡ 10 log10 −
kB |s21rec | MM (out ) G1av T0
2 G1av

(Tc − T0 ) (G1av + · · · + Gnav )
− , (10.75)
T0 G1av
where the kB|s21rec |2 term is characterized in the calibration step at fIF .

Finally, it is important to note that the correction term for Tc = T0 requires the knowl-
edge of the overall available gain, i.e. the sum of the available gains corresponding to all
significant conversions (G1av + · · · + Gnav ).
Measurement example
Y-factor and cold-source SSB noise figure measurement results of three diode-based
mixers are compared in this section. The three mixers have different gain and match
characteristics but the same IF frequency. The measurement setup includes a spectrum
analyzer (PSA E4440), a vector network analyzer (PNA E8358A), a 346B commer-
cial noise source, and two signal generators to measure the necessary gain, match, and
noise powers. The characterization is performed versus LO power since this power can
affect the noise generated by a diode-based mixer [77]. In the calculations, Tc has been
realistically approximated to T0 .
Let us first consider the cold-source procedure defined by (10.75). When the difference
between Tc and T0 is negligible (10.75) tends to (10.76).

Nc Frec (out ) − 1
NFCS ≡ 10 log10 − . (10.76)
kB |s21rec |2 MM (out ) G1av T0 G1av
For the SSB noise figure measurement with theY-factor technique (labeled as NFYF+3 ),
(10.73) is used, where the SSB noise figure is calculated by simply adding 3 dB to the
scalar Y-factor result.
In addition to that, a scalar version of the cold-source technique is also considered for
comparison. NFCS_SCALAR is given

Nc Frec − 1
NFCS_SCALAR ≡ 10 log10 − , (10.77)
kB |s21rec |2 G1ins T0 G1ins
where G1ins is the principal insertion gain measured in a spectrum analyzer and no noise
calibration of the receiver is considered. This scalar approach is a simple and fast solution
to implement in a spectrum analyzer when good match conditions are satisfied.
The results obtained for the three mixers are plotted in Figure 10.21. The first mixer
under test, Mixer 1, (RF = 0.3 GHz, LO = 1.3 GHz, IF = 1 GHz) presents comparable
conversion losses for principal and image frequencies (G1av ≈ G2av ) and conversions
from idler frequencies are negligible (G3av + · · · + Gnav ≈ 0). In addition, output return
losses are better than −10 dB in the measurement range. According to the properties of
the mixer, the three noise figure calculations (NFCS , NFYF+3 and NFCS_SCALAR ) lead to
similar results (Figure 10.21(a)). For this mixer, NFYF+3 provides a good approximation
of the SSB noise figure, due to the favorable match and conversion characteristics of
the device. In addition, no vector corrections or receiver noise calibration are necessary
because of its good match.
Let us now analyze Mixer 2 (RF = 2 GHz, LO = 3 GHz, IF = 1 GHz). In this case, the
sum of gains corresponding to image and idler conversions (G2av + G3av + · · · + Gnav )
is larger than the principal conversion gain G1av . In addition, the output return losses
are again better than −10 dB in the entire measurement band. The measurement results,
plotted in Figure 10.21(b), are again consistent with the characteristics of Mixer 2.
As can be seen, NFYF+3 underestimates the SSB noise figure, as expected from its
conversion losses. However, the two cold-source approaches, NFCS_SCALAR and NFCS
provide identical results due to the good output match of Mixer 2. This result shows that
NFCS_SCALAR can provide a good estimation of the SSB noise figure of mixers with fair
match characteristics.
Finally, the results obtained for the third mixer, Mixer 3 (RF = 2 GHz, LO = 3 GHz,
IF = 1 GHz), are given in Figure 10.21(c). The mixer presents different principal and
image conversion losses and it is poorly matched at the output port (worse than −5 dB in
the whole measurement range). As shown in Figure 10.21(c), three different responses
20 20
NFCS NFCS
Noise figure (dB) 19 NFYF + 3 19 NFYF + 3
Noise figure (dB)

NFCS _ SCALAR NFCS _ SCALAR
18 18
17 17
16 16
15 15
14 14
4 6 8 10 12 14 4 6 8 10 12 14
LO power (dBm) LO power (dBm)
(a) (b)
20
NFCS
19 NFYF + 3
Noise figure (dB)
NFCS _ SCALAR
18
17
16
15
14
4 6 8 10 12 14
LO power (dBm)
(c)
Fig. 10.21 SSB noise figure as a function of LO power [72]: (a) Mixer 1; (b) Mixer 2; (c) Mixer 3. Reprinted
with permission of the IEEE.
have been obtained. In this case, the error associated with NFYF+3 comes from the non-
ideal gain characteristics and from the poor match properties of the mixer. On the one
hand, the gain ratio, i.e. the ratio of the sum of all significant available gains over the
principal one, does not have a 3 dB value. In addition, the mixer is not adequately matched
at the output. For this last reason, NFCS_SCALAR cannot provide an accurate result. As a
conclusion, in this challenging case a procedure that includes mismatch corrections and
receiver noise calibration is now needed for accurate noise figure characterization.
10.7 Conclusion
In this chapter we have explained two popular methodologies for noise figure characteri-
zation:Y-factor and cold-source. These two methodologies have different implications in
terms of measurement complexity, calibration, error correction, accuracy, etc. We have

tried to provide a clear picture of the principles behind each technique, including their
basic equations, calibration processes, and measurement requirements. The most com-
mon sources of error affecting both methodologies have been reviewed. These sources
have been separated into three categories: mismatch, temperature, and measurement
setup. In most cases they affect each technique differently. Comparative analyses have
been performed whenever it was relevant. Noise figure of mixers has also been addressed
in a final section because of the particularities of this kind of characterization in the con-
text of a frequency conversion. The procedures for obtaining a single-sideband noise
figure from Y-factor and cold-source methodologies have been detailed and compared.
References
[1] J. B. Johnson, “Thermal agitation of electricity in conductors,” Phys. Rev., vol. 32, pp. 97–109,
July 1928.
[2] H. Nyquist, “Thermal agitation of electric charge in conductors,” Phys. Rev., vol. 32,
pp. 110–113, July 1928.
[3] W. B. Davenport and W. L. Root, An Introduction to the Theory of Random Signals and
Noise. New York: McGraw-Hill, 1958.
[4] A. Van der Ziel, Noise in Solid State Devices and Circuits. New York: Wiley-Interscience,
1986.
[5] A. B. Carlson, Communication Systems. An Introduction to Signals and Noise in Electrical
Communication, 3rd ed. New York: McGraw Hill, 1986.
[6] S. A. Maas, Noise in Linear and Nonlinear Circuits. Norwood: Artech House Inc., 2005.
[7] D. M. Pozar, Microwave Engineering. New York: John Wiley and Sons, 1998.
[8] H. T. Friis, “Noise figure of radio receivers,” Proc. IRE, vol. 32, no. 7, pp. 419–422, July
1944.
[9] H. A. Haus and R. B. Adler, Circuit Theory of Linear Noisy Networks. New York: John Wiley
and Sons, 1959.
[10] “IRE Standards on methods of measuring noise in Linear twoports, 1959,” Proc. IRE, vol. 48,
no. 1, pp. 60–68, January 1960.
[11] “Description of the noise performance of amplifiers and receiving systems,” Proc. IEEE,
vol. 51, no. 3, pp. 436–442, March 1963.
[12] IRE 7.S2, “IRE standards on electron tubes: Definition of terms, 1957,” Proc. IRE, vol. 45,
no. 7, pp. 983–1010, July 1957.
[13] H. Rothe and W. Dahlke, “Theory of noisy fourpoles,” Proc. IRE, vol. 44, no. 6, pp. 811–818,
June 1956.
[14] H. A. Haus et al. “Representation of noise in linear twoports,” Proc. IRE, vol. 48, no. 1,
pp. 69–74, January 1960.
[15] P. Penfield, “Wave representation of amplifier noise,” IRE Trans. Circuit Theory, vol. 9, no.1,
pp. 84–86, March 1962.
[16] K. Hartmann, “Noise characterization of linear circuits,” IEEE Trans. Circuits Syst., vol. 23,
no. 10, October 1976.
[17] H. Fukui, “Available power gain, noise figure, and noise measure of two-ports and their
graphical representations,” IEEE Trans. Circuit Theory, vol. 13, no. 2, pp. 137–142, June
1966.
[18] M. W. Pospieszalski, “Interpreting transistor noise,” IEEE Microwave Magazine, vol. 11,
no. 6, pp. 61–69, October 2010.
[19] R. Q. Lane, “The determination of device noise parameters,” Proc. IEEE, vol. 57, no. 8,
pp. 1461–1462, August 1969.
[20] M. Mitama and H. Katoh, “An improved computational method for noise parameter
measurement,” IEEE Trans. Microw. Theory Tech., vol. 27, no. 6, pp. 612–615, June 1979.
[21] G. Vasilescu, G. Alquie, and M. Krim, “Exact computation of two-port noise parameters,”
Electron. Lett., vol. 25, no. 4, pp. 292–293, February 1989.
[22] A. Boudiaf and M. Laporte, “An accurate and repeatable technique for noise parameter
measurements,” IEEE Trans. Instrum. Meas., vol. 42, no. 2, pp. 532–537, April 1993.
[23] L. Escotte, R. Plana, and J. Graffeuil, “Evaluation of noise parameter extraction methods,”
IEEE Trans. Microw. Theory Tech., vol. 41, no. 3, pp. 382–387, March 1993.
[24] V. Adamian and A. Uhlir, “A novel procedure for receiver noise characterization,” IEEE
Trans. Instrum. Meas., vol. 22, no. 2, pp. 181–182, June 1973.
[25] A. C. Davidson, B. W. Leake, and E. Strid, “Accuracy improvements in microwave noise
parameter measurements,” IEEE Trans. Microw. Theory Tech., vol. 37, no. 12, pp. 1973–1978,
December 1989.
[26] R. Meierer and C. Tsironis, “An on-wafer noise parameter measurement technique with
automatic receiver calibration,” Microwave Journal, vol. 38, pp. 22–37, March 1995.
[27] G. Caruso and M. Sannino, “Computer-aided determination of two-port noise parameters,”
IEEE Trans. Microw. Theory Tech., vol. 26, no. 9, pp. 639–642, September 1978.
[28] J. M. O´ Callaghan and J. P. Mondal, “A vector approach for noise parameter fitting
and selection of source admittances,” IEEE Trans. Microw. Theory Tech., vol. 39, no. 8,
pp. 1376–1382, August 1991.
[29] J. M. O´ Callaghan, A. Alegret, L. Pradell, and I. Corbella, “Ill conditioning loci in noise
parameter determination,” Electron. Lett., vol. 32, no. 18, pp. 1680–1681, August 1996.
[30] S. Van den Bosch and L. Martens, “Improved impedance pattern generation for auto-
matic noise parameter determination,” IEEE Trans. Microw. Theory Tech., vol. 46, no. 11,
pp. 1673–1678, November 1998.
[31] S. W. Wedge and D. B. Rutledge, “Wave techniques for noise modeling and measurement,”
IEEE Trans. Microw. Theory Tech., vol. 40, no. 11, pp. 2004–2012, November 1992.
[32] J. Randa and D. K. Walker, “On-wafer measurement of transistor noise parameters at NIST,”
IEEE Trans. Instrum. Meas., vol. 56, no. 2, pp. 551–554, April 2007.
[33] “Applications and operation of the 8970A noise figure meter,” Hewlett-Packard Product Note
8970A-1 (Agilent Manual 08970–99000), November 1981.
[34] “Agilent N8973A, N8974A, N8975A NFA series noise figure analyzers,” Agilent Data Sheet
5980–0164E, November 2007.
[35] “Agilent PSA series spectrum analyzers. Noise figure measurements personality,” Agilent
Technical Overview 5988–7884EN, August 2005.
[36] “Application firmware for noise figure and gain measurements R&S®FS-K30 for
R&S®FSP/FSU/FSQ,” Rohde & Schwarz FS-K30 Data Sheet, November 2003.
[37] “Fundamentals of RF and microwave noise figure measurements,” Agilent Application Note
57–1, October 2000.
[38] “Agilent 346A/B/C noise source,” Agilent Operating and Service Manual 00346–90139, July
2001.
[39] “Noise figure measurement accuracy – the Y-factor method,” Agilent Application Note 57–2,
March 2004.
[40] “High-accuracy noise figure measurements using the PNA-X series network analyzer,”
Agilent Application Note 1408–20, September 2010.
[41] “Noise figure measurements, VectorStar, making successful, confident NF measurements on
amplifiers,” Anritsu Application Note 11410–00637, June 2012.
[42] N. Otegi, J. M. Collantes, and M. Sayed,“ Receiver noise calibration for a vector network
analyzer,” 76th Microwave Measurement Symposium (ARFTG), December 2010.
[43] “A new noise parameter measurement method results in more than 100x speed improvement
and enhanced measurement accuracy,” Maury Microwave Application Note 5A–0.42, March
2009.
[44] “Noise figure measurement without a noise source on a vector network analyzer,” Rohde &
Schwarz Application Note 1EZ61_2E, October 2010.
[45] BIPM, IEC, IFCC, ISO, IUPAC, IUPAP, OIML, “Guide to the expression of uncertainty in
measurement,” ISO, 1993, corrected and reprinted 1995.
[46] BIPM, IEC, IFCC, ISO, IUPAC, IUPAP, OIML, “Guide to the expression of uncertainty in
measurement. Supplement 1. Numerical methods for the propagation of distributions,” ISO,
2004.
[47] D. Boyd, “Calculate the uncertainty of NF measurements,” Microwaves & RF, pp. 93–102,
October 1999.
[48] J. Randa, “Uncertainty analysis for noise-parameter measurements at NIST,” IEEE Trans.
Instrum. Meas., vol. 58, no. 4, pp. 1146–1151, April 2009.
[49] A. Collado, J. M. Collantes, L. De la Fuente, N. Otegi, L. Perea, and M.Sayed, “Combined
analysis of systematic and random uncertainties for different noise-figure characterization
methodologies,” IEEE MTT-S Int. Microwave Symp. Dig., pp. 1419–1422, June 2003.
[50] N. Otegi, J. M. Collantes, and M. Sayed, “Uncertainty estimation in noise figure mea-
surements at microwave frequencies,” AMUEM 2005, International Workshop on Advanced
Methods for Uncertainty Estimation in Measurement, May 2005.
[51] “Noise figure corrections,” Anritsu Application Note 11410–00256, November 2000.
[52] “NF accuracy,” Anritsu Application Note 11410–00227, November 2003.
[53] “TheY factor technique for noise figure measurements,” Rohde & Schwarz Application Note
1MA178_0E, May 2011.
[54] Agilent Technologies, Noise Figure Uncertainty Calculator [Online], Available:
http://www.agilent.com.
[55] Rohde & Schwarz, Noise Figure Error Estimation Tool [Online], Available: http://www2.
rohde-schwarz.com.
[56] Anritsu, Noise Figure Uncertainty Calculator [Online], Available: http://www.anritsu.com.
[57] D. Vondran, “Noise figure measurement: Corrections related to match and gain,” Microwave
Journal, vol. 42, pp. 22–38, March 1999.
[58] J. M. Collantes, R. D. Pollard, and M. Sayed, “Effects of DUT mismatch on the noise figure
characterization: A comparative analysis of two Y-factor techniques,” IEEE Trans. Instrum.
Meas., vol. 51, no. 6, pp. 1150–1156, December 2002.
[59] G. F. Engen, “Mismatch considerations in evaluating amplifier noise performance,” IEEE
Trans. Instrum. Meas., vol. 22, no. 3, pp. 274–278, September 1973.
[60] N. J. Khun, “Curing a subtle but significant cause of noise figure error,” Microwave Journal,
vol. 27, no. 6, pp. 85–98, June 1984.
[61] L. F. Tiemeijer, R. J. Havens, R. de Kort, and A. J. Scholten, “Improved Y-factor method for
wide-band on-wafer noise-parameter measurements,” IEEE Trans. Microw. Theory Tech.,
vol. 53, no. 9, pp. 2917–2925, September 2005.
[62] “Agilent N4000A, N4001A, N4002A SNS series noise sources 10 MHz to 26.5 GHz,” Agilent
Product Overview 5988–0081EN, July 2005.
[63] E. W. Herold, “The operation of frequency converters and mixers for superheterodyne
reception,” Proc. IRE, vol. 30, no. 2, pp. 84–103, February 1942.
[64] M. J. O. Strutt, “Noise-figure reduction in mixer stages,” Proc. IRE, vol. 34, no. 12,
pp. 942–950, December 1946.
[65] H. C. Torrey and C. A. Whitmer, Crystal Rectifiers. McGraw-Hill: New York, 1948.
[66] W. L. Pritchard, “Notes on a crystal mixer performance,” IRE Trans. Microw. Theory Tech.,
vol. 3, no. 1, pp. 37–39, December 1955.
[67] G. C. Messenger and C. T. McCoy, “Theory and operation of crystal diodes as mixers,” Proc.
IRE, vol. 45, no. 9, pp. 1269–1283, September 1957.
[68] R. D. Haun, “Summary of measurement techniques of parametric amplifier and mixer noise
figure,” IRE Trans. Microw. Theory Tech., vol. 8, no. 4, pp. 410–415, July 1960.
[69] M. R. Barber, “Noise figure and conversion loss of the Schottky barrier mixer diode,” IEEE
Trans. Microw. Theory Tech., vol. 15, no. 11, pp. 629–635, November 1967.
[70] D. N. Held and A. R. Kerr, “Conversion loss and noise of microwave and millimeter-wave
mixers: Part 1 – theory,” IEEE Trans. Microw. Theory Tech., vol. 26, no. 2, pp. 49–55, February
1978.
[71] D. N. Held and A. R. Kerr, “Conversion loss and noise of microwave and millimeter-wave
mixers: Part 2 – experiment,” IEEE Trans. Microw. Theory Tech., vol. 26, no. 2, pp. 55–61,
February 1978.
[72] N. Otegi, N. Garmendia, J. M. Collantes, and M. Sayed, “SSB noise figure measurements
of frequency translating devices,” IEEE MTT-S Int. Microwave Symp. Dig., pp. 1975–1978,
June 2006.
[73] R. Poore, “Noise in ring topology mixers,” Agilent EEsof EDA, 2000.
[74] S. A. Maas, Microwave Mixers, 2nd ed., Norwood: Artech House, 1992.
[75] “Noise measurement software FS-K3. Noise test system with FSE, FSIQ, or FSP analyzers,”
Rohde & Schwarz Application Note, News from Rohde-Schwarz 167, 2000/II.
[76] J. Dunsmore, “Novel method for vector mixer characterization and mixer test system vector
error correction,” IEEE MTT-S Int. Microwave Symp. Dig., vol. 3, pp. 1833–1836, June 2002.
[77] G. M. Hegazi, A. Jelenski, and K. S. Yngvesson, “Limitations of microwave and millimeter-
wave mixers due to excess noise,” IEEE Trans. Microw. Theory Tech., vol. 33, no. 12,
pp. 1404–1409, December 1985.
11 TDR-based S-parameters
Peter J. Pupalaikis and Kaviyesh Doshi
11.1 Introduction
Many engineers are familiar with the VNA as an instrument for measuring S-parameters.
The VNA’s origins lie in microwave systems analysis and its application has been pri-
marily in the frequency domain. Many are also familiar with the use of TDR for making
qualitative measurements of time domain reflections and other phenomena. TDR has its
origins in signal integrity analysis, as signal integrity is primarily concerned with time
domain effects.
It is less well known that TDR and associated TDT is also a highly useful technique
for precise quantitative measurements in signal integrity and can be used effectively for
S-parameter measurement.
This chapter deals with the measurement of S-parameters using time domain tech-
niques such as found in TDR and TDT. We cover the topic by first describing the
hardware architecture of TDR instruments including the sampling system, the pulser,
and the timebase. Then we describe how time domain TDR and TDT measurements are
converted to raw, uncalibrated, frequency domain S-parameters. We do not deal with cal-
ibration techniques as these are the same for the VNA and TDR once raw S-parameters
have been determined. Then, we quantitatively discuss the main element that effects
the accuracy of time domain measurements: that of noise or SNR. SNR is such a big
problem that it is the major source of error in time domain derived S-parameters and it is
worthwhile understanding the sources of dynamic range degradation in TDR systems and
the key design areas for improvement. We end the chapter with a consideration of how
S-parameter measurements are affected by noise and present equations for determining
measurement uncertainty when noise is the primary source of error.
11.2 TDR pulser/sampler architecture
The TDR module consists of a step source (pulser) and a sampler that can measure
the reflected signal. Figure 11.1 is an idealized schematic of a pulser-sampler. First we
describe the working principles of the sampling system by ignoring the pulse generating
system and its output and then describe the operation of the pulse generating system.
The sampling strobe shown in the lower right corner of the figure is a clock signal that
controls when the signal from the DUT is sampled. The block marked Impulse Generator
consists of a unipolar impulse generator, an amplifier, and a nonlinear transmission line
(NLTL). Input to the NLTL is a slow rise-time amplified impulse that is converted to a
280 Peter J. Pupalaikis and Kaviyesh Doshi
DAC≈ + 2V
C3 C1
Summing
Q→V
converter
D1
14 bit D
To DUT
50Ω
ADC D2
C4 C2
Output S1
Sampling
DAC strobe
Pulse
generating Pulse ≈ − 2V S2
system strobe
Impulse
generator
Bias voltages
Fig. 11.1 TDR sampler schematic.
faster rise-time impulse. The behavior of an NLTL for generating a faster rise-time output
from a slow rise-time step has been studied extensively [1, 2] and the references therein
provide more NLTL details. Capacitors C1 and C2 in Figure 11.1 are AC-coupling
capacitors, whereas the capacitors C3 and C4 store the charge corresponding to the
signal coming from the DUT. Schottky diodes D1 and D2 form a switch that controls
the sampling process. The “summing charge to voltage converter” sums the charge on
C3 and C4 and converts that to voltage, which is then digitized by the ADC.
To understand the operation of the sampler, note that the polarity of the DAC is such
that the two Schottky diodes are reverse biased. Now consider the case when the sampling
strobe maintains the diodes in the reversed bias region, and there is no signal from the
DUT. When the Schottky diodes are reverse biased, they can be thought of as an open.
The DUT in this case sees a matched load of 50 ohms. Capacitor C3 holds the charge
due to the 2 V DAC and the capacitor C4 holds the charge due to the −2V DAC. In this
state, the output of the “summing charge to voltage converter” is zero. The system under
measurement remains undisturbed by the sampler.
Next consider the case when we momentarily forward bias the two Schottky diodes
and there is no signal from the DUT. In this state, capacitor C3 is charged to −0.2 V, the
forward bias voltage drop across D1. Capacitor C4 is charged to 0.2 V, the forward bias
voltage drop across D2. Since there is no signal from the DUT, the node marked D is
at 0 V. Now suppose there was some reflected or through signal from the DUT. In this
case, the voltage at node D is the voltage due to the signal from the DUT. Let this be
x V. Now C3 is charged to x − 0.2 V and C4 is charged to x + 0.2 V. When the forward
TDR-based S-parameters 281
Sampling strobe
+ 1V
− 1V
≈ 100 ns
Impulse train at S1
−8 V
Impulse train at S2
+8 V
Fig. 11.2 Waveforms indicating the biasing of the Schottky diodes in the sampler.
bias is removed, the two capacitors are discharged and the “summing charge to voltage
converter” converts the charge to 2x V (x − 0.2 + x + 0.2 V).
The appropriate biasing of the Schottky diodes is achieved through the sampling
strobe and the block marked Impulse Generator. The sampling strobe is a clock signal
of approximately 10 MHz frequency, shown in Figure 11.2. Details of generating the
TDR strobe are explained in the next section. This sampling strobe is first converted to
an impulse and then amplified. The amplified impulse then passes through the NLTL,
the result of which is a faster rise time impulse as shown by the signals marked S1
and S2 in Figure 11.2. As shown, S1 is an impulse with a peak of approximately −8 V
whereas S2 is an impulse with a peak of approximately 8 V. This high level of voltage
is enough to forward bias the two Schottky diodes. Thus, for the short interval of time
corresponding to the impulse width, the diodes are forward biased and the signal from
the DUT is recorded by capacitors C3 and C4. For the remaining time, the sampler acts
like a matched load for the DUT.
The schematic of the pulser block in Figure 11.1 is shown in Figure 11.3. This is one
of the many possibilities as described in [3]. As shown in Figure 11.3 and described
in [3], the pulser consists of a constant current source, supplying current I0 ; bias volt-
age −V that drives the current source; bias voltage +V that controls the switching of
output diode; fast switch S1 ; and resistor R1 , which is usually 50 . Initially when
the switch is open, there is a constant current flowing through resistor R1 and the
+V
S1
Output
I0 50Ω
−V
Fig. 11.3 TDR pulser schematic.
diode producing the generator baseline voltage of −I0 R1 volts. When switch S1 is
closed, the diode becomes reversed biased, thus disconnecting I0 from R1 . The out-
put voltage changes rapidly to its topline value of zero volts. The pulser has the source
resistance of R1 ohms. A constant current I0 is always drawn from the independent
power supply −V . The rise time of the step is governed by the switching transients
of switch S1 and the charge storage time in the diode. The opening and the closing
of the switch can be controlled by an external pulser strobe signal (not shown here).
Note that the schematic described here is to describe the basic operation of a pulse gen-
erating system. More advanced pulser designs and details can be found in [3, 5] and
references therein.
Referring back to Figure 11.1, when the pulser is active, the sampler will record the
voltage due to the pulser as well as any reflections from the DUT. A sampler-only module
can be constructed without the pulser generating system. More details about the sampler
can be found in [6] and references therein.
11.3 TDR timebase architecture
Since TDR measurements have traditionally been performed on sampling scopes, it is

natural to assume that TDR timebases resemble sampling scope timebases. There are
two types of sampling scope timebases in use and both are used for TDR with adap-
tation for the pulser. The most common style of timebase is the sequential sampling
timebase, which is shown schematically in Figure 11.4. Here we see on the left a ref-
erence clock, a pulser strobe, and an arm signal. This is a very simplistic diagram that
assumes that the pulser strobe generation time is totally arbitrary, which means it might
be free-running or generated manually. If free-running, the arm signal determines when
Reference
clock ↑N
Pulser
strobe Counter Analog Sample
Trigger (coarse fine strobe
Arm delay) delay
DTC
Fig. 11.4 TDR using sequential sampling.
the system is ready to take a sample. Prior to arming, a programmable delay device
called a DTC is set to count off a predetermined time following the firing of the pulser
which is the trigger event. In this manner, whenever the system is programmed with
a delay time and armed, a sampling strobe will be generated a specified time after
the pulser strobe. By repeatedly programming different times, arming the system, and
generating pulser strobes, an equivalent time waveform containing the behavior of the
DUT due to the applied step edges can be recorded. We call this waveform equiva-
lent time because it represents a correct, high sample rate waveform despite the fact
that the actual samples were taken at dramatically different times. For equivalent time
sampling to work in a TDR, the reaction of the DUT to the TDR stimulus must be
identical for every pulser strobe. The delay action performed by the DTC is generally
performed by two elements: one digital and the other analog. The digital element counts
clocks and is the coarse timer. The analog element is fine and depends on some phe-
nomenon like the discharge time of a precharged capacitor. Although the sequential
sampling timebase is the most popular, it has many severe disadvantages for a TDR.
These are:
• The DTC is nonlinear and requires calibration. Even 1 ps of sampling error creates
large inaccuracies in TDR-based S-parameter measurements.
• The system is slow. As we will see later in the chapter, sampling speed is critical
for TDR because of the amount of averaging required. TDR measurements using a
sequential sampling range in an actual sample rate between 40 and 150 kS/s, so for a
40 Kpoint waveform only one to a few waveforms can be acquired each second.
• Because a trigger system is involved, the jitter of the trigger also adds to measurement
error.
• The equivalent time sample rate is dependent on the granularity of the DTC control.
Another type of sampling scope timebase is the CIS timebase. This architecture was
originally proposed by LeCroy [7], [8]. This timebase is illustrated in Figure 11.5. The
CIS timebase generates continuous pulser and sampler strobes that intentionally beat
with each other. The example shown in Figure 11.5 is the timebase arrangement for a
LeCroy SPARQ in “normal" mode, meaning that the pulser is pulsing at 5 MHz or with a
200 ns period. The sampler is placed at the seemingly odd sample rate of 9.884647 MHz
2.53047 GHz
PLL
Φ VCO ÷8
DDS
N = 20719 100 MHz 316.309 MHz

D 65536
9.884647 MHz
÷ 32
sample strobe
100 MHz 5 MHz

Reference ÷ 20
pulser strobe
Fig. 11.5 TDR using coherent interleaved sampling.
through the combination of the PLL and the DDS. Here we see that the input to the
phase detector is both the 100 MHz reference clock and, in steady-state, a 100 MHz
phase locked output from the DDS. The DDS in this example, has been programmed
to multiply its input by 20 719/65 536 ≈ 0.316 so that when the system is locked, the
input to the DDS is approximately 316 MHz, requiring the output of the VCO at eight
times higher frequency to be approximately 2.53 GHz. When locked, therefore, we have
a sample strobe rate of the VCO output divided by 32. If we examine Figure 11.5 we
see that this system produces a sample strobe rate of precisely 2048/20 719 times the
100 MHz reference. Since the pulser is pulsing at a rate of precisely one-twentieth of
the reference clock frequency, we have the following equality:
2048 1 fref 1
fref = ,
20 719 S 20 P
where S refers to an integer number of samples and P refers to an integer number of
cycles of the pulser. Therefore, we have:
S 40 960
= .
P 20 719
This specific arrangement of frequencies means that exactly 40 960 samples corre-
sponds to exactly 20 719 cycles of the pulser. The Stern-Brocot algorithm described in
[9] was used to obtain the rational number equivalent of the ratio of frequencies. There-
fore, a back-end memory system can store consecutive samples using a modulo 40 960
counter and even average many results. When read out, the memory is reordered by tak-
ing the memory index times 24 079 and again counting modulo 40 960. This reordering
produces a 40 960 sample equivalent time waveform of a complete cycle of the pulser
at, for this example, an equivalent time sample rate of 204.8 GS/s.
The operation of this system is shown in Figure 11.6. Here we see samples being taken
of repeating waveforms representing the reaction of the DUT to repeated step edges. In
TDR based S-parameters
Sampler strobes
Pulser strobes
Pulser cycle
”Dirty”
incident edge
Relaxation to Useful portion

quiescent state for TDR
Reflectionat
Incident edge DUT interface
Fig. 11.6 TDR cycles.
the end, an entire waveform is acquired representing one complete cycle. Not all of the
cycle is usable. The useful part consists of the portion just before the application of the
high-speed edge up to the time that the TDR is turned off. Usually, the edge is really fast
in only one direction. Equally important, during the time between the edge application
and when it turns off, the TDR presents an ideally 50 load. The reverse going edge
(the “dirty” edge in Figure 11.6) is slow and during the off time, the system presents an
uncontrolled, non-ideal impedance.
The benefits of the CIS timebase are:
• The sample rate is very high – approximately 10 MS/s in this SPARQ example. The
samples are taken consecutively and continuously.
• The system removes the possibility for timebase nonlinearity.
• Presuming the programming capability of the multiplication and division factors in
the system, there are no practical limitations on record length or equivalent time
sample rate.
There are some drawbacks to this timebase method:
• While the sample rate is high, the system, while not requiring the storage of, does
require slipping over unwanted portions of the pulser cycle. In the end, we usually
want only a portion of half a cycle of TDR, as illustrated in Figure 11.6 (a slightly
smaller portion is shown to account for the potential duty cycle variation).
• While fast, the time between the negative edge and the positive edge must be long
enough for the system to reach a quiescent or fully discharged state. This is because
TDR depends on the assumption of zero energy storage in the system prior to the step
edge. This same assumption must also be met for sequential sampling timebases, but
is easier because of the much slower speeds involved. For DUTs with longer electrical
lengths, longer pulser periods must be programmed.
Generally speaking, the extreme speed benefits and lack of timebase nonlinearity issues
far outweigh the drawbacks of CIS.
11.4 TDR methods for determining wave direction
The VNA operates by sweeping frequencies. At each frequency, a standing-wave is

developed in a DUT. The VNA is attempting to simulate the Fourier transform situa-
tion whereby the sinusoidal stimulus has been present for all time. Because of this, it
is impossible to determine the incident and reflected wave from an acquired voltage
waveform as illustrated in Figure 11.7. In Figure 11.7, we have a buoy man trying to
understand what is going on by the simple up and down motion of the buoy. Because the
buoy only moves up and down and because it has been doing it forever, the buoy man
cannot determine the direction of the underlying forward and reverse propagating waves.
He only senses the sum of the wave effects. Because of this inability to determine wave
direction, the use of directional couplers is necessitated in the VNA to perform the sepa-
ration. Directional couplers are microwave devices that can distinguish the directionality
of waves, but they generally suffer from a number of imperfections and limitations. They
are specified and rated on two key specifications that are interesting for VNA usage: on
directionality and attenuation, especially frequency-dependent effects. Directionality is
the ability of the directional coupler to output waveforms going in only one direction.
The attenuation of the directional coupler goes up as frequency goes down, meaning that
the size of the acquired waveform is small or nonexistent at low frequency giving the
VNA poor dynamic range at low frequency and no DC point. Generally, the outputs of
the directional-couplers are sampled as voltage waveforms and since the results are sinu-
soidal waves at a single frequency, a complex amplitude and phase is determined. Since
the sinusoid is very narrow-band (ideally a single frequency) the resulting amplitude and
phase can be determined very precisely with high dynamic range.
TDR operates by launching an impulsive wavefront in the form of the rising edge of
a step into a DUT. The TDR is attempting to simulate the Laplace transform situation
whereby the system has never been stimulated prior to the arrival of the step edge and
My buoy has been going up and down for all time.

I can’t tell which direction any waves are going.
Buoy motion
Resultant (Sum) wave
Forward propagating wave
Reverse propagating wave
Fig. 11.7 Buoy man and his perception in VNA.
is in a completely quiescent state with no stored energy. In analyzing TDR waveforms,

there are two key assumptions made:
(i) Until the TDR edge arrives, there has never been any incident edge from the source.
(ii) After the TDR edge comes and goes, no other edges are generated by the source.
These assumptions are illustrated by buoy man’s perceptions in TDR in Figure 11.8.
Here, buoy man is waiting and when the first up and down motion of the buoy is detected,
he knows that it is the incident wave because of the first assumption. As the buoy keeps
going up and down, he knows that these must be reflected waves because of the second
assumption. The ability to detect wave direction from a single voltage waveform that is
the sum of forward and reverse propagating waves is one of the keys to TDR and means
that TDR does not require directional couplers. Of course, by examining Figure 11.6,
we see that in order to meet the key assumptions, we must have the repetition rate of the
TDR low enough such that on each cycle, the system is totally relaxed and all energy
has been removed from the system. In other words, the cycle must be long enough for
the effects of the incident edge on the system to die down sufficiently. By the way, this
I had been waiting and suddenly the buoy went up

and then down. I knew this was the incident wave
and which direction it was going since I know
the source is to the left. The buoy continued to
go up and down but I knew these were the reflected
waves and that they are going in the opposite
direction.
Buoy motion
Resultant (sum) wave
Forward propagating wave
Reverse propagating wave
Fig. 11.8 Buoy man and his perception in TDR.
is why TDR is especially bad for handling AC coupled devices because the AC coupling
leads to long time constants that make the meeting of these assumptions very difficult.
To summarize, in TDR it is easy to separate the incident from the reflected waveform.
The incident wave is the rising edge of the step and the reflected wave is everything later.
A perceptive reader might wonder whether the TDR is actually properly accounting
for reflections that occur after the incident waveform has been generated. For example,
it is possible (and it occurs in practice) for waves returning from the DUT to be reflected
from the source and sent back towards the DUT in the forward direction. This possibility
is handled as follows:
(i) The assumption is that at the source side, the sampler and source are sufficiently
co-located such that the sampler is seeing only the reflected waves after the incident
has been generated. In other words, although forward going waves are retransmitted
from reverse going returning waves from the DUT, the system only ever sees the
returning waves. This means that the system might see two returning reflections
from the DUT due to imperfections of the source termination.
(ii) The waves returning from the DUT due to secondary reflections at the source are
either ignored or come out in calibration. In general TDR usage (meaning not for
S-parameter determination) the DUT is often assumed to have only one interface,
as accounting for all the internal reflections within the DUT is difficult. Therefore,
secondary reflections either from the source or within the DUT are moved off the
screen when TDR is being used for qualitative measurements. When TDR is being
used for S-parameter determination, these secondary reflections must be provided
and the waveform must be long enough for all the reflections to die down to essen-
tially zero and be removed through calibration techniques similar to VNA usage.
Remember, this same situation exists in the VNA as well and the standing waves
generated are a function of similar effects.
The accounting for internal reflections within a DUT in TDR usage is commonly
referred to as peeling [10], [11]. Peeling accounts for all reflections by remembering all
of the reflections in the system and applying this memory to classify each reflection that
is seen.
Figure 11.9 shows how TDR is used in principle and how users mentally separate
the incident waveform from the reflected waveform. At the top of Figure 11.9 we see
three overlaid waveforms for the cases of an open, short, and matched load (the load
is the same impedance as the line in which the TDR waveform was propagating). Pay
attention to the different labeling of the y-axis for the top drawing portion. All the
waveforms begin with a step from 0 to half the source voltage level. Here we presume
that the impedance of the source is the same as the impedance of the line. The step
stays at this level until the DUT is encountered after which it either stays the same for
the matched load, jumps to the source voltage level for the open, or drops to zero for
the short.
The next two waveforms separate the incident from the reflected portion. Here we see
that the incident waveform is the step common to all three cases and that the reflected
waveform is either a positive step in the case of the open, a negative step in the case of
the short, or zero in the case of the matched load.
TDR users mentally remove the incident step from their thinking when they view a
TDR waveform. In fact, it is sometimes common to simply shift the incident step edge
slightly off the screen to the left to see essentially only the reflected waveform. This
can be done when qualitative measurements are being performed. In the case of certain
quantitative measurements, like impedance, it is helpful to calibrate the system. This is
often done by applying a single short or open standard, calibrating the voltage drop or
rise, and then assuming that the result of a short is an inverted version of the open or
vice versa.
One thing worth noting here is that the DUT interface is shown in Figure
11.9 as denoting the time in the voltage waveform corresponding to the DUT, but
this is not the waveform at the DUT itself. All TDR waveforms are sampled near the
Acquired waveform
Vs Open
Vs Matched load
2
0 Short
Incident waveform
Vs
2
− Vs
2
Reflected waveform
Vs Open
2
0 Matched load
− Vs Short
2
Pulser/sampler DUT
interface Interface
Fig. 11.9 TDR concept and incident and reflected step separation.
pulser. This means that the time in the waveform corresponding to the DUT is for a
round-trip, meaning that the time corresponds to the time between the launch of the
incident wave and the return of the reflected wave.
11.5 Basic method for TDR-based S-parameter measurement
Since the VNA is directly sampling frequency, the determination of frequency domain
S-parameters is straightforward. The frequency content of the reflected waveform is
divided by the frequency content of the incident waveform, which is one frequency for
each measurement.
In TDR, all frequencies are launched at once and all frequencies are received in a
single acquisition. By separating and converting the incident and reflected time domain
waveforms to the frequency domain through the DFT, the equivalent S-parameter cal-
culation is performed. In fact, once the DFT of the incident and reflected waveforms is
computed, all other calculations, like calibration, proceed exactly like the VNA.
Traditionally in TDR-based S-parameter measurement, the incident portion of the
waveform is not considered. As mentioned previously, if we examine Figure 11.9 we
Step-like waveform at Step-like waveform at

driven port M Port N
d/dt d/dt
Denoising Denoising
Incident Reflect
extractor Extractor
DFT DFT
D N
Raw S-parameter S [N][M]
Fig. 11.10 TDR S-parameter calculation block diagram.
see that if the portion of the acquired waveform containing the incident edge is sim-
ply removed from the waveform (i.e. the time up to and just after the incident edge
is removed) then the resulting waveform resembles a rescaled version of the reflected
waveform. Traditionally, the step-like waveform with the incident portion gated off is
used to compute the frequency content using methods provided by [12] or [13]. Using
these methods, the incident frequency content is assumed to be unity (not a perfect step,
but a perfect impulse) which is not really a problem as downstream calibration will take
care of this. The disadvantage, however is that the calibration must take care of chang-
ing pulser conditions like frequency content and, most important, skew. Another lesser
consideration is that using traditional methods, the error terms contain the confusing
step frequency content that drops at 20 dB/decade† . In other words, traditional methods
cause the error terms to look different to the error terms produced by the VNA.
Here we present an alternate method as shown in Figure 11.11. Here we see that
the step-like waveform at both the driven and measured ports are differentiated (i.e. the
first difference is calculated in discrete terms). They then undergo an optional denoising
step, for example as provided in [14]. Then the incident portion is extracted from the
driven port waveform and the reflected portion is extracted from the measurement port
waveform. It should be apparent that by computing the frequency content of the now
† The way to understand this effect is to realize that the DFT of the impulse response of a system is
essentially the frequency response of a system and tends to be mostly flat. The step response is the integral
of the impulse response and therefore the response drops in frequency content as a system with a pole at
zero frequency.
Acquired derivative waveform
Incident extraction Reflect extraction

window window
Incident derivative waveform
Reflected derivative waveform
Pulser/sampler DUT
interface interface
Fig. 11.11 TDR S-parameter frequency content determination.
separated portions, we have frequency domain versions of both. The preferred method is
to utilize the DFT, FFT, or most preferably, the CZT [15], [16]. The CZT allows arbitrary
end frequency and frequency spacing. The best way to treat the incident and reflected
waveforms prior to conversion to the frequency domain is to zero out the incident and
reflected portions of the original waveform. In other words, the incident waveform is
created by zeroing out the reflected portion and the the reflected waveform is created by
zeroing out the incident portion of the voltage waveform acquired. This operation can be
understood as a windowing operation as shown in Figure 11.11. Because of the derivative
and windowing, both incident and reflected waveforms have no edge discontinuities, but
also very important is the fact that the two waveforms retain exactly the same length and
sample timing. This alleviates the need for special handling of the incident and reflected
waveforms.
As mentioned previously, the frequency-dependent incident waveform content is in
the rising edge of step. This is seen more clearly by looking at the derivative waveforms
as shown in Figure 11.11. Here it is clearly seen that the incident waveform occurs at
an early time and that the entire incident waveform can be formed by simply extracting
the beginning portion where the first impulse occurs. This impulse is so large and rec-
ognizable that it is not interesting to dwell on algorithms used for finding and extracting
it. The waveform portion that occurs after the incident impulse is assumed to contain
only reflected waveform. Note that this waveform portion may contain reflections due
to interactions between returning waves and the pulser/sampler as discussed previously.
While not described here, mathematical analysis shows that this is accounted for using
ordinary VNA calibration algorithms. Note that in TDT, the entire waveform consists of
the reflected portion.
The next and final step is to divide the frequency content of the incident portion into the
reflected portions. This produces what is termed a raw S-parameter. It is an S-parameter
because it is a frequency domain vector of ratios of reflected waves to incident waves. We
say it is raw because it is not yet calibrated. We will say nothing here about calibration
because at this step the results produced are similar to those produced by a VNA (see
Chapter 8 for details on two-port VNA calibration).
To summarize:
• The raw S-parameters were computed using a true ratio of reflect to incident and the
frequency domain content was computed from a derivative waveform.
• The method presented here is entirely insensitive to skew and requires only rudimen-
tary efforts to place the incident edge.
• Because of the derivative action, the error terms produced in calibration look similar
to those found in VNA calibrations.
In fact, the main difference between the S-parameters produced by TDR using these
methods and by the VNA is the dynamic range (i.e. the amount of signal in the incident
waveform relative to the amount of noise). By computing the derivative we have normal-
ized the step frequency content which drops at 20 dB/decade, but have simultaneously
reshaped the noise such that it increases with increasing frequency. This reshaping has
not changed the ratio of the two and the fact remains that TDR generally has SNR, and
therefore a dynamic range that drops at 20 dB/decade.
11.6 Summary of key distinctions between TDR and VNA
To summarize the key distinctions between the TDR and VNA:
(i) In TDR, the pulser produces all frequency content for the incident waveform in
a single acquisition. In other words, while the VNA sweeps frequencies, TDR
produces all frequency content in every pulse. As we will see, it is the broadband
nature of the incident wave that leads to dynamic range degradation in TDR-based
S-parameter measurement.
(ii) In TDR, the frequency content incident on the DUT is frequency dependent.
More specifically, because the incident wave is present in the rising edge of the
step-like waveform, the power drops approximately in a relationship that is inversely

proportional to frequency.
(iii) TDR does not require directional couplers. This is because all the incident energy is
present in the rising edge and with care, the incident and reflected waveform can be
separated in time. The lack of directional couplers means that TDR does not suffer
from dynamic range degradation at low frequency like the VNA.
(iv) TDR makes direct measurements in the time domain, whereas VNA measurements,
while often employing samplers, are directly measuring in the frequency domain
(because each acquisition is a sinusoid at a single frequency).
11.7 Dynamic range calculations
Since SNR or dynamic range is so important in TDR, it is useful to derive it and to

highlight the features that improve and detract from the dynamic range.
Within time domain instruments, we acquire step waveforms, therefore we start with
an acquired signal defined according to:
w [k] = s [k] + ε [k] . (11.1)
In (11.1), w [k] is a sample of the step waveform actually acquired, s [k] is a sample
of the step portion containing the signal of interest, and ε [k] is a sample of the noise
signal that we assume to be white, normally distributed, uncorrelated noise.
The signal content in the step is in the form of the frequency content of the deriva-
tive, so the derivation must consider this. Since during calculation we don’t know the
difference between the noise and the step, we must take the derivative of both. We will
be approximating:
d d d d d
w (t) = [s (t) + ε (t)] = s (t) + ε (t) = x (t) + ε (t) . (11.2)
dt dt dt dt dt
In (11.2), x (t) represents the true desired input signal in the form of an impulsive
wave front which is approximated as a discrete-time waveform with a sample x [k], and
d
dt ε (t) is the time derivative of the noise signal which is approximated as a discrete-time
waveform with a sample ε [k] which will be described in the following.
We are interested in these two signals in the frequency domain:
X ≈ F {x (t)} = DFT (x) ,

7 8
d
E ≈F ε (t) = DFT ε .
dt
We calculate the dynamic range, for each frequency, as an SNR (ratio of signal strength
and expected noise value):
X [n]
SNR [n] = . (11.3)
E [n]
In order to calculate the SNR, we calculate the frequency content of each of these
components separately and take the ratio. We start with the noise component. A noise
signal ε which contains only uncorrelated, normally distributed, white noise, has a mean
of 0 and a standard deviation of σ , which is the same as saying it has an rms value of σ .
We have K points of this signal ε [k], k ∈ 0 . . . K − 1.
If we calculate the discrete-Fourier-transform (DFT) of this noise signal, we obtain
N + 1 frequency points N = K/2, n ∈ 0 . . . N:
nk
1 −j 2π
E [n] = ε [k] e K,
K
k
where the frequencies are defined as
n Fs
f [n] = (11.4)
N 2
and Fs is the sample rate.

By the definition of the rms value and by the equivalence of noise power in the
time-domain and frequency-domain, we know that

Nbw
1 2 2
ε [k]2 = σ = ! E [n] √ , (11.5)
K 2
k n=0
where Nbw is the last frequency bin containing noise due to any band limiting effects.
We define an average value for the noise Eavg that satisfies the following relationship:

2 2 2
Eavg √ =σ = Nbw Eavg √ .
n 2 2
Therefore
1 σ
Eavg = 9 √ ,
fbw K
Fs /2
where fbw is the frequency limit for the noise calculated by substituting Nbw for n
in (11.4).
We, however, are taking the derivative of the signal. The derivative in discrete terms
is defined as
d ε [k] − ε [k − 1]
ε (t) ≈ ε [k] = ,
dt Ts
where Ts = 1/Fs is the sample period. Using the same equivalence in (11.5), we have

Nbw
1 2 2

(ε [k]) = σ =
2 !
E [n] √ . (11.6)
K 2
k n=0
Using the Z-transform equivalent of the derivative in the frequency domain, and an
average value for the noise in it, it can be shown that

⎛
f [n] ⎞2
−j 2π
⎜ 1 − e Fs ⎟
⎜ ⎟
⎜⎜ ⎟
1 ⎜ 2 ⎟
⎟ .
(ε [k])2 = σ = ⎜ E √
2⎟
avg
n ⎜ ⎟
K Ts
k ⎜ ⎟
! ⎝ ⎠
Therefore, the average noise component at each frequency is given by

f [n]
−j 2π
Fs
1 − e

2

E [n] = Eavg √ .
Ts 2
We can make an approximation that gives us a further insight by expanding the
numerator term in a series expansion

f " #
−j 2π 2πf f 3

Fs =
1 − e +O ,
Fs Fs

which allows us to approximate the noise component as

√
2πf [n] /Fs 2 2πf [n] σ 2
Eavg √ = E [n] = √ 9 . (11.7)
Ts 2 fbw
K Fs /2
Note here that σ corresponds to the noise in the step waveform, not the noise in the
derivative waveform, and the noise shaping for the derivative action is accounted for
in (11.7).
Now that we have the noise component of dynamic range, we move to the signal
component.
Without regard to the rise time or the frequency response of the step, which we will
consider later, we define the signal such that, in the discrete domain, the integral of the
signal forms a step
s [k] = s [k − 1] + x [k] Ts ,
where x is an impulse such that x [0] = A/Ts = A · Fs and is zero elsewhere such that
s forms a step that rises to amplitude A at time zero and stays there. X = DFT (x) and
therefore the signal components at each frequency are defined as
A
X [n] = = A · Fs .
Ts
Again, to gain further insight, we define
K · Ts = K/Fs = Td , (11.8)
where Td is the acquisition duration (i.e. the amount of time in the acquired waveform).
Therefore
A
X [n] = .
Td
Using (11.3), the ratio can therefore be expressed as
√ √
X [n] A K fbw
SNR [n] = = √ .
E [n] Td 2πf [n] σ Fs
Since these are voltage relationships, we can express the SNR in dB as

√ √
A K fbw A2 Kfbw
SNR (f ) = 20 log √ = 10 log (11.9)
2Td πf σ Fs 4Td2 π 2 f 2 σ 2 Fs
and using (11.8), finally

A2 fbw
SNR (f ) = 10 log . (11.10)
Td 4π 2 f 2 σ 2
We would like to express the noise in dBm, so we have

NdBm = 20 log (σ ) + 13.010 = 10 log 20σ 2
and therefore
NdBm
10 10
σ2 = . (11.11)
20
Substituting (11.11) in (11.10)
⎛ ⎞

⎜ 20A2 fbw ⎟ 2
SNR (f ) = 10 log ⎜ ⎟ = 10 log 20A fbw − NdBm .
⎝ NdBm ⎠ 4Td π 2 f 2
4Td π 2 f 2 10 10
Then, to clean things up, we extract some constants

20
10 log ≈ −6
8π 2
and therefore
2A2 fbw
SNR (f ) = 10 log − NdBm − 6.
Td · f 2
Now let’s consider some other factors. First, there is a frequency response of the
pulse, and a frequency response of the sampler. These responses can be aggregated into
a single response. Since, in decibels, it is simply the frequency response of the step
calculated by taking the DFT of the derivative of the step (isolating only the sampled
incident waveform) and calculating in dB, this value can simply be added to the dynamic
range. Similarly, we account for cabling and fixturing which we aggregate into a single
response, in decibels of F (f ). The signal must traverse the path through the cabling and
fixturing twice:
2
2A fbw
SNR(f ) = 10 log − NdBm + P (f ) + 2F (f ) − 6. (11.12)
Td f 2
Next, we consider the effects of averaging. Averaging the waveform by an amount
M achieves a 3 dB reduction in noise with every doubling of M. This leads to an
improvement in the dynamic range by
√
20 log M = 10 log (M) . (11.13)
The form of (11.13) allows it to be inserted directly into the numerator in (11.12)
2
2A fbw M
SNR (f ) = 10 log − NdBm + P (f ) + 2F (f ) − 6. (11.14)
Tdur f 2
We really don’t want to consider the dynamic range in terms of a number of averages
and instead prefer to consider the amount of time we are willing to wait. The number of
averages taken in a given amount of time is given by:
Fsa Tw
M= . (11.15)
Td Fse
In (11.15), we now need to distinguish what is meant by sample rate. F se becomes
the equivalent time sample rate and replaces what we previously called Fs . F sa is the
actual rate that samples are acquired at in the acquisition system and Tw is the amount
of time over which acquisitions are taken. Substituting (11.15) in (11.14), we obtain the
complete dynamic range equation shown in (11.16).

2A2 fbw Fsa Tw
SNR (f ) = 10 log − NdBm + P (f ) + 2F (f ) − 6. (11.16)
Td2 f 2 Fse
TDR-based S-parameters dynamic range equation
11.8 Dynamic range implications
The dynamic range equation (11.16) has several implications worth discussing. First
the obvious ones. Regarding frequency, the dynamic range drops at 20 dB per decade
(or 6 dB per octave). This can be considered as the effect of the drop-off in frequency
components of a step. If the waveform utilized could be an impulse, this effect could be
avoided. This effect is counteracted by the expression P (f ) which accounts for practical
step responses.
Next is the obvious fact that the dynamic range is strongly dependent on the step size.
It goes up by 6 dB for every doubling of the step amplitude, although the high frequency
content is also accounted for in P (f ) (which is not concerned with the difference between
pulser or sampler response). In other words, P (f ) is used to account for the product of
the pulser energy content and the sampler response.
The dynamic range is directly proportional to the random noise and also losses in the
cabling and fixturing, but this is also counteracted by a high sample rate. The dynamic
range goes up by 3 dB for every doubling (or 10 dB for every ten times increase) in either
the actual sampler sample rate or the time one waits for acquisitions to transpire.
The dynamic range is strongly affected by the length of the acquisition in time as
indicated by the squared term Td in the denominator. The reason why it is squared is
two-fold. One effect is the amount of noise let into the acquisition. Remember that the
actual signal – the incident wavefront – is contained in a very small time location, yet
the noise is spread over the entire acquisition. As the acquisition length increases, the
amount of noise increases with no increase in signal. If one knew where to look in the
waveform, the effect of long acquisitions could be counteracted by limiting or gating of
the waveform in the time domain. The second effect is the effect on averaging. Longer
acquisitions take more time to acquire.
Now some more complicated considerations that are not necessarily obvious. First is
the effect of the bandwidth limit fbw on the noise. In many cases, noise in equivalent
time sampler arrangements is essentially white. This is especially true if a major source
of the noise comes from quantization effects in the ADC. This means that all the noise
power is present up to the Nyquist rate Fse /2. In this case, fbw = Fse /2 and these terms
cancel so the dependence on noise bandwidth and equivalent time sample rate disappears
from the equation and the dynamic range is completely independent of the equivalent
time sample rate. This may seem counter-intuitive because increasing the sample rate
causes more noise to fall outside the spectrum of interest due to even noise spreading,
but this effect is fully counteracted by the increase in acquisition time and therefore
the decrease in the number of acquisitions that can be averaged. In the case where the
trace noise is specified with a bandwidth limit (as in most cases), the dynamic range is
actually penalized by 10 log (fbw / (Fse /2)), which seems unfair until you consider that
unless the Nyquist rate is set exactly equal to this limit frequency, then acquisitions are
needlessly oversampled (needless in theory, not necessarily in practice due to aliasing
considerations). To make a proper comparison of band limited and non-band limited
noise, one must compare using this adjustment.
From (11.16) therefore, after consideration, we see that there are a few basic ways to
improve the dynamic range in TDR measurements. These are:
(i) Increase the amplitude of the step.

(ii) Increase the actual sample rate of the system.
(iii) Decrease the acquisition length.

(iv) Decrease the portion of the acquisition over which reflections are considered.
(v) Decrease the noise in the pulser/sampler hardware.
(vi) Increase the frequency response of the sampler and the frequency content of the
pulser.
(vii) Decrease the length and losses in the cabling and fixturing.
All of these methods have been utilized to varying degrees in many TDR-based
S-parameter measurement instruments with 11.8 and 11.8 involving improved hard-
ware (as increasing the amplitude generally causes linearity problems). One particularly
interesting technique that effectively accomplishes 11.8 in an algorithmic fashion is the
use of wavelet de-noising techniques for lifting reflections from the noise [14].
11.9 Systematic errors and uncertainty due to measurement noise

in a network analyzer
Non-idealities in the source, receiver, and various interconnections (like direction cou-
plers, internal switches, and cables) introduce systematic errors in the measurements
made by network analyzers. Such systematic errors are modeled in different ways and
the model parameters are calculated by performing a calibration before making the DUT
measurements. The model is referred to as the error-term model. The coefficients of the
error-term model are collectively referred to as the error terms. Once the error terms have
been determined, the uncorrected DUT measurements (referred to as raw measurements)
are then combined with the error terms and the S-parameters of the DUT are calculated.
Calibration algorithms for two-port and n-port VNAs have been described in Chapters
8 and 9, respectively. An algorithm to calculate the S-parameters of a multi-port DUT
is described in [17]. The algorithm is general enough so that it works with any kind of
model for the systematic errors. An important issue that should be considered is how
the systematic error correction interacts with the noise in the measurement system. The
dynamic range is not as high in the TDR-based network analyzer as it is in the frequency-
based network analyzer. In such a system, both the calibration measurement as well as
the raw DUT measurements are corrupted by noise. The error terms calculated by such
noisy measurements are different from the actual systematic errors. In this section, we
provide a method to determine the interaction of the error terms and noise in the raw
DUT measurement when the final DUT S-parameters are calculated. We consider only
the one-port DUT here. For more detailed information refer to [18, 19].
11.9.1 Error propagation for a one-port DUT

Figure 11.12 is a setup for measuring the S-parameters of a one-port DUT. dut represents
the one-port S-parameters of the DUT. Sij represents the S-parameters corresponding
to the error terms of an error-term model of choice. msd represents the uncalibrated or
raw measurement of the DUT. The Sij are calculated by some calibration method. As an
S21
e
S11 S22 Γdut
Γmsd
S12
Fig. 11.12 One-Port DUT and error terms.
example, for the SOLT calibration technique, the Sij are calculated by connecting the
known short, open, and load calibration standards.
For the model in Figure 11.12, S11 corresponds to Ed – the directivity error term; S21
is chosen as one; S12 corresponds to Er – the reflection error term; and S22 corresponds
to Es – the source match error term.
The expression for the raw measured DUT S-parameters can be derived from the
signal flow diagram
S11 − dut (S11 S22 − S21 S12 )

msd = . (11.17)
1 − dut S22
Equation (11.17) can be modified to obtain the expression for DUT S-parameters from
the raw DUT S-parameter measurement and the knowledge of error terms:
msd − S11
dut = . (11.18)
msd S22 − (S11 S22 − S21 S12 )
Equation (11.18) is the expression for the DUT S-parameters when there is no mea-
surement noise and the error terms are known exactly. We would like to consider the
effects of measurement noise on the DUT S-parameters. To simplify the analysis, we
will consider the case when only the raw DUT measurement is noisy. Any noise in
the uncalibrated DUT measurement will cause an uncertainty in the calculation of the
S-parameters of the DUT. Suppose ε is the uncertainty in measuring msd and δ is the
uncertainty in calculating the S-parameters of the DUT, then,
msd + ε − S11
dut + δ = .
(msd + ε)S22 − (S11 S22 − S21 S12 )
Substituting the expression for msd from (11.17), and after some algebraic manipu-
lation, we have
(1 − dut S22 )2
δ = ε . (11.19)
εS22 (1 − dut S22 ) + S21 S12
As expected, if there is no uncertainty in the raw measured DUT S-parameters, i.e.
if ε = 0, then δ = 0, i.e. there is no uncertainty in the calculated DUT S-parameters
(assuming that the known error-terms represented the true systematic errors). Also, if
the error terms were such that there was no systematic error, i.e. S11 = S22 = 0 and
−10
−20
Magnitude response in dB
−30
−40
−50 s12
s22
−60
−70
−80
−90
−100
0 5 10 15 20 25 30 35 40
Frequency in GHz
Fig. 11.13 S22 and S12 example.
S21 = S12 = 1, then = ε, i.e. there is no uncertainty propagation and the uncertainty
in measurement is translated as the uncertainty in the calculated DUT S-parameters.
When none of the above trivial cases is true, i.e. there is an uncertainty in measuring
the raw DUT S-parameters, and the error terms are non-trivial, then (11.19) translates the
uncertainty in DUT measurement to the uncertainty in dut calculation. As an example
consider ε = 0.01, and further consider S22 and S12 as shown in Figure 11.13.
For S21 = 1 and dut corresponding to an ideal short (i.e. dut = −1), and if the
uncertainty in raw DUT measurement is ε, the uncertainty in DUT S-parameters can
now be calculated using (11.19). Figure 11.14 shows the effects of uncertainty prop-
agation for a non-ideal case. Here the trace with circles is the actual dut , the trace
with triangles is the DUT S-parameters with an uncertainty of ±ε (i.e. with ideal error-
terms), while the trace with crosses is the DUT S-parameters with uncertainty for the
non-trivial case.
There are multiple points to be noted for the case described above:
1. The uncertainty expression in (11.19) is a function of the uncertainty in the raw DUT
measurements, i.e. one must know what the uncertainty is in order to determine the
uncertainty in the DUT S-parameters. In general the actual uncertainty in the mea-
surement is not known, but a probability distribution of the uncertainty due to noise
is known. The problem then becomes estimating the distribution of the uncertainty
in the DUT S-parameters.
2. Although it is not directly evident from (11.19), S22 plays an important role in
the uncertainty propagation. As an example, instead of choosing an S22 as shown
0.8
0.6
0.4
0.2
S11in dB
−0.2
−0.4
−0.6
−0.8
0 5 10 15 20 25 30 35 40
Frequency in GHz
Fig. 11.14 dut ± δ for dut = −1, ε = ±0.01, S21 = 1, S12 and S22 as shown in Figure 11.13.
0.8
0.6
0.4
0.2
S11 in dB
−0.2
−0.4
−0.6
−0.8
0 5 10 15 20 25 30 35 40
Frequency in GHz
Fig. 11.15 dut ± δ for dut = −1, ε = ±0.01, S21 = 1, S22 = 10−5 , and S12 as shown in Figure 11.13.
in Figure 11.13, if we have an S22 that is 100 dB down throughout the frequency
of interest (almost an ideal S22 ), then the uncertainty in the DUT S-parameters is
shown by the curve with squares in Figure 11.15. For comparison purposes, the older
uncertainty is shown with crosses.
It is evident that in the high-frequency region, where the two S22 differ, the uncer-
tainty is significantly higher for the non-ideal S22 . The reason for this increase in
amplification is the sign of dut S22 . In the example provided above, the two are
of opposite signs, making the numerator larger than one and thereby increasing the
uncertainty in the DUT.
3. It should be noted that the expression in (11.18) is the uncertainty in DUT S-
parameters only due to uncertainty in the raw DUT S-parameter measurement. A
more general case needs to include the uncertainty propagation due to noise in the
calibration measurements as well. As expected, the complexity of the math increases
with different calibration techniques and the number of ports in the DUT. Complete
software for the evaluation of the uncertainty, taking into account all the contributions,
actually exists [18, 19].
11.10 Conclusions
This chapter has dealt with TDR techniques for network measurements. The hardware
architecture of TDR instruments including the sampling system, the pulser, and the
timebase have been described. The main element that effects the accuracy of time domain
measurements, the noise, has then been quantitatively discussed. It has been shown how
it is the main source of uncertainty in time domain derived S-parameters.
The sources of dynamic range degradation in TDR systems and the key design areas
for improvement have been presented.
Finally, a quantitative consideration of how S-parameter measurements are affected
by noise, when it is the primary source of error, has been given.
Acknowledgments
The authors wish to thank Ron Ramsey of Picosecond Pulse Labs, Dr. Steve Ems,
Dr. James Mueller, and Dr. Leonard Hayden of Teledyne LeCroy for their input in
describing the operation of pulsers and samplers.
References
[1] M. J. W. Rodwell, D. M. Bloom, and B. A. Auld, “Nonlinear transmission line for picosecond
pulse compression and broadband phase modulation,” Electronics Letters, vol. 23, p. 109,
Jan. 1987.
[2] R. J. Baker, D. J. Hodder, B. P. Johnson, P. C. Subedi, and D. C. Williams, “Generation of
kilovolt-subnanosecond pulses using a nonlinear transmission line,” Measurement Science
and Technology, vol. 4, pp. 893–895, 1993.
[3] J. R. Andrews, B. A. Bell, and E. E. Baldwin, “Reference flat pulse generator – Technical
note,” National Bureau of Standards, Boulder, CO. National Engineering Lab, Oct 1983.
Report Number NBS-TN-1067.
[4] A. Agoston, J. B. Rettig, S. P. Kaveckis, J. E. Carlson, and A. E. Finkbeiner, “Dual channel

time domain reflectometer,” July 1988. U.S. Patent 4 755 742.
[5] A. Agoston and J. E. Carlson, “Fast transition flat pulse generator,” July 1988. U.S. Patent
4 758 736.
[6] M. Kahrs, “50 years of RF and microwave sampling,” IEEE Trans. Microw. Theory Tech.,
vol. 51, pp. 1787–1805, June 2003.
[7] R. Miller, “Waveform translator for DC to 75 GHZ oscillography,” June 2001. U.S. Patent
6 242 899.
[8] S. Ems, S. Kreymerman, and P. J. Pupalaikis, “Time domain reflectometry in a coherent
interleaved sampling timebase,” September 2010. U.S. Patent Application 12/888 550.
[9] R. L. Graham, D. E. Knuth, and O. Patashnik, Concrete Mathematics: a foundation for
computer science. Addison-Wesley Professional, 1994.
[10] L. A. Hayden and V. K. Tripathi, “Characterization and modeling of multiple line interconnec-
tions from TDR measurements,” IEEE Trans. Microw. Theory Tech., vol. 42, pp. 1737–1743,
September 1994.
[11] D. A. Smolyansky and S. D. Corey, “PCB interconnect characterization from TDR
measurements,” Printed Circuit Design Magazine, May 1999. TDA Systems App. note
PCBD-0699-02.
[12] W. L. Gans and N. S. Nahman, “Continuous and discrete Fourier transform of step-like
waveforms,” IEEE Trans. Instrum. Meas., vol. IM-31, pp. 97–101, June 1982.
[13] A. M. Nicolson, “Forming the fast Fourier transform of a step response in time domain
metrology,” Electron. Lett., vol. 9, pp. 317–318, July 1973.
[14] P. Pupalaikis, “Wavelet denoising for TDR dynamic range improvement,” in DesignCon,
IEC, February 2011.
[15] P. Pupalaikis, “The relationship between discrete-frequency S-parameters and continuous-
frequency responses,” in DesignCon, IEC, February 2012.
[16] M. T. Jong, Methods of Discrete Signal and System Analysis. McGraw-Hill, 1982.
[17] P. Wittwer and P. J. Pupalaikis, “A general closed-form solution to multi-port scattering
parameter calculations,” in 72nd ARFTG Conference Digest, p. 137, 2008.
[18] A. Ferrero, M. Garelli, B. Grossman, S. Choon, and V. Teppati, “Uncertainty in multiport
S-parameters measurements,” Microwave Measurement Conference (ARFTG), 2011 77th
ARFTG, pp. 1–4, June 2011.
[19] METAS VNA Tools II [Online]. Available: http://www.metas.ch/metasweb/Fachbereiche/
Elektrizitaet/HF/VNATools/VNATools.html.
Part IV
Nonlinear measurements
12 Vector network analysis for
nonlinear systems
Yves Rolain, Gerd Vandersteen, and Maarten Schoukens
12.1 Introduction
The measurement of the nonlinear behavior of microwave systems and components has
evolved a lot over the last years. Starting from instrument prototypes, vector network
analyzers for nonlinear systems (NVNA) have now entered the product lines of all the
major instrumentation vendors. The major challenge for the scientific community is to
embed these devices in the mainstream design and characterization of nonlinear devices
and circuits.
As the NVNA is still young, most currently active professionals did not experience
NVNA technology during their education or their career. Therefore, it is extremely impor-
tant to clearly define what can be expected from an NVNA. There is a need for an explana-
tion of what an NVNA is and is not. Explaining the limitations of the NVNA technology
is also extremely important, as this can avoid false expectations and deceptions.
This text has the ambition to take a small step in this direction. This is why much effort
is spent in the first sections of this chapter in drawing the big picture around the NVNA.
Our hope is that this might help practitioners to position the NVNA and to obtain some
intuition about the actual measurements the NVNA makes.
The remainder of the text explains the ideas behind the different instruments that have
NVNA capability. The setups are very different, but the measurements they make are
very similar. The key idea is that to characterize a nonlinear device under test, one needs
to measure the complete spectrum (amplitude and phase) of all the port quantities (waves
or voltages and currents) that are present at all the ports of the device.
Remember that “A journey of a thousand miles begins with a single step.” To avoid
the reader becoming overwhelmed by new jargon and concepts, we will start from the
S-parameter formalism and the linear time-invariant (LTI) system framework to outline
the similarities and the differences with the nonlinear framework.
12.2 Is there a need for nonlinear analysis?
12.2.1 The plain-vanilla linear time-invariant world

S-parameters have been the driving force behind RF and microwave design and char-
acterization of the last 40 years [1]. Their ability to describe a distributed circuit that is
inherently complex and hard to understand in an intuitive way proves to be an efficient
design tool that can also validate a design or a circuit.
310 Yves Rolain, Gerd Vandersteen, and Maarten Schoukens
Unfortunately, S-parameters also have their limitations. The basic assumption for their
validity is that the circuit or system under test remains linear and time-invariant [1]. Put
in layman’s words, this means that the superposition principle holds: the response of a
system to a sum of two inputs is the sum of the responses to the individual signals and
the response scales proportional to the input(s).
Common sense tells us that this assumption is never valid in general. When the input
power is increased without bounds, any practical system will break down and therefore
is not LTI. Linearity always comes at a price, which is the acceptance of the small-signal
operation paradigm. This type of operation assumes that the input signal is small enough
to ensure that the response of the system stays close to linear.
Taking a step backwards to see the general picture leads to the striking conclusion
that even our most basic tools are not always valid. They come with a set of assumptions
that we have to meet to obtain reliable results. Even if this was probably very clear to
practitioners in the early days of S-parameters, the wide dissemination and the general
success of S-parameter-based design and characterization has diluted the feeling that
these hypotheses do indeed matter.
12.2.2 Departure from LTI

The push of portable telecommunication towards power-efficient designs has continu-
ously weakened the validity of the linearity assumption for practical designs with a long
battery lifetime. The S-parameter framework first broke at the output of power amplifiers,
where S22 was no longer power-independent at the higher power levels [2].
The engineer’s way to overcome this problem is to extend the LTI framework to
include the new situation. Keeping the input power constant was not sufficient to restore
the reliability of the results. The predictive power could be restored at the cost of splitting
S22 into two contributions. The first one is proportional to the incident wave at the output
port. This is the normal S22 term. The addition of a , S22 term that is proportional to the
complex conjugate of the incident waves at port 2 solved the problem. This resulted in the
so-called hot-S22 . This was the first breach in the LTI based S-parameter characterization
framework. It was followed by many others.
12.2.3 Measuring a non-LTI system

This new situation leads to new challenges in the measurement world. Wave ratios
provided by the S-parameters no longer suffice to characterize a system: it becomes
mandatory that waves are measured separately and accurately. This results in a shift of
measurement paradigm from a relative measurement (which is pretty easy) to an absolute
measurement (which is pretty hard).
But more is needed to measure nonlinear systems. The LTI framework ensures that
measurements can be performed one frequency at a time without jeopardizing the quality
of the characterization. The response to any (periodic) signal can be obtained by the
Fourier series decomposition of the input signal and the superposition of the responses
to the individual sine waves in this Fourier series. The most complex experiment one
Vector network analysis for nonlinear systems 311
b1 a2
LTI
a1 b2
Fig. 12.1 Definition of the waves around a two-port device.
needs to set up for a complete characterization of a LTI system boils down to a sine
wave test at one single frequency. There is no need for the measurement of information
at different frequencies at the same time [3].
Once one departs from LTI systems, this is no longer true. Consider the most simple
departure of linearity for a DUT: a polynomial static cubic nonlinearity. The equations
governing the output waves of the ideally matched DUT are given below:
7
b2 (t) = αa1 (t) − βαa13 (t)
(12.1)
b1 (t) = γ a2 (t).
The incident waves ai (t) and the reflected waves bi (t) at port i = 1, 2 are defined as
in Figure 12.1. The constants α, β, and γ ∈ R. A user then measures S21 with a standard
VNA. The VNA performs a frequency sweep on the DUT. It is excited by a sine wave
at the standard power level that is applied at the input port of the device to obtain the
dynamic S21 (ω) response.
A novice instrumentation user is tempted to believe that this measurement does indeed
represent the behavior of the device. It is easy to be fooled by a measurement that has a
high signal-to-noise ratio and is very repeatable. If a second measurement is then taken at
a different power level, it will result in a different behavior. Nonlinearity of the device can
be a possible explanation, and a skilled instrumentation engineer will grasp a spectrum
analyzer to view the complete spectrum.
VNA measurements hence are both extremely powerful but their outcome is very
dependent on the validity of the LTI hypothesis. An engineer’s solution to this problem
is to measure the validity of the LTI hypothesis separately. The VNA is therefore extended
to allow a power sweep at one frequency. Nonlinearity of the device will then result in a
deviation from a constant gain versus the power. Designers need to be able to assess the
magnitude of these perturbations, especially for high-performance designs.
12.2.4 Figures of merit to characterize the nonlinearity

To enable an easy comparison of the order of magnitude of the nonlinear disturbance for
different systems, Figures of Merit (FOMs) have been introduced. The 1 dB compression
point registers the (input or output) power level at which the actual gain is reduced by
1 dB compared to the linear gain. The FOM reduces a complete function of the power
to a single number. Therefore it is clear that the comparison of this FOM over different
systems cannot be completely fair. As a measurement addict, it is tempting to measure the
complete gain versus power dependency instead. This is pretty easy to realize even with a
simple and classical VNA, as the 1 dB compression point is obtained from (relative) gain
measurements taken at the fundamental frequency only. Is the measurement instrument

for the nonlinear behavior a simple extension of the classical VNA?
To show that this is not the case, let’s move to a different FOM, the Third Order
Intercept (TOI) point. Things are a bit more involved here [4]. The TOI is defined as
the intersection of the extrapolation of the linear gain at the fundamental tone at f0 and
the extrapolation of the harmonic response taken at 3f0 . As this result is obtained as the
intersection of two extrapolated curves, it is again clear that it is not a really fair measure
of the nonlinearity either. It does not take into account the saturation of the distortion
at higher power levels. Again, the instrumentation engineer is tempted to measure the
response curve at the fundamental and the third harmonic in a power sweep. This seems
to be a good way to obtain more and better information about the nonlinearity.
This time the measurement is not a simple extension of the VNA measurement. The
TOI requires the combination of measurements taken at two different frequencies: f0 and
3f0 . This requires either a spectrum analyzer measurement or a VNA that is capable of
measuring at the fundamental and at a harmonic frequency simultaneously while exciting
at the fundamental only.
In practice, the intermodulation product is also often used to measure the nonlinearity.
The intermodulation product is measured using a 2-tone excitation signal (two sine waves
whose frequency is really close), as this allows the measurement of the nonlinearity for
narrow-band systems where the harmonic distortion lies out of the pass band of the
device and is therefore attenuated. The measurement then is always performed with a
spectrum analyzer.
12.3 The basic assumptions
The characterization of all nonlinear systems is both a much too ambitious and a pretty
foolish goal. This can most easily be felt if this goal is translated to a totally different
field of science, namely animal biology. It takes a lifetime to understand even a small
part of the behavior of an elephant. Many biologists have spent their careers trying to
understand this. A biologist will therefore certainly never try to describe the biology
of the non-elephants. Even if this example looks stupid, taking one step back shines a
different light on the problem. Because the LTI framework does not meet our demands,
we are tempted to replace it by its complement, the class of nonlinear systems.
This is certainly not a very smart choice, as each of us can think of a system that is
not linear and behaves in a totally crazy way: chaotic systems, uncertain systems, and
systems that contain hysteresis are all nonlinear, but are also all very different from our
well-known LTI class. They may be so different that they are almost impossible to use
in a practical design.
To be successful, we will therefore extend the class of systems in a more directed
way. We will consider systems that are close to the LTI behavior, but allow for saturation
effects and large-signal operation [5].
Selecting a system class is not enough to enable a correct measurement of the behav-
ior of the system. Even for a practical S-parameter measurement, there are conditions
imposed on the excitation signal. Imposing small-signal operation is an often tacit

assumption for a VNA. For a NVNA, we will investigate which assumptions are needed
to obtain the measurements we are after.
12.3.1 Restricting the class of systems: PISPO systems

The class of the LTI systems is defined as the class of systems that obey the superposition
principle. To expand this class gradually, we will remove the superposition principle, and
replace it by a criterion that is sensible for the systems that one encounters in practical
applications.
Consider for example a power amplifier in a telecommunication link. In small-signal
operation, the amplifier output is a sine wave when fed by a (small signal) sine wave.
When we increase the amplitude of the input, the shape of the wave starts to deviate
gently. Increasing the input power increases the distortion in the wave, but it maintains
its periodicity. We can formalize this period-maintaining behavior as follows.
d ef i n i t ion 12.1. A system that obeys the period-maintaining principle belongs to

the PISPO class of systems.
What is a PISPO system?

The PISPO class is used to extend the LTI class in the context of this chapter. Therefore,
it is important that we clearly understand its properties. Intuitively, we see a link between
the PISPO class and the Volterra systems [5]. Volterra systems also allow for a gradual and
gentle departure from linearity. The key idea behind the Volterra model is the extension
of the impulse response of a LTI system to a multilinear impulse response, the kernel
function. For a second-order nonlinearity, the Volterra kernel h2 (t1 , t2 ) links the second-
order output signal y2 (t) to the input signal u (t) as follows:
+∞ +∞
y2 (t) = h2 (τ1 , τ2 )u(t − τ1 )u(t − τ2 )dτ1 dτ2 . (12.2)
−∞ −∞
A good introduction to Volterra systems can be found in [5]. The problem is that
Volterra systems have a bad reputation because of their poor convergence properties for
strongly nonlinear systems. Fortunately, PISPO systems do not suffer this problem. Using
a least-squares fit, a Volterra system can model all PISPO systems with an approximation
that has a perfectly regular behavior. The solution stems from the fact that the PISPO
system approximates the hard nonlinearity in a mean-squared sense. Even if the system
under test is a discontinuous static nonlinearity, the PISPO class provides a least-squares
approximation for it.
Intuitively, one is tempted to believe that the PISPO class only contains systems that
do not modify the frequency content of the excitation signals: namely amplifiers and
attenuators. More often than not however, we also need to characterize systems that
translate the frequency content, such as mixers or detectors.
SRF SIF
kΔƒ
ƒ ƒ
ƒRF ƒRF – ƒLO ƒRF + ƒLO
SLO
ƒ
ƒLO
Fig. 12.2 A PISPO mixer as a three port.
A PISPO mixer
A frequency translating system is not a two-port DUT, but rather a three-port one. Con-
sider the case of an ideal mixer [6] as in Figure 12.2. The RF port is excited by a
multiple-tone periodic signal that has a carrier frequency fRF and a modulation period
TRF . The LO port is excited by a sine wave having a frequency fLO . The power spectra
of the inputs are represented in Figure 12.2.
To determine whether or not this system belongs to the PISPO class, we need to be
able to check whether a periodic excitation leads to a periodic output signal sI F (t) that
has the same period. It turns out that this is not so trivial as it looks.
Mathematically speaking, the input signal can only be a periodic signal if there exists
a joint period for the signal pair sLO (t), sRF (t). This requires that the two signals
simultaneously repeat perfectly after the common period Tin :
-
sLO (t + nTin ) = sLO (t)
n∈Z (12.3)
sRF (t + nTin ) = sRF (t)
This condition can then be translated into a condition on the periods TRF and TLO of
the two input signals:
Tin = kTRF = lTLO k, l ∈ N (12.4)
which means that the period of the input signals taken separately needs to be commen-
surate. In the frequency domain, this can be reformulated to the more commonly used
requirement for commensurate frequencies:
RF LO
fin = = . (12.5)
k l
As a result, we can only determine whether the system belongs to the PISPO class if we
can obtain one common frequency grid with spacing fin for both signals
-
fRF (k) = (lRF + nRF (k)) f
(12.6)
fLO (k) = (lLO + nLO (k)) f
Slow mode
ain bout
Static NL
Fig. 12.3 Slow modes in a PISPO framework.
with lRF , lLO , nRF (k), nRF (k) ∈ N, and f ∈ R. This joint period is now used as the
period of the input signal when we try to determine if the period of the input and the
output waves do match.
PISPO looks beyond static nonlinearity

Many nonlinear models and measurement approaches somehow rely on the presence of
a static nonlinearity [5]. This hypothesis also comes naturally into play for many applica-
tions. In practical circuitry, where active elements are connected to off-chip components,
the linear dynamics introduced by the connections are often an order of magnitude slower
than the on-chip dynamic effects. As a consequence, it is very tempting to assume that
the nonlinearity is static or quasi-static. In practice, this proves to be a strong hypothesis,
as very slow time constants (at a time scale of microseconds) appear around a signal that
has a GHz frequency (time constants at the timescale of nanoseconds or less).
The origin of the slow time constants can be understood by the following simple
example. Consider an amplifier that operates under compression. Besides its input and
output ports, the device is also connected to the outside world via its DC bias port. This
connection’s impedance is not very relevant at RF frequencies, because the RF signal
is carefully blocked by design at this port. However, the IF impedance is also known to
have a significant contribution in the RF operation of the device [7], [8]. This impedance
influences the slow modes of the amplifier, which are dynamic effects that appear around
the RF frequency, but have time constants at the IF time scale. How is this possible?
To explain the slow modes we consider a very simple model, that uses the nonlinearity
introduced before in (12.1). This is used to illustrate the behavior of the main path of the
amplifier. To model the path that connects to the DC bias port, an additional parallel path
is introduced as shown in Figure 12.3. Note that the resulting system is still a PISPO
system.
12.3.2 Influence of the excitation signal

From here on, only PISPO systems are considered. Looking back to the LTI framework,
it is clear that both the system and the excitation signal have to obey restrictions. Now
that the system restrictions are made clear, it is time to take a look at the influence of the
excitation signal on the system behavior.
It is evident that it is neither possible nor useful to build a different model for each
different excitation signal. We will therefore delimit classes of excitation signals for
which the system behaves in a similar way. This leads to sets of signals that are grouped
based on their power spectrum or power spectral density and their PDF.
Does the signal choice matter?

To illustrate the change in behavior that results when changing the excitation signal,
a series of excitation signals is fed to a PISPO system. In the first series of tests, the
system is assumed to be the static nonlinear system, described in (12.1). In a thought
experiment, we excite this system with a sine wave. The sine wave is said to have a
frequency of 1 a.u. and an amplitude that is large enough to excite the nonlinearity up to
high compression levels. Note that the frequency of the sine wave does not matter here
as the system is assumed to be perfectly static. Plotting the input and output time signals
in an X-Y plot yields the plot of Figure 12.4. This type of plot is sometimes called a
Lissajous figure. This shows the nonlinearity of the DUT that is operated in very deep
compression. Power levels even start to decrease with increasing input power! When the
frequency of the excitation is changed, the response of the device remains unaltered.
In a second thought experiment, we cascade the static system of (12.1) with a bandpass
filter. This filter mimics the dynamics of a real RF system. Here, we have chosen a fourth-
order bandpass Butterworth response with a pass band ranging from 1 a.u. to 2 a.u. The
amplitude response of the filter is given in Figure 12.5.
The response of the tandem connection to the sine wave excitation used in the first
experiment is shown in Figure 12.6. There is a clear difference between these responses,
due to the dynamics of the system. If the system were an LTI system, it would now be
fully characterized.
We then perform a second series of experiments to show the dependence of the system
response on the signal class. We will now excite the static and the dynamic system again
with a different excitation signal, namely a narrow-band multisine signal. The signal
consists of 512 equally spaced spectral lines of equal amplitude located in a frequency
band ranging from 0.8 to 0.98 a.u.
0.5
bout [a.u.]
−0.5
−1
−1 −0.5 0 0.5 1
ain [a.u.]
Fig. 12.4 Response of a static PISPO system to a sine wave (response is in arbitrary units [a.u.]).
|H(f)| [dB] 0
−20
−40
−60
0 1 2 3 4 5
f [a.u.]
Fig. 12.5 Frequency response of the cascaded LTI system.
0.5
bout [a.u.]
−0.5
−1
−1 −0.5 0 0.5 1
ain [a.u.]
Fig. 12.6 Response of a dynamic PISPO system to a sine wave (response is in a.u.).
0.5
bout [a.u.]
−0.5
−1
−1 −0.5 0 0.5 1
ain [a.u.]
Fig. 12.7 Response of a static PISPO system to a multisine wave (response is in arbitrary units [a.u.]).
As shown in Figure 12.7, the response of the static system to a multisine or a sine
wave signal is perfectly identical. There is no dependence on the properties of the signal.
On the contrary, the response of the dynamic system, shown in Figure 12.8, no longer
resembles the response obtained for the sine wave!
0.5
bout [a.u.]
−0.5
−1
−1 −0.5 0 0.5 1
ain [a.u.]
Fig. 12.8 Response of a dynamic PISPO system to a multisine wave (response is in arbitrary units [a.u.]).
What does that tell us about nonlinear measurements? It shows that there is a major
difference between an LTI system and a PISPO system: the measurement of the system
response no longer depends on the system class alone, but also on the excitation signal
used. A measurement is only valid and reproducible when both the system class AND
the excitation signal are specified.
This certainly looks like very bad news. Based on the previous results alone, it is
tempting to conclude that the measurement is only valid for the particular excitation
signal used. This would reduce the predictive capacity of the measurement to zero.
Fortunately, the situation is not as bad as it looks.
Specifying the class of signals

A measurement is often the response to an engineer questioning the behavior of the
system that operates or will operate in a certain context. This means that there is a lot of
prior knowledge present about the possible excitation signals of the device.
Most applications will be constructed under limiting assumptions on the power spec-
trum or the power spectral density of the applied excitation. For telecommunication
applications for example, a transmitter is designed to support a certain standard. This
fixes the power spectral density of the signals to the spectral mask that is associated with
that particular modulation. When designing a mixer LO driver for example, the spectrum
of the signal is bounded with known bounds on the frequency and the amplitude.
We know that a general model that is valid for all signals is out of reach because of the
lack of a general theory for the PISPO system. As it does not make sense to measure the
behavior of the system outside this application bound class of signals, we will restrict
the signals to have a fixed power spectral density and will sweep the power of the signal
over the allowable power range.
What if the aim of the measurement itself is the identification of a model for the
system? The challenge there is to capture the behavior that matters to the model and the
use one wants to make of that model. Any model relies on basic hypotheses to be valid.
The challenge in the selection of a signal class is to make the class narrow enough to
meet these hypotheses and wide enough to remain applicable in practice.
The excitation signal for a single experiment needs to have a fixed power spectrum.
For a sine wave excitation, this leaves us with a class containing a single signal: a sine
of fixed amplitude and frequency.
For a modulated signal, there is a larger range left to choose the input signal from.
For a fixed level of the total signal power, some type of modulation signals have a fixed
power spectrum and a data-dependent phase spectrum. Others have a data-dependent
power and phase spectra.
In the context of a measurement, this data dependency is conceptualized as a random
variation of the phase and the amplitude of the excitation signal over a set of possible
values. A single measurement is then performed on one realized signal in this class.
The power spectral density alone is not sufficient to define a signal class. This can
intuitively be understood by the following thought experiment: consider two signals
with the same power spectral density, but a different behavior in the time domain. The
first signal is a swept sine. The second signal is a Gaussian noise source with a fixed
power spectrum. When these waveforms with equal power excite a nonlinear system, the
response of the system will be quite different. The level of the nonlinear contributions in
the output signal can be up to an order of magnitude higher for the swept sine signal. To
understand this behavior, we will look at the histogram of the time signal. This measured
quantity represents a sampled version of the Probability Density Function (PDF) of the
signal. The PDF describes the distribution of the different amplitude levels present in the
signal (both signals are normalized to contain the same power). The histogram is shown
in Figure 12.9 for signals that are 128 000 samples long.
The PDF of the signals has an almost inverse behavior. The Gaussian noise signal
spends most of its time at low amplitude levels. Therefore, it excites the nonlinearity
gently most of the time. From time to time, a peak value appears. The swept sine signal,
on the other hand, spends most of its time at high amplitude levels. The nonlinearity is
therefore strongly addressed during the major part of the excitation signal. This increases
the level of the nonlinearity to much higher values than for the Gaussian noise signal.
10000
hist
5000
0
−4 −2 0 2 4
a [a.u.]
Fig. 12.9 Histogram of the swept sine (full line) and the Gaussian noise (dash-dot line) signal of equal
power. The horizontal axis represents the amplitude in the time domain, the vertical axes the
occurrence count.
This example clearly demonstrates that the PDF of the signal severely influences the
nonlinear distortion too.
12.3.3 The definition of the nonlinear operating point

Repeatable measurements are the cornerstone of any characterization and/or modeling
effort. To characterize a nonlinear system, the conditions that are required to enforce the
repeatability of a characterization are much more involved than in the LTI case. First,
one needs to restrict the class of systems that are to be measured. In this chapter, we
restrict the class of systems that, when excited with a periodic signal, produce a periodic
output that has the same period: the PISPO class. In addition to this, the class of input
signals is to be reduced too. The signal class is reduced to signals with a fixed power
spectrum and a fixed PDF.
It is the coupling of a fixed system class and a fixed signal class that enables one to
obtain repeatable measurements for a nonlinear system. This shift in paradigm looks
artificial at first glance. Note however the large similarity between this way of working
and the setting of an operating point in an S-parameter transistor characterization. To
stress this concept, this coupling of a fixed system and a signal class is called the nonlinear
operating point throughout this work.
12.4 Principle of operation of an NVNA
This section is devoted to the general principles that govern the operation of an NVNA.
First, we will look for an ideal instrument that is capable of nonlinear characterization.
Next, we will shortly touch on the requirements imposed by the use of the discrete
Fourier transform to obtain the spectral measurements. Finally, the challenges posed by
the calibration of the NVNA are covered.
12.4.1 Introduction
Now that the class of the systems that we want to characterize and the possible test signals
are defined, we can find out the influence of these assumptions on the capabilities of the
instrumentation. The major issue lies in the absence of the superposition principle for
the DUT: if the response to a certain class of excitation signals is to be known, the
measurements have to be taken with an excitation that belongs to this class.
Engineering practice learns to start from known techniques to create something new.
Can existing instruments be extended, adapted, or combined to handle the nonlinearity?
The nonlinearity mixes the spectral information of more than one frequency to generate
the output response. Hence measuring the complete spectrum in one single measurement
is the way to go. The most obvious solution lies in a high-speed time-domain measure-
ment of the sampled wave data, combined with a DFT to calculate the spectrum [9, 10].
The behavior of the nonlinearity depends on the properties of the excitation signal.
When an arbitrary signal can be generated, the excitation can be adapted accurately to
the characterization needs. An instrument that combines a time domain data acquisition
and an arbitrary waveform generator exists at IF frequencies. It is called a FFT analyzer

or a DSA [9]. The challenge we face is to port the functionality of this instrument to the
RF domain!
In the microwave frequency range, signals can be acquired in the time or the frequency
domain. Classical frequency-domain instruments are a VNA or a spectrum analyzer. In
time domain measurements, a real-time oscilloscope or a sampling oscilloscope [11], [12]
is commonly used. Signals can be generated by a sine wave generator, a modulated
generator, or an arbitrary waveform generator. Clearly, an oscilloscope and an arbitrary
waveform generator or a modulated generator can mimic a DSA. We show below that
the VNA can also deliver solutions, albeit at the cost of hardware extensions.
Microwave sources have for a long time been limited in their ability to generate
modulated waveforms: see Chapter 3. In the last years however, increasing numbers
of microwave sources have provided at least some modulation capability. Sources are
continuously moving away from CW generation alone. They provide a (complex) mod-
ulation over ever-increasing bandwidths. The kind of signals generated by such a source
is extremely suited to mimicking the behavior of real telecommunication signals while
maintaining the capability to generate purely periodic waveforms. This is a most wel-
come feature for an easy and correct transformation of the waveforms to the frequency
domain. Some acquisition systems require a trigger signal to operate properly. This is a
time marker (often a block pulse) that repeats at the period of the modulated waveform.
This trigger defines a fixed point in the period of the modulated signal. We see below
that this feature is mandatory for some of the instrumentation setups to work properly.
In all these RF measurements, the error correction and assessment is a key issue. It
is clear that not all measurements require precision at the level of a standards lab. But
even if the demand for accuracy is modest, it still needs to be reached. This requires the
existence of a calibration and a verification procedure of some kind. We will also have
a quick look at this aspect of the nonlinear characterization.
The ideal microwave instrument for the nonlinear characterization measures complete
wave spectra without distortion. It has a flexible signal generator, that can simultaneously
impose the spectral content and the PDF of the excitation signal with a high spectral
purity. All these features are also perfectly synchronized in frequency and in phase
through the use of a common reference clock (labeled CLK in Figure 12.10). This ideal
instrument needs now to be approximated with real setups. As a new device deserves a
new name, this device will be called the NVNA from now on.
12.4.2 Basic requirements for nonlinear characterization

A PISPO system is a system with a mixed behavior. Its nonlinearity aspect produces wave-
form shapes that are best characterized in the time domain, while its dynamic behavior is
best visualized in the frequency domain. The NVNA therefore needs to measure waves
in a configuration that allows an easy transformation between both domains.
To transform measurements between the time and frequency domain, the DFT is used.
Most often, we will choose to work with the FFT algorithm that is both numerically
efficient and stable.
4-channel ACQ
a1 b1 a 2 b2
GEN1 GEN2
DUT
CLK
Fig. 12.10 The general NVNA setup. Note that CLK is a common reference clock used to avoid spectral
leakage.
The power of the DFT can hardly be over-estimated, but being a real-world method
it comes with a set of hypotheses that have to be met accurately to avoid problems. The
DFT is prone to two types of errors in the spectral domain.
First, the signals to be transformed have to obey the Shannon-Nyquist theorem. Prac-
tically speaking, this theorem states that the discrete-time representation of a waveform
is only unique (and valid!) if the signal is sampled fast enough. Sampling at a lower rate
results in spectral folding or aliasing. Spectral components then appear at frequencies
that are different to the original frequency. Aliasing is avoided by external filtering before
the conversion of the signal to discrete-time. Sometimes, we will violate this hypothesis
on purpose to obtain very large bandwidth measurements.
Requirement 1 The measured signals are band-limited to avoid unexpected aliasing of
the spectra.
Secondly, the DFT is also prone to spectral smearing or spectral leakage. Leakage
turns discrete spectral lines in a distribution of spectral power over a (larger) number
of spectral lines. This ruins the spectral resolution, and results in errors in the spectrum
that can be as large as 30%. Leakage avoidance is theoretically simple to explain but it
is hard to implement. A leakage-free spectrum is obtained when the measurement time
window is an integer multiple of the period of the excitation signal,
Tacq = mTs = nTexc m, n ∈ R, (12.7)
with Texc the period of the excitation signal and Tacq the time span of the measurement.
This requires synchronization between the acquisition and the generator of the NVNA.
This is seen if (12.7) is transformed to the frequency domain:
Fs Fexc
Facq = = m, n ∈ R. (12.8)
m n
The bad news here is that the DFT is extremely sensitive to the presence of errors in
this equation, even if they are very small.
How can one impose this requirement with high accuracy in practice? This is only
possible if the time reference (the reference clock) used by all the parts of the instrument
is the same. The question is to know whether this condition is also sufficient. This can
best be analyzed using a small example.
Consider an NVNA that excites a DUT with a sine wave at a frequency of 1 GHz. The
wave signals are acquired at a sampling rate Fs = 1/Ts = 5GHz. Generator and acquisition
both run from the same 10 MHz reference clock (this is common for instrumentation).
Assume that 500 samples are acquired, then
⎧
⎪ 500
⎨Tacq = = 0.1 μs
5 GHz (12.9)
⎪
⎩Texc = 1 = 1 ns,
1 GHz
and the condition (12.7) is met. When we use a scope and a sine wave generator to
perform the actual measurements, leakage is very likely to remain present. The scope and
the generator multiply the reference frequency by a factor of 500 and 100, respectively.
Such a high multiplication factor can be obtained by a phase-locked loop. The device
frequencies will be accurately locked to the reference, but the uncertainty on the phase
(also known as phase noise) will be increased by a factor that is roughly proportional to the
frequency multiplication.As a consequence, a (slow) drift of the phase of the acquisition’s
Fs with respect to the generator’s Fexc becomes very hard to avoid in practice. This results
in the presence of leakage again, especially if the period of the signal becomes larger.
Requirement 2 Leakage-free measurements are obtained when the sampling frequency
Fs of the acquisition and the frequency Facq of any generated spectral line obey
Facq = Fms = Fnexc m, n ∈ R. This requires that all the frequencies in a measurement
are commensurate to a fundamental frequency.All these frequencies are also phase-
coherent to the reference frequency of the instrument or the setup.
12.4.3 A calibration for nonlinear measurements

The measurement quality of a RF instrument is always determined by its calibration.
The S-parameter calibration calibrates S-parameters, which means that it removes the
systematic errors from the measured wave ratios. As a nonlinear characterization requires
the knowledge of waves rather than wave ratios we expect that the S-parameter calibration
needs an extension. In which sense is the calibration to be extended?
As a “nonlinear” measurement is no longer a ratio of waves taken at one single fre-
quency, the S-parameter calibration of the VNA will no longer completely ensure correct
measurements. Trouble can be expected because the S-parameter calibration assumes
both the VNA and the DUT to be LTI systems. The linearity of the VNA enables a cali-
bration to be performed one frequency at a time. The linearity of the DUT calls for the
calibration of wave ratios only. As explained in Chapter 9, one entry in the compensation
matrix can therefore be freely chosen.
For the NVNA, we will maximally re-use the already existing calibration. Thereto,
we assume:
Assumption 1: The acquisition part of the NVNA is an LTI system

This is not a strong assumption. It can quite easily be met if we provide attenuation
before the acquisition channels of the NVNA. Imposing the linearity is then a matter
of keeping the power budget at the input of the channel in the zone where that channel
combines a good SNR and a good linearity. This has to be imposed for each measurement.
Assumption 1 implies that the relation between measured and calibrated waves is LTI.
The calibration can therefore still be performed one frequency at a time, and remains
independent between frequencies. We will again use a practical example. Consider a
classical 8-term calibration as in [13]. The wave correction at an angular frequency
of ω is ⎡ ⎤ ⎡ ⎤
a1 b
% & m1
⎢b1 ⎥ ⎢am1 ⎥
⎢ ⎥ = X1 (ω) 0 ⎢ ⎥, (12.10)
⎣a2 ⎦ 0 X2 (ω) ⎣bm2 ⎦
b2 am2
where the suffix m denotes a measured wave. A numerical suffix denotes a calibrated
wave. All quantities are complex numbers (∈ C). The one-port correction matrices X1 (ω)
and X2 (ω) are defined as
% &
l1,2 −h1,2
X1,2 = = k1,2 Y1,2 (12.11)
k1,2 −m1,2
at each frequency separately. The explicit dependence on the frequency is removed from
this expression to reduce notational burden. These matrices will be used extensively in
the following chapters. We can now rewrite (12.10) as follows:
⎡ ⎤ ⎡ ⎤
a1 " # bm1
⎢b1 ⎥ 0 ⎢am1 ⎥
⎢ ⎥ = k1 Y1 ⎢ ⎥ (12.12)
⎣a2 ⎦ 0 k2
Y 2 (ω) ⎣bm2 ⎦ .
k1
b2 am2
The problem is that the outcome of any classical S-parameter calibration only deter-
mines the matrix without k1 . Hence, only seven of these eight error coefficients are
known. Since S-parameters are defined as ratios between waves, k1 (ω) can be freely
chosen in the S-parameter calibration. As long as a wave ratio is calculated, k1 (ω) does
not matter: it appears in the numerator and the denominator of the wave ratio and gets
factored out.
For the TOI as defined earlier, response measurements at ω0 and 3ω0 are to be com-
bined. The values k1 (ω0 ) and k1 (3ω0 ) can no longer be factored out and hence their
ratio appears in the result. The function k1 (ω0 ) therefore needs to be “measured” by an
additional calibration step: the so-called “absolute” calibration.
We will determine the complex function k1 (ω) in two successive steps: first the ampli-
tude of the function is obtained; next the phase function is characterized. Once this
complex function is known at all the test frequencies, the correction of the raw data boils
down to a matrix multiplication as is shown in (12.10).
4-channel ACQ
a1 b1 a2 b2
GEN1 GEN2
REF
Power meter
Fig. 12.11 Measurement for the power calibration.
Measuring |k1 (ω)|

We start from the wave correction relation (12.10). Since the correction equations
between port 1 and port 2 are decoupled (the off-diagonal blocks are zero), we can deter-
mine |k1 (ω)| by a one-port measurement as in Figure 12.11. The exact (but unknown)
incident power at the DUT can be determined from a1 and b1 . Thereto, consider the
incident power to the DUT at port 1
Pin = |a1 |2 − |b1 |2 = |l1 bm1 − h1 am1 |2 − |k1 bm1 − m1 am1 |2 . (12.13)
After substitution of (12.11), we can rewrite this equation as follows:

Pin = |k1 |2 |l1/k1 bm1 − h1/k1 am1 |2 − |bm1 − m1/k1 am1 |2 . (12.14)
In this expression, everything can be calculated besides |k1 (ω)| and Pin . However, if it
is possible to connect a power meter at the input port, |k1 (ω)| can be found since we
then measure the power Pm = Pin ,
Pm
|k1 |2 = .
| 1/k1 bm1 − 1/k1 am1 |2 − |bm1 − m1/k1 am1 |2
l h
Everything is measured (at the same frequency) in this equation besides |k1 (ω)|2 .
The magnitude of the calibration function is therefore known. Note that the procedure
outlined here is the same as the one that will be used for the real-time load pull system,
as will be shown in Chapter 14.
Measuring ∠k1 (ω)

Measuring the phase of k1 (ω) is the most complex part of the calibration, as here the phase
difference between spectral components at different frequencies ω must be calibrated.
Intuitively starting along the lines of a LTI reasoning, it might not be very clear why
there is a need for this phase spectrum calibration at all.
0.5
ain [a.u.]
−0.5
0 500 1000
t [a.u.]
Fig. 12.12 Two multisine signals in a.u. with the same power spectrum and a different phase spectrum. The
black line corresponds to zero phase, the gray one to a random phase spectrum.
To show the need for this alignment of spectral lines at different frequencies, we
consider a thought experiment based on a simple multiple-tone (multisine) signal. The
signal consists of 16 spectral lines, that all carry an equal power. For the first signal of
Figure 12.12, the phase of all spectral components is set to zero (full line). This signal is
pulse shaped and has an extremely high peak amplitude. The second signal has a phase
spectrum that is randomly selected for each line. The phase is drawn from a uniform
distribution ranging from [0, 2π[. The signal now looks very much as a random noise
signal, even though the power spectrum is still the same. It is clear that the first signal
will excite the nonlinearity in a totally different way than the second one.
What we see here, is that the phase spectrum of a multisine signal influences the shape
of the signal in the time domain. Therefore, it is very important to measure the phase
characteristic accurately when the properties of the nonlinearity are to be quantified.
Why is this information not measured by a normal VNA? A VNA measures one
frequency at a time and calculates only wave ratios. It can only obtain the phase difference
between sinusoidal waves that have the same frequency. Hence, there is no way to
measure the phase spectrum: the phase spectrum measures the phase difference between
spectral lines at different frequencies.
How are we going to calibrate the phase spectrum? The idea again is pretty simple to
understand. We will create a signal that is repeatable over a long period of time and is
very well known. This signal will then serve as a calibration element. It will be fed to
one of the ports of the NVNA and will be measured. The known difference between the
measured and the standard’s phase will then be used to correct the measurement.
One of the problems that we face is that the phase standard signal has to contain a
spectral line for each frequency that is involved in the measurement that we want to
calibrate. Since this requires a wideband signal, some kind of a pulse-shaped signal will
be used.
To make all this more practical, we will develop the idea for a sine wave excita-
tion. Following the same lines, the method applies to modulated excitation signals using
an appropriate calibration signal, as shown in Figure 12.13. When a sine wave excites
4-channel ACQ
a 1 b1 a 2 b2
Z0
GEN1 GEN2
REF
Reference signal
Fig. 12.13 Measuring the phase spectral standard.
a PISPO system, the output waves have the same period as the sine wave. A comb
generator fed with a sine wave will create a large number of harmonics of the sine
wave and therefore can act as a reference signal: it contains all the requested frequen-
cies. This comb generator is often called “the golden diode” in the context of nonlinear
characterization.
Again, we start from the error correction equation (12.10). The exact (but unknown)
reflected DUT wave b1 (lω0 )) now contains N harmonic components simultaneously.
The wave bm1 is then also measured at the same N harmonic components and we obtain
a set of N complex equations:
b1 (lω0 ) = k1 (lω0 ) bm1 (lω0 ) − m1 (lω0 ) am1 (lω0 ) for l = 1 . . . N. (12.15)
This can be rewritten to introduce the calibration coefficients as before:

7 8
m1 (lω0 )
b1 (lω0 ) = k1 (lω0 ) bm1 (lω0 ) − am1 (lω0 ) for l = 1 . . . N. (12.16)
k1 (lω0 )
We require that the phase spectrum of the multiple tone wave is equal to the (exactly
known) excitation of the standard signal in the reference plane of port 1 to calculate the
phase ∠k1 :

m1 (lω0 )
∠k1 (lω0 ) = ∠b1 (lω0 ) − ∠ bm1 (lω0 ) − am1 (lω0 ) for l = 1 . . . N (12.17)
k1 (lω0 )
This delivers the phase difference that is to be used to compensate for the dynamic
errors in the spectrum of the measured signals.
12.5 Translation to instrumentation
We are now ready to analyze the behavior of the existing instruments and instrumentation
setups. The perspective that we take here is to compare the setups on a purely technical
basis. We will compare the capabilities and indicate their positive and negative properties.
We will not include the software capability or the integration into existing simulation
packages.
We have chosen to present the setups in chronological order.
12.5.1 Oscilloscope-based receiver setups

The instrumentation setup of the oscilloscope-based NVNA is given in Figure 12.14.
The oscilloscope can either be a real-time [11] or a sampling oscilloscope [12]. Two
generators are needed in this setup if we want to be able to characterize a two-port
dynamic PISPO system for a particular class of excitation signals. It is no longer allowed
to switch off the input source when a reverse excitation is to be applied, as this modifies
the input signal and/or impedance and hence the PISPO approximation of the DUT can
also change.
Note the presence of a common time reference (a reference clock). It is used to avoid
frequency slipping between the timebase of the scope and the period of the generated
signals.
The oscilloscope-based setup is a time-domain setup. As the waveforms are acquired in
discrete time directly in the time domain, all the spectral lines in the signal are measured
simultaneously and therefore they are automatically phase-aligned in the spectral domain.
To obtain distortion-free waveforms, it is therefore sufficient to correct for timebase
distortion and for the own (mainly linear) dynamics of the channel.
For a real-time oscilloscope, the samples are acquired at a rate that is equal to the sam-
pling rate of the scope. As this requires extremely fast ADCs, the sampling oscilloscope
has been proposed to reduce the conversion speed while maintaining the sampling rate
for periodic signals [12].
For a sampling oscilloscope, the discrete time signal consists of samples that are
acquired one at a time and each in a separate period of the signal. Sampling reduces
the actual sample conversion speed to a value that is up to more than a hundred times
slower than the actual sampling speed of the data record. The shape of the waveform is
left untouched (for an ideal setup) as shown in Figure 12.15. The gray line represents
the RF signal. In this example, the signal is sampled using a sample period that is a little
4-channel
scope
a1 b1 a2 b2
DUT
REF GEN1 GEN2
Fig. 12.14 The scope-based NVNA.

x (t)
Fig. 12.15 Sampling of a waveform
x (t)
Fig. 12.16 Influence of a random variation of the trigger point position on the measurement.
larger than the period of the underlying signal. The measured samples are represented
by the circles on the plot. The black line represents the (unknown) continuous time
representation of the sampled signal. Visual comparison of the black and the gray lines
shows the same signal, but represented on a stretched time axis. Sampling did indeed
convert the signal to a lower frequency without distortion of the shape of the wave!
The AD conversion speed for sampling scopes becomes lower and therefore the con-
version can be made more accurate. However, there is no “free lunch” in instrumentation!
The key hypothesis needed to obtain a correct measurement is that the signal has to remain
stationary during the acquisition of the complete signal (all samples in the trace). Time
variations in the behavior of the system must therefore be slow. Theoretically we will
assume a quasi-static behavior to reduce the errors to zero. Practically, the non-stationary
variation of the DUT can only become visible on a timescale that is an order of magni-
tude slower than the acquisition time of a wave. For fast variations of the temperature
for example, this can sometimes be a limitation.
To obtain good measurements, the stability of the position of the trigger point used
by the scope to determine the periodicity of the signal becomes extremely important. To
illustrate this, a thought experiment is again welcome. Starting from the ideally sampled
signal of Figure 12.16, we introduce a small (with respect to the sampling period) random
variation of the trigger point during the sampling process. This is what we call timing
jitter; it introduces an uncertainty in the sampling instant results, and hence the measured
signal varies too. The magnitude of the signal variation is indicated by the error bars on
Figure 12.16. The sensitivity of the measurement to timing errors clearly increases with
the derivative of the signal.
At first sight, this high and signal-dependent sensitivity seems to be a disadvantage.
However, assume now that we want to minimize the measurement errors induced by
the trigger uncertainty. The trigger acts as a window comparator whose timing precision
increases with the signal derivative. The most stable trigger will therefore always be
obtained when the trigger level is set to correspond to the portion of the signal with a
maximal rate of change.
For modulated signals, the signal shape in the time domain is so complex that it is
impossible to trigger the sampling scope directly from the wave signal. As a consequence,
a frame synchronization signal is to be generated by the modulated source to define the
trigger point. It is preferable to select a block-pulse-shaped signal (rather than a sine
wave, for example) to obtain a maximal slew rate of the signal close to the trigger point
and hence minimize the jitter for the trigger.
Calibration issues
In the following we discuss why nonlinear measurements require additional power and
phase calibration to remove the residual errors that are left untouched by the S-parameter
calibration. However, the calibration of an oscilloscope-based setup is a little bit more
tricky: besides the LTI errors that are induced by the bandwidth limitations of the setup,
the scope also requires a calibration to compensate for errors in the time grid of the
acquired samples: the timebase correction [14].
The timebase correction is the most complex problem in the calibration of a scope.
Often, the timebase suffers from nonlinear distortion (as a function of the sampling
instants) or even from the presence of jumps. As a consequence, the data points in
the acquired discrete time signal are no longer perfectly equidistant in time. This
introduces a kind of phase modulation. Even if the waveforms to be measured are per-
fectly periodic at the expected frequencies, significant leakage errors are created in the
DFT spectra. This timebase distortion comes in two types: a systematic distortion that
is independent of the measurement realization and a stochastic component, the time
jitter.
The fast and dirty approach to compensate for timebase errors is to use a non-
rectangular time window to get rid of the leakage. This introduces a significant
measurement error. The clean way to circumvent the problem is to calibrate the timebase
of the scope.
Two classes of calibration method exist in the general literature. The first class per-
forms prior characterization of the timebase errors, measuring a sine wave signal. It
compensates for the systematic timebase error, leaving the jitter untouched [15], [14].
Maximum-likelihood methods that operate completely in the time domain allow for a
distorted input waveform and are available in this context [16]. They do not impose
requirements on the shape of the timebase distortion.
The second class of methods performs the correction during the measurement of the
signals [17]. These require a clean sine source and two additional acquisition channels
to be present in the setup. A sine wave and a cosine wave (obtained by a 90o hybrid) of
appropriate frequency are then fed to the additional channels while the device waves are
measured. A simultaneous acquisition of the sine wave, the cosine wave, and the device
waves then allows for compensation of the systematic timebase distortion and the part
of the jitter that is common to all the acquisition channels.
Conclusion
The oscilloscope-based NVNA is the intuitive solution to the measurement of a PISPO
DUT. It captures the complete time signal directly in the time domain, thereby reducing
the number of processing steps needed to obtain the data. Ideally speaking, it should be
sufficient to take the DFT of the measured samples to obtain the raw data. In practice, the
timebase errors introduce an additional complication in the processing. Their removal
using a timebase calibration is certainly possible, but is rather complicated and pretty
involved.
12.5.2 Sampler-based receiver setups

The main disadvantage of the oscilloscope-based devices lies in the presence of the
timebase errors. In order to avoid these, it is tempting to replace the timebase by
something new. The problem of the timebase is twofold: the stability of the trigger
point (equivalent time) and the lack of coherence between the signal source and the
timebase jeopardize the measurement quality.
We will replace the timebase by a sampling clock generator to impose an equally
spaced sampling grid that is phase coherent with the generator waves by construction.
Good measurement engineering practice teaches that measurement speed and accuracy
seldom come together. The sampling or reduction of the AD conversion speed therefore
is an attractive measurement option. Ideally, this reduces the leakage problems in the
discrete time signal to zero and increases the measurement resolution [18–21].
To find out how to get a timebase replacement, it is important to understand the source
of the potential problems. Sampling scopes rely on the detection of the position of the
trigger point to generate their sampling grid. The position of the trigger is very noise
sensitive because it relies on a point based decision. The trigger results from a comparison
between two signals at a single instant in time. There is no possibility of averaging the
noise in that type of circuitry. What we would like to obtain is that the position of the
trigger point would be determined using the information present in a complete period of
the signal to be acquired.
This calls for a PLL type of solution. The noise can then be reduced drastically by
the narrow bandwidth of the PLL. The use of a phase locked loop to synchronize an
acquisition to the generated signal was already used in the early days of network analysis.
It is the idea behind the vector voltmeter. The phase accuracy that is obtained from the
PLL in the sampling process results in time accuracy levels for equivalent time sampling
that are better than for the real-time scope.
A/D
A/D
A/D
A/D
Fs
Fs
FracN
DUT
REF GEN1 GEN2
Fig. 12.17 The sampler-based setup.
What we aim for here is different. We also need to obtain a very flexible setting of the
sampling clock to accurately define the frequency ratio between the acquisition sampling
clock and the generator frequency grid. Hence, we would rather go for a fractional N
synthesizer, that contains a PLL, to realize the high-resolution clock frequencies that are
needed for the measurements.
The instrument that is obtained after the transformation of the timebase is shown in
Figure 12.17. Note that the clock frequency that is generated by the FracN synthesizer is
much lower than the bandwidth of the RF signal. The spectrum of the discrete time signal
measured by the DFT after the ADC will be periodically repeating over Fs . The measured
discrete time signal is real-valued, and this calls for an auto-conjugate spectrum. All the
spectral lines that are present in the RF signal will therefore fold down in the Nyquist
band (frequency between 0 and Fs/2). As long as the different spectral lines do not overlap,
we will therefore be able to reconstruct the original signal.
Aliasing: a narrow-band solution to broadband measurements

To make this folding process more practical and to develop the intuition of the reader who
is not so familiar with aliasing and playing with sampling frequencies, we will consider
a series of example measurements that show the capabilities of the sampling converter
measurement. We will increase the complexity of the experiments gradually to show the
potential of the method.
In the first example, we show the very basics of aliased measurements using a sine
wave excitation for a PISPO system. Although this case might seem trivial, it enables
us to show the properties of the aliased measurements. We can start from the sampling
scope waveforms obtained above and look for an equivalent measurement obtained by
the sampling converter.
The second example extends the sine wave result to the measurement of the response
of a PISPO system to a modulated excitation signal. We use a narrow-band modulated
signal to show the differences and the similarities between the original RF signal and its
IF converted replica.
Finally, we briefly show that the conversion of a wideband multiple tone signal is also
possible. The additional complication encountered here stems from the large difference
in shape between the time-domain signal before and after the conversion. We show that
the spectral lines of the RF signal can still be recovered, albeit that their frequencies will
be completely scrambled. Frequency engineering allows us to unscramble the measured
spectrum and to restore the original waveform.
The frequency engineering behind aliasing

Before we look at an example, it is important to take a look at the generation of the
clock frequencies that are used in the setup. Assume that the reference clock (REF in
Figure 12.17) is a Fref = 10 MHz clock. This is a common value for the reference
frequency that is used by an RF instrument. The fractional-N synthesizer (labeled FracN
in Figure 12.17) transforms this reference frequency into the sample frequency Fs
p
Fs = Fref p, q ∈ N. (12.18)
q
The ratio p/q is designed such that its maximum value is modest (for example 2) but its
resolution is very high (for example better than 1 Hz). Both the samplers and the ADCs
of the instrument run from the same sampling frequency Fs . While this is not mandatory,
it allows us to obtain a much cleaner explanation.
In a measurement, the samplers and the ADCs acquire Nacq successive data samples.
The time span over which data are acquired is therefore Navq/Fs = Nacq Ts . After taking
a DFT, this results in a folded spectrum that has a spectral resolution that is equal to
Facq = Fs/Nacq .
The easy case: down sampling the PISPO response for a CW excitation
Now, we consider a CW measurement of a PISPO system to understand the basic opera-
tion of the frequency aliasing method. The GEN1 source in Figure 12.17 generates a sine
wave of frequency Fsine . After passing the DUT, the output wave contains a number of
harmonic components of that excitation. Assume for simplicity that harmonics are only
present at 2Fsine and 3Fsine . The spectral mask of the input and output waves is shown in
Figure 12.18. We select the sampling frequency Fs such that the aliased wave measured
by the ADC is a time-dilated version of the RF wave. The time dilatation g results in a
S (f )
Fif 2Fif 3Fif
f
Fsine 2kFsine 3kFsine
kFs 2kFs 3kFs
Fig. 12.18 Spectra of the DUT operating under CW excitation. The input sine wave frequency is Fsine . The
conversion frequency and the IF frequency are shown for the different harmonics.
waveform that varies more slowly: asine (t) = aI F (gt), as was the case in Figure 12.15.
The relation between the spectral lines in the RF wave (fundamental frequency Fsine )
and the folded waveform (fundamental frequency Fif in Figure 12.18) can be expressed
in the frequency domain. It can be reduced to:
⎧
⎪
⎪ = S(Fsine )
⎨S(FI F )
S(2FI F ) = S(2Fsine ) . (12.19)
⎪
⎪
⎩S(3F ) = S(3F )
IF sine
To obtain this, we select Fs and k such that Fsine = kFs + FI F . The fundamental IF
frequency is chosen low enough to ensure that all the harmonics are converted to a
frequency below the Nyquist frequency. Here, we choose FI F < F6s . The spectral line at
the fundamental frequency Fsine appears after folding in the discrete time spectrum at
a frequency of FI F . Since we convert the mth harmonic of FI F with the same sampling
frequency Fs , the harmonic response appears at a frequency of mFI F , as requested.
The measured response waveforms obtained by the sampling converter and the real-time
oscilloscope are therefore identical up to a time-stretching.
To obtain measured waveforms that are free of spectral leakage, we must ensure that we
always measure a complete period of the IF waveforms. This requires that the frequency
of the folded (IF) lines lies on an integer multiple of the spectral resolution of the DFT:
Fs
FI F = l l ∈ N. (12.20)
N
Conclusion: The measurement of the response of a PISPO system to a sine wave excita-
tion is as easy with a sampling converter based instrument as it is with a real-time
oscilloscope. Both instruments will (ideally) measure the same waveform, without
any distortion up to a linear scaling of the time axis. The time-domain shape of
the RF and IF waveforms match perfectly. The spectra will also be equal up to the
frequency compression factor.
Converting a narrow-band modulated signal

Now, let us complicate the situation a little more by considering a narrow-band modulated
excitation signal. The generator GEN1 generates a multisine signal that consists of M
spectral lines that are spaced Fms apart. The complete modulated signal has a modest
bandwidth with respect to Fs . Mathematically, we express this as MFms Fs/2. The
first line of the multisine is assumed to have a frequency of Fms . To avoid unnecessary
notational clutter, and without loss of generality, we assume that Fms = kstart Fms
where again kstart ∈ N. We assume that the output wave of the PISPO system under
test is as shown in Figure 12.19. This spectrum contains out-of-band spectral distortion
components around 2Fms and 3Fms and small spectral regrowth contributions (small
white rectangles around Fms ) only. All the spectral lines lie on the same spectral grid
that is spaced Fms apart.
The sampling process again compresses all the spectral lines that are present in the RF
spectrum in an IF band ranging from F = 0 to F = Fs /2 using the relation F = kFs +Fif .
S (f ) Fs Fs Fs
2 2 2
f
Fms 2kFms 3kFms
kFs 2kFs 3kFs
S (f ) F
s
2
f
Fms
Fig. 12.19 Spectral mask of the modulated output signal. The modulated input waveform is drawn in
dark gray.
The narrow bandwidth of the modulation spectrum assures that both the modulation and
the spectral regrowth around the fundamental frequency Fms are all aliased using the
same value of k. Hence, the modulated lines around each carrier appear in the same order
and with the same spacing in the IF and in the RF spectrum. The same is true for the
modulation around the harmonics at 2Fms and 3Fms . This folding process is illustrated
in the frequency domain in Figure 12.19.
The IF spectrum obtained by the sampling converter is no longer exactly equal to the
spectrum of the real-time scope up to a compression of the frequency axis. The frequency
grid of the modulation Fms is the same in the IF and the RF spectra. However, the
frequency spacing between the fundamental and its harmonics has been compressed.
This means that the time waveforms of the IF and the RF signal are no longer the same;
but, the waveforms are not totally different either. One can show that these waveforms
share the same envelope. Again, we obtain a leakage-free measurement whenever we
measure an integer number of periods of the IF wave.
Conclusion: The measurement of the response of a PISPO system to a narrow-band

modulated excitation with distortion no longer yields exactly the same waveform
for a sampling converter and a real-time oscilloscope. The instruments will (ideally)
measure different signals that share the same envelope, without any distortion. The
measured spectra will (ideally) be exactly equal around each carrier, but the spacing
in between the carriers will be compressed. One can exactly reconstruct the RF
signal based on the measurements of the sampling converter if the spacing between
the carriers is restored to its original value.
Converting a wideband multiple tone signal with harmonics

In the third example, we consider the case of a wideband modulation [22]. The bandwidth
of the modulation signal now exceeds the IF bandwidth of the ADC. In this case, things
are more involved. It becomes impossible to obtain a spectrum for the IF wave that is
identical or even similar to the RF spectrum as before. All the spectral lines that are
present in the RF spectrum are still measured, but they no longer appear in the same
order in the IF and the RF spectra. We can illustrate this on a simple example, again to
avoid notational clutter.
Consider a multisine signal that consists of three lines that are spaced more than the
sampling frequency Fs apart. This signal excites a PISPO system that produces a second
and a third harmonic response only.
For the sake of illustration, we assume that the lines are spaced Fms = Fs + δms
apart. The excited lines are labeled 1, 2, and 3 in Figure 12.20. The smaller diamond
labeled lines in the figure represent the spectral regrowth of the modulated signal around
the fundamental frequency. A careful look at Figure 12.20 shows that the frequency
difference between the diamond labeled lines and a multiple of the sampling frequency
Fs differs from line to line. After folding, all the diamond labeled lines will therefore
appear at a different IF frequency.
The spectral lines that lie around 2Fs are shown in Figure 12.21. The lines that lie
around the fundamental are superimposed on this plot artificially to show the interdepen-
dence between the IF frequencies. The overlay is constructed such that if a fundamental
line is located at ff und (k) = kFs + δk , this line is overlaid with f2Fs (k) = 2kFs + γk .
Again, a closer look at Figure 12.21 shows that the spectral lines that are included in
the second harmonic response will fold to other frequencies than the lines around the
fundamental frequency.
The result of the complete frequency folding is illustrated in Figure 12.22. Note that all
the diamond labeled lines appear twice in the measured spectrum: once at a frequency f
and once at a frequency Fs −f . For the excited lines this was made visible by the labels 1,
2, and 3 and mirrored 1, 2, and 3. The second contribution is always the complex conjugate
S (f )
Fs Fs Fs Fs Fs Fs Fs
kFs
f
4 1 2 3
Fig. 12.20 The response measured around the fundamental frequency Fms . Excited lines are labeled 1, 2, 3.
S (f )
Fs Fs Fs Fs Fs Fs Fs
kFs
f
1 2 3
Fig. 12.21 The response measured around twice the fundamental frequency 2Fms . The fundamental lines
are superimposed on this spectrum, such that kFs overlaps 2kFs . This shows the difference in
frequencies between the fundamental and second harmonic response.
S (f )
Fs
S*
3 1 2 f
1 2 3
4 4
Fig. 12.22 The IF spectrum of the wideband multisine. Lines labeled with a diamond lie close to the
fundamental Fms , triangle labeled lines lie close to 2Fms and dot labeled lines lie around 3Fms .
Dashed lines represent the complex conjugate of corresponding full lines.
of the first one. This additional mirroring symmetry around the Nyquist frequency Fs/2
is imposed by the real-valued character of the time signals.
There is one additional important detail that is visible in Figure 12.22. Some spectral
lines, such as line 4, initially fold to a frequency that is larger than the Nyquist frequency
Fs/2. Of course, these lines also appear mirrored and complex-conjugated in the lower
half of the IF band (line with mirrored 4). If the spectrum is to be measured without
errors, this complex conjugate has to be accounted for in the reconstructed measured
spectrum.
Hence the measurement still contains all the spectral lines contained in the original
signal. A proper choice of the sampling frequency can still impose that the spectral lines
that have different RF frequencies end up at different IF frequencies too. Of course, this
does not come free. The price that one has to pay is that the spectral resolution that is
required in the IF domain is increased. Longer measurement records are therefore to
be acquired and processed. In the example considered here, we have to provide room
for three more lines in between the lines of the fundamental tones, which increases the
resolution by a factor of 4.
Finally, we can now show a blueprint of the IF spectrum that is measured. It is clear that
a significant amount of housekeeping is needed to reconstruct the waveforms properly.
The IF lines have to be repositioned to their original position in the original RF grid
(taking complex conjugates into account) if one is to obtain a distortion-free time domain
waveform.
Conclusion: The measurement of the response of a PISPO system to a broadband mod-
ulated excitation with distortion does not directly yield the same spectrum for a
sampling converter and a real-time oscilloscope. The spectra measured by the two
instruments still contain the same spectral lines if a proper sampling frequency
is selected to avoid overlap of the RF lines in the IF domain. The time-domain
waveforms obtained by a direct transformation of the IF and RF spectra to the time
domain will be totally different however. The original RF time waveform can still
be (perfectly) reconstructed if each spectral line is carefully replaced (watch out for
the complex conjugates) to the frequency it originated from before the conversion
to the time domain is performed.
Calibration issues
Once the measurements are taken, the calibration of the sampling converter is pretty easy
to realize. All the calibration steps can be performed before the start of the measurements,
and only the steps that were explained in the general calibration of the NVNA have to
be taken.
Conclusion
The sampling converter has the advantage that it is freed from the timebase problems and
the consecutive spectral leakage problems that can be present in the oscilloscope-based
receiver. However, this comes at a price. For sine wave measurements and modulated
measurements of low bandwidth (with respect to the sampling frequency) the sampling
frequency can be selected such that the sampled waveforms are time-stretched copies of
the RF waveforms. When a wideband modulation is applied, the measurement remains
possible, but comes at the cost of increased housekeeping to determine the origin of the
measured spectral lines.
12.5.3 VNA-based setups

Up to now, the setups have all been measuring directly in the time domain or the equivalent
time domain. Put in a nutshell, this means that all the spectral lines that are present in
the signal are acquired simultaneously. The main advantage of this way of working lies
in the fact that the different lines are synchronized by construction of the measurement.
This situation changes completely whenever a linear VNA is used to acquire the data
[23, 24]. A VNA can be conceptualized as a frequency-domain acquisition device. A
modern VNA measures the four waves of a two-port device simultaneously, but does so
for one frequency at a time. The measurements that are taken at different frequencies
are not synchronized. Put in a different way, one can state that there is no common
time reference for the measurements that are taken at the different frequencies. Formally
speaking, this means that the measurements contain an additional phase indeterminacy,
⎡ ⎤ ⎡ ⎤
A1 (ωk ) A1 (ω1 )
⎢B1 (ωk ) ⎥ ⎢B1 (ω1 )⎥
⎢ ⎥ He (ωk , ω1 ) ⎢ ⎥
⎣A2 (ωk )⎦ = e
j ϕ(ωk )
⎣A2 (ω1 )⎦ . (12.21)
B2 (ωk ) B2 (ω1 )
In this expression, ϕ(ωk ) is an unknown phase-shift that varies in a random way from
measurement to measurement and hence from frequency to frequency. The complex
matrix He (ωk , ω1 ) ∈ C4×4 describes the exact relationship between the waves. This
means that this is the relationship that would be obtained if the measurement were
performed by an ideal real-time scope. The suffix one in ω1 labels one test frequency
that serves as a reference. This reference can be freely chosen. It is clear that the phase-
shift ϕk is unimportant for S-parameter measurements. Whenever a ratio of waves is
taken at one single frequency, this term disappears.
Whenever the measured quantity contains spectral rays at different frequencies, this
phase must be known. As an example, consider again the amplifier as above. To determine
the third-order intermodulation contribution at the third harmonic one calculates the
influence of the third-order term:
ej ϕ(3ω0 ) B2 (3ω0 ) = ej ϕ(ω0 ) H A21 (ω0 )A∗1 (ω0 ), (12.22)
with H the complex gain associated with the Volterra kernel. The phase of the inter-
modulation product is scrambled by a different unknown phase term in the right
and the left hand sides of the equation. This results in a phase indeterminacy of the
intermodulation components.
How can we get around this? If we could have a trigger signal that is common to all
the single frequency experiments, the problem would be solved. To see this, assume that
the trigger signal consists of a perfect Dirac impulse that indicates the start of the period
of the generated wave. Assume also that the internal sources of the VNA are perfectly
synchronized to that Dirac impulse. A VNA measurement is then only started when the
signal arrives at exactly the same point in the modulated waveform. As a result, ϕ(ωk ) = 0
for each test frequency and the measured spectra coincide perfectly with the exact ones.
Now we can translate this idea into a real device: the synchronizer. This device gener-
ates a periodic impulse train out of a periodic signal. As far as the system is concerned,
there are a number of alternatives that could be tempting when designing a synchronizer.
The first and most obvious choice is to use a comb generator. This idea was first suc-
cessfully implemented in [23]. When a step recovery diode [25] is fed by a sine wave, it
will ideally generate a large impulse at a fixed position along the sine wave. When tuned
and packaged properly, this generates tens to hundreds of harmonics over a bandwidth
that is wide enough to cover current instrumentation needs. A second possibility is to use
a nonlinear transmission line. Proper design allows the steepening of one of the edges
Rec D
Rec C
Rec E
Rec A
Rec B
Gensweep
DUT
REF GEN1 GEN2
SYNC
Fig. 12.23 The VNA-based setup.
(positive or negative going edge) of the applied signal to extremely high slew rates.
The advantage of this approach is that these nonlinear transmission lines can easily be
matched. This enables us to use input waveforms that vary over a bandwidth that is much
larger than the input bandwidth of the comb generator. A third possible alternative is to
use an extremely wideband pulse-shaping amplifier that conceptually acts as an ideal
comparator for the applied input signal. From a user perspective, the actual implemen-
tation does not matter, as the only demand is that a known spectral line is present at each
frequency of interest.
The synchronizer needs to be fed by a reference signal that is somehow related to
the period of the input signal generated by the source. If the NVNA uses the built-in
sources of the VNA, this is quite easy to realize. One of the sine wave sources is then split
and fed in parallel to the synchronizer and the DUT, as shown in Figure 12.23. When a
modulated signal is used as an excitation, the synchronization is to happen on the period
of the modulated signal.
Calibration issues
The phase calibration of the VNA-based setup follows the general case as is explained
in Section 12.4.3. It requires a second synchronizer besides the “measurement synchro-
nizer” that is shown to be needed for the measurement itself. The second signal is needed
because the measurement synchronizer is connected to an additional receiver path. This
path is not calibrated during the calibration of the VNA. The measurement of the addi-
tional reference signal enables the compensation for all the phase errors of that additional
signal path.
Conclusion
The VNA-based setup has the advantage that it can rely on a very large installed base of
network analyzers.As theVNA is the most popular RF measurement device of the last half
century, it also can count on an RF community that is well skilled and feels comfortable
using it. However, the VNA-based setup comes at a cost: it requires the presence of an
additional receiver and a synchronizer device. This device generates a reference signal,
Fs
IQADC
IQADC
IQADC
IQADC
IQADC
LOacq
REF
DUT
IQAWG LOgen SYNC
Fig. 12.24 The IQ modulation-demodulation based setup.
hence it is an active device and it needs regular re-calibration to keep the accuracy of the
measurement. The disadvantage is that a malfunctioning of the synchronizer is hard to
notice in normal operation and therefore it remains a potential source of measurement
errors.
12.5.4 IQ-modulator based setups

The last setup in the list mimics the architecture of a receiver/transmitter pair to realize
a measurement device [26, 27]. The structure of the setup is shown in Figure 12.24. The
main idea behind this setup is to measure modulated signals with a moderate modulation
bandwidth. The generator consists of an IQ-modulated source. The IQ-modulated signal
can be expressed as a complex time signal as
a1c (t) = m(t) ej ωc t , (12.23)
where the modulation envelope of the signal is called m(t), while the carrier of the IQ
signal is labeled cgen (t).The signal that is fed to the DUT is then
a1 (t) = R (a1c (t)) + I (a1c (t)) . (12.24)
This signal passes through the DUT. Its output b2 (t) is then measured and is demodu-
lated by the signal cacq (t). Ideally, the carrier signals at the generation and the acquisition
are equal cacq (t) = cgen (t). For some generators, the presence of a static phase error can-
not be avoided in practice and cacq (t) = cgen (t) ej ϕacq . This is assumed in the remainder
of this section.
To make things more practical, we again use an example to make the argumentation
more concrete. We use the same example as for the VNA setup. We measure the third-
harmonic response of the simple PISPO system defined before, and obtain:
ej ϕacq (3ωgen ) B2 (3ω0 ) =j ϕacq (ωgen ) H A21 (ω0 ) A∗1 (ω0 ), (12.25)
for any spectral line B2 (3ω0 ) located in the third-harmonic frequency band and A1 (ω0 )
located in the fundamental band. Any time a new demodulation frequency is selected,
the phase difference ϕacq can change randomly. This is due to the hybrid nature of the
IQ instrument: time and frequency domain approaches are indeed combined in a single
setup. When the IQ-modulation frequency of the generator or the acquisition is changed,
a random phase error can appear in a way that is identical to the VNA setup, unless the
generator and the acquisition are made phase-coherent by construction. All the spectral
rays in the demodulated envelope are automatically synchronized because the acquisition
behaves as a real-time scope for the complete envelope signal.
Phase calibration considerations

This kind of measurement setup simplifies the phase calibration of modulated measure-
ments quite a bit. The phase calibration of a modulated wideband signal is hampered by
signal-to-noise ratio issues: each line that can potentially be measured has to be ener-
gized on a frequency grid that has a line density that is set by the frequency resolution of
the modulated signal. A huge number of spectral lines that will never be used have to be
present in the reference signal anyway to maintain general applicability. In the case of the
IQ modulator, this is no longer the case. The calibration of the envelope measurements
requires a timebase correction and a calibration of the acquisition in the IF domain. The
calibration between the harmonics calls for a comb generator that only needs to carry
energy at the harmonics of the fundamental IQ-modulation frequency. Of course, like
the VNA, the IQ modulator requires an additional measurement channel to measure this
reference signal simultaneously with the other acquired waves.
Conclusion
The IQ-modulator setup attempts to combine the advantages of the frequency and the
time domain approaches. It avoids the timebase problems and the triggering problems
of the RF oscilloscopes, but the ability to measure the complex envelope in the time
domain makes the measurements quite easy to set up. The analysis bandwidth of the
envelope is narrow enough to allow direct conversion to a discrete time signal in the
envelope domain and thus this device avoids the complications of the sampling receivers
with spectral folding. It also avoids the frequency by frequency sweep inherent in the
frequency-domain methods.
To measure the response between the fundamental and harmonic bands, the IQ-
modulator uses a frequency-domain approach of sweeping the frequency and therefore
also needs a synchronizer if the phase relation of the carrier and its harmonics is to be
measured. This increases the complexity of the setup and requires an additional receiver,
exactly as with the VNA setup.
12.6 Conclusion, problems, and future perspectives
We arrive now at the end of this bird’s eye view of the measurement of the nonlinear
PISPO systems operating at microwave frequencies. Of course, there is much more to
tell and there are many more issues and problems. They are discussed extensively in the
literature, in a stream of new and exciting papers. Our goal here is neither to be complete
nor encyclopedic. Triggering the curiosity of the reader and providing a glimpse into the
world of nonlinear characterization using the NVNA seems already very ambitious.
Where to go from here? It is our belief that we are now at a point where the basic
measurement capability that allows us to characterize the most commonly used systems
is indeed present.
What are the next stepping stones and where will they lead to? Of course, prediction
of the future is hard and a matter of opinion. Our belief is that the NVNA will become
really useful in the next decade, as it slowly finds its way into mainstream RF design. An
NVNA-like device will allow us to port and extend the available nonlinear design theory
into practice. It will allow us to close the design loop from a nonlinear point of view.
However, there are a number of very challenging issues that remain to be solved. There
is a lot more work needed to demonstrate the usefulness of the nonlinear information in
a design world that is still mainly dominated by S-parameters.
References
[1] S. F. Adam, “Microwave Theory and Applications.” Englewood Cliffs, NJ: Prentice-Hall,
1969.
[2] D. Rytting, “Network analyzers from small signal to large signal measurements,” PROC.
67th ARFTG Conference, 2006 , pp. 11–49, 16-16, June 2006.
[3] Y. Rolain, W. Van Moer, G. Vandersteen and J. Schoukens, “Why are nonlinear microwave
systems measurements so involved?” IEEE Trans. Instrum. Meas., vol. 53, no. 3, pp. 726–729,
June 2004.
[4] P. Wambacq, “Symbolic analysis of large and weakly nonlinear analog integrated circuits,”
Ph.D. dissertation, K.U. Leuven, Belgium, Jan. 1996.
[5] M. Schetzen, “The Volterra and Wiener Theories of Nonlinear Systems.” New York: Wiley,
1980.
[6] S. A. Maas, “Nonlinear Microwave Circuits”. New York: Wiley, 1988.
[7] W. Van Moer and Y. Rolain, “Measuring the sensitivity of microwave components to bias
variations,” IEEE Trans. Instrum. Meas., vol. 53, no. 3, pp. 787–791, June 2004.
[8] A. Alghanim, J. Benedikt and P. J. Tasker, “Investigation of electrical base-band memory
effects in high-power 20W LDMOS transistors using IF passive load pull,” 3rd International
Conference on Information and Communication Technologies, vols 1–5, Syria, 2008.
[9] E. Oran Brigham, “The Fast Fourier Transform.” Englewood Cliffs, NJ: Prentice-Hall, 1974.
[10] J. Schoukens and R. Pintelon, “Identification of Linear Systems.” London, U.K.: Pergamon
Press, 1991.
[11] M. Sipila, K. Lehtinen and V. Porra, “High-frequency periodic time domain waveform mea-
surement system,” IEEE Trans. Microw. Theory Tech., vol. 36, no. 10, pp. 1397–1405, Oct.
1988.
[12] G. Kompa and F. van Raay, “Error-corrected large-signal wave-form measurement system
combining network analyser and sampling oscilloscope capabilities,” IEEE Trans. Microw.
Theory Tech., vol. 38, no. 4, pp. 358–365, Apr. 1990.
[13] A. Ferrero and U. Pisani, “Two-port network analyzer calibration using an unknown ‘thru’”,
IEEE Microwave Guided Waves Lett., vol. MGWL-2, pp. 505–507, Dec. 1992.
[14] J. Verspecht, “Broadband sampling oscilloscope characterization with the nose-to-nose cali-
bration procedure: A theoretical analysis,” in Proc. 1994 IEEE Instrumentation Measurement
Technology Conf., Hamamatsu, Japan, May 1994, pp. 526–529.
[15] J. Verspecht and K. Rush, “Individual characterization of broadband sampling oscilloscopes
with the nose-to-nose calibration procedure,” IEEE Trans. Microw. Theory Tech., vol. 43, no.
2, pp. 354–374, Apr. 1994
[16] G. Vandersteen and R. Pintelon, “Maximum likelihood estimator for jitter noise models for
HF sampling scopes,” IEEE Transactions on Instrumentation and Measurement, vol. 49,
no. 6, pp. 1282–1284, Dec. 2000.
[17] P. D. Hale, C. M. Wang, D. F. Williams, K. A. Remley and J.D. Wepman, “Compensation
of random and systematic timing errors in sampling oscilloscopes,” IEEE Transactions on
Instrum. Meas., vol. 55, no. 6, pp. 2146–2154, Dec. 2006.
[18] W. Van Moer, Y. Rolain and A. Geens, “Measurement based nonlinear modeling of spectral
regrowth,” IEEE Trans. Instrum. Meas., vol. 50, no. 6, pp. 1711–1716, Dec. 2001.
[19] W. Van Moer, Y. Rolain and J. Schoukens, “An automatic harmonic selection scheme for
measurements and calibration with the nonlinear vectorial network analyser,” IEEE Trans.
Instrum. Meas., vol. 51, no. 2, pp. 337–341, Apr. 2002.
[20] W. Van Moer and Y. Rolain, “Proving the usefulness of a 3-port nonlinear vectorial net-
work analyser through mixer measurements,” IEEE Trans. Instrum. Meas., vol. 52, no. 6,
pp. 1834–1837, Dec. 2003.
[21] K. M. Youngseo, P. Roblin, Myoung Sukkeun, J. Strahler, F. De Groote, and J. P. Teyssier,
“Multi-harmonic broadband measurements using a large signal network analyzer,” PROC.
75th ARFTG, pp. 1–6, 28-28, May 2010.
[22] M. El Yaagoubi, G. Neveux, D. Barataud, T. Reveyrand, J.-M. Nebus, F. Verbeyst, F. Gizard,
and J. Puech, “Time domain calibrated measurements of wideband multisines using a large-
signal network analyzer,” IEEE Trans. Microw. Theory Tech., vol. 56, no. 5, pp. 1180–1192,
May 2008.
[23] U. Lott, “Measurement of magnitude and phase of harmonics generated in nonlinear
microwave two-ports,” IEEE Trans. Microw. Theory Tech., vol. 37, no. 10, pp. 1506–1511,
Oct. 1989.
[24] P. Blockley, D. Gunyan and J. B. Scott, “Mixer-based, vector-corrected, vector signal/network
analyzer offering 300 kHz-20 GHz bandwidth and traceable phase response,” IEEE MTT-S
International Microwave Symposium Digest, pp. 4 pp., 12-17 June 2005.
[25] “Microwave Components 0.001-40 GHz,” Herotek, Datasheets [Online]. Available:
www.herotek.com.
[26] M. Thorsell and K. Andersson, “Fast multiharmonic active load–pull system with waveform
measurement capabilities,” IEEE Trans. Microw. Theory Tech., vol. 60, no. 1, pp. 149–157,
Jan. 2012.
[27] H. Jie, K. G. Gard, N. B. Carvalho and M. B. Steer, “Dynamic time-frequency waveforms
for VSA characterization of PA long-term memory effects,” ARFTG Conference, 2007 69th,
pp. 1–5, 8-8, June 2007.
[28] T. Van den Broeck and J. Verspecht, “Calibrated vectorial nonlinear network analyzer,” in
Proc. IEEE MTT-S, San-Diego, CA, 1994, pp. 1069–1072.
13 Load- and source-pull techniques
Valeria Teppati, Andrea Ferrero, and Gian Luigi Madonna
13.1 Introduction
Chapter 12 has explained how S-parameter measurements in small-signal conditions are

not adequate to characterize an active device for a relevant number of applications. In
large-signal conditions the performance of the active device under test (e.g. output power,
gain, and efficiency) does not only depend on the chosen bias point and the excitation
frequency, but also on the input power level and the loading conditions at the input and
output ports.
The “Rieke diagrams” were already used in the early 1940s to show how the per-
formances of microwave tubes and oscillators vary as a function of load impedance
[1]. The oscillating frequency and output power used to be plotted on the Smith chart
as contours at constant level. The modern term of “load-pull” comes from these early
applications, where a change of the load impedance would have “pulled” the output
oscillator frequency.
More precisely, the term “source and load-pull systems” refers nowadays to the set of
instrumentation (at microwaves and at lower frequencies) needed to address two main
issues:
• setting and monitoring the DUT loading conditions,

• measuring the DUT performance of interest.
Chapter 12 has already described in detail the most modern large-signal measurement
techniques to address the second point. However, it will be clear in this chapter that
the two problems are strictly correlated, as the solutions chosen for the former have a
considerable impact on the latter. In particular, considerations of cost and test duration,
complexity of the measurement setup, as well as the availability of equipment and trained
personnel might pose strong constraints on the load-pull techniques that can be used.
Eventually, the choice of the final solution needs to balance these aspects with the required
load-pull functionalities and accuracy.
Complete and automated source and load-pull systems appeared in the early 1970s
[2, 3]. They are today able to quickly characterize microwave devices under large-signal
excitation and they are used for a wide range of applications:
• matching network design, to achieve optimal performance from a device operating in

nonlinear conditions (typically, for amplifier or oscillator design);
346 Valeria Teppati, Andrea Ferrero, and Gian Luigi Madonna
Input Output
matching matching
network network
DUT ΓS DUT
aS ΓL
Source Load
(previous ΓS ΓL (next
stage) a1 b2 stage)
b1 a2 ΓIN ΓOUT
(a) (b)
Fig. 13.1 Definition of the quantities of interest. ©2001 IEEE. Reprinted, with permission, from [7].
• large-signal device modeling, where load-pull data is used directly in a behavioral

model, or indirectly extract and tune a device model;
• technology process development, where load-pull measurements are used to determine
the effects of varying process parameters on device large-signal performances;
• device reliability study and identification of failure conditions (e.g. ruggedness testing
of modules, where the device is measured with a specified output VSWR at all phases);
• quick device evaluation for production, where specific load and source impedances
are set for a single pass/fail test.
This list is not exhaustive, as source and load-pull can be in general applied to the
evaluation of any nonlinear device. Further examples can be found in literature for mixer
design [4], oscillator measurements [5], and diode characterization [6].
To explain the basic principle, let us consider the situation in Figure 13.1(a). It refers
to the typical example of an amplifier design, where a microwave transistor is connected
to the previous and following stages by two linear and tunable matching networks.
The basic design target is to find the proper load reflection coefficients L and S
to meet the required specifications. If the transistor operated in linear conditions, the
solution could be found with the only knowledge of its S-parameters. For instance, the
maximum output power would be achieved by designing the output matching network
so that L = OU ∗
T . If the transistor were considered unilateral (i.e. S12 = 0), then the
condition would simply be L = S22 ∗ .
In large-signal conditions, the transistor nonlinearities play a fundamental role and

the optimum loading conditions may be significantly different from the linear case [8],
[9]. Load-pull systems allow finding the proper load values experimentally.
In the simplest implementation, the active device is driven by the microwave source
at a single frequency and its performances are measured while physically changing L
(for load-pull) or S (for source-pull). The monitored quantities are typically:
• input and output power (PI N and POU T );

• operating and transducer gain (GOP and GT ), along with the corresponding
compression;
• PAE or drain efficiency;
• noise figure and noise parameters.
Load- and source-pull techniques 347
Moreover, by driving the device with two tones or properly modulated signals, inter-
modulation or ACPR measurements can be performed to investigate the linearity of the
amplifier.
In the simplest case, it is sufficient to control the source or load impedance on a
relatively small bandwidth around the excitation frequency. This is true for a relatively
large number of applications. However, if the active device is driven into strong nonlinear
conditions (e.g. as in the case of high-PAE amplifiers [10, 11]), the spectral content of
the output signal can be relevant at harmonic frequencies, too; the corresponding load
conditions can significantly affect the device performance [12,13], and harmonic source
and load-pull systems are therefore used to experimentally investigate these effects.
A first classification of the source/load-pull systems refers to the techniques used to
control the reflection coefficients S and L . Passive load-pull systems use mechanically
tunable elements (the so-called “tuners”), while active systems synthesize the desired
loads electronically. Reflection coefficients with magnitude up to unity can be reached
at the DUT ports by the active techniques, thus overcoming the limitations of the passive
tuners due to fixture and probe losses (for details on fixturing and probing issues see
Chapter 2). Passive tuners are described in Section 13.2, while active load-pull systems
are the subject of Sections 13.3 and 13.4. Fundamental and harmonic load tuning are
discussed, along with the advantages that combining passive and active tuners can have
for some specific applications.
A second classification of the load-pull techniques refers to the DUT measurement
principle. Non-real-time techniques are typically simpler; they are used only with pas-
sive systems and they are based on the tuner pre-characterization. Real-time systems
can exploit all types of loads – active, passive, or a combination of them – and their
measurement accuracy does not rely on the mechanical repeatability of the tuner posi-
tion. The peculiarities of real-time and non-real-time systems are described in detail in
Section 13.5.
Section 13.6 focuses on the most recent advances in the load-pull technology, which
combine harmonic load-pull, mixed-mode signals, and time domain waveform measure-
ment. These techniques require test setups that are usually not available off-the-shelf;
nevertheless, they are becoming increasingly important in the R&D labs for accurate
nonlinear device modeling and to assist in the design of advanced microwave nonlinear
circuits.
One final remark: in this chapter we generically refer to load-pull systems. Source-pull
measurements use, in principle, the same techniques to synthesize the source reflec-
tion coefficient; however, the presence of a generator term in the source equivalent
circuit poses additional challenges in the accurate measurement of the source reflection
coefficient S . This is the topic of Section 13.7.
13.2 Setting the load conditions: passive techniques
Passive techniques are based on mechanical, tunable devices (usually referred to as

tuners) to generate the required impedance at the ports of the DUT.
13.2.1 Basics
The most common tuner type is the “slide screw” tuner. As shown in Figure 13.2, it is
based on a slab line, consisting of two parallel ground planes with a center main line,
plus a reflective element (a conductive “probe”). When the tuner is used as the output
matching network of Figure 13.1(a), the load reflection coefficient L is controlled by
setting the position of the tuning probe along the longitudinal and vertical axes of the
slab line. When the probe is completely retracted, the line impedance – typically 50 –
is not perturbed. When the probe moves closer to the main line, an impedance step is
introduced, corresponding to the probe length. In particular, the presence of the probe
reduces the impedance of the corresponding slab-line portion.
The mismatch introduced by the probe peaks at the frequency where the probe length
corresponds to λ/4, where λ is the wavelength in the slab-line. In order to increase
the frequency range, commercially available tuners [14, 15] typically provide two or
three probes.
In a very first approximation, changing the probe position along the vertical direction
causes a change of the reflection coefficient magnitude, while the movement along the
longitudinal axis changes the reflection coefficient phase. The reflection coefficient con-
trol is no longer straightforward if two or more probes are simultaneously used, due to
their combined effect on the slab line.
The movement can be manually set by micrometer positioners, or it can be automati-
cally controlled by precise stepper motors. Manually controlled tuners are usually simpler
and cheaper. Automated tuners, however, allow reduced measurement time and greater
accuracy thanks to the precise stepped motors, making them nowadays the preferred
solution [16].
Ideally the probes should be non-contacting, touching neither the ground planes nor
the center conductor. This enables the motors to move the probes quickly and precisely,
with no perceptible wear or drift over time, thus providing longer tuner life and excellent
Tuner
probe
Main
line
Fig. 13.2 A slide screw tuner, with a conductive probe moving in two directions within a slab line.
Pre-matching Losses Passive Losses Pre-matching Passive

network tuner or tuner tuner or
active load active load
(a) (b)
Fig. 13.3 Pre-matching at device level (a) or at measurement system level (b).
repeatability. Some mechanical tuners realize a sliding contact between the probe and
the ground planes. Contacting probes are easier to design, but at the cost of shorter tuner
life, slower operation, and worse repeatability.
Due to the tuner’s intrinsically passive nature, the synthesized reflection coefficients
are limited in magnitude by the unavoidable losses of the test setup (due to cables,
on-wafer probes, etc.). Highly reflective loads cannot be realized at the DUT reference
planes, especially in the on-wafer environment. For instance, consider that an insertion
loss of 1 dB between the tuner and the probe tip transforms an ideal || = 1 at the tuner
output into a || = 0.8 at the DUT port. Even if tuner losses are completely removed, the
actual DUT load can be unsuitable for highly mismatched devices, especially at higher
frequencies (where losses are larger), thus precluding the investigation of interesting
regions of the Smith chart.
In order to overcome the problem, some fully passive solutions are available [17],
based on pre-matching networks at the device level – as shown in Figure 13.3(a) – or
on pre-matching tuners at the measurement system level – as shown in Figure 13.3(b).
The most recent versions of the mechanical passive tuners, integrate programmable pre-
matching capabilities [18]. As shown in Figure 13.3, in both pre-matching configurations
it is possible to use an active load instead of a tuner. This is treated in Sections 13.3
and 13.4.
13.2.2 Harmonic load-pull with passive tuners

As mentioned in Section 13.1, harmonic load-pull involves the control of the load reflec-
tion coefficient at a finite (usually small) number of harmonically related frequencies.
Figure 13.4 shows three different methods to use passive tuners for harmonic load-pull,
based on [19, 20], namely:
• cascaded tuners,
• triplexers with normal tuners,
• stub resonators.
The first solution refers to Figure 13.4(a): it simply cascades different tuners to increase
the number of degrees of freedom and allow fundamental and harmonic load tuning.
It increases the bench complexity and losses, but at the same time exhibits a greater
flexibility for frequency control and does not need any special hardware. Obviously, no
independent harmonic control is possible since the movement of one of the tuners affects
the impedance at both fundamental and harmonic frequencies.
Passive
tuner 1
Passive Passive
tuner 1 tuner 2 Triplexer
Passive
tuner 2
Passive
tuner 3
(a) (b)
Traditional Harmonic
probe probe
Harmonic
tuner
(c)
Fig. 13.4 Different types of harmonic tuning with passive tuners. Cascaded tuners (a), triplexer solution
(b), and stub resonator technique (c). ©2007 IEEE. Reprinted, with permission, from [21].
The second solution is shown in Figure 13.4(b). It uses different traditional tuners
for each harmonic, with a filter triplexer to separate the fundamental and harmonic
signals, so that they may be tuned independently [19]. In this way it is easier to change
bands and the entire Smith chart may be covered with independent controls, but the
insertion loss of the triplexer considerably limits the attainable reflection coefficient
magnitude.
Finally, harmonic tuners with a resonant probe are shown in Figure 13.4(c). They are
a compact, elegant solution that in principle allows high reflection coefficients along
with independent harmonic control. They are, however, relatively narrow-band and they
require cumbersome procedures to change the operating band (i.e. disassembling the
tuner, changing the slug, repeat the pre-characterization with a VNA). This might result
in reduced repeatability and reflection coefficient control.
13.3 Setting the load conditions: active, open-loop techniques
The passive tuners described in Section 13.2 provide a simple, mostly effective and
economic way to control the load conditions at the lower microwave frequencies and for
connectorized DUTs. However, in the presence of larger losses (i.e. at higher frequencies
and for on-wafer applications) passive load-pull systems do not allow to reach highly
reflective loading conditions. The problem is especially evident for harmonic tuning,
Microwave
source
Variable Phase
attenuator shifter
Input
DUT f0
amplifier DUT Circulator
Triplexer
2 1 2f0
3
Output 3f0
ΓL ΓL
amplifier
a1 b2 a1 b2
b1 a2 b1 a2
(a) (b)
Fig. 13.5 Active, open loop load-pull, with single source (a) [28], and with multiple sources (b) [29].
©2001 IEEE. Reprinted, with permission, from [7].
since the optimum harmonic termination is typically on the edge of the Smith chart [22].
Active-load systems were originally introduced at the end of the 1970s as a solution
to this problem; these days they are commercially available in different forms [23–27].
They probably represent the most reliable scheme for microwave and millimeter wave
load-pull test-sets.
A simple way to electronically synthesize a load reflection coefficient is to inject
a coherent signal into the DUT output. The concept can be explained by referring to
Figure 13.5. If the DUT delivers an out-going traveling wave b2 , controlling the in-going
wave a2 simply translates into setting the load reflection coefficient L to the value a2 /b2 .
The a2 signal can be taken from the same input source (as shown in Figure 13.5(a))
or from other external signal sources, coherent with the excitation signal (see
Figure 13.5(b)).
This technique was originally introduced by Takayama in 1976 [28] and it has
been widely used in industrial and research environments [29, 30]. It can be easily
extended to harmonic tuning, with the help of frequency multipliers [30], or with the
use of additional sources tuned to the desired harmonic frequencies [29], as shown in
Figure 13.5(b).
In the case of Figure 13.5(a), the synthesized load is controlled by the variable
phase-shifter and attenuator, and it is constant as long as the DUT outgoing wave b2
does not change with respect to a2 . For instance, the attenuator setting must be con-
tinuously adjusted during an input power sweep, to compensate for the output power
change.
In general, computer-controlled measurements are mandatory to achieve a constant
load. Iterative algorithms continuously monitor the actual load and properly control
the settings of the attenuator and phase-shifters in the configuration of Figure 13.5(a)
or the settings of the microwave sources in the configuration of Figure 13.5(b). This
increases the measurement time, as well as the possibility of failures that are potentially
destructive for the DUT (e.g. as in the case of load reflection coefficient magnitudes
higher than unity).
13.4 Setting the load conditions: active-loop techniques
The active loop is a closed-loop technique for microwave impedance synthesis in

load-pull measurements. Originally introduced in 1982 [32], it improves the control
of the load impedance compared to the open-loop techniques described in Section 13.3.
In contrast to them, it generates the in-going wave a2 by amplifying and phase-shifting
the out-going wave b2 , instead of generating it by a separate source.
13.4.1 Active loop: basics

The active-loop principle is shown in Figure 13.6(a), which shows its implementa-
tion at the output of a generic DUT. The out-going wave b2 produced by the DUT
itself is coupled through a directional coupler, attenuated, phase-shifted, amplified, and
re-injected back into the DUT as a2 . The “losses” block at the DUT output represents
all the attenuation between the DUT reference plane and the loop directional coupler
input, e.g. due to the fixture, probes, and cable losses. Under the assumption that all
the components are ideally matched to the same reference impedance, the synthesized
reflection coefficient L can be approximated by:
L = l 2 · k · α · G · e−j φ , (13.1)
Variable Phase Variable Phase

attenuator shifter attenuator shifter
ΓOUT Γ0 ΓOUT Γ0
DUT Circulator DUT S

Losses 2 1 Losses 2 1
3 3
Loop Loop
coupler Loop Loop
ΓL ΓAL ΓAL coupler
amplifier ΓL amplifier
(a) (b)
YIG
Variable Phase Variable Phase tunable
attenuator shifter attenuator shifter filter
ΓOUT ΓOUT
DUT Circulator DUT Circulator
Losses 2 1 Losses 2 1
3 3
Loop Loop
coupler Loop coupler Loop
ΓL amplifier ΓL amplifier
(c) (d)
Fig. 13.6 Active-loop principle schematic (a), the two possible oscillation types (b), and further
improvements to avoid oscillations (c), (d).
where l represents the losses between the DUT and the active loop (effectively as a
gain lower than one), k is the directional coupler coupling factor, α represents the vari-
able attenuation setting, G is the amplifier gain (all expressed in real, linear units),
while φ is the value of the loop phase-shift. The variable attenuator is used to change
the L magnitude, while the phase-shifter is used to change its phase. These two con-
trols act separately on the magnitude and on the phase, respectively, through α and φ
of (13.1).
Additionally, it is clear from (13.1) that L does not depend in first approxima-
tion on the output power of the DUT, i.e. no adjustments of the variable attenuator
are needed during automated input power sweeps. Along with the independent control
of the reflection coefficient magnitude and phase, this property makes the active-loop
technique advantageous over the open-loop methods in terms of simplicity, safety, and
ease of use.
However, active loops have two main drawbacks: namely, their potential instability
and the relatively large phase-shift inside the loop bandwidth. These are addressed in the
next sections.
13.4.2 Stability analysis of the active loop

The active loop of Figure 13.6(a) could be unstable for two main reasons [33].
The first one is related to the finite isolation of the loop directional coupler. This
creates a feedback path from the output of the loop amplifier back into the attenuator and
phase-shifter, indicated as a gray line in Figure 13.6(b). Oscillations can arise if the loop
gain exceeds unity – which is possible because of the gain G. Calling I the directional
coupler isolation, this “internal” loop gain is equal to I · α · G.
We will now find a condition to avoid this type of instability. The maximum required
gain G is obtained when the synthesized |L | is 1. In that case G = αl12 k . As α ≤ 1
always, then:
1
G≥ 2 . (13.2)
l k
To avoid oscillations, the loop gain must be lower than 1, i.e. I · α · G < 1. This is
always verified if G < I1 . From (13.2), we have
1 1
≤G< . (13.3)
l2k I
This means that to avoid oscillations it must hold that
1 k
2
< = D, (13.4)
l I
where D is the directional coupler directivity.

In conclusion, the larger the losses between the DUT and the active loop, the higher
the coupler isolation must be.
A second – and often more critical – reason for the loop instability is related to the
reflection coefficient O that the active loop sees at its input. For this reason, this issue is
sometimes referred to as the “external” loop stability. Referring again to Figure 13.6(b)
(dashed line), instability can arise when
|O | · |AL | ≥ 1. (13.5)
By calling S the scattering matrix of the “losses” block, the quantity O is:
S12 S21 OU T
O = S22 + . (13.6)
1 − S11 OU T
The condition for the stability is then:

S22 + S12 S21 OU T · |AL | < 1, (13.7)
1 − S11 OU T
which is always verified if:

S12 S21 OU T
|S22 ||AL | + · |AL | < 1.
(13.8)
1 − S11 OU T
Unfortunately, the quantity |AL | can be greater than one, because it must compensate
for the losses between the DUT and the active loop. In particular, if the “losses” block
is not well matched – i.e. |S22 | is considerably different to zero – and if the required
|OU T | is high (close to unity), this condition could be false and oscillations could be
triggered.
A possible remedy consists of moving the directional coupler as near as possible to the
DUT output. This solution was patented in 1999 [34] and it is shown in Figure 13.6(c).
Ideally, the stability condition is now
|OU T | · |L | < 1. (13.9)
In principle this is always true, as |OU T | < 1 and the optimum |L | of interest

cannot be larger than unity. Moreover, the reflection coefficient at the DUT reference
plane becomes:
L = l · k · α · G · e−j φ , (13.10)
while the “internal” loop gain is now I · α · G · l.
Moving the losses inside the loop therefore has multiple advantages:
• it improves the “external” loop stability, according to (13.9);
• it improves the “internal” loop stability, as its loop gain is multiplied by a factor l,
with l < 1;
• finally, the loop amplifier gain G needed to obtain the same L is lower, as can be
seen by comparing (13.1) and (13.10).
13.4.3 Practical active-loop implementations

Various practical implementations of the active-loop-based load-pull systems are
described in the literature [29, 30], [35–37] and the first systems have begun to be
commercially available in recent years [23–27].
In Section 13.4.2, the stability of the active loop was analyzed at its operating fre-
quency – i.e. the frequency at which the loop is intended to synthesize the desired
reflection coefficient. However, the loop is stable only if the same criteria are fulfilled
at all the frequencies inside the pass-band of the various loop components (attenuator,
phase-shifter, and amplifier) – particularly in the lower range, where the loop gain G is
usually higher.
The common solution consists of attenuating the loop gain outside the frequencies of
interest as much as possible, by adding a narrow band-pass filter centered at the loop
operating frequency. In particular,Yittrium Iron Garnet (YIG) filters combine very narrow
bandwidth (around 0.5% of the central frequency) with excellent tuning capabilities.
As shown in Figure 13.6(d), the YIG filter is usually located after the attenuator and
phase-shifter to minimize its input power.
Besides improving the loop stability, the YIG filter in the active loop enables a number
of interesting features.
First of all, it naturally makes the loop frequency-selective. Inside the YIG filter pass-
band the synthesized load is controlled in magnitude and phase by the attenuator and
phase-shifter settings (as described in Section 13.4.1). Outside the YIG pass-band, the
synthesized load is a match, as long as the circulator is well matched and the loop coupler
has a very low coupling factor (between −25 and −30 dB, so that the YIG filter out-
of-band high input SWR has no effect). This is a great advantage with respect to basic
passive tuners, that have uncontrolled and unpredictable behavior in frequency. As an
important side effect, an active loop can be designed to cover several octaves (by a proper
choice of attenuators, phase-shifters, and amplifiers) and its operating frequency can be
run-time reconfigured by controlling the YIG coil DC current.
Moreover, if the YIG coil current control is fine enough, it is possible to use it to
slightly tune the phase of the YIG filter S-parameter S21 . This phase-shift control is
faster and more repeatable than any mechanical phase-shifting. Figure 13.7(a–b) shows
the typical response of a YIG filter, magnitude and phase, while varying the main coil
current. In Figure 13.7(c–d), the responses are normalized to one of them, taken as a
reference. At 3 GHz central frequency, for example, a total phase variation of 160–170◦
is possible, while the S21 magnitude is still in the 3 dB bandwidth. Modern active load-
pull test-sets exploit this electronic phase-shifting capability to considerably speed up
the measurements.
Given the active-loop frequency selectivity, the realization of harmonic active loads
is straightforward. To achieve fundamental and harmonic tuning at the same time, or
multiple harmonic tuning, more loops can be combined using wideband components, as
in the example of Figure 13.8. In this implementation, all the passive components and the
loop amplifier must have at least a bandwidth of one octave. The loads at fundamental and
harmonic frequencies are controlled in a completely independent way. As we have seen
–15
YIG current 150 YIG current
–20 variation variation
–25 100
–30 50
|S21| (dB)
∠S21 (°)
–35
0
–40
–45 –50
–50 –100
–55 –150
–60
2.97 2.98 2.99 3 3.01 3.02 3.03 2.97 2.98 2.99 3 3.01 3.02 3.03
Frequency (GHz) Frequency (GHz)
(a) (b)
15
150
10
100
∠(S21/S21ref) (º)
|S21/S21ref| (dB)
5 50
0 0
–5 –50
–100
–10
–150
–15
2.97 2.98 2.99 3 3.01 3.02 3.03 2.97 2.98 2.99 3 3.01 3.02 3.03
Frequency (GHz) Frequency (GHz)
(c) (d)
Fig. 13.7 Typical YIG filter response versus main coil current, magnitude (a), and phase (b). In (c) and (d),
the same YIG responses are normalized with respect to the central one.
in Section 13.2.2, the realization of harmonic loads with passive tuners is possible, but
more complicated, due to the intrinsic wideband characteristics of these passive devices.
Similarly, more loops can be combined through hybrid couplers to obtain
differential/common-mode active loads [38–41], as shown in Figure 13.9. If the
microwave hybrids are ideal, the two loops independently tune the differential and
common modes. Harmonic differential- and common-mode loads are obtained sim-
ply by changing the YIG filter frequency. Differential/common-mode loads are the basis
of mixed-mode load-pull systems, which are described in Section 13.5.3. Mixed-mode
load-pull techniques are of increasing importance, as differential active devices (e.g.
transceivers and amplifiers) are being extensively used in many applications for the
reduction of the effects of external disturbances.
13.4.4 Wideband load-pull

Besides CW signal stimuli, modern applications also require testing devices under
two-tone excitations or wideband modulated signals, such as for multicarrier W-CDMA
applications.
YIG
Variable Phase tunable
attenuator shifter filter
Power
splitter Power
@ f0
combiner
YIG
@ 2f0
DUT Circulator
Losses
Loop
coupler Loop
amplifier
Fig. 13.8 Fundamental and harmonic tuning through active loop. In this implementation all the passive
components and the loop amplifier need to have a bandwidth of one octave. ©2007 IEEE.
Reprinted, with permission, from [21].
Unfortunately, the electrical delay introduced by a passive tuner and the connecting
cables makes the phase of the reflection coefficients presented to the DUT reference
planes not constant in frequency. Active techniques also suffer from the same problem,
especially if a YIG filter is present – as in the active loop. Under wideband excitation,
this means that the reflection coefficient varies inside the signal bandwith. At harmonic
frequencies, the in-band phase change of L is even larger. This is usually not acceptable,
as it does not correctly represent the loading conditions of the device when used in a
real circuit. In this case the variations of L and S with frequency are much smaller,
due to reduced dimensions of the circuit components (often approximated as lumped
elements).
Active techniques can be modified to overcome this problem, and Chapter 14 is
dedicated to the detailed description of such wideband systems.
13.4.5 Combining passive tuners and active techniques

The concept of combining passive tuners and active loops has been well known since the
late 1970s [19]. There are several ways to mix the two techniques and a vast literature
on the topic; see for example [19, 33, 35–37].
The idea is that, even when using a passive element (a tuner or a taper as in [36], or
a sliding short circuit as in [35]), it is always possible to inject an additional signal to
“boost” the effective reflection coefficient seen by the DUT. This additional signal can
Common-mode loop
YIG
Variable Phase Loop
tunable Circulator
attenuator shifter amplifier
filter
1 2
3
Differential-mode loop
1 2
3
Σ Δ Loop input
180° hybrid
180° 0°
Port out 1
Loop Δ Σ
coupler 1
0° 180°
Loop output
Loop 180° hybrid
coupler 2
Port out 2
Fig. 13.9 Differential active loop implementation. ©2006 IEEE. Reprinted, with permission,
from [38].
come from the same microwave source used for the device excitation (as in the case of
the open-loop techniques of Section 13.3) or from the device output itself (thus creating
a loop, similar to that described in Section 13.4). It can be coupled through directional
couplers, combiners, or circulators, and can be amplified or not.
Some examples are shown in Figure 13.10. The simple configuration of (a) is exploited
in [36], where the passive element is a taper. The hybrid-load configuration with a
directional coupler of (b) is described in [35], where the passive element is a sliding
short. The hybrid load with feedback loop of (c) appears in [37] (without amplifier) as
well as in [19]. Finally, the hybrid configuration with active loop of (d) is described in
detail in [33].
In all cases, the advantages of the hybrid loads are the following:
• higher reflection coefficients than using a passive element alone;

• less power needed to reach the same reflection coefficient, as compared to the active
load alone [33, 36].
Microwave
source
Port 2 Port 2 Passive

Passive Microwave
element source element
DUT DUT
ΓL ΓL
(a) (b)
YIG
Port 2 Passive (Amplifier) Port 2

element Circulator Circulator
2 1
3
DUT DUT
Passive
element
ΓL ΓL
(c) (d)
Fig. 13.10 Simplest passive/active hybrid load [36] (a). Hybrid load configuration with a directional
coupler [35] (b). Hybrid load with feedback loop [37, 19] (c). Hybrid configuration with active
loop [33] (d).
13.5 Measuring the DUT single-frequency characteristics
Sections 13.2 to 13.4 dealt with the problem of presenting a desired load reflection
coefficient L at the DUT ports, at single and multiple frequencies, as well as in
broadband cases. We now focus on the measurement of the DUT quantities, including the
synthesized loading conditions. Many of the concepts already described in Chapter 12
for generic nonlinear measurements are resumed here and analyzed from the perspective
of the practical load-pull test-sets.
Regarding the DUT measurement technique, there are two main types of load-pull
systems:
• non-real-time systems, which rely on the off-line pre-characterization of the tunable

loads;
• real-time systems, which measure the DUT loading conditions on-line, by using
properly calibrated instrumentation (usually, but not exclusively, a vector network
analyzer).
13.5.1 Real-time vs. non-real-time load-pull measurements

Figure 13.11(a) shows a typical implementation of a basic non-real-time load-pull mea-
surement system. The DUT input and the output power levels are measured by power
meters while sweeping the tuners in different positions. The S-parameter matrix of all
the components between the power meters and the DUT reference planes are measured
during a pre-calibration phase. In this way, after proper de-embedding, the power meter
measurements can be referred to the DUT reference planes. Since the tuners are included
in this calibration path, they must be measured for all the different positions that will
be used during the actual measurements. This is a time-consuming procedure, but since
the tuner is placed very near the DUT, the effect of losses is reduced and the reflection
coefficient at the reference plane can be close to unity.
The alternative is to swap the tuning device and the directional couplers, as shown in
Figure 13.11(b), to obtain a real-time system. The relation between the waves measured
through the couplers (am1 , bm1 , am2 , bm2 ) and the waves at the reference plane (a1 , b1 ,
a2 , b2 ) is now unique, and does not depend on the load or source impedances. If a VNA
Port 2
VNA
Port 1 Power 1 2
meters Input Output
Microwave fixture fixture
Power
source meter
3
Input variable Input DUT
attenuator amplifier Input Input Γs ΓL Output Output
block tuner tuner block
(a)
Switching network
am1 bm1 bm2 am2
Port 1 Port 2
Port 2
VNA
Port 1
Input variable Input Input DUT Output

a1 a2
attenuator amplifier load load
b1 b2
ΓIN ΓOUT
ΓS ΓL
(b)
Fig. 13.11 Simplified scheme of a non-real-time pre-calibrated load-pull system (a), and of a VNA based,
real-time load-pull system (b). ©2007 IEEE. Reprinted, with permission, from [21].
YIG
ΓOUT
DUT Circulator
2 1
Loop 3
coupler Loop
ΓL amplifier
bm2 am2
Fig. 13.12 Active loop in a real-time load-pull system.
is connected to the couplers, it becomes possible to apply a calibration procedure to

perform accurate measurements at the DUT reference planes.
Originally, the real-time configuration could be used only with active loads, since the
coupler losses set a severe limitation on the tuner reflection coefficients. In particular,
the best configuration to improve the loop stability is the one shown in Figure 13.12,
where the measurement couplers are placed inside the loop (see also Section 13.4.2).
More recently, the introduction of very low loss, ultra-wideband directional couplers
[42,43] with an insertion loss in the range 0.1–0.2 dB up to 15 GHz, has made it possible
to use the real-time configuration with passive tuners.
An analysis of the uncertainty contributions and a comparison of the uncertainty
budget for real-time and non-real-time techniques up to 40 GHz has been presented in
[44]. Experiments and simulations have shown how real-time techniques are far more
accurate than non-real-time, especially as the frequency increases.
The calibration procedure typically used for real-time systems is an extension of the
traditional ones for S-parameter measurements. Since it deserves some explanation, it
will be described in detail in the next section.
13.5.2 Calibration of real-time systems

The error-correction theory for load-pull network analyzer-based measurements has been
described mainly in three fundamental papers by Tucker et al. [45], Hecht [46], and
Ferrero et al. [47]. Even though it can be seen as a subset of the calibration techniques
for the nonlinear vector network analyzers described in Chapter 12, it is here summarized
to understand its implications to the setup of a real-time load-pull system.
The error coefficients of a traditional, two-port VNA calibration are the eight elements
of the matrices X1 and X2 defined as [48]:

a1 bm1
= X1 (13.11)
b1 am1
and
a2 bm2
= X2 , (13.12)
b2 am2
where
l1,2 −h1,2
X1,2 = = k1,2 Y1,2 . (13.13)
k1,2 −m1,2
If all the eight error coefficients are known, it is possible to find the actual power at
the DUT input reference plane as
Pin = |a1 |2 − |b1 |2 = |l1 bm1 − h1 am1 |2 − |k1 bm1 − m1 am1 |2 .
Only seven out of these eight coefficients are the outcome of any classical S-parameter
calibration, since the S-parameters are defined as ratios between waves. In particular,
the coefficients have to be normalized with respect to k1 and only the quantities l1 /k1 ,
m1 /k1 , h1 /k1 and k2 /k1 , l2 /k1 , m2 /k1 , h2 /k1 , are known. The input power Pin can be
computed only if the last coefficient k1 is known, using the following formula:

Pin = |k1 |2 |l1 /k1 · bm1 − h1 /k1 · am1 |2 − |bm1 − m1 /k1 · am1 |2 = |k1 |2 Pin
n
.
If it is possible to connect a power meter at the input port, the magnitude of the
coefficient k1 can be found during the calibration. If Pm is the reading of the power
meter, it holds that:
Pin = Pm = |k1 |2 Pin
n
, (13.14)
and therefore the error coefficient k1 is computed as
Pm
|k1 |2 = n. (13.15)
Pin
This is a straightforward solution for connectorized DUTs, where a power meter

is likely to be available with the same type of connector as the DUT. However, the
technique needs some adaptation for on-wafer measurements, as the direct connection
of an accurate power sensor at the DUT reference planes is not possible.
The technique described in [47] does not use any on-wafer power sensor; it is shown
in Figure 13.13. It is based on the assumption that the output reflectometer has typically
an on-wafer port (port 2) and a connectorized one (port 3, e.g. a coaxial port). Thus, the
reflectometer measurements am2 and bm2 have a fixed relationship with port 2 waves (a2 ,
b2 ), and a different – but still easy to find – relationship with the port 3 waves (a3 , b3 ).
The calibration is performed in three conceptual steps.
1. First, a traditional, two-port, on-wafer calibration (e.g. TRL) is performed at the DUT
ports, with the system configured as in Figure 13.13(a).
2. Second, a generic two-port DUT (for instance, a thru) is connected to the on-wafer
ports and a one-port SOL calibration is performed at the coaxial reference plane port
3, as shown in Figure 13.13(b).
Calibration phase I: S-parameter calibration
Switching network
am1 bm1 bm2 am2
Port 1 Port 2 Port 3

Port 2
VNA
Two-port CAL
Port 1 standards
a1 a2
b1 b2
(a)
Calibration phase II: coaxial port 3 calibration
Switching network
am1 bm1 bm2 am2

Short
VNA
Port 1 Thru a3 Open
b3
a1 a2 Load
b1 b2
(b)
Calibration phase III: power meter measurement at port 3
Switching network
am1 bm1 bm2 am2

VNA
Power
Port 1 Thru
a3 meter
b3
a1 a2
b1 b2
(c) T
Fig. 13.13 Steps for the real-time load-pull system calibration.
3. Finally, the power level at the port 3 reference plane is measured by a coaxial reference
power meter, as in Figure 13.13(c).
In other words, this procedure replaces the on-wafer power reading at port 1 with a
coaxial power measurement at port 3, at the price of an additional one-port calibration at
the coaxial port. The device connected between ports 1 and 2 during the coaxial port cal-
ibration should allow the transmission of some power to excite the output reflectometer,
but the knowledge of its S-parameter is in principle not needed.
With an error box notation similar to (13.11) and (13.12) we can write

a3 bm2
= X3 , (13.16)
b3 am2
where
l3
l3 −h3 − hk33
X3 = = k3 k3 = k3 Y3 . (13.17)
k3 −m3 1 − mk33
Matrix Y3 can be obtained from the one-port calibration procedure in Figure 13.13(b).
However, the output reflectometer is a two-port, reciprocal network. Its transmission
matrix T , shown in Figure 13.13(c), imposes the condition

a2 T11 T12 a3 a3
= =T . (13.18)
b2 T21 T22 b3 b3
By substituting (13.16) in (13.18) and then in (13.12) we obtain

bm2 bm2
X2 = T X3 (13.19)
am2 am2
and therefore
k3
(Y2 )−1 T Y3 = I , (13.20)
k2
where I is the (2 × 2) identity matrix. By extracting the determinant of both sides of
(13.20) we get
k2 det Y3
=± , (13.21)
k3 det Y2
being det T = 1, as the reflectometer is reciprocal.
The power meter reading Pmt at port 3 allows us to compute |k3 |:
Pmt
|k3 |2 = n , (13.22)
Pmt
where
n
Pmt = |l3 /k3 · bm2 − h3 /k3 · am2 |2 − |bm2 − m3 /k3 · am2 |2 (13.23)
is computed from the error coefficients in Y3 and the network analyzer readings am2 ,
bm2 when the power meter is connected.
In conclusion we obtain:

det Y3 Pmt
|k2 | =
n . (13.24)
det Y2 Pmt
This eventually allows us to compute k1 and to solve the power calibration problem,
because the ratio k2 /k1 is known from the two-port S-parameter calibration.
13.5.3 Mixed-mode, harmonic load-pull systems

The first experiments on harmonic load-pull measurements were performed in the early
1990s [30, 49, 13]. In [49] the system was based on six-port-reflectometer measurements
and the influence of the harmonic termination at 2f0 on the output power was for the
first time demonstrated for a MESFET device. Similarly, the influence of harmonic
terminations on the transistor PAE was demonstrated in [13] by using a VNA-based
receiver. In the following years the technique was further refined and many innovative
results obtained.
This section describes one of the most recent achievements of active-loop-based load-
pull techniques [38] and demonstrates the potentials of active harmonic load-pull for
differential device characterization (in particular, a Bluetooth transceiver). The com-
plete measurement system is shown in Figure 13.14. The active differential load is the
one already shown in Figure 13.9, with the possibility of tuning the loads at funda-
mental or harmonic frequencies by simply changing the YIG tuning frequency. Again,
Spectrum
analyzer
Sampler 1
and phase lock
Common-mode loop
1 2
3
1 2
3
Σ 180° Δ
hybrid
180° 0°
Port 1
b1 Differential-mode loop Δ 180° Σ
hybrid
a1 0° 180°
Bluetooth bm1 am1
transceiver a2
b2
bm2 am2
Port 2
Meas switch
Sampler 2
Fig. 13.14 Differential/common-mode harmonic load-pull system. ©2006 IEEE. Reprinted, with
these features can be achieved only by active-loop-based systems, as passive tuners can-
not independently tune the differential- and common-mode loads at fundamental and
harmonics.
The measurement system consists of two reflectometers, a VNA, and a spectrum
analyzer. The measured performances are:
D ≡ bD /aD
C ≡ bC /aC
(13.25)
PD ≡ |aD | − |bD |2 = |aD |2 (1 − |D |2 )
2
PC ≡ |aC |2 − |bC |2 = |aC |2 (1 − |C |2 ),
where the differential- and common-mode quantities are defined as:

aD 1 1 −1 a1
≡√ (13.26)
aC 2 1 1 a2

bD 1 1 −1 b1
≡√ . (13.27)
bC 2 1 1 b2
The system can be calibrated with any classical load-pull calibration at ports 1 and 2,
as shown in Section 13.5.2.
Furthermore, the use of a spectrum analyzer allows the measurement of spurious
common-mode as well as differential power ratios at all the frequencies of interest.
The transceiver includes an internal VCO and a power amplifier. It is mounted on
a connectorized evaluation board (including power supply and control signals) and it
generates a differential output signal at 2.402 GHz.
Figure 13.15 shows the load-pull contours of differential output power and pulling
(a spurious signal, generated by the VCO harmonics), when C at all frequencies is
ΓD @ f0 map PD, dBm ΓD @ f0 map Pulling, dBc

4 −30
2 −35
0
−40
−2
−4 −45
−6
−50
−8
−10 −55
(a) (b)
Fig. 13.15 D @ f0 load-pull contour map of differential output power (a) and pulling (b). ©2006 IEEE.
5
0 PD, dBm
Pulling, dBc
–5
–10 ΓD @ 2f0 map
–15
–20
–25
–30
–35
–40
–45
–50
–200 –150 –100 –50 0 50 100 150 200
ΓD @ 2f0 phase, deg
(a)
5
0 PD, dBm
Pulling, dBc
–5
–10 ΓD @ 2f0 map
–15
–20
–25
–30
–35
–40
–45
–50
–200 –150 –100 –50 0 50 100 150 200
ΓD @ 2f0 phase, deg
(b)
Fig. 13.16 Effect of differential-mode (a) and common-mode (b) harmonic load on pulling and output
differential power (D @ f0 = 0.3 ∠ 115◦ ). ©2006 IEEE. Reprinted, with permission, from [38].
kept constant to a value close to zero. These are typical load-pull maps, that allow the
identification of optimum D @ f0 loads for output differential power or pulling.
The influence of the harmonic differential termination on output power and pulling is
shown in Figure 13.16. While sweeping the harmonic loads, D @ f0 is kept constant
at the optimum value for output differential power, i.e. 0.3 ∠ 115◦ . This was the first
experimental verification of the effect of C @ 2f0 (around 10 dB variation) and of
the very limited influence of D @ 2f0 on the pulling signal. This demonstrates the
innovation potential of harmonic mixed-mode load-pull measurements.
13.6 Measuring the DUT time domain waveforms
As seen in the previous section, the quantities of interest for load-pull are typically input
and output power, gain, PAE, intermodulation products and ACPR, oscillator pulling,
etc. – i.e. quantities mainly defined in the frequency domain. In addition, the measure-
ment of the complete voltage and current time domain waveforms at the DUT input
and output ports is being given increasingly more attention. As examples, time domain
waveform information can help in the design of high-efficiency power amplifiers [50,
51], in building and validating more sophisticated nonlinear models [52, 53], and – in
general – in reaching a deeper understanding of the active DUT behavior [54].
As extensively discussed in Chapter 12, a nonlinear device excited at a fundamental
frequency f0 in large-signal conditions generates distorted waveforms, i.e. signals with
non-negligible harmonic content. Their measurements at microwave frequencies can be
carried out in two ways:
• directly in the time domain, by sampling the periodic waveforms (often “sub-Nyquist”)
as in traditional sampling oscilloscopes;
• in the frequency domain, by measuring the harmonic components of the signal in
magnitude and phase (e.g. with a VNA), and by applying an inverse Fourier transform
to obtain the time domain waveforms.
In both cases, the measurement still implies addressing the usual issues for load-pull
characterization: setting of the DUT loading conditions and calibrating the measure-
ment equipment. While the passive and active load techniques described in the previous
sections are still applicable, the calibration algorithms presented so far need to be
extended to implement systematic error correction for the time domain waveforms. As
will be clear in the following, it is still convenient to perform this operation in the
frequency domain, given the linearity assumption for the measurement setup.
13.6.1 Load-pull waveform techniques in the time domain

A first family of load-pull waveform techniques measures the DUT signals with
high-speed digital oscilloscopes or similar sampling equipment. Among the possible
instrumentation, it is worth mentioning the Microwave Transition Analyzer [55] (now
discontinued, but still used in research laboratories) and the Large Signal Network Ana-
lyzer [56, 57], which can be considered as the evolution of the MTA: very similar in
principle, but much faster, with an increased dynamic range and more channels. Still,
any sampling oscilloscope with at least two channels can be generally employed in a
time domain waveform load-pull setup.
In the first implementations, the sampling oscilloscope was used in a very simple
way, as shown in Figure 13.17(a). It was possible to monitor the output voltage on the
oscilloscope screen while varying the load, typically with passive techniques. No vector
error correction was possible (as only two out of four DUT waves were measured) and
the load could be measured with a VNA during a pre-calibration phase. Of course, this
solution was used to provide a very quick, but only qualitative evaluation tool.
Sampling
oscilloscope
Ref Test
Tuner
DUT
(a)
Time-domain
waveform
receiver
Ref Test
am1 am2
bm1 bm2
Tuner and/or
active loop
a1 a2
b1 b2
Port 1 Port 2
(b)
Fig. 13.17 Time domain waveform load-pull in its simplest implementation (a) and in the vector-corrected
configuration (b). ©2008 IEEE. Reprinted, with permission, from [58].
A more complete solution is shown in Figure 13.17(b). The time domain receiver “test”
channel is used to measure all the four incident and reflected waves by a microwave
multiplexing switch. Obviously, if the time domain receiver had four test channels, the
four waves could be measured simultaneously, the switch would be no longer needed,
and the measurement would be much faster.
As anticipated, it is still convenient to apply the systematic error-correction procedure
in the frequency domain. In particular, the sampled waveforms are first transformed in
the frequency domain via FFT. Since the signals are periodic, they can be represented by
a discrete set of phasors. We will refer to them as am1,n , am2,n , bm1,n , and bm2,n , where
n = 1, 2, ... represents the order of the different harmonic components.
The error model is still the same one as described in Section 13.5.2, here reported for
completeness:

a1,n bm1,n
= X1,n (13.28)
b1,n am1,n
and
a2,n bm2,n
= X2,n , (13.29)
b2,n am2,n
where a1,n , a2,n , b1,n , and b2,n are the phasors at the DUT reference planes. The frequency
dependence of the two error coefficient matrices is made explicit by the n subscript.
If the error coefficients are known for all the frequencies of interest, the phasors at the
DUT reference planes can be computed from the measured signals. The corresponding
voltage and current phasors are given by
|Z |
Vi,n = √ ref (ai,n + bi,n )
{Zref }
|Zref | (13.30)
Ii,n = √ (ai,n − bi,n )
Zref {Zref }
and the time domain waveforms are eventually reconstructed as

$
vi (t) = N |V | cos(2π kf0t + ∠Vi,n )
$n=1 i,n (13.31)
ii (t) = Nn=1 |Ii,n | cos(2π kf0t + ∠Ii,n )
From these equations it turns out that the phase of the different harmonic components
(i.e. ∠Vi,n , ∠Ii,n ) has to be error-corrected, as much as the corresponding magnitudes.
The calibration procedure described in Section 13.5.2 already allows the use of the
VNA as a selective power meter, by computing the magnitude of the error coefficient k1,n
with a power meter reading. Based on the previous considerations, it is straightforward to
recognize that this is not sufficient for time domain waveform reconstruction; the phase
of k1,n also needs to be found during the calibration procedure.
This additional step – often referred to as phase calibration – is generally performed
with the measurement of a pre-characterized DUT (a golden device), which produces
traceable time domain waveforms at its reference planes. This procedure was originally
introduced in 1989 [59], and the principle is still under improvement[60]. NIST traceabil-
ity is obtained by measuring the golden device with a sampling oscilloscope, previously
calibrated using the “nose-to-nose” [61] technique.
13.6.2 Load-pull waveform techniques in the frequency domain

A second family of load-pull waveform techniques uses a network analyzer to vector-
ially measure (i.e. in magnitude and phase) the harmonic components of the signals
in the frequency domain, instead of relying on time domain sampling. Given the
high frequency-selectivity of a VNA with respect to a broadband oscilloscope, this
theoretically results in a much higher dynamic range and measurement accuracy.
However, a VNA cannot be directly used to measure load-pull time domain waveforms,
mainly because the coherency of the phase measurement is not maintained during the
frequency sweep. Some modifications have to be introduced in the measurement setup.
We will try to explain this concept with the help of Figure 13.18.
It is well known that the basic, conceptual block of a VNA is the vector voltmeter. This
is basically a super-heterodyne receiver with two channels (namely, test and reference),
f Bandpass
Reference f filter
VTO APC
Bandpass Voltmeter Phasemeter
filter
Test
Iinear DUT
(a)
10 MHz ref
Bandpass
f0 f f0 2f0 3f0 4f0 filter
Reference
VTO APC Bandpass Voltmeter Phasemeter

filter
Test
Non-linear DUT
f0 2f0 3f0 4f0

(b)
Comb
generator
10 MHz ref
f0 f Bandpass
f0 2f0 3f0 4f0 filter
Reference
VTO APC Voltmeter Phasemeter

Bandpass
filter
Test
Non-linear DUT
f0 2f0 3f0 4f0

(c)
Fig. 13.18 Simplified scheme of a vector voltmeter (a). In (b) the reference signal sweeps through the
various harmonics of the distorted signal under test, and it is not possible to have a stable reading
of the phase of the harmonics. In (c), the reference signal is taken from a comb generator having
fixed harmonic phases; now the measurement of the distorted signal is possible.
which is tuned to measure the test and reference phasor magnitude at a certain frequency,
along with the phase of the test phasor with respect to the reference one.
For reflection coefficient measurements, the reference channel typically measures a
signal coupled from the excitation signal, as depicted in Figure 13.18(a). If the mea-
surement and excitation frequency changes, the phase of this reference signal changes
randomly, which generally is not an issue. The same setup could be used to measure the
harmonics of a distorted test signal, generated by a nonlinear DUT one-by-one. In prin-
ciple, this could be done as in Figure 13.18(b), by simply changing the reference signal
frequency with a second, auxiliary microwave source, sharing the same reference clock
with the main one. In practice however, the auxiliary source usually does not maintain
the same phase reference (i.e. the same “origin of the time axis”) while sweeping its fre-
quency. Eventually, the measured phase values of the test signal harmonic components
are not meaningful.
In order to correctly measure the harmonic phases, a stable reference signal, contain-
ing all the harmonics of interest, must be fed into the reference channel, as shown in
Figure 13.18(c). This “comb” signal can be generated by any kind of nonlinear device
and must be coherent with the excitation source.
The calibration of the complete load-pull system, shown in Figure 13.19, can be
performed in the same way as described in Section 13.6.2 for sampling-oscilloscope-
based systems, i.e. with the additional measurement of a pre-characterized, traceable
nonlinear device to compute the phase of the error coefficient k1,n . The waveforms are
then reconstructed using (13.30) and (13.31). In 2005, a VNA-based load-pull system,
exploiting this measurement technique, was presented. It had a very large bandwidth (300
kHz–20 GHz) which allowed the reconstruction of complex modulated waveforms, with
80 dB dynamic range [62]. More recently, a 4-port measurement system, with differential-
and common-mode load tuning capabilities and time domain waveform measurements
has been proposed [41].
VNA
Ref Test
Comb signal
4 to 1 switch
am1 am2
bm1 bm2
Tuner and/or
active loop
a1 DUT a2
b1 b2
Port 1 Port 2
Fig. 13.19 Time domain waveform load-pull with frequency domain receiver and phase lock on a comb
generator. ©2008 IEEE. Reprinted, with permission, from [58].
13.6.3 Other calibration approaches

The load-pull waveform techniques described in the previous sections rely on the use of
a golden nonlinear device (e.g. a golden diode) to calibrate the phase of the harmonic
components. This can be avoided if an extra sampling oscilloscope/MTA/LSNA channel
is available during calibration.
In fact, the golden device typically used in phase calibration is nothing but a transfer
standard: it transfers the accuracy of the oscilloscope that has characterized it to the
load-pull system. If, however, an additional calibrated time domain receiver is available
(i.e. a calibrated channel of an oscilloscope, MTA or LSNA), it can be directly used for
the phase calibration. This method is more robust with respect to the transfer standard
and it removes the effects of the connection repeatability. The calibration accuracy and
traceability are then directly related to those of the auxiliary time domain receiver, which
must be calibrated at its reference plane (e.g. by a “nose-to-nose” calibration [61] or with
the NIST pulse standard).
Similar to the setup described in Section 13.5.2, this method makes use of an auxiliary
port (port 3) during calibration. The S-parameter and power calibration steps are the same
as described in Figure 13.13(a–c). The only difference is the receiver type, which is no
longer aVNA but it can be any kind of time domain receiver. The step for phase calibration
is shown in Figure 13.20. The source generates a CW signal, whose frequency is stepped
through the values of interest (i.e. the intended fundamental and harmonic frequencies).
At the generic k-th frequency, the measurements of the raw waves am1,n , bm1,n , am2,n ,
and bm2,n are acquired simultaneously with the voltage V3,k at port 3 by the auxiliary
time domain receiver channel.
Since a SOL calibration is already performed at port 3 according to Figure 13.13(b),
the reflection coefficient a3,n /b3,n of the calibrated auxiliary channel can be computed,
and b3,n is obtained by (13.30) (with i = 3). Moreover, the elements of the scattering
matrix Sn of the passive network between ports 2 and 3 in Figure 13.20 can be easily
computed.
Calibration phase IV: time domain measurement at port 3
Switching network
am1 bm1 bm2 am2
Time domain Time domain
ref channel test channel
Port 1 Port 2 Port 3
Thru
a3
b3 Time domain
a1 a2 auxiliary test
b1 b2 channel
Fig. 13.20 Phase calibration using a calibrated auxiliary time domain receiver.
By comparing the b2,n wave at the DUT reference plane 2 obtained with the two
receivers:
S12,n
b2,n = b3,n a = k2,n bm2,n − m2,n am2,n , (13.32)
S22,n b2,n
2,n
− n
where m is S11,n S22,n − S12,n S21,n , it is possible to find the phases ∠k2,n .
The calibration problem is eventually solved by computing ∠k1,n , since the complex
ratio k2,n /k1,n is known from the S-parameter calibration at ports 1 and 2.
13.6.4 Measurement examples

Figure 13.21 shows the time domain drain currents of a microwave FET in common
source configuration during a power sweep, for two different classes of operation –
namely, class A and B [58, 63]. As an example, this type of plot allows the observation
of overshooting of the instantaneous current values that could degrade the transistor
efficiency, and to reshape them by properly choosing the fundamental and harmonic
loading conditions [54, 64].
Similarly, the plots in Figure 13.22 show the trajectory of the instantaneous point
in the ID -VDS plane for different power levels (solid lines). The transistor DC charac-
teristics are superimposed for reference (dotted lines). These plots allow the observation
of potential loading conditions that could bring the transistor to operate in unsafe areas
of the I-V plane for a certain class of operation.
Class A
800
600
Id, mA
400
200
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Class B
800
600
Id, mA
400
200
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time, ns
Fig. 13.21 Time domain drain current waveforms for a FET, biased in class A and class B. ©2008 IEEE.
(%)
Class A 28
800 27
26
25
700
24
23
600 22
21
20
500 19
iDS, mA
PAE @ dB compression load map

400 ΓLopt = 0.26, –20 deg
300
200
100
0 10 20 30 40 50 60
VDS, V
(a)
Class B
800 (%)
54
52
700 50
48
46
600 44
42
500 40
38
36
iDS, mA
400 34
300 PAE @ 3 dB compression load map

ΓLopt = 0.54, 13 deg
200
100
0 10 20 30 40 50 60
VDS, V
(b)
Fig. 13.22 Dynamic load lines for increasing power, on the best load for PAE, in class A (a) and B (b).
©2008 IEEE. Reprinted, with permission, from [58].
13.7 Real-time source-pull techniques
Source-pull systems have important applications in low noise amplifier design, where the
lowest noise figure and, in general, the optimal transistor noise parameters are found by
applying different source impedance values [65, 68] (for this topic, see also Chapter 10).
Moreover, it is well known that the harmonic source impedance can heavily influence
the power-added efficiency in power amplifiers [69].
Source-pull measurements use, in principle, the same techniques to synthesize the
source reflection coefficient as described in Sections 13.2 to 13.4. However, the pres-
ence of a generator term in the source equivalent circuit poses additional challenges in
the accurate measurement of the source reflection coefficient S . We will refer to the
simplified test-set scheme shown in Figure 13.11(b) to explain this concept. The VNA
and two reflectometers measure the waves at the reference planes of the DUT, while two
tuners set the source and load conditions, respectively, at the input and output ports.
After the calibration procedure described in Section 13.5.2, the reflectometer is able
to measure the input reflection coefficient of the circuit connected to its test port. For
example, the port 1 reflectometer in Figure 13.11(b) allows calibrated measurement of
the DUT input reflection coefficient I N as
b1
I N = . (13.33)
a1
Equation (13.33) defines the relationship – set by the DUT – between the waves at the
input reference plane. On the other side, the microwave source imposes
a1 = aS + S b1 (13.34)
where S is, by definition, the source reflection coefficient. From (13.34), it results

a1 aS
S = 1− . (13.35)
b1 a1
Therefore, a single measurement of a1 and b1 is not sufficient to compute the source

reflection coefficient. In fact, S is equal to the ratio a1 /b1 only if aS = 0, i.e. the internal
generator is switched off.
Hughes et al. [70] proposed the solution shown in Figure 13.23(a). First, the source
switch is set to position 1 and the DUT input gamma is computed by (13.33). Then,
the source switch is turned to position 2 and a second acquisition of waves a1 and b1 is
performed. From (13.35), the source reflection coefficient is simply the ratio S = a1 /b1 ,
since the source term is null. This simple technique relies on two basic assumptions. First,
the DUT is not unilateral, so that a significant portion of the source signal from port 2 can
reach the input reflectometer. Moreover, the reflection coefficient of the source switch
SW does not change while turning the switch from position 1 to 2.
An entirely different approach is described in [71] and shown in Figure 13.23(b).
Here the signal from the microwave source is summed with the wave reflected by the
tuning element and injected into the DUT. The reflectometer is used in an unconventional
configuration (referred to as reverse) and it directly monitors the tuner coefficient T .
After a proper calibration procedure, S is directly available, but – this time – it is the
DUT reflection coefficient I N that cannot be determined.
ΓSW
Microwave ΓS
source
1
Source Input Reflectometer
switch tuner ΓIN
(a)
ΓT ΓS
Input Reflectometer
tuner ΓIN
Microwave
source
(b)
ΓT ΓS
Input Reflectometer
tuner
Microwave
source ΓIN
(c)
Tree-sampler
reflectometer
ΓT ΓS
Input
tuner
Microwave ΓIN
source
(d)
Fig. 13.23 Existing solutions for source reflection coefficient measurement. ©2001 IEEE. Reprinted, with
The method shown in Figure 13.23(c) solves the latter problem in two steps [72]. First,
the microwave signal is injected before the reflectometer and the DUT input characteristic
is computed. Then, it is switched immediately after, and the source reflection coefficient
is measured by the reflectometer in the reverse configuration. Again, the switch reflection
coefficient is assumed constant while changing the switch position.
As a common feature, all the previous techniques measure the DUT and the source
reflection coefficients by two different steps. For fast and automatic characterization of
active devices, that can be time consuming. The technique proposed in [73] is based
on the concept of a three-sampler reflectometer (see Figure 13.23(d)), which allows the
simultaneous determination of source and DUT input gamma. This technique is indeed
fast and accurate, but it is based on an unconventional error model and it requires a
special-purpose calibration procedure.
An alternative, simple, yet rigorous method is shown in [74] for determining the source
reflection coefficient while characterizing active devices. Briefly, it consists in measuring
the waves at the input reference plane under two different DUT bias conditions. The
variations of the DUT input waves due to the bias change give enough information to
compute the source reflection coefficient with sufficient accuracy for most applications.
13.8 Conclusions
Far from being exhaustive, this chapter has shown the main techniques and issues of the
so-called “load-pull” measurements.
Passive, non-real-time, fundamental load-pull systems still provide a simple and robust
solution for a wide range of applications. They are mainly used for basic characterization
of microwave transistors, as well as to experimentally find the optimal loads for the design
of relatively simple microwave nonlinear circuits. However, they suffer from a number
of issues. The maximum magnitude of the synthesized load is typically limited by the
setup losses. They rely on often tedious pre-characterization procedures of the tuners,
which have a limited repeatability and affect the measurement accuracy [44]. Finally,
the load at harmonic frequencies is not well controlled (if at all).
Real-time load pull has the advantage of improved accuracy. The vector measurements
of a VNA enable the implementation of rigorous error-correction and calibration methods
and allow on-line monitoring of the loading conditions. However, this introduces a first
step in cost as well as in complexity of the measurement system, which can be handled
only by properly trained personnel. Recent advances have improved the losses in passive
systems to implement real-time techniques, but active loads are still the best option for
measurement speed and maximum reflection coefficient magnitudes.
Among the active loads, the open-loop techniques are not affected by oscillation
issues that could occur in active-loop systems. They allow wideband characterization
(e.g. in the presence of wideband modulated signals), as is shown in Chapter 14, but
they require a more complicated control to synthesize and maintain the desired load.
However, load setting is much easier and more robust with active-loop techniques, which
also allow straightforward control of common vs. differential-mode load impedances in

mixed-mode measurements.
If the DUT operates in strong nonlinear conditions, such as for high-efficiency ampli-
fiers, harmonic load-pull becomes important to keep the loading conditions at harmonic
frequencies under control. Similarly, if the device is used in wideband conditions (i.e.
with bandwidths of some MHz) it is important that the load remains constant within the
input signal bandwidth to resemble the real operating conditions of the device. This topic
is discussed in detail in Chapter 14. In both cases, however, the system measurement
setup is complicated and no off-the-shelf solutions are available on the market.
When the complete signal waveforms at the DUT input and output ports are needed
(e.g. for nonlinear model extraction), load-pull time domain waveform techniques are
applied. Time domain sampling-based techniques allow fast waveform acquisition but
suffer from limited dynamic range related to the broadband oscilloscope input stages.
VNA-based techniques measure the different harmonic components in the frequency
domain separately, which generally slows down the measurement. However, the high
dynamic range instrinsic to the VNA narrow-band measurement principle makes this
technique suitable for millimeter-wave applications. As a final consideration, load-pull
waveform techniques provide the most comprehensive tool for thorough large-signal
device characterization, but always at the price of a considerable test setup complexity,
which makes them mainly suitable for advanced R&D labs.
References
[1] D. Hamilton, J. Knipp, and J. Kuper, Klystrons and Microwave Triodes. New York: McGraw-
Hill, 1948.
[2] G. Heiter, “Characterization of nonlinearities in microwave devices and systems,” IEEE
Trans. Microw. Theory Tech., MTT-21, pp. 797–805, Dec. 1973.
[3] J. Cusack, S. Perlow, and B. Perlman, “Automatic load contour mapping for microwave
power transistors,” IEEE Trans. Microw. Theory Tech., MTT-22, pp. 1146–1152, Dec. 1974.
[4] D. Le and F. Ghannouchi, “Multitone characterization and design of FET resistive mixers
based on combined active source-pull/load-pull techniques,” IEEE Trans. Microw. Theory
Tech., MTT-46, pp. 1201–1208, Sept. 1998.
[5] F. Ghannouchi and R. Bosisio, “Source-pull/load-pull oscillator measurements at
microwave/mm wave frequencies,” IEEE Trans. Microw. Theory Tech., MTT-41, pp. 32–35,
Feb. 1992.
[6] E. Davis, “Rieke diagrams for avalanche diodes show performance as a function of load,”
Proc. IEEE, 55, pp. 1521–1522, Aug. 1967.
[7] A. Ferrero, V. Teppati, G. Madonna, and U. Pisani, “Overview of modern load-pull and other
nonlinear measurement systems,” in 58th ARFTG Conference Digest, San Diego, USA, Nov.
2001.
[8] S. C. Cripps, “A theory for the prediction of GaAs FET load-pull power contours,” in IEEE
MTT-S Intl. Microwave Symp. Dig., Boston, MA, May 1983, pp. 221–223.
[9] S. C. Cripps, “Old-fashioned remedies for GaAs FET power amplifier designers,” IEEE
MTT-S Newsletters, pp. 13–17, Summer 1991.
[10] F. Sechi, “High efficiency microwave FET power amplifier,” Microwave J., pp. 59–63, Nov.
1981.
[11] I. Bahl, E. Griffin, A. Geissberger, C. Andricos, and T. Brukiewa, “Class-B power MMIC
amplifiers with 70 percent power-added efficiency,” IEEE Trans. Microw. Theory Tech.,
MTT-37, pp. 1315–1320, Sept. 1989.
[12] D. Snider, “A theoretical analysis and experimental confirmation of the optimally loaded and
overdriven RF power amplifier,” IEEE Trans. Electron Devices, ED-14, pp. 851–857, Dec.
1967.
[13] A. Ferrero and U. Pisani, “Large signal 2nd harmonic on wafer MESFET characterization,”
in 36th ARFTG Conf. Dig., Monterrey, CA, Dec. 1990, pp. 101–106.
[14] Maury Microwave Corporation [Online]. Available: http://www.maurymw.com.
[15] Focus Microwaves Inc. [Online]. Available: http://www.focus-microwaves.com.
[16] F. Sechi, R. Paglione, B. Perlman, and J. Brown, “A computer controlled microwave tuner
for automated load pull,” RCA Review, 44, pp. 566–572, Dec. 1983.
[17] J. F. Sevic, “A sub 1 load-pull quarter wave prematching network based on a two-tier TRL
calibration,” IEEE Trans. Antennas Propag., 47, no. 2, pp. 389–391, Feb. 1999.
[18] C. Tsironis, “Prematched programmable tuners for very high VSWR testing,” in IEEE μAPS
Microwave Application & Product Seminars, Anaheim, CA, June 1999.
[19] R. Stancliff and D. Poulin, “Harmonic load-pull,” in IEEE International Microwave
Symposium Digest, Orlando, FL, Apr. 1979, pp. 185–187.
[20] Focus Microwaves, “An affordable harmonic load pull setup,” Microwave J., pp. 180–182,
Oct. 1998.
[21] V. Camarchia, V. Teppati, S. Corbellini, and M. Pirola, “Microwave measurements part II:
nonlinear measurements,” IEEE Instrum. Meas. Mag., 10, no. 3, pp. 34–39, June 2007.
[22] P. Colantonio, F. Giannini, and E. Limiti, “Nonlinear approaches to the design of microwave
power amplifiers,” International Journal of RF and Microwave Computer-Aided Engineer-
ing, 14, pp. 493–506, Nov. 2004.
[23] Verspecht-Teyssier-DeGroote s.a.s. [Online]. Available: http://www.vtd-rf.com.
[24] Antwerta Microwave [Online]. Available: http://www.anteverta-mw.com.
[25] NMDG NV [Online]. Available: http://www.nmdg.be.
[26] Mesuro Ltd. [Online]. Available: http://www.mesuro.com.
[27] High Frequency Engineering [Online]. Available: http://www.hfemicro.com.
[28] Y. Takayama, “A new load-pull characterization method for microwave power transistor,” in
IEEE MTT-S Intl. Microwave Symp. Dig., Cherry Hill, NJ, June 1976, pp. 218–220.
[29] J. Benedikt, R. Gaddi, P. J. Tasker, and M. Goss, “High-power time domain measurement
system with active harmonic load-pull for high-efficiency base-station amplifier design,”
IEEE Trans. Microw. Theory Tech., 48, no. 12, pp. 2617–2624, Dec. 2000.
[30] F. Ghannouchi, F. Larose, and R. Bosisio, “A new multiharmonic loading method for large-
signal microwave and millimeter-wave transistor characterization,” IEEE Trans. Microw.
Theory Tech., 39, pp. 986–992, June 1991.
[31] J. Nebus and J. V. P. Bouysee, J. M. Coupat, “An active load-pull setup for the large signal
characterization of highly mismatched microwave power transistors,” in Instrumentation and
Measurement Technology Conference, 18–20 May 1993, pp. 2–5.
[32] G. Bava, U. Pisani, and V. Pozzolo, “Active load technique for load-pull characterization at
microwave frequencies,” Electronic Lett., 18, pp. 178–179, Feb. 1982.
Trans. Instr. Meas., 57, no. 11, pp. 2640–2646, Nov. 2008.
[34] A. Ferrero, “Active load or source impedance synthesis apparatus for measurement test set
of microwave components and systems,” U.S. Patent 6 509 743, 12 June 2000.
[35] J.-M. Nebus, J.-M. Coupat, and J.-P. Villotte, “A novel, accurate load-pull setup allowing
the characterization of highly mismatched power transistors,” IEEE Trans. Microw. Theory
Tech., MTT-42, no. 2, pp. 327–332, Feb. 1994.
[36] Z. Aboush, J.Lees, J. Benedikt, and P. Tasker, “Active harmonic load-pull system for charac-
terizing highly mismatched high power transistors,” in IEEE MTT-S Symp. Dig., Long Beach,
CA, 12–17 June 2005, pp. 1311–1314.
[37] F. Ghannouchi, M. Hashmi, S. Bensmida, and M. Helaoui, “Loop-enhanced passive source
and load-pull technique for high reflection factor synthesis,” IEEE Trans. Microw. Theory
Tech., MTT-58, no. 11, pp. 2952–2959, Nov. 2010.
[38] A. Ferrero and V. Teppati, “A novel active differential/common-mode load for true mixed-
mode load-pull systems,” in International Microwave Symposium, San Francisco, CA, USA,
June 2006, pp. 1456–1459.
[39] M. P. van der Heijden, D. Hartskeerl, I. Volokhine, V. Teppati, and A. Ferrero, “Large-signal
characterization of an 870 MHz inverse class-F crosscoupled push-pull PA using active
mixed-mode load-pull,” in Radio Frequency Integrated Circuits (RFIC) Symposium, 2006
IEEE, San Francisco CA, USA, June 2006, pp. 389–392.
[40] V. Teppati, M. Garelli, V. Camarchia, A. Ferrero, and U. Pisani, “Advanced load-pull tech-
niques – from single-ended to multiport/differential measurement systems,” in European
Microwave Week Workshop WSW6, Munchen, Oct. 2007, pp. 105–120.
[41] V. Teppati, A. Ferrero, M. Garelli, and S. Bonino, “A comprehensive mixed-mode time
domain load- and source-pull measurement system,” IEEE Trans. Instrum. Meas., 59, no. 3,
pp. 616–622, Mar. 2010.
[42] V. Teppati, M. Goano, A. Ferrero, V. Niculae, A. Olivieri, and G. Ghione, “Broad-band
coaxial directional couplers for high-power applications,” IEEE Trans. Microw. Theory Tech.,
MTT-51, no. 3, pp. 994–997, Mar. 2003.
[43] V. Teppati and A. Ferrero, “A new class of non-uniform, broadband, non-symmetrical
rectangular coaxial-to-microstrip directional couplers for high power applications,” IEEE
Microwave and Wireless Components Letters, 13, no. 4, pp. 152–154, Apr. 2003.
[44] V. Teppati and C. R. Bolognesi, “Evaluation and reduction of calibration residual uncertainty
in load-pull measurements at millimeter-wave frequencies,” IEEE Trans. Instrum. Meas., 61,
no. 3, pp. 817–822, Mar. 2012.
[45] R. Tucker and P. Bradley, “Computer-aided error correction of large-signal load pull
measurements,” IEEE Trans. Microw. Theory Tech., MTT-32, pp. 296–300, Mar. 1984.
[46] I. Hecht, “Improved error-correction technique for large-signal load-pull measurements,”
IEEE Trans. Microw. Theory Tech., MTT-35, pp. 1060–1062, Nov. 1987.
[47] A. Ferrero and U. Pisani, “An improved calibration technique for on-wafer large-signal
transistor characterization,” IEEE Trans. Instrum. Meas., IM-47, pp. 360–364, Apr. 1993.
[48] A. Ferrero, U. Pisani, and K. Kerwin, “A new implementation of a multiport automatic
network analyzer,” IEEE Trans. Microw. Theory Tech., MTT-40, pp. 2078–2085, Nov. 1992.
[49] F. Larose, F. Ghannouchi, and R. Bosisio, “A new multi-harmonic load-pull method for
nonlinear device characterization and modeling,” in IEEE MTT-S Intl. Microwave Symp.
Dig., 1, May 1990, pp. 443–446.
[50] D. Barataud, C. Arnaud, B. Thibaud, M. Campovecchio, J. Nebus, and J. Villotte, “Measure-
ments of time domain voltage/current waveforms at RF and microwave frequencies based on
the use of a vector network analyzer for characterization of nonlinear devices — application
to high-efficiency power amplifiers and frequency-multipliers optimization,” IEEE Trans.

Instrum. Meas., IM-47, pp. 1259–1264, Oct. 1998.
[51] D. Barataud, M. Campovecchio, and J.-M. Nebus, “Optimum design of very high-efficiency
microwave power amplifiers based on time domain harmonic load-pull measurements,” IEEE
Trans. Microw. Theory Tech., 49, no. 6, pp. 1107–1112, June 2001.
[52] R. Gaddi, J. Pla, J. Benedikt, and P. Tasker, “LDMOS electro-thermal model validation from
large-signal time domain measurements,” IEEE MTT-S Symp. Dig., 1, pp. 399–402, 2001.
[53] O. Jardel, F. D. Groote, T. Reveyrand, J.-C. Jacquet, C. Charbonniaud, J.-P. Teyssier, D. Flo-
riot, and R. Quere, “An electrothermal model for AlGaN/GaN power HEMTs including
trapping effects to improve large-signal simulation results on high VSWR,” IEEE Trans.
Microw. Theory Tech., MTT-55, no. 12, pp. 2660–2669, Dec. 2007.
[54] D. Williams, J. Leckey, and P. Tasker, “A study of the effect of envelope impedance on
intermodulation asymmetry using a two-tone time domain measurement system,” IEEE MTT-
S Symp. Dig., 3, pp. 1841–1844, June 2002.
[55] Hewlett-Packard Co., The microwave transition analyzer: a versatile measurement set for
bench and test. Rohnert Park, CA: HP Product Note 70820-1, 1991.
[56] J. Verspecht, F. Verbeyst, and M. V. Bossche, “Large-signal network analysis: Going beyond
S-parameters,” in URSI 2002, Maastricht, Netherlands, Aug. 2002.
[57] J. Verspecht, “Large-signal network analysis,” Microwave Magazine, 6, no. 4, pp. 82–92,
Dec. 2005.
[58] V. Teppati, A. Ferrero, V. Camarchia, A. Neri, and M. Pirola, “Microwave measurements part
III: Advanced nonlinear measurements,” IEEE Instrum. Meas. Mag., 11, no. 6, pp. 17–22,
Dec. 2008.
[59] U. Lott, “Measurement of magnitude and phase of harmonic generated in nonlinear
microwave two-ports,” IEEE Trans. Microw. Theory Tech., MTT-37, no. 10, pp. 1506–1510,
Oct. 1989.
[60] L. Gommé and Y. Rolain, “Design and characterisation of an RF pulse train generator for
large-signal analysis,” Measurement Science and Technology, accepted for publication, 2009.
[61] J. Verspecht, “Broadband sampling oscilloscope characterization with the “nose-to-nose”
calibration procedure: a theoretical and practical analysis,” IEEE Trans. Instrum. Meas., 44,
no. 6, pp. 991–997, Dec. 1995.
[62] P. Blockley, D. Gunyan, and J. Scott, “Mixer-based, vector-corrected, vector signal/network
analyzer offering 300kHz–20GHz bandwidth and traceable phase response,” in IEEE
International Microwave Symposium Digest, Long Beach, CA, June 2005.
[63] V. Teppati, V. Camarchia, A. Ferrero, M. Pirola, and S. D. Guerrieri, “A comprehensive GaN
HEMT characterization for power amplifier design,” in IEEE Topical Workshop on Power
Amplifiers for Wireless Communications, San Diego, USA, 16–17 Jan. 2006, pp. 1–2.
[64] P. Colantonio, F. Giannini, E. Limiti, and V. Teppati, “An approach to harmonic load- and
source-pull measurements for high-efficiency pa design,” IEEE Trans. Microw. Theory Tech.,
MTT-52, no. 1, pp. 191–198, Jan. 2004.
[65] V. Adamian and A. Uhlir, “A novel procedure for receiver noise characterization,” IEEE
Trans. Instr. Meas., IM-22, pp. 181–182, June 1973.
[66] G. Caruso and M. Sannino, “Determination of microwave two-port noise parameters through
computer-aided frequency-conversion techniques,” IEEE Trans. Microw. Theory Tech., MTT-
27, pp. 779–783, Sept. 1979.
[67] A. Davidson, B. Leake, and E. Strid, “Accuracy improvements in microwave noise parameter
measurements,” IEEE Trans. Microw. Theory Tech., MTT-37, pp. 1973–1978, Dec. 1979.
[68] D. Le and F. Ghannouchi, “Noise measurements of microwave transistor using an uncalibrated

mechanical stub tuner and a built-in reverse six-port reflectometer,” IEEE Trans. Instrum.
Meas., IM-44, pp. 847–852, Aug. 1995.
[69] P. Colantonio, F. Giannini, E. Limiti, and V. Teppati, “An approach to harmonic load- and
source-pull measurements for high-efficiency PA design,” IEEE Trans. Microw. Theory Tech.,
MTT-52, no. 1, pp. 191–198, Jan. 2004.
[70] B. Hughes and P. Tasker, “Improvements to on-wafer noise parameter measurements,” in
36th ARFTG Conf. Dig., Monterrey, CA, Nov. 1990, pp. 16–25.
[71] D. Le and F. Ghannouchi, “Source-pull measurements using reverse six-port reflectomenters
with application to MESFET mixer design,” IEEE Trans. Microw. Theory Tech., MTT-42,
pp. 1589–1595, Sept. 1994.
[72] G. Berghoff, E. Bergeault, B. Huyart, and L. Jallet, “Automated characterization of HF power
transistor by source-pull and multiharmonic load-pull measurements based on six-port
techniques,” IEEE Trans. Microw. Theory Tech., MTT-46, pp. 2068–2073, Dec. 1998.
[73] G. Madonna, M. Pirola, A. Ferrero, and U. Pisani, “Testing microwave devices under different
source impedances: a novel technique for on-line measurement of source and device reflection
coefficents,” in IMTC/99 Conf. Proc., Venezia, Italy, May 1999, pp. 130–133.
[74] G. Madonna and A. Ferrero, “Simple technique for source reflection coefficient measurement
while characterizing active devices,” in 53nd ARFTG Conf. Dig., Anaheim, CA, June 1999,
pp. 104–106.
14 Broadband large signal
measurements for linearity
optimization
Marco Spirito and Mauro Marchetti
14.1 Introduction
The recent introduction of high-performance modulation schemes (e.g. (W)-CDMA and

OFDM) provides the capability of realizing high-data rate communication links (i.e. up
to 100 Mbps from 20 MHz spectrum)[1]. The broadband nature of those signals together
with the large difference between the peak and the average power across the modulation
bandwidth requires a large number of spectral components to accurately represent the
signal statistics. The modulated signal should be amplified by the transmitting chain
without loss of information (i.e. low EVM) and with little out-of-band-distortion to avoid
interference with adjacent transmitting channels. The quality of the communication link
can be translated into specification parameters of the active element of the transmission
chain, in the case of the PA, through the device IM 3 and ACPR level. In general, it is
very difficult to link the technology parameters of an active device directly to its linearity
performance, since the linearity achieved for a given PA is the result of its interaction
with the surrounding circuitry. For this reason, most attempts to improve the linearity of
PAs are currently made at the circuit level.
In order to properly compare different technologies (e.g. SiGe and III-V) or device
technology generations, one must provide the optimum loading conditions, at funda-
mental, harmonic, and baseband frequencies, to the active device during the evaluation
phase, ideally under the same driving signal of the final application. This measurement
task is intrinsically complex since the broadband nature of the signal of interest conflicts
with the narrow-band nature of the currently employed high dynamic range receivers (i.e.
narrow IF bandwidth super-heterodyne receivers). Moreover, the capability of conven-
tional load-pull setups (active and passive) allows the accurate control of the reflection
coefficient only in a narrow frequency bandwidth. In this chapter we review the active
load-pull techniques and test-benches that have been developed in the last years [2–4]
to facilitate large signal measurements with digitally modulated (standard compliant)
broadband signals. These test-benches allow optimization of the device large-signal per-
formance, e.g. saturated output power, PAE as well as its linearity performance, such as
IM 3 and ACPR, providing the capability to link technology generations with linearity
performance.top
The chapter is organized as follows: first, the problem statement of electrical delay
in load-pull systems is introduced. Second, the architecture of two broadband load-pull
test benches (i.e. closed-loop and mixed-signal open-loop) and their various building
Broadband large signal measurements for linearity optimization 385
blocks are analyzed. Then the power and linearity requirements for the amplifiers used
in the active loads are analyzed. In conclusion, some experimental results presenting a
linearity-optimized SiGe PA and a high power LDMOS are given.
14.2 Electrical delay in load-pull systems
When testing the large-signal performance of devices or circuits with modulated signals,
it is important to control the reflection coefficient offered to the DUT, not only within the
modulation bandwidth (i.e. at the fundamental frequency), but also within the frequency
bands where the nonlinear device generates power, e.g. adjacent channels, harmonics
Tuning element
Electrical delay
Tuner
Passive
Coaxial cable
ΓS/L
DUT reference
plane
Active
(a)
–10
–20
Phase delay [deg]
–30
–40
–50
–60
Freq 1 MHz
–70 Freq 3 MHz
–80 Freq 5 MHz

Freq 25 MHz
–90
0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00
Length [m]
(b)
Fig. 14.1 (a) Phase delay caused by the electrical lengths of the cables present between the tuning element
and the DUT (b) phase rotation of the reflection coefficient at the DUT reference plane as a
function of cable length and signal bandwidth.
386 Marco Spirito and Mauro Marchetti
and baseband frequencies [5, 3]. The reason for this stringent requirement comes from
the fact that load-pull can be an accurate predictive tool of the device performance in the
application environment (i.e. IC or board level), only when the impedance provided by
the measurement setup closely resembles those offered to the DUT in the final operating
conditions.
In the final circuit implementations, matching networks are placed in close proximity to
the DUT with distances in the order of a few hundred microns for ICs to a few millimeters
for PCB assemblies. These distances are negligible, in terms of electrical lengths, with
respect to the modulation bandwidths of the signals employed today, e.g. up to hundreds
of MHz. However, in all conventional load-pull setups the actual tuning element (i.e.
tuner or active loop) is always located at some distance from the DUT (Figure 14.1(a)),
which is much larger than for any practical matching network. This distance, as well
as any physical length within the tuning element itself (e.g. the length of the active
feedback loop, or the position of the probe inside the mechanical tuner), yields very
large electrical delays, causing rapid phase changes of the reflection coefficients versus
frequency. In Figure 14.1(b) the phase variation of the reflection coefficient versus the
length of the coaxial cable is shown for four modulation bandwidths (i.e. 1, 3, 5, and
25 MHz). Typical values for these phase fluctuations start from about a few degrees
per MHz for a fundamental passive mechanical tuner and tend to be much higher for
active-loop systems.
Consider now that a three carrier W-CDMA signal with adjacent channels provides a
total bandwidth of 25 MHz. This results in a reflection coefficient, offered by a passive
tuner, varying more than 50◦ in phase over the signal bandwidth (assuming 2◦ /MHz phase
variation). Such large phase variation, which translates in non-realistic impedance condi-
tions provided to the device, will cause measurement errors, when compared to the device
response in its final application environment, such as IM 3 asymmetry, spectral re-growth,
and PAE degradation [4]. These considerations clearly highlight the difficulties as well
as the need of realizing accurate wideband load-pull measurement setups.
14.3 Broadband load-pull architectures
Test-benches that can make load-pull measurements with realistic broadband signals
require specific choices, mainly for the measurement approach and the load control. In
this section we analyze the requirements of the detection scheme, the RF front-end, the
baseband control, and the high-frequency active loads, needed to implement accurate
large-signal broadband systems.
14.3.1 Detection scheme

The detection schemes for high-frequency signals can be divided into two major
categories: direct acquisition (i.e. real-time or sub-sampling oscilloscope based) and
down-converted acquisition (i.e. heterodyne mixer or sampler-based architectures).
While real-time oscilloscopes are starting to be available with high sampling rates and
high-frequency front-end capabilities [6], i.e. above 26 GHz, the 8-bit vertical resolution
in combination with the higher noise level (due to the large bandwidth where the noise
is received) make such receivers unsuitable for device linearity characterization. Sub-
sampling oscilloscopes, at the other end of the scale, provide large vertical resolution for
the low-frequency ADC (i.e. better than 14 bits) with high input frequency capabilities
(i.e. higher than 60 GHz) [7]. Nevertheless, the low sampling speed of the sampler (i.e.
below 1 MHz) combined with the higher noise bandwidth, requires a long measuring
time and large averaging to reach a dynamic range in the order of 70 dB. Moreover,
when considering realistic modulated signals (i.e. more than 20 000 frequency bins in a
4 MHz span), the limiting factor is the memory depth of the instrument. For these rea-
sons, large-signal load-pull test-benches employing this detection scheme have usually
been confined to multitone excitations with a limited number of tones [8]. The rest of
this section is devoted to a more in-depth analysis of heterodyne architectures that allow
the detection of standard compliant modulated signals.
These large-signal test-benches rely on a tuned-receiver architecture to provide high
signal sensitivity and large dynamic ranges. In such detection schemes, the high-
frequency signals are translated to lower intermediate frequencies. Narrow-band filtering
of the down-converted signal is employed to reduce noise and increase the detection
dynamic range. When employing broadband signals, narrow-band filtering is avoided
and the full ADC bandwidth or an external spectrum analyzer is used to acquire the mod-
ulated signal. The signal down-conversion to an intermediate frequency can be achieved
with either high-frequency mixers or sub-sampling-based systems.
When employing high-frequency mixers, the LO is swept over the signal harmonics
down-converting bandwidths up to few GHz for commercially available mixers, with an
input third-order intercept point above 20 dBm [9], [10]. Such components provide a high
system linearity so that the intermodulation distortion products of linear amplifiers can
be properly detected. Note that the harmonic distortion products should be at least 18 dB
below the harmonic distortion level of the DUT to guarantee low linearity measurement
uncertainties [11]. Recalling that the relation between the IM 3 products expressed in
dBc (IM 3 ) and the OIP3 is given by [12]:
IM 3
OIP3 = Pout + , (14.1)
2
the mixer OIP3 needs to be at least 9 dB higher than the device under test OIP3 .
Moreover, when used in combination with wideband ADCs [13] these detection
schemes allow the down-conversion and sampling of a large portion of the spectrum
around the carrier frequency and the harmonics (i.e. three carrier W-CDMA signals with
adjacent and alternate channels providing a total bandwidth of 35 MHz).
Sub-sampling based systems employ a sampler down-conversion, driven by a precise
low-frequency signal (typically 10–25 MHz) phase locked to the 10 MHz crystal that
provides the frequency reference to the signal driving the DUT. The sampling pulses are
created by a step recovery diode [14] or a nonlinear transmission line [15] and allow
the down-conversion of the entire system RF bandwidth to the ADC acquisition band.
When considering modulated signals, the frequency of the LO signal driving the samplers
must be chosen properly to avoid overlapping of different tones on the same baseband
frequency. The minimum frequency windows, and hence LO frequency, to properly
down-convert a modulated signal with NSSB single-sideband tones and considering NH
harmonics of interest, is given by [16].
fBW = fLOmin = (2NH + 1) (2NSSB + 1) fRES , (14.2)
where fRES is the required resolution frequency, which is the inverse of the required
measurement time. Consider for example the measurement of a modulated signal with
the following test conditions: a standard compliant W-CDMA signal including the upper
and lower adjacent channels, resulting in a 15 MHz wide signal (i.e. 5000 tones per
channel assuming a 1 KHz spacing between tones) and a measurement time of 0.5 msec
(i.e. translating into a 500 Hz frequency resolution).
When we substitute these values in (14.2), we obtain:
NH =3
NSSB =3 × 2500
(14.3)
fRES =500H z
fBW =fLOmin ≈ 52.5MH z.
Note the 0.5 msec window chosen here is given as an example representing a reasonable
measurement time to allow for sufficient averaging to reach the required measurement
dynamic range. These results indicate that when choosing realistic measurement times
and memory depths (i.e. fRES not too small) typical sampler architectures (i.e. fLO
10–25 MHz) are not indicated for measurements of standard compliant modulated
signals. For this reason, in the rest of the chapter we will only consider mixer-based
architectures.
14.3.2 RF front-end
When measuring devices are matched for optimum linearity, it is worth noting that there
will be a dramatic difference between power levels in the fundamental and harmonic
frequency bands. This is due to the high OIP2 and OIP3 of such devices.
Using a single mixer to down-convert the entire frequency band poses severe
challenges to an accurate detection for the following reasons:
• using an attenuator to optimize the power level at the mixer RF port will increase
the system noise floor (see Figure 14.2(a)), preventing the measurement of the “low-
power” harmonic components;
• when no attenuation is used, the nonlinearities of the mixer itself, which is over-
driven at the fundamental tone, prevent the correct measurement of these “low-power”
harmonic components (see Figure 14.2(b)).
40
20
PTone1(f1)
0
PTone1(f1), HT2, IM3 [dBm]
IM3
–20 HT2 (2f1)
–40
Fundamental mixer noise floor
–60
–80
–100
Harmonic mixer noise floor
–120
–140
5.0 10.0 15.0 20.0 25.0 30.0
Pavl @ DUT [dBm]
(a)
25
PTone1(f1)
0 HT2 (2f1) with filter

HT2 (2f1) without filter
PTone1(f1), HT2, [dBm]
–25
–50 Generated by mixer

non-linearities
–75
–100
–125
Harmonic mixer noise floor
–50 –40 –30 –20 –10 0
Pavl @ DUT [dBm]
(b)
Fig. 14.2 Two-tone measurement on calibration Thru (fc = 2.14 GHz, f = 0.2 MHz), showing the
corrected measured power at the fundamental; (a) amplitude of fundamental, IM 3 and H T2
components measured by the HP 8510 (high power levels), (b) 2nd -harmonic, with and without
high-pass filter (low power levels). © [2004] IEEE. Reprinted, with permission, from [17].
To overcome this restriction, a multi-branch mixer configuration was first introduced

in [18]. Using this approach each of the waves (i.e. a and b) coupled out by the input
and output reflectometers can be routed to two mixers; see Figure 14.3. A fundamental
mixer, used to measure the fundamental frequency band, employing an attenuator to
maximize the system dynamic range for the high power (i.e. fundamental) tones. All
the higher harmonics are measured by the harmonic mixer, using a high-pass filter in
the signal path. Figure 14.2 (b) [17] presents the calibrated data for a two-tone power
Coupler section
Fundamental
Fundamental
Harmonic
Harmonic
a1 f 0 a1 2f 0 b1 f 0 b1 2f 0
To LO drive
Fig. 14.3 Block scheme describing the implementation of the multi-branch mixer. © [2006] IEEE.
Reprinted, with permission, adapted from [3].
sweep on a calibration thru using a high-pass filter (pass band 3.0–26.5 GHz) providing a
rejection of 40 dB below 3 GHz. The filter blocks out the fundamental signal, avoiding the
generation of higher harmonics by the harmonic mixer (where no attenuation is present)
and significantly increasing the dynamic range for higher-frequency components.
The choice of a high IP3 mixer in combination with a multi-branch mixer configu-
ration provides a broadband capability to accurately measure the reflection coefficients
offered to the DUT in the various control bands. Two large signal broadband setups
were presented in [3] and [4] and a simplified block diagram, illustrating the common
components in the two test-benches is shown in Figure 14.4.
The depicted system is composed of high-frequency signal sources (i.e. RF to drive
the DUT and LO to drive the down-converting mixers). The system input RF section is
based on three reflectometers configured to measure simultaneously the input and source
reflection coefficient [19]. The traveling waves are detected in a real-time fashion using
a traditional four-coupler configuration [20]. The coupled a− and b − waves are down-
converted to a lower IF to be digitized. The system presented in [3] employs the HP 8510
receiver unit to process both the high-frequency and low-frequency (baseband) signal
components. All the signals have to be down-converted (RF) or up-converted (baseband)
to the first IF frequency of the instrument (i.e. 20 MHz). This is done through RF and
baseband (BB) mixers. Note in Figure 14.4 the baseband mixers are not explicitly shown.
The LO synthesizer provides the required down-converting or up-converting signal.
Mechanical switches are used to route these IF signals to the HP 8510 mainframe.
The system described in [4] employs wideband AD converters (100 MS/s sampling
frequency) to digitize the IF signals. This architecture enables the direct measurement
of the device reflection coefficients over a wide bandwidth of 40 MHz in a single data
acquisition. With this hardware, wider bandwidths, up to 120 MHz, and the frequency
content in the harmonic bands can be measured by stepping the frequency of the LO that
AD Converters
10 MHz ref.
aref f0 LO
Downconv.
a1 a1 a1 b1 b1 b1 a2 a2 a2 b2 b2 b2
BB f0 2f0 BB f0 2f0 BB f0 2f0 BB f0 2f0
aref f0 LO mixers
RF Source
LO mixers DC Supply DC Supply
a1 BB BB BB
Impedance Impedance a2 BB
Ref. channel
b1 BB control control b2 BB
Input section DUT Output section

Active loads Active loads
f0, 2f0, etc. f0, 2f0, etc.
IN Cal.
reference
planes
Harmonic
Harmonic
Harmonic
Harmonic
Fundamental
Fundamental
Fundamental
Fundamental
a1 f 0 a1 2f 0 b1 f 0 b1 2f 0 b2 f 0 b2 2f 0 a2 f 0 a2 2f 0
LO mixers
LO mixers
Fig. 14.4 Load-pull architecture for linearity optimization under broadband stimuli. © [2006] IEEE. Reprinted, with permission, adapted from [3].
drives the down-converting mixers. The large bandwidth of the AD receiver allows the
baseband components to be measured directly without using up-converting mixers. In
this configuration, IF electronic switches route the different signals to the receiver.
14.3.3 System calibration

Broadband system calibrations follow those of traditional large-signal setups and are,
usually, a combination of the techniques described in [21] and [22], performed at all the
frequency tones of interest at the fundamental and harmonic bands. Interpolation can
be used when standard compliant modulated signals (composed by a large number of
frequency bins) are employed.
14.4 Broadband loads
To facilitate measurements with wideband excitations it is important to optimize the

load performance in terms of linearity, frequency response (i.e. electrical delay within
the signal bandwidth), and power-handling capabilities. Traditionally, the tuning of the
load and source impedances has been implemented using passive mechanical tuners. In
order to minimize the phase delay (see Figure 14.1) and maximize the reflection coeffi-
cient that can be provided to the DUT, tuners are mounted directly at the DUT interface
(i.e. connectors or wafer probes). This requires calibration data-files to retrieve the tuner
S-parameters while real-time reading of the provided reflection coefficient can only be
achieved when low-loss bi-directional couplers are placed between the tuner and DUT.
When aiming for linearity characterization and optimization [5], the second harmonic
load and source impedance also need to be properly controlled. The path insertion losses
at the higher harmonics (i.e. diplexer, coupler sections, and eventually wafer-probes)
call for an active load implementation when targeting linearity characterization and opti-
mization. In the rest of the section we describe in detail the architecture and performance
of closed-loop and mixed-signal open-loop active loads.
14.4.1 Closed-loop active loads

Closed-loop topologies are shown in Chapter 13. The use of an off-the-shelf phase-
shifter, variable attenuator, and narrow-band filter often comes at the price of a large
electrical delay provided by the loop. Moreover, since the physical distance of the active
load to the DUT should also be minimized, compact active-loop implementations are
required. A more compact active loop (i.e. smaller electrical length) can be achieved by
integrating the phase-shifter and variable attenuator in a unique block by employing 90
degrees coupler-based IQ modulators, as shown in Figure 14.5 (a) [23]. The impedances
that can be achieved by the IQ modulator topology are shown in 14.5 (b), presenting
the provided reflection coefficients over the IQ plane. At RF frequencies (i.e. below 10
GHz) coupler-based IQ modulators can be realized using low-cost PCB technologies and
PIN diodes. In Figure 14.6 [3], a board-level implementation of such an IQ modulator,
operating at 2.14 GHz, is shown. The in-phase and quadrature resistances are tuned via
RI
-3dB
Isolated -3dB
Input
Input -3dB
Output
Isolated -3dB I
Input
Isolated -3dB
-3dB
RQ
(a) (b)
Fig. 14.5 (a) Simplified block diagram of an IQ modulator based on quadrature hybrids, (b) simulation of
the IQ modulator constellation showing the I and Q signal required for a given phase delay and
amplitude attenuation (arrow).
current-controlled PIN diodes. The operation of the modulator can be summarized as

follows. The first 90 degrees coupler splits the signal into two orthogonal components
(i.e. I and Q). The two signals are now amplitude-adjusted by changing the variable
resistances (i.e. the PIN diodes) at the output ports of the two quadrature power splitters.
As an example, when the terminating resistances are set to 50 Ohm no power is reflected
and the I or Q signal provided at the input of the power combiner is zero; when the
terminating resistances are low (high), the signals are strongly reflected and provided to
the power combiner with phase inversion (no phase inversion). The final combination
between the I and Q components achieves a 360-degree phase-shift and variable attenua-
tion. A fixed attenuator at the input of the IQ board can be used to optimize the maximum
power at the input of the modulator in order not to overdrive the diodes and thus avoid
linearity degradation. During calibration, the IQ modulator can then be stepped in dif-
ferent positions in order to characterize the loop response and allow the user to request
a specific reflection coefficient at the DUT reference plane.
Note that since the 90 degree couplers are implemented using transmission lines,
higher-frequency implementations will be smaller (i.e. 61 × 65 mm at 4.28 GHz versus
96 × 101 mm at 2.14 GHz), which partially compensates the increased electrical delay
provided by the fixed dimensions interconnects to the DUT.
As mentioned before, to minimize the overall ∠ variation with frequency, the electri-
cal length of the various sections of the active load-pull system must be kept as short as
possible. In this respect, the small dimensions of the coupler-based IQ modulators allow
D D
Pin diodes
Input
Output
Pin diodes
D D
Fig. 14.6 Board-level implementation of IQ modulator using pin diodes to control electronically the
in-phase and quadrature resistance. Board dimension is 96 x 101 mm (WxH). © [2006] IEEE.
<Γfund variation
3 MHz span
14.55°
<Γ2nd variation
3 MHz span
13.8°
Fig. 14.7 ∠ variation with frequency at the DUT reference plane for active loop topology at the output
(frequency span 3 MHz). © [2006] IEEE. Reprinted, with permission, from [3].
for very compact active loops. The integration of a compact coupler-based IQ modula-
tor placed in close proximity to the DUT [3] achieves the ∠ variation with frequency
shown in Figure 14.7. The impacts of the individual contributions to the ∠ variation
with frequency of the various system components are summarized in Table 14.1.
Table 14.1 Active Loop ∠ variation
Active Loop [◦ /MHz]
input/output
probe + cable 0.9 / 0.9

coupler section 1.2 / 1.0
diplexer 0.4 / 0.4
active loop f0 2.55 / 2.55
active loop 2f0 2.3 / 2.3
Total f0 5.05 / 4.85

Total 2f0 4.8 / 4.6
AWG f0
I f0 Load
AWG 2f0
2f0 Load
10 MHz ref.
I Active
load
Q
AWG Nf0
I Nf0 Load
Fig. 14.8 Simplified block diagram of the wideband active loads with phase-coherent frequency
up-conversion. © [2008] IEEE. Reprinted, with permission, from [4].
14.4.2 Mixed-signal active loads

In Chapter 13, wideband open-loop techniques were briefly introduced, in this section
we discuss in details an open-loop concept based on signals generated by baseband
AWGs and up-converted, to the desired RF frequency, using IQ mixers [4]. Although
conceptually simple, this method requires high speed and high dynamic range DACs to
generate the a-waves with high dynamic range (i.e. 70 dB spurious free dynamic range)
and with a wide bandwidth to cover the needs of modern communication signals (i.e.
larger than 100 MHz to cover a three carrier W-CDMA signal with adjacent and alternate
channels at the third harmonic). Moreover, the signal injection at the various ports needs
to be phase-coherent at both the RF frequencies as well as at the baseband.
The mixed-signal architecture described in [4, 24] integrates all the AWGs in a PCI
extensions for instrumentation (PXI) express platform (enabling 5.5 GByte/s transfer
capacity to the PXI backplane [13]), sharing the same timebase, making them fully
synchronized. The source signal and all injection signals needed to synthesize the
user-defined reflection coefficients at the DUT reference planes, are originating from
fully synchronized (400 MS/s) arbitrary waveform generators, shown in Figure 14.8.
as
a1,n S21# b2,n
a1 inject
Software ΓS,n(fn) S11# ΓL,n(fn) Software

iteration iteration
DUT a2 inject
b1,n a2,n
Si,j# (i,j = 1,2): large signal param. ΓX,n(fn):User defined

ax,n, bx,n : measured waves freq.dependent
reflection coeff.
Driving signal
Device-generated wave
System injected wave
Fig. 14.9 Principle of the mixed-signal open-loop active load-pull approach. When the nonlinear DUT is
excited with a user-defined modulated signal as , it generates signals in the baseband,
fundamental, and higher harmonic frequency bands. By measuring the device-generated waves
(b1,n and b2,n ), as well as the incident waves, the waves to be injected are estimated by
successive iterations. When the required reflection coefficient versus frequency (at every
controlled band) is achieved, the iteration has converged and the large signal parameters (e.g.
PAE, Pout , IM 3 , IM 5 ) are measured. © [2008] IEEE. Reprinted, with permission, from [4].
In order that the source and all the injected signals (fundamental and harmonics) are
phase coherent, or in other words, are perfectly locked in phase and exhibit no phase
drift among each other, IQ up-conversion is preferred over digital IF techniques [25]. By
employing IQ up-conversion, a single synthesizer can be used to drive the local oscilla-
tor port of the IQ mixers (Figure 14.8). Frequency multipliers are used to obtain the LO
signals driving the IQ mixers at the higher harmonic bands (i.e. a × N multiplier is used
for generating the N th harmonic). This approach guarantees that the active loads and the
driving signal are fully phase coherent.
The principle of operation of the mixed-signal broadband architecture is shown in
Figure 14.9. As in the classical open-loop approach, in the mixed-signal case only the
content of the driving waveform (as ) is known prior to the acquisition. All other injection
signals (a1inject,n and a2inject,n ), containing all the frequency components of the signal of
interest, need to be created from scratch, and any desired reflection coefficient behavior
vs. frequency can be generated. This final result is obtained by iteratively adjusting the
amplitude and phase of the injected waveforms independently at each frequency band
of interest. To obtain a specific reflection coefficient, an injection signal, based on the
linear relation shown in (14.4), is required at all the frequency components of interest.
ax,n (fn ) = bx,n (fn ) · x,n (fn ), (14.4)

where ax,n and bx,n are the incident and reflected waves at port x and harmonic index n,
while x,n represents the user defined reflection coefficient versus frequency at port x
and harmonic index n. By monitoring the deviation of the measured reflection coefficient
from the desired one for each frequency bin, the injected wave can be optimized and
is found by subsequent iterations. The error checking (i.e. the distance between the
obtained value and the required reflection coefficient value) and the optimization are
done in the frequency domain, while the actual injection signals are loaded and acquired
in the time domain. The open-loop approach guarantees that no sustained oscillations
can occur. In practice, when the user-defined reflection coefficients (input or output)
force the device to operate in an instable region, the system simply fails to converge. In
all other situations the optimization algorithm converges as normal. Finally, computer
controlled attenuators and high power amplifiers are placed in the signal path, after the
IQ up-converters, in order to level the power of the injection signals. This approach
makes full use of the maximum dynamic range of the AWGs at all times, which is an
essential step to meet (in generation) the spectral requirements of modern wideband
communication signals.
The IQ up-conversion approach requires, compared to other known signal genera-
tion techniques, relatively limited length of the data records (i.e. limited to the DAC
up-converted bands) yielding a significant speed advantage, when standard models of
complex modulated signals are employed in the measurements.
Figure 14.10 shows the functionality of the setup presented in [4], where a test signal
composed of 161 sinusoidal tones in the bandwidth between 2060 MHz and 2220 MHz
is fed to a calibration thru, while the output active load is set to provide an open condition
over the whole 160 MHz bandwidth. The measured reflection coefficient at the output
reference plane of the thru is plotted as a function of frequency. Figure 14.10 provides
clear evidence that, using an open-loop mixed-signal technique, the desired reflection
coefficient (L = 1) can be set, without any phase delay or amplitude unbalance, over a
wide modulation bandwidth (i.e. 160 MHz).
Wideband Signal Generation

When working with complex modulated signals a good place to start is the modulation
test standard [26, 27]. According to the standard a test signal is created, which consists
of a finite sequence of IQ data samples specified in the time domain. In conventional lab
instrumentation such as vector-signal generators, this sequence for a given standard (e.g.
W-CDMA) is typically embedded in the instrument. In testing operation these signals are
uploaded in the internal AWGs and up-converted with IQ mixers yielding the modulated
RF signal. In practice, these test records are sequentially repeated yielding a large but
finite number of discrete spectral components in the frequency domain. More precisely,
the number of samples, in combination with the sampling speed at which the signal is
generated, result in an effective frequency bin size (fAWG ), or frequency resolution of
1.02
Magnitude
0.98
2.06 2.08 2.1 2.12 2.14 2.16 2.18 2.2 2.22

Frequency [GHz]
(a)
0.75
Phase [degrees]
–0.75
2.06 2.08 2.1 2.12 2.14 2.16 2.18 2.2 2.22

Frequency [GHz]
(b)
Fig. 14.10 Measured reflection coefficient at the output reference plane of the DUT for a signal composed
of 161 sinusoidal tones in a 160 MHz bandwidth. © [2008] IEEE. Reprinted, with permission,
from [4].
the generated signal,

f sAWG 1
fAWG = = , (14.5)
NAWG TMOD
where fAWG represents the frequency bin size of the generated signal, f sAWG and NAWG
are, respectively, the sampling frequency and the number of samples used by the AWGs
to construct the waveform, and TMOD is the time period of the source signal that is needed
to meet the requirements of the modulation standard according to the given test model.
To provide the reader with an example, a W-CDMA signal has a channel bandwidth of
5 MHz, a chip rate of 3.84 Mcps, 2560 chips/slot, and 15 slots/frame. When considering
one frame the complex waveform is 10 ms long or in other words it will have a frequency
resolution of 100 Hz if we then consider a single slot, the frequency resolution becomes
1.5 kHz. This frequency representation allows us to analyze modulated communication
signals such as “classical” multi-tone signals, but now with a very large number of
frequency tones (e.g. more than 23 000 frequency tones when considering a bandwidth
of 35 MHz).
Figure 14.11 shows the frequency-binned spectral content of the I and Q signals to
be delivered to the IQ modulator (Figure 14.11(a)). This block generates the RF source
signal that drives the DUT with a given modulation (e.g. W-CDMA). Due to the ever
present nonlinearities of the active device under test, the DUT-generated waves (b1,f und
0 0
Δf I Q
–40 –40
BW/2
–80 –80
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14
(a)
Amplitude [dBm]
–30
IM3 IM3
–60 BW
IM5 IM5
–90
0 20 40 60 80 100
(b)
–30 –30
I Q
–60 –60 IM3
IM3 IM5
IM5
–90 –90
0 10 20 30 40 0 10 20 30 40
Frequency [MHz]
(c)
Fig. 14.11 Illustration of the generated and acquired signals in the proposed load-pull system.
(a) Frequency-binned spectral content of the I and Q waveforms for generating the drive signal
of the DUT. (b) Down-converted low IF representation of the spectrum in the fundamental band
at the output of the DUT. (c) Spectral content of the I and Q waveforms for generating the active
load injection signal to achieve the user-defined reflection coefficient over the fundamental band.
© [2008] IEEE. Reprinted, with permission, from [4].
and b2,f und ) contain intermodulation sidebands besides the desired fundamental signal.
Moreover, the spectral content generated by nonlinearities is also present in the baseband
and harmonic frequency bands. When considering fundamental operation, the down-
converted RF signal with intermodulation sidebands is given in Figure 14.11(b). In order
to realize the desired reflection coefficients over the total bandwidth where spectral
content is present, the I and Q injection signals must now include the third- and fifth-
order intermodulation distortion (IM 3 and IM 5 ) sidebands (Figure 14.11(c)). Failing
to provide the proper signal at the IM 3 and IM 5 frequency bands creates an unrealistic
50 termination for those DUT-generated signals, invalidating the linearity performance
evaluation. The I and Q baseband injection signal at all the controlled harmonics must
also include the third and fifth harmonic to provide a realistic reflection coefficient in
these frequency bands.
Note that, when combining high-speed AD converters [4] with mixed-signal active
loads, the time span and hence the frequency bin size used for the data acquisition must
be compatible with the generated test signal, as described by the following equation,
f sAD fAWG 1
fAD = = = , (14.6)
NAD k k · TMOD
where, fAD is the resulting frequency bin size of the acquired signals; f sAD and NAD
are, respectively, the sampling frequency and the number of samples used by the AD
converters; and k is an integer. For a correct measurement, the frequency bins of the
acquisition and the generation should match; thus the frequency resolution of the AD
converter should be set equal (k = 1), or an integer factor better (smaller frequency bin
size) than that of the generated signals.
14.5 System operating frequency and bandwidth
To properly describe the bandwidth of the presented broadband architecture we will

consider the following different bandwidths:
• signal detection bandwidth,
• signal generation bandwidth,
• modulation bandwidth.
The signal detection bandwidth is determined by the RF front-end bandwidth, i.e. by

the down-converting mixer performance, just as in a traditional network analyzer.
The signal generation bandwidth is limited by the frequency handling capabilities
of the amplifiers inside the active loads, which are commercially widely available up
to 40 GHz, and are less widely available up to 110 GHz. Moreover, depending on
the active-load topology (i.e. closed-loop or mixed-signal open-loop), discussed in the
previous section, the operating frequency of the analog components within the active
load defines the active-load frequency range, i.e. 90 degree coupler-based IQ modula-
tors (intrinsically narrow-band) for the closed-loop topology and IQ mixer (available in
wideband configurations) for the mixed-signal open-loop topology.
Finally, the maximum modulation bandwidth is determined, for the closed-loop con-
figuration, by the system configuration (as shown in Table 14.1) and the user-specified
maximum ∠ variation with frequency that can be tolerated at the DUT reference plane.
For the mixed-signal open-loop configuration, the maximum analog frequency of the
arbitrary waveform generators is the only limit for the bandwidth over which the reflec-
tion coefficients can be controlled, since the ∠ variation with frequency can be canceled
by the mixed-signal open-loop approach.
14.6 Injection power and load amplifer linearity
After reviewing the load-pull architectures and the relative active-load topologies that
allow us to perform large-signal characterization employing broadband signals, in this
section we analyze the power and linearity requirements of the active-load amplifier to
properly synthesize a specific L at the DUT reference plane.
To better analyze the problem, the active load can be described with its Thevenin
equivalent as depicted in Figure 14.12, where EDUT and ZDUT and ESYS and ZSYS are
the equivalent voltage source and output impedance of the DUT and of the measurement
I2
+
b2
ZDUT ZSYS
V2
+ a2 +
EDUT ESYS
– – –
V2 – Z0· I2
ΓL =
I2 + Z0 · I2
Fig. 14.12 Thevenin equivalent schematic of an active load-pull configuration. The load impedance offered
to the DUT at the reference plane is varied by adjusting the equivalent voltage source ESYS in
amplitude and phase. The related power needed to synthesize specific impedances depends
strongly on the equivalent system impedance (ZSYS ). © [2010] IEEE. Reprinted, with
system, respectively. The equivalent voltage sources can be expressed in terms of the
transmitted and incident waves to the DUT as,
9
bDUT · (ZDUT + Z0 )
EDUT = √ bDUT = 2 · Pb2 · (1 − |DUT |2 )
Z0
(14.7)
9
bSYS · (ZSYS + Z0 )
ESYS = √ bSYS = 2 · Pa2 · (1 − |SYS |2 ).
Z0
(14.8)
With reference to the schematic of Figure 14.12, the required injected power needed
to achieve a certain L , or in other words a certain impedance ZL = V2 /I2 , can be
calculated as,
(1 − |DUT |2 ) |ZDUT + Z0 |2 |ZL − ZSYS |2

Pa2 = Pb2 · · · . (14.9)
(1 − |SYS |2 ) |ZSYS + Z0 |2 |ZDUT + ZL |2
As is clear from (14.9), the required injected power not only depends on the output
power of the DUT and the desired L , but also on the output impedance of the device.
When considering high power devices, with output impedances in the order of a few ,
the required injection power to cover the desired Smith chart area can be extremely high
in a 50 system (e.g. 2 to 10 times higher than the maximum output power of the DUT).
Applying (14.9) to the case of a high power amplifier with an output impedance of 2
and an available output power of 200 W results in a required injection power larger than
2 kW to synthesize a load impedance of 1 in a 50 system. Clearly this represents a
strong limitation of active loads. This is usually overcome by employing pre-matching
circuitry, converting the 50 impedance of the system to a value that is much closer to
the output impedance of the DUT. This widely used technique (also applied in passive
load-pull) not only reduces the losses, but also lowers the power requirement of the
load injection amplifier [29,30]. When using a pre-matching fixture reducing the system
impedance to 10 , the required injection power for the same load condition (i.e. 1 )
reduces from 2 kW to 360 W, while with a pre-match to 5 the required injection power
is only 142.2 W.
When considering multi-tone or modulated signals, the situation becomes more com-
plicated, as the linearity of the injection amplifier also needs to be taken into account [3].
To study the linearity constraints on the injection amplifier, consider a two-tone test sig-
nal, for which the power injected by the load amplifier at the third-order intermodulation
frequency products of the two-tone test signal is given by,
Pa2 ,IM 3 (dBm) = 3 · Pa2 ,f und(dBm) − 2 · IP3,a2 =

(1 − |DUT |2 )
= 30 · log[Pb2 ,f und(mW ) · ·
(1 − |SYS |2 ) (14.10)
|ZDUT + Z0 |2 |ZL − ZSYS |2
· · ] − 2 · IP 3,a2
|ZSYS + Z0 |2 |ZDUT + ZL |2
where Pb2 ,f und is the available power coming out of the DUT at the fundamental tones,
and Pa2 ,f und and IP3,a2 are the power injected by the load amplifier at the fundamen-
tal tones and its output third-order intercept point, respectively. A harmonic balance
simulation with an Agilent ADS is performed using the simple schematic illustrated in
Figure 14.13. In this schematic an amplifier component based on a polynomial model
is used to simulate the DUT and the injection amplifier linearity. The same DUT as for
the single tone considerations is used, with the same output impedance of 2 , while its
OIP3 is set in this simulation to 63 dBm. For this device the output power is set equal
to 50 W per tone, in order to achieve the same peak voltage as in the single tone case.
These conditions yield an actual IM 3 of the DUT of −30.35 dBc. The results of the
simulation are shown in Figure 14.14, where the apparent IM 3 of the DUT is plotted as
a function of the decreasing OIP3 of the injection amplifier, for different pre-matching
conditions of the system impedance. The dotted line is the actual IM 3 level as would
be achieved with a passive circuit. The dot-dash line represents the IM 3 level due to the
Pa2 ,IM 3 as approximated by (14.10). From Figure 14.14, we can observe that the correct
vin vout
I_Probe 2 1
P_nTone P_nTone
PORT1 I_dut2 PORT2
Amplifier2 S2P_Eqn Amplifier2
Num=1 Num=2
DUT S2P1 Inj_Amp
Z=50 Ohm Z=50 Ohm
S21=38.456 S[1,1]=0 S21=dbpolar(50,0)
Freq[1]=2 GHz Freq[1]=2 GHz
S11=polar(0,0) S[1,2]=1 S11=polar(0,0)
Freq[2]=2.1 GHz Freq[2]=2.1 GHz
S22=polar(0.9231,180) S[2,1]=1 S22=polar(0,180)
P[1]=dbmtow(7.7472) P[1]=polar(dbmtow(power),180)
S12=0 S[2,2]=0 S12=0
P[2]=dbmtow(7.7472) P[2]=polar(dbmtow(power),180)
Z[1]=50
Z[2]=5
Fig. 14.13 ADS schematic for the evaluation of the required injection amplifier linearity by active load-pull.
−20
Zsys = 5Ω
Zsys = 7Ω
−25
Zsys = 10Ω
IM3 level DUT [dBc]
Passive impedance
Calculated
−30
−35
−40
−45
75 73 71 69 67 65 63 61 59
OIP3 Iinjection amplifier [dBm]
Fig. 14.14 Harmonic balance simulated IM 3 level of the DUT vs. decreasing OIP3 of the injection amplifier
for different impedance pre-match values. The dotted line is the actual IM 3 level as would be
achieved with passive matching techniques. The dot-dash line represents the IM 3 level only due
to the Pa2 ,IM 3 as approximated by (14.10). A polynomial model was used for the amplifier
linearity. © [2010] IEEE. Reprinted, with permission, from [28].
IM 3 level is only achieved when the injection amplifier OIP 3 is sufficiently high. When
the injection amplifier is less linear, it introduces significant IM 3 products that cause an
error in the measurements, such as IM 3 increase or cancellation effects. Consequently,
to have reliable linearity measurements in a load-pull setup, even when pre-matching
is used, the injection amplifier (and thus its peak power) needs to be at least 10 times
higher than the one of the DUT.
At this point it is important to note that while the peak power requirements apply
for both closed-loop and mixed-signal active loads, the linearity requirements strictly
apply only for closed-loop active loads. The reason for this is to be found in the iterative
convergence approach employed by the mixed-signal approach. Since the iteration on the
required reflection coefficient involves the signal channel as well as its adjacent channels
(intermodulation products), this procedure compensates, in the injection signal, for most
of the nonlinearities of the active-loop amplifier. For this reason it is difficult to indicate
the linearity requirement for the mixed-signal approach in a closed formula since it
also depends on the convergence algorithm used. In general it can be stated that such
a technique allows the use of the active-load amplifier, even under modulated signal
excitation, much closer to its compression point.
14.7 Baseband impedance control
The importance of controlling the baseband impedance, when aiming for optimum device
linearity, was proven theoretically in [31] and [5] as well as experimentally in [32]
To Baseband control
(active or passive) and DC
Low frequency bridge

To RF ADC/Sampler
To RF Mixer/Sampler
To RF active loads DUT Reference

f0,2f0 etc. plane
RF coupler section
Fig. 14.15 Low-frequency detection bridges placed on the bias line to couple the low-frequency traveling
waves to the broadband receiver, i.e. sampler or AD.
and [33]. When attempting to properly control the baseband impedance two difficulties
arise when employing conventional high-frequency system architectures, namely:
1. how to obtain calibrated measurements of the baseband reflection coefficients,

2. how to obtain impedance control over the wideband (baseband) frequency.
The first point in the above list is due to the limited bandwidth of the reflectometers,
often implemented as coupled line couplers, which are usually employed in the test sets
of VNA and load-pull architectures (see Figure 14.4). These components provide a min-
imum operating frequency (linked to the dimension of the component) often in the order
of few hundred MHz. To circumvent this limitation in [34] and [35] low-frequency (i.e.
resistive bridges) couplers were employed in the bias line. Inserting the low-frequency
detection bridges on the bias path, as shown in Figure 14.15, avoids high-frequency sig-
nals being routed through these components which would provide very high losses at RF
frequencies. The low-frequency traveling waves at the output of the detection bridges
can be directly sampled by the receiver, as was shown in [34], using a microwave sam-
pling oscilloscope architecture or by the broadband AD employed in a heterodyne mixer
based architecture as was presented in [35]. Using low-frequency bridges also allows
us to employ the conventional 12 error-terms calibration techniques for the calibrated
measurement of the baseband reflection coefficients. Moreover, the use of dedicated low-
frequency detection bridges allows us to optimize the accuracy of the controlled baseband
reflection coefficients, due to the high performance of the low-frequency bridges (i.e.
high directivity). Finally, employing similar calibration techniques the ones used at RF
allows us to share the same calibration standards for both the RF as well as the baseband
calibration (e.g. short, open, load, and thru), reducing the overall calibration time.
In order to enable baseband impedance control in [34] an additional AWG generating
a signal coherent (i.e. sharing the same clock) with the RF signal driving the device, was
employed. As this is an open-loop topology on the baseband path, low-frequency ampli-
fiers need to be employed. When standard compliant modulated signals are employed
(i.e. W-CDMA) DC coupled amplifiers should be employed to provide a controlled
impedance through the entire baseband frequency range. In [17] a simplified baseband
impedance control was presented that employed a simple resistive switch bank. While
this method allows only a resistive impedance control, when implemented in a small form
factor (i.e. using SMD components and PCB dedicated layout) this approach allows a
simple control of the baseband resistance over a broad frequency range (i.e. up to 5 MHz),
which can be employed for device linearity improvement [3].
14.8 Broadband large signal measurement examples
In this section some measurement examples demonstrating the large-signal char-

acterization capabilities of the broadband architectures presented in the previous
sections given.
14.8.1 IMD asymmetries measurements

As was reported in [5], the out-of-band source termination can cause asymmetries
between the upper and lower IMD products, shown in Figure 14.16, where the maximum
38
OIP3 lower band
36 OIP3 upper band
OIP3up, OIP3Io [dBm]
34
32
30
28
26
12 14 16 18 20 22 24 26 28
ZS,BB = ZS,2nd [Ohm]
Fig. 14.16 Measured maximum OIP3 levels for upper and lower IM 3 components versus resistive
ZS,BB = ZS,2nd using a swept Icq bias conditions (f0 = 2.14 GHz, f = 0.5 MHz). © [2006]
IEEE. Reprinted, with permission, from [3].
measured upper and lower band OIP3 (achieved at each Ic sweep) are plotted as a function
of ZS,BB = ZS,2nd [33]. This means that the IMD at 2f1 − f2 has a different magnitude
to that at 2f2 − f1 . Since the linearity performance is limited by the highest IMD level,
this asymmetry leads to a degradation in linearity performance compared to the opti-
mum symmetrical case. This asymmetry between the IM 3 signals versus tone spacing
is often referred to as the memory effect. These memory effects can be divided into two
classes: thermal memory effects (up to a few MHz) and electrical memory effects (caused
mostly by the source and load termination, including the biasing network impedances).
The mechanism generating these memory effects can be quite complex, and standard
available models are quite often not able to predict them. This highlights the importance
of properly characterizing these effects through measurements, which can provide the
information required for their cancellation. A clear example of the large variation of
IM 3 upper and lower tone power versus the tone spacing of the input signal is given in
Figure 14.17. Here a bipolar device matched for optimum linearity was stimulated with
a two-tone signal with an increasing tone spacing.
The importance of properly characterizing these intermodulation distortion mecha-
nisms becomes clear when studying (digital) pre-distortion techniques to linearize the
PA behavior. The information required from the measurements is the power versus fre-
quency of the IMD components and their phases [36,37]. This has been obtained through
different approaches, both in the frequency domain [36] using sinusoidal inputs as well
as in the time domain [38] using a waveform with a complex time dependence. In most
cases, however, this information was not coupled with full control of the source and load
termination offered at the device under test. The broadband system architecture described
in the previous sections provides the capability to characterize the intermodulation
distortion as a function of tone spacing, while controlling the high-frequency termination
and the baseband impedance.
–44
–46
–48
IM3up IM3Io [dBc]
–50
–52
–54
–56
–58
–60 IM3 upper band
–62 IM3 lower band
0.0 500.0k 1.0M 1.5M 2.0M 2.5M 3.0M

Tone spacing [Hz]
Fig. 14.17 Measured upper and lower band IM 3 components versus tone spacing for fixed bias and load
conditions.
14.8.2 Phase delay cancellation

When employing mixed-signal loads it is possible to demonstrate, directly on the DUT
performance degradation, the effects of phase delays when employing broadband signals.
In [4], an NXP Gen 6 LDMOS device with a gate width of 1.8 mm is measured using
wideband modulated signals in load-pull test conditions. The drain current and voltage
are set to 13 mA and 28 V, respectively. First, the optimum fundamental load and source
matching conditions are found using conventional single tone load-pull measurements,
namely: L,f 1 = |0.6| ∠45◦ and S,f 1 = |0.5| ∠90◦ . The input and output baseband
impedances are set to enforce a short condition; and the input and output 2nd harmonics
are set to an open circuit condition (L,f 2 = S,f 2 = |0.95|) to optimize the efficiency
[39]. The electrical delay-free operation, provided by mixed-signal loads is compared
with the results (under phase delay) which would be achieved by employing the closed-
loop loads presented in Section 14.4.1. For the driving signal, a two-channel W-CDMA
signal (centered at 2.135 GHz and 2.145 GHz) is chosen, and the input and output
reflection coefficients are set to the above defined optimal conditions providing a phase
delay given by the following two cases:
1. without electrical delay,

2. with an electrical delay of 4.85◦ /MHz for the fundamental source and load and
4.6◦ /MHz for the 2nd harmonic source and load.
Figure 14.18 illustrates the source and load matching conditions provided to the active
device under test for the two different cases. Note that the filled markers represent the
source and loading conditions for the two-carrier W-CDMA signal without any elec-
trical delay, yielding points that are completely overlapping versus frequency in the
Smith chart. As shown in Figure 14.18, for the case with electrical delay the funda-
mental load trajectory has been shifted such that the optimum matching condition is
now centered at 2.135 GHz. This was required to avoid the unstable region of the active
device.
The comparison is to the “best known case” of a closed-loop load, since in practical
closed-loops there are amplitude variations within the control frequency band that are
not accounted for. Moreover, oscillation conditions in closed-loop systems for very large
bandwidths are difficult to avoid, due to the use of wideband loop filters. Passive loads
with harmonic tuning will have a comparable or even worse phase variation of the
reflection coefficients versus frequency, depending on the distance of the tuner from the
DUT reference plane.
The measurement results are summarized in Table 14.2. There is significant perfor-
mance degradation for the active device when measured with an electrical delay present
in the reflection coefficients. This is also evident from Figures 14.19(a) and (b) which
show the power spectral density at the device output reference plane for the fundamental
and 2nd harmonic frequency bands. Note that a 5 dB output power drop and close to an
8% degradation of the PAE can be observed, when compared to the situation with no
electrical delay.
Table 14.2 Measurement results comparison in the

two cases with and without electrical delay
MEASUREMENT RESULTS
Without With
electrical delay electrical delay
PAE 24.2 % 16.3 %

POUT Ch. 1 20.3 dBm 20.5 dBm
POUT Ch. 2 20.6 dBm 15.4 dBm
ACLR1 Ch. 1 −43.9 dBc −43.0 dBc
ACLR2 Ch. 1 −42.2 dBc −41.6 dBc
ACLR1 Ch. 2 −42.1 dBc −41.8 dBc
ACLR2 Ch. 2 −39.6 dBc −39.2 dBc
Γsource,f1, delay Γload,f1, delay

Γsource,f1, no delay Γload,f1, no delay
Γsource,f2, delay Γload,f2, delay
Γsource,f2, no delay Γload,f2, no delay
Fig. 14.18 Source and load reflection coefficients at the device reference plane in the fundamental
(2.1225 GHz – 2.1575 GHz) and harmonic (4.245 GHz – 4.315 GHz) frequency range, with
electrical delay (open symbols) and without electrical delay (filled symbols). © [2008] IEEE.
14.8.3 High power measurements with modulated signals

As previously explained in Section 14.6, in an active load-pull system, to synthesize a
specific L at the DUT reference plane, it is necessary to inject a certain amount of
power into the DUT. When working at high power levels with wideband signals an extra
complication arises because the linearity of the injection amplifier needs to be taken into
account.
–10
Output power spectral density [dBm/Hz]
–20
–30
–40
–50
–60
Without delay
With delay
–70
2.125 2.13 2.135 2.14 2.145 2.15 2.155
Frequency [GHz]
(a)
–20
Output power spectral density [dBm/Hz]
–30
–40
–50
–60
–70
Without delay
With delay
–80
4.25 4.26 4.27 4.28 4.29 4.3 4.31
Frequency [GHz]
(b)
Fig. 14.19 Measured output power spectral density (dBm/Hz) vs. frequency (GHz) of a NXP GEN 6
LDMOS device (gate width 1.8 mm) in the proposed load-pull setup (a) at the fundamental
frequency band using a 3 kHz resolution bandwidth. (b) at the 2nd harmonic frequency band
using a 6 kHz resolution bandwidth. The measurement is shown for the two cases with (dashed
line) and without electrical delay (drawn line). The reflection coefficients offered to the device
under test are given in Fig. 14.18. © [2008] IEEE. Reprinted, with permission, from [4].
As was shown in Figure 14.14, to have reliable linearity measurements in an active

load-pull setup, even when pre-matching is used, the injection amplifier linearity (and
thus its peak power) needs to be at least 10 times higher than that of the DUT, which
becomes extremely expensive and impractical when working with devices with peak
envelope powers as high as 200 W, e.g. designed for base-station applications.
0
Average PAE [%] for 30 W output power
ACPR [dB] for 30 W output power
−0.1
−0.2
−0.3
33
−0.4 .7
32
.3
30
.9
29
−0.5 28 .5
.
26 1
25 .8
.4
2 −34.9
−0.6 22 4
.
21 6 −34.3
.
19 2 −33.8
.8
−33.3
−0.7 −32.7
−32.2
−31.
2
−0.8
−0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1
(a)
0
Average PAE [%] @ 3 dB PAR reduction
Average Pout [W] @ 3 dB PAR reduction
−0.1
−0.2
−0.3
−0.4
32
22.5
23.7
.3
24.91
30
26.
−0.5 2 .9
27.3
2 9.5
28. 8
26 8.0
29.
30.
32.2
2 .6
33.4
2 5
9
2 3 .2
−0.6 21 2.3.8
1 .0
189.5
.1
−0.7
−0.8
−0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1
(b)
Fig. 14.20 (a) Load-pull contours, on a 5 ohm normalized Smith-chart, of average PAE and ACPR for an
average output power of 30 W. (b) Load-pull contours, on a 5 normalized Smith-chart, of
average PAE and average output power at 3 dB of peak-to-average ratio reduction. The related
peak to average power (PEP) is as high as 150 W. © [2010] IEEE. Reprinted, with permission,
from [28].
In this section the capability of a mixed-signal architecture to work with complex

modulated signals at power levels that are typically in use for base station applications
(e.g. peak envelope power ∼ 200 W), is reviewed [28].
As previously mentioned, when using mixed-signal loads, an iterative process is per-
formed to optimize the reflection coefficient of each individual frequency component of
the wideband signal (e.g. 23 362 tones with 1.5 kHz spacing for a W-CDMA considering
a total bandwidth of 35 MHz). Due to these iterations, the injection amplifier is basically
pre-distorted for its own nonlinearities, allowing the use of an injection amplifier with
a much lower linearity compared with what is typically required in closed-loop archi-
tectures (i.e. 10 times higher than that the of the DUT). Figure 14.20 shows the ACPR,
average PAE, and output power for a single-channel W-CDMA signal at 2.14 GHz with
a peak-to-average ratio of 9.5 dB, employing an injection amplifier with 200 W and an
associated 60 dBm output I P3 . In these measurements, the nonlinearity of the injection
amplifier does not affect the measurement results because the L is controlled to the
user-defined value in and out of the band.
References
[1] Motorola, Long Term Evolution (LTE): A Technical Overview. Motorola technical white
paper, 2010.
[2] H. Arthaber, M. Mayer, and G. Magerl, “A broadband active harmonic load-pull setup with
a modulated generator as active load,” in Microwave Conference, 2004. 34th European, 2,
Oct. 2004, pp. 685–688.
[3] M. Spirito, M. Pelk, F. van Rijs, S. Theeuwen, D. Hartskeerl, and L. de Vreede, “Active
harmonic load-pull for on-wafer out-of-band device linearity optimization,” IEEE Trans.
Microw. Theory Tech., 54, no. 12, pp. 4225–4236, Dec. 2006.
[4] M. Marchetti, M. Pelk, K. Buisman, W. Neo, M. Spirito, and L. de Vreede, “Active harmonic
load-pull with realistic wideband communications signals,” IEEE Trans. Microw. Theory and
Tech., 56, no. 12, pp. 2979–2988, Dec. 2008.
[5] M. van der Heijden, H. de Graaff, and L. de Vreede, “A novel frequency-independent third-
order intermodulation distortion cancellation technique for BJT amplifiers,” IEEE J. Solid-
State Circuits, 37, no. 9, pp. 1176–1183, Sept. 2002.
[6] (2012) Agilent technologies website. [Online]. Available: http://www.home.agilent.
com/agilent/product.jspx?nid=-33821.0.00&cc=US&lc=eng
[7] (2012) Tektronix website. [Online]. Available: http://www.tek.com/oscilloscope/dsa8300-
sampling-oscilloscope
[8] M.Akmal,V. Carrubba, J. Lees, S. Bensmida, J. Benedikt, K. Morris, M. Beach, J. McGeehan,
and P. Tasker, “Linearity enhancement of GaN HEMTs under complex modulated excita-
tion by optimizing the baseband impedance environment,” in Microwave Symposium Digest
(MTT), 2011 IEEE MTT-S International, June 2011, pp. 1–4.
[9] (2012) Miteq inc. website. [Online]. Available: http://www.miteq.com/
[10] (2012) Marki microwave website. [Online]. Available: http://www.markimicrowave.com/
[11] Agilent Technologies, Optimizing Dynamic Range for Distortion Measurements. Santa Rosa,
CA: Agilent 5980-3079EN, 2000.
[12] D. M. Leenaerts, J. van der Tang, and C. S. Vaucher, Circuit Design for RF Transceivers.
Boston, Dordrecht, London: Kluwer Academic Publishers, 2001.
[13] (2012) National instruments website. [Online]. Available: http://www.ni.com/
data-acquisition/
[14] S. Vandenplas, J. Verspecht, F. Verbeyst, F. Vandamme, and M. Bossche, “Calibration issues
for the large signal network analyzer (LSNA),” in ARFTG Conference Digest, Fall 2002,
Washington, DC, Dec. 2002, pp. 99–106.
[15] K. Noujeim, J. Martens, and T. Roberts, “A frequency-scalable NLTL-based signal-source
extension,” in 41st European Microwave Conference (EuMC), 2011, Manchester, UK, Oct.
2011, pp. 476–479.
[16] P. Roblin, Nonlinear RF Circuits and Nonlinear Vector Network Analyzers: Interactive
Measurement and Design Technique. Cambridge UK: Cambridge University Press, 2011.
[17] M. Spirito, L. C. N. de Vreede, M. de Kok, M. Pelk, D. Hartskeerl, H. F. F. Jos, J. E.
Mueller, and J. Burghartz, “A novel active harmonic load-pull setup for on-wafer device lin-
earity characterization,” in Microwave Symposium Digest 2004, Fort Worth, TX, June 2004,
pp. 1217–1220.
[18] D. D. Poulin, J. R. Mahon, and J. P. Lanteri, “A high power on-wafer pulsed active load pull
system,” IEEE Trans. Microw. Theory Tech., 40, pp. 2412–2417, Dec. 1992.
[19] G. Madonna, A. Ferrero, M. Pirola, and U. Pisani, “Testing microwave devices under different
source impedance values - a novel technique for on-line measurement of source and device
reflection coefficients,” IEEE Trans. Instrum. Meas., 49, pp. 285–289, Apr. 2000.
[20] G. P. Bava, U. Pisani, and V. Pozzolo, “Active load technique for load-pull characterisation
at microwave frequencies,” Electronics Letters, 18, pp. 178–180, Feb. 1982.
[21] D. Rytting, “Network analyzer error models and calibration methods,” White Paper,
September, 1998.
[22] A. Ferrero and U. Pisani, “An improved calibration technique for on-wafer large-
signal transistor characterization,” IEEE Trans. Instrum. Meas., 42, no. 2, pp. 360–364,
Apr. 1993.
[23] L. Devlin and B. Minnis, “A versatile vector modulator design for MMIC,” in Microwave
Symposium Digest, 1990., IEEE MTT-S International, May 1990, pp. 519–521, vol. 1.
[24] M. Squillante, M. Marchetti, M. Spirito, and L. de Vreede, “A mixed-signal approach for
high-speed fully controlled multidimensional load-pull parameters sweep,” in Microwave
Measurement Conference, 2009 73rd ARFTG, Jun. 2009, pp. 1–5.
[25] W. C. E. Neo, J. Qureshi, M. J. Pelk, J. R. Gajadharsing, and L. C. N. de Vreede, “A mixed-
signal approach towards linear and efficient N-Way doherty amplifiers,” IEEE Trans. Microw.
Theory Tech., 55, no. 5, pp. 866–879, May 2007.
[26] “UTRA (BS) FDD; Radio Transmission and Reception,” Technical Specification Group
Radio Access Networks – 3rd Generation Partnership Project, 3GPP TS 25.104 V4.1.0
(2001-06), 2001.
[27] “Base station conformance testing (FDD),” Technical Specification Group Radio Access
Networks – 3rd Generation Partnership Project, 3GPP TS 25.141 V4.1.0 (2001-06), 2001.
[28] M. Marchetti, R. Heeres, M. Squillante, M. Pelk, M. Spirito, and L. de Vreede, “A mixed-
signal load-pull system for base-station applications,” in Radio Frequency Integrated Circuits
Symposium (RFIC), 2010 IEEE, May 2010, pp. 491–494.
[29] Z. Aboush, C. Jones, G. Knight, A. Sheikh, H. Lee, J. Lees, J. Benedikt, and P. Tasker,
“High power active harmonic load-pull system for characterization of high power 100-watt
transistors,” in 2005 European Microwave Conference, 1, Oct. 2005, p. 4.
Trans. Instrum. Meas., 57, no. 11, pp. 2640–2646, Nov. 2008.
[31] V. Aparin and C. Persico, “Effect of out-of-band terminations on intermodulation distor-
tion in common-emitter circuits,” in Microwave Symposium Digest, 1999 IEEE MTT-S
International, 3, 1999, pp. 977–980, vol. 3.
[32] D. Williams, J. Leckey, and P. Tasker, “Envelope domain analysis of measured time domain
voltage and current waveforms provide for improved understanding of factors effecting lin-
earity,” in Microwave Symposium Digest, 2003 IEEE MTT-S International, 2, June 2003,
pp. 1411–1414, vol. 2.
[33] M. Spirito, M. van der Heijden, M. Pelk, L. de Vreede, P. Zampardi, L. Larson, and
J. Burghartz, “Experimental procedure to optimize out-of-band terminations for highly linear
and power efficient bipolar class-AB RF amplifiers,” in Proceedings of the Bipolar/BiCMOS
Circuits and Technology Meeting, 2005, Oct. 2005, pp. 112–115.
[34] A. Alghanim, J. Lees, T. Williams, J. Benedikt, and P. Tasker, “Using active IF load-pull to
investigate electrical base-band induced memory effects in high-power ldmos transistors,”
in APMC 2007, Asia-Pacific Microwave Conference, 2007, Dec. 2007, pp. 1–4.
[35] M. Mirra, M. Marchetti, F. Tessitore, M. Spirito, L. de Vreede, and L. Betts, “A multi-step
phase calibration procedure for closely spaced multi-tone signals,” in 2010 75th ARFTG
Microwave Measurements Conference (ARFTG), May 2010, pp. 1–5.
[36] J. Vuolevi, T. Rahkonen, and J. Manninen, “Measurement technique for characterizing mem-
ory effects in RF power amplifiers,” in Radio and Wireless Conference, 2000. RAWCON
2000. 2000 IEEE, 2000, pp. 195–198.
[37] J. Vuolevi, J. Manninen, and T. Rahkonen, “Cancelling the memory effects in RF power
amplifiers,” in The 2001 IEEE International Symposium on Circuits and Systems, 2001.
ISCAS 2001. 1, May 2001, pp. 57–60, vol. 1.
[38] P. Draxler, I. Langmore, T. Hung, and P. Asbeck, “Time domain characterization of power
amplifiers with memory effects,” in Microwave Symposium Digest, 2003 IEEE MTT-S
International, 2, June 2003, pp. 803–806, vol. 2.
[39] D. Hartskeerl, I. Volokhine, and M. Spirito, “On the optimum 2nd harmonic source and load
impedances for the efficiency-linearity trade-off of RF LDMOS power amplifiers,” in Radio
Frequency integrated Circuits (RFIC) Symposium, 2005. Digest of Papers. 2005 IEEE, June
2005, pp. 447–450.
15 Pulse and RF measurement
Anthony Parker
15.1 Introduction
Circuits exhibit a variety of operational traits that are far from the behavior presented in
introductory circuit design textbooks. Transistor characteristics curves vary significantly
depending on how they are measured and on the history of electrical conditions. The
characteristics are not always repeatable, which raises the dilemma of the choice of
which characteristic to base a design upon.
The central idea behind pulse measurements is that the high-frequency characteristics
of a device are a function of a quiescent operating condition. Pulse techniques attempt
to determine these in an invariable operating condition. If short enough pulses are used,
a pulse measurement at a specific condition of operation gives the characteristics that
a high-frequency signal would encounter. This is a simple idea, but there is a practical
limit to how short the pulses can be, so it is then necessary to draw upon radio-frequency
techniques to probe past higher-frequency anomalies in the characteristics.
Dynamic processes and interactions in active elements produce seemingly complicated
electrical characteristics that are best explored with pulse and RF techniques. These pro-
cesses can be traced to mechanisms of self-heating by power dissipation, bias-dependent
change in trapped charges, and to impact ionization and breakdown.
This chapter covers a set of topics that provide a foundation for understanding the
pulse measurement technique augmented with RF measurements. Pulse characterization
techniques dovetail with RF and nonlinear techniques to explore transistor dynamics for
small-signal and nonlinear applications.
15.2 Dynamic characteristics
Many devices exhibit characteristics that change with time, frequency, and with oper-
ating conditions such as temperature and terminal bias. All characteristics are affected,
including terminal current, linearity, and charge state. The pulse measurement technique
is one of the more powerful and insightful methods for characterizing these dynamics in
transistors and circuits [1].
The characteristics vary considerably with the time taken to measure each point in the
curves. This is well illustrated in Figure 15.1, which shows characteristics that are typical
for GaAs pHEMT transistors. They are reproducible and are consistent from wafer to
wafer in a mature fabrication process. The dilemma that this dynamic behavior presents
is the question of what are the characteristics seen by signals in a circuit.
Pulse and RF measurement 415
100 ns
0.9
2 μs
0.8 10 μs
0.7
Drain current (A/mm)
VGS = 0.0 V 1 ms
0.6 10 ms
100 ms
0.5
0.4
VGS = −0.5 V
0.3
0.2 VGS = −1.0 V

0.1
VGS = −1.5 V
0
0 1 2 3 4 5 6 7 8
Drain potential (V)
Fig. 15.1 Pulse characteristics of a 150μm GaAs/AlGaAs pHEMT with 2.0V pinch-off. The pulses
emanate from a bias at 4.0 V on the drain and –2.0 V on the gate. Six pulse widths were used,
with a 100ms pulse repetition period.
Pulse characteristics also change with the bias condition used between pulse measure-
ments, as shown in Figure 15.2. This raises the question of what is the bias and how does
it relate to the quiescent condition relative to the rate of change in characteristics. A large
signal may see a fixed set of characteristics set by the average bias, or various characteris-
tics corresponding to conditions set by the signal during its excursions along the load line.
This is a question that relates to the characteristic frequencies of the processes involved.
In practice the electrical characteristics of field-effect transistors change with the
timing of signals, bias condition, and frequency. From the perspective of high-frequency
signals, the characteristics are time-variant, or a function of memory of previous signals.
Pulse measurements attempt to establish a history at a fixed bias and then measure the
characteristics quickly enough, so the bias is not perturbed significantly. Each bias point
has a corresponding set of static characteristics.
Changes in static characteristics with bias come from changes in physical processes
associated with the operation of the transistor. Temperature and charge state at trapping
sites are important factors, or state variables, that change with bias. The drain current
is a time-invariant function of terminal potentials and state variables such as junction
temperature and the potential of trapped charge.
Each state variable has its own response to the terminal conditions. From a broadband
perspective, these responses are slowly varying signals that also control the characteris-
tics of the transistor. From a narrow-band perspective, the responses are state variables
dependent on past signals that control the static characteristics.
In many devices the variations caused by changes in state variables can be quite
dramatic and affect basic performance parameters such as the intrinsic gain of a transistor,
as shown in Figure 15.3.
416 Anthony Parker
0.8
VGS = 0.0 V
0.6
VGS =−0.5 V
0.4
(−0.5,4.0)
VGS = −1.0 V
0.2
VGS = −1.5 V
0
Bias (−2.0,1.0) Bias (−2.0,4.0)
0 1 2 3 4 5 6 7 8
Drain potential (V)
Fig. 15.2 Pulse characteristics of the 150 μm GaAs/AlGaAs pHEMT shown in Fig. 15.1 from three bias
points. The 100 ns pulse characteristics vary considerably with the long-term bias condition
established between pulses. The arrows show the transition from each bias point to (VGS = −0.5,
VDS = 2.4).
18
VDS = 4.5 V
16
Intrinsic gain (y21/y22)
14
VDS = 3.5 V
12
VDS = 2.5 V
10
VDS =1.5 V
8
100 Hz 10 kHz 1 MHz 100 MHz 10 GHz

Frequency
Fig. 15.3 Intrinsic gain of the 150 μm GaAs/AlGaAs pHEMT shown in Fig. 15.1. There is variation
across several decades of frequency that depends significantly on drain bias. There is a break in
the data at 10 MHz below which pulse measurements were used and above which RF
measurements were used.
15.3 Large-signal isodynamic measurements
The principle behind a typical use of pulse measurements is maintaining a fixed bias
relative to any dynamic processes that change the device’s electrical characteristics. A
nominal bias is held for a long period between very short pulses during which mea-
surements are made. Ideally the bias period would be longer than the response times
of the transistor’s or circuit’s dynamic processes and the pulses would be shorter than
these response times. For a FET, drain current measurements during a set of pulses, each
to a different potential, provide a pulsed drain-current characteristic. In the ideal case,
there would be no response recorded from the transistor’s dynamic processes, so the
characteristics are considered to be isodynamic.
Isodynamic characteristics are those for constant state variables. Each set of state
variable values has a corresponding isodynamic characteristic and variation of transistor
characteristics with operating condition are a result of changes in the state variables.
That is, the dynamic behavior of the transistor is described by variation of the state
variables. The processes that link state variables to terminal or operating conditions are
the dynamic processes of the transistor. True DC characteristics are those for which the
state variables have reached steady state at each point.
An isodynamic pulse measurement is illustrated in Figure 15.4, which also shows a
true DC measurement for comparison. The latter requires measurement after a long time
at each point to ensure that all dynamic processes have reached a steady state. Tradi-
tional step-and-sweep measurements can be too fast for this and exhibit manifestations
of transistor dynamics, which are also shown in Figure 15.4. The timing of true DC
1.2 Isodynamic
100 μs sweep
10 ms sweep
1
True DC
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7 8
Drain-source potential (V)
Fig. 15.4 Comparison of step-and-sweep measurements with estimates of isodynamic and true DC
characteristics. The isodynamic characteristics are from a bias at VDS = 5, IDS = 0.36 A/mm
where they overlap the DC characteristic. The gate-source potential from –2.0 to 0.5 V in 0.5V
steps is the parameter.
418 Anthony Parker
measurements can be in the range 10-100 ms for Gallium Arsenide devices to many
hundreds of seconds in wide band-gap devices, such as Gallium Nitride.
Reliance on DC data alone is problematic and leads to inconsistency between bias
and small-signal characteristics. For example, the bent curves in step-and-sweep mea-
surements suggest an apparent change in drain conductance that is an artifact of the
measurement sequence and timing that is not observed during radio-frequency operation.
The DC and isodynamic characteristics are coincident at their common bias point but
differ considerably at other points. In most radio-frequency applications the operating
frequency of the transistor is higher than that of the response of its dynamic processes, so
the isodynamic characteristics provide a better view of device operation. That is, signal
excursions and small-signal parameters at radio frequency should be determined from
the isodynamic characteristics rather than the DC characteristics.
15.3.1 Measurement outside safe-operating areas

The safe-operating area, defined by limits beyond which the device may be damaged,
varies with signal condition. The safe area is bounded by: a maximum voltage; a
maximum temperature and hence maximum power-time product; a maximum current;
and a maximum current-time product. Exceeding these limits can cause breakdown or
permanent physical alteration of the device.
Time constants associated with damage limitations allow pulse measurements to
extend beyond the safe-operating area of DC measurements. Temperature rise can be
slow enough to allow very high power levels to be achieved for short periods. After a
short time, the device must be returned to a low power condition to cool down. The larger
safe-operating area for shorter pulses is invaluable for investigating devices that may be
used in pulsed applications, such as radar.
15.3.2 Pulsed-RF characteristics

Pulsed measurement can be accompanied by pulsed-RF measurements. Pulsed-RF net-
work analyzers have been developed to measure the performance of the transistor during
its pulsed operating condition. These systems measure the microwave characteristics
of devices in isodynamic bias conditions. The result is a set of RF characteristics in a
transient operating condition that corresponded to a different quiescent condition. This
can be quite different from static S-parameters that are measured over a range of DC bias
conditions because there will be changes in RF behavior linked to the dynamic processes
in the transistor.
Pulsed-RF network analyzers also allow RF characterization at bias points outside the
safe-operating area. Pulsed and radar applications can operate in these regions.
15.4 Dynamic processes
Characterizing and modeling the dynamics of transistor systems requires an understand-

ing of the processes in the device that cause the observed variation. This understanding
is essential for the design of pulse and linearity measurements, the interpretation of
the resulting data, and for the development of circuit models that predict the observed
behavior.
The two dominant mechanisms are temperature dependence and charge trapping. A
description of these is presented in this section to provide the understanding required to
interpret measurements and model transistor dynamics.
15.4.1 Temperature and self-heating

Temperature has a significant impact on the nature of all aspects of transistor operation.
The dynamic processes, their response rates, and the current transport process all vary
with temperature. These processes influence a transistor’s current and hence power dis-
sipation, which then affects temperature via a thermal feedback. This feedback is known
as self-heating and it is an inescapable aspect of all dynamic processes in any transistor
circuit.
Over a reasonable temperature range, say 250 K to 400 K, a linear approximation
might be assumed. Across this temperature range, the drain current of a FET would
follow
iD = iDO (1 − λ T ), (15.1)
where λ [K−1 ] is a thermal coefficient at temperature TN [K], and iDO [A] is the current
at temperature TN [2, 1]. The temperature rise T [K] is the difference between TN and
the channel temperature. A temperature rise is produced when there is power dissipated
in the channel, pD (t) [W], which is the product of the channel current, iD , and potential,
vDS [V]. The temperature at any time, t [s], is

T (t) = pD (τ ) hT (t − τ )dτ, (15.2)
which is a convolution of the power dissipation and the impulse response, h(t), of the
thermal path from the channel to ambient [3].
Thermal response
Power dissipation and temperature are related by the specific heat capacity, c [J/K·kg],
and thermal conductivity, k [W/m·K], of the structure. Consider heat flow into a material,
dQ 2
dt [W/m ], through a small cross section to a heat sink. The product of area density and
heat capacity, ρA c [J/K·m3 ], is a measure of the energy that can be stored in a region of
material for a given temperature, which is analogous to charge in a capacitor for a given
voltage. Thermal conductivity relates temperature difference or gradient through the
region of material, ∇T [K/m], to the heat flow into it. The thermodynamic rate equation
relates these quantities as
dQ d
= ρA c T + k ∇T . (15.3)
dt dt
The time constant of this thermal response is the ratio of the mass-heat capacity product
and thermal conductivity, τT = ρA c/k.
420 Anthony Parker
A small region in the vicinity of the channel of a transistor has dimensions in the
order of fractions of microns with correspondingly small area density and high thermal
conductivity. The rate of temperature rise of such a region is in the order of nanoseconds.
The whole transistor and its surroundings form a distributed thermal path that draws heat
from the channel. The rate of heat flow for the larger structure is slower because the net
thermal conductivity is lower and mass is larger.
Transient measurements show heating to be a sub-first-order phenomenon. That is a
gradual response over several decades of time, which contrasts a first-order response that
occurs over about one decade of time. A detailed solution involves fractional calculus,
which can confirm that the response of a regular distributed thermal path is of the order
of one half. In the frequency domain, the response of a transistor’s thermal path closely
conforms to:
RT 1
HT (ω) = , (15.4)
(1 − j ω/ωT ) (1 − j ω/ω0 )1−nT
nT
where nT is the order of the response, which will be near to or less than 0.5, and
RT [K/W] is the thermal resistance of the thermal path, which is the inverse of its thermal
conductivity. The characteristic frequency of the response is ωT , which is the inverse of
the time constant of the whole path from the channel to ambient. This path includes the
total mass and thermal conductivity of the device, so the characteristic frequency is in
the relatively low Hz to kHz range. The channel region where heat is generated has a
finite, albeit small, size that has a characteristic frequency, ωo in (15.4), in the order of
10 to 100 GHz. Above this frequency, a first-order response is appropriate because the
heat source is distributed throughout this region. The frequency response, HT (ω), and
impulse response hT (t), form a Fourier transform pair.
Since all aspects of the dynamics of transistors depend on temperature, the character-
ization and simulation of temperature variation with time is important. At any instant in
time, there is an instantaneous channel temperature, which is a function of the history
of power dissipation. In the steady state, the temperature rise in (15.1) is the product of
the power dissipation and the thermal resistance, RT (that is, HT (0)). The reduction in
drain current is then given by iD = iDO (1 − λ RT iD vDS ).
The dynamic response for time-varying current and voltage is a function of the time-
varying temperature rise, which is a convolution of the time-varying power dissipation
with the thermal impulse response, (15.2). Although the thermal response can then be
inferred from a transient measurement of drain current, there will be some ambiguity
because the initial thermal response in the channel region is too fast for 100 ns pulse
equipment. The response can be more readily analyzed and measured in the frequency
domain in terms of the characteristic frequencies and order, nT , of the response [3, 4].
This approach to the extraction of heating parameters is discussed later in Section 15.7.3.
15.4.2 Charge trapping

Within the structure of transistors there are regions, or sites, that trap charge in mid-band
energy states [5]. Mid-band states are always present at surfaces and interfaces and can be
included in bulk regions to control or pin certain process parameters. Rather than moving
between valence and conduction band, charges can move to the mid-band state. Once in
the mid-band state, further movement is delayed, so the charge is trapped temporarily.
Deep-level states are those close to the middle of the semiconductor band gap and these
trap charges for longer time periods.
The classic manifestation of charge trapping is gate lag. This is an additional rise in
drain current that occurs a few milliseconds after stepping the gate potential to turn-on.
The size of the current increase and the delay vary with the initial bias and the destination
of the gate step.Variations over several orders of magnitude in response time are observed,
but the lag is a first-order response, that is a response over one or two decades of time. The
lag can be extremely long, with tens of minutes being typical in devices fabricated with
wide band-gate materials such as Gallium Nitride. Trapping in passivation layers can
be responsible for long-term alteration of transistor characteristics. In high electric field
conditions, such as at high drain-source potentials, the lag can be faster than the 100 ns
resolution of pulse measurements. This is dealt with in more detail in Section 15.5.
Because the trapping process imposes an inherent delay, its influence on drain current
is a function of the past bias conditions. This is a memory effect with a bias and frequency
dependence related to the occupation and charging rates of the trap centers [6–8].
Trap rates
The extent and period of trapping is well described by capture and emission processes in
terms of carrier concentrations and energy bands in the semiconductor [9]. Drawing an
electric circuit analogy of a trap center provides a description that is readily understood by
engineers working with FET circuits and that can be implemented in a circuit simulator
[10, 11].
Charge in a trap center is analogous to charge in a capacitor, CT [F]. The ionization
potential of the trap, vT [V], is analogous to the potential across the capacitor. The ion-
ization potential is always restricted between zero, for neutral charge, and the potential,
VO [V], of the fully depleted trap. The latter is positive or negative depending on the
ionization polarity of the trap.
The capacitor representing the trap center is charged by a nonlinear controlled current
source given by
: qv ;
I
iT = ωe (T ) CT VO − vT − vT exp , (15.5)
kT
where q [C] is the electron charge and vI [V] is a control voltage that accounts for the
change in Fermi level due to electric fields [4], and

ET
ωe (T ) = AT T 2 exp − , (15.6)
kT
where AT T 2 [s−1 ] is an Arrhenius factor, k [eV/K] is the Boltzmann constant, and ET

[eV] is the trap’s activation energy.
In the steady state with zero net current, the characteristic frequency for the trapping
process is
ET : qv ;
I
ωE (vI , iJ , T ) = AT T exp −
2
1 + exp . (15.7)
kT kT
422 Anthony Parker
Control of the trap by transistor terminal currents is by vI through appropriate

functions.
Three characteristics of trapping can be noted. The first is that the trapping response
has a single time-constant for fixed vI . The second is that the time constant is set by
the value of the target vI and is independent of the initial bias. The third is that tem-
perature significantly affects the trap response, so trapping effects include evidence of a
simultaneous heating response.
15.4.3 Impact ionization

Observable impact ionization in a FET requires a sufficiently strong electric field, which is
governed by the drain-source potential. Avalanche breakdown can result when a cascade
of ionizations occurs where the additional carriers go on to generate even more.
Experiments show a reasonably logarithmic relationship between the impact ionization
rate and the inverse of electric field strength [5]. This experimental evidence is the basis
for the following expression for impact ionization rate, RI , with slope and intercept
corresponding to two fitting parameters A and B [V/m]:
RI (F ) = A e−B/F , (15.8)
where F is the electric field strength [V/m]. Increased temperature reduces the impact
ionization rate relative to the electric field [12, 13].
In terms of the electron current from the source, the total current is thus:
iDS
iD = . (15.9)
1 − RI
RI
The hole current returning toward the source is iD RI = iDS 1−R I
and some fraction
of this will tunnel to the gate or surface. Tunneling is more probable when the gate bias
is negative, which is attractive to the holes, but reduces exponentially with increasing
gate potential [14]. The measured gate current from impact ionization, as shown in
Figure 15.5, increases as the drain current increases, but varies with the gate potential
because the tunneling probability varies.
The hole tunneling is easily measured with low-frequency semiconductor parameter
analyzers. Pulse system, which has less dynamic range, only detects high levels asso-
ciated with breakdown. However, the kink in the drain current characteristics is quite
distinct in pulse measurements and has a response time consistent with a trapping process.
Measurement of the kink effect

A more substantial increase in drain current is produced by the positive potential of
accumulated holes that have reached the surface, usually at the source end of the channel.
Those that occupy trap states cause a response according to trap occupation and charging
rates and the polarity of the trap. These rates are observed to increase dramatically
with drain potential, such that the effect is much faster than even the shortest pulse
measurements [15].
1A
Drain current
10 mA
VGS = −1.5 V
100 μA
Current
1 μA
VGS = −1.5 V Gate current
10 nA
VGS = −1.0 V
100 pA VGS = 0.0 V

Impact ionization model
0 1 2 3 4 5 6
Drain potential (V)
Fig. 15.5 Gate and drain DC characteristics of a typical pHEMT. Hole current generated by impact
ionization is clearly evident in the gate current. The dashed line is the product of drain current
and (15.8).
Thus, the increased drain current comes from positive feedback via a field effect. The
transconductance of the FET amplifies the effect of a relatively low trapping potential,
so the observed drain current increase is significantly greater than the contribution from
ionization alone [6].
At higher power bias points the traps respond at increasingly faster rates. The effect is
slow at modest drain-source potentials, so pulse measurements are able to observe a rise
in drain current as the traps ionize. The position of the kink in the drain current moves
to higher drain potentials as the width of the measurement pulses reduces, as shown in
Figure 15.1. This is because the kink is centered at the drain potential where the trapping
rate is comparable to that of the pulse length. That is, the traps have time to ionize on
the high side of the kink, but not on the low side.
15.5 Transient measurements
The typical use of pulse characterization is to avoid dynamic processes in the transistor
rather than to analyze them. This overlooks considerable information in a complete step
response over a longer time period. The picture is completed by repeated measurements
from different biases, which can be illustrated in an investigation of gate lag.
15.5.1 Measurement of gate lag

Gate lag is a delayed additional rise in drain current that occurs a short time after a FET
is turned on, as discussed in Section 15.4.2. A measurement of gate lag is shown in
424 Anthony Parker
From..
150
VGS = −0.4 V
145 VGS =−0.8 V

Drain current (mA)
140 VGS =−1.2 V
135 VGS =−1.6 V
130
VGS =−2.0 V
125
1 μs 10 μs 100 μs 1 ms 10 ms
Time after turn-on step
Fig. 15.6 Time domain response after stepping to zero gate-source potential, vGS = 0 V, from various
initial gate biases, VGS , as annotated. The drain potential is fixed at VDS = 1.5 V.
Figure 15.6. Here the device is switched to zero gate-source potential without changing
the drain potential. After a few milliseconds the drain current settles to the same value
irrespective of the initial gate potential, which is a reasonable expectation.
When the FET is switched on to zero gate-source potential, the current rises from the
much lower value that it had prior to zero time. The current rise causes heating with
a response over several decades in time, as described in Section 15.4.1. This produces
the reduction in current over several decades of time that is common to all the turn-on
transients.
The current rise at about 300 μs can be explained by a hole trap in the substrate.
Before the transient the negative gate bias prior to the transient injects electrons into the
substrate, which ionizes the trap. The extent of ionization increases with more negative
gate bias, so the height of the gate-lag current rise is proportional to the initial gate
bias. When the gate potential is stepped to zero, the trap potential increases as holes
are captured through a first-order process with time constant set by the FET’s terminal
potentials.
The dependence on initial drain bias is shown in Figure 15.7. Increasing the drain bias
offsets the influence of negative gate potential on the electric fields in the substrate.
The time constant of the gate lag is set by the destination point of the transient. The
variation in timing with destination drain-source potential is shown in Figure 15.8.
A complication in the gate-lag characteristic is the temperature dependence of the
capture process. This increases as the temperature does, so the responses interact. Con-
sequently, the transitions in Figure 15.8 vary from the one-decade rise of a single time
constant process to a faster response when heating is coincident.
150
145
Drain current (mA)
140
From..
VDS = 2.5 V
135
130 VDS = 2.0 V
125 VDS = 1.5 V
1 μs 10 μs 100 μs 1 ms 10 ms
Time after pulse step
Fig. 15.7 Time domain response after stepping to vDS = 1.5 V, vGS = 0 V (the same destination as
Fig. 15.6) from VGS = −2.0 V and various drain potential biases, VDS , as annotated. The curve
for VDS = 1.5 V is the same as the bottom curve in Fig. 15.6.
170 To.. VDS = −0.5 V
160
VDS = −1.0 V
150 VDS = −1.5 V

Drain current (mA)
140 VDS = −2.0 V
130
VDS = −2.5 V
120
110
100
1 μs 10 μs 100 μs 1 ms 10 ms
Time after pulse step
Fig. 15.8 Time domain response after stepping to vGS = 0 V and various vDS from VGS = −2.0 V,
VDS = 1.5 V. The center curve is the same as the bottom curves in Figs 15.7 and 15.6.
To study the trap ionization process, the pulse measurement can be repeated with
varying times at the initial bias [16]. That is, the gate lag is measured as a function of
how long the trap is allowed to ionize. This is an inverse pulsing technique, where the
426 Anthony Parker
bias time between pulses is varied and the pulses are set to a long period to measure the
transient.
15.5.2 Time evolution characteristics

Each bias point about which pulse measurements are made establishes a temperature
and trapping state. There is a unique set of isodynamic characteristics for each state, and
therefore for each bias point. The three sets of pulse measurements corresponding to
three bias points in Figure 15.2 illustrate the dependence of isodynamic characteristics
on the bias state.
Two aspects of pulse and transient characteristics are worth noting. One is that the bias
point needs to be well established prior to an isodynamic measurement. The isodynamic
characteristic will vary depending on how long the bias has been established, that is, the
large-signal characteristics undergo bias evolution. The second is that any pulse is the
start of a step change to a new bias point at the potential of the pulse, so if the pulse width
is extended the measurement will be influenced by the transition to a new bias point.
Transient responses over a set of long pulses produce data that shows a time evolution
of large-signal characteristics.
The concept of bias evolution is that an isodynamic characteristic, seen by a radio-
frequency signal, immediately after turning on a transistor is different to that when
the turn-on bias has settled. For example, in Figure 15.2 if the transistor was off at bias
(–2.0, 4.0) and turned on to bias (–0.5, 4.0), then the isodynamic behavior of the transistor
would initially be close to that of the off-state and then over a finite time it would evolve
to that of the on-state. In this example, the large-signal current swing at radio frequency
would reduce during the microsecond to millisecond switch-on transient. Except in the
case of wide band-gap devices, the timescale of the transient is too short to permit
repeated measurements by conventional pulse systems of isodynamic characteristics at
regular intervals during the turn-on period.
Time evolution of large-signal characteristics shows the transition from a point on an
isodynamic characteristic to the establishment of a new bias condition at that point. Two
sets of time-evolution characteristics are illustrated in Figure 15.9. Each line in these
graphs is a transient response from the initial bias to a new bias [17].
The surfaces provide convincing evidence of dynamic processes in the device. There
are time constants that clearly depend on the destination point and that vary over many
orders-of-magnitude in time. Each set of time-evolution characteristics is strongly linked
to the initial bias condition. Self-heating and trapping related to impact ionization are
prominent in Figure 15.9(a) because the initial bias dissipates no power and the trap state
is at the pinch-off extreme. In Figure 15.9(b), the initial bias condition already establishes
a degree of heating and a moderate trap state.
A notable feature of the data in Figure 15.9 is the dependence of time constants on
the destination drain potential. There is a straight-line relationship between log-time and
potential. This is explored from an alternative small-signal vantage in Section 15.8.
Detailed information about specific dynamics is present in a select set of transient
200 200
150 150
Drain Drain
current 100 current 100
(mA) (mA)
50 50
0 0
100 ms 100 ms
1 ms 1 ms
10 μs 10 μs
1 3 5 7 100 ns 1 3 5 7
Time 100 ns Time
Drain Drain
potential (V) potential (V)
(a) Low bias (b) High bias
Fig. 15.9 Time-evolution large-signal characteristics of two pHEMTs. Each plot shows drain current
versus time after a step from a fixed bias point (large ◦ at 100 ns). There are four surfaces
corresponding to gate-source potential from –1.5 to 0.0 V in 0.5V steps as the parameter.
measurements in the time-evolution data and from the dependence of these on a range
of initial bias conditions.
15.6 Pulsed measurement equipment
There are many pulse measurement systems reported in the literature offering a variety of
options. They can be assembled from individual instruments that provide bias and pulses,
synchronize timing, and measure current and voltage. In advanced systems, pulsed radio
sources and network parameter measurements are incorporated, particularly for mea-
suring high power devices intended for pulsed-radar applications. Pulse techniques also
facilitate nondestructive investigation of transistor breakdown regions [18]. A degree of
sophistication is achieved with arbitrary pulse patterns and arbitrary control of initial
bias and pulse timing [19, 20].
Commercial systems are available that are capable of sub-microsecond pulses. Exam-
ples are Accent Opto’s Dynamic I(V) Analyser [21], Auriga Measurement System’s
Pulse IV/RF System, Focus Microwave’s modular pulse system, and systems by Agi-
lent Technologies, Keithly Instruments, and Amcad Engineering. These systems fill the
instrumentation gap between semiconductor parameter analyzers and nonlinear vec-
tor network analyzers. Typical pulse systems offer a large-signal capability and speeds
sufficient to give near isodynamic characteristics at low drain biases.
15.6.1 System architecture

The functional diagram of a pulsed measurement system, shown in Figure 15.10, includes
both pulsed-I/V and pulsed-RF subsystems. Pulse and bias sources, voltage and current
428 Anthony Parker
Pulse sources
+ Bias source +
A V Current/voltage sampling V A
DUT
Bias networks
VNA
a1 b1 b2
Mixer-based test set
Phase lock
Pulsed RF source LO Source
Fig. 15.10 Simplified diagram of a generic pulsed measurement system. Alternative connections provide
load terminations when there is no pulsed-RF test set.
sampling blocks, and associated timing generators form the pulsed-I/V subsystem. A
pulsed-RF source and mixer-based vector network analyzer form the pulsed-RF sub-
system. The test device is connected directly to the pulsed-I/V subsystem, or to bias
networks that connect the pulsed-RF subsystem or RF terminations.
Pulsed-I/V system
Steady-state semiconductor parameter analyzers provide a source-monitor unit for each
terminal of the test device. The unit sources one of voltage or current while monitoring
the other. In a pulsed measurement system, a pulsed voltage is added to a bias voltage and
applied to the device. A precise measurement grid is rarely obtainable by pulse systems
because of the transient response limitations of pulse equipment. Thus, actual terminal
conditions, both voltage and current must be recorded. It is essential to recognize that the
pulse data do not lie on a regular grid of values, so a naive plot of characteristics curves
can be misleading because each line will not correspond to a constant control potential.
The position of the voltage and current sensors between the pulse source and the test
device is affected by transmission line effects associated with the cabling between the
sensing points. These affect the transient response and performance of the pulse system.
An additional complication is introduced when the test device must be terminated for RF
stability. A bias network is required but this introduces its own transient response to the
measured pulses. The initial 100 ns transient in many pulsed measurements is dominated
by the bias network.
Current can be measured by various methods, which trade between convenience and
pulse performance. Hall-effect/induction probes placed near the test device can sense
terminal current. These probes have excellent common-mode immunity but tend to drift
and add their own transient response to the data. A stable measurement of current is
possible with a series sense resistor. This requires a differential input with very good
common-mode rejection at high frequencies.
Pulsed-RF system
Pulsed-RF test sets employ vector network analyzers with a wideband intermediate
frequency (IF) receiver and an external sample trigger [20, 22]. The systems need two
RF sources and a mixer-based S-parameter test set. One source provides a continuous
phase reference for the mixers and samplers, while the other provides a pulse-modulated
RF output.
The pulsed bias must be delivered through bias networks. During a pulsed-I/V mea-
surement, the RF source is disabled and the RF test set provides terminations for the test
device. Pulsed-RF measurements are made one pulse point at a time. With the pulsed bias
applied, the RF source is gated for a specified period during the pulse and the network
analyzer is triggered to sample the RF signals. The same pulse point is repeated to work
through a required frequency list and averaging setting.
Bias networks
The bias network that connects the test device needs to provide stable high-frequency
termination while passing pulse stimuli. In addition, there need to be current and voltage
measurement ports. The trade-off between these requirements necessarily limits the
maximum rise time of the pulse. If faster than 100 ns pulses are required, then the pulse
source must be connected directly to the test device [23].
An enhanced bias network that allows reasonable length cables to the samplers is
shown in Figure 15.11. The DC-blocking capacitor is reduced, so that it does not draw
current for a significant portion of the pulsed bias. At the same time it provides adequate
passage at RF frequencies. The isolating inductor must be small enough to pass the pulsed
bias while providing adequate RF isolation. In the figure, the DC-blocking capacitor and
isolating inductor values are an order of magnitude smaller than are those in conventional
bias networks. The network provides a good RF path for frequencies above 500 MHz
and does not significantly disturb pulses longer than 100 ns. Modifying the network to
provide a RF path at lower frequencies will disturb longer pulses.
The pulsed bias is fed to the bias network in Figure 15.11 through a cable that introduces
transmission line transients. A snubber is added to control these. The values shown are
suitable for suppressing the 10 ns transients associated with a 1 m cable.
Voltage sampling in Figure 15.11 is through a frequency-compensated network that
provides isolation between the RF path and the cable connected to the voltage sampling
digitiser. Without this isolation, the capacitance of the cable would load the pulsed-bias
waveform, significantly increasing its rise time. The voltage sample point should be as
430 Anthony Parker
Current sampler
Induction probe
Pulse source
via 50 Ω cable
50 Ω 15 nF
50 Ω High Z voltage
Snubber
270 pF sampler via
950 Ω 50 ohm cable
10 pF
70 nH
RF Pulse/Bias/RF
Termination to DUT
Fig. 15.11 An enhanced bias network that allows voltage and current measurement instruments to be
connected via reasonably long cables. The bias network is designed to cut-off at 500 MHz, to
allow pulses through the bias port to the DUT.
close as possible to the DUT to reduce the effect of reflected pulses. The network in this
example sets a practical limit of about 15 cm on the length of the cable connecting the
transistor under test to the bias network.
Induction current probes introduce their own time constants to the measurement that
is visible in the time domain transient record. Current measurement with series sense
resistors ameliorates this, but adds to the output impedance of the pulse source. Usually a
capacitance of a few picofarads is associated with the sense or bias network that restricts
the choice of resistance value for a specified rise time.
Series-resistor sensing requires a floating differential amplifier operating over the
range of pulse potentials. The common-mode gain of the amplifier is higher for short
time intervals, so some of the step change in potential is recorded as a current transient.
Placing a sense resistor in the ground return is an alternative, but the transmission-line
effects of the connection between the pulser and DUT need to be considered.
15.6.2 Timing
The most critical aspect of pulse measurement is sample timing. In many cases the sample
will be gathered at some point in a time-dependent dispersion process, so it is important
to consider the timing relative to the time-constants of these processes. In general, full
information can only be gathered by a time domain pulse-profile measurement. Equally,
the time spent establishing the quiescent condition before the pulse must be long enough
that there are no residual effects from a previous pulse. At least two orders of magnitude
less than the time constants of the dynamic processes is recommended.
Measurement equipment capable of pulsing to points on the I/V-plane in a random
sequence provides a powerful display for verifying isodynamic timing. If the time at
quiescence between pulses is insufficient, then the pulse measurement will be dependent
upon the particular history of previous pulse points. Step-and-sweep sequencing gener-
ates a monotonic change in history, so dynamic effects are not obvious because adjacent
points have similar pulse histories. However, if a random sequence is employed, the adja-
cent points will have different pulse histories and the corresponding effect of dynamic
processes will be visibly different.
Interpolation and iteration

Often, measurements are desired at a particular pulse point or on a regular grid of points.
For a target pulse voltage, the actual voltage at the transistor will usually be different due
to interactions with the pulse amplifier output impedance and amplifier time constants, as
well as cabling and bias network transients. These could be compensated for in advance
with known current, but this current is being measured. This is why pulsed voltages need
to be measured at the same time as the device currents.
If measurements are desired at specific voltage values, then one of two approaches can
be used. Firstly, over successive pulses, the target voltage values can be adjusted to iterate
to the desired value. This necessarily involves a measurement control overhead and can
require considerable time for many points. The second approach is to establish a look-up
table to calibrate the pulse setting to give the target voltage. For true step-response data,
this correction needs to be adjusted throughout the pulse period to compensate for the
transient response of the pulse source [24].
15.7 Broadband RF linearity measurements
The conventional wisdom was to consider dynamic processes to be slow enough to

be irrelevant during the high-frequency operation of transistor circuits. Other than an
anomalous shift of bias conditions, the slow processes do not affect performance at high
enough operating frequencies, which need only be a few tens of MHz. Kinks, hysteresis,
or memory seen in low-frequency and slow-pulse drain characteristics are not evident in
the bias dependence of high-frequency small-signal parameters. Thus, a set of fast pulse
measurements that provide drain characteristics at various bias points can be indicative
of the high-frequency operation. Little information on the nature of slow dynamics is
required to predict high-frequency small-signal parameters.
This is not true for low-signal, or weak nonlinearity scenarios. In particular, when there
are two tones in the signal, any nonlinearity generates distortion products including inter-
modulation products near the signal frequencies. Intermodulation exhibits a significant
variation with bias and this bias dependence changes with the difference between the
frequencies used in a two-tone intermodulation measurement. Contemporary wisdom
attributes this to slow processes in the transistor and low-frequencies impedances in the
circuit [7, 25, 26]. If the difference-frequency between two tones resonates with the char-
acteristic frequencies of the dynamic processes in a transistor, then these processes will
be excited and affect the high-frequency characteristics. For certain bias conditions, the
variations of intermodulation level and asymmetry between the upper and lower products
432 Anthony Parker
can be several orders of magnitude. Thus generation of distortion and intermodulation

is a crucial aspect of the slow dynamic processes in a transistor.
The requirement to control linearity at microwave frequencies is probably the most
compelling reason to characterize the full dynamics of heating and trapping processes in
transistor models. At the very least, a priori understanding of these processes is essential
to the correct interpretation of pulse measurements. Even so, the contribution to linearity
often spans frequencies beyond the reach of pulse equipment. It is necessary to draw on
characterization techniques based on intermodulation and high-frequency small-signal
measurements to bridge the gap.
15.7.1 Weakly nonlinear intermodulation

Generation of intermodulation can be analyzed with a weakly nonlinear model, which
is often a small-signal model with nonlinear elements described by a Taylor series. An
expression sufficient to illustrate intermodulation generation by slow dynamic processes
in the transistor is:
2 3
id (vg , vd ) ≈ gm vg + gm vg + gm vg + gmd vg vd , (15.10)
where the signal potentials are vg = vG − VD and vd = vD − VD . The transconductance

= 1 d 2 i , and g = 1 d 3 i . A
and its derivatives are given by gm = dvdG iD , gm 2 dv 2 DG m 6 dv 3 D G
cross-conductance gmd = 12 dvdD dvdG iD is also included because it is key to understanding
the interaction with slow dynamics. All the conductances are evaluated at vG = VG and
vD = VD .
(15.10) is a very simple description that neglects the drain conductance and any terms
higher than third order. The bias drain current ID is neglected, so that (15.10) gives the
signal current, id , in terms of the signal potentials, vg = vG − VD and vd = vD − VD .
A load impedance, ZL , driven by the drain current, as shown in Figure 15.12, develops
a drain signal given by
2 3
vd ≈ −gm vg ZL − gm vg ZL − gm vg ZL − gmd vg vd ZL , (15.11)
where the value of ZL varies with the frequency of each current component.
G D
vg id vd ZL
S S
Fig. 15.12 Simple weakly nonlinear unilateral model of a FET driving a load impedance.
Intrinsic nonlinearity
An intrinsic level of intermodulation is produced by (15.10) when the gate is driven
by a two-tone signal. For the case of tones at frequencies ω1 and ω2 the third-order
intermodulation product will be at frequencies of ω1 − ω and ω2 + ω. Drain currents
v 3 , and by the second-
at these frequencies will be generated by the third-order term, gm g
order term, gmd vg vd , in (15.10) [27, 28].
For the case of two tones close to ω such that ω = ω1 − ω2 ω, the frequen-
cies of the intermodulation currents will also be close to ω, so the drain potential for
the intermodulation products will be proportional to the load impedance at ω, which
is ZL (ω).
The third-order intermodulation current associated with the cross-conductance term,
gmd vg vd , comes from second-order components in vd at ±ω as well as at ±2ω. The
latter is usually not significant in slow-rate dynamics and the former comes from drain
currents at frequencies ±ω generated by the second-order term, gm v 2 Z , in (15.11),
g L
so is proportional to the load impedance at the difference frequency, ZL (±ω).
The amplitude of the drain potential’s third-order intermodulation product, Vd<IMD> ,
in terms of the gate signal amplitude, Vg , is the sum of contributions from the
transconductance and cross-conductance terms:
1

Vd<IMD> ≈ − Vg 3 ZL (ω) gm − gmd gm ZL (±ω) . (15.12)
4
This is an intrinsic level of intermodulation because slow dynamic processes have not
+ g
yet been considered. It is proportional to (gm md gm ZL (±ω)) where ZL (−ω) and
ZL (+ω) affect the upper and lower intermodulation products at ω1 −ω and ω2 +ω,
respectively. There is a conjugate relationship between the load impedances at positive
and negative frequencies, such that ZL (−ω) = ZL ∗ (ω), so there is an asymmetry
between the bracketed terms in (15.12) for the upper and lower intermodulation products,
which is often observed in measurements [25,26, 29,30]. This asymmetry can be removed
if ZL (ω) is real. Also, a suitable choice of bias and load impedance at the difference-
frequency, such that gmd gm Z (ω) = −g , can eliminate the intrinsic intermodulation
L m
altogether. There are useful operating regions where an optimal load can be realized in
practice because gm < 0, as shown in Figure 15.13.
Note that the case above is simplified to illustrate a dominant intermodulation

mechanism. It overlooks other intermediate second-order products that contribute to
intermodulation [31–33].
15.7.2 Intermodulation from self-heating

Self-heating is a response to instantaneous power that has been described in
Section 15.4.1. The power is the product of the drain current and drain potential, iD vD ,
which can be expressed in terms of the bias, ID and VD , and the signal components, id
and vd , to enable separation of bias and instantaneous power terms.
pD = (ID + id ) (VD + vd ) = ID VD + (id VD + vd ID ) + id vd . (15.13)

434 Anthony Parker
100
gm [mA/V]
80 Ids [mA]
Current / conductance
60
40
20
gmd [mA/V2]
0
–20
–40
–1.5 –1 –0.5 0 0.5

Gate bias (V)
Fig. 15.13 Typical conductance and derivatives for a microwave transistor. This example is a model for a
small-signal pHEMT.
In a weakly nonlinear scenario with two closely spaced tones, the signal currents and
potentials will be dominated by the fundamental frequency components with frequencies
near ω separated by ω. The first term in (15.13) is the bias power component, which
influences the bias temperature. The second, bracketed, term in (15.13) will be dominated
by fundamental components with frequencies near ω. In most radio applications this
frequency will be high, so there will be little, though not necessarily negligible, thermal
response to this power component.
The last term in (15.13) is a product of fundamental tones, so will have components
at frequencies ±ω and ±2ω. The latter is likely to be too high to excite a significant
thermal response. However, for closely spaced tones, the former is capable of producing
a significant thermal response.
The component of drain current that is directly affected by the thermal response
through the self-heating mechanism is −id λ pD HT , which is derived from (15.1) and
(15.2). Third-order intermodulation in the self-heating current comes from the product
of the fundamental components in id with the difference-frequency components in pD
(from the last term in (15.13)). Expressing drain current and potential in terms of gate
potential and load impedance gives the amplitude of the self-heating contribution to
third-order intermodulation:
1
<IMD>
vD ≈ − Vg 3 gm 3 ZL2 (ω) HT (±ω), (15.14)
4
where HT (−ω) and HT (+ω) affect the upper and lower intermodulation products,
respectively. The intermodulation product given by (15.14) is not negligible and can be
easily observed at bias and load conditions that reduce the intrinsic nonlinearity.
15.7.3 Measuring heating response

The level of the thermal contribution to intermodulation relative to difference frequency
mirrors the thermal response of the transistor. That is, vD<IMD> versus ω is proportional
to HT (ω) given by (15.4). However, to observe the thermal response in a linearity

measurement, other dynamic processes need to be reduced with a suitable bias condition
and frequency-independent load. If the load is not constant then its frequency dependence
can mask that of other slow-rate dynamics.
Intrinsic intermodulation, (15.12), varies with ZL (ω) and the self-heating intermod-
ulation (15.14), varies with HT (ω). Presenting a constant-impedance load to the drain
holds the intrinsic contribution constant, so any observed variation with frequency is
from heating or other transistor dynamics.
The load shown in Figure 15.14 presents a 50 load for all frequencies when
ZL = 50 and L = 100 C. The capacitance should be large enough to allow the spec-
trum analyzer, ZL , to observe the lowest frequency of interest; approximately 5 nF for
100 MHz. This is the frequency of the intermodulation product, which is high, and not
the low difference-frequency.
A measurement of intermodulation generated by self-heating is shown in Figure 15.15
for a typical pHEMT. The transistor is biased near the zero crossing of gm , which occurs
in the region of pinch-off. This zero crossing is illustrated in Figure 15.13. The intermod-
ulation measurement that reveals the thermal response is found by choosing a bias that
gives a minimal level at high difference-frequencies [3]. The high difference-frequency
point has the lowest thermal contribution because HT (ω) is small, so a bias that min-
imizes this point is one where the other dynamic processes are cancelled out. As the
difference frequency is reduced, the thermal contribution increases.
In the example shown, the characteristic frequency of thermal response is 2.7 kHz
and the order of the response is near 0.5, or only 10 dB per decade. The implication of
this low order is that the magnitude of the heating contribution at 1 GHz is reduced by
only 60 dB, which is not necessarily a negligible level. This dynamic process compli-
cates the linearity of the broadband circuit that deals with a wide range of difference
frequencies.
L R
DUT VDD
Z ZL
Fig. 15.14 Constant impedance bias network for measurement of slow dynamics within the DUT. The load
presented to the drain is Z = ZL = 50 for all frequencies when L = 100 C.
436 Anthony Parker
Intermodulation level (dBc) –25
VGS =−2.14 V
–30
VGS = −2.15 V
–35
VGS = −2.16 V
–40 HT (Δω)
–45
VGS = −2.18 V
VGS = −2.17 V
–50
100 Hz 1 kHz 10 kHz 100 kHz 1 MHz 10 MHz
Difference frequency (Δω)
Fig. 15.15 Two-tone intermodulation measurement centered at 200 MHz of a typical pHEMT for
VDS = 2.5 V. A heating response, HT , given by (15.4) is shown for nT = 0.5 and
ωT = 2π × 2700 Hz. The load in Fig. 15.14 was used. Charge trapping accounts for the
inflection near 80 kHz.
There is another dynamic process observed in Figure 15.15 as an inflection near

80 kHz. This process has a first-order response and a frequency that varies with drain
bias, which would be expected if it were linked to a charge trapping process.
15.7.4 Measuring charge trapping response

Charge trapping provides a feedback path that can generate intermodulation in a mecha-
nism similar to that for self-heating. The ionization potential of trap sites imparts an addi-
tional control over the channel current, which is influenced by drain and gate potentials.
Since the trapping is a first-order low-pass phenomenon with characteristic frequency
given by (15.7), it will respond readily to difference frequencies in the drain signal.
In principle, if other mechanisms are reduced, it is possible to observe a trap’s fre-
quency response, HE (ω), in the variation of intermodulation level with difference
frequency. A significant variation of the frequency of this response with bias is a
clear identifying feature of a trapping mechanism. However, there are two issues that
complicate this measurement.
Measurement of trapping intermodulation is an issue at biases where the trap is either
fully ionized, vT ≈ Vo , or at neutral charge, vT ≈ 0. In these regions the trap cannot
become more ionized than fully ionized or more neutral than neutral, so there is little
change in trap potential to contribute to intermodulation. However, there is substantial
trap related gate lag and hysteresis when the trap state is pulsed to and from these regions.
There is also an issue with the temperature of the trap site. Although the dependence
of ωE on the trap control potential is exponential for vI > 0, this does not account for
the frequency dependence of the intermodulation process because the trap potential is
near zero in this region. Rather, the bias dependence of ωE is mainly due to increasing
temperature as power dissipation increases. As given by (15.7), the variation of ωE with
temperature is significant.
Pulse and transient step responses can be more suited to the characterization of trap-
ping. Transient responses reveal trapping in regions where there is little trap-related
intermodulation. The transients also include the effect of temperature change varying
the trapping rate. This produces a variation with bias of the apparent order of the response,
which can range from an order of one-half order if heating is coincident to a first-order
response of trapping alone.
15.7.5 Measurement of impact ionization

Impact ionization works in conjunction with charge trapping to generate intermodulation
that is highly dependent on bias. The response follows the combined transfer function of
the first-order trap site and the second-order nonlinearity of impact ionization, derived
from (15.8). This transfer function is proportional to the impact ionization rate, so the
associated intermodulation contribution is only generated if the drain-source bias poten-
tial is high enough. It is a first-order response that depends on temperature and bias, so
ωI increases at a rate of about one decade per volt as drain bias increases. At moderate
drain potentials, such as vDS = 7 to 10 V, the response can be in the order of 1 to 10 GHz.
Clear evidence of impact ionization is seen in the striking variation of intermodulation
with difference frequency and bias shown in Figure 15.16. To observe this, a bias in a
–40 VDS = 1.75 V
VDS = 1.25 V
Intermodulation level (dBc)
VDS = 1.50 V VDS = 1.50 V

–50
VDS = 1.25 V
–60
VDS = 3.75 V
VDS = 1.75 to 3.25 V
VDS = 3.50 V
–70
100 Hz 10 kHz 1 MHz 100 MHz
Difference frequency (Δω)
Fig. 15.16 Two-tone intermodulation centered at 500 MHz for a FET in common-source configuration with
VGS = 0 V. Note that the waterfall effect does not start till the drain-source potential reaches
1.75 V.
438 Anthony Parker
region of low gm is required to reduce the intrinsic nonlinearity, such as a gate bias near
zero as shown in Figure 15.13. This bias also produces substantial drain current, which is
a prerequisite for impact ionization, while permitting hole tunneling to the surface states.
As the difference-frequency increases in Figure 15.16, the intermodulation reduces
dramatically. This occurs at this gate bias because the intrinsic nonlinearity is canceled.
This is a first-order response, which is consistent with charge trapping. The rate increases
with drain potential at about one decade per volt due to the increased temperature, impact
ionization rate, and reduced trap ionization. The rate of about 10 kHz at the onset of impact
ionization (VDS = 1.75 V) is slow enough to be observed in pulse measurements, such
as in Figure 15.1. There is significantly less variation in intermodulation at low drain
potentials in Figure 15.16 because the impact ionization rate is negligible.
15.8 Further investigation
Dynamic processes become faster with increasing drain bias, quickly falling to nanosec-
ond scales. This renders pulsed techniques at fractions of a microsecond too slow to
capture isodynamic characteristics. However, radio frequency measurements can bridge
this gap.
Small-signal radio-frequency measurements cover a wide spectrum to more than
100 GHz. Network analyzers can provide small-signal isodynamic current and charge
storage characteristics routinely. Extracting intrinsic gain from these measurements
clearly reveals the frequency response of the dynamics processes in the FET [34].
Small-signal intrinsic gain, Ai , is voltage gain into an open-circuit load. It is easily
calculated from measured network parameters:
y21 2s21
Ai = = . (15.15)
y22 s22 + s22 s11 − s11 − s12 s21 − 1
Surfaces of intrinsic gain versus frequency and bias are shown in Figure 15.17. There
are features in these surfaces that can be correlated with pulse data, such as the time-
evolution responses in Figure 15.9. The time-evolution starts at around 100 ns, which is
a frequency resolution of only 1.6 MHz, whereas the small-signal parameters continue
to vary for a further four or five decades.
Covering the wide spectral range of the intrinsic gain often requires more than one
instrument. The low-frequency data can be measured with a low-frequency analyzer or
similar test fixture [35], or could be derived from pulse data.
With an established correlation between the time-evolution response and intrinsic
gain, it is possible to extrapolate the characteristics of dynamic processes to higher
frequencies [15].
The frequency-independence of intrinsic gain above 1 GHz in Figure 15.17, suggests
that the response is isodynamic above this frequency. The isodynamic region falls to
around 1 MHz at low drain bias potentials, which is well in reach of pulse measurement.
Above three volts, significantly higher frequencies are required for isodynamic charac-
terization. This is evident in the time-evolution data of Figure 15.9, where the transients
20 20
Intrinsic gain
Intrinsic gain
15 15
10 10
5 5
100 GHz 100 GHz

1 GHz 1 GHz
1 MHz 1 MHz
1 kHz 4 5 1 kHz 4 5
2 3 2 3
Frequency 10 Hz 1 Frequency 10 Hz 1
Drain bias (V) Drain bias (V)
(a) VGS = 0.0 V (b) VGS = –1.0 V
Fig. 15.17 Measured small-signal intrinsic gain surfaces versus frequency and drain-source bias for two
gate-source biases. This is a pHEMT similar to that used for Fig. 15.9.
at high drain bias clearly start from higher points at times faster than those that were
measured. At even higher drain potentials, the isodynamic region is pushed to beyond
10 GHz. This can have an impact on radio-frequency operation at these bias conditions.
A drop in gain at low frequencies is the most obvious feature of the intrinsic gain
surface. This is caused by charge trapping and heating. Substrate trapping at low drain
biases overlaps with impact ionization at higher drain biases. The kink in the time-
evolution data at around three volts correlates to the fall in gain at around 1 MHz. This
becomes exponentially faster with increasing drain bias as the emission rate increases
with temperature and capture rates increase.
The peak at one volt comes from an interaction between substrate trapping and self-
heating. These shift the knee of the drain current characteristic, which increases the drain
conductance. A drop in gain occurs at extremely low frequencies, evident at 10 Hz in
Figure 15.17(b), because the knee walks out to high drain potentials. In wideband-gap
devices, the time constants can be so long that a significant reduction in gain is sustained
for a very long period.
References
[1] A. E. Parker, J. G. Rathmell, and J. B. Scott, “Pulsed measurements,” ch. 8, in The RF

and Microwave Handbook, 2nd edn., ser. Electrical Engineering Handbook, M. Golio, Ed.
New York, USA: CRC Press LLC, 2008, 40, pp. 8.1–8.30.
[2] A. E. Parker, J. G. Rathmell, and J. B. Scott, “Pulsed measurements,” ch. 6, in The RF
and Microwave Handbook, 1st edn., ser. Electrical Engineering Handbook, M. Golio, Ed.
New York, USA: CRC Press LLC, 2000, 22, pp. 4–68–4–95.
[3] A. E. Parker and J. G. Rathmell, “Broad-band characterization of FET self-heating,” IEEE
Trans. Microw. Theory Tech., 53, no. 7, pp. 2424–2429, July 2005.
440 Anthony Parker
[4] A. E. Parker and J. G. Rathmell, “Comprehensive model of microwave FET electro-thermal

and trapping dynamics,” in Workshop on Applications in Radio Science, ser. Commission D,
G. James, Ed., URSI. Gold Coast, QLD, Australia: National Committee for Radio Science,
10–12 Feb. 2008, pp. 1–9, http://www.ncrs.org.au/wars/wars2008/sessions.htm.
[5] S. M. Sze, Physics of Semiconductor Devices, 3rd edn. New York, USA: Wiley InterScience,
2006.
[6] M. H. Somerville, A. Ernst, and J. A. del Alamo, “A physical model for the kink effect
in InAlAs/InGaAs HEMT’s,” IEEE Trans. Electron Devices, 47, no. 5, pp. 993–930, May
2000.
[7] J. Brinkhoff and A. E. Parker, “Charge trapping and intermodulation in HEMTs,” in IEEE
MTT-S IMS Digest, K. Varian, Ed. Fort Worth, Texas, USA: IEEE, Inc., New York, NY, USA,
6–11 June 2004, pp. 799–802.
[8] A. E. Parker and J. G. Rathmell, “Contribution of self heating to intermodulation in FETs,”
in IEEE MTT-S IMS Digest, K. Varian, Ed. Fort Worth, Texas, USA: IEEE, Inc., New York,
NY, USA, 6–11 June 2004, pp. 803–807.
[9] W. Shockley and W. T. Read, “Statistics of the recombinations of holes and electrons,” Physics
Review, 87, no. 5, pp. 835–842, 1 Sept. 1952.
[10] C.-T. Sha, “The equivalent circuit model in solid-state electronics—Part 1: The single energy
level defect centers,” Proc. IEEE, 55, no. 5, pp. 654–671, May 1967.
[11] J. G. Rathmell and A. E. Parker, “Circuit implementation of a theoretical model of trap centres
in GaAs and GaN devices,” in Microelectronics: Design, Technology, and Packaging III, ser.
Proceedings of SPIE, V. K. V. Alex J. Hariz, Ed., 6798, Canberra, Australia, 4–7 Dec. 2007,
pp. 67 980R (1–11).
[12] M. Yee, W. K. Ng, J. P. R. David, P. A. Houston, C. H. Tan, and A. Krysa, “Temperature
dependence of breakdown and avalanche multiplication in In0:53 Ga0:47As diodes and het-
erojunction bipolar transistors,” IEEE Trans. Electron Devices, 50, no. 10, pp. 2021–2026,
Oct. 2003.
[13] C. Groves, R. Ghin, J. P. R. David, and G. J. Rees, “Temperature dependence of impact
ionization in GaAs,” IEEE Trans. Electron Devices, 50, no. 10, pp. 2027–2031, Oct. 2003.
[14] R. T. Webster, S. Wu, and A. F. M. Anwar, “Impact ionization in InAlAs/InGaAs/InAlAs
HEMTs,” IEEE Electron Device Lett., 21, no. 5, pp. 193–195, May 2000.
[15] A. E. Parker and J. G. Rathmell, “Bias and frequency dependence of FET characteristics,”
IEEE Trans. Microw. Theory Tech., 51, no. 2, pp. 588–592, Feb. 2003.
[16] J. G. Rathmell and A. E. Parker, “Characterization and modeling of substrate trapping in
HEMTs,” in Proceedings of European Microwave Integrated Circuits Conference, K. Beilen-
hoff, Ed. Munich, Germany: The European Microwave Association, 8–10 Oct. 2007,
pp. 64–67.
[17] A. E. Parker and J. G. Rathmell, “Measurement and characterization of HEMT dynamics,”
IEEE Trans. Microw. Theory Tech., 49, no. 11, pp. 2105–2111, Nov. 2001.
[18] J. Teyssier, J. Viaud, and R. Quere, “A new nonlinear I(V) model for FET devices including
breakdown effects,” IEEE Trans. Microw. and Guid. Wave Lett., 4, no. 4, pp. 104–106, Apr
1994.
[19] J. B. Scott, J. G. Rathmell, A. E. Parker, and M. M. Sayed, “Pulsed device measurements and
applications,” IEEE Trans. Microw. Theory Tech., 44, no. 12, pp. 2718–2723, Dec. 1996.
[20] J.-P. Teyssier, P. Bouysse, Z. Ouarch, D. Barataud, T. Peyretaillade, and R. Quere,
“40-GHz/150-ns versatile pulsed measurement system for microwave transistor isothermal
characterization,” IEEE Trans. Microw. Theory Tech., 46, no. 12, pp. 2043–2052, Dec. 1998.
[21] C. P. II Baylis, L. Dunleavy, and J. Daniel, “Direct measurement of thermal circuit parameters
using pulsed IV and the normalized difference unit,” Microwave Symposium Digest, 2004
IEEE MTT-S International, 2, pp. 1233–1236, Vol.2, June 2004.
[22] J. Scott, M. Sayed, P. Schmitz, and A. E. Parker, “Pulsed-bias/pulsed-RF device measure-
ment system requirements,” in 24th European Microwave Conference, 1. Cannes, France:
European Microwave Association, London, 5–8 Sept. 1994, pp. 951–961.
[23] A. N. Ernst, M. H. Somerville, and J. A. del Alamo, “Dynamics of the kink effect in
InAlAs/InGAs HEMTs,” IEEE Electron Device Lett., 18, no. 12, pp. 613–615, Dec. 1997.
[24] S. A. Albahrani and A. E. Parker, “Impact of the pulse-amplifier slew-rate on the pulsed-IV
measurement of GaN HEMTs,” in 75th Microwave Measurements Conference (ARFTG),
J. Wood, Ed. Anaheim, CA: IEEE Inc, 28 May 1994, pp. 1–7.
[25] N. B. de Carvalho and J. C. Pedro, “A comprehensive explanation of distortion sideband
asymmetries,” IEEE Trans. Microw. Theory Tech., 50, no. 9, pp. 2090–2101, September
2002.
[26] J. Brinkhoff and A. E. Parker, “Effect of baseband impedance on FET intermodulation,” IEEE
Trans. Microw. Theory Tech., 51, no. 3, pp. 1045–1051, Mar. 2003.
[27] S. A. Maas, “How to model intermodulation distortion,” in IEEE Trans. Microw. Theory
Tech., 1, Boston, MA, 1991, pp. 149–151.
[28] S. A. Maas, Nonlinear Microwave and RF Circuits, 2nd edn. Norwood, MA: Artech House,
2003.
[29] G. Passiopoulos, D. R. Webster, A. E. Parker, D. G. Haigh, and I. D. Robertson, “Effect
of bias and load on MESFET nonlinear characteristics,” Electronic Letters, 32, no. 8,
pp. 741–742, 11 Apr. 1996.
[30] D. R. Webster, A. E. Parker, D. G. Haigh, and J. B. Scott, “Effect of circuit parameters and
topology on intermodulation in MESFET circuits,” in IEEE GaAs IC Symposium Digest. San
Jose, CA: IEEE, UK, 10–13 Oct. 1993, pp. 255–258.
[31] G. Qu and A. E. Parker, “Validation of a new HEMT model by intermodulation charac-
terization,” in IEEE MTT-S IMS Digest, ser. 1998 IEEE MTT-S International Microwave
Symposium Digest, R. Meixner, Ed., 2. Baltimore, Maryland, USA: IEEE, Inc., New York,
NY, USA, 7–12 June 1998, pp. 745–748.
[32] G. Qu and A. E. Parker, “New model extraction for predicting distortion in HEMT and
MESFET circuits,” IEEE Trans. Microw. Guid. Wave Lett., 9, no. 9, pp. 363–365, Sept. 1999.
[33] A. E. Parker and G. Qu, “Intermodulation nulling in HEMT common source amplifiers,”
IEEE Microw. Wireless Compon. Lett., 11, no. 3, pp. 109–111, Mar. 2001.
[34] A. E. Parker and J. G. Rathmell, “Novel technique for determining bias, temperature and
frequency dependence of FET characteristics,” in IEEE MTT-S IMS Digest, ser. 2002
IEEE MTT-S International Microwave Symposium Digest, R. Hamilton, Ed., 2. Seattle,
Washington, USA: IEEE, Inc., New York, NY, USA, 2–7 June 2002, pp. 993–996.
[35] J. Brinkhoff and A. E. Parker, “Device characterization for distortion prediction including
memory effects,” IEEE Microw. Wireless Compon. Lett., 12, no. 3, pp. 171–173, Mar. 2005.
Index
abrupt transition, 28, 29 circuits, 21, 23

accuracy, 221, 224 CIS timebase
active load-pull, 347, 350, 358 advantages, 285
adapter disadvantages, 286
cascaded effective efficiency, 150 operation, 284
measuring power through, 150–151 schematic, 283
uncertainty, 153 closed loop active load-pull, 352–353, 358
adjacent channel power (ACP), 93, 95 closed-loop topologies, 392, 394–395
aging, 42, 43 coaxial line, 18
aliasing, 333 coherent interleaved sampling, see CIS timebase
alignment marks, 37 cold FET, 106, 108
all-sideband noise figure, 267–268 cold-source, 246, 249–251, 270–272
aluminum pads, 36 cold temperature, 247
amplifier, 45, 53, 57 color mapping, 87
linearity, 347 comb generator, 372
amplitude, 44, 45, 51, 56, 58, 59, 60, 61 common-mode load, 367
analog-to-digital converters (ADCs), 64, 65, 78, common mode S-parameters, 230
85, 89 “compact” circuit model, 22
ArbitraryWaveform Generator (AWG), 395 complete reflectometer, 225
automated tuners, 348 complex measurement set-up, 345
Automatic Level Control (ALC), 56, 59, 106 compression point, 403
Automatic Network Analyzer (ANA), 98 conductor losses, 33
available gain, 243 conversion, 61
average power, 133 efficiency, 116
converted LO phase noise, 117
coplanar waveguide, 19, 28
bandwidth, 166
mode, 28
baseband (BB) mixers, 390
couplerf
bias-evolution, 426
directivity, 12
bias networks, 428, 429–430, 435
isolation, 12
block diagram, 48, 53
crosstalk, 27–28, 220, 224
bolometric detectors, 135, 136, 137–138
bond wire, 27
DC characteristics, 417, 418
bypass capacitors, 30
DC substituted power, 138
decibel, 132
calibration, 22, 24, 33, 34, 36–37, 207, 220, de-embedding, 30, 33, 38, 220, 223
221–224, 248, 250 delay, 24
factor, 144 device models, 22
standards, 199, 224 device repeatability, 153–154
calorimetric equivalence correction, 146 devices, 21, 22
carrier, 43, 46, 61 device under test (DUT), 21, 22, 196
characteristic impedance, 4–6 DFT frames, 71–73, 76–79, 81–83, 85, 86, 92
characteristics, 142 dielectric loss, 25
charge trapping, 420–422, 436–439 differential circuit, 32
Index 443
differential device characterization, 365 Fast Fourier Transform (FFT), 70, 84

differential load, 367 field-programmable gate arrays (FPGAs),
differential load-pull, 365 186–189
differential S-parameters, 230 figure of merit, 1 dB compression point, 311, 312
digital applications, 219 filter
digitally pre-distorted (DPD) amplifiers, 95, 96 channels, 49
Digital Phosphor (DPX), 86–88, 95 loop, 53, 60
Digital Phosphor Spectrum, 66, 86 lowpass, 43, 50
digital signal processing (DSP), 64, 65, 66, 90, 92 PLL, 53, 60
digital-to-analog converter (DAC), 50, 51, 52, 56, 57 switched, 43, 47, 61
diode, 45, 61 tunable, 43
diode-based power sensors, 135 YIG, 61
diode detectors, 139–142 fixture performance measures
diode type sensors, 136 crosstalk, 27–28
direct comparison power system, 154–156 delay, 24
direct digital synthesizer (DDS), 50, 51, 52, electromagnetic discontinuity, 29
55, 105 loss, 24–25
directional bridge, 109 mismatch, 25–27
directional coupler directivity, 200 multiple-modes, 28–29
directional couplers, 11–12, 109, 200, 221, follower mode, 105
225, 230 Fourier Transforms, 65, 73–76, 78, 85
directions, of current, 29, 181 Fractional-N structures, 105
directivity, 34, 109 frequency
Discrete Fourier Transform (DFT), 66, 69–73, 83, 84 accuracy, 42, 57
dispersion, 24 adjustment, 43
display frame, 87 bandwidth, 59
distortion, signal, 24 coverage, 48
distributed power system, 31 deviation, 59
doubler, 43, 44 difference, 42
double-sideband noise figure, 267–268 divider, 48, 52, 55
double-sideband noise temperature, 269 division, 49, 52
DPX density trigger, 92 domain, 46, 47, 51
DPX®Spectrum, 66, 85 doubler, 43, 44
dry load calorimeter, 146–147 drift, 42
DUT performance, 345 error, 42, 57
dynamic processes, 417, 418, 419, 423, 426, fluctuations, 44
427–431 increments, 57
dynamic range, 73, 74, 119 instability, 44
calculation, 295 mixer, 48, 51
definition, 294 mixing, 54, 55
equation, 298 modulation, 59
implications, 298 multiplier, 41
offset, 46, 53
effective efficiency, concept of, 143, 144 plan, 49
Eisenhart, R. L., 29, 34 range, 42, 52, 55, 56, 57, 60, 61
electrical delays, 386 resolution, 42, 48, 52, 53, 55, 62
electrical models, 199 source, 54
electromagnetic discontinuity, 29 step, 49, 50, 53, 55
electromechanical switches, 108 sweep, 41, 57
enhanced line-reflect-reflect-match (eLRRM), 37 synthesizer, 41, 42, 52, 62
error box, 197 tuning, 50, 53
error coefficients, 198, 223 frequency bin, 398
error correction, 195 frequency mask trigger (FMT), 88, 90–94
error model, 196, 198, 223 Friis formula, 246, 252
estimation, of lumped parasitic values, 25–27 full leaky multiport VNA, 224
excess noise ratio (ENR), 248 fundamental mixer, 115, 389
444 Index
gain variations, 55 load reflection coefficients, 346

gallium nitride, 418, 421 lock
gate lag, 421, 423–426, 436 frequency, 52
Gauss instruments, 68 indicator, 41
golden diode, 370, 373 logic analyzer, 93–94
graphical system design, 189 looking, 135
ground, 30, 32 loop filter bandwidth, 53, 60
ground plane, 32 loop oscillations, 353–354
ground-signal-ground (GSG), 36 loss, 24–25, 33, 59
conductor, 33
harmonic conversion, 115 dielectric, 25
harmonic load-pull, 347, 349–350 skin-effect, 25
harmonic mixer, 389 lossless transmissions lines, 6
harmonics, 43, 44, 49, 53 LO waveform modification, 115
heterodyne, 386 low-frequency bridges, 404
hot temperature, 247 LRM calibration, 211–215
housing, 41 lumped element, 22
hybrid circuits, 34
hybrid coupler, 232
manual tuners, 348
Mason’s gain, 27
image responses, 116 matching network, 345
IMD products, 405 maximum output power, 346
impact ionization, 422, 426, 437–439 maximum power, 149
impedance, 27, 56 measurement accuracy, 33–34
incident power, 132 measurements, 134
injection power, 401 measurement time, 171, 172
insertion gain, 249, 252 memory effects, 95, 96, 406, 415, 421, 431
in situ test, 37 microcalorimeter, 145–146
instrumentation standards, 163 micro-positioner, 36
interaction, 22 microstrip, 18
intermodulation, 431–438 microstrip mode, 29
intrinsic gain, 415, 438, 439 microwave
IQ (In-phase and Quadrature), 61, 66, 93–94 equipment, 41
data conditioning, 184 frequencies, 52, 53, 61
modulator, 392 instrumentation, 39
up-conversion, 396 probes, 219
isodynamic characteristics, 417, 418, 426, 427, 438 probing, 34–38
isolation chains, 114 synthesizers, 41, 62
systems, 44, 55
jitter, 45 transistor, 346
joint time-frequency analysis, 174 microwave power, measurement of, 132
mismatch, 22, 25–27, 33
Kaiser window, 73–76, 85 corrections, 149
kink effect, 422–423, 431, 439 factor, 144, 149–150
mixed-mode
LAN eXtensions for Instrumentation (LXI), 164 load-pull, 365
large-signal, 345 S-matrix, 232
device modeling, 346 S-parameters, 230–235
latency, 166 mixed-signal, 395, 397
launchers, 219 mixer, 41, 43, 45, 48, 49, 51, 54, 55, 57, 61
linearity, 419, 431, 432 mixer-based architectures, 388
Line-Reflect-Line (LRL), 22 modulated signals, measurements of, 134–136
list mode, 55, 57 modulation, 41, 42, 55, 58, 59, 60, 61, 134
loading conditions, 345 modulation bandwidth, 397, 400
load-pull systems, 345, 346 modulation events, 66
power calibration, 362, 365 modulator, 59, 61
Index 445
monolithic microwave integrated circuit, 34 peak envelope power, 133

multiple-input multiple-output (MIMO) technology, peak power, 133
179, 180 persistence, 86–87
multiple-modes, 28–29 persistence spectrum, 66, 85–86
multiport measurement, 219 phase
multiport probes, 219 array, 183
multiport VNA, 221 calibration, 370
coherency, 176
native oxide, 36 delay, 407
net power, 132 detector, 41, 52, 53
NIST Type IV power meter, 138 modulation, 58, 59, 60
NLTL, see nonlinear transmission line noise, 42, 44, 45, 46, 48, 50, 52, 53, 54,
noise, 24, 241 55, 62
noise added, 243, 267 stable cables, 34
noise circles, 245 truncation, 51
noise factor, 243, 266 phase-lock-loop (PLL), 43, 44, 47, 52, 53, 54, 55,
noise figure, 242 57, 60
noise parameters, 244–246 PIN diode, 106, 108
noise temperature, 243–244, 268–270 PISPO
non-CW excitation, 135 mixer, 314
non-ideal ground, 32 signal class, 315, 320
nonlinear, 92 system class, 313
nonlinear circuit, 65 port voltage, 28
nonlinear conditions, 345 position, 131
nonlinearity, 433, 437–438 power, 131, 134
nonlinear operating point, 320 power-ground fixture, performance measures,
nonlinear transmission line, 280 30–32
non-real-time load-pull, 347 power levels, 143
non-real-time systems, 359 power measurement, 142–143, 149
NVNA technology, 309 process, 154
IQ modulator based, 341 type of, 132
measurement synchronisation, 322, power meter, 137, 139, 142
339 uncertainty, 152
oscilloscope-based, 328 power-reference, 139, 151
phase calibration, 325 oscillator, 151
power calibration, 325 power spectral density, 318
sampler based, 331 pre-matching, 349
VNA, 338 prescaler, 53
Nyquist criteria, 69, 78, 83, 85 probability density function (PDF), 319
probe design, 35
OIP3 , 387 probe station, 35
open-loop, 139 probing, 34
active load-pull, 351 process development, 346
optimum load, 346 producing, 139
oscillation, 30 propagation constant, 5, 22, 33
oscillator, 41, 42, 44, 47, 52, 55, 57, 60 propagation of uncertainties, 151
oscilloscopes, 86, 88, 91, 93–94 pulling, 366
output, signal, 30 pulse, 134
oversampled structure, 118 characteristics, 415, 416
measurements, 414, 415
partially leaky multiport VNA, 224 pulsed-I/V, 427, 428–429
partial reflectometer, 221, 225 pulsed-RF, 418, 428, 429
passive load-pull systems, 347 pulse power, 133
passive tuners, 347, 349 pulser
PCI eXtensions for Instrumentation (PXI), operation, 281
164–165 schematic, 281
446 Index
range, 143 skin-effect loss, 25

real-time, 66, 68, 69, 76, 85, 90, 92–94 slot-line mode, 29
load-pull, 347 slow modes, 315
load-pull calibration, 361–365 small-signal parameters, 431, 432, 438
real-time signal analyzer (RTSA), 64, 66–69, 74–75, Smith Chart, 13–16, 245
80, 85, 87, 88, 91, 93, 94 software-designed systems, 171
trigger system, 90 SOLR calibration, 210–211
real-time systems, 359 SOLT calibration, 215–216
receivers, 221 source-locking, 105
reference planes, 21, 22, 24 source-pull, 345, 375–378
references, 134 source reflection coefficient, 376
reference temperature, 242 sources, 221
reflected power, 132 S-parameter calculation (from TDR), 291
reflection coefficient, 7–8 error propagation, 301
reflectometer, 12, 23, 197, 201, 221 S-parameters, 101, 309
relative vs. absolute measurement, 310 common mode, 230
remains, 143 differential, 230
residual directivity, 111 hot-S22 , 310
resistive switch bank, 405 invalidation, 310, 311
resolution bandwidth (RBW), 65, 74–76, 78, 80, mixed-mode, 230–235
83, 85 single ended, 230
resonance, 28 spatial multiplexing, 179, 180
results, 146 spectral purity, 42, 43
return loss, 7–8 spectrum, 43
Rohde & Schwarz GmbH & Co., 68 spectrum analyzers, 64, 69–70, 74, 75, 78, 83, 86,
round robin, 221 88, 95
RSS combination, 151 spectrum monitoring, 69, 76
splitter, 156
spurs, 43, 44, 49, 51, 53, 55, 62
safe-operating area, 418 square-law detection region, 140
sampler-based architectures, 386, 388 standard compliant W-CDMA, 388
samplers, 115 standing wave, 7
operation, 280 standing wave ratio (SWR), 23
schematic, 280 state variables, 415, 417
sampling oscilloscope, 368 step-and-sweep measurements, 417, 418, 431
scattering parameters, 8–11 streaming, 184–185
Schottky diodes, 280 stripline, 18
second stage correction, 248, 250, 252 strong nonlinear conditions, 347
self-heating, 419, 426, 433–436, 439 sub-harmonics, 43, 44
sequential sampling timebase substituted power, 144
disadvantages, 283 suckout, 27, 33, 34
operation, 283 superposition, 310
schematic, 282 super-resolution spectral analysis, 174
Short-Open-Load-Thru (SOLT) calibration, 37 swept tuned spectrum analyzer (SA), 64, 65, 68
signal detection bandwidth, 400 switch correction, 223
signal flow graphs, 16–18 switched filter bank, 43, 61
signal generation bandwidth, 400 synchronous detection, 118
signal integrity, 219
signal propagation, 3, 4–6 Tektronix, 68, 85–87, 89
signals, 135, 136 thermal noise, 241
signal-to-noise ratio (SNR), 242, 268 thermal response, 419, 420, 434, 435
single-channel W-CDMA, 411 thermistor detector, 137
single ended S-parameters, 230 thermocouples, 138
single-events, 66, 69, 76 thermoelectric detectors, 135, 136, 138–139
single-sideband noise figure, 267–268 three carrier W-CDMA, 386
single-sideband noise temperature, 269 Thru-Reflect-Line (TRL), 22
skin-depth, 25 calibration, 208–210
Index 447
time constant, 420 unit, 142

time domain events, 66 use, frequencies of, 25
time domain reflectometry (TDR)
incident wave extraction, 289 varactor-tuned VCOs, 105
reflected wave extraction, 289 Vector Network Analyzer (VNA), 24, 98, 195
time domain waveform, 347 Vector Signal Analyzer (VSA), 64, 65–66
calibration, 366, 370 virtual instruments, 161–163
load-pull, 368 VMEbus eXtensions for Instrumentation (VXI), 163
time-evolution characteristics, 426–427, VNA calibrations, 195
438 voltage, 131
timing, 42 voltage controlled oscillator (VCO), 41, 52, 53, 54,
trace noise, 120 55, 57, 60
traditional instruments, 162 Volterra, 313
transfer switch, 108
transfer system, 154
transient, 90–93, 96 Watt, 132
transient characteristics, 426 Wave Analyzers, 64
transient detection, 69, 76 wave direction
transient events, 65, 66, 69 for TDR, 287
transient measurements, 423, 437 for VNA, 286
transient nonlinear, 92 waveform, 47, 50, 51
transient signals, 66, 78 load-pull, 368, 374
transistor, 57 waveguide, 18
characterization, 37–38 wavelength, 3
transmission lines, 3, 18, 21, 22 wavelet analysis, 174
model, 4, 5 white noise, 241
trap center, 421 wideband load-pull, 356–357
trapping, 421, 423, 426 wideband signal generation, 397–400
trapping response, 422 window function, 73–76, 78, 80, 81, 85, 92
triggering, 88–92
trigger signal, 66 Y-factor, 246–249, 270–272
Yttrium-Iron-Carnet (YIG)
uncertainty, 22 filters, 61, 355, 365
uncertainty in the effective efficiency, 152 oscillators, 105
undersampled structure, 118 Y-Z de-embedding, 38

Modern RF and Microwave Measurement Techniques (Valeria Teppati, Andrea Ferrero and Mohamed Sayed, 2013) Livro - 476 Páginas

Uploaded by

Copyright:

Available Formats

Modern RF and Microwave Measurement Techniques (Valeria Teppati, Andrea Ferrero and Mohamed Sayed, 2013) Livro - 476 Páginas

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Modern RF and Microwave Measurement Techniques (Valeria Teppati, Andrea Ferrero and Mohamed Sayed, 2013) Livro - 476 Páginas

Uploaded by

Copyright:

Available Formats

more information - www.cambridge.

This comprehensive, hands-on review of the most up-to-date techniques in RF and

VAL ERIA T E PPAT I

© Cambridge University Press 2013

This publication is in copyright. Subject to statutory exception

First published 2013

Library of Congress Cataloguing in Publication data

ISBN 978-1-107-03641-3 Hardback

Cambridge University Press has no responsibility for the persistence or

Preface page xvii

Part I General concepts 1

1 Transmission lines and scattering parameters 3

2 Microwave interconnections, probing, and ﬁxturing 21

2.3 Signal-path fixture performance measures 24

Part II Microwave instrumentation 39

4 Real-time spectrum analysis and time-correlated measurements applied to

4.8.4 Memory effects in digitally pre-distorted (DPD) amplifiers 95

5 Vector network analyzers 98

5.7.3 Using reversing couplers 126

6 Microwave power measurements 130

7 Modular systems for RF and microwave measurements 160

Part III Linear measurements 193

8 Two-port network analyzer calibration 195

9 Multiport and differential S-parameter measurements 219

10 Noise ﬁgure characterization 240

11 TDR-based S-parameters 279

11.10 Conclusions 304

Part IV Nonlinear measurements 307

12 Vector network analysis for nonlinear systems 309

13 Load- and source-pull techniques 345

13.5 Measuring the DUT single-frequency characteristics 359

14 Broadband large-signal measurements for linearity optimization 384

15 Pulse and RF measurement 414

15.4.3 Impact ionization 422

We thus tried to make an effort to produce a book that could:

ACLR Adjacent Channel Leakage Ratio

DWT Discrete Wavelet Transform

QSOLT Quick Short Open Load Thru

VSWR Voltage Standing Wave Ratio

1.2 Fundamentals of transmission lines, models and equations

where ν is velocity (= c in air), f is frequency and λ is wavelength, the wavelength of

150 million kilometers

Fig. 1.3 Two-wire line.

1.2.2 Propagation and characteristic impedance

V (z, t) = (Rz + j ωLz)I (z, t) (1.2)

I (z, t) = (Gz + j ωCz)V (z, t). (1.3)

V = V1 e−j γ z + V2 e+j γ z , (1.7)

I = I1 e−j γ z + I2 e+j γ z . (1.8)

V (z, t) = V (z)ej ωt I (z, t) = I (z)ej ωt (1.13)

and the circuit equations rewritten as

Again, differentiating gives

1.2.3 Terminations, reﬂection coefﬁcient, SWR, return loss

with Vincident a constant and Vreflected a function of ZL , EEmax

1.2.4 Power transfer to load

1.3 Scattering parameters

Fig. 1.5 Generic n-port network.

Z-matrix (impedance matrix):

“Scattering parameters,” which are commonly referred to as S-parameters, are a param-

1.4 Microwave directional coupler

1.4.1 General concepts

V (z, t) = (Rz + j ωLz)I (z, t) (1.2)

I (z, t) = (Gz + j ωCz)V (z, t). (1.3)