X-Parameters

This is the denitive guide to X-parameters, written by the original inventors and
developers of this powerful new paradigm for nonlinear RF and microwave components
and systems.
Learn how to use X-parameters to overcome intricate problems in nonlinear RF and
microwave engineering, as the general theory behind X-parameters is carefully and
intuitively introduced, and then simplied down to specic, practical cases, providing
you with useful approximations that will greatly reduce the complexity of measuring,
modeling, and designing for nonlinear regimes of operation.
Containing real-world case studies, denitions of standard symbols and notation,
detailed derivations within the appendices, and exercises with solutions, this is the
denitive stand-alone reference for researchers, engineers, scientists, and students
looking to remain on the cutting edge of RF and microwave engineering.

David E. Root is an Agilent Research Fellow at Agilent Technologies. He co-led the

Agilent research and technical development of X-parameters through its commercial-
ization. He is a Fellow of the IEEE, and co-editor of Nonlinear Transistor Model
Parameter Extraction Techniques (2011).
Jan Verspecht is a Master Research Engineer at Agilent Technologies. He invented
X-parameters in 1996, and is a Fellow of the IEEE.
Jason Horn is an Expert Design Engineer at Agilent Technologies, who has been
heavily involved in the development of X-parameter measurements.
Mihai Marcu is a Senior Consultant at Agilent Technologies, who is deeply involved
in the development and application of X-parameters for nonlinear modeling.
The Cambridge RF and Microwave Engineering Series

Series Editor
Steve C. Cripps, Distinguished Research Professor, Cardiff University

Peter Aaen, Jaime Pl and John Wood, Modeling and Characterization of RF and
Microwave Power FETs
Dominique Schreurs, Mirtn ODroma, Anthony A. Goacher and Michael Gadringer,
RF Amplier 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, Advanced RF MEMS
Patrick Roblin, Nonlinear RF Circuits and the Large-Signal Network Analyzer
Matthias Rudolph, Christian Fager and David E. Root, Nonlinear Transistor Model
Parameter Extraction Techniques
John L. B. Walker, Handbook of RF and Microwave Solid-State Power Ampliers
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, Modern RF and Microwave
Measurement Techniques
Nuno Borges Carvalho and Dominique Schreurs, Microwave and Wireless
Measurement Techniques
David E. Root, Jan Verspecht, Jason Horn and Mihai Marcu, X-Parameters

Forthcoming
Richard Carter, Theory and Design of Microwave Tubes
Hossein Hashemi and Sanjay Raman, Silicon mm-Wave Power Ampliers and
Transmitters
Earl McCune, Dynamic Power Supply Transmitters
Isar Mostafanezad, Olga Boric-Lubecke and Jenshan Lin, Medical and Biological
Microwave Sensors
X-Parameters
Characterization, Modeling, and Design of
Nonlinear RF and Microwave Components

D AV I D E . R OO T
Agilent Technologies, Inc.

J A N VE R S P E C H T
Agilent Technologies, Inc.

J A S O N HO R N
Agilent Technologies, Inc.

M I H A I MA R C U
Agilent Technologies, Inc.
University Printing House, Cambridge CB2 8BS, United Kingdom

Published in the United States of America by Cambridge University Press, New York

Cambridge University Press is part of the University of Cambridge.

It furthers the Universitys mission by disseminating knowledge in the pursuit of
education, learning and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9780521193238
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
Printing in the United Kingdom by TJ International Ltd. Padstow Cornwall
A catalog record for this publication is available from the British Library
Library of Congress Cataloging in Publication data
Root, David E.
X-parameters : characterization, modeling, and design of nonlinear RF and microwave components / David E.
Root, Agilent Technologies Inc., Jan Verspecht, Agilent Technologies Inc., Jason Horn, Agilent Technologies
Inc., Mihai Marcu, Agilent Technologies Inc.
pages cm (The Cambridge RF and microwave engineering series)
Includes bibliographical references.
ISBN 978-0-521-19323-8 (Hardback)
1. Microwave circuitsDesign and constructionMathematics. 2. Electric circuits, NonlinearDesign and
constructionMathematics. 3. Parametric devicesDesign and constructionMathematics.
4. Differential equations. I. Verspecht, Jan. II. Horn, Jason. III. Marcu, Mihai. IV. Title.
TK7876.R66 2013
621.38410 2dc23 2013013915
X-parameters is a trademark of Agilent Technologies, Inc.
ISBN 978-0-521-19323-8 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.
For Marilyn, with thanks for her patience, support, and, most of all, her love.
David

In memory of Petrus Verspecht.

Jan

To Jessica, Jonathan, and Elise, my inspiration.

Jason

To Domnica, for the patience shown in the many evenings and weekends
that I have spent away from her.
Mihai
Just as the S-parameters revolutionized linear microwave circuit engineering nearly 60 years
ago, the relatively new development of the X-parameters and the mixer-based VNA provides
a truly scientic approach to nonlinear RF and microwave circuit design. This book, written
by experts, contains a wealth of information about the characterization and modeling of
nonlinear components as well as their applications to various types of designs. I can only wish
that such capability and textbook had been available when I was a design engineer.
Les Besser
Founder of Compact Software and Besser Associates

S-parameters revolutionized linear RF and Microwave design in the 1970s and X-parameters are
doing the same for non-linear design today. Starting with the familiar foundation of S-parameters,
the text guides the reader through the additional non-linear terminology needed to provide
a clear and practical view of X-parameters. Many practical examples show how to apply
them in real world designs and answers are provided to some of the more subtle concepts of
cross-frequency phase and memory effects. In a world where wireless is proliferating, this
book will be an invaluable reference for any RF designer to reduce design turns and improve
their rst-pass designs.
Mark Pierpont
Agilent Technologies
 Contents

Preface page xiii

Acknowledgments xv

1 S-parameters a concise review 1

1.1 Introduction 1
1.2 S-parameters 1
1.3 Wave variables 2
1.4 S-parameter measurement 5
1.5 S-parameters as a spectral map 7
1.6 Superposition 8
1.7 Time invariance of components described by S-parameters 10
1.8 Cascadability 11
1.9 DC operating point 12
1.10 S-parameters of a nonlinear device 12
1.11 Additional benets of S-parameters 15
1.11.1 S-parameters are applicable to distributed components at high
frequencies 15
1.11.2 S-parameters are easy to measure at high frequencies 15
1.11.3 Interpretation of two-port S-parameters 15
1.11.4 Hierarchical behavioral design with S-parameters 16
1.12 Limitations of S-parameters 16
1.13 Summary 18
Exercises 18
References 18
Additional reading 19

2 X-parameters fundamental concepts 20

2.1 Overview 20
2.2 Nonlinear behavior and nonlinear spectral mapping 20
2.3 Multi-harmonic spectral maps 22
2.4 Load- and source-mismatch effects 25
2.5 Cascading DUTs 25
2.6 Example: cascading two RF power ampliers with independent bias 27
viii Contents

2.7 Relationship to harmonic balance 29

2.8 Cross-frequency phase 30
2.8.1 Commensurate signals 30
2.8.2 Denition of cross-frequency phase 30
2.9 Basic X-parameters for multi-harmonic multi-port stimulus 34
2.9.1 Time invariance and related properties of Fp,k(.) functions 35
2.9.2 Denition of X-parameters and X-parameter behavioral model 36
2.9.3 Example: a set of X-parameters 37
2.10 Physical meaning of the basic X-parameters 38
2.10.1 Reference stimulus and response 38
2.10.2 Physical interpretation 39
2.11 Using the X-parameter behavioral model 39
2.11.1 Example: amplier with source and load mismatch 40
2.12 Summary 43
Exercises 44
References 44
Additional reading 44

3 Spectral linearization approximation 45

3.1 Simplication of basic X-parameters for small mismatch 45
3.1.1 Non-analytic maps 46
3.1.2 Large-signal operating point 48
3.2 Adding small-signal stimuli (linearized nonlinear spectral mapping) 50
3.2.1 Small-signal interactions: the RF terms 51
3.2.2 Small-signal interactions: the DC terms 52
3.3 Physical meaning of the small-signal interaction terms 55
3.4 Discussion: X-parameters and the spectral Jacobian 60
3.5 X-parameters as a superset of S-parameters 60
3.6 Two-stage amplier design 64
3.7 Amplier matching under large-signal stimulus 68
3.7.1 Output matching and hot-S22 69
3.7.2 Input matching 78
3.8 Practical application a GSM amplier 80
3.9 Summary 84
Exercise 84
References 87
Additional reading 87

4 X-parameter measurement 88
4.1 Measurement hardware 88
4.1.1 Hardware requirements 88
4.1.2 Mixer-based systems 88
 Contents ix

4.1.3 Sampler-based systems 91

4.1.4 Stimulus requirements 93
4.2 Calibration 93
4.2.1 Scalar-loss correction 94
4.2.2 S-parameter calibration 94
4.2.3 NVNA calibration 96
4.3 Phase references 97
4.3.1 Phase-reference signals 97
4.3.2 Measurement considerations 99
4.3.3 Practical phase references 100
4.4 Measurement techniques 101
4.4.1 Large-signal response measurements 101
4.4.2 Small-signal response measurements 101
4.4.3 Practical measurement considerations 105
4.4.4 Simulation-based extraction 106
4.5 X-parameter les 106
4.5.1 Structure 107
4.5.2 Naming conventions 107
4.5.3 Example le 108
4.6 Summary 110
References 110
Additional reading 111

5 Multi-tone and multi-port cases 112

5.1 Introduction 112
5.2 Commensurate signals large A1,1 and large A2,1: load-dependent
X-parameters 113
5.2.1 Time invariance, phase normalization, and commensurate
two-tone LSOP 114
5.2.2 Spectral linearization 115
5.3 Establishing the LSOP using a load tuner: passive load pull 116
5.4 Additional considerations for commensurate signals 118
5.4.1 Extraction of X-parameter functions under controlled loads 118
5.4.2 Harmonic superposition 118
5.4.3 Limitations of passive load pull for load-dependent X-parameters 119
5.4.4 Sampling of the three-RF-variable space dening the refLSOPS 119
5.4.5 Hardware setup for load-dependent X-parameters 119
5.4.6 Calibrating out uncontrolled harmonic impedances 119
5.5 Arbitrary load-dependent X-parameters of a GaAs FET 120
5.5.1 Load-dependent X-parameter model of a GaN HEMT:
estimating the effect of independent harmonic
impedance tuning 123
x Contents

5.6 Design example: Doherty power amplier design and validation 129
5.6.1 Doherty power amplier 129
5.6.2 X-parameter characterization of the transistors 130
5.6.3 X-parameter model validation 132
5.6.4 Doherty power amplier design using X-parameters 135
5.6.5 Results 136
5.7 Incommensurate signals 138
5.7.1 Notation for incommensurate two-tone X-parameters 138
5.7.2 Time invariance for incommensurate two-tone X-parameters 140
5.7.3 Reference LSOP 141
5.7.4 Spectral linearization 141
5.7.5 Discussion 143
5.7.6 When intermodulation frequencies are negative 143
5.7.7 X-parameter models of mixers 144
5.8 Summary 147
Exercises 148
References 148
Additional reading 148

6 Memory 150
6.1 Introduction 150
6.2 Modulated signals: the envelope domain 151
6.3 Quasi-static X-parameter evaluation in the envelope domain 151
6.3.1 Quasi-static two-tone intermodulation distortion from a static
one-tone X-parameter model 152
6.3.2 ACPR estimations using quasi-static approach 159
6.3.3 Limitations of quasi-static approach 160
6.3.4 Advantages of quasi-static X-parameters for digital modulation 161
6.4 Manifestations of memory 161
6.5 Causes of memory 163
6.5.1 Self-heating 163
6.5.2 Bias modulation 163
6.6 Importance of memory 167
6.6.1 Modulation-induced baseband memory and carrier memory 167
6.6.2 Dynamic X-parameters 168
6.6.3 Identication of the memory kernel: conceptual motivation 171
6.6.4 Step response of the memory kernel 172
6.6.5 Application to real amplier 173
6.6.6 Validation of memory model 175
6.6.7 Interpretation of dynamic X-parameters 181
6.6.8 Wide-band X-parameters (XWB) 182
References 187
Additional reading 188
 Contents xi

Appendix A: Notations and general definitions 189

A.1 Sets 189
A.2 Vectors and matrices 189
A.3 Signal representations 190
A.3.1 Time-domain representation (real signal) 190
A.3.2 Complex representation (complex envelope signal) 190
A.4 Fourier analysis 191
A.5 Wave denitions 192
A.5.1 Generalized power waves 192
A.5.2 Voltage waves 194
A.6 Linear network matrix descriptions 194
A.6.1 S-parameters 195
A.6.2 Z-parameters 195
A.6.3 Y-parameters 195
References 195

Appendix B: X-parameters and Volterra theory 196

B.1 Introduction 196
B.2 Mathematical notation and problem denition 196
B.3 Application of the Volterra theory 197
B.4 Derivation of the McLaurin series 198
B.5 McLaurin series for the DC output 200
B.6 Conclusions 200
References 201

Appendix C: Parallel Hammerstein symmetry 202

References 203

Appendix D: Wide-band memory approximation 204

Appendix E: Solutions to exercises 206

Index 216
Preface

The need for a rigorous, yet practical, framework for characterization, modeling, and
design of nonlinear electronic components at high frequencies has never been more
urgent. The communications revolution is inexorably forcing active devices into more
and more strongly nonlinear regimes of operation. This is a consequence of the
relentless drive for more efciency in order to save power, extend battery life, and
minimize cooling. The price for efciency is nonlinearity. Dealing with nonlinearity
means that new measurement instrumentation and new modeling and design method-
ologies are required that go far beyond linear S-parameters. Fortunately, there is an
overarching, interoperable paradigm combining all these pieces of the nonlinear puzzle
together, seamlessly. The new paradigm is called X-parameters,1 and that is what this
book is about.
The book is intended as a comprehensive introduction to X-parameters. It is aimed at
a diverse audience with a wide range of backgrounds. This is quite a challenging
undertaking! We are targeting professional microwave engineers, device modeling
engineers and scientists, RF and microwave circuit designers, electronic and communi-
cations engineers, CAE professionals developing simulator algorithms, and microwave
and RF professionals developing new high-speed instrumentation for a wide range
of nonlinear characterization applications. The inherent interdisciplinary nature of
X-parameters is the prime reason we seek to appeal to this broad audience. The practical
solutions based on X-parameters deployed by industry over the past several years
depend on contributions in all of these areas.
With this diverse audience in mind, we have chosen a particular sequence with which
to introduce the subject. We start with a concise summary of the well-known time-
invariant linear theory, namely S-parameters. We choose this context, familiar to many
readers, to introduce more advanced concepts that will be needed for the remainder of
the book. Chapter 2 introduces X-parameters, based on multi-tone nonlinear spectral
maps dened on a harmonic grid, and goes into signicant detail about the application
and implications of the constraint of time invariance. Chapter 3 simplies the general
discussion to simple practical cases, based on the application of spectral linearization, a
useful approximation that reduces complexity, enabling practical applications. Several
examples are presented demonstrating the power, utility, and relative simplicity of these

1
X-parameters is a trademark of Agilent Technologies, Inc.
xiv Preface

simplest X-parameters. The origins of conjugate terms in the spectral linearization are
discussed. Chapter 4 is devoted to how X-parameters are measured, and also to how
they are computed (generated) from within a circuit simulator. The functional block
diagram of the main instrument (the nonlinear vector network analyzer NVNA) is
discussed, and the application of measurements using a pulse generator phase reference
to obtain the key X-parameter quantities is reviewed. Chapter 5 extends the treatment of
X-parameters to multiple large signals and multiple ports, as is necessary in the
treatment of many mixers, the treatment of intermodulation with phase, and the large-
signal response of power ampliers as nonlinear functions of both input power and
reections of electrical signals back into the device due to large mismatch, going
beyond the rst spectral linearization approximation of Chapter 3. Finally, Chapter 6
extends the treatment of X-parameters to dynamic memory effects, important phe-
nomena exhibited by practical modern high-speed devices in response to wide-band
communication signals, for example. Several appendices are provided for detailed
derivations, standard symbol and notational denitions, and further elaboration of some
parts of the main text to help serve as a reference for workers in the eld.
The book is appropriate as a text for an advanced undergraduate or graduate course in
electrical engineering. In fact, we perceive an acute need to make X-parameters a
standard part of the electrical engineering curriculum. The book may also be appropriate
for applied mathematicians and scientists with an interest in rigorous and practical
foundations for applications to a wide range of nonlinear systems well beyond
electronics.
The background needed by readers of this book is not much more than rst-year
calculus, basic circuit theory, and simple Fourier analysis. Rudimentary knowledge of
electronic power ampliers and transistors, S-parameter fundamentals, differential
equations, circuit design, and circuit simulation would certainly be helpful.
Acknowledgments

The authors are profoundly grateful to our many dedicated and talented colleagues who
collaborated with us to develop and deploy X-parameter technology, products, support,
and services. We are grateful to our many customers, academic researchers, and
practicing professionals for their thoughtful feedback, stimulating discussions, and
creative applications of this technology. We thank Agilent management for their vital
support. Finally, we thank the staff at Cambridge University Press for their commitment
to this project, their cheerful professionalism, and their patience.
1 S-parameters a concise review

1.1 Introduction

This chapter presents a concise treatment of S-parameters, meant primarily as an

introduction to the more general formalism of X-parameters. The concepts of time
invariance and spectral maps are introduced at this stage to enable an easier general-
ization to X-parameters in the ensuing chapters. The interpretations of S-parameters as
calibrated measurements, intrinsic properties of the device under test (DUT), IP-secure
component behavioral models, and composition rules for linear system design are
presented. The cascade of two linear S-parameter components is considered as an
example to be generalized to the nonlinear case later. The calculation of S-parameters
for a transistor from a simple nonlinear device model is used as an example to introduce
the concepts of (static) operating point and small-signal conditions, both of which must
be generalized for the treatment of X-parameters.

1.2 S-parameters

Since the 1950s, S-parameters, or scattering parameters, have been among the most
important of all the foundations of microwave theory and techniques.
S-parameters are easy to measure at high frequencies with a vector network analyzer
(VNA). Well-calibrated S-parameter measurements represent intrinsic properties of the
DUT, independent of the VNA system used to characterize it. Calibration procedures [1]
remove systematic measurement errors and enable a separation of the overall values into
numbers attributable to the device, independent of the measurement system used to
characterize it. These DUT properties (gain, loss, reection coefcient, etc.) are famil-
iar, intuitive, and important [2]. Another key property of S-parameters is that the
S-parameters of a composite system are completely determined from knowledge of
the S-parameters of the constituent components and their connectivity. S-parameters
provide the complete specication of how a linear component responds to an arbitrary
signal. Therefore designs of linear systems with S-parameters are predictable with
absolute certainty. S-parameters dene a complete behavioral description of the linear
component at the external terminals, independent of the detailed physics or specics of
the realization of the component. S-parameters can be shared between component
vendors and system integrators freely, without the possibility that the component
2 S-parameters a concise review

implementation can be reverse engineered, protecting IP and promoting sharing and

reuse. Indeed, one may ask the question, are S-parameters measurements, or do they
constitute a model? The answer is really both.
S-parameters need not come only from measurements. They can be calculated from
physics by solving Maxwells equations, by linearizing the semiconductor equations, or
computed from matrix analysis of linear equivalent circuits. In this way, the many
benets of S-parameters can be realized, starting from a more detailed representation of
the component from rst principles or from a complicated linear circuit model.
Graphical methods based around the Smith chart were invented to visualize and
interpret S-parameters, and graphical design methodologies soon followed for circuit
design [2][3]. These days, electronic design automation (EDA) tools provide simulation
components S-parameter blocks and design capabilities using the familiar
S-parameter analysis mode.
One of the great utilities of S-parameters is the interoperability among the measurement,
modeling, and design capabilities they provide. One can characterize the component with
measured S-parameters, use them as a high-delity behavioral model of the component
with complete IP protection, and design systems with them in the EDA environment.

1.3 Wave variables

The term scattering refers to the relationship between incident and scattered (reected
and transmitted) traveling waves.
By convention, in this text the circuit behavior is described using generalized power
waves [2]. There are alternative wave denitions used in the industry. These are briey
reviewed in Appendix A, together with the general notations used in this text.
The wave variables, A and B, corresponding to a specic port of a network, are
dened as simple linear combinations of the voltage and current, V and I, at the same
port, according to Figure 1.1 and equations (1.1):
V Z0I
A p ,
2 Z0
V  Z0I 1:1
B p :
2 Z0
The reference impedance for the port, Z0, is, in general, a complex value. For the
purpose of simplifying the concepts presented, the reference impedance is restricted to
real values in this text.

Figure 1.1 Wave denitions.

 1.3 Wave variables 3

I1 I2
+ + + +
DUT
A1 A2
V1 Port 1 Port 2 V2
B1 B2

- - - -

Figure 1.2 Incident and scattered waves of a two-port device.

The currents and voltages can be recovered from the wave variables, according to
equations (1.2):
p
V Z 0 A B,
1
I p A  B: 1:2
Z0

Here, A and B represent the incident and scattered waves, V and I are the port voltage and
current, respectively, and Z0 is the reference impedance for the port. A typical value of Z0 is
50 by convention, but other choices may be more practical for some applications. Avalue
for Z0 closer to 1 is more appropriate for S-parameter measurements of power transistors,
for example, given that power transistors typically have very small output impedances.
The variables in equations (1.1) and (1.2) are complex numbers representing the
RMS-phasor description of sinusoidal signals in the frequency domain (see Appendix A
for a more detailed discussion of the notations used). Later this will be generalized to the
envelope domain by letting these complex numbers vary in time.
A, B, V, and I can be considered RMS vectors, the components of which indicate the
values associated with sinusoidal signals at particular ports labeled by positive integers.
Thus Aj is the incident wave RMS phasor at port j and Ik is the current RMS phasor at
port k. For now, Z0 is taken to be a xed real constant, in particular, 50 .
A graphical representation of the wave description is given in Figure 1.2.
To retrieve the time-dependent sinusoidal voltage signal at the ith port, the complex
value of the phasor and also the angular frequency, , to which the phasor corresponds,
must be known. The voltage is then given by
pk jt
vi t RefV i e g, 1:3
pk
and similarly for the other variables, where Vi
are peak values. Equation (1.3) for the
voltage, and a similar equation for the time-dependent current, can be used to dene
real, time-dependent wave quantities using the same linear combinations as in (1.1):
1  
at p vt Z 0 it ,
2 Z0
1   1:4
bt p vt  Z 0 it ;
2 Z0
4 S-parameters a concise review

p  
vt Z 0 at bt ,
1   1:5
it p at  bt :
Z0

It is convenient to keep track of the frequency associated with a particular set of phasors
by rewriting (1.1) according to (1.6), and (1.2) according to (1.7), where the port
indexing notation is made explicit:
1  
Ai p V i Z 0 I i ,
2 Z0
1   1:6
Bi p V i  Z 0 I i ;
2 Z0
p  
V i Z 0 Ai Bi ,
1   1:7
I i p Ai  Bi :
Z0

For each angular frequency, , (1.6) is a set of two equations dened at each port.
The assumption behind the S-parameter formalism is that the system being described
is linear and therefore there must be a linear relationship between the phasor repre-
sentation of incident and scattered waves. This is expressed in (1.8) for an N-port
network as follows:
X
N
Bi S ij Aj , 8i 2 f1, 2,. . . , N g: 1:8
j1

The set of complex coefcients, Sij(), in (1.8) denes the S-parameter matrix or, simply, the
S-parameters at that frequency. Equation (1.8), for the xed set of complex S-parameters,
determines the output phasors for any set of input phasors. The summation is over all
port indices, so that incident waves at each port, j, contribute in general to the overall
scattered wave at each output port, i. For now we consider all frequencies to be positive
( > 0). Note that contributions to a scattered wave at frequency come only from incident
waves at the same frequency. This is not the case for the more general X-parameters,
where a stimulus at one frequency can lead to scattered waves at different frequencies.
The set of equations (1.8) represents a model of the network under study. However, this
model is valid only if the network has the topological connections shown in Figure 1.2.
For example, the model might not accurately represent the behavior of the network
when connected as shown in Figure 1.3 because potential losses between the reference
pins of the two ports are not individually identied in the set of S-parameters in (1.8).
For the purpose of creating a model for the network, all ports should be referenced to
the same pin, as shown in Figure 1.4.
Such connectivity is the natural option for the measurement and modeling process
of a three-pin network (like a transistor), but it has to be extended in the general case
of an arbitrary network, and it is necessary for all networks, linear and/or nonlinear.
This connectivity convention is considered by default (unless otherwise specied) for
the remainder of this text.
 1.4 S-parameter measurement 5

I1 I2

+ + + +
DUT
A1 A2
V1 Port 1 Port 2 V2
B1 B2

- - - -

Figure 1.3 Potential losses between reference pins are not individually identied by the model in (1.8).

I1 I2
DUT
+ + + +

A1 Port 1 Port 2 A2
V1 - - V2
B1 B2

- -

Figure 1.4 All ports should be referenced to the same pin for modeling purposes.

Using the topological connection in Figure 1.4, the set of equations (1.8) represents a
complete model of the network under test.
From (1.8) we note that a stimulus (incident wave) at a particular port j will produce a
response (scattered wave) at all ports, including the port at which the stimulus is applied.
Equation (1.8) shows that the scattered waves are linear functions of the complex
amplitudes (the phasors) of the incident waves. The dependence on frequency of the
S-parameter matrix elements, Sij(), can be highly nonlinear, even for a linear device.
For example, an ideal band-pass lter response is linear in the incident wave variable,
but the lter response is a nonlinear function of the frequency of the incident wave.
This is shown in Figure 1.5.

1.4 S-parameter measurement

By setting all incident waves to zero in (1.8), except for Aj, one can deduce the simple
relationship between a given S-parameter (S-parameter matrix element) and a particular
ratio of scattered to incident waves according to (1.9):

Bi 
S ij  : 1:9
Aj  Ak 0
8k6j
6 S-parameters a concise review

DUT
1 2
P_1Tone + Ref + Term
PORT1 Term2
Num=1 Num=2
z=50 Ohm Z=50 Ohm
P=polar(dbmtow(Pwr_dBm),0)
Freq=RFfreq

Output power vs. input power @ 2.5 GHz Output power vs. frequency @ 0 dBm input power
10 50
Output power (dBm)

Output power (dBm)

8 0
50
6
100
4 150
2 200
250
0
300
2 350
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Input power (dBm) Frequency (GHz)

Figure 1.5 A linear network has a linear behavior when plotted versus input power level, but the
dependence on frequency is usually not linear.

Figure 1.6 S-parameter experiment design: (a) forward transmission; (b) reverse transmission.

Equation (1.9) corresponds to a simple graphical representation shown in Figure 1.6 for
the simple case of a two-port component. In Figure 1.6(a), the stimulus is a wave incident
at port 1. The fact that A2 is not present (A2 0) is interpreted to mean that the B2 wave
scattered and traveling away from port 2 is not reected back into the device at port 2.
Under this condition, the device is said to be perfectly matched at port 2. Two of the four
complex S-parameters, specically S11 and S21, can be identied using (1.9) for this case of
exciting the device with only A1. Figure 1.6(b) shows the case where the device is stimu-
lated with a signal, A2, at port 2, and assumed to be perfectly matched at port 1 (A1 0).
The remaining S-parameters, S12 and S22, can be identied from this ideal experiment.
 1.5 S-parameters as a spectral map 7

It is important to note that the ratio on the right-hand side of (1.9) can be computed
from independent measurements of incident and scattered waves for actual components
corresponding to any non-zero value for the incident wave, Aj. The value of this ratio,
however, will generally vary with the magnitude of the incident wave. Therefore, the
identication of this ratio with the S-parameters of the component is valid for any
particular value of incident Aj only if the component behaves linearly, namely according
to (1.8). In other words, the values of the incident waves, Aj, need to be in the linear region
of operation for this identication to be valid. For nonlinear components, such as
transistors biased at a xed voltage, the scattered waves eventually do not increase
as the incident waves become larger in magnitude (this is compression). Therefore,
different values of (1.9) result from different values of incident waves. A better denition
of S-parameters for a nonlinear component is a modication of (1.9), given by (1.10):

Bi 
S ij  lim  : 1:10
jAj j!0 Aj  Ak 0
8k6j

That is, for a general component, bias at a constant DC stimulus, the S-parameters are
related to ratios of output responses to input stimuli in the limit of small input signals.
This emphasizes that S-parameters properly apply to nonlinear components only in the
small-signal limit.

1.5 S-parameters as a spectral map

If there are multiple frequencies present in the input spectrum, one can represent the
output spectrum in terms of a matrix giving the contributions to each output frequency
from each input frequency.
An example in the case of three input frequencies is given by equation (1.11):
2 3 2 32 3
B1 S 1 0 0 A1
4 B2 5 4 0 S 2 0 54A2 5: 1:11
B3 0 0 S 3 A3
Here we assume a single port, for simplicity, and therefore drop the port indices. It is
clear from (1.11) that S-parameters are a diagonal map in frequency space. This means
that each output frequency contains contributions only from inputs at that same fre-
quency. Or, in other words, each input frequency never contributes to outputs at any
different frequency.
A graphical representation is given in Figure 1.7 for the case of forward transmission
through a two-port network with matched terminations at both ports.
The interpretation of Figure 1.7, mathematically represented by (1.11), is that
S-parameters dene a particularly simple linear spectral map relating incident to scattered
waves. S-parameters are diagonal in the frequency part of the map, namely they predict a
response only at the particular frequencies of the corresponding input stimuli. It will be
demonstrated in later chapters that X-parameters provide for richer behavior.
8 S-parameters a concise review

DUT
1 2
+ Term1
A1
Ref
B2
Term2 +

0 0
20
Power of A1 (dBm)

20

Power of B2 (dBm)
40 40
60 60
80 80
100 100
120 120
3.0 3.2 3.4 3.6 3.8 4.0 3.0 3.2 3.4 3.6 3.8 3.0
Frequency (GHz) Frequency (GHz)

S(w 1).A(w 1) + 0.A(w 2) + 0.A(w 3) = B(w 1)

0.A(w 1) + S(w 2).A(w 2) + 0.A(w 3) = B(w 2)
0.A(w 1) + 0.A(w 2) + S(w 3).A(w 3) = B(w 3)

Figure 1.7 Linear spectral map through S-parameters matrix.

For signals with a continuous spectrum, the diagonal nature of the S-parameter
spectral map can be written as equation (1.12):
N
X
Bi S ij  0 Aj 0 d0 , 8i 2 f1, 2, . . . , N g: 1:12
j1

Performing the integral over input frequencies in (1.12) results in equation (1.8), the
form usually given for S-parameters.

1.6 Superposition

Any linear theory, such as S-parameters, enables the general response to an arbitrary
input signal to be computed by superposition of the responses to unit stimuli.
Superposition enables great simplications in analysis and measurement. Superpos-
ition is the reason S-parameters can be measured by independent experiments with
one sinusoidal stimulus at a time, one stimulus per port per frequency using (1.9).
The general response to any set of input signals can be obtained by superposition
using (1.8).
An example of superposition is shown in Figure 1.8, with all signals represented in
both time and frequency domains.
 1.6 Superposition 9

(a)
Signal 1 - incident Signal 1 - incident Signal 1 - transmitted Signal 1 - transmitted
magnitude of A 1 (dBm)

magnitude of B2 (dBm)
10 0.08 0.08 10

10 10

b2 (sqrt(W))
0.04 0.04

a1 (sqrt(W))
30 30
0.00 0.00
50 50
70 0.04 DUT 0.04 70

90 0.08
1 2 0.08 90
0 1 2 3 4 37 38 39 40 41 42 43 Ref 37 38 39 40 41 42 43 0 1 2 3 4
frequency (GHz) time (ns) time (ns) frequency (GHz)

Signal 1 - reflected Signal 1 - reflected

10 0.4
magnitude of B1 (dBm)

b1 (sqrt(W)*1e-3)

10
0.2
30
0.0
50
70 0.2

90 0.4
0 1 2 3 4 37 38 39 40 41 42 43
frequency (GHz) time (ns)

(b) Signal 2 - reflected Signal 2 - reflected

magnitude of B2 (dBm)
0.08 10

10

b2(sqrt(W))
0.04
30
0.00
50
0.04 70
DUT 0.08 90
1 2 37 38 39 40 41 42 43 0 1 2 3 4
Ref time (ns) frequency (GHz)
Signal 2 - transmitted Signal 2 - transmitted Signal 2 - incident Signal 2 - incident

magnitude of A2 (dBm)
10 0.004 0.08 10
magnitude of B1 (dBm)

0.003
10 10
b2 (sqrt(W))
0.002 0.04
b1 (sqrt(W))

30 0.001 30
0.000 0.00
50 0.001 50
0.002 0.04 70
70
0.003
90 0.004 0.08 90
0 1 2 3 4 37 38 39 40 41 42 43 37 38 39 40 41 42 43 0 1 2 3 4
frequency (GHz) time (ns) time (ns) frequency (GHz)

(c)
Signal 1 + Signal 2 scattered @ port 2 Signal 1 + Signal 2 scattered @ port 2
Signal 1 - incident Signal 1 - incident
magnitude of A 1 (dBm)

10 0.15 0.4 10
0.3 magnitude of B2 (dBm)
10 0.10 10
a1 (sqrt(W))

0.2
b2 (sqrt(W))

+
0.05 0.1
30 30
0.00 0.0
50 0.1 50
0.05 DUT 0.2
70 0.10 1 2 0.3
70
90
0 1 2 3 4
0.15
37 38 39 40 41 42 43
Ref 0.4
37 38 39 40 41 42 43
90
0 1 2 3 4
frequency (GHz) time (ns) time (ns) frequency (GHz)
Signal 1 + Signal 2 scattered @ port 1 Signal 1 + Signal 2 scattered @ port 1 Signal 2 - incident Signal 2 - incident
magnitude of B1 (dBm)

10 0.015 0.3 10
magnitude of A 2 (dBm)

10 0.010 0.2 10
b1 (sqrt(W))

+
b2 (sqrt(W))

0.005 0.1
30 30
0.000 0.0
50 50
0.005 0.1
70 0.010 0.2 70
90 0.015 0.3 90
0 1 2 3 4 37 38 39 40 41 42 43 37 38 39 40 41 42 43 0 1 2 3 4
frequency (GHz) time (ns) time (ns) frequency (GHz)

Figure 1.8 Superposition example. (a) Stimulus signal 1 ) response 1. (b) Stimulus signal
2 ) response 2. (c) Stimulus 2 (signal 1) 3 (signal 2); response 2 (response 1) 3
(response 2).

This example uses two signals, each containing two frequency components, as
stimuli incident at each port, independently, with the other port perfectly matched.
The example shows that the response to a linear combination of the stimuli is the same
linear combination of the individual responses.
As always, the caveat is that the component actually behaves linearly over the range
of signal levels used to stimulate the device. There is no a-priori way to know whether a
component will behave linearly without precise knowledge about its composition or
physical measured characteristics.
10 S-parameters a concise review

1.7 Time invariance of components described by S-parameters

A DUT description in terms of S-parameters dened by (1.8) naturally embodies an

important principle known as time invariance. Time invariance states that if y(t) is
the DUT response to an excitation x(t), the DUT response to the time-shifted excitation,
x(t  ), must be y(t  ). This must be true for all time shifts, . That is, if the input is
shifted in time, the output is shifted by the corresponding amount, but is otherwise
identical with the DUT response to the non-shifted input. This is stated mathematically
in equation (1.13), where O is the operator taking input to output:
   
8 2 R yt O xt ) yt  O xt  : 1:13
Time invariance is a property of common linear and nonlinear components, such as
passive inductors, capacitors, resistors, and diodes, and active devices, such as transis-
tors. Examples of components not time invariant (in the usual sense) are oscillators and
other autonomous systems.
The proof follows from elementary properties of the Fourier transform, where a phase
shift by e j in the frequency domain corresponds to a time shift of in the time domain.
The time-domain waves incident at the ports (the stimuli) are ak(t), and their Fourier
pk
transforms are Ak , as in equation (1.14):
 pk
F ak t Ak : 1:14
The time-domain waves scattered from the ports (the response) are bi(t), and their
pk
Fourier transforms are Bi , as in equation (1.15):

 pk
X
N
pk
X
N 
F bi t Bi S ik Ak S ik F ak t : 1:15
k1 k1

If all stimuli are delayed with the same time delay, , the response becomes
( ) ( )
X
N  X
N 
F 1 S ik F ak t  F 1 S ik F ak t e j
k1
n k1 o
1pk
F Bi e j bi t  : 1:16

Equation (1.16) proves that S-parameters are automatically consistent with the principle of
time invariance. Therefore, any set of S-parameters describes a time-invariant system.
Unlike the case for S-parameters, a more general (e.g. nonlinear) relationship
between incident A waves and scattered B waves is not automatically consistent with
the property (1.13) of time invariance. This will be demonstrated in Chapter 2. There-
fore, in order to have a consistent representation of a nonlinear time-invariant DUT, the
time-invariance property is manifestly incorporated into the mathematical formulation
of X-parameters relating input to output waves. A representation of a time-invariant
DUT by equations not consistent with (1.13) means the model is fundamentally wrong,
and can yield very inaccurate results for some signals, even if the model tting
(or identication) appears good at time t.
 1.8 Cascadability 11

1.8 Cascadability

Another key property of S-parameters is that a linear circuit or system can be designed
with perfect certainty knowing only the S-parameters of the constituent components and
their interconnections. The overall S-parameters of the composite design can be calcu-
lated by using (1.8) in conjunction with Kirchhoffs voltage law (KVL) and Kirchhoffs
current law (KCL), applied at the internal nodes created by connections between two or
more components.
This is illustrated for a cascade of two 2-ports in Figure 1.9(a). Each component is
characterized by its own S-parameter matrix, S(1) and S(2), respectively. The output port
(subscript number 2) of the rst component is connected to the input port (subscript
number 1) of the second component, creating an internal node.
It is a straightforward exercise to write the two equations that follow from applying
KVL and KCL at the internal node. These are given by equations (1.17) and (1.18):
1 2
B2 A1 , 1:17
1 2
A2 B1 : 1:18
Each 2-port relates two input and two output variables. When cascaded, equations
(1.17) and (1.18) can be used to compute the overall S-parameter matrix of the
composite. The result is given in equation (1.19):

(a)

(b)

Figure 1.9 The cascade of two 2-ports. (a) Cascade of two DUTs. (b) The equivalent network,
yielding the same response as the cascade of two DUTs.
12 S-parameters a concise review

0 1 1 1 2 1
1 2 S 12 S 21 S 12 S 12
B S 11 S 11 1 2 1 2 C
B 1  S 22 S 11 1  S 22 S 11 C
B C
S composite
B 1 2 2 2 C: 1:19
B S 21 S 21 S 12 S 21 2 C
@ 1
S 22 S A
1 2 1 2 22
1  S 22 S 11 1  S 22 S 11

A network characterized by the Scomposite is equivalent to the cascade of the two DUTs,
yielding the same response under the same stimuli. This is shown in Figure 1.9(b).

1.9 DC operating point

For devices such as diodes and transistors, the S-parameter values depend very strongly
on the DC bias conditions dening the component operating conditions. For example,
a eld-effect transistor (FET) biased in the saturation region of operation will have
S-parameters appropriate for an active amplier (where jS21j  1  jS12j), whereas
when biased off (say VDS 0) the S-parameters will represent the characteristics of a
passive switch (S21 S12 with jSijj < 1).

1.10 S-parameters of a nonlinear device1

It is possible to derive the S-parameters of a transistor starting from a simple nonlinear

model of the intrinsic device. The process illustrates the concept of linearizing a non-
linear mapping about an operating point. This concept, suitably generalized, is used as a
foundation for X-parameters in Chapter 2.
The equivalent circuit of a simple, unilateral FET model is shown in Figure 1.10.
More complete model topologies and equations are discussed in [4].
A current source, iDS, represents the nonlinear bias-dependent channel current from
drain to source, and a simple nonlinear two-terminal capacitor, qGS, represents the
nonlinear charge storage between gate and source terminals.
The large-signal model equations corresponding to this equivalent circuit are given
by (1.20) and (1.21), where the stimulus applied to the device is the set of port voltages,
vGS and vDS, and the device response is the set of port currents, iGS and iDS:
 
dqGS vGS t
iG t , 1:20
dt
 
iD t iDS vGS t , vDS t : 1:21

Equations (1.20) and (1.21) are evaluated for time-varying voltages assumed to have a
xed DC component and a small sinusoidal component at a single RF or microwave

1
This section can be omitted on rst reading.
 1.10 S-parameters of a nonlinear device 13

+ G D +

qGS iDS

Port 1 Port 2

- S -

Figure 1.10 Simple nonlinear equivalent circuit model of a FET.

angular frequency, 2 f. Port 1 is associated with the gate and port 2 is associated
with the drain terminals, referenced to the source.
For a stimulus comprising a combination of DC and one sinusoidal signal, the
voltages are formally expressed as shown in (1.22):
DC
v i t vi V i cos t i : 1:22
Expressions in (1.23) and (1.24) are obtained by substituting (1.22) for i 1,2 into
(1.20) and(1.21), and evaluating the result to rst order in the real quantities Vi:
DC
i1 t i1 0 i1 t 1

d @
DC  dqGS 
qGS v1 V 1 cos t 1 A
dt dV 1 vDC , vDC 1:23
1 2


DC d  
cGS v1 V 1 cos t 1 ;
dt


DC DC DC
i2 t i2 i2 t iDS v1 ,v2


 1:24
DC DC DC DC
gm v1 ,v2 V 1 cos t1 gDS v1 ,v2 V 2 cos t 2 :

The following denitions of the linearized equivalent circuit elements have been used:

 dq 
GS 
DC
cGS v1 ,
dv1 vDC
 i 
1

DC DC DS 
g m v1 , v 2 , 1:25
v1 vDC , vDC

 i   1 2

DC DC DS 
g DS v1 , v2 :
v2 vDC , vDC
1 2

By equating the zeroth-order terms in Vi, the operating point conditions are
obtained as follows:
DC
i1 0,

DC DC DC 1:26
i2 iDS v1 , v2 :
14 S-parameters a concise review

Equating the rst-order terms in Vi leads to


DC d  
i1 t cGS v1 V 1 cos t 1 ,
dt



DC DC DC DC
i2 t g m v1 , v2 V 1 cos t 1 g DS v1 , v2 V 2 cos t 2 :
1:27
Equation (1.27) is expressed in the frequency domain by dening complex phasors
I() and V() according to
 
ii t Re I i e jt ,   1:28
vi t Re V i e jt Re V i e ji e jt :
Equations (1.27) can be rewritten for the complex phasors in matrix notation as


I 1 DC DC V 1
Y v1 , v2 , , 1:29
I 2 V 2
where
2
 3
DC


jc GS v 0
Y v1 , v2 , 4
DC DC 
DC DC 1

 5: 1:30
DC DC
g m v1 , v2 gDS v1 , v2

Equation (1.30) denes the (common source) admittance matrix of the model. The
matrix elements are evidently functions of the DC operating point (bias conditions) of
the transistor, and also the (angular) frequency of the excitation.
Since the phasors representing the port voltage and currents in (1.29) can be
re-expressed as linear combinations of incident and scattered waves using (1.7),
it is possible to derive the expression for the S-parameters in terms of the
Y-parameters (admittance matrix elements). This results in the well-known conversion
formula (1.31) [2]:

S I  Z 0 Y I Z 0 Y 1 : 1:31
Here I is the 2  2 unit matrix, Z0 is the reference impedance used in the wave
denitions in (1.6), Y is the two-port admittance matrix, and S is the corresponding
S-parameter matrix. Substituting (1.30) into (1.31) results in an explicit expression,
(1.32), for the S-parameters corresponding to this simple model in terms of the linear
equivalent circuit element values given in (1.25):
0
 1
DC
1  jcGS v1 Z0
B
 C
B 0 C
B 1 jcGS v1
DC
Z0 C

 B C
DC DC B

 C
S v1 ,v2 , B DC DC DC DC C:
B 2g m v1 ,v2 Z0 1  gDS v1 ,v2 Z0 C
B C
B
 
 
 C
@ DC DC
Z0 A
Z 0 1 gDS v1 ,v2
DC DC DC
1 gDS v1 ,v2 Z0 1 jcGS v1

1:32
 1.11 Additional benefits of S-parameters 15

In summary, the S-parameters of a nonlinear component can be derived or computed by

linearizing the full nonlinear characteristics of the circuit equations around a static (DC)
operating point dened by the voltage or current bias conditions. The S-parameters dene
a linear relationship between the incident and scattered waves at a xed DC operating
point of the device and xed frequency for the incident waves. The S-parameters are an
accurate description of how the device responds to signals, provided the signal amplitude
is sufciently small that the DC operating point is not signicantly affected by the signal.
This will almost always be the case for signals of sufciently small amplitude.

1.11 Additional benefits of S-parameters

1.11.1 S-parameters are applicable to distributed components at high frequencies

S-parameters can accommodate an arbitrary frequency dependence in the linear spectral
mapping. S-parameters therefore apply when describing distributed components for
which lumped approximations are not very accurate or efcient. This is especially true
for high-frequency microwave components when the typical wavelengths of the stimu-
lus approach and become smaller than the physical size of the component. The simplest
example is the case of linear transmission lines. Another common example is the case of
an active device, for which measured S-parameters of a transistor can be much more
accurate than those computable from the linearized lumped nonlinear model. This is
especially true as the frequency approaches and exceeds the device cutoff frequency, fT,
beyond which a distributed representation is generally required.

1.11.2 S-parameters are easy to measure at high frequencies

S-parameters contain no more information than the familiar Y- and Z-parameters of
elementary linear circuit theory, yet they have great practical advantages. S-parameters
are much easier to measure at high frequencies. Y- and Z-parameters require short- and
open-circuit boundary conditions, respectively, on the components for a direct meas-
urement. Short- and open-circuit conditions are hard to achieve at microwave frequen-
cies, and so are impractical. Moreover, such conditions presented to a power transistor
can create oscillations that can destroy the component.
It will be demonstrated that X-parameters combine the accuracy and ease of meas-
urement of a frequency-domain approach based on wave variables with the ability to
handle nonlinearities that go beyond the linear relationship assumed by equation (1.8).

1.11.3 Interpretation of two-port S-parameters

S-parameters relate to familiar quantities, such as gain, return loss, and output match. They
provide insight into the component behavior. Table 1.1 lists the four complex S-parameters
and their corresponding interpretation for generic linear 2-ports [2]. The third column
expresses common amplier quantities in terms of the corresponding S-parameters.
16 S-parameters a concise review

Table 1.1 S-parameters of a generic 2-port and the corresponding figures of merit of a two-port
amplifier

Generic 2-port (with input and

S-parameter output ports properly terminated) Amplier gure of merit
S11 input reection coefcient return loss: dBjS11j
S12 reverse transmission coefcient isolation: dBjS12j
S21 forward transmission coefcient gain: dBjS21j
S22 output reection coefcient output match: dBjS22j

1.11.4 Hierarchical behavioral design with S-parameters

S-parameters can be measured at any level of the electronic technology hierarchy, from
the transistor (in small-signal conditions) to the linear circuit, or all the way to the
complete linear system. S-parameters provide a complete and accurate behavioral
representation of the (linear) component at that level of the design hierarchy for efcient
design at the next. S-parameters can dramatically reduce the complexity of a multi-
component linear design by eliminating all internal nodes through the application of
linear algebra, as discussed in Section 1.8. S-parameters protect the intellectual property
(IP) of the design by providing only the behavioral relationship at the external terminals,
with no information about the internal realization of the functionality in terms of circuit
elements arranged in a particular topology (e.g. the schematic). S-parameters are
therefore a black-box behavioral approach, requiring no a-priori information about the
component, except that it is linear.
S-parameters are hierarchical. A system of two connected linear components can be
represented by the overall S-parameters of the composite and used as a behavioral
model at a higher level of the design hierarchy. The composite S-parameters can be
obtained from direct measurement of the cascaded system, or by composition of the
S-parameters of the constituent components according to the procedure of Section 1.8.
Any subset of interacting linear components can be represented by an S-parameter
block and inserted into a larger design. The process can be repeated iteratively from one
level to the next level.

1.12 Limitations of S-parameters

Interestingly, S-parameters are still commonly used for nonlinear devices, such as
transistors and ampliers. The problem, often forgotten or taken for granted, is that
S-parameters only describe properly the behavior of a nonlinear component in response
to small-signal stimuli, for which the device behavior can be approximated as linear
around a xed DC, or static, operating point. That is, only when the nonlinear device is
assumed to depend linearly on all RF components of the incident signals is the
S-parameter paradigm valid. S-parameters contain no information about how a non-
linear component generates distortion, such as that manifested by energy generated at
 1.12 Limitations of S-parameters 17

harmonics and intermodulation frequencies by the component in response to excita-

tion by one or more tones (sinusoidal signals). S-parameters are inadequate to
describe nonlinear devices excited by large signals, despite many ad hoc attempts.
Attempts to generalize S-parameters to the nonlinear realm have led to a wide
variety of problems, including measurements that appear non-repeatable, and the
inability to make predictable design inferences from such measurements or simula-
tions. Techniques such hot S-parameter measurements and modeling nonlinear
components using large-signal S-parameter analysis are examples of incomplete,
insufcient, and ultimately incorrect approaches. Under large-signal conditions,
totally new phenomena, for which there are no analogs in S-parameter theory, appear
and must be taken into account.
For example, when a nonlinear component, such as a transistor or power amplier,
is stimulated simultaneously by two sinusoidal signals at different frequencies, the
output spectrum is not consistent with the superposition principle discussed in
Section 1.6. Rather, the response of a nonlinear system to two or more excitations
generally contains more non-zero frequency components than were present in the
input signal (see Figure 1.11). These terms are generally referred to as intermodulation

+ V_DC
-

3
1 DUT 2
+ Term1 Gnd Term2 +
A1 A2

B1 B2
- -

20 20
10 10
0 0
A1 (dBm)

B2 (dBm)

10 10
20 20
30 30
40 40
0 1 2 3 4 5 0 1 2 3 4 5
frequency (GHz) frequency (GHz)

Figure 1.11 Intermodulation spectrum of typical transistor output in response to a simultaneous

excitation by two sinusoidal signals at different frequencies. Superposition is not obeyed in
general.
18 S-parameters a concise review

distortion.2 It is clear that this phenomenon cannot be described by S-parameters. The

more general framework of X-parameters is required to address this type of behavior.

1.13 Summary

S-parameters are linear time-invariant spectral maps dened in the frequency domain.
They represent intrinsic properties of the DUT, enabling the hierarchical design of linear
circuits and systems given only the S-parameters of the constituent functional blocks
and their topological arrangement in the design. S-parameters can be dened, measured,
and calculated for nonlinear components, but they are valid descriptions only under
small-signal conditions. Phenomena such as harmonic and intermodulation distortion,
generated by nonlinear components, require a more comprehensive framework for their
consistent description and application.

Exercises

1.1 Prove that (1.17) and (1.18) are equivalent to the circuit laws. Hint: Start by
expressing KVL and KCL in the time domain, use (1.4) to write the laws in wave
variables, and then transform them to the frequency domain for the corresponding
phasors.
1.2 Derive (1.19). Hint: Eliminate the four internal variables appearing in (1.17)
1 2 1 2
and (1.18). The matrix producing B1 and B2 from A1 and A2 is the required
result.
1.3 Derive (1.31) from (1.1) and the denition of the Y-matrix (reviewed in
Appendix A).
1.4 Derive (1.32), the explicit form of the simple FET model S-parameter matrix,
from the model of Figure 1.10.

References

[1] D. Rytting, Calibration and Error Correction Techniques for Network Analysis, IEEE
Microwave Theory and Techniques Short Course, 2007; available at http://ieeexplore.ieee.
org/servlet/opac?mdnumber=EW1062.
[2] G. Gonzalez, Microwave Transistor Ampliers, 2nd edn. Englewood Cliffs, NJ: Prentice Hall,
1984.
[3] S-parameter techniques for faster, more accurate network design, Hewlett-Packard applica-
tion note AN 951, 1968.

2
In fact, there are even cases where passive connectors cause intermodulation distortion [5].
 Additional reading 19

[4] D. E. Root, J. Xu, J. Horn, and M. Iwamoto, The large-signal model: theoretical foundations,
practical considerations, and recent trends, in Nonlinear Transistor Model Parameter
Extraction Techniques, M. Rudolf, C. Fager, and D. E. Root, Eds. Cambridge: Cambridge
University Press, 2012, chap. 5.
[5] J. J. Henrie, A. J. Christianson, and W. J. Chappell, Linear-nonlinear interaction and
passive intermodulation distortion, IEEE Trans. Microw. Theory Tech., vol. 58, no. 5,
pp. 12301237, May 2010.

Additional reading

K. Kurokawa, Power waves and the scattering matrix, IEEE Trans. Microw. Theory Tech.,
vol. 13, no. 2, pp. 194202, Mar. 1965.
2 X-parameters fundamental
concepts

2.1 Overview

The chapter begins with a brief review of nonlinear DUT behavior that cannot be
described by linear S-parameters. In a simple but important case, it is demonstrated that
a DUT description in terms of a nonlinear spectral mapping of signals dened on a
harmonic frequency grid overcomes this limitation. It is shown that cascadability of
nonlinear components described in this way follows from the circuit laws with no
further approximation. Practical considerations of such a description are discussed. The
principle of time invariance is invoked to obtain the nal form for single-tone
X-parameters. Key applications are presented.

2.2 Nonlinear behavior and nonlinear spectral mapping

It is useful to start the discussion about nonlinear behavior by observing the output of an
amplier when a relatively small sinusoidal signal, of frequency f1, is applied at its
input, port 1:


a1 t Re A1  e j1 t , A1 2 C: 2:1
The output contains an amplied replica of the sinusoidal signal and a number of
additional sinusoidal signals at frequencies that are harmonics of f1. These waves are
displayed in a graphical form in Figure 2.1 for two power levels, 50 dBm and
20 dBm, of the incident wave a1(t) (that contains energy only at frequency 1).
The harmonic frequencies generated by the DUT at the output port have negligible
levels if the input signal is small (as is the case at 50 dBm in Figure 2.1). As usual in
engineering applications, the level below which the harmonics are negligible varies
from one application to another. Once their contribution is negligible, the behavior of
the amplier is accurately described by the S-parameters, as discussed in Chapter 1,
provided also the ratio between response and stimulus is independent of the stimulus
power. Compression of the fundamental frequency response is a nonlinear process that
also precludes the applicability of S-parameters.
Levels of scattered waves at harmonic frequencies rise and become non-negligible as
the level of the input signal increases. The wave scattered from the output port, b2(t),
 2.2 Nonlinear behavior and nonlinear spectral mapping 21

1 2
+ A1
Gnd
B2

0 0
A1 B2
20 20
A1 (dBm)

B2 (dBm)
40 A11 = 50 dBm 40
for A11 = 50 dBm
60 60

80 80

100 100
0 1 2 3 4 5 0 1 2 3 4 5
frequency (GHz) frequency (GHz)
0 0
A1 B2
20 A11 = 20 dBm 20 for A11 = 20 dBm
A1 (dBm)

B2 (dBm)

40 40

60 60

80 80

100 100
0 1 2 3 4 5 0 1 2 3 4 5
frequency (GHz) frequency (GHz)

Figure 2.1 Small (top row) and large (bottom row) stimulus and response.

then contains signicant energy at harmonic frequencies. Generally, the energy is

distributed in the frequency domain on a grid, k, dened by the stimulus, 1:
k k1 , k 2 N: 2:2
Two indices are required to specify the location (port) and frequency component of the
variables in the frequency domain: the rst index identies the port and the second
index identies the component on the frequency grid.
The complex amplitude (hence the magnitude and the phase) of the response at each
harmonic at any of the ports depends on the magnitude and on the phase (and hence on
the complex amplitude) of the stimulus, A1,1. This dependence is formally represented
in equation (2.3), which has been generalized to an arbitrary number of ports, N:
Bp, k F p, k A1, 1 , p 1, 2, . . ., N ,
2:3
k 1, 2, . . ., K:
22 X-parameters fundamental concepts

0 20
A1 DUT B2
10 1 2 10
+ Ref
20 A1 B2 0
A1

B2
30 10

40 20
50 30
0 1 2 3 4 5 0 1 2 3 4 5
frequency (GHz) frequency (GHz)

Figure 2.2 Nonlinear spectral map.

The notation Fp,k in equation (2.3) species the scattered wave behavior of the amplier
at port p and harmonic index k as a particular nonlinear function of the input stimulus.
There are N  K such functions if we consider N ports and K harmonics per port.
Equation (2.3) represents a nonlinear spectral mapping of the stimulus, A1,1, to the
responses of the system, Bp,k; this is depicted graphically in Figure 2.2.

2.3 Multi-harmonic spectral maps

The generation of scattered waves with energies at harmonic multiples of the incident
stimulus frequency raises important questions for the cascading of multiple DUTs. The
output waves of DUT1 become the incident waves for DUT2 in the simple cascade of
2-ports shown in Figure 2.3.
Fortunately, for incident signals with spectral components on a harmonic grid, the
scattered waves will also have spectral components precisely on the same frequency
grid. This is true for components such as diodes, transistors, and power ampliers. For
now, we only consider these types of devices that do not produce outputs in the absence
of input stimuli (that is, unlike an oscillator).1
This means that the general problem, for this class of excitations and DUTs, reduces
to characterizing the functional dependence of the scattered waves on all the spectral
components of the incident signals. That is, the inputoutput functional relationship that
completely characterizes the DUT steady-state nonlinear behavior can be formulated as
shown in equation (2.4):
Bp, k F p, k A1, 1 , A1, 2 , . . ., A1, K , A2, 1 , A2, 2 , . . ., A2, K , . . . , AN , 1 , AN , 2 , . . ., AN , K :
2:4
For an N-port device, for given bias conditions at all DUT ports, there are N  K
complex-valued functions, Fp,k, each of which is dened on N  K complex amplitudes
representing the magnitudes and phases of the spectral components incident at each port.

1
Phenomena such as sub-harmonic oscillations are neglected for now.
 2.3 Multi-harmonic spectral maps 23

Reflected

1 2

Incident

Figure 2.3 Cascaded 2-ports showing incident and scattered waves.

Considering that the DC-bias stimuli at each of the N ports, DCSp, p 1, 2,. . ., N,
may vary, these additional independent variables must also be considered, as shown in
(2.5). The DC response, DCRp, at each of the N ports of the DUT also depends on all of
these arguments and is identied separately in (2.5):
DCRp DCF p DCS 1 , .. .,DCS N , A1, 1 , ...,A1, K , A2, 1 , . .., A2, K , .. .,AN , 1 , AN , 1 , ..., AN , K ,
Bp, k F p, k DCS 1 , . .., DCS N , A1, 1 , ...,A1, K , A2, 1 , ..., A2, K , .. .,AN , 1 , AN , 1 , ...,AN , K :
2:5
Equation (2.5) can be written in the more compact form shown in (2.6), where the
generic notations q 1, N and h 1, K are used to show that q takes all integer values
between 1 and N, and k takes all integer values between 1 and K:



DCRp DCF p DCS q , q 1, N , Ar, h , r 1, N , h 1, K ,


 2:6
Bp, k F p, k DCS q , q 1, N , Ar, h , r 1, N , h 1, K :
Though explicit and unambiguous, the notation of (2.6) is still cumbersome. Therefore,
unless specically needed, the index ranges will be dropped, reducing (2.6) to the
simpler form (2.7):

 
DCRp DCF p DCS q , fAr, h g ,

  2:7
Bp, k F p, k DCS q , fAr, h g :
The arguments of the functions in (2.7) must be understood to include all the DC stimuli
and all the RF incident waves at all ports, as explicit in (2.6). There are distinct
functions DCFp(.) and Fp,k(.) corresponding to each value of the indexes p and k. The
DC stimuli, DCSp, may be either voltages or currents. The DUT response at DC at each
of its ports, DCRp, will be current or voltage, whichever complements the stimulus at
the respective port.
For a DC-voltage bias at port p, the functions DCFp(.) are denoted FIp(.), and the
response is the current, Ip, as shown in (2.8):
I p FI p :: 2:8
For a DC-current bias at port p, the functions DCFp(.) are denoted FVp(.), and the
response is the current, Vp, as shown in (2.9):

V p FV p :: 2:9
24 X-parameters fundamental concepts

I2
+

I1
+
A1,k B2,k
V2
V1

B1,k A2,k

- -

Figure 2.4 Characterization of nonlinear 2-port in AB / IV space on harmonic grid.

As a result, the functions DCFp(.) (which could be either FVp(.) or FIp(.)) are real
valued, and both sets of functions DCFp(.) and Fp,k(.) are dened on N  K complex
amplitudes and N real-valued bias conditions. Equations (2.7), in mixed DC-bias and
wave variables, provide a complete representation for the N-port, equivalent to using
spectral components of voltages and currents.
There is an additional constraint on the physically admissible functions dening
relations (2.6). This constraint follows directly from the DUT property of time invari-
ance. This constraint will be applied after simplifying the general expression (2.6) for
practical applications.
As an example, consider the characterization of a transistor as a nonlinear 2-port in
the AB / IV space (the generalized power waves and DC currentvoltage), as depicted
in Figure 2.4.
The set of equations describing the steady-state behavior of the transistor, considering
only three harmonics have signicant energy levels, is shown in (2.10):

V 1 FV 1 I 1 , V 2 , A1, 1 , A1, 2 , A1, 3 , A2, 1 , A2, 2 , A2, 3 ,

B1, k F 1, k I 1 , V 2 , A1, 1 , A1, 2 , A1, 3 , A2, 1 , A2, 2 , A2, 3 ,
2:10
I 2 FI 2 I 1 , V 2 , A1, 1 , A1, 2 , A1, 3 , A2, 1 , A2, 2 , A2, 3 ,
B2, k F 2, k I 1 , V 2 , A1, 1 , A1, 2 , A1, 3 , A2, 1 , A2, 2 , A2, 3 :

Note that, for port 1, current is the independent variable and voltage is the dependent
variable, while, for port 2, the roles are reversed: voltage is the independent variable
and current is the dependent variable. This is consistent with the most commonly
used biasing scheme for a bipolar junction transistor (BJT), the device used in this
example. Mathematically, however, either choice for the independent and dependent
DC-bias variables can be selected at each port, independently, to describe accurately
the behavior of the device, regardless of the actual biasing scheme used in the
application.
 2.5 Cascading DUTs 25

I3
+
V3
-
3
1 2

+ A1
Gnd
B2 +
(Bg, Gs) GL

- -
B1 A2

Figure 2.5 Load and source mismatch.

2.4 Load- and source-mismatch effects

The set of equations (2.7) is used to characterize the steady-state behavior of a nonlinear
system under source- and load-mismatch conditions. This case is shown in Figure 2.5,
where both S and L may have any arbitrary value. The wave generated and scattered
by the source is marked by Bg in this gure.
The boundary conditions for source and load show the waves in complex-vector
notation, as in the example for the A1 wave in (2.11):

A1 A1, 1 , A1, 2 , A1, 3 : 2:11

The system of equations completely describing such behavior under mismatch

conditions is composed by equations (2.7) plus the source and load conditions, as
shown in (2.12), where only three harmonics are considered:

A1, k Bg, k S , k B1, k ,

A2, k L B2, k ,
I3 FI 3 V 3 , A1, 1 , A1, 2 , A1, 3 , A2, 1 , A2, 2 , A2, 3 , 2:12
B1, k F 1, k V 3 , A1, 1 , A1, 2 , A1, 3 , A2, 1 , A2, 2 , A2, 3 , k 1, 2,
B2, k F 2, k V 3 , A1, 1 , A1, 2 , A1, 3 , A2, 1 , A2, 2 , A2, 3 , k 1, 2:

The solution to the system of equations (2.12) is the steady-state behavior of the system
under the specied multi-harmonic source and load mismatch.

2.5 Cascading DUTs

The set of equations (2.7) is also used to nd the steady-state behavior of a cascade of
nonlinear systems. A generic example is shown in Figure 2.6. When two components
are cascaded, an internal node is created at which the circuit laws KVL and KCL must
be satised.
26 X-parameters fundamental concepts

(a)
I1(1) I2(1) I1(2) I2(2)
1 [F(1)] 2 1 [F(2)] 2
+ + + +
A1,k (1) B2,k (1) A1,k (2) B2,k (2)

V1 (1) V2 (1) V1 (2) V2 (2)

B1,k (1) - - A2,k (1) B1,k (2) - - A2,k (2)

(b)

I1(1) I2(2)
1 2

(1)
+ +
A1,k B2,k (2)
[Fcomposite]
V1 (1) V2 (2)

B1,k (1) - - A2,k (2)

Figure 2.6 Cascade of two nonlinear systems. (a) Cascade of two DUTs. (b) The equivalent
network, yielding the same response as the cascade of two DUTs.

It is a straightforward exercise to write the equations that follow from applying KVL
and KCL at the internal node for wave variables in the frequency domain, and DC
voltage and current. These relationships are given in (2.13),

1 2
V2 V1 ,
1 2
I 2 I 1 ,
1 2 2:13
B2, k A1, k ,
1 2
A2, k B1, k :

which dene a set of 2  (K 1) equations:

 two real-numbered equations, one for DC voltage and one for DC current;
 2  K complex equations, two per harmonic index.
The set of equations (2.13) is augmented by

 the boundary conditions, at the source and at the load, and

 the model (2.6) for each of the two nonlinear elements in the cascade.
The boundary conditions are determined by the external interactions at the source and
load. The source boundary conditions, considered known, are shown in (2.14):
 2.6 Example: cascading two RF PAs 27

1
V 1 V source ,
1
I 1 I source ,
1 2:14
A1, k Asource ,
1
B1, k Bsource :

The load boundary conditions, considered known, are shown in (2.15):

2
V 2 V load ,
2
I 2 I load ,
2 2:15
A2, k Aload ,
2
B2, k Bload :

The behavior of the rst nonlinear element in the cascade is modeled in (2.16):
n o n o
1 1 1 1
DCRp DCF p DCS q , Aq, k ,
n o n o 2:16
1 1 1
Bp, k F p, k DCS 1q
, A q, k :

The behavior of the second nonlinear element in the cascade is modeled in (2.17):
n o n o
2 2
DCR2
p
DCF 2
p
DCS q
, A q, k ,
n o n o  2:17
2 2 2
Bp, k F p, k DCS 2q
, A q, k :

The solution to the system of nonlinear equations (formed by (2.13), (2.14), (2.15),
(2.16), and (2.17)) is the steady-state solution of the cascade at the internal node. This
contains the waves and the DC voltages and currents at the internal node, identied
in (2.13).
The entire system of equations is very similar to the one formulated for the linear
operation. Unlike the case of Chapter 1, where the device was linear in all spectral
components (except DC), elimination of the internal variables at the internal node
dened by a cascade now requires the solution of nonlinear equations. If each wave
variable has several harmonic components, there are multiple nonlinear equations in
many variables, expressed by (2.7) in the general case, to be solved self-consistently.

2.6 Example: cascading two RF power amplifiers with independent bias

The mathematical formalism used to describe the nonlinear behavior should be exible
enough to allow formulation of the cascade equations when the DC and RF are present
at different physical ports. A typical example is shown in Figure 2.7, where the
harmonic indices are suppressed and the A and B waves are considered as vectors at
the given numbered ports.
28 X-parameters fundamental concepts

I3
+
V3
-

3
1 2
A1 Gnd B2

B1 A2

Figure 2.7 Two-port (w.r.t. RF) component.

I3(1) I3(2)
+ +
V3(1) V3(2)
- -

3 3
1 2 1 2

+ A1(1) Gnd B2(1) A1(2) Gnd B2(2) +

- B1(1) A2(1) B1(2) A2(2) -

(Bg, GS) GL
[F(1)] [F(2)]

Figure 2.8 Cascading two ampliers.

The situation of Figure 2.8 is obtained when cascading two such components, which
depicts the common case of a cascade of two ampliers working on a passive load.
The techniques described in both Sections 2.4 and 2.5 are used to solve for the
steady-state behavior of the cascade under mismatched source and load conditions.
The internal node is now described by (2.18):
 
1 1 1 1 1 2
B2 F 2 V 3 , A1 , A2 A1 ,
  2:18
2 2 2 2 2 1
B1 F 1 V 3 , A1 , A2 A2 :

The source and load conditions are described by (2.19):

1 1
A1 Bg B1 S ,
2 2
2:19
A2 B2 L ,
 2.7 Relationship to harmonic balance 29

where Bg is the wave generated by the source when terminated on its own matched
load, S .
For a complete description of the cascade, the DC equations need to be added, as
shown in (2.20):
 
1 1 1 1
I 3 DCF 3 V 3 , A1 , A2 ,
  2:20
2 2 2 2
I 3 DCF 3 V 3 , A1 , A2 :

For the particular example shown in Figure 2.8, solving (2.18) and (2.19) gives the
complete RF solution for the cascade, independent of (2.20), given that there is no
interaction through the bias network between the two ampliers. Equations (2.20) can
then be used to determine the current through each amplier at the corresponding
individual operating conditions.
In the general case, interaction through the power distribution network may also be
present, and then (2.18), (2.19), and (2.20) must be solved simultaneously, together with
the other equations describing the behavior of the network connecting the bias ports of
the two ampliers.
The only assumption about cascading nonlinear blocks dened this way is that
enough harmonics are used to characterize the nonlinear mappings (2.4), and that
enough samples of the values of the dependent and independent variables are obtained
(if measurement-based) to ensure accuracy over the ranges of signals likely to be used in
an application. Of course, here we are still also assuming that the signals have spectral
components on a harmonic grid and that each component is properly characterized on
the same grid. This restriction will be relaxed in a later chapter.
It is important to understand that no approximation is made in formulating the system
of equations (2.18) and (2.19). A complete solution of wave values (voltages/currents)
at all interfaces in the cascade is determined for arbitrary values of the amplitude and
phase of all harmonics in the stimulus.

2.7 Relationship to harmonic balance

Solving for cascaded nonlinear blocks as described in Section 2.5 is exactly analogous
to the process a harmonic balance simulator uses to solve the nonlinear algebraic circuit
equations for designs with multiple components in the frequency domain. The only
difference is that here we are dening the components nonlinear relationships in AB
space for the wave variables, whereas most simulators work in IV space.
Harmonic balance analysis, introduced in the mid 1980s, was a great advance for the
steady-state circuit simulation of RF and microwave nonlinear analog circuits [1]. It
provided more efcient solutions compared to conventional time-domain methods (e.g.
SPICE) for most practical nonlinear microwave circuits containing time constants
spanning many orders of magnitude. A big difference, however, is that for most
harmonic balance simulators the model description of the nonlinear components
transistors and diodes is still formulated as time-domain nonlinear ordinary
30 X-parameters fundamental concepts

differential equations. The harmonic balance algorithm therefore has to transform these
model equations into nonlinear algebraic equations in the frequency domain, and back
again, for every iteration of the nonlinear solver.
Modeling and measurement methods have lagged far behind the simulation algo-
rithms. One can say that by measuring or otherwise prescribing the functions (2.4), one
is writing the model equations directly in the mathematical language native to the
simulation algorithm that most efciently solves the problem. The X-parameter
approach is precisely this for large-signal multi-tone steady-state (and later modulated)
signals.

2.8 Cross-frequency phase

Equation (2.3) includes a phase relationship between signals at different frequencies. It

is important to clarify the denition of this phase relationship, called cross-frequency
phase. This is discussed in Section 2.8.2, but the concept of commensurate signals
needs to be introduced rst.

2.8.1 Commensurate signals

Sinusoidal signals in a set are commensurate if all signals in the set are located on a
frequency grid f k kf f und , k 2 N, dened by one frequency, ffund, called the funda-
mental. This concept will be revisited in Chapter 5 for the case of multiple
fundamentals.
A set of commensurate sinusoidal signals allows the denition of the cross-frequency
phase between any pair of two signals in the set.
For a nonlinear amplier under the stimulus of a single large tone, all responses are a
combination of the harmonics of the stimulus. All harmonics are sinusoids located
on a frequency grid dened by the stimulus, as shown in (2.2); hence they are
commensurate.

2.8.2 Definition of cross-frequency phase

In general, any combination of commensurate signals is periodic, with the period set by
the fundamental frequency of the set, in this case ffund.
The cross-frequency phase of any signal in a set of commensurate signals is the phase
of that signal as measured at a moment in time when the signal at the fundamental
frequency has a phase of 0o.
The cross-frequency phase is not dened for non-commensurate signals.
One important aspect is that the signals in the set do not have to coexist in the same
node of the network. They just need to coexist in time so that they can be measured
simultaneously.
For example, let us consider the measurement process of the steady-state behavior of
an amplier under a single-tone stimulus. The input and output waves exist at different
 2.8 Cross-frequency phase 31

1 2
+ A1
Gnd
B2

20 20
A1 B2
10 10
A1 (dBm)

B2 (dBm)
0 0

10 10

20 20

30 30
0 1 2 3 4 0 1 2 3 4
frequency (GHz) frequency (GHz)
0.2 0.2

0.1 0.1
a1 (sqrt(W))

a1 b2
b2 (sqrt(W))

0.0 0.0

0.1 0.1

0.2 0.2
0.6 1.0 1.4 1.8 2.2 2.6 0.6 1.0 1.4 1.8 2.2 2.6
time (ns) time (ns)

Figure 2.9 Example used in understanding the cross-frequency phase denition.

nodes of the network (input and output), but they coexist at the same time. All these
waves form a set of commensurate signals, with the fundamental frequency set by the
stimulus signal (which exists at the input port).
Once the waves are generated, the cross-frequency measurement process is not related
to the DUT itself (the amplier), but simply captures the relationship between the signals
themselves. This process is not inuenced in any way by the origin of the waves.
The DUT itself and its behavior have no inuence on how the cross-frequency phase
is measured. The cross-frequency phase is a very important aspect in the characteriza-
tion of the DUT behavior.
The output scattered wave, B2, is used as an example in this discussion, but the
concepts are applicable to all waves scattered from the ports of a network under the
excitation of a single-tone stimulus.
Figure 2.9 shows a simple example of an amplier used as the system under test for a
better understanding of the cross-frequency phase denition. The stimulus, A1, and the
32 X-parameters fundamental concepts

response at port 2, B2, are shown in both frequency and time domain. The time-domain
view is helpful in understanding the cross-frequency phase relationship as used in the
analysis of the behavior of nonlinear systems.
The stimulus, A1, denes the fundamental frequency of the set of commensurate
signals contained in the response of the amplier.
Without any loss of generality, only three harmonics are used in this discussion,
considering all higher harmonics have negligible power levels.
The waves A and B considered in Figure 2.9 are measured in either the time domain
or the frequency domain.
If the measurement is carried out in the time domain, the waveforms are captured for
one complete period of the fundamental frequency, starting at an arbitrary moment in
time, marked as tm in Figure 2.10, considered the time reference. Figure 2.10 displays
two periods for a better understanding of the overall phenomenon. The period captured
for measurement is highlighted. Due to the periodic nature of the response, a Fourier-
series decomposition of the time-domain signals yields the frequency-domain represen-
tation, complete with amplitude and phase information.
The coefcients of the Fourier series are calculated as shown in (2.20), where T is the
period of the fundamental:
t m
T T
m 2 m 2 m
b2, k t ejk T ttm b t t m ejk T t dt
2 2
B2 , k dt 2:20
T T 2, k
tm 0
m
The phase of each individual spectral line, B2, k ,
reported by (2.20), is the phase
measured at the moment tm, the beginning of the time interval over which the integra-
tion is performed for the Fourier analysis, as also shown in Figure 2.10. In general, the
moment tm can have any arbitrary value because the measurement moment is not
synchronized with the waveform itself.
The time-domain waveform is related to the frequency-domain spectrum through the
Fourier series, as shown in (2.21):
!
X 3
m jk t
b2 t Re B2, k e : 2:21
k1
m m
In (2.21), all are peak values. The spectral components B2, k are obtained directly,
B2, k
with both magnitude and phase, if the waves A and B considered in Figure 2.9 are
measured in the frequency domain at the moment in time tm. The process of measuring
these waves is discussed in detail in Chapter 4.
m
The superscript m in (2.21) indicates that B2, k is the measured wave component at an
arbitrary moment in time, the moment when the measurement is taken. This superscript
is used to distinguish the variables actually measured from the variables that will be
used to characterize the intrinsic properties of the system under test, which will not take
the superscript m.
In the case shown in Figure 2.10(a), the time reference (the beginning of the time
interval used for the integration period in the Fourier analysis) is located at an arbitrary
moment in time, tm. The Fourier series in (2.21) can be reformatted as shown in (2.22)
by explicitly showing the phase of each frequency component:
 2.8 Cross-frequency phase 33

(a) (b)
0.2 0.2

0.1 b2 0.1 b2

0.0 0.0

0.1 0.1

0.2 0.2
0.6 1.0 1.4 1.8 2.2 2.6 0.6 1.0 1.4 1.8 2.2 2.6
time (ns) time (ns)

0.2 0.2
a1=a1,1 a1=a1,1

0.1 0.1

0.0 0.0

0.1 0.1

0.2 0.2
0.6 1.0 1.4 1.8 2.2 2.6 0.6 1.0 1.4 1.8 2.2 2.6
time (ns) time (ns)
TimeReference TimeReference
0.2 0.2
b2,1 b2,2 b2,3 b2,1 b2,2 b2,3
0.1 0.1

0.0 0.0

0.1 0.1

0.2 0.2
0.6 1.0 1.4 1.8 2.2 2.6 0.6 1.0 1.4 1.8 2.2 2.6
time (ns) time (ns)

Figure 2.10 Shifting of time reference for proper denition of cross-frequency phase (two periods
are displayed, the measured period is highlighted). (a) Arbitrary time reference. (b) Time shift
such that fundamental a1,1 has phase 0 .

      
 m  jm j1 t  m  jm j2 t  m  jm j3 t
bt Re B2, 1  e 2 , 1 e B2, 2  e 2 , 2 e B2, 3  e 2 , 3 e : 2:22

The phases may have any values because the reference is an arbitrary moment in time.
The time reference can be shifted to another position in time, which is specic
and more convenient: the moment in time when the fundamental has a phase of 0 .
34 X-parameters fundamental concepts

This time shift can be achieved mathematically, without any additional measurements,
as shown below.
In this treatment of the topic of nonlinear behavior, a cosine-based phase denition is
used. As a result, the new time reference, delayed by , is the moment in time when the
fundamental sinusoid has its peak value. This situation is shown in Figure 2.10(b).
The Fourier series becomes
  m
  m
  m

 m   m   m 
bt  Re B2, 1  e j 2, 1 1 e j1 t B2, 2  e j 2, 2 2 e j2 t B2, 3  e j 2, 3 3 e j3 t ,

2:23
where the time-shift property of the Fourier series was used.
The time shift, , is determined by the condition of the fundamental phase to be 0o
after the time shift, as shown in equation (2.24):
m
m 2 , 1
2, 1  1 0 ) : 2:24
1
The phases of all other harmonics after the time shift can now be calculated, and, after
using (2.24) and (2.2), the new Fourier coefcients are shown in equation (2.25):
  m m
 
 m   
B2, k B2, k  e j 2, k k2, 1 B2, k  e j2, k : 2:25

The phases of the Fourier-series coefcients in equation (2.25) are the cross-frequency
phases with respect to the fundamental. These values can now be used to determine the
cross-frequency relative phase between any two signals in the set of commensurate signals.

2.9 Basic X-parameters for multi-harmonic multi-port stimulus

The functions Fp,k have been used until now to describe the behavior of the system. This
method requires knowledge of the functions for all possible combinations of values for their
multiple arguments (the DC stimuli and the complex multiple incident RF waves at all ports).
The X-parameters have been developed such that they are determined on a minimal
but comprehensive set of conditions, and they can still describe the intrinsic behavior of
the system under any other conditions of use.
In the general case, multiple harmonics are applied simultaneously to all the ports of
the DUT. They are all commensurate signals on the same harmonic grid.
The DUT responds to this multi-harmonic stimulus with scattered waves at its ports,
on all harmonics, Bp,k. This response depends on all incident waves applied by the
stimulus, in their complex form.
To clarify these ideas, a two-port DUT is considered here together with a stimulus
containing only the rst two harmonics, A1,1, A1,2, A2,1, and A2,2, simultaneously
incident on each of the two ports, as shown in (2.26):

Bp, k F p, k A1, 1 , A1, 2 , A2, 1 , A2, 2 : 2:26

 2.9 Multi-harmonic multi-port stimulus 35

2.9.1 Time invariance and related properties of Fp,k(.) functions

As already discussed in Chapter 1, we consider the case of time-invariant systems.
The time-invariance condition imposes some useful properties upon the nonlinear
functions Fp,k.
Considering the response at port 2 of the DUT, b2(t), due to the stimulus applied
at port 1, a1(t), the behavior of the amplier is described by an operator O, as shown
in (2.27):
 
b2 t O a1 t : 2:27
If the stimulus is delayed by a time , the response must be delayed by the same time
duration , as shown in (2.28):
 
b2 t  O a1 t  : 2:28
The wave b2(t) is the summation of the waves at port 2 on all harmonic frequencies,
which means that all harmonics would be delayed by the same amount of time, :
X
K
b2 t  b2, k t  : 2:29
k1

In the frequency domain, the delayed version of the response can be expressed as
shown in (2.30):

Bp, k ejk F p, k A1, 1 ej1 , A1, 2 ej2 , A2, 1 ej1 , A2, 2 ej2 : 2:30

Combining (2.30) and (2.26), and using the frequency-grid denition in (2.2), the
functions Fp,k must satisfy the relationship shown in (2.31):
 
F p, k A1, 1 ej1 , A1, 2 ej1 , A2, 1 ej1 , A2, 2 ej1
2 2
2:31
F p, k A1, 1 , A1, 2 , A2, 1 , A2, 2 ejk :

This leads us to the conclusion that the nonlinear operators Fp,k must satisfy the general
property expressed in (2.32):

2
2 
F p, k A1, 1 e j , A1, 2 e j , A2, 1 e j , A2, 2 e j

k 2:32
F p, k A1, 1 , A1, 2 , A2, 1 , A2, 2 e j , 8 2 R:

This property can now be used to separate the dependence on the magnitude and phase
of the fundamental frequency, as shown in (2.33):

2
2 
F p, k A1, 1 ejA1, 1 , A1, 2 ejA1, 1 , A2, 1 ejA1, 1 , A2, 2 ejA1, 1

k 2:33
F p, k A1, 1 , A1, 2 , A2, 1 , A2, 2 ejA1, 1 , 8 2 R:

The notation P is introduced by (2.34):

P e jphaseA1, 1 : 2:34
36 X-parameters fundamental concepts

Using the notation introduced in (2.34) and considering A1, 1 ejA1, 1 jA1, 1 j, equation
(2.33) can be reformulated as shown in (2.35):
  
F p, k A1, 1 , A1, 2 , A2, 1 , A2, 2 F p, k A1, 1 , A1, 2 P2 , A2, 1 P1 , A2, 2 P2 Pk : 2:35

Equation (2.35) and its implications are very signicant. First, the property of taking the
dependence of the phase (A1,1) outside the arguments of the Fp,k functions enforces the
time invariance of the model formulation. This also shows that not any complex
function can represent a time-invariant system. Only functions for which (2.35) is
possible may be considered in the nonlinear system modeling process.
A second signicance of (2.35) is that the set of functions Fp,k depends on one fewer
real argument than might be assumed by counting the arguments of (2.26). This
reduces the dimensionality of the state space over which the system needs to be
characterized. This becomes a very important practical benet, saving characterization
time and measured data le size, and improving the ultimate speed of simulation when
the model is used in design.

2.9.2 Definition of X-parameters and X-parameter behavioral model

The X-parameters for the multi-harmonic stimulus can now be dened as shown in (2.36):
  
FB
X p, k A1, 1 , A1, 2 P2 , A2, 1 P1 , A2, 2 P2 F p, k A1, 1 , A1, 2 , A2, 1 , A2, 2 : 2:36
Pk
The set of X-parameters can now be used to formulate the equations of a behavioral
model that can calculate the response of the system to any steady-state stimulus
containing energy at any of the three harmonics initially considered. The model
equations are shown in (2.37), where the number of harmonics has been extended to K:
( )!
X
K
FB  
B p X 
A1, 1 , A q, k P k
Pk k : 2:37
p, k
k1 q, k 6 1, 1

The X-parameters are not dependent on the phase of the fundamental due to the
multiplication of each harmonic complex-valued argument with the coefcient Pk,
which time shifts each harmonic to the moment when the fundamental has a phase of 0o,
effectively removing the dependence on the phase of the fundamental.
Although the X-parameters do not depend on the phase of the fundamental, the
X-parameter behavioral model in (2.37) does capture the dependence of the response,
FB
Bp(), on the phase of A1,1 through the multiplication of X p, k with the factor Pk.
The DC behavior must be added to the set of equations shown in (2.37) to obtain a
complete description of the nonlinear behavior of the DUT. This is shown in (2.38):
( )!
FDCR

 k
DCRp X p DCS q , jA1, 1 j, A q, k P ,
( q , k 6 1, 1
)! 2:38
FB
 k
Bp, k X p, k DCS q , jA1, 1 j, A q, k P P :
k

q, k 6 1, 1
 2.9 Multi-harmonic multi-port stimulus 37

The parameters XFDCR

p
in (2.38) are dened directly by the functions DCFp(.) in (2.6),
with the forms shown in (2.39) for voltage bias and in (2.40) for current bias, as per
(2.8) and (2.9):

XFDCR
p
: XFI
p
: FI p :, 2:39

XFDCR
p
: XFV
p

: FV p :: 2:40

The system of equations in (2.38) is a behavioral model that completely characterizes

the nonlinear behavior of the DUT. This model is known as the X-parameter model. We
use the term model with no implication of any approximation of the system descrip-
tion beyond the postulate of time invariance.
The model is valid for all power levels and phases of the K harmonics. Note that the
X-parameter values depend only on the amplitude of the fundamental, but they depend
on both the amplitude and phase of the harmonics.
The set of basic X-parameters is thus composed of the terms identied in (2.41):
( (
DCS p X FDCR
p
stimulus ; X -parameters FB p 1, N , k 1, K : 2:41
Ap, k X p, k

2.9.3 Example: a set of X-parameters

For example, the general model in (2.38) can be explicitly written as shown in (2.42) for
the following specic case:

 a two-port DUT,
 with three-harmonic stimulus applied (incident) at both port 1 and port 2,
 with only two harmonics containing signicant energy in the response (scattered
from ports 1 and 2),
 with voltage bias on port 1, and
 with current bias on port 2.
FI

I1 X 1 VDC 1 , IDC 2 , jA1, 1 j, A1, 2 P2 , A1, 3 P3 , A2, 1 P1 , A2, 2 P2 , A2, 3 P3 ,
FB

B1, 1 X 1, 1 VDC 1 , IDC 2 , jA1, 1 j, A1, 2 P2 , A1, 3 P3 , A2, 1 P1 , A2, 2 P2 , A2, 3 P3 P1 ,
FB

B1, 2 X 1, 2 VDC 1 , IDC 2 , jA1, 1 j, A1, 2 P2 , A1, 3 P3 , A2, 1 P1 , A2, 2 P2 , A2, 3 P3 P2 ,
FV

V2 X2 VDC 1 , IDC 2 , jA1, 1 j, A1, 2 P2 , A1, 3 P3 , A2, 1 P1 , A2, 2 P2 , A2, 3 P3 ,
FB

B2, 1 X 2, 1 VDC 1 , IDC 2 , jA1, 1 j, A1, 2 P2 , A1, 3 P3 , A2, 1 P1 , A2, 2 P2 , A2, 3 P3 P1 ,
FB

B2, 2 X 2, 2 VDC 1 , IDC 2 , jA1, 1 j, A1, 2 P2 , A1, 3 P3 , A2, 1 P1 , A2, 2 P2 , A2, 3 P3 P2 :
2:42

The set of basic X-parameters for this example is shown in (2.43), where the arguments
(the independent variables) are not shown for convenience of notation.
38 X-parameters fundamental concepts

8 8 FI
>
> >
> X
> VDC 1
> > 1FB
>
>
> >
> X 1, 1
>
> A1, 1 >
>
>
> >
> FB
>
> A1, 2 >
> X
< < 1, 2
A1, 3
stimulus ; X -parameters FV : 2:43
>
> IDC 2 >
> X2
>
> >
>
>
> A2, 1 >
> FB
>
> >
>
X 2, 1
>
> A2, 2 >
>
>
> > X FB
>
: A2, 3 : 2, 2

2.10 Physical meaning of the basic X-parameters

The basic X-parameters for the case of multi-harmonic stimulus of a multi-port network
are dened by the mathematical relationships shown in (2.36), and the X-parameter
behavioral model (which calculates the nonlinear behavior of the system based on the
X-parameters) is shown in (2.38). It is very useful to understand the physical meaning
of these mathematical formulations.

2.10.1 Reference stimulus and response

Insight into this physical interpretation can be gained by observing that the response of
the system is identical to the values of the basic X-parameters if the coefcient P has a
value of 1, as shown in (2.44):
8 ( )!
>
>

> FDCR k
< DCRp X p
> DCS q , jA1, 1 j, A q, k P
( q , k 6 1, 1
)!
for P 1 )
>
> FB

>
> k
: Bp, k X p, k DCS q , jA1, 1 j, A q, k P :
q, k 6 1, 1
2:44
Considering the denition (2.34) of the factor P, condition P 1 practically means that
the phase of the fundamental, phase (A1,1), must be 0o. The stimulus for this particular
case is called the reference stimulus, and it is shown in (2.45):
( )


DCS q , jA1, 1 j, A q, k Pk : 2:45
q, k 6 1, 1

The response of the system to the reference stimulus is called the reference response.
Any test stimulus with an arbitrary value of phase(A1,1) but with the same phase
relationships between the harmonics is a delayed version of the reference stimulus and
results in a delayed version of the reference response due to the time-invariance
condition discussed in Section 2.9.1.
The above property can be used to dene the concept of equivalent stimuli. Two
stimuli are said to be equivalent if they are delayed versions of each other.
 2.11 Using the X-parameter behavioral model 39

2.10.2 Physical interpretation

From (2.44) it follows that the physical interpretation of the basic X-parameters is that
they are the frequency-domain components of the reference response generated by the
reference stimulus equivalent to the test stimulus.
It also follows that the X-parameter behavioral model in (2.38) has an easy physical
interpretation. The steady-state response of the system to an arbitrary test stimulus can
be calculated by applying a corresponding time shift to the reference response generated
by the reference stimulus equivalent to the test stimulus.

2.11 Using the X-parameter behavioral model

Lets have a closer look at how the response of the system is calculated by the
X-parameter behavioral model.
Once the phase of the fundamental, phase(A1,1), in the test stimulus is known,
the reference stimulus is obtained by shifting the test stimulus back in time with the
duration corresponding to phase(A1,1), as determined in (2.46):

phaseA1, 1
: 2:46
1
This time shift is equivalent to a phase shift on each harmonic, k, of the test stimulus, as
determined in (2.47):

phase shift on harmonic k k : 2:47

This phase shift is implemented in the model as a multiplication with the coefcient
Pk, as shown in (2.48), where the negative sign for the exponent means a shift back
in time:
( )

 k
DCS q , jA1, 1 j, A q, k P : 2:48
q, k 6 1, 1

The set of signals in (2.48) comprises the reference stimulus equivalent to the test
stimulus.
Considering the X-parameters of the system under test are known, they are the
frequency-domain components of the reference response (the response to the reference
stimulus (2.48)), as shown in (2.49):
(
X FDCR
p
FB : 2:49
X p, k

The response to the test stimulus is thus obtained by time shifting the reference response
(2.49) forward in time with the same value , thus bringing it back in time alignment
with the test stimulus. This is equivalent to phase shifting each harmonic of the
F
reference response, X p, k , with a corresponding phase k. The phase shift is
40 X-parameters fundamental concepts

implemented as a multiplication with the coefcient Pk, as shown in (2.50), where the
positive sign of the exponent means a shift forward in time:
(
X FDCR
p
FB : 2:50
X p, k P k

The entire sequence of steps described in (2.46) to (2.50) is represented mathematically

by the X-parameter behavioral model in (2.38). Also, this sequence of steps represents
the conceptual methodology for measuring X-parameters, which is explained in detail in
Chapter 4.
The X-parameters are thus a description of the intrinsic steady-state behavior of the
system under test, and they can be used to calculate the steady-state response to an
arbitrary stimulus of the same harmonic content as the reference stimulus that generated
the set of X-parameters.

2.11.1 Example: amplifier with source and load mismatch

To clarify these ideas, let us consider the example of an amplier under a single-tone
stimulus working under mismatched conditions at both ports (source and load mis-
match), as shown in Figure 2.11.
Although a single tone is applied by the signal source at port 1 only, the stimulus
comprises incident waves at both ports 1 and 2. The incident wave at port 2, A2, occurs
due to the load mismatch, which reects energy on all harmonics back towards the
amplier output.
Note that wave A1, incident on the input port, is not a pure sinusoidal signal, although
the source itself applies a single large tone. This is due to the nonlinear behavior of the
amplier itself combined with the source mismatch. The amplier internally generates
harmonics which are scattered from both ports 1 and 2. Some of the harmonic energy
scattered from port 1 is reected back towards the amplier due to the source mismatch,
thus generating the distortion of the wave A1 observed in Figure 2.11.
The behavior of the amplier is observed (or measured) at the arbitrary moment in
time, tm, as depicted in Figure 2.11(a).
If the measurement takes place in the frequency domain, this means that both
stimulus and response spectra are measured at the moment tm. If the measurement takes
place in the time domain, this means both stimulus and response waveforms are
measured (recorded) starting at the moment tm for a duration equal to a time period of
the stimulus.
The amplier is characterized by its X-parameters, which are the frequency-domain
representation of the response to the reference stimulus. The Fourier decomposition
of the test stimulus reveals the harmonic content, as shown in Figure 2.11(b).
The reference stimulus is obtained by time shifting the test stimulus until the funda-
mental has a 0o phase at the initial moment of the observation interval, as shown
in Figure 2.11(c).
The response to the reference stimulus is known from the X-parameter values.
 2.11 Using the X-parameter behavioral model 41

+ b2 +
a1
Z=(20j 10) Ohm Z=(10+j 20) Ohm
b1 a2
Freq[1]=1GHz
P[1]=polar(dbmtow(0),0) - -

(a)
0.050 0.10

0.025 0.05
a1 (sqrt(W))

a2 (sqrt(W))
0.000 0.00

0.025 0.05

0.050 0.10
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
time (ns) time (ns)

(b)
0.050 0.10

0.025 0.05
a2 (sqrt(W))
a1 (sqrt(W))

0.000 0.00

0.025 0.05

0.050 0.10
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
time (ns) time (ns)

(c)
0.050 0.10

0.025 0.05
a1 (sqrt(W))

a2 (sqrt(W))

0.000 0.00

0.025 0.05

0.050 0.10
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
time (ns) time (ns)

Figure 2.11 (a) Test stimulus, complete waveforms. (b) Test stimulus, Fourier decomposition.
(c) Reference stimulus, Fourier decomposition.
42 X-parameters fundamental concepts

(a)
0.050 0.2

0.025 0.1
b1 (sqrt(W))

b2 (sqrt(W))
0.000 0.0

0.025 0.1

0.050 0.2
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
time (ns) time (ns)

(b)
0.2 0.2

0.1 0.1
b1 (sqrt(W))

b2 (sqrt(W))

0.0 0.0

0.1 0.1

0.2 0.2
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
time (ns) time (ns)

(c)

0.2 0.2
b2 (sqrt(W))
b1 (sqrt(W))

0.1 0.1

0.0 0.0

0.1 0.1

0.2 0.2
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
time (ns) time (ns)

Figure 2.12 (a) Reference response, Fourier decomposition. (b) Reference response, complete
waveforms. (c) Response to test stimulus, complete waveforms.

The reference response is shown in Figure 2.12(a), using its Fourier decomposition. It
is obtained by converting to the time domain the spectral components represented by the
X-parameters of the amplier. The full waveforms of the reference response, B1 and B2,
are shown in Figure 2.12(b), after combining all spectral components.
 2.12 Summary 43

+ a1 b2 +
Z=(20j 10) Ohm Z=(10+j 20) Ohm
Freq[1]=1 GHz b1 a2
P[1]=polar(dbmtow(0),0) - -

0.050 0.10
a1 (sqrt(W))

0.025 0.05

a2 (sqrt(W))
0.000 0.0

0.025 0.05

0.050 0.10
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
time (ns) time (ns)
0.2 0.2

0.1 0.1
b1 (sqrt(W))

b2 (sqrt(W))

0.0 0.0

0.1 0.1

0.2 0.2
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
time (ns) time (ns)

Figure 2.13 Steady-state behavior.

Finally, the response to the test stimulus is obtained by shifting the reference response
back in time until the initial moment in time of the fundamental period corresponds to
the measurement moment, tm, as shown in Figure 2.12(c).
The steady-state behavior of the amplier, complete with both stimulus and response,
is shown in Figure 2.13.

2.12 Summary

Nonlinear components exhibit behavior, such as compression and harmonic generation,

which is not describable by S-parameters. For the important practical case of signals dened
on a harmonic grid, nonlinear spectral mappings dened on incident-wave variables and
DC-bias conditions can describe the steady-state DUT behavior completely. Nonlinear
system behavior can be predicted without further approximation by cascading the
nonlinear description of the constitutive components. One-tone X-parameters are
derived by applying the principle of time invariance to the nonlinear mapping.
44 X-parameters fundamental concepts

Exercises

2.1 A nonlinear spectral map dened by (2.51),

B2, 1 GA1, 1  A31, 1 , 2:51
is proposed to model a time-invariant DUT. Prove that (2.51) is not time
invariant.
2.2 Prove that an alternative model, given by (2.52), denes a time-invariant map:

B2, 1 GA1, 1  jA1, 1 j2 A1, 1 : 2:52

2.3 Reduce (2.52) to the form given by (2.53) and nd an explicit expression
for X2,1(.):

B2, 1 X 2, 1 jA1, 1 jP1 : 2:53

References

[1] J. Verspecht, Describing functions can better model hard nonlinearities in the frequency
domain than the Volterra theory, Ph.D. thesis annex, Vrije Univ. Brussel, Belgium, Nov.
1995; available at http://www.janverspecht.com/skynet/Work/annex.pdf.

Additional reading

K. Kundert, The Designers Guide to SPICE and Spectre. Norwell, MA: Kluwer Academic
Publishers, 1995.
D. E. Root, J. Horn, T. Nielsen, et al., X-parameters: the emerging paradigm for interoperable
characterization, modeling, and design of nonlinear microwave and RF components and
systems, in IEEE Wamicon2011 Tutorial, Clearwater, FL, Apr. 2011.
D. E. Root, J. Verspecht, D. Sharrit, J. Wood, and A. Cognata, Broad-band, poly-harmonic
distortion (PHD) behavioral models from fast automated simulations and large-signal vectorial
network measurements, IEEE Trans. Microw. Theory Tech., vol. 53, no. 11, pp. 36563664,
Nov. 2005.
J. Verspecht and D. E. Root, Poly-harmonic distortion modeling, IEEE Microwave, vol. 7, no.
3, pp. 4457, June 2006.
3 Spectral linearization approximation

3.1 Simplification of basic X-parameters for small mismatch

The considerations used for the denition of X-parameters apply to the steady-state
behavior of time-invariant nonlinear components with incident (and hence also scat-
tered) waves on a harmonic frequency grid. The generality of the formalism comes at
the cost of considerable complexity. Each spectral map is a nonlinear function of every
applied DC-bias condition and all the magnitudes and phases of each spectral compon-
ent of every signal at every port. Sampling such behavior in all variables for many ports
and harmonics would be prohibitive in terms of data acquisition time, data le size, and
model simulation speed.
Fortunately, in most cases of practical interest, only a few large-signal components
need to be considered with complete generality, while most spectral components can be
considered small and dealt with by methods of perturbation theory. Specically, we can
apply the following methodology.

Identify the few large tones that drive the main nonlinear behavior of the system.

Identify the nonlinear spectral map dened when only these few large tones drive the
system (without any of the other tones identied as small-level signals). This non-
linear spectral map represents a specic steady state the system arrives at due to the
large-signal stimulus.

Linearize the spectral maps around this specic steady state dened by the few
important large tones.

Consider all the rest of the signals (the small signals) to be treated linearly based on
the linearized map.
This methodology approximates the complicated multi-variate nonlinear maps with
much simpler nonlinear maps, dened on the much smaller number of selected large
tones only, and simple linear maps accounting for the contributions of the many tones
with small amplitudes. These approximations result in a dramatic reduction of com-
plexity while providing an excellent description of many important practical cases.
The simplest such approximation is to linearize around a DC operating point and
assume all RF spectral components are treated linearly. This is just the S-parameter
approximation.
In this section, it is assumed that there is one large RF signal that must be treated
generally and that the remaining spectral components can be treated as small. This is
46 Spectral linearization approximation

often the case for power ampliers, at least those nearly matched to 50 , where the
main stimulus signal is a narrow-band modulation around a xed carrier incident at the
input port. The DUT response includes all the harmonics. If the DUT is not perfectly
matched at the output due to an imperfect load impedance, there will be reections at
the fundamental and the harmonics, back into the DUT. These mismatches will be
considered small perturbations to simplify the description. Similarly, scattered waves at
the harmonic frequencies at the input could be reected back into the DUT if it is not
perfectly matched at the input due to an imperfect source impedance. These incident
signals can also be treated as small in many applications. The validity of the approxi-
mation depends on whether these signals are small enough such that their contribution
appears only linearly. In later sections, this assumption will be relaxed, in particular to
deal with arbitrary load dependence at the output port.
The practical example described above of an amplier operating under mismatched
conditions is the same case studied in Section 2.10. It will be demonstrated how the
complex multi-variate nonlinear map can be simplied when only one incident tone is
large, the fundamental applied at the input. All other tones in the stimulus (all harmonics
incident on the ports) have small amplitudes due to the small source and load
mismatches.
Before proceeding with linearizing the spectral map around the nonlinear steady state,
some properties of the functions establishing the nonlinear spectral map need to be
claried, specically the analyticity of these functions. This is important because the
linearization requires the differentiation of these complex functions of complex variables.

3.1.1 Non-analytic maps

Physically, the large sinusoidal incident signal at the fundamental causes the device to
vary periodically in time. The simultaneous presence of small incident signals at other
frequencies will produce output signals at all sum and difference frequencies of the large
and small tones. Since all incident tones are on a harmonic grid, the self-mixing of
each small tone will produce an output at each harmonic. As will be seen in this section,
the complex amplitude of the spectral response at each harmonic will be a function of
the complex amplitude and the complex conjugate of the amplitude of the small tone.
The physical situation is equivalent to a mixer with a large LO signal and small RF
tone. To clarify the ideas, consider the case of the ideal mixer in Figure 3.1.
Without any loss of generality, the RF and LO input signals are sinusoidal and can be
written as in (3.1):
ARF e j0 t ARF ej0 t
RF in t ReARF e j0 t ,
2
3:1
ALO e j1 t ALO ej1 t
LOin t ReALO e j1 t :
2

The output (the IF) is the product of the time-domain signals, as shown in (3.2):
IF out t LOin t  RF in t: 3:2
 3.1 Simplification of small mismatch 47

Complex amplitude = ALOARF

Complex amplitude = ALOA*RF

0 0
10 10
20 20
30 + LO
+ 30

B2 (dBm)
A1 (dBm)

40 A1 B2 40
50 50
60 + 60
70 70
80 A3 80
90 90
100 100
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
frequency (GHz) frequency (GHz)

0
10
Complex amplitude = ARF 20
A3 (dBm)
30
40
50
60
Complex amplitude = ALO 70
80
90
100
0.0 0.5 1.0 1.5 2.0 2.5 3.0
frequency (GHz)

Figure 3.1 Ideal mixer demonstrating non-analytic frequency mappings.

After simple algebraic transformations using (3.1), it can be shown that (3.2) can be
expressed as shown in (3.3):
BIF e j t BIF ej t BI F e j t BI F ej t
IF out t : 3:3
2 2
The complex amplitudes BIF and BIF, together with the corresponding output fre-
quencies, are shown in (3.4):
ALO ARF
BIF ,
2
ALO ARF
BIF , 3:4
2
1 0 ,
 1  0 :
Expression (3.3) shows that the output contains two frequencies, and , as
emphasized after converting (3.3) into (3.5):



IF out t Re BIF e j t Re BIF e j t : 3:5
This well-known behavior of the ideal mixer is displayed graphically in Figure 3.1,
where the complex amplitudes of the input and output waves are shown.
So, in the frequency domain, we can write the mapping in (3.6):
BIF ALO ARF FALO , ARF : 3:6
The important conclusion is that if we associate the complex variables ALO and ARF with
the (positive) single-sided spectrum of incident signals, the complex map F(.), in (3.6),
48 Spectral linearization approximation

associated with the (positive) single-sided output spectrum, is non-analytic because of

the appearance of the conjugate term, ARF .
This is important when we linearize these maps in the perturbation theory to follow.
We therefore consider single-sided spectral maps that depend independently on the real
and imaginary parts of each of their complex arguments, as shown in (3.7):
BIF ALO ARF
ReALO jImALO  ReARF  jImARF 3:7
FReALO , ImALO , ReARF , ImARF :
Equivalently, the independent variables of the spectral maps may be considered func-
tions of their arguments and their conjugates independently, as shown in (3.8):
BIF ALO ARF FALO , ALO , ARF , ARF : 3:8
There is one more important observation. The argument of non-analyticity of the
nonlinear spectral mapping functions F(.) holds true even for the case of the RF
and LO having the same frequency. When 1 0,  0 rad/s and the mixer has a
DC response. Based on the above mathematical analysis, the amplitude of the DC
response still depends on the complex-conjugate term ARF , as shown in (3.6). This
proves that, in the most general sense, the nonlinear mapping functions are non-
analytic for the DC response also when considered as functions of the complex
amplitudes ALO and ARF.
An alternative description can be given in terms of double-sided spectral maps that
are analytic in each complex variable.

3.1.2 Large-signal operating point

Let us consider the response of the multi-port system to a multi-harmonic stimulus, as
described by the X-parameter model derived in Section 2.9.2 and shown in (2.38). For
convenience, this model is shown again here, in (3.9):
0 8 91
  < =
DCRp X FDCR @ DCS q , jA1, 1 j, A q, k P k A
,
p
: ;
q, k 6 1, 1
0 8 91 3:9
  < =
Bp, k X p, k @ DCS q , jA1, 1 j, A q, k Pk APk :
FB
: ;
q, k 6 1, 1

The simplication described in the preceding section refers to the case where all signals
are small, except one. For convenience, and without any loss of generality, we assume
this signal is applied at port 1, on the fundamental frequency, so it has the complex
envelope A1,1.
At the limit, all the small signals have zero power, which means that the response of the
system in the absence of the small signals depends only on the large-signal magnitude, as
shown in equation (3.10). It is easy to recognize that this is the X-parameter model for a
multi-port system under one large-tone stimulus, as shown in (3.10):
 3.1 Simplification of small mismatch 49

  
DCRLS
p X FDCR
p DCS q , jA1, 1 j, 0, 0, . . . , 0 ,
  3:10
LS FB  
Bp, k X p, k DCS q , jA1, 1 j, 0, 0, . . . , 0 Pk :

The superscript LS in (3.10) identies this as the response to the large-signal stimulus
only. This is the large-signal steady state of the system. The small signals will create a
perturbation of the system around this state. Due to its signicant role in the perturb-
ation theory discussed next, we need to identify this large-signal steady state.
The large-signal steady state, comprising the totality of the large-signal stimulus and the
large-signal response, is the large-signal operating point, usually referred to as the LSOP.
The LSOP is thus the state identied in (3.10) and formally shown in (3.11), in which
RFS stands for RF stimulus:
8  
>
> DCS LSOP DCS q ,
>
>
>
> LSOP
A1, 1 ,
< RFS
  
LSOP LSOP FDCR 3:11
>
> DCR X DCS q , jA 1, 1 j ,
>
>
p p
 
>
> FB  
: BLSOP X DCS q , jA1, 1 j Pk :
p, k p, k

As discussed in Section 2.10, a time shift does not change the state of the system. In other
words, equivalent stimuli generate equivalent responses, which are all delayed versions
of each other and thus represent the same basic large-signal steady state of the system.
In a more general sense, multiple large signals may be applied simultaneously, without
any of the small signals. For example, the fundamental may be applied at both input and
output simultaneously, as is the case for power-amplier devices that work on a large load
mismatch, so there is a large wave reected by the load which becomes incident on the
output port. The combination of all the large signals creates a large-signal steady state
around which the small signals (the other harmonics which are not large) create a perturb-
ation. The LSOP in this case is dened in the same way: the totality of the large-signal
stimulus and the corresponding response. The only difference between this case and the
case of a single large tone is the higher complexity of both the stimulus and the response.
Formally, the LSOP for the general multi-large-tone stimulus can be identied as
shown in (3.13), where refRFS (LSOP) is the reference stimulus equivalent to RFS(LSOP)
(i.e. the time-delayed version of RFS (LSOP) that has phase(A1,1) 0):
 
refRFS LSOP all large Aq, k Pk , 3:12
8  
>
> DCS LSOP DCS q ,
>
>  
>
> LSOP
all large Aq, k ,
< RFS
 
LSOP LSOP FDCR LSOP LSOP 3:13
>
> DCR X DCS , refRFS ,
>
>
p p
 
>
>
: BLSOP X
FB
DCS LSOP , refRFS LSOP Pk :
p, k p, k

There is one more convention to be set before proceeding further. This relates to some
simplied notations that we use sometimes from here on. As shown in (3.13), the formal
50 Spectral linearization approximation

identication of the LSOP may be quite cumbersome. In order to simplify the formulae,
the following conventions are used where needed.

The notation LSOPS is used for the large-signal stimulus (which is the stimulus that is
part of the LSOP), which is formally identied in (3.14):
(  
DCS LSOP DCS q ,
LSOPS   3:14
RFS LSOP all large Aq, k :

The notation LSOPR is used for the large-signal response (which is the response that
is part of the LSOP), which is formally identied in (3.15):
(

DCRLSOP
p X FDCR
p DCS LSOP , refRFS LSOP ,
LSOPR LSOP FB
 3:15
Bp, k X p, k DCS LSOP , refRFS LSOP Pk :

The LSOP is the totality of LSOPS and LSOPR, as shown in (3.16):

LSOP fLSOPS, LSOPRg: 3:16
Note that the LSOP is set by the stimulus (i.e. by the LSOPS), but it contains both the
stimulus and the response (i.e. both LSOPS and LSOPR).
It is sometimes useful to refer directly to the specic reference stimulus that identies
the LSOP, complete with DC and RF stimulus. This is identied as shown in (3.17):
n o
refLSOPS DCS LSOP , refRFS LSOP : 3:17

Another useful observation is that the LSOP contains both the DC operating point and
the RF operating point. Sometimes it is useful to identify separately these elements of
the LSOP. The notations shown in (3.18) are used when needed:
(
DCS LSOP DCS p ,
DCOP
DCRLSOP
p X FDCR
p refLSOPS ;
(  
RFS LSOP all large Aq, k , 3:18
RFOP LSOP FB
Bp, k X p, k refLSOPS Pk ;
LSOP fDCOP, RFOPg:
The case of a single large tone and multiple small tones is used in the following
application of the perturbation theory, due to its simpler form, but the derivation is
easily extendable to an arbitrary number of large and small tones.

3.2 Adding small-signal stimuli (linearized nonlinear spectral mapping)

The objective in this section is to nd a more convenient way of formulating the model
shown in (3.9) by taking advantage of the knowledge that some of the signals and the
stimulus are small. This leads to the formulation of a linearized spectral map around the
LSOP set by the large-signal stimulus.
 3.2 Adding small-signal stimuli 51

The linearization process leads to the introduction of additional X-parameter terms.

Although the number of terms in the set of X-parameters increases signicantly, the
nal model is simpler and easier to formulate and use in a simulation environment. In
addition, the new terms introduced in the linearization process provide signicantly
more insight into the nature of the nonlinear behavior and how the system evolves from
one large-signal state to another one.
The LSOP is the steady state achieved by the system when no small-signal perturb-
ations are applied.
When small perturbations are applied on harmonic frequencies around the LSOP
stimulus, the system will respond with variations of the scattered RF waves and its DC
responses around the values at the LSOP.
The variations of the scattered waves are captured by new X-parameter terms that are
referred to as RF-interaction terms.
The variations of the DC responses are captured by new X-parameter terms that are
referred to as RF-to-DC-interaction terms.
They will be studied individually.

3.2.1 Small-signal interactions: the RF terms

The case of a single large-tone stimulus is presented here to allow simplication of the
mathematical expressions. This does not represent any loss of generality. The results can
easily be extended to the more general case of a multi-harmonic stimulus.
The LSOP stimulus is shown in (3.19):
n  o
LSOPS DCS q , A1, 1 : 3:19

The corresponding reference stimulus is shown in (3.20):

n  o
refLSOPS DCS q , jA1, 1 j : 3:20

The scattered waves emerging from the system operating at the LSOP are shown in
(3.21):
( )!
Bp, k F p:k LSOPS, A q, k : 3:21
q, k 6 1, 1

The application of the perturbation theory evaluates the complex function Fp.k(.) of the
complex variables Aq,kPk, while the refLSOPS remains unchanged. The value of this
function is estimated around the xed refLSOPS using a series expansion based on the
partial derivatives. The calculation of the partial derivatives requires the Fp.k(.) function
to be analytical with respect to all independent variables.
The analysis of a mixer, presented in Section 3.1.1, showed that these nonlinear
mapping functions, Fp.k(.), are in fact non-analytic in the general case. As a result, they
must be considered complex functions depending on both sets of complex variables
Aq,kPk and (Aq,kPk)* in order to be analytic functions and thus allowing the series
expansion based on the partial derivatives.
52 Spectral linearization approximation

The scattered waves can then be represented around the LSOP as shown in (3.22):
0 8 91
< =
Bp, k F p, k @refLSOPS, A q, k Pk APk
: ;
q, k 6 1, 1
F p, k refLSOPS P k

qN 2   3 3:22
X F p, k  
l K
4 F p, k 

 Aq, l Pkl

 
l  
Aq, l Pkl 5:
q1
Aq, l Pl  Aq, l P
refLSOPS refLSOPS
l 1
q,l 6 1,1

The partial derivatives in (3.22), of the Fp.k(.) nonlinear mapping function, are used to
dene two sets of RF-interaction X-parameters of type S and of type T, as shown in
(3.23) and (3.24), respectively:
 
S F p, k  F p, k 
X p, k;q, l 
  Pl , 3:23
Aq, l Pl  Aq, l 
refLSOPS refLSOPS

 
F p, k  F p, k 
T 
X p, k;q, l 
    Pl : 3:24
Aq, l Pl refLSOPS Aq, l 
refLSOPS
FB
In addition, the rst term in (3.22) can easily be expressed using the X p:k , based on the
denition in equation (2.36).
With the above denitions and observations, (3.22) can be rewritten as shown in
(3.25), where the dependence of all X-parameters on the LSOP is also recognized:

FB
Bp, k X p:k refLSOPS Pk
qN
X
lK h i

S T
X p, k;q, l refLSOPS Aq, l Pkl X p, k;q, l refLSOPS Aq, l Pkl : 3:25
q1
l1
q, l 6 1, 1

3.2.2 Small-signal interactions: the DC terms

The system response to the small RF perturbations also includes small variations of the
DC-response components.
For DC-voltage-biased ports, the response is the DC current. Considering the den-
ition (2.39) of the X pFI term, the current response for an arbitrary port p can be written
as shown in (3.26), where all RF signals are small except for A1,1:


I p X pFI refLSOPS, A1, 2 P2 , A1, 3 P3 , A2, 1 P1 , A2, 2 P2 , A2, 3 P3

 3:26
FI p refLSOPS, A1, 2 P2 , A1, 3 P3 , A2, 1 P1 , A2, 2 P2 , A2, 3 P3 :
 3.2 Adding small-signal stimuli 53

The variation of the DC current Ip as the system is exercised around the LSOP by the
small RF perturbation can be evaluated by a series expansion of (3.26) using partial
derivatives.
In the same discussion presented in Section 3.1.1 for the mixer, it was proven
that the functions FIv(.) are non-analytic with respect to the variables Aq,k, so they
have to be considered functions of both Aq,k and their complex conjugates Aq, k in
order to ensure analyticity and make possible a series expansion based on partial
derivatives.
We consider a single-tone large-signal excitation for a simplied mathematical
formulation. This does not represent any loss of generality, and the results can be easily
extended to the more general case of a multi-harmonic stimulus.
The LSOP stimulus and corresponding reference stimulus for this case are shown
in (3.27):
n  o
LSOPS DCS q , A1, 1 ,
n  o 3:27
refLSOPS DCS q , jA1, 1 j :

The current response of the system operating at the LSOP is shown in (3.28):
( )!
I p FI p refLSOPS, A q, l : 3:28
q, l 6 1, 1

When the small-signal perturbation is superimposed over the LSOP stimulus, the
current will vary around its LSOP value, and its total value can be expressed as shown
in (3.29):
0 8 91
< =
I p FI p @refLSOPS, A q, l Pl A
: ;
q, l 6 1, 1
FI p refLSOPS
qN 2   3
lXK
FI  FI p 

4
p  l  l  5

l 
Aq, l P

 
  Aq, l P
q1
A q, l P Aq, l Pl 
refLSOPS refLSOPS
l1
q, l 6 1, 1

FI p refLSOPS
qN 0   1
lXK
FI  FI p 
@ 
Aq, l A:
p
 Aq, l   3:29
q1
A q, l  Aq, l 
refLSOPS refLSOPS
l1
q, l 6 1, 1

One important property of the mapping functions FIp(.) is that they always evaluate to a
real-numbered value because the result of the evaluation is a DC current.
The rst term in (3.29), FIp(refLSOPS), is a real number because it represents the DC
current at the LSOP (without any small-signal perturbations).
54 Spectral linearization approximation

It then follows that the sum of all the terms must also be a real number. However, it is
observed that each individual small-signal perturbation on each of the harmonics can be
applied individually, by itself, independent of all the other small signals on any other
harmonic frequencies. It is thus required that the sum of the two terms for each of the
harmonics must be a real number, as shown in (3.30):
0   1
 
@FI p  FI p 
Aq, l   Aq, l A 2 R
Aq, l  Aq, l  3:30
refLSOPS refLSOPS

8 p, q 1, N , l 1, K :
It can be mathematically proven that, for (3.30) to be true for any value Aq, l 2 C, it is
necessary and sufcient for (3.31) to be true:
!
FI p FI p
: 3:31
Aq, l Aq, l
Expression (3.30) can be rewritten as in (3.32) when using (3.31):
 !
FI p 
2  Re Aq, l 2 R: 3:32
Aq, l refLSOPS

A new X-parameter term is dened as shown in (3.33):


Y FI p 
X p;q, l refLSOPS 2 : 3:33
Aq, l refLSOPS

The total DC response of the DUT to the superposition of the small RF perturbation
over the large RF signal can thus be calculated as shown in (3.34):
( )!
I p FI p refLSOPS, A q, l Pl
q, l 6 1, 1
qN
X
lK   3:34
Y
X pFI refLSOPS Re X p;q, l refLSOPS Aq, l :
q1
l1
q, l 6 1, 1

For the case of current-bias ports, the DUT response is the port voltage. Following a
Z
similar process, a new X-parameter term, X p;q, l , is dened as shown in (3.35):

Z FV p 
X p;q, l refLSOPS 2 : 3:35
Aq, l refLSOPS
The FVp(.) functions must also satisfy condition (3.36) because the DC voltage is also a
real-numbered value:
!
FV p FV p
: 3:36
Aq, l Aq, l
 3.3 Physical meaning of the small-signal interaction terms 55

It follows that the total voltage response can be calculated as shown in (3.37):
0 8 91
< =
V p FV p @refLSOPS, A q, l Pl A
: ;
q, l 6 1, 1
qN
  3:37
X
lK
Z
X FV
p

refLSOP Re X p;q, l refLSOPS Aq, l :
q1
l1
q, l 6 1, 1

3.3 Physical meaning of the small-signal interaction terms

For practical use of the X-parameter model formulated by (3.25), (3.34), and (3.37),
it is important to understand the physical meaning of the small-signal interaction
terms.
The distinct X-parameters dening the spectral mappings linearized around a single-
tone LSOP can be interpreted as follows.
FB
The terms X p, k represent the set of mappings from a single-tone RF input to the
scattered output waves at port p and harmonic frequency index k for a system perfectly
matched at the output port and perfectly matched at each harmonic at all ports.
S T
There are two sets of functions, X p, k and X p, k , that determine the sensitivity to
mismatch of the system at ports p and harmonic index k. It is noted that, unlike
linear S-parameters, there are independent contributions to the scattered waves
S
proportional to both Aq,l and Aq, l . The terms X p, k are related to the so-called hot
S-parameters, but contain generalizations including harmonics. It will be seen later
T
the importance of the terms X p, k as the magnitude of the input drive, jA1,1j, becomes
large.
Y
The terms X p;q, l give the sensitivity of the DC-bias currents at port p to RF
Z
mismatches at port q at the harmonic index l. Similarly, the terms X p;q, l give the
sensitivity of the DC-bias voltages at port p to mismatches at port q at the harmonic
index l. These terms are essential for proper modeling of the match-dependent power-
added efciency (PAE) under drive, for example.
We note that, unlike linear S-parameters, X-parameters characterize port-to-port and
frequency-to-frequency interactions. That is, a mismatch at a given harmonic at one port
can affect the response at another port at each harmonic frequency as well as the DC-
bias condition.
Let us analyze how the scattered waves can be estimated by the model around the
LSOP using (3.25). When the coefcient P has a value of 1, meaning that the phase of
the fundamental, phase(A1,1), is 0 (the system is in the reference-LSOP state), the
waves can be expressed as the sum of various X-parameter terms. The small-signal
interaction terms are scaled by the small-signal stimuli or their complex-conjugate
value. This is shown in (3.38):
56 Spectral linearization approximation

P 1):
qN
X
l K h i
FB S T
Bp, k X p:k ref LSOPS X p, k;q,l ref LSOPS Aq,l X p,k;q,l ref LSOPS Aq,l :
q1
l 1
q,l 6 1,1

3:38
FB
The resulting wave, Bp,k, is composed of the LSOP term, the X p:k , plus contributions
due to the small-signal stimulus on each individual frequency component, Aq,l. Each of
these small-signal contributions are independent of each other and they depend linearly
only on the stimulus applied on the individual frequency. They can thus be separated by
applying the small-signal stimuli one at a time, on each frequency and port.
For example, let us consider that all small-signal stimuli are zero, except for the one
applied at port q on harmonic l, Aq,l. The wave scattered at port p on harmonic k, Bp,k, is
then calculated as shown in (3.39):
)
Ar, m 0, 8r, m 6 q, l
)
Aq, l 6 0
FB S T
Bp, k X p:k refLSOPS X p, k;q, l refLSOPS Aq, l X p, k;q, l refLSOPS Aq, l :
3:39
The small-signal response (i.e. the part of the response that occurs due to the small-
signal stimulus) is identied as bp,k, and it is the difference between the total response,
FB
Bp,k, and the large-signal response, X p:k refLSOPS , as shown in (3.40):
)
Ar, m 0, 8r, m 6 q, l
)
Aq, l 6 0
FB S T
bp, k Bp, k  X p:k refLSOPS X p, k;q, l refLSOPS Aq, l X p, k;q, l refLSOPS Aq, l :
3:40
S
From the formulation shown in (3.40), it follows that the X p, k;q, l term is an intrinsic
property of the system under test and that it represents the small-signal transfer function
from harmonic l applied at port q to harmonic k occurring at port p when the system
operates at its reference LSOP (refLSOP).
T
In a similar manner, from the formulation shown in (3.40), it follows that the X p, k;q, l
term is an intrinsic property of the system under test and that it represents the small-
signal transfer function with respect to the conjugate of the small signal from harmonic l
applied at port q to harmonic k occurring at port p when the system operates at its
reference LSOP (refLSOP).
As an example, let us consider an amplier, as shown in Figure 3.2. The stimulus A1
contains only large-signal energy on the fundamental frequency A1,1, and the stimulus
A2 contains only small-signal energy on the second harmonic A2,2. This case represents
a typical situation of an amplier under a stimulus coming from a very good source
(with negligible spectral content) and with a mismatch on the second harmonic.
 3.3 Physical meaning of the small-signal interaction terms 57

1 2
+ A1 Gnd B2

B1 A2

20 20
A1 A2
10 10
A1,1
0 0
A2,2

A2 (dBm)
A1(dBm)

10 10
20 20
30 30
40 40
50 50
0 1 2 3 4 0 1 2 3 4
frequency (GHz) frequency (GHz)

Figure 3.2 Amplier under second-harmonic load-pull conditions.

In many practical situations, this represents the setup for the second harmonic load
pull of an amplier.
The total response of the amplier under these conditions is calculated using
X-parameters. As an example, the wave scattered from port 2 (the output of the
amplier) on the fundamental frequency, B2,1, is observed as a function of the phase
of the small-signal stimulus, A2,2.
The actual response of the amplier is shown in Figure 3.3. As the phase of A2,2 varies, the
A2,2 vector describes a circle because its amplitude remains constant during this experiment.
The vector B2,1 does not describe a circle as the phase of A2,2 varies, as shown in
Figure 3.3. A zoomed-in view of the tip of the vector is shown on the lower left side of
the gure to provide a better understanding of the physical phenomenon.
It is important to understand how the X-parameter model reproduces this effect. The
calculation of B2,1 using the X-parameter behavioral model is shown in (3.41):
FB S T
B2, 1 X 2, 1 ref LSOPS X 2, 1;2, 2 ref LSOPS A2, 2 X 2, 1;2, 2 ref LSOPS A2, 2 : 3:41
According to this model, the total response is the sum of three vectors: FX, SX, and TX,
as shown in Figure 3.4, where the three vectors are identied in (3.42):
FB
FX X 2, 1 refLSOPS ,
S
SX X 2, 1;2, 2 refLSOPS A2, 2 , 3:42
T
TX X 2, 1;2, 2 refLSOPS A2, 2 :
58 Spectral linearization approximation

A2,2

B2 +
0.015 0.005
A2
-

B2,1
B2,1

0.01 0.01 0.20 0.15 0.10 0.05 0.00 0.05 0.10 0.15 0.20

Figure 3.3 Actual B2,1 response of the amplier versus phase of A2,2.

As the phase of the A2,2 stimulus varies, the SX and TX vectors rotate in opposite
directions, as shown in Figure 3.5. This is because the A2, 2 rotates in the opposite
direction to A2,2. The end result is that X S A2, 2 X T A2, 2 describes a contour, as
shown in Figure 3.4 and Figure 3.5. The center of this contour is dened by the FX
term, which does not vary when the phase of A2,2 changes because it represents the
LSOP response.
Figure 3.5 illustrates how the two components rotate to generate the contour centered
on the tip of the LSOP vector.
For more complex situations, when multiple small signals are applied at different
ports, the locus of the tip of the B2,1 vector may have various shapes. These shapes
are calculated by the corresponding X-parameter behavioral model by summing
multiple vectors proportional to the individual small-signal stimuli and their complex
conjugates.
 A2,2

B2 +
0.015 0.005
A2
-

B2,1
B2,1

SX

TX

0.01 0.01 0.20 0.15 0.10 0.05 0.00 0.05 0.10 0.15 0.20

FX

Figure 3.4 B2,1 response of the amplier versus phase of A2,2 calculated by the X-parameter
behavioral model.

B2,1

SX

0.01 0.01
TX

FX

Figure 3.5 The summation of the three vectors yielding the total B2,1 response calculated by the
X-parameter behavioral model. The circular arrows show the direction in which the vectors rotate
as the phase of A2,2 increases.
60 Spectral linearization approximation

3.4 Discussion: X-parameters and the spectral Jacobian

We see that the spectral linearization in (3.25), (3.34), and (3.37) reduces the complexity
of the nonlinear functions dening the scattered waves and DC responses to functions of
many fewer variables than the general (harmonic grid) case, like the one shown in
Section 2.9.3. Only DC-bias variables and the magnitude of the fundamental incident
wave at the fundamental frequency appear as arguments to nonlinear functions. All
other spectral components appear linearly in the equations.
The X-parameter functions (3.23), (3.24), (3.33), and (3.35) are related to partial
derivatives of the corresponding nonlinear functions, F(.), evaluated at the LSOP
established by the bias variables and the single large-tone stimulus and response.
These denitions provide a means of computing the X-parameters from any standard
nonlinear model, such as the example of the FET model in Chapter 1. Any model that
is dened in terms of conventional nonlinear equivalent circuit elements, with well-
dened and differentiable constitutive relations for currents and charges (or voltages
and uxes), can be converted into X-parameters. Since most component models, such
as transistor and diode models, are dened in terms of IV and QV relations, one can
transform these equations into the spectral domain and associate specic X-parameter
terms with the spectral Jacobian in AB space (wave variables). Since nonlinear
simulators compute the spectral Jacobian (in IV space) from knowledge only of these
constitutive relations and their partial derivatives, it is evident that X-parameters can
be computed in simulators by solving only the single-tone harmonic-balance problem.
That is, knowledge of the one-tone nonlinear map and the spectral Jacobian is
sufcient to compute the X-parameters that provide the DUT multi-linear response
to many RF signals, provided all but the one large tone can be considered to contribute
linearly to the output.

3.5 X-parameters as a superset of S-parameters

The X-parameter model describes the intrinsic behavior of a nonlinear system. Because
the nonlinear behavior is dependent on the level of the large-signal stimulus, it is
common practice to extract the X-parameters as a function of this variable (the level
of the large-signal stimulus).
At sufciently low levels, the stimulus becomes small enough to be considered a
small signal and the system behaves linearly. It is thus natural that X-parameters become
S-parameters at the limit situation of small-signal stimulus.
We consider the example of an amplier, as shown in Figure 3.6, to gain a better
understanding of this aspect.
The X-parameters of this amplier are extracted for a single-tone large-signal stimu-
lus applied at its input (port 1) with a power sweep for A1,1 from 40 dBm to 10 dBm
and considering ve harmonics.
The large-signal reference response for the rst three harmonics, as captured by the
X-parameters, is shown in Figure 3.7.
 3.5 X-parameters as a superset of S-parameters 61

Figure 3.6 An amplier working with a 50 source and a 50 load.

Figure 3.7 Large-signal reference response versus level of the large-signal stimulus,
as captured by the X-parameters.

As the level of the large-signal stimulus decreases, the levels of the harmonics
decrease faster than the fundamental. Below a certain level, the harmonics become
negligible with respect to the fundamental (say below 30 dBm), and the behavior of
the system should be well predicted by the linear network theory (S-parameters).
Also, as the level of the large-signal stimulus decreases, the response at each
harmonic should reach a point from where it should vary with a constant slope
proportional to the order of the harmonic. Figure 3.8 shows the slope with respect to
the level of the large-signal stimulus versus the level of the large-signal stimulus for the
rst three harmonics.
As expected, the slopes become constant at small signals (below approximately
30 dBm) and equal to 1 dB/dB, 2 dB/dB, and 3 dB/dB for the rst, second, and third
harmonic, respectively. This is the behavior expected in the linear region of operation.
62 Spectral linearization approximation

Figure 3.8 Slope with respect to the level of the large-signal stimulus versus large signal stimulus
of the X (F) terms for the rst three harmonics.

Figure 3.9 shows the typical variation of the magnitude of X (S) and X (T) parameters
versus the level of the large-signal stimulus. Two representative cases are shown:

parameters representing transfer functions at the same frequency (small-signal

response at the same frequency as the small-signal stimulus), and

parameters representing cross-frequency transfer functions (small-signal response at a
frequency different from that of the small-signal stimulus).
The two cases shown in Figure 3.9 are for:
S T

the X 2, 1;2, 1 and X 2, 1;2, 1 , which represent transfer functions from the wave at the
fundamental frequency incident at the output port to the wave at the fundamental
frequency scattered from the output port, and
S T

the X 2, 2;2, 1 and X 2, 2;2, 1 , which represent cross-frequency transfer functions from the
wave at the fundamental frequency incident at the output port to the wave at the
second-harmonic frequency scattered from the output port.
S
At low levels, the magnitude of X 2, 1;2, 1 becomes constant, while the magnitude of
T S
X 2, 1;2, 1 keeps decreasing, thus becoming negligible with respect to the X 2, 1;2, 1 .
S T
Similarly, the cross-frequency terms, X 2, 2;2, 1 and X 2, 2;2, 1 decrease continuously in
S
magnitude, thus becoming also negligible with respect to X 2, 1;2, 1 .
It is now important to observe what happens to the phase of the same-frequency X (S)
S
terms. The example of X 2, 1;2, 1 is shown in Figure 3.10. As expected, its phase also
becomes constant, independent of the level of the large-signal stimulus.
All of the above proves that at low levels of the large-signal stimulus the behavior
becomes linear, and this is accurately captured by the X-parameter model.
It is also expected that the values of the X (S) terms at the same frequency become
identical with the corresponding classical S-parameters. In our example, the output
reection coefcient and reverse isolation, both with respect to small-signal perturb-
S S
ations around the LSOP, are the X 2, 1;2, 1 and X 1, 1;2, 1 parameters, which become equal to
the S22 and S12 when the large-signal drive level becomes small enough.
 3.5 X-parameters as a superset of S-parameters 63

Figure 3.9 X (S) terms at the same frequency become constant in magnitude, and all cross-frequency
terms, both X (S) and X (T), become negligible at low levels of the large-signal stimulus.

Figure 3.10 The phase of the X (S) terms at the same frequency becomes constant at low levels
of the large-signal stimulus.

In addition, the large-signal transfer functions also become equal to the correspond-
ing S-parameter terms when the large-signal drive level becomes small enough. For our
LS
particular example, the large-signal gain, G21 , and the large-signal input reection
LS
coefcient, in , are dened as shown in (3.43):
64 Spectral linearization approximation

F
LS X 2, 1
G21 ,
AN 1, 1
F 3:43
LS X 1, 1
in :
AN 1, 1

Figure 3.11 shows a comparison between the X-parameters at low levels of the
large-signal stimulus and an independent measurement of the S-parameters. The
X-parameters not shown in Figure 3.11 become zero at low drive levels.
It is thus clear that X-parameters represent a complete description of the behavior of
the network, in both the nonlinear and linear regions of operation. X-parameters are thus
a superset of S-parameters. At low drive levels all X-parameters become zero except for
the ones that correspond to the S-parameters.

3.6 Two-stage amplifier design

One of the important design tasks is to predict the behavior of cascaded systems when
the models for the individual systems are known. As an example, the result of cascading
two ampliers, as depicted in Figure 3.12, is very well predicted when the system
operates in the linear region. This is achieved using the S-parameter formalism.
When the system is pushed into nonlinear operation, S-parameters are insufcient for
an accurate description of the cascade, but X-parameters have the full capability of
covering the compression region, in addition to the linear (small-signal) region.
Figure 3.12 shows a power sweep that extends up to approximately 3 dB gain
compression of the overall system.
The cascade of the X-parameter models of the two ampliers, shown in Figure 3.13,
predicts very well all harmonics, in both magnitude and phase, for all power levels
where the models have been measured.
As an example, Figure 3.14 shows the rst three harmonics scattered from the output
of the system into the load, in both magnitude and phase. The response predicted by the
cascaded X-parameter models is superimposed over the actual response of the system,
showing the excellent accuracy of the models.
It is important to understand and control the inter-stage interaction during the design
and manufacturing process of the system. The behavioral models must predict with
accuracy both the forward- and reverse-propagating waves at the interface between the
two ampliers. Figure 3.15 shows the rst three harmonics of the forward-propagating
waves, and Figure 3.16 shows the reverse-propagating waves. Both magnitudes and
phases are accurately predicted by the cascade of the X-parameter models.
This level of accuracy is also achieved under simultaneous source and load mismatch.
Such a performance has not been previously achieved by any other behavioral model
preceding X-parameters.
One of the questions often asked is how important are the X (T) parameters for the
accuracy of the model. One of the relatively simple tests that can emphasize their
Figure 3.11 X-parameters (lines) become S-parameters (dots) for low levels of the large-signal
stimulus.
66 Spectral linearization approximation

Figure 3.12 Two cascaded ampliers system gain in compression.

Figure 3.13 Two cascaded ampliers represented by their corresponding X-parameter models.

Figure 3.14 Two cascaded ampliers output harmonic waves predicted by X-parameters (line)
versus system response (symbols).
 3.6 Two-stage amplifier design 67

Figure 3.15 Two cascaded ampliers inter-stage forward waves predicted by X-parameters (line)
versus system response (symbols).

Figure 3.16 Two cascaded ampliers inter-stage reverse waves predicted by X-parameters (line)
versus system response (symbols).

importance is to use a model that does not have them. In the above example, the
X-parameter models of the two ampliers have been simplied by removing all the X (T)
parameters from the original models. A simulation has been run using the simplied
models with a source impedance of ZS (80  15j) and a load impedance of ZL
(30 25j) . The results for the inter-stage forward propagating waves are shown in
Figure 3.17.
There is a signicant inaccuracy in predicting the behavior of the system when the
X (T) parameters are not considered. This is true for all nodes in the cascade and for both
forward- and reverse-propagating waves. Two main reasons contribute to these
inaccuracies.
The rst reason is the way the model handles the load interactions. The impact of the
load mismatch across all harmonics is improperly evaluated by the model without the
X (T) terms. The rst amplier works on the input impedance of the second amplier,
which may show signicant mismatch, and the second amplier works on the system
68 Spectral linearization approximation

Figure 3.17 Two cascaded ampliers inter-stage forward waves predicted without X (T)
parameters (line) versus system response (symbols).

load, which may also show signicant mismatch at all harmonics. Thus, both models
cannot ensure good accuracy with respect to the load without the X (T) terms.
The second source of inaccuracies is the impact of the source harmonics. The rst
amplier creates harmonic content in the input scattered wave (the B0 in Figure 3.13),
due to its nonlinear behavior, which is then reected back into the system by the source
mismatch, thus contributing to the incident wave A0 (in Figure 3.13). In a similar way,
the second amplier contributes to the harmonic content of the B1 wave, which reects
back on the output impedance of the rst amplier and adds to the harmonic content
generated by the nonlinear behavior of the rst amplier itself. The model needs to
process correctly the entire harmonic content of the source, and this cannot be done
accurately without the X (T) parameters.
The above comments explain, and are themselves re-enforced by, the inaccuracies
observed in Figure 3.17.
Such attempts at formulating a simple extension of S-parameters into the nonlinear
region have been used in the past with limited success. Two such examples are large-
signal S-parameters and hot S-parameters, which are distinct concepts. Each of them
may describe the system with reasonable accuracy, but only over very restricted
conditions of applicability, such as very good terminations and inter-stage match, which
are rarely seen in practice. Two of the sources of their inaccuracy under broad range
conditions come from the missing X (T) terms and the missing cross-frequency inter-
action terms.
The full formulation of the X-parameters (i.e. including both X (S) and X (T) terms) is
thus essential to ensure the accuracy required by a successful design process.

3.7 Amplifier matching under large-signal stimulus

One of the strengths of the X-parameter model is the capability of accurately handling
the potential mismatch that occurs at various nodes in the system for both the input and
output ports of an amplier.
 3.7 Amplifier matching under large-signal stimulus 69

3.7.1 Output matching and hot-S22

Performance of an amplier under small-signal conditions is completely determined by
its S-parameters. The source and load conditions (source and load match) are uniquely
identied for optimum operation, and a well-dened design process based on S-
parameters has been successfully used for many years.
A similar approach has been sought for the nonlinear behavior, but it has not been
successful until recently due to the inaccuracy of the modeling capabilities offered by
technologies prior to X-parameters.
One important design objective is to understand how to achieve the output match and
how best to measure and model it.
The output-match measurement has to be carried out under the so-called hot
conditions. This means that the output reection response to an output incident signal
must be measured when the amplier is driven into the nonlinear behavior by a signal at
its input port. This requirement poses the practical difculty of separating the two
components of the output signal:

the large signal that appears due to the input signal, and

the output signal that appears due to the reection of the output incident wave on the
output impedance (the output reection coefcient) of the amplier.

This is difcult because the two components occur at the same frequency, and older
network analyzers do not have the capability of separating these waves. This feature is
now supported by the nonlinear network analyzer, which is the basic instrument
supporting X-parameter measurement.
Figure 3.18 depicts the output waves of an amplier under large-signal drive condi-
tions from a single tone applied at its input port.
A small-signal incident tone is applied as an incident wave on the output port at
the same frequency with the large signal scattered from the output due to the large-
signal stimulus at port 1. This is identied as small-signal wave a2,1. The phase
of a2,1 is swept for a constant magnitude, and thus its complex envelope vector
describes a circle.
The scattered wave response at port 2 varies in both magnitude and phase, and its
complex amplitude vector describes an ellipse. The X-parameter model accurately
predicts this response, as shown in Figure 3.18.
The small-signal response at port 2, identied as b2,1, is the additional response
occurring at wave B2,1 on top of the LSOP (established by the large-signal drive) due to
the incident wave a2,1. As the phase of a2,1 varies, b2,1 varies in both magnitude and
phase; its complex amplitude vector describes an ellipse centered on the tip of the
complex amplitude vector representing the LSOP response.
Knowing the incident and reected waves, a2,1 and b2,1, respectively, the output
reection coefcient of the amplier under large-signal drive conditions can be deter-
mined. The result is displayed in Figure 3.18.
It is very interesting to observe that the output reection coefcient is not constant: it
depends on the phase of the incident wave a2,1, which is a condition outside the
70 Spectral linearization approximation

Small-signal stimulus at fundamental output

+ A1 A2 +

Small-signal a2,1
- -

0.0010

0.0012
0.0000

0.0002

0.0004

0.0006

0.0008
0.0012

0.0010

0.0008

0.0006

0.0004

0.0002
B1 B2

Small-signal output response at fundamental Output response at fundamental

Small-signal b2,1

B2,1

0.0006 0.0004 0.0002 0.0000 0.0002 0.0004 0.0006 0.20 0.15 0.10 0.05 0.00 0.05 0.10 0.15 0.20

Output reflection coefficient at fundamental

1.0
0.9

1.2
0.8

1.4
0.7

1.6
0.6

1.8
2.0
0.5

0.4
Output reflection coefficient

3.0
0.3
4.0

5.0
0.2

10
0.1

20
0.1

0.2

0.3

0.4

0.5

0.6

0.7
0.8
0.9
1.0

1.2

1.4
1.6
1.8
2.0

3.0

4.0
5.0

10

20

20

0.1
10

0.2 5.0

.0
4
.3
0
.0
3
.4
0
.5

.0
0

2
.8
0.6

1
1.6
0.7

1.4
0.8

1.2
0.9

1.0

Figure 3.18 Output reection coefcient at fundamental frequency of an amplier under single-
tone large-signal drive, predicted by X-parameters (line) versus system response (symbols).
 3.7 Amplifier matching under large-signal stimulus 71

Small-signal output response at fundamental Output reflection coefficient at fundamental

1.0
0.9

1.2
0.8

1.4
0.7

1.6
0.6

1.8
2.0
0.5
Output reflection coefficient
0.4

3.0

0.3
4.0

5.0
Small-signal b2,1

0.2

10
0.1

20

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8
0.9

2.0

3.0

4.0

5.0
1.0

1.2

1.4

1.6
1.8

20
10
0.0006 0.0004 0.0002 0.0000 0.0002 0.0004 0.0006
20

0.1
10

0.2 5.0

4.0
.3
0
.0
3
.4
0

.5

.0
0

2
.8
0.6

1
1.6
0.7

1.4
0.8

1.2
0.9

1.0
System response

X-parameters without X(T) terms

Figure 3.19 Left: response to small-signal perturbations around the LSOP predicted without X (T)
parameters (line) versus system response (symbols). Right: output reection coefcient predicted
without X (T) parameters (single point) versus system response (symbols) limitations of hot S22.

amplier itself. The output reection coefcient describes a circle in the Smith chart as a
function of the phase of a2,1.
It is thus clear that the output reection coefcient is not an intrinsic property of the
nonlinear amplier under large-signal drive because it depends on variables outside the
amplier itself (it depends on the load reection coefcient, L). This raises important
questions regarding what design process to follow when implementing the output-
matching networks for nonlinear systems.
It is also interesting to observe how far away the model would be if the X (T)
parameters were not considered. This scenario is shown in Figure 3.19.
This is a condition similar to the hot-S22 measurement that was intensively used
before the X-parameter model was introduced. As shown, the small-signal response,
b2,1, describes a circle as a function of the phase of a2,1, and the output reection
coefcient of the amplier under large-signal drive becomes a single point, thus
incorrectly suggesting it would be an intrinsic property of the amplier.
Based on this discussion, it is clearly concluded that the output reection coefcient
is not an intrinsic property of nonlinear systems and that load mismatch is one of the
external conditions that might inuence its value.
The consideration of both X (S) and X (T) parameters together makes it possible for the
X-parameter formalism to characterize accurately and completely the output reection
properties with respect to small-signal perturbations around the LSOP (like the load
mismatches).
Once again, the full formulation of the X-parameters proves to be essential in ensuring
the necessary accuracy in modeling the output reection coefcient of nonlinear systems.
72 Spectral linearization approximation

Figure 3.20 Modulated signal applied on the output of an amplier driven by a large
input tone.

The small-signal output reection coefcient of an amplier also depends on

the LSOP set by the large-signal stimulus. As an example, let us consider the
reection of a modulated signal, with an arbitrary complex modulation, applied
to the output of an amplier when the amplier is driven by a large single tone
applied at its input. The amplier is thus working on an active load with internal
impedance of 50 and with small-signal modulated available power, as shown in
Figure 3.20.
The time-domain waveform and the trajectory described by the complex envelope of
the modulated power wave incident on the output port of the amplier are shown in
Figure 3.21. In order to illustrate the complete generality of these concepts, the
modulation is ctitious and it has been synthesized such that the trajectory (the polar
plot) of its complex envelope resembles a smiley face.
The time-domain modulation waveform is a periodic signal, and the time-domain
representation in Figure 3.21 captures one period.
It is interesting to observe how the modulation of the reected signal (the small-signal
scattered wave at port 2, part of the B2,1 wave) is distorted as the large signal drives
the amplier into nonlinear behavior (into compression). This is shown in Figure 3.22
(a summary of which is depicted on the front cover).
The plots shown in Figure 3.22 represent the behavior of the actual amplier.
It is important to demonstrate that the X-parameter model shown in (3.25)
accurately reproduces this behavior. The response in Figure 3.22 is the B2,1
scattered wave.
 Figure 3.21 Small-signal modulated power wave incident on the output port of the amplier.

(a)

Figure 3.22 Small-signal output scattered wave at different levels of compression.

74 Spectral linearization approximation

(b)

Figure 3.22 (cont.)

Considering that in a hot-S22 experiment a small-signal stimulus is applied only at the

fundamental frequency, all Aq,l waves are zero, except for A2,1. The X-parameter model
for the B2,1 wave reduces to the form shown in (3.44):
FB S T
B2, 1 X 2, 1 refLSOPS P1 X 2, 1;2, 1 refLSOPS A2, 1 P0 X 2, 1;2, 1 refLSOPS A2, 1 P2 :
3:44
In order to demonstrate this capability, the X-parameter model of the amplier in Figure 3.20
S
is extracted as a function of the input incident power level. The variation of the X 2, 1;2, 1 and
T
X 2, 1;2, 1 parameters versus the power level of the input drive signal is shown in Figure 3.23.
S
As expected, at low drive levels, the X 2, 1;2, 1 has the dominant value when compared
T T
with the X 2, 1;2, 1 parameter. As the drive level increases, the X 2, 1;2, 1 parameter increases
S
at a faster rate than the X 2, 1;2, 1 and becomes comparable with it. The continuous
 3.7 Amplifier matching under large-signal stimulus 75

S T
Figure 3.23 X 2, 1;2, 1 and X 2, 1;2, 1 versus input drive level.

variation of the relative values between these two parameters, combined with the
formulation of the model itself, is the reason why the X-parameter model is capable
of reproducing the nonlinear behavior of the amplier with accuracy, as described in the
following gures.
The wave A2,1 incident on port 2 at the fundamental frequency is only the small signal
modulated with the ctitious modulation of the smiley face shown in Figure 3.21.
The small-signal response of the amplier to the small-signal stimulus A2,1 is
predicted by the model, as shown in (3.45):
FB S T
B2, 1  X 2, 1 ref LSOPS P1 X 2, 1;2, 1 ref LSOPS A2, 1 P0 X 2, 1;2, 1 ref LSOPS A2, 1 P2 :
3:45
At the power level of 20 dBm, the X (S) and X (T) parameters have the values shown in
(3.46):
AN 1, 1 20 dBm;
S
X 2, 1;2, 1 0:270=14 deg; 3:46
T
X 2, 1;2, 1 0:004=108 deg:
The small-signal response predicted by the model in (3.45) at 20 dBm input drive
level is shown in Figure 3.24. The amplier has a predominantly linear behavior. The
76 Spectral linearization approximation

Figure 3.24 Small-signal response at 20 dBm input drive level predicted by X-parameter
model (line) versus amplier response (symbols)

Figure 3.25 Small-signal response at 7 dBm input drive level predicted by X-parameter model
(line) versus amplier response (symbols).

T S
X 2, 1;2, 1 has negligible value as compared to X 2, 1;2, 1 , so the response is a linear
transformation of the stimulus, showing a magnication and phase rotation.
At the power level of 7 dBm, the X (S) and X (T) parameters have the values shown in
(3.47):
AN 1, 1 7 dBm,
S
X 2, 1;2, 1 0:425=0 deg, 3:47
T
X 2, 1;2, 1 0:090=120 deg:
The small-signal response predicted by the model in (3.45) at 7 dBm input drive level
T
is shown in Figure 3.25. The amplier has a slightly nonlinear behavior. The X 2, 1;2, 1 is
 3.7 Amplifier matching under large-signal stimulus 77

Figure 3.26 Small-signal response at 3 dBm input drive level predicted by X-parameter model
(line) versus amplier response (symbols).

S
no longer negligible compared to X 2, 1;2, 1 , so the response is a more complex trans-
formation of the stimulus, showing a magnication and a phase rotation that both
depend on the instantaneous phase of the stimulus. The smiley face is now taller
than it is wide.
At the power level of 3 dBm, the X (S) and X (T) parameters have the values shown in
(3.48):
AN 1, 1 3 dBm,
S
X 2, 1;2, 1 0:367=7 deg; 3:48
T
X 2, 1;2, 1 0:183=146 deg:

The small-signal response predicted by the model in (3.45) at 3 dBm input drive level is
shown in Figure 3.26. At this drive level, the amplier has a signicantly nonlinear
S T
behavior. The X 2, 1;2, 1 and X 2, 1;2, 1 parameters have comparable values, which result in a
signicant distortion of the waveform.
At the power level of 5 dBm, the X (S) and X (T) parameters have the values shown in
(3.49); they are almost equal in amplitude and have almost opposite phases (there is a
146o phase difference between them):
AN 1, 1 5 dBm,
S
X 2, 1;2, 1 0:272=9 deg, 3:49
T
X 2, 1;2, 1 0:214=155 deg:

The small-signal response predicted by the model in (3.45) at 5 dBm input drive level is
shown in Figure 3.27. At this drive level, the amplier has a strongly nonlinear
behavior. The response looks attened due to the almost equal and almost out-of-
S T
phase values of the X 2, 1;2, 1 and X 2, 1;2, 1 parameters.
In all cases, the X-parameter model accurately predicts the behavior of the amplier,
which proves once again the strength of this technology.
78 Spectral linearization approximation

Figure 3.27 Small-signal response at 5 dBm input drive level predicted by X-parameter model
(line) versus amplier response (symbols).

3.7.2 Input matching

In a very similar manner, it can be proven that the input match of a nonlinear system is
also dependent on the phase of the incident waves. While this was generally accepted
for the frequency at which the large signal was applied, it can now be proven for all the
harmonic frequencies.
Figure 3.28 shows the input waves on the second harmonic at the input of an
amplier under a single-tone large-signal drive.
A small-signal incident wave is applied, a12, and the response on the second
harmonic, B12, is monitored.
The response describes an ellipse when the phase of the small-signal stimulus a1,2
varies. The X-parameter model accurately predicts this response, as shown in Figure 3.28.
Separating the small-signal component of the response, b2,1, allows for an accurate
calculation for the input reection coefcient on the second harmonic frequency.
As shown in Figure 3.28, the reection coefcient is not constant with respect to the
phase of the incident wave, which leads to the conclusion that the input reection
coefcient is not an intrinsic property of the nonlinear system.
It would again be interesting to observe how different the result would be if the X (T)
parameters were not considered in the model. The results are shown in Figure 3.29.
This is a condition similar to the extension of S-parameters measured under hot
conditions that was attempted with limited success before the X-parameter model was
introduced. As shown, the small-signal response, b1,2, describes a circle as a function of
the phase of a1,2, and the input reection coefcient becomes a single point, thus
incorrectly suggesting it would be an intrinsic property of the amplier.
Based on this discussion, it is clearly concluded that the input reection coefcient is
not an intrinsic property of nonlinear systems. Source mismatch is one of the external
conditions that might inuence its value.
 3.7 Amplifier matching under large-signal stimulus 79

Small-signal stimulus on second harmonic input

+ A1 A2 +
Small-signal a1,2

- -
0.014

0.012

0.010

0.008

0.006

0.004

0.002

0.000

0.002

0.004

0.006

0.008

0.010

0.012

0.014
B1 B2

Input response on second harmonic Input small-signal response on second harmonic

Small-signal b1,2
B1,2

0.0005

0.0004

0.0003

0.0002

0.0001

0.0000

0.0001

0.0002

0.0003

0.0004

0.0005
0.04 0.03 0.02 0.01 0.00 0.01 0.02 0.03 0.04

Input reflection coefficient on second harmonic

1.0
0.9

1.2
0.8

1.4
0.7

1.6
0.6

1.8
2.0
0.5
Input Reflection Coefficient

0.4
3.0
0.3
4.0

5.0
0.2

10
0.1
20
0.1

0.2

0.3

0.4

0.5

0.6

0.7
0.8
0.9
1.0

1.2

1.4
1.6
1.8
2.0

3.0

5.0
4.0

10

20

20

0.1
10

0.2
5.0
.0
4
.3
0
.0
3
.4
0
.5

.0
0

2
.8
0.6

1
1.6
0.7

1.4
0.8

1.2
0.9

1.0

Figure 3.28 Input reection coefcient on second harmonic frequency of an amplier under
large-signal drive, predicted by X-parameters (line) versus system response (symbols).
80 Spectral linearization approximation

Input small-signal response on second harmonic Input reflection coefficient on second harmonic

1.0
0.9

1.2
0.8

1.4
0.7

1.6
0.6

1.8

0
0.5

2.
Input Reflection Coefficient
4
0.

3.0
0.3
4.0

5.0
Small-signal b1,2

0.2

10
0.1

20

0.1

0.2

0.3

0.4

0.5

0.6

0.7
0.8
0.9
1.0

1.2

1.4
1.6
1.8
2.0

3.0

4.0
5.0

10

20
0.0003 0.0002 0.0001 0.0000 0.0001 0.0002 0.0003
20

0.1
10

0.2 5.0

.0
4
.3
0
.0
3
.4
0

.5

.0
0

2
.8
0.6

1
1.6
0.7

1.4
0.8

1.2
0.9

1.0
System response

X-parameters without X(T) terms

Figure 3.29 Left: response on second harmonic to small-signal perturbations around the LSOP
predicted without X (T) parameters (line) versus system response (symbols). Right: input reection
coefcient at second harmonic predicted without X (T) parameters (single point) versus system
response (symbols) limitations of hot-S-parameters.

All the examples presented so far show that the X-parameter model does represent the
intrinsic properties of the nonlinear system and that it accurately reproduces its behavior
under various external conditions.

3.8 Practical application a GSM amplifier

The ability of X-parameters to describe correctly the output match of devices under
nonlinear operating conditions is valuable because there are many practical examples of
devices that are not adequately described by other methods such as hot-S22. One
example of such a device is a GSM handset amplier. Because GSM uses a constant
envelope, ampliers are typically run in saturated mode in order to achieve high
efciency. This results in the amplier always operating in a nonlinear state, regardless
of output power level. As a handset amplier, the output is connected to an antenna, and
the load seen by the amplier varies depending on the environment in which the phone
is operated. Because of this, the characterization of the ampliers behavior under a
wide range of match conditions is very important. Due to the highly nonlinear operating
condition of the amplier, X-parameters are needed in order to describe this behavior
accurately.
This section describes the use of X-parameters to characterize the Skyworks
SKY77329 Power Amplier Module, a commercially available module used in mass-
produced phones such as the Sony Ericsson W810. It is a multi-chip module which
 3.8 Practical application a GSM amplifier 81

VBATT Module
HBT
DCS/PCS_IN Match PA Match DCS/PCS_OUT

BAND SEL CMOS

TX EN collector
voltage
amplitude
VAPC controller

GSM_IN Match PA Match GSM_OUT

GND

Figure 3.30 Functional block diagram of the SKY77329 Power Amplier Module. 2005
Skyworks Solutions Inc. Reproduced, with permission, from [2].

Figure 3.31 Measurement system used to characterize the SKY77329.

contains two RF ampliers as well as analog and digital control circuitry. Additional
details about the module and its characterization can be found in [1]. Figure 3.30 shows
the block diagram as well as the available control lines.
Since GSM handsets operate under pulsed conditions, the device should be charac-
terized under pulsed conditions as well. A nonlinear vector network analyzer (NVNA)
system was congured with pulse generators and triggering to provide a 1/8 duty cycle
pulse with a period of 4.615 s to provide GSM-like operating conditions. Figure 3.31
shows the measurement-system conguration. This system was used to measure the
X-parameters of the module, as well as the behavior of the module under different load
conditions.
Measurements of DC voltage and current, taken during the pulse along with the RF
measurements, were included in both the X-parameters and the validation measure-
ments. This is required in order to characterize and validate power-added efciency, an
important metric for handset power ampliers. Separate measurements were taken for
the high-band and low-band ampliers, including sweeping the battery voltage on pin
82 Spectral linearization approximation

Figure 3.32 Complete X-parameter-based representation of the module in the circuit simulator.

Output Voltage
15
10
5
Volts

0
5
10
15
TX_Enable
4
2
Volts

0
2
VAPC
2.0
1.5
Volts

1.0
0.5
0.0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Time (ms)

Figure 3.33 Accurate reproduction of the module output behavior in response to digital and analog
control and modulation.

VBATT and the amplitude control voltage on pin VAPC to characterize the device across
its entire range of power levels. The two models were then combined in the circuit
simulator to create a single component representing the module, as shown in Figure 3.32.
By adding simple logic on the digital control pins TX_EN and BAND_SEL, the
component is able to reproduce accurately the behavior of the module, even under
relatively complex stimulus conditions such as those shown in Figure 3.33.
Because the X-parameters include measurements of the device behavior under
matched conditions, simulation results correlate extremely well with validation
 3.8 Practical application a GSM amplifier 83

40

Output power (dBm)

30

20

Data Sheet
10
Simulation

0
0 0.5 1 1.5 2
10

20
VAPC

Figure 3.34 Comparison of simulated and specied output power versus VAPC.

Figure 3.35 Comparison of measured, X-parameter-predicted, and hot-S22-predicted response to a

phase-swept stimulus tone at the output port.

measurements. In order to verify the measurements, simulation results were also

compared to the datasheet. Figure 3.34 shows a comparison of simulated and specied
output power versus voltage on the VAPC amplitude control pin.
In order to validate that the X-parameters accurately capture the device dependence
on load match, independent validation measurements were made with an additional tone
applied to the output. The phase of this tone was swept while the magnitude was
kept constant, resulting in a set of stimulus conditions similar to that described in
Figure 3.18. The response of the device was measured and compared to the response
predicted by the measured X-parameters as well as the response predicted by measured
hot-S22. Figure 3.35 shows the results of this comparison.
84 Spectral linearization approximation

It is immediately apparent from the elliptical shape of the measured response that hot-
S22 is not capable of predicting this behavior. The measured X-parameters, however, are
in excellent agreement with the measurements and provide an accurate characterization
of the device behavior. The resulting X-parameter-based model is therefore able to
predict accurately the behavior of the module under a wide range of environmental and
stimulus conditions, making it a valuable tool for tasks such as evaluating the module or
designing it into a system.

3.9 Summary

This chapter presented a useful and systematic simplifying approximation to the multi-
variate time-invariant spectral maps dened on the harmonic grid that were introduced
in Chapter 2. The complicated multi-variate nonlinear maps were approximated by
much simpler nonlinear maps dened on a small number of selected large tones only,
and simple linear maps accounting for the contributions of the many tones with small
amplitudes. This approach results in a dramatic reduction of complexity while providing
an excellent description of many important practical cases. One-tone X-parameters are
derived by linearizing the nonlinear spectral map around the periodic large-signal
operating point established by the DUT response to a large single-tone excitation, and
applying the principle of time invariance to the nonlinear mapping. Examples of
cascading two nonlinear ampliers demonstrate the accuracy of the approximations
and the advantages over methods based on ad hoc generalization of S-parameters. The
new X (T) conjugate terms that arise in linearized X-parameter theory are shown to
contribute signicantly to output matching under strong input drive, even for small
mismatch conditions.

Exercise

3.1 The ellipse described by the sum of X (S) and X (T) parameters. Consider a non-
linear multi-port under the stimulus of a single large-signal tone of frequency 0,
applied at port 1, A1,1. Prove that the tip of the vector Bp,k representing the
response at port p on the frequency k k0, describes an ellipse when a
small-signal perturbation is applied at port q on frequency l with constant
amplitude and variable phase.
The solution is presented in Appendix E, but the main steps (1)(5) are
summarized here.
(1) The response at port p on the harmonic k is described by the X-parameter model, as
shown in (3.50):
FB S
Bp, k X p:k refLSOPS Pk X p, k;q, l refLSOPS Aq, l Pkl
T
X p, k;q, l refLSOPS Aq, l Pkl : 3:50
 Exercise 85

B2,1

B2,1
0.01 0.01 0.01 0.01

Figure 3.36 Small-signal terms almost aligned on the major axis (center is on the LSOP response).

There are three terms in the response Bp,k: one large-signal and two small-signal
terms.
(2) There is a specic phase of Aq,l for which the two small-signal terms in (3.50) are
aligned, resulting in a maximum displacement from the large-signal term. Prove
that the direction on which they are aligned has the phase shown in (3.51):
   
S T
phase X p, k;q, l phase X p, k;q, l
phaseMajorAxis k  phaseA1, 1 : 3:51
2  
 S 
The maximum distance from the large-signal term is X p, k;q, l 
 
 
 T   
X p, k;q, l   Aq, l , the sum of the magnitudes of the two small-signal terms. This
makes for a maximum distance on the trajectory as shown in (3.52), which is the
major axis of the ellipse, as shown in Figure 3.36:
   

 S   T 
MajorAxis 2 X p, k;q, l  X p, k;q, l   jAq, l j: 3:52

(3) There is a specic phase of Aq,l for which the two small-signal terms in (3.50) are
counter-aligned, resulting in a minimum displacement from the large-signal term.
Prove that the direction on which they are aligned has the phase shown in (3.53):
   
S T
phase X p,k;q,l phase X p,k;q,l
phaseMinorAxis k phaseA1,1 : 3:53
2 2
The phases in (3.51) and (3.53) identify two perpendicular axes.  
 S 
The minimum distance from the large-signal term is X p, k;q, l  
 
 
 T   
X p, k;q, l   Aq, l , the difference of the magnitudes of the two small-signal terms.
86 Spectral linearization approximation

B2,1

B2,1
0.01 0.01 0.01 0.01

Figure 3.37 Small-signal terms almost counter-aligned on the minor axis (center is on the LSOP
response).

This makes for a minimum distance on the trajectory as shown in (3.54), which is
the minor axis of the ellipse, as shown in Figure 3.37:
   

 S   T 
MinorAxis 2 X p, k;q, l   X p, k;q, l   jAq, l j: 3:54

(4) Find the two points located on the major axis for which the sum of the distances
from one of the extremes of the minor axis to them is equal to the sum of
the distances from one extreme of the major axis to them and equal to
   
 
 S   T   
2 X p, k;q, l  X p, k;q, l   Aq, l . Prove that these two points are symmetrically
placed with respect to the intersection of the two axes at distances of
 r
   
   S   T 
2Aq, l  X p, k;q, l   X p, k;q, l . These are the foci of the ellipse.
(5) Prove that the distances from the tip of the total-response vector, Bp,k, to the two
foci are determined by the expressions in (3.55):
     

   S   T 
d 1 Aq,l  X p,k;q, l  X p,k;q,l 
 r
   
  

   S   T  T
2Aq,l  X p, k;q,l  X p, k;q,l  cos phase Aq,l phase X p,k;q, l ,
     
3:55
   S   T 
d 2 Aq,l  X p,k;q, l  X p,k;q,l 
 r
   
  

   S   T  T
2Aq,l  X p, k;q,l  X p, k;q,l  cos phase Aq,l phase X p,k;q,l :

The sum of these two distances is thus constant with respect to the phase of the
small-signal perturbation, (a), and has the value
     

   S   T 
d d 1 d 2 2Aq, l  X p, k;q, l  X p, k;q, l  : 3:56
 Additional reading 87

This proves that the tip of Bp,k describes an ellipse when the phase of Aq,l varies.
The characteristics of this ellipse have been determined during the exercise.

References

[1] J. Horn, J. Verspecht, D. Gunyan, L. Betts, D. E. Root, and J. Eriksson, X-parameter

measurement and simulation of a GSM handset amplier, in EuMIC 2008, Amsterdam,
Oct. 2008.
[2] Skyworks, SKY77329 PA module for quad-band GSM / EDGE, SKY77329 Datasheet,
Oct. 2005.

Additional reading

D. E. Root, J. Horn, T. Nielsen, et al., X-parameters: the emerging paradigm for interoperable
characterization, modeling, and design of nonlinear microwave and RF components and
systems, in IEEE Wamicon2011 Tutorial, Clearwater, FL, Apr. 2011.
D. E. Root, J. Verspecht, D. Sharrit, J. Wood, and A. Cognata, Broad-band, poly-harmonic
distortion (PHD) behavioral models from fast automated simulations and large-signal vectorial
network measurements, in IEEE Trans. Microw. Theory Tech., vol. 53, no. 11, pp. 36563664,
Nov. 2005.
J. Verspecht, Describing functions can better model hard nonlinearities in the frequency domain
than the Volterra theory, Ph.D. thesis annex, Vrije Univ. Brussel, Belgium, Nov. 1995;
available at http://www.janverspecht.com/skynet/Work/annex.pdf.
J. Verspecht and D. E. Root, Poly-harmonic distortion modeling, IEEE Microwave, vol. 7,
no. 3, pp. 4457, June 2006.
J. Verspecht, M. V. Bossche, and F. Verbeyst, Characterizing components under large signal
excitation: dening sensible large signal S-Parameters?!, in 49th ARFTG Conf. Dig., Denver,
CO, 1997, pp. 109117.
4 X-parameter measurement

One of the key features that led to the wide adoption of S-parameters was the availabil-
ity of hardware and calibration techniques capable of making quick, accurate, and
repeatable S-parameter measurements. S-parameters can also be easily extracted in
simulation from device or circuit models. In either case, the resulting S-parameters
can immediately be used in simulation or design tools. In order to achieve similar
success in the nonlinear domain, X-parameters must be easily measured and also easily
extracted from simulation.

4.1 Measurement hardware

X-parameters, like S-parameters, represent the steady-state behavior of a device in the

frequency domain. The linearity assumption of S-parameters, however, greatly simpli-
es the measurement requirements. The additional capabilities of X-parameters come at
the cost of additional complexity in the measurement system as well as the modeling
paradigm.

4.1.1 Hardware requirements

X-parameters include cross-frequency terms that capture the distortion products gener-
ated by device nonlinearities. In order to measure these terms, the measurement
hardware must be capable of measuring coherent cross-frequency phase. Because
S-parameters include no cross-frequency interaction (a consequence of the linearity
assumption), each frequency may be measured independently with no need for a
consistent time base or cross-frequency phase. Since all S-parameters are ratios of
waves, there is also no need for accurate measurement of absolute power only the
relative power is needed. As a result, the hardware and calibration techniques developed
for S-parameter measurement generally are not sufcient for X-parameter measurement
and must be extended to include cross-frequency phase and calibrated absolute power.

4.1.2 Mixer-based systems

The most common instrument architecture used to measure S-parameters is the mixer-
based vector network analyzer (VNA). The mixer-based VNA (Figure 4.1) owes its
popularity to its extremely high achievable dynamic range and modern calibration
 4.1 Measurement hardware 89

B1 A1 A2 B2
LO

DUT

Figure 4.1 Mixer-based VNA diagram.

techniques that enable highly accurate and repeatable measurements. The basic idea
behind the architecture is the use of mixers to convert high-frequency signals that are
difcult to measure to a lower intermediate frequency (IF) which is much easier to handle.
The measurement frequency is selected by adjusting the frequency of the local
oscillator (LO) driving the mixers so the IF remains constant. Since only one frequency
is measured at a time, narrow-band lters can be used to eliminate noise and spurs at
other frequencies and enable very high dynamic-range measurements. Calibration is
critical for accurate measurements, since the effect of the mixers must be removed in
addition to the cables and couplers.
Although this approach is extremely well suited to S-parameter measurements, it is
missing the phase coherence required to measure the cross-frequency terms present in
X-parameters. Because each frequency is measured individually, there is no consistent
time base or phase relationship between consecutive measurements. Even if each meas-
urement were accurately time stamped, the phase shift of the LO is generally unknown as
it shifts to a new frequency. This causes measurement phase coherence to be lost.
Figure 4.2 shows measurements taken of the fundamental, second, and third harmon-
ics at arbitrary times and the resulting measured phases of each signal. If these phase
measurements are used to reconstruct the waveform, it looks nothing like the actual
signal. If a phase-reference signal with known harmonic phase relationships (in this
case, a pulse train with known phase of 0 at all harmonics) is measured at the same time
as each measurement, however, the phase offset of each measurement can be deter-
mined and removed. The reconstructed waveform using corrected phases matches the
original waveform.
Equation (4.1) shows how this correction is accomplished:


corr
n meas
n  nref meas  nref char : 4:1
The phase correction factor, dened as the difference between the measured phase of the
reference at harmonic n, nref meas , and the known or pre-characterized phase of the
reference at harmonic n, nref char , is subtracted from the measured phase of the DUT at
harmonic n, meas
n , to obtain the corrected phase of the DUT at harmonic n, corrn .
In order to enable measurement of coherent cross-frequency phase in a mixer-based
system, an additional receiver can be added to measure a phase-reference signal that
90 X-parameter measurement

(a)
0.2
Actual Waveform

0.1

0.0

0.1

0.2
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
time (ns)
0.2
Fundamental

0.1

0.0 Measured Phase:

126.28 degrees
0.1

0.2
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
time (ns)
0.06
Second Harmonic

0.04
0.02
Measured Phase:
0.00
90.60 degrees
0.02
0.04
0.06
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
time (ns)
0.03
0.02
Third Harmonic

0.01
Measured Phase:
0.00
114.08 degrees
0.01
0.02
0.03
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
time (ns)
0.12
0.10
Phase Reference

Measured Phase:
0.08
Fundamental: 12.24 degrees
0.06
Second Harmonic: 74.88 degrees
0.04
0.02 Third Harmonic: 143.28 degrees
0.00
0.02
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
time (ns)

Figure 4.2 Phase coherence. (a) Measurements of fundamental, second, and third harmonics; also
shown are the original waveform and the phase-reference signal. (b) Waveform reconstructed from raw
phase measurements. (c) Waveform reconstructed from phase-corrected measurements; this matches
the actual waveform in (a). Corrected phase measured phase measured phase-reference phase.
 4.1 Measurement hardware 91

(b)
Reconstructed B2 - Phase Uncorrected

0.2

0.1 Raw Phases:

Fundamental: 126.28 degrees

0.0
Second Harmonic: 90.60 degrees
0.1
Third Harmonic: 144.08 degrees
0.2

0.3
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
time (ns)

(c)
Reconstructed B2 - Phase Corrected

0.2

Corrected Phases:
0.1
Fundamental: 138.52 degrees
0.0 Second Harmonic: 165.48 degrees
Third Harmonic: 29.20 degrees
0.1

0.2
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
time (ns)

Figure 4.2 (cont.)

has spectral content at all frequencies of interest with a consistent phase relationship. By
measuring the phase of the incident and scattered waves at each frequency relative to
this phase reference, phase coherence is restored. Although it is necessary to know the
cross-frequency phase response of the phase reference during calibration in order to
make accurate cross-frequency phase measurements, only consistency is required to
restore phase coherence in the measurement system. Calibration and phase-reference
selection will be discussed in detail in Sections 4.2 and 4.3.
This type of system, shown in Figure 4.3, is referred to as a nonlinear vector network
analyzer (NVNA) because it is capable of making the cross-frequency phase measure-
ments necessary to characterize nonlinear devices. A VNA with enough receivers can
be converted to an NVNA by adding a phase reference and the necessary software
to handle measurement and calibration. The NVNA architecture has many of the
characteristics of the VNA architecture from which it is derived, including the high-
dynamic-range capability and repeatability.

4.1.3 Sampler-based systems

Mixer-based systems measure signals by accurately capturing the magnitude and phase
of all spectral components of a signal. Once the complete spectrum is known, the
inverse Fourier transform can be used to compute the waveforms in the time domain.
Conversely, if the time-domain waveforms are known, the Fourier transform can be
used to compute the complete spectrum including coherent cross-frequency phase.
92 X-parameter measurement

B1 A1 A2 B2 F
LO

Phase
DUT Reference

Figure 4.3 Mixer-based NVNA diagram.

Real Signal and Samples

3
6
8
1
4
5
2 7

DSP Sample Reordering

3
6
8
1
Equivalent Time
4
5
2
7

Figure 4.4 Equivalent-time samplers.

Therefore any measurement system that is capable of capturing time-domain waveforms

can be used for X-parameter measurements.
Sampler-based systems measure signals by sampling the time-domain waveforms,
either in real time or in equivalent time. Real-time samplers measure sequential
samples of a signal and, assuming the sampling rate is sufciently high, can recon-
struct the signal using some form of interpolation. Real-time samplers are capable of
measuring a repetitive signal within a single period of the signal, and can also be
used to measure non-repetitive signals or transient events. Equivalent-time samplers,
which can only measure repetitive signals, measure one or more samples per period
over the span of many periods of the signal, then reconstruct a single period of the
signal by mapping the actual time of each sample to the equivalent time within the
single period (see Figure 4.4).
In either case, the sampler-based system captures complete time-domain waveforms.
These waveforms can be transformed into the complete spectrum (magnitude and phase
 4.2 Calibration 93

Small-Signal
Source 1

Power
DUT
Combiner
Small-Signal
Source 2

Large-Signal
Source 1

Figure 4.5 Example X-parameter stimulus conguration.

of all components) that is required for X-parameter measurements. Calibration is

typically still required at high frequencies in order to account for the effects of the test
set and non-ideal frequency response of the sampler. Although a phase reference is not
required in a sampler-based system in order to make coherent cross-frequency phase
measurements, a well-characterized calibration phase reference may be required for
accuracy at high frequencies.

4.1.4 Stimulus requirements

Independent of the hardware architecture chosen to measure the stimulus and response
spectrum and/or waveforms, the measurement system must also contain hardware
capable of providing the appropriate stimulus for X-parameter measurements. This
includes both the large-signal stimulus appropriate for the device under test (DUT)
and an additional stimulus needed to inject the small-signal stimulus that is required for
extraction of the small-signal (X (S), X (T), X (Y), and X (Z)) terms.
The large-signal stimulus required depends on the DUT and the application for which
the X-parameters are being measured. It may be as simple as DC only for an oscillator,
or may include several DC and RF sources as well as source and load tuners for more
complex circuits. The small-signal stimulus may be provided by a single-tone conti-
nuous wave source with a frequency range covering all frequencies of interest for the
DUT. It must be able to be applied at any port simultaneously with the large-signal
stimulus. This can be achieved by switching a single source between ports and
combining with the large-signal stimulus, or by using multiple sources to cover the
different ports, as shown in Figure 4.5.

4.2 Calibration

Calibration, or the mathematical correction for systematic measurement errors, is

required for accurate measurements at high frequencies. Calibration procedures typic-
ally require a system-error model, a method for identifying error terms in that model,
and a method of removing the modeled system error from measurements.
94 X-parameter measurement

Figure 4.6 Scalar-loss correction.

4.2.1 Scalar-loss correction

A simple example of a calibration method commonly used for large-signal measure-
ments is scalar-loss correction, in which the path between the DUT and the measure-
ment hardware is modeled as a simple scalar loss. Figure 4.6 shows an example
scalar-correction error model as well as a method used to identify the error terms and
apply the correction to a measurement.
This correction only affects the magnitude of the measurement, and neglects errors
due to effects other than loss (such as mismatch). More complex error models can be
used to improve calibration quality, but more complex measurements are typically
required to identify the error terms in these models.

4.2.2 S-parameter calibration

Two key reasons for the success of S-parameters are that they represent the intrinsic
DUT characteristics (independent of the measurement system used) and they can be
measured relatively easily with a high degree of accuracy. This is largely due to the
calibration techniques that have been developed to ensure accurate removal of
system error, as well as the high level of automation of the calibration procedure.
A calibration wizard that guides users through the measurements needed to identify
error terms and automatic measurement calibration is a standard feature in all modern
network analyzers.
Figure 4.7 shows a network analyzer error model with 12 error terms. Because only
ratioed waves are needed for S-parameter measurements, the absolute power and phase
 4.2 Calibration 95

Figure 4.7 S-parameter error-correction model without normalization.

Figure 4.8 Ten-term S-parameter error-correction model.

of the waves do not need to be measured. One term can therefore be normalized to unity
and some terms do not need to be isolated, reducing the number of unknown terms to
ten and thus simplifying the identication process. This ten-term error model, shown in
Figure 4.8, is commonly used for two-port S-parameter calibration in modern network
analyzers.
Several methods exist for identifying the terms of the error models by measuring
known standards such as open circuits, shorts, loads, and transmission lines. This error
model is capable of accounting for test set losses, phase shifts, imperfect directivity, and
mismatches as well as many systematic errors caused by the measurement hardware
itself (mixers, ampliers, lters, etc.). It is typically applied in such a way that the
calibration and S-parameter extraction from the measurements are combined into one
step. In this case, the imperfect load match seen by the DUT is also accounted for. The
application of the error model to a measurement is shown in (4.2):
96 X-parameter measurement

0 10 1 0 1
S 011  e00 S 0
 e 00
S 0 0
S
@ A@1 22
1f 2r 11 A
e2r  e11 @ 21 12 A
2f
e10 01
1f e1f e10
2r e2r
01
e01 01
2f e1r
S 111 0 10 1 0 1,
S 011  e00 S e
0 00 0 0
@1 1f 11 A@
e1f 1 2210 012r e11 A  e11 e11 @S 21 S 12 A
2r 1r 2f
e10 e
1f 1f
01 e e
2r 2r e01 01
2f e1r
0 10 1
00  
S 0
@ 21 A@1 22 S 0
 e 2r
e2r  e2f A
11 11
e01
2f e10 01
2r e2r
S 121 0 10 1 0 1,
S 011  e00 S 0
 e 00
S 0 0
S
@1 1f 11 A@
e1f 1 2210 012r e11 A  e11 e11 @ 21 12 A
2r 1r 2f
e10 e
1f 1f
01
e 2r e 2r e01 01
2f e1r
0 10 1 4:2
00  
0 S 0
 e
@S 12 A@1 11 1f 11 A
1f  e1r
e11
e01
1r e 10 01
e
1f 1f
S 112 0 10 1 0 1,
S 011  e00 S 0
 e 00
S 0 0
S
@1 1f
e11 A@1 22 2r 11 A
e2r  e11 11 @ 21 12 A
1f 1r e2f
e10 01
1f e1f e10
2r e2r
01
e01 01
2f e1r
0 10 1 0 1
S 0
 e 00 S 0
 e 00
S 0 0
S
@ 22 2r A@ 11
1 10 01 e11
1f A  e11 @ 21 12 A
1f 1r
e10 e
2r 2r
01
e e
1f 1f e01 01
2f e1r
S 122 0 10 1 0 1:
S 011  e00 S 0
 e 00
S 0 0
S
@1 1f
e11 A@1 22 2r 11 A
e2r  e11 11 @ 21 12 A
1f 1r e2f
e10 01
1f e1f e10
2r e2r
01 e01 01
2f e1r

Note that this combined correction/extraction relies on the assumption that the DUT
behaves linearly, and is not valid for measurement of devices behaving nonlinearly.

4.2.3 NVNA calibration

A nonlinear vector network analyzer obviously does not assume that the DUT is
behaving linearly, so some modication of the calibration procedures used for
S-parameters is clearly required. In addition to the linearity issue, an NVNA should
be able to measure absolute power and cross-frequency phase, not just ratioed waves, so
all terms must be individually identied, and normalization to unity is not allowed for
any term. Furthermore, many nonlinear measurements (including X-parameters) include
injection of signals at both ports simultaneously, so separate models for forward and
reverse sweeps that assume constant load match are no longer sufcient.
To address these issues, a single eight-term error model that can be applied to forward,
reverse, and combined stimulus congurations can be used (see Figure 4.9). The extraction
of terms must also be modied to include absolute magnitude and the appropriate cross-
frequency phase corrections, as well as isolation of all terms. This can be achieved by
measuring an absolute power reference (such as a power meter) and a well-characterized
phase reference in addition to the standards required for S-parameter calibration.
 4.3 Phase references 97

Figure 4.9 Eight-term error-correction model.

The resulting error model can be used to correct wave measurements, as shown in (4.3):
2 10 01 3
e e  e00 e11 e11
" # 6 n n 01 n n n
7" 0#
a1n 6 en n 7 an
e01
6 6
7 : 4:3
b1n 4 e00 n 1 7 0
5 bn
e01
n e01
n

Both the stimulus and the response at both ports are measured. Corrections for any
imperfections in the stimulus, such as source harmonics or imperfect load match, are not
made at this time. Such corrections, which essentially predict what the device will do
under a condition other than that under which it was measured, involve modeling
assumptions about the device and are left to the model extraction. For X-parameters,
this will be discussed in Section 4.4.

4.3 Phase references

The procedure for NVNA calibration includes measuring the same characterized pas-
sive standards used for S-parameter calibration, measuring an absolute power standard
such as a power sensor, and measuring a phase standard. Unlike S-parameter calibration
and power calibration, phase calibration is not commonly used in small- or large-signal
measurements outside of NVNA, and may be a new topic even for those with a great
deal of experience with high-frequency measurements. The key component of phase
calibration is a well-characterized phase reference with an output spectrum that
includes tones at all frequencies of interest with consistent, known phase relationships.

4.3.1 Phase-reference signals

From a theoretical standpoint, any well-dened signal with a spectrum that contains
energy at multiple frequencies can be used as a phase reference. Figure 4.10 shows the
amplitude spectrum of several well-known types of waveforms. The waveforms shown
in Figure 4.10 are realistic in that they have nite rise times of 10 ps and are therefore
fairly representative of what can be generated by currently available technologies.
For example, the spectrum of an ideal square wave has harmonic content at all odd
harmonics in addition to the fundamental frequency. The spectral components at all
frequencies are well dened, both in magnitude and phase. A square wave could
therefore be used as a phase-reference signal as long as it is very nearly ideal or any
98 X-parameter measurement

Waveform Spectrum
(a)
5 30
20
4
10

power (dBm)
voltage (V)

3 0
10
2 20
30
1
40
0 50
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 0 5 10 15 20 25 30 35 40 45 50
(b) time (ns) frequency (GHz)
5 30
20
4 power (dBm) 10
voltage (V)

3 0
10
2 20
30
1
40
0 50
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 0 5 10 15 20 25 30 35 40 45 50
time (ns) frequency (GHz)
(c)
5 30
20
4
10
power (dBm)
voltage (V)

3 0
10
2 20
30
1
40
0 50
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 0 5 10 15 20 25 30 35 40 45 50
(d) time (ns) frequency (GHz)
5 30
20
4
10
power (dBm)
voltage (V)

3 0
10
2 20
30
1
40
0 50
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 0 5 10 15 20 25 30 35 40 45 50
time (ns) frequency (GHz)

Figure 4.10 Spectra of various signals. (a) Triangle wave; (b) saw wave; (c) square wave;
(d) pulse train.
 4.3 Phase references 99

Measurement or Model or Characterization of

specification of golden device that supports
phase-reference specified stimulus
stimulus

Calibrate
receiver /
test port

Golden Receiver /
device test port

Phasereference
stimulus

Figure 4.11 Golden device as phase reference.

non-idealities (such as nite rise/fall times) have been well characterized. Since the
spectrum of the square wave contains only odd harmonics, however, it is difcult or
even impossible to choose a square wave that has energy at all frequencies of interest for
many types of measurements. Triangle waves share this limitation. Saw waves and
repetitive pulse trains contain energy at all harmonics, making them better candidates
for practical application. Pulse trains have the advantage that an ideal pulse train has
harmonic content that does not roll off with frequency. Although real pulse trains do roll
off due to nite rise times and non-ideal pulse shapes, they typically perform better at
high harmonic orders than available alternatives.
Although the signals discussed so far are dened in terms of their time-domain
waveforms and have spectra that can be analytically calculated, other types of
signals can be used as well. For example, a well-characterized golden nonlinear
component, such as a saturated amplier, can be used to generate a reference signal
(see Figure 4.11). As long as the output of the component is known, including the phase
of harmonics and/or intermodulation products, the resulting signal can be used as a
phase reference.

4.3.2 Measurement considerations

Although many signals can theoretically be used as phase references, there are practical
considerations that must be taken into account for real measurements. In order to make a
good phase reference, the signal must be able to be consistently and repeatably
generated by a real source or component, consistently and repeatably measured by a
real measurement system, and well characterized at the calibration connector.
Real signal generators do not generate ideal signals, so non-idealities of the source
(such as distortion and phase noise or jitter) and their impact on the quality of the
reference signal must also be considered. Distortion that is repeatable and known or can
be characterized has no impact on the quality of the phase reference, but any variation in
distortion will translate to uncertainty in the phase reference. Phase noise and jitter also
introduce uncertainty, and there may be a large multiplicative factor for high-frequency
100 X-parameter measurement

measurements with a low-frequency reference signal. For example, 0.36 degrees

of phase noise at 10 MHz or 100 ps of jitter translate to 360 degrees of phase noise
at 10 GHz.
The phase-reference signal must be known in order to be used for NVNA calibration,
so characterization of the real signal generator is important regardless of the signal
chosen. The complexity of the characterization required, however, may depend on both
the signal chosen and the technology used to produce the signal. A saw wave, for
example, may have different distortion that must be characterized at different repetition
frequencies. The signal generator may also depend on other factors such as ambient
temperature or, if applicable, input power. The more factors the signal generator is
sensitive to, the more complex characterization is needed to enable its use as an accurate
phase reference.
Real measurement systems have limited dynamic range and a noise oor. Signals that
have too little energy at frequencies of interest will not be measurable on a real
measurement system, and therefore are not appropriate phase-reference signals. The
spectral power of saw, square, and triangle waves decreases as harmonic index
increases, so these signals may be difcult to use when many harmonics are required.
Pulse trains have constant spectral power versus frequency, but there may be tradeoffs
between pulse width and minimum/maximum frequency of operation. The pulse
height is limited by the maximum input of the measurement system, so pulse width
determines maximum power per pulse. Wider pulses give more power at low
pulse repetition frequencies, but limit the maximum frequency since the signal must
settle between pulses.

4.3.3 Practical phase references

Repetitive-pulse generators are the most commonly used phase references in non-
linear measurement systems for a number of reasons. Several technologies exist that
are capable of generating high-quality, repeatable pulse trains that can be used at a
variety of pulse-repetition frequencies and contain sufcient energy at high frequency
to enable accurate measurements. A pulse train is also well characterized if the shape
of a single pulse is known, regardless of the pulse-repetition frequency. For technolo-
gies that are insensitive to other factors such as temperature and input power,
a relatively simple characterization is therefore sufcient for a wide range of uses
(see Figure 4.12).
Well-characterized golden devices are not commonly used in commercial nonlinear
measurement systems because accurate characterization of such devices over wide
ranges of stimulus conditions is difcult. There are several advantages to this approach,
however, that justify its use in some cases. The golden device can be stimulated at the
same frequencies as the DUT, guaranteeing that spectral energy is present in the
reference signal at all frequencies of interest. When using a repetitive-pulse generator,
the stimulus frequencies are restricted to be on a harmonic grid with some minimum
tone spacing. These restrictions are removed when using a golden device.
 4.4 Measurement techniques 101

Single-pulse
characterization
6

5
Compute
4
harmonic
voltage (V)

2
response at
1
stimulus
0
frequency
1 Calibrate
0 20 40 60 80 100
time (ps)
120 140 160 180 200
receiver /
test port

Phase Receiver /
reference test port

Repetition
frequency
Phase-reference
stimulus

Figure 4.12 Phase-reference characterization.

4.4 Measurement techniques

In order to measure X-parameters, a test plan that denes the stimulus conditions and
measurements from which the X-parameters will be extracted as well as an algorithm to
extract X-parameters from those measurements are required. A method for measuring
the large-signal terms (X (F), X (FI), and X (FV)) is presented in this section, along with two
alternative methods for measuring the small-signal terms (X (S), X(T), X(Y), and X(Z)).
The methods assume the measurements to be phase normalized to refLSOP, as discussed
in Chapter 3. All methods are initially described as they would operate on an ideal
measurement system, assuming that the desired stimulus is applied precisely with no
unwanted spectral content; practical measurement considerations that apply to all
methods are discussed at the end of this section.

4.4.1 Large-signal-response measurements

The large-signal response contained in the X(F) terms is conceptually the easiest to
measure and identify. The measurement consists of simply applying the desired large-
signal stimulus with no small-signal stimulus present and measuring the response at all
frequencies of interest. The X(F) term at any specied port and frequency corresponding
to the applied large-signal stimulus is equal to the measured response at that frequency.

4.4.2 Small-signal-response measurements

From a theoretical perspective, any set of small-signal stimuli that is sufciently rich
to stimulate all small-signal behavior of interest and sufciently varied that all terms
can be independently identied can be used to extract X-parameters. Practical
102 X-parameter measurement

20 20
10 10
a1 (dBm) 0 0

b2 (dBm)
10 10
20 20
30 30
40 40
50 50
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
frequency (GHz) a1 b2 frequency (GHz)

DUT

b1 a2
20 20
10 10
0 0
X 1F,k = b 1,k
b1 (dBm)

a2 (dBm)
10 10
20 20
30 F 30
X 2,k
= b 2,k
40 40
50 50
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
frequency (GHz) frequency (GHz)

Figure 4.13 Large-signal-response measurements and identication.

measurement considerations, however, have led to the adoption of the offset-

frequency and offset-phase measurement techniques for general use. Both of these
techniques are optimal in the sense that X-parameters can be extracted from the
theoretical minimum number of required measurements, and both have been demon-
strated to be capable of making accurate X-parameter measurements on real meas-
urement systems.

4.4.2.1 Offset-frequency measurements

The small-signal stimulus used in the offset-frequency method consists of, as the name
suggests, tones slightly offset in frequency from the large-signal stimulus and its
harmonics and/or mixing products. Since the small-signal stimulus is applied at the
same time as the large-signal stimulus, the DUT is operating in a nonlinear state and
mixing occurs. A single tone slightly offset from one large-signal response frequency
will generate output at upper and lower sidebands of all large-signal response frequen-
cies, as shown in Figure 4.14.
Because of this slight offset, the DUT response to the small-signal stimulus also
falls at frequencies offset from the large-signal stimulus, and can easily be separated
from the large-signal response by ltered tuning of the measurement frequency.
Furthermore, at frequencies other than DC, each measured upper sideband response
corresponds directly to a single X (S) term, and each lower sideband response corres-
ponds directly to a single X (T) term, so all terms can be independently identied as
simple ratios of (phase-normalized) scattered waves to incident waves, as shown in
Figure 4.15 and (4.4):
 4.4 Measurement techniques 103

20
20

10
10
20
0
0
10

b2 (dBm)
10

b2 (dBm)
10
0
20
20
a1 (dBm)

10
30
30
20
40
40
30
50
50
40 1.99998 1.99999 2.00000 2.00001 2.00002
0.99998 0.99999 1.00000 1.00001 1.00002
50 frequency (GHz)
frequency (GHz)
1.99998 1.99999 2.00000 2.00001 2.00002

frequency (GHz)
20

20 20 10
10 10
0
0 0

b2 (dBm)
a1 (dBm)

10 10 10

b2 (dBm)
20 20
30 30 20
40 40
30
50 50
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
40
frequency (GHz) frequency (GHz)
a1 b2 50
2.99998 2.99999 3.00000 3.00001 3.00002

20
DUT frequency (GHz)

10
20
b1 a2
20

0 10 10
0 0
b1 (dBm)

10
b (dBm)

10

a2 (dBm)
10

20
1

20
20
30 30
30 40 40
50 50
40 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
frequency (GHz) frequency (GHz)
50
0.99998 0.99999 1.00000 1.00001 1.00002
frequency (GHz)

20 20

10 10

0 0

10
b1 (dBm)

b1 (dBm)

10

20 20

30 30

40 40

50 50
1.99998 1.99999 2.00000 2.00001 2.00002 2.99998 2.99999 3.00000 3.00001 3.00002

frequency (GHz) frequency (GHz)

Figure 4.14 Offset-frequency stimulus and response.

Incident Wave Scattered Waves

bLS

bSSU
aSS bSSL

fmf fm fm+f fkf fk fk+f

Figure 4.15 Offset-frequency measurements and identication.

S bSSU
p, k
X p, k , q, m ,
aSS
q, m

bSSL 4:4
T p, k
X p, k , q, m   :
SS
aq , m

The offset-frequency measurement technique relies on the fact that the on-frequency
small-signal terms are equal to the limit of the offset-frequency terms as f goes to 0. As
long as the offset chosen is small enough, good agreement is guaranteed. What qualies
as small enough, however, is device dependent, so, without a-priori knowledge about
the device being tested, there is always some risk that too large an offset might be used.
104 X-parameter measurement

Figure 4.16 Offset-phase measurements and identication.

Very small offsets can also complicate the measurement system requirements and result
in longer measurements due to lter settling time requirements.

4.4.2.2 Offset-phase measurements

Unlike the offset-frequency method, the small-signal stimulus of the offset-phase
method is always applied at precisely the same frequencies at which the large-signal
stimulus and its harmonics and/or mixing products fall. Since the response to the small-
signal stimulus falls at the same set of frequencies as the response to the large-signal
stimulus, separation of the large- and small-signal response must be handled by the
extraction instead of ltering in the measurement system. Furthermore, the on-
frequency small-signal response may contain contributions from both the X (S) and
X (T), and these contributions must be identied separately in the extraction. This is
achieved in the offset-phase method by subtracting the large-signal stimulus and
response from the combined response to identify the small-signal part, and by making
measurements under two small-signal conditions where a phase offset is introduced into
the second small-signal stimulus, as shown in Figure 4.16.
The X (S) and X (T) terms can then be identied by solving (4.5):
 
S T
bSS1
p, k X a SS1
p, k , q, m q, m X p, k , q, m a q, m ,
SS1
  4:5
S T
bSS2
p, k X a SS2
p, k , q, m q, m X p, k , q, m a q, m :
SS2

Although this measurement technique has twice as many small-signal stimulus condi-
tions as the offset-frequency technique, it still has the same optimal number of
measurements since half as many measurements (on-frequency versus upper- and
lower-sideband measurements) are taken at each stimulus condition. This method is
described in detail in [1].
 4.4 Measurement techniques 105

4.4.3 Practical measurement considerations

Both methods discussed in the preceding sections would work awlessly on an ideal
measurement system capable of producing exactly the requested stimulus and making
precise, noise-free, spur-free measurements. Unfortunately, real measurement systems
are not ideal, so some consideration must be given to making the measurement
technique robust to the errors present in real measurement systems.
Non-idealities in the measurement system can come from two main sources: meas-
urement errors (such as noise and spurs) and stimulus errors (such as source harmonics
and mismatch). Good measurement technique, including both experiment design and
extraction, can minimize the sensitivity of the extracted parameters to measurement
errors and correct for stimulus errors to the greatest extent possible. Minimizing
sensitivity to measurement errors, while important, will not be discussed in great detail
here. It is a problem common to many measurements, and the techniques used to
address it for X-parameter measurement (taking multiple measurements and aver-
aging/tting, designing experiments for optimal conditioning, etc.) are common. Cor-
recting for imperfect stimulus is of particular interest, however, since it can be handled
in a unique way in X-parameter measurements.
Linear measurements are routinely corrected for imperfect stimulus through load-
match correction built into most standard S-parameter measurement routines. This
correction mathematically removes the effect of the non-ideal load match seen by the
DUT during the S-parameter measurement assuming that the DUT is behaving linearly.
As a result, true 50 (or other specied impedance) S-parameters can be obtained on a
system without an ideal 50 port match.
Most nonlinear measurements are not capable of carrying out stimulus correction
because not enough is known about the DUT. For example, in order to correct an
amplier third-harmonic measurement specied into a 50 impedance that is meas-
ured under actual impedances that vary between 45 and 55 at the fundamental,
second, and third harmonics, the measurement system must know how the third
harmonic responds to incident power at the fundamental, second, and third harmonics.
There is simply not enough information to determine this from a simple forward
measurement into a passive load.
X-parameters, on the other hand, contain exactly the needed information in the
S T S T S T
X 2, 3;2, 1 , X 2, 3;2, 1 , X 2, 3;2, 2 , X 2, 3;2, 2 , X 2, 3;2, 3 , and X 2, 3;2, 3 terms. Since these terms are
extracted from small-signal stimulus to which the DUT is responding in a spectrally
linear way, linear correction techniques (similar to S-parameter load match) can be used
to extract them in a way that fully corrects for imperfect stimulus. Once extracted, these
terms can be applied to the large-signal-stimulus measurements to extract X (F) terms
that are fully corrected for imperfect stimulus and represent the DUT behavior under the
specied conditions.
X-parameters enable, for the rst time, fully corrected measurements of nonlinear
gures of merit under specied conditions. This can either translate to better measure-
ments with the same quality of measurement system currently used, or similar measure-
ment quality with less expensive measurement equipment. It also results in signicantly
106 X-parameter measurement

Desired Measured Imperfections Ideal

Stimulus Response
Measured
Desired
Stimulus

Stimulus Stimulus

XS,XT

Measured
Imperfection
Response

Response Response

Figure 4.17 X-parameter correction for imperfect stimulus.

less system-to-system variation of measurements, since the system-dependent stimulus

imperfections have been corrected out to identify the intrinsic behavior of the DUT.

4.4.4 Simulation-based extraction

X-parameters can be extracted in simulation by using any of the measurement tech-
niques discussed in the preceding sections. Practical measurement considerations can
often be ignored in high-dynamic-range simulators, so the simple extraction techniques
described generally work very well on their own. In some simulators, however, there are
more efcient ways to extract X-parameters.
When simulating the DUT under the large-signal stimulus, harmonic-balance simula-
tors have access to the harmonic Jacobian in addition to the waveforms at the solution.
Since the small-signal terms of the X-parameters are the partial derivatives of scattered
waves at various ports and frequencies with respect to incident waves at various
ports and frequencies, there is a direct relationship between these terms and the harmonic
Jacobian which is described in [2]. By applying the correct boundary conditions, the
small-signal terms can be extracted directly from the solution of the large-signal-stimulus
simulation without the need to perform additional nonlinear circuit simulations with
small-signal stimulus present. This can result in signicant performance improvements.

4.5 X-parameter files

In order to ensure seamless interoperability between measurement hardware,

simulation-based extraction tools, and circuit simulators capable of using X-parameters,
a standard le format is required. The open .xnp le format serves this purpose,
providing the exibility needed to contain complex, multi-dimensional X-parameter
as well as the structure needed to enable automatic interpretation of the data by circuit
simulators or other processing tools.
 4.5 X-parameter files 107

Table 4.1 Independent variable naming conventions used in the .xnp file format

Independent-
variable name Meaning
fund_k kth fundamental frequency
VDC_p DC voltage applied to port p as part of LSOPS
IDC_p DC current applied to port p as part of LSOPS
AN_p_f magnitude of a large-reference-signal incident wave applied to port p at
frequency f as part of LSOPS, where f is represented by its harmonic or
mixing indices (AN_p_n for one fundamental frequency,
AN_p_n1_n2_. . . for multiple fundamental frequencies)
AM_p_f magnitude (AM) and phase in degrees (AP) of a large-signal incident
AP_p_f wave (which is not a reference signal) applied to port p at frequency f
as part of LSOPS
GM_p_f magnitude (GM) and phase in degrees (GP) of the reection coefcient of
GP_p_f the load at port p at frequency f as part of LSOPS
GX_p_f real (GX) and imaginary (GY) parts of the reection coefcient of the
GY_p_f load presented at port p at frequency f as part of LSOPS
ZM_p_f magnitude (ZM) and phase in degrees (ZP) of the load impedance
ZP_p_f presented at port p at frequency f as part of LSOPS
ZX_p_f real (ZX) and imaginary (ZY) parts of the load impedance presented at
ZY_p_f port p at frequency f as part of LSOPS

4.5.1 Structure
The .xnp le format is based on and strictly adheres to the generic MDIF le format,
documented in [3], which is capable of storing multi-dimensional data with arbitrary
numbers of independent and dependent variables. Three types of blocks are used in .xnp
les: the XParamAttributes block, the XParamPortData block, and the XParamData
block. An .xnp le contains one XParamAttributes block that includes the le format
version of the le, the number of ports of the device described by the le, and the
number of fundamental frequencies included in the LSOP. One XParamPortData block
is also included in the le, and this contains the reference impedance and name of each
port. These two blocks, which are informational only, are typically included at the
beginning of the le. Comments describing the measurement or simulation setup may
also be included. Finally, the X-parameters are contained in one or more XParamData
blocks in the le.

4.5.2 Naming conventions

Because of the exibility of the X-parameter framework, different les may contain
different types of data. For example, a measurement on an amplier may include
X-parameters at multiple DC-bias voltages with current output, while a measurement
on a mixer may have no bias at all but multiple fundamental frequencies. Naming
conventions for independent and dependent variables are used to identify what infor-
mation is present in the le. These conventions are described in Table 4.1 and Table 4.2.
108 X-parameter measurement

Table 4.2 Dependent variable naming conventions used in the .xnp file format

Dependent-variable
Name Meaning
FB_p_f X (FB) scattered wave at port p, frequency f due to LSOP
FI_p X (FI) DC current at port p due to LSOP
FV_p X (FV) DC voltage at port p due to LSOP
S_pOut_fOut_pIn_fIn X (S) and X (T) terms relating input wave at port pIn, frequency fIn to
T_pOut_fOut_pIn_fIn scattered wave at port pOut, frequency fOut
XY_pOut_pIn_fIn X (Y) term relating input wave at port pIn, frequency fIn to DC current
at port pOut
XZ_pOut_pIn_fIn X (Z) term relating input wave at port pIn, frequency fIn to DC voltage
at port pOut

! fund_1 = 1 GHz NumPts = 1

! VDC_3 = 10 NumPts=1
! GM_2_1 = [100e-03->800e-03] NumPts=8
! GP_2_1 = [100->170] NumPts=7
! AN_1_1 = [10e-06(-20.000000dBm)->100e-06(-10.000000dBm)] NumPts=6

Figure 4.18 Comment lines.

4.5.3 Example file

The current version of the X-parameter le format is version 2.0, as of this publication.
The older version 1.0 may still be found in the industry, from older measurements.
The major difference between version 2.0 and its predecessor version 1.0 is the type
of waves that are used for the X-parameter description: version 1.0 used voltage waves,
whereas version 2.0 uses generalized power waves.
In addition, only the X-parameter data blocks (identied as XParamPortData in
version 2.0) existed in version 1.0. The XParamAttributes block and the XParamPort-
Data block did not exist in version 1.0.
Explanatory comment lines were also not very common in the earlier version 1.0,
although they were supported. Beginning with version 2.0, comment lines are com-
monly used to clarify the measurement conditions under which the X-parameters have
been extracted.
An example of a version 2.0 X-parameter le of an amplier with one DC power
supply is presented here. The content of a typical X-parameter le is described in the
order in which the information is commonly listed in an actual le.
Comment lines, identied by an exclamation mark, may be used for conveying
information in a human readable format, and they are ignored by the simulator. An
example is shown in Figure 4.18.
 4.5 X-parameter files 109

BEGIN XParamAttributes
% Index(int) Version(real) NumPorts(int) NumFundFreqs(int)
0 2.0 3 1
END

Figure 4.19 XParamAttributes block.

BEGIN XParamPortData
% PortNumber (int) RefZ0(complex) PortName(string)
1 50 0 RFin
2 50 0 RFout
3 50 0 DCbias
END

Figure 4.20 XParamPortData block.

Comment lines may be included anywhere in the le (at the beginning, in the body, or
at the end of the le).
Based on the data in Figure 4.18, a user can determine the following:

 the model has one fundamental frequency (it was extracted with a single large tone
present in the stimulus) at 1 GHz;
 a 10 V bias was applied at port 3;
 a load-dependent model was extracted versus the load presented to port 2 of the DUT;
 the load at port 2 was specied as a reection coefcient;
 the magnitude of the reection coefcient was swept from 0.1 to 0.8 with eight
points;
 the phase of the reection coefcient was swept from 100o to 170o with seven
points;
 a tone was applied at port 1 at the fundamental frequency and it was used as the phase
reference;
 the power of the tone applied at port 1 was swept from 20 dBm to 10 dBm with
six points.
The XParamAttributes block species the le version, the number of ports of the DUT,
and the number of fundamental frequencies used in the X-parameter model. An example
is shown in Figure 4.19.
The XParamPortData block species the reference impedance and the port name for
each of the ports of the DUT. Complex reference impedances are supported, and they
are specied in units of measure of ohms, in realimaginary format. An example is
shown in Figure 4.20.
110 X-parameter measurement

VAR fund_1(real) = 1000000000

VAR VDC_3(real) = 10
VAR GM_2_1(real) = 0.2
VAR GP_2_1(real) = 30
BEGIN XParamDate
% AN_1_1(real) FI_3(real) FB_1_1(complex) FB_1_2(complex) FB_2_1(complex)
% FB_2_2(complex) S_1_1_1_1(complex) T_1_1_1_1(complex) S_1_2_1_1(complex) T_1_2_1_1(complex)
% S_2_1_1_1(complex) T_2_1_1_1(complex) S_2_2_1_1(complex) T_2_2_1_1(complex) XY_3_1_1(complex)
% S_1_1_1_2(complex) T_1_1_1_2(complex) S_1_2_1_2(complex) T_1_2_1_2(complex) S_2_1_1_2(complex)
% T_2_1_1_2(complex) S_2_2_1_2(complex) T_2_2_1_2(complex) XY_3_1_2(complex) S_1_1_2_1(complex)
% T_1_1_2_1(complex) S_1_2_2_1(complex) T_1_2_2_1(complex) S_2_1_2_1(complex) T_2_1_2_1(complex)
% S_2_2_2_1(complex) T_2_2_2_1(complex) XY_3_2_1(complex) S_1_1_2_2(complex) T_1_1_2_2(complex)
% S_1_2_2_2(complex) T_1_2_2_2(complex) S_2_1_2_2(complex) T_2_1_2_2(complex) S_2_2_2_2(complex)
% T_2_2_2_2(complex) XY_3_2_2(complex)
...
Numerical values
...
END

Figure 4.21 XParamData block.

The X-parameter data blocks follow next. The values of all independent variables are
listed for each block, as shown in the example in Figure 4.21.

4.6 Summary

Given the proper tools, X-parameters can be extracted from models in simulation or
from measurements taken on the device itself. A measurement system must be able to
provide the appropriate large-signal stimulus and small-signal stimulus, and must be
able to measure cross-frequency phase (or, equivalently, time-domain waveforms) in
order to support X-parameter measurements. At the time of publication, several such
systems are commercially available. X-parameter measurement can be highly auto-
mated and requires little user expertise beyond that required for S-parameter measure-
ment. This enables the X-parameters to provide for nonlinear devices a powerful
behavioral modeling paradigm with ease of use approaching that of the standard
paradigm for linear devices, S-parameters.

References

[1] D. B. Gunyan, D. E. Root, L. C. Betts, and J. M. Horn, Large signal scattering functions
from orthogonal phase measurements, U.S. Patent 7 671 605, Mar. 2, 2010.
[2] D. E. Root, D. D. Sharrit, and J. Wood, Behavioral model generation, U.S. Patent 8 170
854, May 1, 2012.
 Additional reading 111

[3] Agilent Technologies, Inc., Advanced Design System documentation (ADS2009U1 or

later) Working with data les X-parameter GMDIF format. Available at http://edocs.
soco.agilent.com/display/ads2009U1/Working+with+Data+Files#WorkingwithDataFiles-
XparameterGMDIFFormat.

Additional reading

Applying error correction to network analyzer measurements, Agilent AN 12873, Agilent

Technologies, Inc., Santa Clara, CA, 2002.
T. Van den Broeck and J. Verspecht, Calibrated vectorial nonlinear-network analyzers, in 1994
MTT-S IMS Dig., May 1994, vol. 2, pp. 10691072.
J. Verspecht, Method and a test setup for measuring large-signal S-parameters that include the
coefcients relating to the conjugate of the incident waves, U.S. Patent 7 038 468, May 2,
2006.
5 Multi-tone and multi-port cases

5.1 Introduction

In this chapter, the general X-parameter formalism is applied in two different cases of
great practical interest beyond the simplest case of nearly matched components under
large drive considered in Chapter 3. Specically, X-parameter expressions are
developed by spectral linearization around an LSOP dened by the DUT in response
to two large incident signals, as opposed to the case of just one large signal considered
in Chapter 3.
The rst case is an example of two signals whose frequencies, f1 and f2, are the same,
but are incident with distinct phases at different ports. This is an important subcase of a
commensurate frequency relationship, namely f1/f2 n/m for n and m integers. The
second case considers incommensurate frequencies, where the ratio of the two signal
frequencies is irrational. Whether or not the signals are commensurate affects the
dimension of the LSOP as well as the form of the time-invariant X-parameter equations,
as will be evidenced below.
The rst extension, to arbitrary load-dependent X-parameters, describes the case
where the DUT behaves fully nonlinearly with respect to both A1,1 and A2,1, the phasors
corresponding to waves at the same fundamental frequency, incident at two distinct
ports, port 1 and port 2, respectively. This is usually the condition of operation for bare
transistors and highly mismatched high-power ampliers under strong input drive. The
spectrum generated by a DUT subject to such excitations still lies on a harmonic grid,
but now the LSOP, and therefore the X-parameter nonlinear functions, are more
complicated than the case considered in Chapter 3. This extension includes most aspects
of continuous-wave (CW) load pull, historically used to perform scalar characterization
of the behavior of a nonlinear DUT as a function of the complex load impedance. It also
includes waveform measurements at DC, fundamental and harmonic frequencies,
sometimes referred to as time-domain load pull. These load-dependent X-parameters
can be measured using an NVNA with passive tuners, or active injection, or some
combination of tuners and active injection (hybrid load pull). X-parameters provide a
powerful measurement-based nonlinear design ow. They can be used immediately to
design nonlinear multi-stage ampliers, including high-efciency power ampliers such
as Doherty and other amplier classes that may even require harmonic tuning.
The second extension compared to Chapter 3 deals with two large signals at incom-
mensurate fundamental frequencies incident at the same port or at distinct ports of the
 5.2 Commensurate signals 113

DUT. Here too the LSOP must be dened in terms of two tones. The frequencies at
which the X-parameter functions must be dened do not lie on a harmonic grid in this
case, but rather on a more complicated, discrete but dense, frequency spectrum corres-
ponding to all intermodulation products of the fundamental tones. An extension of the
notation labeling the spectral components and the corresponding X-parameter functions
is introduced. The example chosen to illustrate this capability is that of three terminal
mixers, where the local oscillator (LO) and a radio-frequency (RF) signal are treated as
incommensurate large input signals at distinct ports and frequencies, producing inter-
modulation terms at the IF and the other ports. Powerful measurement-based nonlinear
three-port models of mixers, including intermodulation products with both magnitude
and phase, as nonlinear functions of bias, and power of each distinct tone, can be
handled with this approach. Mismatch, for example at the IF port of a mixer, can
be accounted for by spectral linearization around the multi-tone LSOP.
The two extensions of X-parameters presented in this chapter, when used together, enable
end-to-end measurement-based design approaches to nonlinear RF circuits and systems.

5.2 Commensurate signals large A1,1 and large A2,1: load-dependent

X-parameters

There are several cases of great practical importance where it becomes necessary to
relax the simplifying approximation made in Chapter 3 that the scattered waves, Bp,k,
depend nonlinearly only on the applied bias conditions and A1,1, the single large RF
component incident at port 1 at the fundamental frequency. An example is a bare
transistor presented with a very large output mismatch at the fundamental frequency.
In this case, the large B2,1 wave generated by the device in response to a large incident
A1,1 will reect from the mismatch and create a large incident wave at the output port,
also at the fundamental frequency, given by (5.1):
A2, 1 2, 1 B2, 1 : 5:1
In (5.1), 2,1 is the complex reection coefcient at port 2 at the fundamental frequency.
The nearly matched case, treated in Chapter 3, is sufcient for j2,1j  1, since, for
small enough reection coefcient, the magnitude of A2,1, from (5.1), will be small
enough such that the spectral linearization of the DUTs response with respect to A2,1
around A2,1 0 will be valid. Thus, under well-matched conditions, A2,1 can be small
enough, even for devices with large power gain, that its inuence on the LSOP of the
DUT can be considered negligible, and its contribution to the scattered waves, Bp,k, can
S
be modeled as approximately linear in A2,1 and A2,1 , with coefcient functions X p,k;2,1
T
and X p, k;2,1 , respectively.
On the other hand, large mismatch means that the condition j2,1j  1 is no longer
satised and values of j2,1j even greater than unity can occur with active loads. For
large A1,1 (strong input drive), a large output mismatch means A2,1 can be sufciently
large, from (5.1), that it changes the DUT LSOP, and hence the spectral linearization
used previously is not valid.
114 Multi-tone and multi-port cases

It is important to emphasize, however, that large mismatch (large j2,1j) by itself does
not necessarily mean that the spectral linearization with respect to A2,1 around A2,1 0
of Chapter 3 is not valid. For example, if the DUT is biased in an off state (e.g.
a transistor biased at pinchoff ), B2,1 can be very small, so, even for a large 2,1, A2,1
can still be small enough from (5.1) that linearization around A2,1 0 can still be valid.
If the DUTs scattered waves depend very weakly on A2,1, then even a large jA2,1j may
not signicantly affect the DUT LSOP; the formalism of Chapter 3 can still apply. It is
necessary to go beyond the scope of Chapter 3 only when both the mismatch is large
and the sensitivity of the DUT to mismatch is large simultaneously [1].
It is impossible to know a priori how large an jA2,1j is required to invalidate the
considerations of Chapter 3 for a given device. Of course, this is quite analogous to the
question of how large a source power level can be used to make an accurate set of
conventional S-parameter measurements of an active device before the assumption that
the device is operating linearly breaks down. Certainly, for unmatched power transistors
and high efciency power ampliers at moderate signal levels and above, the nonlinear
dependence on A2,1 of the DUT response must be taken into account.
One approach for large jA2,1j is to take terms of higher order than unity in A2,1 and
A2, 1 to approximate their nonlinear contributions to Bp,k [2]. Instead, in what follows,
the full nonlinear dependence on A2,1 will be treated in a manner consistent with that of
A1,1. That is, A2,1 becomes an argument (independent variable) of the X-parameter
functions. No polynomial basis is required nor is an expansion coefcient extraction
method needed. The accuracy of the resulting model depends only on measurement
accuracy and sampling density in the space of the independent variables.

5.2.1 Time invariance, phase normalization, and commensurate two-tone LSOP

The condition created by the large A1,1 and large A2,1 is shown in Figure 5.1. Two large
signals, A1,1 and A2,1, and the DC bias are considered as part of the LSOPS, as shown
in (5.2):


LSOPS DCS p , A1, 1 , A2, 1 : 5:2

The response to additional incident waves at harmonics of the fundamental frequency,

considered small, is treated by spectral linearization. The treatment of nonlinear spectral
maps of a time-invariant DUT on a frequency grid, presented in Chapter 2, is fully
applicable here. Since both the A1,1 and A2,1 signals are at the same frequency, they
always have a xed relative phase, 2,1, independent of time shift, given by (5.3), with
P e jA1, 1 :

e j2, 1  e jA2, 1 A1, 1  e jA2, 1 P1 : 5:3

The reference LSOP stimulus (refLSOPS), dened in the same way as in Chapter 2, is
therefore given by



refLSOPS DCS p , jA1, 1 j, jA2, 1 j, 2, 1 DCS p , jA1, 1 j, A2, 1 P1 : 5:4
 5.2 Commensurate signals 115

A1,k A2,k

B1,k B2,k

Figure 5.1 Two large tones (circled) at the same fundamental frequency and small tones at
harmonic frequencies incident at ports 1 and 2 of a nonlinear DUT. Scattered waves have
frequency components on the same frequency grid.

5.2.2 Spectral linearization

The general multi-tone nonlinear map from the harmonic grid into itself is now
spectrally linearized around the refLSOPS dened by (5.4). Here both A1,1 and A2,1
are large, dening the LSOP, and all the other signals at each port and harmonic, Ap,k,
are considered small. In addition to the large A2,1 tone dening the operating point, the
presence of an additional small A2,1 is considered. Its signicance will be apparent
later. For simplicity, in this discussion A1,1 is always taken to be zero. That is, no
perturbation is considered at the input port at the fundamental frequency.
LSOP LSOP
For clarication, the large-signal waves at the LSOP are A1, 1 and A2, 1 . The total
large-signal waves, including the small-signal perturbation around the LSOP, are then as
shown in (5.5), considering no small-signal perturbation is applied on A1,1 in this exercise:
LSOP LSOP
A1, 1 A1, 1 A1, 1 A1, 1 ,
LSOP 5:5
A2, 1 A2, 1 A2, 1 :
The same principles as in Chapters 2 and 3 are applied. The spectral mapping must be time-
invariant, which means that the nonlinear maps, Fp,k, can be evaluated at a time-shifted
argument, and then time shifted back by appropriate phase multiplication of the result. The
time shifts of the mappings arguments and resulting values are applied rst, and then the
nonlinear functions are spectrally linearized around the refLSOPS given by (5.4). The result
for the scattered waves is presented in (5.6), with some of the intermediate steps in the
derivation included, and the results for the DC-bias responses are shown in (5.7) or (5.8):
 
LSOP
Bp, k F p, k DCS q ,A1,1 ,A2, 1 A2,1 ,A1,2 ,A2, 2 , ...,AP,K
   
LSOP
F p,k DCS q ,A1,1 P1 , A2,1 A2,1 P1 ,A1, 2 P2 ,A2, 2 P2 , ...,AP, K PK Pk
 
F p, k  F p,k 
F p, k refLSOPSP
k  A2, 1 P 
k1  A2,1 Pk1
A2,1 P1  A 
P 
refLSOPS 2, 1 refLSOPS
2 3
 
X 6 F p,k  F p,k  
 
0 07
4  Ap0 ,k 0 Pkk   Ap0 ,k 0 Pkk 5
k 0
  0

p 1,::,P Ap0 , k 0 P Ap0 ,k 0 P
0 k
refLSOPS refLSOPS
k 0 2...,K
F
X p, k refLSOPS Pk
S T
X p,k;2,1 refLSOPS A2, 1 Pk1 X p,k;2,1 refLSOPS A2, 1 Pk1
X h S 0 T 0
i
X p, k;p0 ,k 0 refLSOPS Ap0 ,k 0 Pkk X p, k;p0 ,k 0 refLSOPS Ap0 ,k 0 Pkk ; 5:6
p0 1,::,P
k 0 2...,K
116 Multi-tone and multi-port cases

I p X I
p refLSOPS
n o
Y
Re X p;2,1 refLSOPS A2,1 P1
X n o 5:7
Y
Re X p;q, l refLSOPS Aq,l Pl ;
q1,...,P
l 2,...,K

V p X V
p refLSOPS
n o
Z
Re X p;2,1 refLSOPS A2,1 P1
X n o 5:8
Z
Re X p;q, l refLSOPS Aq,l Pl :
q1,...,P
l 2, ...,K

The terms proportional to A2,1 and A2, 1 are listed separately in (5.6)(5.8), for
emphasis, because they are related to one of the large-tone variables, A2,1, dening
the LSOP.
For clarity, we rewrite (5.6) in the following form:
F
Bp, k ALSOP
1, 1 ,A2, 1
LSOP
A2, 1 ,A1, 2 , ...,Ap, K X p, k ref LSOPSPk
   
S T
X p, k;2, 1 ref LSOPS A2, 1 ALSOP 2, 1 Pk1 X p, k;2, 1 ref LSOPS A2, 1 ALSOP 2, 1 Pk1
X h S 0 T 0
i
X p, k;p0 , k 0 ref LSOPSAp0 , k 0 Pkk X p, k;p0 , k 0 ref LSOPSAp0 , k 0 Pkk :
p0 1,::,P
k 0 2...,K

5:9

5.3 Establishing the LSOP using a load tuner: passive load pull

The LSOP in (5.9) need not be established by two independent signal sources. In the
common method of passive load pull, for example, a single signal source is used to
stimulate the device at the input port and a mechanical tuner is used at the output to
reect the DUT-produced B2,1 back into the output port as A2,1, using (5.1). The
physical experiment controls the complex reection coefcient, 2,1, so each of the
measurements is tabulated as a function of 2,1. The tuner is considered ideal for now,
meaning that only at the fundamental frequency is there a non-zero reection coef-
cient, 2,1. The constraint of constant 2,1, for all power values incident at port 1, makes
the reection coefcient the independent variable, whereas A2,1 becomes a dependent
variable.
To be specic, consider the case p 2, k 1 in (5.9). Non-zero values of Ap0 , k 0 , for
0
k > 1, such as the ones that may be generated by reections of DUT-generated
harmonics, for example, cause changes in B2,1 through the terms inside the summations
in (5.9). According to (5.1), this means that there must be a corresponding small change,
A2,1, in A2,1 around its value at the LSOP. The terms A2,1, and therefore also A2, 1 by
 5.3 Establishing the LSOP using a load tuner 117

virtue of the non-analytic nature of the spectral mappings, contribute additional changes
S T
to all the Bp,k values through the X-parameter functions X p, k;2, 1 and X p, k;2, 1 ,
respectively.
For passive load-pull-based X-parameter measurement, re-parameterizing (5.9) by
2,1 instead of A2,1 results in the equivalent expression (5.10) for the X-parameter model
of the DUT:
F
Bp, k X~ p, k jA1, 1 j,2, 1 Pk
S
 F

X~ p, k;2, 1 jA1, 1 j,2, 1 A2, 1 2, 1 X~ 2, 1 jA1, 1 j,2, 1 Pk1
T
 F

X~ p, k;2, 1 jA1, 1 j,2, 1 A2, 1 2, 1 X~ 2, 1 jA1, 1 j, 2, 1 Pk1
X h S T
i
X~ p, k;p0 , k 0 jA1, 1 j,2, 1 Ap0 , k 0 Pkk X~ p, k;p0 , k 0 jA1, 1 j,2, 1 Ap0 , k 0 Pkk :
0 0

p0 1,::,P
k 0 2...,K

5:10
The following constraint (5.11) is appended to the model equations (5.10):
A2, 1 2, 1 B2, 1 : 5:11
F
The functions X~ :, : are related to X (F)(.,.) by (5.12):
F
 
F
X~ p, k jA1, 1 j, 2, 1 X p, k jA1, 1 j, Aref
2, 1
LSOPS

F
5:12
for Aref
2, 1
LSOPS
2, 1 X~ 2, 1 jA1, 1 j, 2, 1 :
S T
The sensitivity functions X~ and X~ are related to X(S) and X(T) by (5.13):
S
 
S
X~ p, k;p0 , k 0 jA1, 1 j, 2, 1 X p, k;p0 , k 0 jA1, 1 j, Aref
2, 1
LSOPS
,
T
 
T
X~ p, k;p0 , k 0 jA1, 1 j, 2, 1 X p, k;p0 , k 0 jA1, 1 j, Aref
2, 1
LSOPS
5:13
F
for Aref
2, 1
LSOPS
2, 1 X~ 2, 1 jA1, 1 j, 2, 1 :
The ~ symbols indicate that the functional forms when 2,1 is the independent
variable are different from the functional forms when A2,1 is the independent variable.
This is common, if somewhat precise, mathematical notation for two functions related
to each other in that they have the same range but their domains are related by a
coordinate transformation.
In (5.10), use was made of the fact that, at the refLSOPS, all the Ap0 , k 0 0
F
and, through (5.11), that Aref
2, 1
LSOPS
2, 1 X 2, 1 . It is the values of the functions
F S T
X~p, k , X~p, k;p0 , k 0 , and X~p, k;p0 , k 0 that therefore appear in the measured .xnp le when a
DUTs X-parameters are extracted using passive load-pull acquisition with 2,1 the
control variable (independent variable) of the experiment. It is important to stress,
however, that the tuner impedance values do not need to be directly perturbed around
the specied, xed values of 2,1. For example, the tuner can be used just to establish
the LSOP, and perturbations can be applied using active source injection of small
signals. The sensitivity information can be extracted from the responses to these
118 Multi-tone and multi-port cases

perturbations around the LSOP. Regardless of how the measurement is done, the X (S)
and X (T) parameters represent sensitivities of the response to perturbations in the
A-waves around the LSOP established by the reection coefcient of the load.
In words, (5.12) and (5.13) mean that the numerical values of the DUTs X-parameter
functions in the data le measured under xed 2,1 conditions are the same as if the
X-parameter functions were measured without a tuner (e.g. by active source injection) at
F
a xed ALSOP
2, 1 value corresponding to 2, 1 X~ 2, 1 jA1, 1 j, 2, 1 .
Even though (5.10) and (5.11) describe the DUT parameterized by 2,1, the equations
represent the DUT properly in any circuit into which it is placed, whether or not the
DUT model is presented with a xed impedance. For example, consider the DUT model
characterized under controlled and swept 2,1 conditions, presented with applied inde-
pendent signals Aapp app
1, 1 and A2, 1 in a particular circuit. The simulator has to evaluate (5.10)
consistently with (5.11) for A1, 1 Aapp app
1, 1 and A2, 1 A2, 1 . The model looks up and
interpolates the values of the X-parameters as functions of 2,1 iterating on the value
of 2,1 until (5.10) is satised for a particular value of reection coefcient, conv 2, 1 , such
conv ~
that also Aapp 2, 1 2, 1 B 2 , 1 conv
2, 1 . The values of (5.10) under these conditions are
precisely the correct DUT scattered waves corresponding to incident A1, 1 Aapp 1, 1 and
A2, 1 Aapp
2, 1 phasors. The equations (5.10) and (5.11) are solved simultaneously with all
the other equations of the circuit into which the DUT model is placed. The result is the
correct solution of the circuit consistent with the correct inputoutput relationships of
the DUT.

5.4 Additional considerations for commensurate signals

5.4.1 Extraction of X-parameter functions under controlled loads

The constraint (5.1) makes the extraction of each X-parameter function using passive
load-pull measurements more complicated than the ideal experiment design outlined in
Chapter 3 in terms of independent incident A-waves. These details are beyond the scope
of this work.

5.4.2 Harmonic superposition

When the X-parameter model functions are measured using independent active source
injection for both large signals A1,1 and A2,1, then, on evaluating the model for
S T
A2, 1 Aapp
2, 1 A2, 1 , the X-parameter functions X p, k;2, 1 and X p, k;2, 1 do not contribute
LSOP

to the scattered waves in (5.9) since A2, 1  ALSOP

2, 1 0. The linear terms in (5.9)
correspond to the signal components at the harmonics of the fundamental. This is the
harmonic superposition principle [3]. It too can be relaxed, if necessary, by using more
independent spectral frequencies in the nonlinear mapping (less spectral linearization).
It will be shown in the following how effective this approximation is, even for DUTs
where independently designed harmonic terminations are important to the device
operation.
 5.4 Additional considerations, commensurate signals 119

5.4.3 Limitations of passive load pull for load-dependent X-parameters

Passive tuners by themselves can provide reection coefcients no larger than unity.
The active or hybrid load-pull setups are therefore necessary for some applications
where the operating conditions of the amplier change from an off-state to an on-state
depending on other signals in the circuit. This is the case for Doherty ampliers, for
example, where the auxiliary amplier is off in the low-input-power range and turns on
only at large input powers [4]. This will be treated in Section 5.6
Passive load pull requires the DUT to be generating signicant B2,k from the A1,1
drive signal, to be reected as A2,1 using (5.1). If the device (transistor) is off, such as
when biased at pinchoff, B2,k is essentially zero. For applications at or near a DUT off-
state, active source injection to produce a signicant A2,1 is therefore required. This is a
different paradigm compared to conventional load pull, where the DUT is usually
considered operating always in an on-state. However, active injection for X-parameter
characterization is completely analogous to linear S-parameter measurements where a
source is used to inject an A2,1. S-parameters are routinely measured when the device is
biased off as well as on. So an active-source-injection system for X-parameter measure-
ments is really just a natural extension of the linear case, needed to characterize fully the
nonlinear DUT over its entire range of operation.

5.4.4 Sampling of the three-RF-variable space defining the refLSOPS

There are three independent RF variables specifying the device refLSOPS dened by
(5.4). For xed powers of each incident RF signal, changing the relative phase angle,
2,1, changes the state of the DUT. It is generally required to vary the magnitudes of
each tone, jA1,1j and jA2,1j, independently over their full ranges, and also to sweep 2,1
over the complete range from 0 to 2 radians. This requires correspondingly more data
than for the case of the one-tone refLSOPS considered in Chapters 2 and 3. Sufciently
dense sampling of both the magnitudes and the phase angle is required for accuracy
over the complete range of DUT operation. The actual sample density required for a
xed accuracy depends on the particular nonlinearity of the DUTs responses to these
excitations, which cannot be known a priori.

5.4.5 Hardware setup for load-dependent X-parameters

A typical setup for measuring load-dependent X-parameters is shown in Figure 5.2. The
passive tuner at the input is used primarily to achieve a better power match to the device
from the RF source. It is not necessary, in general, if sufcient power is available from the
source without an impedance transformer. The passive tuner at the output port is used to
vary the load at the fundamental frequency over a range of values across the Smith chart.

5.4.6 Calibrating out uncontrolled harmonic impedances

Consider the tuners in the measurement setup of Figure 5.2. Real tuners are not ideal
in the sense of Section 5.3. Specically, as the fundamental load is varied from one
specied state to the next, the harmonic impedances presented by the tuner may vary in
120 Multi-tone and multi-port cases

GPIB

Bias
NVNA
Tees

Tuner
NVNA
DC Supply Control
Firmware
Software

USB
Passive Passive
DUT
Tuner Tuner

Figure 5.2 NVNA with load tuner for load-dependent X-parameter characterization.

uncontrolled ways. Provided both the DUT-generated harmonics and the harmonic
reection coefcients are not very large, these imperfections can be calibrated out by
using the X-parameter characterization and identication algorithms. The technique is
quite analogous to that of Chapter 3, whereby X-parameters calibrate out the source
harmonic imperfections and multi-frequency mismatch of the instrument. X-parameter
extraction can also correct for errors in the actual value of the fundamental load
impedance of the tuner. That is, it is possible to correct the measurements to compensate
for the fact that the actual measured impedance presented to the device may not be
precisely the specied impedance. This enables the corrected data to be tabulated on a
precise grid of specied impedances in the complex plane. This is important for the
subsequent interpolation of the measured X-parameters by the simulator when using the
model in a nonlinear design.

5.5 Arbitrary load-dependent X-parameters of a GaAs FET

A bare GaAs transistor, part number WJ FP2189, with typical output power of 1 W, was
characterized by load-dependent X-parameters over a wide range of incident power. The
loads were established by a mechanical tuner from Maury Microwave Corp. The
harmonic impedances presented by the tuner at the output port were measured and
recorded in a le for subsequent validation. Scattered waves were measured, including
ve harmonics and DC-bias values. Perturbation tones were used up to the third
harmonic to obtain harmonic load sensitivities.
Delivered-power contours, in dBm, are plotted in Figure 5.3 obtained from simulations
using the resulting X-parameter model and from independently measured load-pull data.
The measurements and X-parameter extraction are automated and straightforward.
More complicated is the nonlinear validation. To validate the model, the actual meas-
ured harmonic loads presented by the tuner need to be presented to the model of the
device in the simulator environment for an accurate comparison.
Characteristics, such as those shown in Figures 5.45.6, could only be simulated,
using compact transistor models, prior to the availability of X-parameters.
 5.5 Load-dependent X-parameters of a GaAs FET 121

Delivered-Power Contours (dBm)

1.0
0.9

1.2
0.8

1.4
0.7

1.6
0.6

1.8
2.0
0.5
4
0.
0
3.
0.3
4.0

5.0
0.2

0.1 10

20
0.1

0.2

0.3

0.4

0.5

0.6

0.7
0.8
0.9
1.0

1.2

1.4
1.6
1.8
2.0

3.0

4.0
5.0

10

20
20

0.1 10

0.2
5.0

4.0
.3
0
.0
3
.4
0
.5

.0
0

2
.8
0.6

1
1.6
0.7

1.4
0.8

1.2
0.9

1.0

Figure 5.3 Delivered-power contours from load-dependent X-parameter model (line) and
independent load-pull measurements (symbols).

16
14
12
Voltage (V)

10
1.0
0.9

1.2
0.8

1.4
0.7

1.6
0.6

1.8

8
2.0
0.5

0.4

3.0
6
0.3
4.0

5.0
4
GammaLoad[1]

0.2

2
0.1 Z1 10

0.0 0.2 0.4 0.6 0.8 1.0

20
0.1

0.2

0.3

0.4

0.5

0.6

0.7
0.8
0.9
1.0

1.2

1.4
1.6
1.8
2.0

3.0

4.0
5.0

10

20

20
Time (ns)
0.1
10
0.5
0.2 5.0

.3
4.0 0.4
0
Current (A)

.0
3
.4
0
0.3
.5

.0
0

2
.8
0.6

1
1.6
0.7

1.4
0.8

0.2
1.2
0.9

1.0

0.1

0.0
0.0 0.2 0.4 0.6 0.8 1.0
Time (ns)

Figure 5.4 Voltage and current from load-dependent X-parameter model (line) and independent
load-pull measurements (symbols) at Z2,1 Z1.
122 Multi-tone and multi-port cases

20

15

Voltage (V)
10

1.0
0.9

1.2
0.8

1.4
0.7

1.6
0.6

1.8
5

0
0.5

2.
4
0.
3.0
0.3
0
Z2 4.0

5.0
GammaLoad[1]

0.2

0.1
10
5
20 0.0 0.2 0.4 0.6 0.8 1.0
0.1

0.2

0.3

0.4

0.5

0.6
0.7
0.8
0.9

2.0

3.0

4.0
5.0
1.0

1.2
1.4
1.6
1.8

20
10
20 Time (ns)
0.1 10
0.30
0.2 5.0

.3
4.0 0.25
0

Current (A)
.0
3
.4
0
0.20
.5

0
0

.
2
.8
0.6

1
1.6
0.7

1.4
0.8

0.15
1.2
0.9

1.0

0.10

0.05
0.0 0.2 0.4 0.6 0.8 1.0
Time (ns)

Figure 5.5 Voltage and current from load-dependent X-parameter model (line) and independent
load-pull measurements (symbols) at Z2,1 Z2.
1.0
0.9

1.2
0.8

1.4
0.7

1.6
0.6

1.8
2.0
5
0.

4
0.
3.
0
0.30
0.3
Z2 4.0

5.0
GammaLoad[1]

0.2
0.25
Current (A)

10
0.1
20
0.20
0.1

0.2

0.3

0.4

0.5

0.6
0.7
0.8
0.9

2.0

3.0

4.0
5.0
1.0

1.2
1.4
1.6
1.8

20
10

20
0.1
10 0.15
0.2 .0

.0
5

4
0.10
.3
0
.0
3
.4
0
0.15
.5

10 12 14 16 18
.0

2 0 2 4 6 8
0

2
1.8
0.6

1.6
0.7

1.4
0.8

1.2
0.9

Voltage (V)
1.0

Figure 5.6 Dynamic load-line: measured (symbols) versus simulated (line) at Z2,1 Z2.
 5.5 Load-dependent X-parameters of a GaAs FET 123

The advantage of X-parameters is that the measured characteristics of this FET can be
used in a nonlinear circuit design immediately after obtaining the X-parameter model
from a measurement process on the hardware DUT.
The device model can be placed in an environment with even moderately large
harmonic mismatches, and the resulting DUT behavior is predicted accurately, even
without having controlled the harmonic impedances in the X-parameter measurement
process. This capability to predict beyond the precise measured harmonic termination
conditions is quite analogous to the situation discussed in Chapter 3, where 50
X-parameter measurements could predict well the variation of the DUT behavior with
load impedance for a moderate amount of mismatch.

5.5.1 Load-dependent X-parameter model of a GaN HEMT: estimating the effect

of independent harmonic impedance tuning
A load-dependent X-parameter model using (5.6) and (5.7) was developed for a
10 W GaN packaged transistor in [1]. The part is a Cree CGH40010 GaN HEMT [5].
A hardware conguration like that of Figure 5.2 was used. A xed number of nine load
states was used in the characterization for X-parameter extraction. The harmonic
terminations were uncontrolled during the data acquisition for the X-parameter model;
the tuner provides various values of harmonic impedances as it moves from one
fundamental complex load state to the next.
Validation measurements were made using a full multi-harmonic load-pull system.
Three output tuners were used for the validation measurements to control, independ-
ently, the rst, second, and third harmonic impedances. The hardware conguration for
validation is shown in Figure 5.7. Nine tuner states for each harmonic were used to
present the set of independently specied fundamental and harmonic impedances to the
device output, for a total of 93 729 load states.

GPIB

Bias
NVNA
Tees

Tuner
DC Supply NVNA
Control
Firmware
Software
USB

Passive Passive Passive Passive

Tuner DUT Tuner Z1 Tuner Z2 Tuner Z3

Figure 5.7 Multi-harmonic load-pull hardware conguration for validation measurements of

X-parameter model of GaN HEMT.
124 Multi-tone and multi-port cases

70
60
Load conditions at harmonic frequencies 50

PAE (%)
1.0
40

2.0
5
0.
30
Reflection coefficient

20
5.0
0.2
10 10
20 6 8 10 12 14 16 18 20 22 24
0.5

2.0

5.0
1.0

20
10
Input power (dBm)
20
Fundamental 10
2 100
0. .0
Harmonic_2nd 5
80
Harmonic_3rd
60

Voltage (V)
.5

.0
0

40
1.0

20
0
20
40
0.0 0.5 1.0 1.5 2.0 2.5
Load line Time (ns)
1.4 1.4
1.2 1.2
1.0 1.0
Current (A)
Current (A)

0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
0.2 0.2
40 20 0 20 40 60 80 0.0 0.5 1.0 1.5 2.0 2.5
Voltage (V) Time (ns)

Figure 5.8 Harmonic load impedances set A: validation of X-parameter GaN HEMT model using
multi-harmonic load pull. Results from simulations using X-parameter model extracted into
controlled fundamental impedance only (line) versus measurements (symbols).

In the validation process, the set of harmonic impedances, controlled in the validation
experiments, is taken into the nonlinear simulator and presented to the X-parameter
model. The model was constructed by controlling only the fundamental load and
extracting the harmonic sensitivity functions at nominally 50 . The model uses the
sensitivity functions in (5.6) and (5.7) to predict the effects of tuning the harmonic loads
to any value, even outside the Smith chart. Comparison is made of the measured and
simulated time-domain voltage and current waveforms, the dynamic load lines, and the
power-added efciency (PAE) from small to large input power for each of the harmonic
load impedances specied in the top left of Figure 5.8Figure 5.11. Figure 5.12 and
Figure 5.13 demonstrate the cases when the second and third harmonic impedances are
swept around the edge of the Smith chart, meaning that the reection coefcient has
 5.5 Load-dependent X-parameters of a GaAs FET 125

70
60
Load conditions at harmonic frequencies 50

PAE (%)
1.0
0.9

1.2
0.8

1.4
40

0.7

1.6
0.6

1.8
2.0
0.5
0.
4
0
30
3.
Reflection coefficient

0.3
4.0 20
0.2 5.0

10
10
Harmonic_2nd
0.1
20 6 8 10 12 14 16 18 20 22 24
0.1

0.2

0.3

0.4

0.5

0.6
0.7
0.8
0.9

2.0

3.0

4.0
5.0
1.0

1.2
1.4
1.6
1.8

20
10
20
Input power (dBm)
0.1
Fundamental 10

80
0.2 .0
5
.0
4

Harmonic_3rd
0
.3
.0
3
60

Voltage (V)
.4
0
.5

40
.0
0

2
1.8
0.6

1.6
0.7

1.4
0.8

1.2
0.9

1.0

20

20
0.0 0.5 1.0 1.5 2.0 2.5
Load line Time (ns)
1.5 1.5

1.0 1.0
Current (A)
Current (A)

0.5 0.5

0.0 0.0

0.5 0.5
40 20 0 20 40 60 80 0.0 0.5 1.0 1.5 2.0 2.5
Voltage (V) Time (ns)

Figure 5.9 Harmonic load impedances set B: validation of X-parameter GaN HEMT model
using multi-harmonic load pull. Results from simulations using X-parameter model extracted
into controlled fundamental impedance only (line) versus measurements (symbols).

magnitude 1 in these cases. The results demonstrate that the X-parameter model,
extracted by controlling only the fundamental impedance value and without controlling
harmonic impedance, nevertheless predicts accurately the DUT response to the inde-
pendently tuned (controlled) fundamental and harmonic impedances over the entire
Smith chart.
The harmonic load-pull characterization required three independent load tuners,
one each for the fundamental, second, and third harmonics, respectively, at the
output. This demonstration is a direct experimental validation of the harmonic
superposition principle [3] that led to (5.6) and (5.7). This predictive capability
of X-parameters dramatically reduces the hardware complexity, the number of
126 Multi-tone and multi-port cases

70
60
Load conditions at harmonic frequencies 50

PAE (%)
1.0
0.9

1.2
0.8
40

1.4
0.7

1.6
0.6

1.8
2.0
0.5
4
0. 30
0
3.
Reflection coefficient

0.3
4.0 20
0.2 5.0

Harmonic_2nd 10
10
6 8 10 12 14 16 18 20 22 24
0.1
20
0.1

0.2

0.3

0.4

0.5

0.6
0.7
0.8
0.9

2.0

3.0

4.0
5.0
1.0

1.2
1.4
1.6
1.8

20
10
20
Input power (dBm)
0.1
Fundamental 10

100
0.2 .0
5
.0
4 80
0
.3
Harmonic_3rd .0
3
60

Voltage (V)
.4
0
.5

.0
0

40
1.8
.6
0

1.6
0.7

1.4
0.8

1.2
0.9

1.0

20
0
20
40
0.0 0.5 1.0 1.5 2.0 2.5
Load line Time (ns)
1.4 1.4
1.2 1.2
1.0 1.0
Current (A)
Current (A)

0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
0.2 0.2
40 20 0 20 40 60 80 0.0 0.5 1.0 1.5 2.0 2.5
Voltage (V) Time (ns)

Figure 5.10 Harmonic load impedances set C: validation of X-parameter GaN HEMT model using
multi-harmonic load pull. Results from simulations using X-parameter model extracted into
controlled fundamental impedance only (line) versus measurements (symbols).

measurements, and the resulting data le size for the design of high-efciency
ampliers where harmonic terminations may still have signicant inuence on the
DUT behavior. Should the reection coefcient at a harmonic and the DUT sensi-
tivity to harmonic injection both be very large, the X-parameter framework can treat
the harmonic incident waves without the linearity assumption by suitable modica-
tion of (5.6) and (5.7).
The ideal number of X-parameter measurements per power and frequency can be
computed from (5.14):

N X -par 1 LSOP two ports  two phases per portN Z 1 1 2  2N Z 1 :

5:14
 5.5 Load-dependent X-parameters of a GaAs FET 127

70
60
Load conditions at harmonic frequencies 50

PAE (%)
1.0
40
Harmonic_2nd

2.0
5
0.
30
Reflection coefficient

20
5.0
0.2
10 10
20 6 8 10 12 14 16 18 20 22 24
0.5

2.0

5.0
1.0

20
10
Harmonic_3rd Input power (dBm)
20
Fundamental 10
2 80
0. .0
5
60

Voltage (V)
40
.5

.0
0

2
1.0

20
0
20
40
0.0 0.5 1.0 1.5 2.0 2.5
Load line Time (ns)
1.4 1.5

1.0 1.0
Current (A)
Current (A)

0.5 0.5

0.0 0.0

0.5 0.5
40 20 0 20 40 60 80 0.0 0.5 1.0 1.5 2.0 2.5
Voltage (V) Time (ns)

Figure 5.11 Harmonic load impedances set D: validation of X-parameter GaN HEMT model using
multi-harmonic load pull. Results from simulations using X-parameter model extracted into
controlled fundamental impedance only (line) versus measurements (symbols).

For the harmonic load-pull validation, the number of experiments is given by

(5.15):
N harm l-p N Z 1  N Z 2  N Z 3 : 5:15
For this example, we have N Z i 9 for i 1, 2, 3, so only 45 measurements are required
for the X-parameter model compared to 729 for the harmonic load-pull measurements.
Despite a factor of 16 fewer measurements, the X-parameter model is able to match
closely the actual independently tuned harmonic load-pull results. This saves over an
order of magnitude of measurement time and le size. It also improves the simulation
speed of the resulting model.
128 Multi-tone and multi-port cases

75

Load conditions at harmonic frequencies

50

PAE (%)
1.0
0.9

1.2
0.8

1.4
0.7

1.6
0.6

1.8
2.0
0.4 0.5

Harmonic_3rd 3.0 25
Reflection coefficient

0.3
4.0

0.2 5.0

0.1
10

20
0
0 50 100 150 200 250 300 350 400
0.1

0.2

0.3

0.4

0.5

0.6
0.7
0.8
0.9

2.0

3.0

4.0
5.0
1.0

1.2
1.4
1.6
1.8

20
10
0.1
20
Phase of Load Gamma on H2 (deg)
10
Fundamental
0.2 5.0

4.0
46
.3
0
.0
3
.4
0
44
.5

.0

Pdel (dBm)
0

2
.8
.6

1
1.6
0

0.7

42
1.4
0.8

1.2
0.9

1.0

Swept 2nd harmonic load 40

(unmarked circles)
38

36
0 50 100 150 200 250 300 350 400
Phase of Load Gamma on H2 (deg)

Figure 5.12 PAE and power delivered (Pdel) versus phase of swept second-harmonic (H2) load
impedance at the edge of the Smith chart for xed fundamental and third-harmonic load
impedances. Results from simulations using X-parameter model extracted into controlled
fundamental impedance only (line) versus measurements (symbols).

5.5.1.1 Harmonic tuning may not require harmonic load-pull characterization

The above example is important for a number of reasons. The rst is that the harmonic
terminations presented to a transistor can signicantly affect its properties. This is
evident in the waveforms shown in Figure 5.8Figure 5.11 and is a well-known fact
in amplier design [6]. The key here is that just measuring, through the X-parameter
formalism, the harmonic sensitivities (X (S) and X (T) functions) as functions of the
fundamental load is often sufcient to predict very accurately the effects of independ-
ently controlling these harmonic loads, as they vary over the entire Smith chart and even
beyond.

5.5.1.2 Source pull is not required

Another important consequence of the X-parameter model is that it will give correct
results when placed into an environment with arbitrary impedance at the source,
provided that, when the X-parameter model is extracted, the power is swept from low
values to the maximum power delivered. This is derived in [1]. This means that it is not
necessary to perform an additional fundamental source-pull characterization in order to
have the X-parameter model give the correct response when embedded in an arbitrary
port 1 impedance environment in the circuit design.
 5.6 Doherty power amplifier design and validation 129

75

Load conditions at harmonic frequencies

50

PAE (%)
1.0
0.9

1.2
0.8

1.4
0.7

1.6
0.6

1.8

0
0.5

2.
Harmonic_H2
4

25
0.
3.0
Reflection coefficient

0.3
4.0

0.2 5.0

0.1
10
0
50 100 150 200 250 300 350 400
20
0
0.1

0.2

0.3

0.4

0.5

0.6
0.7
0.8
0.9

2.0

3.0

4.0
5.0
1.0

1.2
1.4
1.6
1.8

20
10
20 Phase of Load Gamma on H3 (deg)
0.1
Fundamental 10

0.2 5.0

4.0
46
.3
0
.0
3

0.
4
44
.5

Pdel (dBm)
.0
0

2
.8
0.6

1
1.6

42
0.7

1.4
0.8

1.2
0.9

1.0

Swept 3rd harmonic load 40

(unmarked circles)
38

36
0 50 100 150 200 250 300 350 400
Phase of Load Gamma on H3 (deg)

Figure 5.13 PAE and Pdel power versus phase of swept third harmonic load impedance at the edge
of the Smith chart for xed fundamental and second-harmonic load impedances. Results from
simulations using X-parameter model extracted into controlled fundamental impedance only
(line) versus measurements (symbols).

5.6 Design example: Doherty power amplifier design and validation

This section presents a complete Doherty power amplier design, entirely within the
circuit simulator, based on X-parameter models obtained from NVNA measurements of
unmatched GaN power transistors [4]. A key feature of the approach is the application
of an active source-injection-based X-parameter measurement setup. This is demon-
strated to be necessary for providing sufcient data for an accurate X-parameter model
of the transistor used for the auxiliary amplier over the range it will be exercised in
Doherty operation.
The rst-pass design success only one fabrication build was required to meet the
design specications is conrmed by independent nonlinear measurements of the
fabricated Doherty power amplier.

5.6.1 Doherty power amplifier

The Doherty power amplier is a high-efciency architecture based on combining two
or more ampliers, typically operating at different classes, together with coherent input
and output phasing, to provide an overall characteristic that is quite efcient, and also
130 Multi-tone and multi-port cases

main amplifier

Input Output
Transistor
Match Match

GMAIN
0
90 deg, Z0
RF IN Splitter
90
auxiliary amplifier
90 deg, Z0 /2
Input Output
Transistor RF OUT
Match Match

GAUX

Figure 5.14 Ideal two-way Doherty power amplier block diagram.

reasonably linear, over a wide range of input power. Details of the operating principles
of the Doherty amplier can be found in [6] and [7]. The design example discussed here
is taken from [4].
A simple block diagram for a two-way Doherty power amplier is given in Figure
5.14. The two ampliers, with transistors biased for different classes of operation,
present loads to one another that are designed to vary as a function of total applied
input power. This mutual load pulling of the constituent active devices denes a
coupled nonlinear problem for the overall performance of the Doherty power amplier.
It cannot be solved by independent scalar load-pull characterizations of each of the
individual transistors. The details of the phase relationships of the load pulling are
critical for the operation of the overall Doherty design.
The main and auxiliary amplier matching networks were designed using fundamen-
tal and harmonic impedance tuning of the measured transistor X-parameter models in
the simulator. Proper splitter and combiner circuitry was also designed and implemented
to achieve optimum Doherty output power and power-added efciency.
While ideal design equations can help with an initial design, [6], an optimized design,
based on non-ideal components, requires accurate nonlinear models of both main and
auxiliary power ampliers and their constituent transistors. This design example starts from
packaged, unmatched transistors, from which the main and auxiliary ampliers, including
their input- and output-matching networks, must be designed. Transistor X-parameters,
measured under active-source-injection conditions, provide a natural measurement-based
modeling approach, from which the entire Doherty amplier can be designed, enabling the
details of the mutual interaction between the active components to be solved self-
consistently and design parameters to be optimized in the simulator environment.

5.6.2 X-parameter characterization of the transistors

The transistors used in this work are unmatched 10 W high-electron-mobility transistors
(HEMT) (CGH40010F from CREE). These transistors are biased in deep class AB and
class C operation, for the main and auxiliary ampliers, respectively.
 5.6 Doherty power amplifier design and validation 131

The transistor used for the main amplier is biased in an on-state, and can therefore be
characterized by X-parameters measured using the NVNA under passive load-pull condi-
tions as described in Section 5.3. However, the transistor used for the auxiliary amplier is
biased off in the absence of RF power, and will therefore experience a load outside the
Smith chart for the low-power region of operation. This is because, while off, the transistor
does not generate any B2,1 of its own, yet sees signicant A2,1 coming from the output of the
main amplier. This means that the nonlinear properties of the transistor in this range
cannot be characterized under passive load-pull conditions, and active-source-injection
characterization for the X-parameter modeling must be used, as discussed in Section 5.4.3.

5.6.2.1 Nonlinear measurement setup

The measurement system used to characterize the transistor biased for the auxiliary
amplier is shown in Figure 5.15. The gure inset shows the transistor mounted in a
xture for measurement. In this setup, the NVNA employs active source injection to
enable device characterization both inside and outside the Smith chart. The auxiliary
amplier transistor will experience extreme load conditions at input power levels below
and near transistor turn-on. It is critical for the Doherty amplier design to capture
data in this transition region. It is impossible to acquire such data with measurement
setups using passive tuners only. The measurement algorithm controls load conditions
by adjusting the amplitude and phase of the RF vector signal generator. The source
applying stimulus on port 2 is controlled to achieve specied A2,1 values. The
X-parameter functions are tabulated versus the complex values of the incident waves
at each port, as discussed in Section 5.2.

PNA-X/NVNA

DC Power Analyzer Multimeter

A1 B2
Tuner
High-power Active RF
DUT
B1 A2 Load-pull Test Set Source

Current probe

Figure 5.15 Active load-dependent X-parameter measurement setup.

132 Multi-tone and multi-port cases

All measurements are performed under pulsed continuous-wave test conditions using
a 1.3 GHz carrier frequency with 100 s pulse width and 10% duty cycle. Pulse
generators and RF pulse modulators internal to the PNA-X network analyzer are used
to provide the required test conditions. Two DC power analyzer modules provide
constant gate and drain DC bias to the transistor under test. The drain current is
measured using a current probe, which is placed between the transistor drain terminal
and large electrolytic capacitors at the transistor test xture. Several in-pulse current
samples are acquired and averaged by the NVNA to ensure very accurate power-added
efciency (PAE) measurements.

5.6.3 X-parameter model validation

Prior to the full Doherty design, the X-parameter models of the individual transistors are
validated extensively. They are compared to DC and small-signal S-parameter meas-
urements over the range of bias conditions. Then they are validated against large-signal
measurements, at load conditions included in the extraction data (to verify that the
model plays back large-signal measurements correctly) and independent load conditions
not used during model extraction (to validate interpolation capabilities within the
measured load grid). Example large-signal validation results are shown in Figure 5.16

(a)
40

30

20

10
12 14 16 18 20 22 24 26 28 30 32
(b)
40
Power (dBm)

20

20
40
12 14 16 18 20 22 24 26 28 30 32
(c)
40
20
0
20
40
60
12 14 16 18 20 22 24 26 28 30 32
Available input power (dBm)

Figure 5.16 Validation results for class C transistor intended for auxiliary amplier.
Simulated using measurement-based X-parameter model (solid lines) and measured (symbols).
(a) Fundamental frequency; (b) second harmonic; and (c) third harmonic.
 5.6 Doherty power amplifier design and validation 133

(a)

1.2
Simulated
1 Measured
Drain current (A)

0.8
0.6
0.4
0.2
0
0.2
0 10 20 30 40 50 60
Drain voltage (V)
(b)

Simulated
0.6 Measured
Drain current (A)

0.5
0.4
0.3
0.2
0.1
0
12 14 16 18 20 22 24 26 28 30 32
Available input power (dBm)

Figure 5.17 Validation results for class C transistor intended for auxiliary amplier. Simulated
using measurement-based X-parameter model (solid lines) and measured (symbols). (a) Dynamic
load line; (b) DC-bias current versus input power.

and Figure 5.17. In Figure 5.16, a xed value of A2,1 with magnitude 12 dBm and phase
75 degrees relative to the phase of A1,1 is simultaneously incident at port 2 of the
transistor. The 12 dBm output power below transistor cutoff (at about 18 dBm input
power) in Figure 5.16(a) is therefore due to the nearly complete reection of this wave
from the transistor output impedance.
Figure 5.18 shows measured and simulated fundamental output power versus input
power in the presence of an additional A2,1 incident wave of varying magnitude and
phase angle. The output impedance of the transistor is a highly nonlinear function of
transistor input drive. At power levels below turn-on (approximately 18 dBm), the
transistor output impedance is very large. When no A2,1 is present (lower trace of Figure
5.18(a)), the transistor transmits negligible B2,1. For levels of A2,1 from 10 dBm to
30 dBm (remaining traces of Figure 5.18(a)), the large B2,1 below turn-on is the result of
the almost complete reection of the injected A2,1 wave (Figure 5.18(a)). At higher input
power levels, above turn-on, the transistor output impedance changes both output
reection magnitude and phase. At certain input power levels and A2,1 phase angles,
injected and transmitted powers cancel almost completely. These cancellations are hard
134 Multi-tone and multi-port cases

(a)

40

30
B2,1(dBm)

20

10 No A2,1 present
A2,1 = +10 dBm
0 A2,1 = +20 dBm
A2,1 = +30 dBm
10
12 14 16 18 20 22 24 26 28 30 32
Available input power (dBm)
(b)

40

30
B2,1(dBm)

20
(A2,1) = 0 deg
(A2,1) = 5 deg
10
(A2,1) = 25 deg
(A2,1) = 50 deg
0
12 14 16 18 20 22 24 26 28 30 32
Available input power (dBm)

Figure 5.18 Validation results for class C transistor intended for auxiliary amplier. Simulated
using measurement-based X-parameter model (solid lines) and measured (symbols). Fundamental
output power versus input power at different A2,1 magnitudes (a) and phases (b).

to predict accurately, but are important because they result in highly nonlinear transfer
characteristics with input power (Figure 5.18(b)). The X-parameter model (solid line)
captures very accurately these measured transfer characteristics (traces with symbols)
and thereby enables accurate design of the Doherty output combiner with proper output
alignment of the main and auxiliary ampliers.
In a typical high-efciency power amplier design, tuning of the higher-order har-
monic loads, at both the input and the output of the active device, is necessary in order to
achieve optimum performance. Although the measurement setup depicted in Figure 5.15
controls fundamental terminating impedances only, the X-parameter active-injection-
measurement process extracts the DUT harmonic load sensitivity functions, X (S) and
X (T), at both the input and output ports, as discussed in Section 5.3. These terms enable the
X-parameter model to take into account independent harmonic impedance tuning in the
design of the input- and output-transistor-matching networks. The harmonic tuning
capabilities of the X-parameter model are therefore essential to achieve an optimal design.
The effect of the X (S) and X (T) terms is implicitly recognized from the comparison
between the simulated and measured fundamental load-pull results presented in Figure
5.19. Here, measured results (symbols) are accurately reproduced by the model (lines)
 5.6 Doherty power amplifier design and validation 135

Measured optimum:
40.4 dBm @ (33+j*12)

Simulated optimum:
40.4 dBm @ (33+j*12)

Figure 5.19 Measured (symbols) and simulated (lines) contours for power delivered into the load.

only when the harmonic impedances of the measurement system are embedded properly
in the circuit simulator and presented to the X-parameter model. In fact, measured
second- and third-harmonic impedances presented to the power transistor drain terminal
were relatively far from 50 during the characterization. The X-parameter measure-
ment process calibrates out these uncontrolled harmonic impedances, accounting for
them using the harmonic sensitivity terms. The resulting X-parameter model can
therefore properly represent the DUT performance when presented with a wide range
of harmonic loads in a design, even though the harmonic impedances were not
controlled during the X-parameter measurements.

5.6.4 Doherty power amplifier design using X-parameters

The ADS schematic used for the Doherty power amplier design is shown in Figure
5.20. The X-parameter models for the main and auxiliary ampliers are at the center.
Harmonic power at the output of the main and auxiliary ampliers is generally
unwanted and should be reected back to the active devices with proper phase angles.
Several matching topologies may fulll such requirements. Here, low-pass lter struc-
tures are employed and realized using microstrip elements on a high-quality Rogers
Duroid 6010 laminate.

5.6.4.1 Output match

Design objectives are to maximize output power and drain efciency while also
minimizing harmonic content at the transistor outputs. The fundamental load impedance
is thus chosen as a compromise between the power delivered to the load (Pdel) and the
PAE. To improve third-harmonic suppression, open-circuited stub elements are
designed with electrical lengths corresponding to short circuits at the third harmonic.
Final transformation to the transistor drain terminal is achieved using a DC feed with
an 80o electrical short at the second harmonic. Impedance transformations are all
136 Multi-tone and multi-port cases

Main amplifier X-parameter

Auxiliary amplifier X-parameter

Input Match
Output Match

Figure 5.20 Schematic of Doherty design using X-parameter models from measurements
of individual power ampliers.

performed within a 0.6 quality factor circle to reduce element sensitivity to PCB
fabrication tolerances. Proper phase alignment of the two amplier outputs (and
power-leakage reduction) is achieved with transmission line transformers.

5.6.4.2 Input match

The purpose of the input-matching network is to transform the source impedance to
an appropriate gate impedance that yields optimum power gain and stability. At lower
frequencies, the transistor has higher gain and is prone to oscillate. A resonator is added
to limit the source reection coefcient for frequency octaves lower than the funda-
mental frequency. The DC gate-bias feed is constructed using a high-impedance
quarter-wavelength transformer with degraded Q in order to lower the Q of the reec-
tion coefcient presented at the second-harmonic frequency. A branch-line hybrid
coupler with meandered series arms drives the two amplier inputs in quadrature. The
meander lines save board space, and additional electromagnetic coupling is accounted
for using EM simulations.

5.6.5 Results
The nal manufactured Doherty power amplier is shown in Figure 5.21. Simulated
and measured amplier delivered output power versus available power (Pavs) and
drain efciency versus delivered output power is shown in Figure 5.22. Measured
 5.6 Doherty power amplifier design and validation 137

Gate bias Drain bias

RF input

RF output

Termination

Gate bias Drain bias

Figure 5.21 Manufactured Doherty power amplier mounted on an aluminum heat sink.

(a)
50
Simulated
45 Measured
Pdel (dBm)

40
Simulated sat. Pdel = 44.5 dBm
35
Measured sat. Pdel = 44.2 dBm
30
15 20 25 30 35
Pavs (dBm)
(b)
Drain efficiency (%)

80 Simulated
Measured
60
Simulated max. eff. = 76.9 %
40
Measured max. eff. = 75.7 %
20

25 30 35 40 45
Pdel (dBm)

Figure 5.22 Simulated (lines) and measured (symbols) Doherty AM-to-AM and drain efciency
versus delivered power.

saturated output power is 44.2 dBm (simulation predicts 44.5 dBm), and the
measured maximum drain efciency is 75.7% (simulation predicts 76.9%). Overall,
the simulation and measurement results agree very well. Measurements of the fabri-
cated power amplier are within 0.3 dB and 1.2% between simulated and measured
saturated output power and maximum drain efciency, respectively. The characteristic
double-peak behavior of Figure 5.22(b) is an indication that the design is working
in Doherty mode.
138 Multi-tone and multi-port cases

5.7 Incommensurate signals

The treatment of X-parameters up to now has been restricted to nonlinear multi-tone

maps from signals on a harmonic grid onto the same grid of frequencies. This is
sufcient for nonlinear devices being stimulated by a single large tone and generating
harmonic spectral responses at multiple ports. It also includes effects of large mismatch
at the fundamental frequency, and also at the harmonics, as can happen for waveform
engineering of high-efciency ampliers by specic harmonic-matching conditions. All
signals, including the incident and scattered waves, are necessarily periodic signals with
the same period as the inverse of the grid spacing in the frequency domain.
In this portion of the chapter, the treatment is extended to include two incommensur-
ate incident signals. The signals can be incident either on the same port or on different
ports of the DUT.
Two sinusoidal signals, with (non-zero) frequencies f1 and f2, are incommensurate if
the ratio of their frequencies is irrational. Equivalently, f1 and f2 are incommensurate
if there are no two non-zero integers for which the commensurability condition of
Chapter 2 is fullled. Mathematically, this is indicated by (5.16):
m, n 2 such that mf 1 nf 2 0 ) m n 0: 5:16
An example of the incident and scattered waves of a nonlinear 2-port subject to
stimulation at the input with two incommensurate frequencies is given in Figure 5.23.
Only a nite number of spectral tones in the output spectrum are shown.
Because the output spectrum generated by a nonlinear DUT can be reected back into the
device due to mismatch, or propagate through other linear and nonlinear components, spectral
maps dened on all intermodulation products of the large incident signals need to be dened.

5.7.1 Notation for incommensurate two-tone X-parameters

The case of two large fundamental incommensurate tones and their intermodulation
products is presented here. The generalization to N large incommensurate fundamental
tones is straightforward, but the notation is somewhat tedious.

A1 B1

f1 f2 f1f2 f1 f2 2f1 2f2

2f1f2 2f2f1

Figure 5.23 Incident and simplied scattered-wave spectra for nonlinear 2-port excited by two
large incommensurate tones.
 5.7 Incommensurate signals 139

All non-zero signals considered in this section exist on a discrete subset of real
intermodulation frequencies, f, given by
f nf 1 mf 2 , 5:17
where n and m are integers. The case where f evaluates to a negative number is treated
later. Each wave variable is specied by a complex phasor indexed by a port index, p, a
positive integer indicating the port at which the stimulus is applied, and two integers
specifying the frequency of the incident signal in terms of contributions from specic
orders of each of the two fundamental tones at frequencies f1 and f2. This is shown in
(5.18) for incident and scattered waves:
Ap, n, m ,
5:18
Bp0 , n0 , m0  :
Note that there can be more ports than large tones. For example, a mixer may have
three ports but there may be only two incident large tones, one at the LO and the other
at the RF.

Example: A3,[2, 1] is the incident wave at port 3 corresponding to a frequency of

2f1  f2.
The same convention presented in the preceding chapters is used for DC stimulus and
response. If the DC stimulus at port q is a voltage, DCSq Vq, the DC response is a
current, DCRq Iq. Conversely, if the DC stimulus at port p is a current, DCSq Iq, the
DC response at that port is a voltage, DCRq Vq.
The complete multi-port multi-tone stimulus includes both the DC and RF stimuli.

Example: Equation (5.19) denes a stimulus with DC voltages applied at ports 1 and 2, a
DC current source at port 3, and incident signals applied at each port at different
intermodulation frequencies of the underlying large signals:
stim DCS; RFS  V 1 , V 2 , I 3 ; A1, 0, 1 , A3, 1, 0 A2, 1, 1 , A3, 1, 2 : 5:19

The corresponding response generally has more terms given multiple intermodulation
products produced by the DUT, such as that given in (5.20):
resp DCR; RFR  I 1 , I 2 , V 3 ; B1, 1, 0 , B1, 0, 1 , B1, 1, 1 , . . . , B1, N , M  , B2, 1, 0 , B2, 0, 1 , . . . , B3, N , M  :
5:20
A maximum order of N for harmonics of the rst fundamental tone and M harmonics of
the second tone is generally specied. This means that each of the components in the
response, (5.20), is a nonlinear function of all of the elements of the stimulus, (5.19).
These functions are generally nonlinear, so the general case is complicated.
140 Multi-tone and multi-port cases

5.7.2 Time invariance for incommensurate two-tone X-parameters

The concept of cross-frequency phase for commensurate signals was central to imple-
menting time-invariant maps from incident to scattered waves. For incommensurate
signals, things are fundamentally different, requiring a different approach.
Cross-frequency phase is an ill-dened concept for incommensurate signals. Each
successive zero crossing of the phase of the rst signal at frequency f1 corresponds to a
unique value of the phase of the second signal, 2, at frequency f2.1 The innite
sequence of 2 values is dense on the interval [0, 2), with no nite gaps; roughly
speaking, this means that all possible real values of phase are approached arbitrarily
closely. A corollary is that it is possible to nd, simultaneously, the phase of the rst
fundamental tone equal to zero and the phase of the second fundamental tone arbitrarily
close to zero, just by waiting a long enough time.
An important implication of this fact, related to taking appropriate limits in the space
of such signals, is that a two-tone steady-state signal, in the frequency domain, can be
uniquely identied by specifying only the magnitudes of the individual signals. This is
one fewer degree of freedom compared to the commensurate case considered in the rst
part of this chapter, where a phase angle was required in addition to the two magnitudes.
In an important sense, the signal dened by phasors each with phase 0 will dene our
reference for enforcing a time-invariant description on the map from inputs to outputs.
Consider now the two fundamental signals as well as signals at any other intermo-
dulation frequency. Each intermodulation frequency is manifestly commensurate with
respect to the two fundamental tones. This follows trivially from (5.21) and (5.22).
Equation (5.21) denes an equation with integer coefcient among three frequencies
that is clearly satised when f3 is given by (5.22):
nf 1 mf 2  1f 3 0, 5:21
f 3 f reqfAp, n, m g nf 1 mf 2 : 5:22
The time-invariant process is as follows. Apply independent phase shifts to each of the
fundamental tones until each has zero phase. This is shown in (5.23),

Ap, 1, 0 ! Ap, 1, 0 P1

1, 0 and Ap0 , 0, 1 ! Ap0 , 0, 1 P1
0, 1 : 5:23

where P[1,0] and P[0,1] are the phases of the two large tones. This translates the large-
signal components to the reference state.
Any signal component at an intermodulation frequency labeled [n,m] is commensurate
with respect to the fundamental tones, and therefore has a well-dened cross-frequency
phase with respect to both of the fundamental tones. The same reasoning of Chapter 2
can now be invoked to deduce that the phase of the incident waves with indices [n,m]
must be coherently shifted by the sum of the integral powers of the individual phases
of the large tones according to the respective indices. This is shown in (5.24):
Aq, n, m ! Aq, n, m Pn m
1, 0 P 0, 1 : 5:24

1
This is not the case for commensurate signals, where there is only a nite number of possible values for the
phase of the second signal, 2 [8].
 5.7 Incommensurate signals 141

Since the scattered waves (the response of the nonlinear system) also have components
only on the same intermodulation spectrum, we must likewise phase shift the scattered
waves by the same terms. This is given in (5.25) for the reference response:
Bq, n, m ! Bq, n, m Pn m
1, 0 P 0, 1 : 5:25
Therefore, under a transformation to the reference state dened by (5.23) and (5.24),
where the reference response is given by (5.25), one must have the relationships given
in (5.26) for the DUT in order that the equivalent phase shifts of the intermodulation
terms preserve their cross-frequency phase. Here the phase factors in (5.25) have been
taken to the right-hand side of (5.26):


Bq, n, m F q, n, mDCS, Ap1 , 1, 0 , Ap2 , 0, 1, , . . . , Ap0 , n0 , m0  
n0 m0 5:26
F q, n, m DCS, Ap1 , 1, 0 P1 1
1, 0 , Ap2 , 0, 1 P 0, 1 , . . . , Ap0 , n0 , m0  P 1, 0 P 0, 1 P 1, 0 P 0, 1 :
n m

Not all functions have the transformation properties of (5.26). Only those functions F(.)
with the property (5.26) are admissible spectral mappings for a time-invariant DUT
being excited by signals at all intermodulation frequencies associated with two funda-
mental tones at incommensurate frequencies.
Noting that the rst two arguments at the reference state are real, the introduction of
the X-parameter function is dened according to (5.27):
 
FB 0
m0
X q, n, m DCS, jAp1 , 1, 0 j, jAp2 , 0, 1 j, . . . , Ap0 , n0 , m0  Pn P
1, 0 0, 1


F q, n, m DCS, Ap1 , 1, 0 , Ap2 , 0, 1 , . . . , Ap0 , n0 , m0  5:27a
n m :
P1, 0 P0, 1

This means that the DUT X-parameter functions can be dened on a manifold with two
fewer real dimensions than the general multi-tone spectral mappings.
Combining the property expressed in (5.26) with the denition in (5.27a) leads to the
X-parameter model to be formulated, as shown in (5.27b).
 
FB 0
m0
Bq, n, m X q, n, m DCS, jAp1 , 1, 0 j, jAp2 , 0, 1 j, ...,Ap0 , n0 , m0  Pn
1, 0 P 0, 1 P 1, 0 P0, 1 : 5:27b
n m

5.7.3 Reference LSOP

The discussion in Section 5.7.2 means it is possible to choose the reference LSOP for
the state of a DUT excited by two incommensurate tones to be identied with the
phasors of the two signals, each of which corresponds to zero phase. That is, refLSOPS
depends only on the magnitudes of Ap1 , 1, 0 and Ap2 , 0, 1 , where p1 and p2 are the ports at
which the rst and second tones are incident, respectively.

5.7.4 Spectral linearization

The nonlinear map (5.27b) depends on many variables, and is therefore very complex to
deal with in general. Just as in earlier sections, useful approximations can be developed
to simplify things dramatically for practical applications.
142 Multi-tone and multi-port cases

It is considered that, out of all the possible signal components represented by Ap,[n,m]
incident on the DUT, only the fundamental tones, Ap1 , 1, 0 and Ap2 , 0, 1 , are large, and all
other components at intermodulation frequencies are considered small. This is the case,
for example, in mixers with nearly matched ports. It is also the case for power ampliers
with ports nearly matched to 50 . In these cases, the rich intermodulation spectra will
be reected, but the terms will be small enough that the spectral linearization process
will still be a good approximation to taking their effects into account.
Equation (5.26) is therefore spectrally linearized around the reference LSOPS given
by (5.28):


ref LSOPS DCS p , jAp1 , 1, 0 j, jAp2 , 0, 1 j, 0, . . . , 0 : 5:28
Spectral linearization with respect to all other signals is applied around the refLSOPS of
(5.28). As all other previous cases, the maps (5.27b) are non-analytic, and so we need to
differentiate with respect to the intermodulation phasors and their complex conjugates,
independently. The now familiar procedure is applied, and the resulting X-parameter
model equations take the form (5.29)(5.31). The sum is over all ports, q, and all
integers n0 and m0 :
F
Bp, n, m X p, n;m ref LSOPS Pn1, 0 Pm 0, 1
X S 0
mm0
X p, n;m;q, n0 ;m0  ref LSOPS Pnn 1, 0 P 0, 1 Aq, n0 ;m0 
q, n0 , m0 5:29
X T
nn0 mm0 
X p, n;m;q, n0 ;m0  ref LSOPS P1, 0 P0, 1 Aq, n0 ;m0  ,
q, n0 , m 0

X  
Y 0
m0
I p X pI ref LSOPS Re X p;q, n0 , m0  ref LSOPS Pn P
1, 0 0, 1 A q, n 0 , m0  ,

q, n0 , m0

5:30
X  
Z 0
m0
V p X pV ref LSOPS Re X p;q, n0 , m0  ref LSOPS Pn
1, 0 P 0, 1 Aq, n0 , m0  :
q, n0 , m 0

5:31
The X-parameter sensitivity functions are therefore identied as the partial derivatives
of the scattered waves with respect to complex signals at the intermodulation frequen-
cies (and their conjugates) evaluated at the refLSOPS:

S Bp, n, m  0 0
X p, n, m;q, n0 , m0    Pn1, 0 Pm0, 1 ,
Aq, n0 , m0  ref LSOPS
 5:32
T Bp, n, m  n0 m0
X p, n, m;q, n0 , m0    P P :
Aq, n0 , m0  ref LSOPS 1, 0 0, 1

5.7.4.1 Examples
A two-tone multi-port model (with no explicit DCS dependence), which describes a
mixer or amplier, can be dened by (5.33):
 5.7 Incommensurate signals 143


Bp, n, m X F p, n, m jAp1 , 1, 0 j, jAp2 , 0, 1 j Pn1, 0 Pm 0, 1
X

X S p, n, m;q, j, k jAp1 , 1, 0 j, jAp2 , 0, 1 j Pnj mk
1, 0 P0, 1 Aq, j, k
q;j, k 5:33
X

X T p, n, m;q, j, k jAp1 , 1, 0 j, jAp2 , 0, 1 j Pnj mk 
1, 0 P 0, 1 A q, j, k :
q;j, k

For p1 1 and p2 2, this example describes a mixer with the RF at port 1 and LO at
port 2. The IF corresponds to port 3. The refLSOPS is specied only by the magnitude
of the RF and LO signals (assumed incommensurate).
For both p1 1 and p2 1, (5.33) describes a power amplier with a two-tone
stimulus at the input port (port 1). In this case, the port index, p, of the scattered
B-waves goes from 1 to 2. The refLSOPS is specied by the magnitudes of the two
input tones.

5.7.5 Discussion
Finite and independent phase shifts are used to shift signals to the reference LSOP and
then to shift back the scattered waves appropriately. This is the procedure used to
implement time invariance in the incommensurate case. It is noteworthy to consider that
the actual time shift necessary to reach the refLSOP (where both fundamental signals
have zero phase) may be innite. This fact can be established, rigorously, considering
sequences of time shifts and the convergence in function spaces. This level of detail is
beyond the scope of this work.

5.7.6 When intermodulation frequencies are negative

X-parameters operate in the frequency domain, mapping the spectrum of the input
stimulus to the spectrum of the DUT response. The input and output signals considered
in the time domain are real. This means that the spectra in the frequency domain are
double sided with conjugate symmetry, and hence contain redundant information by a
factor of 2. For efciency, only half the spectra need to be considered in the equations,
and only those X-parameters mapping the chosen spectral components need to be stored
in the les. In many applications, it is typical that only positive frequency and DC
components are considered. For X-parameters, the frequencies of the signals are
determined by the indices n and m and by the values of the fundamental frequencies,
fi, through (5.17). If the frequencies, fi, are swept, the frequency of the intermodulation
term corresponding to a particular set of n and m can become negative. Specically, for
the choice f1 3 GHz and for f2 swept from 2.9 GHz to 3.1 GHz in two steps, the
intermodulation term [1, 1] will correspond rst to 100 MHz and then to 100 MHz.
It is much more convenient to have a consistent indexing scheme independent of the
numerical values of the fi. Therefore, the convention is adopted that the spectral
components considered correspond to those sets of n and m where the rst non-zero
value in the set of these integers is positive. For any two index vectors [n,m] and
[n,m], each of which corresponds to the same jfj, one will be consistent with this
144 Multi-tone and multi-port cases

convention and the other will be excluded. If the frequency corresponding to the
convention on indices is negative, the corresponding positive frequency is considered
for the signal and the complex conjugate of the corresponding phasor is taken.

5.7.6.1 Examples
Wave variables and X (F) functions with intermodulation indices [2,3] are selected for
processing and storage, whereas those indexed by [3,5] are not, because, in the latter
case, the rst non-zero integer is negative. So, in this case, the corresponding inter-
modulation vector [3,5] is considered instead. If 3f1  5f2 is positive, the complex
value of the phasor at that frequency is considered. If 3f1  5f2 is negative, the complex
conjugate of the corresponding phasor is considered and its contribution attributed to
the spectral component with the corresponding positive frequency.
X-parameter functions labeled by two sets of port and intermodulation vectors have
rules to remove redundancy that can be deduced in a similar way. The details are not
described here, but the following is given as an example:
S
the parameter X p, 2, 3;q, 0, 2 is an allowed (and saved) quantity, but the parameter
S
X p, 3, 2;q, 0, 2 is not allowed.

5.7.7 X-parameter models of mixers

An example of the incommensurate two-tone X-parameter is given by the characteriza-
tion and modeling of a commercial mixer component (Mini-Circuits LAVI-22VH
(TB-433) double-balanced mixer). A schematic of the component is given in Figure 5.24.
The RF signal is applied at port 1, the LO at port 2. The conversion gain of a mixer
with IF matched to a perfect 50 termination is given by the simple X-parameter
expression (5.34):
F
jX 3, 1, 1 j
conversion gain : 5:34
jA1, 1, 0 j

electrical schematic
LO RF
FET QUAD#1

FET QUAD#2

IF

Figure 5.24 Mini-Circuits double-balanced mixer used for three-port X-parameter model. 2013
Scientic Components Corporation d/b/a Mini-Circuits. Used with permission.
 5.7 Incommensurate signals 145

Conversion gain (fixed IFfreq = 400 MHz)

Conversion gain (dB) 6

10

12

14

16
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Frequency (GHz)

Figure 5.25 Comparison of measurement-based X-parameter mixer model and manufacturers

datasheet for conversion gain (in dB) versus frequency.

LOtoIF Isolation
70
LOtoIF Isolation (dB)

60

50

40

30
0.0 0.5 1.0 1.5 2.0 2.5 3.0
LO frequency (GHz)

Figure 5.26 Comparison of measurement-based X-parameter mixer model and manufacturers

datasheet for LOIF isolation (in dB) at the IF port versus frequency.

Here we assume that the IF signal of interest is at the difference of the RF and LO
frequency (where LO frequency is lower than the RF frequency).
The X-parameters of the Mini-Circuits mixer were measured. The conversion gain
was computed from (5.34) and compared with values from the manufacturers data-
sheet. This is shown in Figure 5.25. The LO-to-IF isolation term, at the IF port, can be
computed from a similar expression given in (5.35) (see Figure 5.26):
jA2, 0, 1 j
LO-to-IF isolation F
: 5:35
jX 3, 0, 1 j

The X-parameter model has much more information than just conventional datasheet
information. An example is given by the phase of the IF signal, which is a well-dened
quantity when normalized to the phase of the RF and LO signals using (5.29). For the
Mini-Circuits mixer, this information is shown in Figure 5.27. The expression to
compute it is given simply by (5.36):
146 Multi-tone and multi-port cases

IF phase (fixed IFfreq = 400 MHz)

0
10
IF Phase (deg)

20
30
40
50
60
0.0 0.5 1.0 1.5 2.0 2.5 3.0
RF frequency (GHz)

Figure 5.27 Phase of IF signal versus frequency as contained in the X-parameter mixer model.

10
mag of 2 x LOfreq (dBm)

20
30
40
50
60
70
0.0 0.5 1.0 1.5 2.0 2.5 3.0
LO frequency (MHz)

Leakage at IF port
phase of 2 x LOfreq (deg)

200
150
100
50
0
50
100
150
200
0.0 0.5 1.0 1.5 2.0 2.5 3.0
LO frequency (MHz)

Figure 5.28 Leakage of second-harmonic signal of LO at IF port, in magnitude and phase, versus
LO frequency (LO power 21 dBm, xed IF frequency 400 MHz).

h i
F
IF phase phase X 3, 1, 1 : 5:36

Additionally, the signal at the second harmonic of the LO signal can be observed at the
F
IF port. This complex quantity, given by X 3, 0, 2 (normalized in phase to LO and RF at
the fundamental), is plotted in Figure 5.28. Many performance parameters of mixers can
be expressed in terms of X-parameters. Some examples are given in Table 5.1.
 5.8 Summary 147

Table 5.1 Mixer behavior in terms of X-parameters

Large RF at port 1, large LO at port 2, IF output at port 3

Mixer terminology X-parameter expressions or terms

F
Conversion gain (difference frequency) jX 3, 1, 1 j
jA1, 1, 0 j
F
LO leakage at IF port for perfectly matched IF X 3, 0, 1
S T
Mismatch terms at IF frequency and IF port X 3, 1, 1;3, 1, 1 X 3, 1, 1;3, 1, 1
S T
IF-to-LO isolation terms X 2, 1, 1;3, 1, 1 X 2, 1, 1;3, 1, 1
S T
LO-second-harmonic-to-IF isolation X 3, 1, 1;2, 0, 2 X 3, 1, 1;2, 0, 2

5.8 Summary

X-parameters spectrally linearized about an LSOP established by two large incident

tones were considered in two important cases. The rst case corresponds to two
signals incident at different ports at the same fundamental frequency. For power
ampliers, this entails measuring X-parameters as functions of complex-valued load
as well as power. This case includes what is sometimes known as time-domain load
pull. It was demonstrated that harmonic sensitivities (X (S) and X (T) terms) measured
as functions of the complex load at the fundamental frequency were sufcient to
predict accurately the effect of tuning the fundamental and harmonic load imped-
ances independently. This provides nearly the accuracy of harmonic time-domain
load pull, but requires only a small fraction of the measurements and time, in a
hardware conguration that is much simpler and less expensive. The resulting X-
parameter model was shown to produce excellent results for a GaN HEMT amplier.
This approach has been directly applied to more complicated high-efciency appli-
cations, for example the Doherty amplier in [4], and discussed as a complete design
example in Section 5.6.
The second case considered in this chapter corresponds to the case of two signals at
incommensurate fundamental frequencies incident at the same or distinct ports of a
nonlinear time-invariant DUT. The example provided to illustrate this formalism was a
mixer, where RF and LO signals are applied at different ports. The resulting model from
NVNA measurements on a commercial mixer showed excellent results compared to the
product datasheet, with many more interaction terms provided that could not be
obtained by any other means. The same X-parameter two-tone measurement procedure
can also be applied to two-port ampliers, producing X-parameter models where the
phases of the intermodulation products are captured along with their magnitudes. This
can enable advanced circuit-design techniques, such as derivative superposition [9], in
order to cancel distortion terms. Previous to the availability of two-tone X-parameters,
this was possible only using nonlinear compact transistor models, the accuracy of which
can be questionable.
148 Multi-tone and multi-port cases

Exercises

5.1 Prove that two sinusoidal signals, corresponding to the incident waves at ports 1
and 2 at the same fundamental frequency, always have the same relative phase,
independent of time shift, given by (5.3).
5.2 Using the ideal X-parameter experiment design of Chapter 3, derive expressions
(5.14) and (5.15) for the number of measurements required as functions of the
number of independently controlled load impedances.
5.3 Derive (5.29) for the spectral linearization around an incommensurate two-tone
LSOP.

References

[1] J. Horn, D. E. Root, and G. Simpson, GaN device modeling with X-parameters, in IEEE
Compound Semiconductor Integrated Circuit Symp. (CSICS), Monterey, CA, Oct. 2010,
pp. 14.
[2] S. Woodington, R. Saini, D. Williams, J. Lees, J. Benedikt, and P. J. Tasker, Behavioral
model analysis of active harmonic load-pull measurements, in 2010 IEEE MTT-S Int.
Microwave Symp. Dig., Anaheim, CA, may 2010, pp. 16881691.
[3] J. Verspecht, M. V. Bossche, and F. Verbeyst, Characterizing components under large signal
excitation: dening sensible large signal S-Parameters?!, in 49th ARFTG Conf. Dig.,
Denver, CO, 1997, pp. 109117.
[4] T. S. Nielsen, M. Dieudonn, C. Gillease, and D. E. Root, Doherty power amplier design in
gallium nitride technology using a nonlinear vector network analyzer and X-parameters, in
IEEE CSICS Dig., La Jolla, CA, Oct. 2012, pp. 14.
[5] Cree CGH40010 GaN HEMT; available at http://www.cree.com/rf/products/sband-xband-
cband/packaged-discrete-transistors/cgh40010.
[6] S. C. Cripps, RF Power Ampliers for Wireless Communications, 2nd edn. Norwood, MA:
Artech House, 2006.
[7] W. H. Doherty, A new high efciency power amplier for modulated waves, Proc. IRE,
vol. 24, no. 9, pp. 11631182, Sept. 1936.
[8] S. Strogatz, Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemis-
try, and Engineering. Reading, MA: Perseus Press, 1994.
[9] D. Webster, J. Scott, and D. Haigh, Control of circuit distortion by the derivative superpos-
ition method, IEEE Trans. Microw. Guid. Wave Lett., vol. 6, no. 3, pp. 123125, 1996.

Additional reading

M. S. Hashmi, F. M. Ghannouchi, P. J. Tasker, and K. Rawat, Highly reective load-pull, IEEE

Microwave, vol. 12, no. 4, pp. 96107, 2009.
J. Horn, S. Woodington, R. Saini, J. Benedikt, P. J. Tasker, and D. E. Root; Harmonic load-
tuning predictions from X-parameters, in IEEE PA Symp., San Diego, CA, Sept. 2009.
 Additional reading 149

D. E. Root, J. Horn, T. Nielsen, et al., X-parameters: the emerging paradigm for interoperable
characterization, modeling, and design of nonlinear microwave and RF components and
systems, in IEEE Wamicon 2011 Tutorial, Clearwater, FL, Apr. 2011.
R. S. Saini, S. Woodington, J. Lees, J. Benedikt, and P. J. Tasker, An intelligence driven active
loadpull system, in IEEE Microwave Measurements Conf. (ARFTG), Anaheim, CA, 2010,
pp. 14.
G. Simpson, J. Horn, D. Gunyan, and D. E. Root, Load-pull NVNA enhanced X-parameters
for PA designs with high mismatch and technology-independent large-signal device models,
in IEEE ARFTG Conf., Portland, OR, Dec. 2008, pp. 8891.
P. Tasker, Practical waveform engineering, IEEE Microwave, vol. 10, no. 7, pp. 6576, 2009.
J. Verspecht, D. Gunyan, J. Horn, J. Xu, A. Cognata, and D. E. Root, Multi-tone, multi-port, and
dynamic memory enhancements to PHD nonlinear behavioral models from large-signal meas-
urements and simulations, in 2007 IEEE MTT-S Int. Microwave Symp. Dig., Honolulu, HI,
June 2007, pp. 969972.
6 Memory

6.1 Introduction

X-parameters were introduced in Chapter 2 as frequency-domain nonlinear mappings.

The X-parameter functions are dened on the complex amplitudes of the incident
CW signals, returning the complex amplitudes of the scattered CW signals. For
xed power, bias, and load conditions, these input and output amplitudes are constant
(in time) complex numbers. As such, X-parameters were dened only for steady-state
conditions.1
Real applications deal with signals that vary in time and the corresponding time-
varying responses of DUTs to such stimuli. In particular, signals used in wireless
communications can have independent time-varying values for the amplitudes and
phases associated with each of the multiple carriers involved. These time-dependent
modulations contain the information associated with the signal. Much of communi-
cation systems design is devoted to trying to preserve the information content of these
signals while efciently amplifying, transmitting, and then demodulating them at the
receiver to recover the information.
This chapter introduces the notion of dynamics into the signals and into the
mappings between input and output stimuli that dene the DUT as a nonlinear
dynamic system. The envelope domain is introduced to dene the class of signals
considered. Quasi-static applications of static X-parameters are introduced as methods
for estimating the DUTs response to modulated signals, and limitations are dis-
cussed. Examples are provided of phenomena that require the abandonment of the
static nonlinear system description. This introduces the notion of memory, where the
response of the system depends not only on the instantaneous value of the input
signal, but also on the history of the input signal. The concept of modulation-induced
baseband memory is articulated.
The second part of the chapter is devoted to a detailed exposition of dynamic
X-parameters. This is a fundamental extension of X-parameters to modulation-induced
baseband memory effects. The theory is rigorously developed, starting from a few basic
assumptions that can be independently validated. Applications to actual DUT complex
stimulusresponse characterization and modeling for practical wide-band modulated

1
This does not include special cases of pulsed X-parameters encountered in Chapter 2.
 6.3 Quasi-static X-parameter evaluation 151

signals are presented to validate the approach. A signicant simplication of the general
dynamic X-parameters, valid in the limit of wide-band modulation, is presented under
the name wide-band X-parameters (XWB).

6.2 Modulated signals: the envelope domain

Modern communication signals can be represented efciently by a sum of high-

frequency carriers with complex amplitudes that vary in time. These signals can be
described mathematically by (6.1):
( )
X
jn t
At Re An te : 6:1
n

Here, An(t) is the time-varying complex amplitude associated with the nth carrier having
frequency n. The frequencies may or may not be commensurate. The signal A(t) can
have components at each port, but port indices are suppressed for simplicity.
The scattered waves from a nonlinear DUT stimulated with signal A(t) can also be
assumed to take a form equivalent to that of (6.1).2 This is shown in (6.2):
( )
X
jm t
Bt Re Bm te : 6:2
m

Here, {m} constitutes the set of all intermodulation frequencies produced by the DUT
in response to the signal. There can be more spectral components generated by the
nonlinear DUT, and therefore more time-varying envelopes, Bm(t), than were present in
the incident signal.
Complex envelopes are represented by their magnitudes and phases, each of which
varies in time. This is detailed in (6.3) for the incident envelopes, with a similar
expression, (6.4), holding for the envelope of each scattered wave:

An t jAn tje jAn t , 6:3

Bm t jBm tje jBm t : 6:4

The dynamic modeling problem considers how to dene the mappings from the
{An(t)} of the incident waves to the {Bm(t)} of the scattered waves.

6.3 Quasi-static X-parameter evaluation in the envelope domain

In this section, the case is considered where a static X-parameter model can be evaluated
in the envelope domain to calculate the DUTs response to a modulated input signal,
under the quasi-static approximation. This approximation asserts that, at any time t, the

2
The case of a chaotic system, where the output spectrum can be qualitatively different, is not considered.
152 Memory

actual output time-dependent envelopes can be approximated by applying the static

X-parameter mappings to the values of the input amplitudes at that same instant of time.
To see how this works, one starts with an X-parameter model that was identied from
measurements presumed to have been taken under steady-state conditions. For example,
consider a simple one-tone X-parameter steady-state mapping dened by (6.5):
F  
Bp, k X p, k jA1, 1 j e jA1, 1 : 6:5
F
All mismatch terms are neglected, for simplicity, and therefore only the X p, k terms are
considered. The time variation of the input envelope is specied by either an explicit
time-dependent IQ signal or one of several modulated source components available in
the simulator. The settings for the envelope simulation are chosen to allow adequate
time sampling of the envelope corresponding to the input signal and to allow a suitably
long period for the simulation.
In the quasi-static approximation, the DUT response, at time t, is computed to be the
value of the static X-parameter mapping, as described in (6.5), applied to the value of
the input envelope at the same time instant, t. This results in (6.6):

F  
Bp, k t X p, k A1, 1 t e jA1, 1 t : 6:6

This procedure produces time-varying envelopes for each of the harmonics, indexed by
integer k, produced by the DUT in response to the modulated amplitude around the
carrier at the fundamental frequency.
The quasi-static approximation is an extrapolation from steady-state conditions to a
dynamic (time-varying) condition. It is certainly valid for sufciently slowly varying
An(t), since it reduces to the static mapping as the instantaneous envelope amplitudes
become constant. It is an accurate approximation to the actual time-dependent response
provided that any underlying system dynamics are sufciently fast that, at any instant of
time, the DUT is nearly in its steady-state condition determined by the value of the input
amplitude at that same time. That is, the system adiabatically tracks the input from one
steady state to the next, parameterized by the time. The quasi-static approximation
breaks down as the signal modulation rate increases and becomes comparable to or
faster than timescales for which other dynamical effects become observable. In particu-
lar, electro-thermal effects, bias-line interactions, and other phenomena come into play
for which a more elaborate dynamical description of the DUT is required. This will be
discussed in more detail in the following.

6.3.1 Quasi-static two-tone intermodulation distortion from a static one-tone

X-parameter model
In the following, the theory is illustrated using an actual amplier, namely a Mini-
Circuits ZFL-11AD. This amplier is characterized by performing a one-tone
X-parameter model at a carrier frequency of 1750 MHz. The two-tone experiments
are performed with a tone spacing of 1200 Hz, whereby the tones are symmetrically
placed around the carrier frequency. A time-domain representation of the two-tone input
 6.3 Quasi-static X-parameter evaluation 153

0.2

0.1
input (V)

0.0

0.1

0.2
0 500 1000 1500 2000 2500
time (s)

Figure 6.1 Two-tone input signal in the time domain.

0.2

0.1
input (V)

0.0

0.1

0.2
0.000 0.005 0.010 0.015 0.020
time (s)

Figure 6.2 Two-tone input signal in the time domain (zoomed in).

signal is shown in Figure 6.1. The carrier oscillation period is so small compared to the
millisecond range of the time axis of Figure 6.1 that the signal looks like a solid black area.
The individual oscillations can only be seen on a higher-resolution timescale expressed
in nanoseconds rather than milliseconds. This is illustrated in Figure 6.2, which represents
a zoom on the rst 20 nanoseconds of the 2.5 milliseconds shown in Figure 6.1.
This two-tone sinusoidal signal can be represented as a single carrier with a modu-
lated complex amplitude A(t) according to (6.7). Equal magnitudes, A1, for the two
tones, are considered for simplicity:
 

A1 cos 1 t A1 cos 2 t 2A1 cos t cos 0 t
2
8 9
<   = 6:7

Re 2A1 cos t e j0 t
RefAte g:
j0 t
: 2 ;
154 Memory

0.2

0.1
re & im input (V)

0.0

0.1

0.2
0 500 1000 1500 2000 2500
time (s)

Figure 6.3 Two-tone input signal in the envelope domain: real part (solid line) and
imaginary part (dashed line).

Here, 2  1 and 0 (1 2)/2. For our example, 1 and 2 correspond to

(1750 MHz  600 Hz) and (1750 MHz 600 Hz), respectively.
From (6.7), it follows that the envelope representation A(t) of any two-tone signal,
using as carrier frequency the middle of the two frequencies, 0, is simply given by
 

At 2A1 cos t : 6:8
2
The real and imaginary parts of A(t) are shown in Figure 6.3.
Consider now the transmission of this two-tone signal through a system that is
described by a simple X-parameter model that only takes into account the transmission
F
of the fundamental signal. In other words, only the presence of a term X 2, 1 :, which is
here more simply denoted by XF(.), is considered in the following. The amplier
response B(t) using the quasi-static approximation is calculated according to (6.6),
whereby A(t) is given by (6.8). The result is given below:
 
Bt X F jAt j e jAt 6:9
with
  
 
jAtj 2A1  cos t , 6:10
2
   

At arccos sgn cos t : 6:11
2

The amplitude jA(t)j and phase (A(t)) are shown in Figure 6.4 and Figure 6.5,
respectively.
Note that (6.11) is simply a way to express that (A(t)) toggles between 0 and 180 ,
whereby e j(A(t)) toggles between 1 and 1, depending on the sign of the function
 6.3 Quasi-static X-parameter evaluation 155

0.20

0.15
|A(t)| (V)

0.10

0.05

0.00
0 500 1000 1500 2000 2500
time (ms)

Figure 6.4 Input signal envelope amplitude.

150
(A (t)) ()

100

50

0
0 500 1000 1500 2000 2500
time (s)

Figure 6.5 Input signal envelope phase.

cos(t/2). The amplitude and phase of the X-parameter function X (F)(.) of our
amplier example are shown in Figure 6.6 and Figure 6.7, respectively.
Note that Figure 6.6 corresponds to a classic AM-to-AM characteristic, whereas
Figure 6.7 corresponds to a classic AM-to-PM characteristic. The corresponding com-
pression characteristic is shown in Figure 6.8.
Equations (6.9)(6.11) are used to evaluate and analyze the DUT response. For the
amplier example considered, the real and imaginary parts of the two-tone response B(t)
are shown in Figure 6.9.
Note that the output signal B(t), regardless of the shape of the nonlinear function X(F)(.),
has exactly the same period as the function cos(t/2), namely a period equal to 4/.
The periodic output envelope B(t) can therefore be expanded into a complex Fourier
series, with harmonic frequencies equal to an integer times /2, as shown in (6.12):
156 Memory

0.4
output amplitude (V)

0.3

0.2

0.1

0.0
0.00 0.05 0.10 0.15 0.20 0.25
input amplitude (V)

Figure 6.6 Amplitude of X (F)(.) versus input amplitude (AM-to-AM).

3.0

2.5
output phase ()

2.0

1.5

1.0

0.5

0.0
0.00 0.05 0.10 0.15 0.20 0.25
input amplitude (V)

Figure 6.7 Phase of X (F)(.) versus input amplitude (AM-to-PM).

10

9
gain (dB)

30 25 20 15 10 5
input amplitude (V)

Figure 6.8 Compression characteristic.

 6.3 Quasi-static X-parameter evaluation 157

0.4

0.2
re & im output (V)

0.0

0.2

0.4
0 500 1000 1500 2000 2500
time (s)

Figure 6.9 Calculated output signal in the envelope domain.

X


B t Bk e jk 2 t : 6:12
k

The coefcients of this series, Bk, have some interesting properties. Based on the
symmetry conditions
 
2
B  t Bt 6:13

and
Bt Bt , 6:14
which can readily be veried using (6.9)(6.11), the resulting series Bk will always be a
symmetric odd harmonic series. In other words, Bk 0 if k is an even integer, and
Bk Bk : 6:15
These odd components Bk, having harmonic indices 1, 3, 5,. . . or 1, 3, 5,. . .,
correspond to the upper and lower intermodulation products, respectively, generated by
the two-tone signal. Note that the spacing between the harmonic frequencies, as one
would expect for a two-tone signal, equals . The lower third-order intermodulation
(IM3) product corresponds to k 3, the lower fth-order intermodulation (IM5)
corresponds to k 5, the upper IM3 corresponds to k 3, the upper IM5
corresponds to k 5, etc.
By making use of the symmetry conditions (6.13) and (6.14) and the denition of the
complex Fourier series, the following simple relationship between the one-tone X-parameter
function X (F)(.) and the quasi-statically generated intermodulation products Bk results:

=2
2  
Bk X F 2A1 cos cos kd, 6:16

0

only valid, of course, for k equal to an odd integer.

158 Memory

tone output power (dBm)

4

10

12

22 20 18 16 14 12 10 8
tone input power (dBm)

Figure 6.10 Measured and calculated tone output power.

20
IM3 & IM5 (dBm)

30

40

50

60
22 20 18 16 14 12 10 8
tone input power (dBm)

Figure 6.11 Measured (symbols) and calculated IM3 (solid line) and IM5 (dashed line).

Experimental validation of the described quasi-static approach is provided by com-

paring amplitude values of the intermodulation products Bk as calculated by using
(6.16) with actual measured amplitude values, for several different input power levels.
Figure 6.10 shows the calculated output power, per tone, as a function of the input
power, per tone (solid line), together with actual measured data acquired from the
example two-tone experiments (dots).
Figure 6.11 shows the calculated values of B3 and B5, which correspond to IM3 and
IM5, respectively, together with the measured data points. The measured and calculated
values correspond very well, thereby proving the validity of the quasi-static approach.
Equation (6.16) reveals another important characteristic of quasi-statically generated
intermodulation products, namely that Bk only depends on A1 and X (F)(.), but is
 6.3 Quasi-static X-parameter evaluation 159

60
spectral density (dBm /Hz)

70

80

90

100

110
30 20 10 0 10 20 30
frequency (MHz)

Figure 6.12 Input signal spectrum.

independent of . In other words, the intermodulation products are independent of the

spacing between the two tones of the input signal.

6.3.2 ACPR estimations using quasi-static approach

In a similar manner to the case of the sinusoidal amplitude modulation considered
above, it is possible to estimate other nonlinear gures of merit (FOMs), such as the
adjacent channel power ratio (ACPR) and others common to digital communications
circuits, from a simple static X-parameter model in an envelope analysis. For digitally
modulated signals, it is easiest to use built-in modulated sources available in the
simulator. The simulator simply evaluates the static X-parameter model at each sampled
time according to (6.6). This is illustrated using an actual device, namely a 25 W
GaN monolithic microwave integrated circuit (MMIC) power amplier (CREE
CMPA2560025F). The device is rst characterized by a static X-parameter model. This
is equivalent to measuring the AM-to-AM AM-to-PM characteristic of the device. Next,
a long-term-evolution (LTE) input signal with a 2.5 GHz carrier frequency is applied to
the device. The spectrum of the input signal is shown in Figure 6.12.
The output signal is predicted by applying the static X-parameter model to the
measured input envelope. Then the spectrum of the modeled output signal is compared
to the actual measured output spectrum. The result is shown in Figure 6.13. Note the
signicant spectral re-growth in both the measured and the modeled output spectra.
The relevant FOM is computed from the simulated output spectrum according to the
specic protocol appropriate to the modulation format. Examples of quasi-statically
estimated ACPR from X-parameters for a real amplier and independent experimental
validation are provided in Table 6.1.
Note that there is a good correspondence between the measurements and the values
derived from the static X-parameter model.
160 Memory

Table 6.1 ACPR estimated from static X-parameters versus independent measurements

Output power Lower sideband Upper sideband

(dBm) ACPR (dB) ACPR (dB)
X-parameter model 37.9 18.6 18.7
Measurements 38.0 19.0 19.1

30
spectral density (dBm /Hz)

40

50

60

70

80
30 20 10 0 10 20 30
frequency (MHz)

Figure 6.13 Output spectra: static X-parameter model (dotted line) and measurements (solid line).

6.3.3 Limitations of quasi-static approach

As explained in Section 6.3.1 the quasi-static evaluation of a static X-parameter model
will produce an intermodulation spectrum with identical levels for upper and lower
sidebands from a two-tone input signal. Moreover, the simulated levels are independent
of the modulation rate the frequency separation, f, between the two incident tones.
That is, the distortion spectrum shows no bandwidth dependence. In the envelope
domain, it can be shown that the time-varying output envelopes are symmetric with
respect to their peaks.
In the limit of slowly varying input envelopes, (6.6) reduces to evaluating a set of
independent steady-state mappings at different power levels. That is, (6.6) reduces to
(6.5), provided the latter was characterized over the full range of amplitudes covered
by the time-varying input signal of (6.6). Simulations for modulated signals become
exact as the modulation rate, or signal bandwidth (BW), approaches zero (the narrow-
band limit).
For the two-tone case, it is possible actually to measure true steady-state X-parameters
as functions of an LSOP that depends on both large input tones, using the approach
discussed in Chapter 5. Two-tone X-parameters provide exact intermodulation spectra, in
magnitude, phase, and their dependence on the frequency separation of the tones.
 6.4 Manifestations of memory 161

It should not be surprising that full information about the DUT response to two
steady-state tones cannot be obtained from measurements taken with an LSOP set by
only a single incident CW signal. It is clear there is more information in a set of two-
tone measurements than in a single-tone measurement.
At high modulation bandwidths, the actual time-dependent scattered waves are no
longer accurately computable from (6.6). That is, the quasi-static approximation of
going from (6.5) to (6.6) becomes less valid. Any FOMs, such as third-order intercept
(IP3) or ACPR, derived from the scattered waves simulated under the quasi-static
approximation will therefore not be in complete quantitative agreement with the actual
performance characteristics of the DUT.

6.3.4 Advantages of quasi-static X-parameters for digital modulation

Key nonlinear FOMs are usually scalars (numbers). For example, IP3, ACPR, and
(error vector magnitude) EVM are numbers, typically computed from measurements
made with a spectrum analyzer. The FOMs of particular components, such as individual
stages of a multi-stage amplier, cannot generally be used to infer the overall FOM for
the composite system. That is, just knowing the FOMs of individual parts is not enough
to design a nonlinear RF system for lowest ACPR.
X-parameters, on the other hand, enable at least a quantitative estimate of any nonlinear
FOM for the full system just from knowledge of the X-parameters of the constitutive
components. The X-parameters of the composite system follow from the nonlinear
algebraic composition of the component X-parameters according to the general treatment
in Chapter 2. The quasi-static approximation can be used easily in the envelope domain to
evaluate the overall system response to modulated signals of various kinds.
Improved estimates of DUT behavior in response to wide-band modulated signals
require a more careful treatment of dynamical behavior that begins in the following
section.

6.4 Manifestations of memory

Before dening memory and dealing with it precisely, the topic is introduced with a set
of frequently observed manifestations, or signatures, of memory. These can be con-
sidered evidence of underlying dynamical interactions that cause the DUT to behave
differently from quasi-static inferences. None of these manifestations of memory are
sufcient on their own to characterize the dynamics in a complete way, but rather they
are evidence of the existence of dynamics for which a general treatment is required.
Real ampliers exhibit behaviors more complicated than those predicted by the
techniques of Section 6.2 for the case of rapidly varying input envelopes. For
example, the intermodulation spectrum of a power amplier in response to a large-
amplitude two-tone excitation will show asymmetric upper and lower levels, with
levels depending strongly on bandwidth (frequency separation of the tones). These
162 Memory

-11

IM3 (dBm) -12

-13

-14

-15

-16
-2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5
two-tone offset frequency (MHz)

Figure 6.14 Frequency-dependent and asymmetric intermodulation spectrum.

2.5
2.0
|B(t)| (Vpeak)

1.5
1.0
0.5
0.0
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
|A(t)| (Vpeak)

Figure 6.15 Multi-valued instantaneous AM-to-AM amplier characteristics.

effects are induced by nonlinear behavior and are especially visible under high levels
of compression. Some simulated results are shown in Figure 6.14.
For this simulation, a constant input amplitude of 6 dBm per tone is chosen, whereby
the two-tone spacing is swept from nearly 0 Hz up to 4 MHz. The carrier frequency for
the experiment is xed at 2.6 GHz. The device is a radio-frequency integrated circuit
(RFIC) amplier, whereby the nonlinear memory effects are caused by the presence of a
6 nH inductor in series with the DC power supply. Note that the x-axis scale indicates
the offset frequency of each of the two tones, thereby corresponding to half the
tone spacing. The power at the output per tone is about 9 dBm. The magnitude of
intermodulation distortion depends signicantly on the separation frequency of the two
excitation tones. Also evident is that the lower frequency sideband, at the maximum
separation frequency, is about 1 dBm larger than the upper sideband. Another mani-
festation of memory is that the instantaneous AM-to-AM characteristic exhibits multi-
valuedness. These effects are illustrated by performing a simulation whereby one uses
bandwidth-limited noise as the input signal. The instantaneous output envelope
amplitude jB(t)j versus the instantaneous input envelope amplitude jA(t)j is depicted
in Figure 6.15.
It is evident that the output power is not uniquely determined by the input power at
the same time instant. Rather, the output depends on the history of the input signal.
 6.5 Causes of memory 163

6.5 Causes of memory

Primary causes of memory effects include dynamic self-heating, bias-modulation

effects, charge-storage (junction capacitances) and transit-time effects in active devices
(transistors and diodes), and trapping effects in some semiconductor devices (e.g. drain
lag in GaN FETs).

6.5.1 Self-heating
Self-heating involves the self-consistent coupling of electrical power dissipation in the
device (transistor or amplier) to a temperature change that in turn modulates the gain
and therefore changes the output waveforms. Typically, self-heating in a transistor is
modeled by two coupled equivalent circuits, one electrical and one thermal. The
electrical circuit has current sources and capacitors, the values of which depend on
voltages in the electric circuit and also the dynamic junction temperature, Tj, of the
thermal circuit. The thermal circuit calculates the junction temperature, Tj, as a response
to electrical power dissipation computed in terms of the voltages and currents of
the electrical circuit, and thermal resistance and thermal capacitance parameters in the
thermal circuit [13].
A hallmark of dynamic-thermal effects is that they vary at timescales typically many
orders of magnitude slower than those of the RF signals. As the modulation rate
becomes comparable to thermal-relaxation timescales, signicant dynamical effects
from the mutual electrical and thermal coupling become observable.

6.5.2 Bias modulation

Bias modulation occurs when a nonlinear device produces energy at baseband frequen-
cies by intermodulation of higher-frequency RF components. The low-frequency signal
components thus produced can pass through the bias lines, modulating the voltages and
currents, causing them to vary in time, which in turn modulates the RF-scattering
properties of the DUT. This is depicted in Figure 6.16 for the case of a pulsed RF
excitation.
The gure illustrates how an inductor that is present in series with the DC power
supply will cause memory effects. Consider an RFIC amplier whereby one suddenly
applies a signicant amount of RF power at time zero, with no RF being present before
time zero. As such, the envelope representation of the input signal A1(t) looks like a
perfect step. The amplitude after the step will be noted as A0. So what is the envelope
representation of the corresponding output signal B2(t)? As will be explained in the
following, the B2(t) will typically look like a step with an exponential relaxation part on
top, as shown in Figure 6.16. This can be explained as follows. In general, B2(t) will be
a function, F(.), of both A1(t) and the bias voltage. This is expressed as follows:

B2 t F A1 t , V bias t : 6:17
164 Memory

Figure 6.16 Illustration of biasing effects.

The bias voltage Vbias(t) depends dynamically on the bias current Ibias(t), expressed by
dI bias t
V bias t V DC  L : 6:18
dt
Finally, the bias current Ibias(t) is a function I(.) of both A1(t) and Vbias(t):
 
I bias t I A1 t , V bias t : 6:19
So what happens if one applies a step RF input? Before the step is applied, the bias
current has reached a steady-state value, I0, and the bias voltage equals VDC. The value
of I0 is given by
I 0 I 0, V DC : 6:20
Right after the step, at an innitesimal time , the bias current Ibias() will still be equal
to I0 as the bias current through the inductor can only vary at a nite rate, as shown by
(6.18). The bias voltage right after the step, Vbias(), will be given by solving (6.21) as
an implicit equation:
 
I 0 I A0 , V bias : 6:21
For most circuits, Vbias() will be smaller than VDC. The output amplitude right after the
step, B2(), is calculated using (6.17):
 
B2 F A0 , V bias : 6:22
Immediately after the step, the bias current Ibias(t) starts to increase at a rate given
by (6.23):
dI bias 1  
V DC  V bias : 6:23
dt L
As time goes by, Ibias(t) and Vbias(t) will gradually, in a relaxation-like manner, approach
their steady-state values corresponding to a CW excitation signal with amplitude A0.
The CW steady-state solutions are given by
 6.5 Causes of memory 165

lim V bias t V DC , 6:24

t!

lim I bias t I A0 , V DC 6:25

t!

and
lim B2 t F A0 , V DC : 6:26
t!

The shapes of the Vbias(t), Ibias(t), and B2(t) are qualitatively depicted in Figure 6.16.
A natural question then arises as to how this relates to the validity of the quasi-static
X-parameter approach, as explained in Section 6.3. Suppose that such an inductor is
present under what circumstances will one be able to use the quasi-static X-parameter
approach? The answer is present in (6.23), which shows that the derivative of the bias
current is proportional to the reciprocal of the inductance L. In other words, the smaller
the value of L, the faster all transitions occur. If any transitions happen faster than any
noticeable variations of the input envelope amplitude, the resulting behavior will still
look like a quasi-static map. The conclusion is that the quasi-static approach will still
work for small enough L or for slow enough modulation.
But if that is not the case, what are our options? If one has direct access to the voltage
bias node shown in Figure 6.16, one can still use CW X-parameters. The procedure is to
measure directly the functions F(.) and I(.), whereby one not only sweeps the RF
amplitude, A, but also Vbias, and whereby one measures the dependences of the RF output,
B(t), and the bias current, Ibias(t), not only on the RF amplitude, but also on the bias
voltage. Note that the relationships (6.17) and (6.19) remain valid under a CW RF
excitation, so the functions F(.) and I(.) can be measured under CW operating conditions.
In practice, this can be done, for example, by using an Agilent NVNA. The Agilent NVNA
controls the biasing instruments and can be made to sweep the applied DC bias over a
range of values large enough to cover any set of conditions expected to be encountered in
the actual application. The DC currents (voltages), as well as their sensitivities to power
and load, are simultaneously measured during the X-parameter extraction. They are
represented by the X(I) and X(Y) functions that depend on power, bias, and RF frequency.
The extracted model will have an external port for the bias. The bias-modulation
effects can be simulated by adding a circuit model of the bias network between this pin
and the battery. This network can be a simple inductor, an L-R network, or more
complicated L-R-C networks. Now the model can be simulated with a modulated signal.
The low-frequency components generated by the device will cause the node voltage at
the X-parameter bias pin to vary in time. This is another application of the quasi-static
approximation, where the previously xed voltages were once just parameters specify-
ing the mapping from incident As to scattered Bs, but have now become time-varying
variables that modulate the mapping. An example of such a circuit, and the results, are
given in Figure 6.17.
The circuit that is used to illustrate the principle is an RFIC amplier that needs two
bias voltages. A rst bias supply enters through the pin marked v3ic and a second bias
supply enters the RFIC through the RF output (as such, a DC block is needed). The rst
bias does not draw any current and is not contributing to any memory effects. The
166 Memory

DC_Feed
+ Vdc = 6.0 V .o +
Vdc = 3 V
- . -
L = 2.0e-6 H

v3ic

v1ic v2ic
DUT

+
DC_Block DC_Block

Z=50 Ohm
R=50 Ohm
- Freq[1]=Freq1
Freq[2]=Freq2

Figure 6.17 Schematic used for simulating bias-line-induced memory.

10

12
IM2 (dBm)

14

16

18
7 5 3 1 1 3 5 7
output tone frequency offset (MHz)

Figure 6.18 IM3: comparison between circuit schematic and X-parameter ( circuit schematic,
x X-parameter model).

second bias draws all of the supply current. For the purpose of demonstration, a series
2 H inductor is connected between the bias pin and the 3 V DC supply. As a rst step,
through simulation, an X-parameter model is extracted that contains the dependency of
the bias current and the RF output on the bias voltage (pin v2ic). Next, a set of two-
tone simulations is performed, whereby the simulation is performed on the original
circuit schematic as well as on a schematic in which the RFIC has been replaced by the
extracted X-parameter model. The results for IM3 and the output power per tone are
shown in Figure 6.18 and Figure 6.19, respectively. It is evident that the method of
 6.6 Importance of memory 167

output power per tone (dBm)

9.1
9.0
8.9
8.8
8.7
8.6
8.5
8.4
2.5 1.5 0.5 0.5 1.5 2.5
output tone frequency offset (MHz)

Figure 6.19 Two-tone output power: comparison between circuit schematic and X-parameter
( circuit schematic, x X-parameter model).

using bias-dependent X-parameters can produce the tone-separation dependence as well

as the (slight) asymmetry of the magnitudes of the intermodulation spectral products.

6.6 Importance of memory

Memory effects make it much more difcult to quantify nonlinearity, since the standard
gures of merit, such as IM3 and ACPR, depend sensitively on the modulation
bandwidth and format. In particular, these FOMs cannot be a-priori calculated just from
knowledge of static X-parameters (e.g. AM-to-AM and AM-to-PM characteristics). It is
therefore more difcult to design circuits and systems trading off efciency for linearity
with nonlinear components demonstrating signicant memory effects.
Efciency is of such paramount importance that the power amplier market is ready
to sacrice open-loop linearity, preferring to linearize these nonlinear devices externally,
primarily with digital pre-distortion (DPD) and other techniques. Pre-distortion in the
presence of strong memory effects is one of the major problems facing the power
amplier industry today. Control, elimination, or (ultimately) exploitation of memory
effects would enable a dramatic improvement in performance from existing technolo-
gies. In all these cases, a good model of memory is a key component of a solution.

6.6.1 Modulation-induced baseband memory and carrier memory

Two distinct kinds of memory effects are signicant for power ampliers. They will be
referred to as carrier-memory effects and modulation-induced baseband-memory effects.
Carrier memory is caused by dynamical effects that physically occur at the carrier
frequency, typically at RF and microwave frequencies. Group-velocity effects and gain
variation with frequency fall into this category. Simply sweeping the frequency of a
single RF tone across the band of interest for an amplier and noting the frequency
response indicates RF carrier memory. Such effects have to be taken into account when
one wants to construct a model that remains accurate across a bandwidth that is
signicant when compared to the RF carrier frequency. The addition of RF carrier
168 Memory

memory to X-parameters is ongoing; it is not described here [14]. Note that this kind of
memory does not require any nonlinear effects. It shows up for a linear lter as well as
for an amplier under small-signal operating conditions if there is a signicant variation
of the scattering parameters across the modulation bandwidth. For this reason, this kind
of memory is often also called linear memory. It is sometimes called short-term
memory, referring to the fact that the timescale of this kind of memory effect is
typically in the nanosecond range, comparable to the period of the RF carrier frequency.
When a high-Q resonating lter structure is involved, however, carrier memory
can introduce long time transients. Under such conditions, the terminology short-
term memory is misleading. Therefore, the term short-term memory is not used in
this book.
Modulation-induced baseband memory is of a different nature. It is caused by
dynamical effects that physically happen in the baseband, at timescales many orders
of magnitude slower than the RF stimuli, such as microseconds, milliseconds, or even
seconds. Examples include the time variation of the bias settings, dynamic self-heating,
and trapping effects. Although the physical effects happen relatively slowly, their
inuence shows up at the high-frequency carrier through the process of modulation.
Dynamic temperature variations and dynamic bias settings may slowly modulate the
compression and AM-to-PM characteristics of the amplier. Such modulation-induced
baseband-memory effects can become signicant even at relatively small modulation
bandwidths, for example a few kilohertz or megahertz. A general way to treat such
memory effects in the framework of X-parameters is described in the following
sections. Modulation-induced baseband memory is sometimes called long-term
memory, complementary to the term short-term memory. As previously explained,
this terminology is misleading as some carrier-memory effects may actually correspond
to transitions that have longer settling times than some modulation-induced baseband-
memory effects.

6.6.2 Dynamic X-parameters

6.6.2.1 Overview
The dynamic X-parameter method [3] is an original approach taken to extend
X-parameters to include modulation-induced baseband-memory effects. Starting from
a few basic assumptions, which can be validated by independent experiments, a full
theory is developed by rigorous derivation. The approach results in a new term
subsuming all long-term memory effects that simply adds to the existing static
X-parameter expressions developed previously. The core of the new term is a memory
kernel function that is uniquely identied from a new type of complex envelope
transient measurement that can be measured on an NVNA or vector signal analyzer
(VSA). The approach is extremely powerful in that within the original assumptions the
resulting model is valid for any type of modulated signal, independent of format, power
level, or details of the probability density function (PDF) dening the statistics of the
signal. That is, the resulting behavioral model is highly transportable [5], [8]. The
dynamic X-parameter method generically abstracts the intrinsic dynamics and
 6.6 Importance of memory 169

nonlinearities of the DUT. It can therefore be directly used to predict accurately the
DUT response to a wide variety of modulated signals, including those with high
peak-to-average ratios (PARs) and wide modulation bandwidths, without the need to
re-extract the model for different formats or average power conditions. Independent
experimental validation of the model, applied to a real power amplier subjected to very
different types of modulated signals, conrms the theory. The model and its identica-
tion has a compelling and intuitive conceptual interpretation as well.

6.6.2.2 Motivation
The hysteretic behavior of the instantaneous gain depicted in Figure 6.15 suggests the
starting point for the approach. The fact that the output RF amplitude is multiple valued
means there must be additional independent variables that need be specied besides the
RF input amplitude. Moreover, the dependence on these auxiliary variables must be
identied by the modeling process.
For simplicity, only the case of a unilateral, perfectly matched, inputoutput map
around a single carrier is considered here. The approach can be easily extended to
multiple ports with mismatch, harmonics, and intermods.

6.6.2.3 Assumptions
The general time-dependent complex envelope of the scattered wave, B(t), is assumed to
be a nonlinear function of the input envelope, A(t), and an arbitrary number of unknown
real-valued dynamical variables, hi(t), for i 1,2,. . .. These hidden variables repre-
sent underlying physical processes such as time-varying temperature, bias voltages and
currents, or trapping states in semiconductors. The unknown constitutive relation
mapping these time-varying inputs to the output is given by
  
Bt X At, h1 t, . . . , hN t e jAt : 6:27

The rst assumption is that only the magnitude of A(t) enters the nonlinear function in
(6.27), with the phase an overall multiplicative factor. This was demonstrated to be a
good approximation in [7]. Recall that it is rigorously true that the phase is a purely
multiplicative factor in the static one-tone case, as derived in Chapter 2.
The second assumption is that the values of the hidden variables never depart greatly
from their values under steady-state conditions. That is, hi(t), dened in (6.28), is small:
 
hi t hi t  si At : 6:28

Here, si(jA(t)j is the quasi-statically mapped value of hi(t), namely the value that hi(t)
would have under steady-state conditions corresponding to an input amplitude magni-
tude of value jA(t)j (see the discussion of Section 6.3). Small hi(t) means (6.27) can be
linearized around the steady-state values of the hidden variables, simplifying the
description and subsequent analysis considerably.
Substituting for hi(t) using (6.28) in (6.27), the function X is linearized around its
value, with all hidden variables taking their (quasi-) static values, to rst order in hi(t),
to obtain the expression in (6.29):
170 Memory

 
     
Bt X At ,s1 At  h1 t ,...,sN At hN t e jAt
0 1

B           XN
X  C
B C
 BX At ,s1 At  ,...,sN At     
hi t Ce jAt
@ i1
hi       
 
A
At ,s1 At ,...,sN At

!
  X
N  
 X F At  Di At  hi t e jAt :
i1

6:29
The rst term on the right-hand side of (6.29) is identied as the familiar static
X-parameter function, X (F), because, under steady-state conditions, hi(t) 0 for each
hidden variable, and this corresponds to the quasi-static case.
Specic assumptions about the explicit form of the hidden variables are now intro-
duced in order to evaluate the new terms that appear in addition to the static
X-parameter function in (6.29).
Each hidden variable is expressed according to (6.30):

hi t Pi jAt  jk i d: 6:30
0

The function Pi(.) can be interpreted as a source term, describing how the devices
nonlinear dependence on the input amplitude excites the hidden variable. The function
ki(.) describes the dynamics of the process, namely how the source term at some time in
the past affects the present value of the hidden variable. As such, ki(t) represents an
impulse response of a linear lter governing the relaxation of the particular phenomenon
represented by the ith hidden variable. Equation (6.30) is a convolution representing the
causal response of the hidden variable to the source term generated by the input-signal
envelope at all points in the past, according to the impulse response governing that
dynamical process.
The fact that Pi(.) depends on the magnitude of the complex amplitude describes
the generic phenomena of conversion between the rst zone signal and the baseband
auxiliary variables. This type of feedback architecture has been considered in [10]
and [11].
The term hi(t) appearing in (6.29) can now be evaluated using (6.28) and (6.30). For
a constant input value of A (a true CW signal), the value of the hidden variables can be
calculated from (6.30) to be

 
si A Pi jAj k i d: 6:31
0

The quasi-static value of the hidden variable is therefore given by

   
si jAtj Pi jAtj k i d: 6:32
0

From Equations (6.28), (6.30), and (6.32), (6.29) is re-expressed as (6.33), where the
order of summation and integration have been interchanged:
 6.6 Importance of memory 171

 !
 
X
N       
Bt  X F At    
Di At Pi At   Pi jAtj k i e jAt
0 i1
   
F 
     
X     
At G At , At  , d e jAt :
0
6:33
In (6.33), the convenient notation G(x,y,t) has been introduced, dening the memory
kernel as follows:
X
N  
Gx, y, t Di x Pi y  Pi x k i t: 6:34
i1

It follows immediately from (6.34) that, for all values of x and t, the memory kernel
vanishes when the rst two arguments are equal. That is,
Gx, x, t 0: 6:35
Condition (6.35) corresponds to the case that the present and past values of the
amplitude are the same, which is the condition dening the steady state. In steady state,
the memory contribution must vanish since the entire contribution is already accounted
for by the X (F) term in (6.33).

6.6.2.4 Discussion
The memory kernel, G(x,y,t), depends nonlinearly on the instantaneous amplitude of the
input signal, the past value of the input-signal envelope, and explicitly on the time at
which the past value of the input occurred.
The remarkable fact, to be demonstrated below, is that the memory kernel can be
identied uniquely from a specic set of nonlinear measurements. Once the memory
kernel is measured, it can be substituted into (6.33), and the contributions to the
scattered waves from all hidden variables can be accounted for, no matter how many
variables are assumed to be present in (6.27), in particular without the need to identify
the functions Pi and ki separately for each variable.

6.6.3 Identification of the memory kernel: conceptual motivation

Under a one-tone steady-state excitation at microwave frequencies, physical variables
such as temperature assume specic constant values. That is, the temperature does not
vary in time with the frequency of the RF signal, but rather assumes a xed value based
on the power dissipation averaged over a much longer timescale. It can be said that
temperature has a low-pass dynamical behavior.3 Steady-state measurements at dif-
ferent power levels (different values of A) therefore generally correspond to different
values of DUT temperature. Care must be taken to vary the amplitude slowly enough so

3
The particular value of the temperature depends on the power-added efciency of the device at that power
level and the effective thermal resistance.
172 Memory

that sampled measurements at different power levels will reect the characteristics of
the system at steady state.4 Alternatively, the power can be stepped from one value to
the next, but sufcient time must be allocated after each transition, such that the system
reaches its steady-state value before making an RF measurement of the output. That is,
each step in amplitude causes a thermal transient, where the RF and thermal variables
are initially decoupled. Eventually the thermal transient dies away and the system
returns to the new steady state corresponding to the value to which the RF amplitude
was stepped.
The foregoing considerations motivate the procedure for identifying the memory
kernel. Each hidden variable can be made to depart from its steady-state value by
stepping the value of the input-signal amplitude. The step transition is related to the
source term, and the time-dependent relaxation is related to the linear lter term of
(6.30), respectively. In general, the response to such steps between steady-state values
depends on both the initial state and the nal state of the transition.

6.6.4 Step response of the memory kernel

A stepped-input complex-amplitude signal dened by (6.36) is now applied:

A1 , t  0,
At 6:36
A2 , t > 0:
Substituting (6.36) into (6.33), the step response, denoted by Bstep(A1, A2, t), is evalu-
ated for t > 0:
 t 
F
     
B A1 , A2 , t X
step
jA2 j G jA2 j, jA2 j, d G jA2 j, jA1 j, d e jA2 :
0 t
6:37
The second term on the right-hand side of (6.37) vanishes due to (6.35). The memory
kernel can be isolated from (6.37) by taking the time derivative of both sides. The result
is given by
d step
GA2 , A1 , t  B A1 , A2 , tejA2 : 6:38
dt
It is apparent from (6.38) that the memory kernel is completely identied from the set of
complex-amplitude step measurements from initial states A1 to nal states A2. The time
derivative converts the step response to an impulse response. Since G must be dened
for all values of its arguments, the required measurements must sample transitions from
all possible initial values of A to all possible nal values. That is, if the range of
amplitudes from A 0 to A Amax is sampled at N discrete values, there are N2
transitions that must be considered. For each transition, the time-dependent response, in
magnitude and phase, must be sampled for all times from immediately after the step to
times long enough for the DUT to have relaxed to its steady-state value.

4
This type of variation is known as adiabatic.
 6.6 Importance of memory 173

Figure 6.20 Step inputs and responses.

A2
A(t)
A1
t
B(t) FB
X 2,1 (A1) FB
X 2,1 (A2)
t
dB(t) G(A2,A1,t)
dt
t
dBLS(A1,A2,t)
=G(A2,A1,t)
dt

Figure 6.21 Identication of memory kernel from step responses.

The method is presented graphically in Figure 6.20 and Figure 6.21 for different steps
and corresponding responses.

6.6.5 Application to real amplifier

The methodology described in Section 6.6.4 was applied to a Mini-Circuits ZFL11AD
amplier [4]. The carrier frequency used was 1.75 GHz. Fifteen different values of
amplitudes from 0.01Vpeak to 0.2Vpeak were selected for the set of step envelope
measurements. Figure 6.22 depicts input steps from two different initial values of A1,
a small value (top left) and a larger value (bottom left), to all other levels. The right side
of the gure shows the magnitudes of the time-dependent responses corresponding to
transitions to each of the nal values of A2 from the low initial value (top right) and
from the higher initial value (bottom right).
174 Memory

| A(t) | | B(t) |
(Vpeak) 0.20 (Vpeak) 0.5

0.15 0.4

0.3
0.10 A2 swept
0.2
0.05
A1=0.01V peak 0.1
0.00
0.0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time(s) time(s)
(Vpeak) 0.20 (Vpeak) 0.5

0.4
0.15
A1=0.14V peak
0.3
0.10 A2 swept
0.2
0.05
0.1
0.00 0.0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time(s) time(s)

Figure 6.22 Step-response magnitudes from two initial amplitudes for Mini-Circuits
ZFL11AD amplier.

A1 = 29 dBm

Magnitude of B(t) Phase of B(t)

0.5
(dBm) (deg)
113
1.0 A1 = 3.6 dBm
114
1.5
115
2.0
A1 = 3.6 dBm 116
2.5
117
3.0 118
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time (ms) time (ms)
A1 = 29 dBm

Figure 6.23 Magnitude and phase responses from various initial states to a xed nal step
value of A2 10.6 dBm.

The magnitude response and phase response corresponding to transitions to a xed

nal value of A2 10.6 dBm from a variety of initial values are shown in Figure 6.23.
The steady-state X-parameter unilateral transfer function at the fundamental fre-
quency can be obtained from the long-term response of the step, or alternatively from
the static X-parameter measurements described in Chapter 2. For the Mini-Circuits
amplier of this example, the results are presented in Figure 6.24.
The memory kernel, G(A2, A1, t), is plotted versus A2 and A1 at a xed time (50 ns)
in Figure 6.25.
 6.6 Importance of memory 175

115
10
magnitude (dB)

9 116

f (deg)
8 117

7
118
6
30 25 20 15 10 5
30 25 20 15 10 5
A (dBm)
A (dBm)

Figure 6.24 Static X-parameter gain function for Mini-Circuits ZFL11AD amplier.

|G ( A2, A1, 50 ns)|

V
ms 0.2 0.20
0.1 0.15
0.0
0.10 A (V
0.05 1 peak)
0.10 0.05
0.15
A2(Vpeak) 0.20

Figure 6.25 Memory kernel at xed t 50 ns for Mini-Circuits ZFL11AD amplier.

The magnitude of the memory kernel as a function of time for a particular step
transition is shown in Figure 6.26.

6.6.6 Validation of memory model

The memory model thus obtained from the step envelope measurements is implemented
in the Agilent Advanced Design System (ADS) simulator.
The model is validated by comparing the actual measured DUT response to simula-
tions using the model stimulated with the same signals. That is, the actual measured
signals are brought into the simulator to excite the model of the DUT for a more
accurate comparison. Any difference between measured response and simulated
response is therefore due entirely to the model, and not due to slight differences between
the ideal and actual modulated stimuli. The modulated signals are applied using an
Agilent signal generator. The response measurements for validation are made on the
Agilent NVNA. The measured response data are taken back into ADS to make the
comparison with the simulated performance.
Envelope-domain validation results are shown in Figure 6.27, Figure 6.28, and
Figure 6.29. Figure 6.27 shows the amplitude of the input envelopes for two-tone
experiments at a frequency spacing of 19.2 kHz and for four different power levels. In
176 Memory

V 0.00
ms 0.05

0.10

0.15

0.20
0 10 20 30 40 50 60

V 0.00
ms
0.05

0.10

0.15

0 1 2 3 4 5
time (ms)

Figure 6.26 Memory kernel (V/s) versus time (s) for a step transition from low (0.01) to high
(0.2) input amplitude. Real part (solid line); imaginary part (dotted line). The lower plot is a
zoomed-in area of the upper plot.

0.20
input amplitude (Vpeak)

0.15

0.10

0.05

0.00
0 20 40 60 80 100 120 140
time (ms)

Figure 6.27 Dynamic X-parameter model validation: measured input envelope amplitudes for a
two-tone frequency difference of 19.2 kHz.

Figure 6.28, the measured and simulated output waveforms are shown. The output
envelopes corresponding to higher power levels are clearly not symmetric about their
peak values, despite the fact that the input amplitudes are symmetric. In Figure 6.29, the
amplitude of the error is shown between the measured and the modeled output envel-
opes. The simulated and measured output amplitudes agree very well. The memory
model is therefore able to go far beyond the quasi-static application of the static
X-parameter model discussed earlier. The improvement in accuracy is purely attribut-
able to the second term in (6.33), involving the memory kernel.
 6.6 Importance of memory 177

0.45
output amplitude (Vpeak) 0.40

0.30

0.20

0.10

0.00
0 20 40 60 80 100 120 140
time (ms)

Figure 6.28 Measured and modeled output envelope amplitudes (measurement: solid line,
model: dots).

0.010

0.008
model error (Vpeak)

0.006

0.004

0.002

0.000
0 20 40 60 80 100 120 140
time (ms)

Figure 6.29 Amplitude of difference between measured and modeled output envelopes versus time.

Figure 6.30 shows the measured amplitudes of the output envelopes versus the
instantaneous input amplitude; this is the so-called dynamic compression characteris-
tic. For the higher input-power levels, one can clearly see the looping of the character-
istics. Figure 6.31 shows the differences between the model and the measurements
(versus instantaneous input amplitude). One concludes that the dynamic X-parameter
model accurately models the looping.
The memory model can also predict the detailed dependence of DUT gain and
intermodulation distortion characteristics on modulation bandwidth. This is demon-
strated in Figures 6.326.34. In Figure 6.32, the measured and simulated gains of the
lower and upper signals are compared as a function of tone spacing over a wide range of
frequencies. The sharp resonances and asymmetric values are well predicted.
In Figures 6.33 and 6.34, a comparison is presented between the measured and
simulated frequency dependences of the two third-order intermodulation sidebands.
178 Memory

0.45

output amplitude (Vpeak)

0.40

0.30

0.20

0.10

0.00
0.00 0.05 0.10 0.15 0.20
input amplitude (Vpeak)

Figure 6.30 Measured dynamic compression characteristics at four different power levels.

0.009
0.008
error amplitude (Vpeak)

0.006

0.004

0.002

0.000
0.00 0.05 0.10 0.15 0.20
input amplitude (Vpeak)

Figure 6.31 Amplitude of difference between measured and modeled output envelopes versus
instantaneous input amplitude.

Gain (lower) Gain (upper)

(dB) 8.2 (dB) 8.2

8.0 8.0
7.8 7.8
7.6 7.6
7.4 7.4
7.2 7.2
7.0 7.0
6.8 6.8
3 4 5 6 7 3 4 5 6 7
tone spacing (Hz, log10 scale) tone spacing (Hz, log10 scale)

Figure 6.32 Gain versus tone spacing: measured and simulated with dynamic X-parameters.
 6.6 Importance of memory 179

IM3 (lower) IM3 (upper)

IM3 20 IM3 20
(dBm) 25 (dBm) 25
30 30
35 35
40 40
45 45 sweet spot
50 50
55 55
3 4 5 6 7 3 4 5 6 7
tone spacing (Hz, log10 scale) tone spacing (Hz, log10 scale)

Figure 6.33 Magnitude of IM3 versus tone spacing: measured and simulated with dynamic
X-parameters.

IM3 (lower) IM3 (upper)

Phase Phase
(deg) 250 (deg) 200

200 150

150 100

100 50
3 4 5 6 7 3 4 5 6 7
tone spacing (Hz, log10 scale) tone spacing (Hz, log10 scale)

Figure 6.34 Phase of IM3 versus tone spacing: measured and simulated with dynamic
X-parameters.

The model is able to predict very accurately both the magnitudes and the phases of the
intermodulation products, including the sharp resonance.
The IMD sweet spot an operating condition where the device behaves more
linearly than at other nearby conditions, is clearly identied by the memory model.
Another demonstration of this is given in Figure 6.35. The 60 kHz tone spacing
corresponds to a peak in the gain and a minimum in the power of the third-order
distortion product. The instantaneous gain, while still showing considerable hysteresis,
is evidently more linear than the characteristics at 120 kHz tone spacing that compresses
more with incident power.
Another stringent test of the memory model is how well it can predict the measured
DUT characteristics in response to a wide-band digitally modulated signal. The same
dynamic X-parameter model obtained for the Mini-Circuits amplier was used to predict
the modulated output envelope and spectral re-growth in response to a wide-band code-
division multiple-access (WCDMA) signal. Comparisons of the dynamic X-parameter
model simulation with measurements are shown in Figures 6.36 and 6.37. Also
compared are the predictions of the static (CW) X-parameter model. The CDMA signal
180 Memory

(a) (b)

Figure 6.35 Measured (a) and modeled (b) dynamic compression characteristics (120 kHz tone
spacing: solid line; 60 kHz tone spacing: dashed line).

Figure 6.36 Validation of dynamic X-parameter and quasi-static X-parameter model simulations
for CDMA input signal. Waveforms: solid line, measured; dotted line, input signal; x, static
X-parameters; , dynamic X-parameters.

10

20
spectrum (dBm)

30

40

50

60

70
7.5 6.0 4.5 3.0 1.5 0.0 1.5 3.0 4.5 6.0 7.5
frequency (MHz)

Figure 6.37 Validation of dynamic X-parameter and quasi-static X-parameter model simulations
for CDMA input signal. Spectral re-growth: solid line, measured; , dynamic X-parameters,
dotted line, static X-parameters.
 6.6 Importance of memory 181

Figure 6.38 Interpretation of dynamic X-parameters in terms of superposition of memory effects.

was generated and uploaded to an Agilent ESG, and the waveforms were measured on
the NVNA in envelope mode.

6.6.6.1 Discussion
It should be noted that both the two-tone signal (sinusoidally modulated carrier) and the
WCDMA signals are completely different types of signals compared to the complex
step excitations used to derive the memory kernel. Nevertheless, the dynamic
X-parameter model is able to reproduce very accurately the actual characteristics of
the measured DUT responses to each signal. In the case of the two-tone signal, the
model is able to resolve the narrow resonance in gain and distortion with frequency
separation of the tones. For the digitally modulated signal, the model does an excellent
job describing the DUT output envelope time-dependent waveforms in the envelope
domain and spectral re-growth in the frequency domain. In both cases, the simulations
with the memory model are in much better agreement with the detailed measurements
than the predictions of the CW X-parameter model evaluated in the quasi-static
approximation. This is a powerful validation of the claim of transportability of the
memory model.

6.6.7 Interpretation of dynamic X-parameters

An interesting and useful interpretation of the dynamic X-parameter theory can be made
in terms of a principle of superposition of nonlinear impulse responses. It has been
demonstrated that step responses or the impulse response given their time derivative
generate a departure from the steady state that eventually relaxes. The linearity assump-
tion around the steady state is equivalent to the notion that the departure from steady
state induced by a sequence of steps (impulses) is the superposition of the departures
from steady state induced by each impulse separately. This is graphically illustrated in
Figure 6.38. It is possible to check the applicability of dynamic X-parameter theory by
seeing if the DUTs response to multiple pulses is indeed the superposition of the
responses to individual pulses.
A key difference between these sets of impulses and other similar looking theories,
such as [6] or [11], is that these impulses depend nonlinearly on both the initial and nal
182 Memory

Existing models Dynamic X-parameters

B(t) B(t)

always asymmetry
symmetric is possible
B(t)
B(t)

Figure 6.39 Complementary step responses are non-physical consequences of some impulse
response envelope models, but not dynamic X-parameters.

values for the steps. The form of the resulting expression for B(t), given by (6.33), is
much more general. In particular, the turn-on and turn-off transients in a real power
amplier are completely different from each other. An actual turn-on transient will
occur rapidly but not settle into its steady-state value for, say, microseconds until the
device temperature reaches its steady-state value. When turned off, however, the RF
amplitude becomes zero on timescales shorter than the RF frequency (limited only by
transit time and stored charge in transistors or group delay considerations in ampliers).
These characteristics are accurately predicted by dynamic X-parameters. Other
approaches, such as the nonlinear complex impulse response model (NCIRM), and
more classical models, including the parallel Hammerstein model, can be shown to
produce symmetric turn-on and turn-off transients. A schematic representation of this is
shown in Figure 6.39. A mathematical proof of this statement for the parallel Hammer-
stein model is presented in Appendix C. Note that the calculation for an NCIRM model
can be performed in a similar way.

6.6.8 Wide-band X-parameters (XWB)

6.6.8.1 Motivation
In the limit of narrow-band modulation, the system becomes arbitrarily close to its
steady-state value and the contribution from the memory kernel vanishes. The device
can be described in the narrow-band limit by the static (CW) X-parameters of Chapter 2.
In many practical cases, the opposite limit, namely wide-band modulation, can also
result in a simplication of the general dynamic X-parameter formalism [4].
The indication that this is possible is given by the following observations. A two-tone
stimulus at various modulation rates (offset frequencies) is applied to the Mini-Circuits
amplier in Figure 6.40.
At 1.2 kHz offset (Figure 6.40(a)), the instantaneous AM-to-AM characteristic is
essentially static. There is no hysteresis, which was a key signature of memory effects.
At 30 kHz (Figure 6.40(b)), things are quite different. There is substantial hysteresis,
indicative of signicant dynamic memory effects. At a still higher frequency offset of
8.75 MHz (Figure 6.40(c)), the behavior qualitatively changes again. The looping is
nearly gone and the device appears approximately static. That is, the AM-to-AM
 6.6 Importance of memory 183

Figure 6.40 Nonlinear behavior as a function of signal bandwidth for Mini-Circuits

ZFL-11AD amplier.

characteristic is essentially single valued again. However, there is still a substantial

difference between Figures 6.40(a) and (c). At large frequency offset, the effective static
characteristics are quite different from the actual static characteristics. In particular,
there is effectively less compression at high offset frequencies so the system appears
more linear than could be inferred from the actual static characteristics.
This can be understood in terms of timescales associated with hidden variables. At
very low modulation rates, the hidden variables track the RF input-amplitude variation.
At intermediate modulation rates, resonances in the dynamics of the hidden variables
may be excited. The hidden variables depart from their quasi-static values, but their
values vary in time, causing the hysteresis observed in Figure 6.40(b). Finally, at very
high modulation rates, the hidden variables can no longer respond to the rapidly varying
input envelope signal. Thermal variables have a low-pass characteristic as described in
Section 6.6.3. Here the hidden variables assume constant values given by a suitable
average over the amplitude variation.
An example of these considerations is furnished by a self-heating transistor. Static
AM-to-AM and AM-to-PM characteristics are traced out by sampled measurements at
different power levels. Measurements at each power level are made after waiting for the
DUT to equilibrate thermally. The actual DUT junction temperature5 is different at
every different CW power level. For class B and class C ampliers, the device typically
gets hotter as the RF power increases.6 The gain of the amplier typically decreases
with increasing temperature. If, instead, one ramped up the input power at a rate faster
than the typical thermal response, one would get a higher value for the gain at the same
instantaneous power, since the temperature would be lower than at the steady-state
condition. One therefore expects more compression from static AM-to-AM and AM-to-
PM measurements than from signals characterized by rapidly varying instantaneous
power levels. In particular, for a two-tone input signal, with tones separated more
widely than the inverse thermal relaxation time, the instantaneous AM-to-AM and
AM-to-PM characteristics are essentially iso-thermal. The effective temperature in this
case is somewhere between the low and high power values for the static case. From

5
Temperature is dened everywhere in space and time. For simplicity, and according to standard practice, it is
approximated by a single value that may vary in time.
6
This is not true for class A ampliers, which become cooler with increased applied power.
184 Memory

these simple considerations, the qualitative characteristics of Figure 6.40(a) and (c) can
be explained. In particular, the wide-band DUT response is more linear than can be
inferred from static characteristics.
Starting from the dynamic X-parameter formalism, one can derive the following
simplication appropriate in the wide-band modulation limit. If the amplitude values in
the past were always varying rapidly, their effects can only be manifested in the
scattered waves according to constant time-averaged values. That means that the
memory kernel appearing in (6.33) can be replaced by its time-averaged value, as
indicated in (6.39):
0 1

   
G jAtj, jAt  j, d  @ G jAtj, a, d Aada: 6:39
0
0 0

The average can be calculated by integrating the possible amplitudes over their prob-
ability distribution, (a), as indicated by the last equality of (6.39).7 The advantage of
the latter expression is that it is valid also for stochastic information signals that are
dened only in terms of their PDFs.8 Further discussion of this point is given in
Appendix D.
The inner integral in (6.39) is evaluated by substituting for G the time derivative of
the envelope response to the step excitations using (6.38), where the fact is used that, at
innite time, the step response from any initial condition to A(t) is precisely the static
X-parameter steady-state function X(F)(jA(t)j):

  d step  
G jAtj, a, d  B a, jAtj, d
0 d
0     6:40
Bstep a, jAtj, Bstep a, jAtj, 0
   
X F jAtj Bstep a, jAtj, 0 :
Substituting (6.40) into (6.39), and adding the result back into (6.33), the nal expres-
sion for wide-band X-parameters is obtained as follows:
0 1
 
  
Bt  @ X W B jAt j, a adaAexp j At , 6:41
0

whereby the bivariate model kernel XWB(.) is dened as follows:

X W B x, y Bstep y, x, 0: 6:42
The above equation implies that only the output value at time zero, immediately after a
step, determines the behavior of the system under wide-band modulated excitation
signals. A pictorial representation is given in Figure 6.41.

7
A lower-case a is used here to avoid confusion with the instantaneous amplitude denoted by A.
8
This is justied by the ergodic principle, namely that the time average is equal to the statistical average.
 6.6 Importance of memory 185

Figure 6.41 Wide-band X-parameters can be identied from the initial step response.

Bmeas (t)
0.4
Amplitude (Vpeak)

0.3

0.2 BCW (t) BWB (t)

0.1

0.0
0 100 200 300 400 500
Time (ns)

Figure 6.42 Validation of wide-band X-parameters for a two-tone signal.

6.6.8.2 Results for wide-band X-parameters

Figure 6.42 and Figure 6.43 demonstrate the results for wide-band X-parameters on the
Mini-Circuits amplier. A two-tone experiment is performed with a tone spacing of
3.75 MHz. Wide-band X-parameters, denoted with the WB subscript, are compared to
the NVNA measured data and the static X-parameters, which are here denoted with the
CW subscript (for continuous wave). Figure 6.42 shows the fundamental output time-
varying envelope of the amplier produced in response to a two-tone input signal.
Figure 6.43 shows the gain compression and magnitude of the third-order intermodula-
tion products versus input tone power.
Tables 6.2 to 6.4 contain the measured and modeled two-tone gain, IM3 products,
and IM5 products for the four input power levels.
Except for the lowest power level IM5 product, which is near the noise oor of
the measurements, the wide-band X-parameter model predicts gain, IM3, and IM5
products with an error smaller than 1 dB; the errors of the static model are signicantly
higher.
186 Memory

Table 6.2 Comparison of measured and modeled gain

Input power (dBm per tone) 20 15 12 10

Measured gain (dB) 10.41 9.84 8.85 7.97
WB X-parameter gain (dB) 10.39 9.71 8.60 7.65
CW model gain (dB) 10.35 9.52 8.19 7.07

Table 6.3 Comparison of measured and modeled IM3

Input power (dBm per tone) 20 15 12 10

Measured IM3 (dBm) 53.0 35.5 28.9 28.0
WB X-parameter IM3 (dBm) 52.4 35.1 28.5 27.2
CW model IM3 (dBm) 48.0 31.0 23.1 20.1

Table 6.4 Comparison of measured and modeled IM5

Input power (dBm per tone) 20 15 12 10

Measured IM5 (dBm) 82.0 68.4 42.8 36.1
WB X-parameter IM5 (dBm) 71.2 67.5 42.8 36.9
CW model IM5 (dBm) 67.4 53.8 44.0 33.5

Figure 6.43 Validation of wide-band X-parameter: gain compression and intermodulation

(measured, dots; WB model, solid line; CW model, dashed line).

It is observed that the wide-band X-parameters dene an effective quasi-static model

that is different from the actual static (CW) X-parameters. There is no hysteresis in the
predicted instantaneous AM-to-AM, for instance. The predictions with wide-band
X-parameters are much closer to the actual measured characteristics of the DUT than
 References 187

those predicted by the CW X-parameters. The results are quite consistent with the
qualitative reasoning presented in Section 6.6.8.1. In particular, the device appears to be
more linear for wider bandwidths based on its less compressed AM-to-AM charac-
teristics and also its decreased intermodulation distortion values.

6.6.8.3 Discussion
Wide-band X-parameters still require stepped complex amplitude step responses to
characterize the bivariate kernel appearing in (6.41). However, this represents a drastic
simplication of the trivariate kernel of the full dynamic X-parameters required for
(6.38). This results in many fewer measurement points, since only the initial time point
after the transition needs to be measured, rather than the entire time record of the
response all the way to steady state.
It is also evident from (6.41) that this effective static model requires knowledge only
of the bivariate step response, and is therefore valid for any (fast) modulated signal. It is
independent of format, PDF, average power, or other details of the signals.
The accuracy of the approach is expected to be well suited for continuously modu-
lated very wide-band signals, as is characteristic of digitally modulated communications
signals. Even if some of the energy in such broad-band signals happens to fall near a
narrow dynamical resonance, it represents a tiny fraction of the total energy of the signal
and is not expected to require the full dynamic X-parameter approach.
Of course, in general, the dynamic X-parameter theory is a superset of both the CW
and WB X-parameters, and reduces to them in the appropriate static and wide-band
limits, respectively.

References

[1] M. Rudolph, C. Fager, and D. E. Root, Eds., Nonlinear Transistor Model Parameter Extrac-
tion Techniques. Cambridge: Cambridge University Press.
[2] A. Soury and E. Ngoya, Handling long-term memory effects in X-parameter model, in
IEEE Int. Microwave Symp. Dig., Montreal, Canada, June 2012, pp. 13.
[3] J. Verspecht, J. Horn, L. Betts, C. Gillease, and D. E. Root, Extension of X-parameters to
include long-term dynamic memory effects, in 2009 IEEE MTT-S Int. Microwave Symp.
Dig., Boston, MA, June 2009, pp. 741744.
[4] J. Wood, D. E. Root, and N. B. Tullaro, A behavioral modeling approach to nonlinear
model-order reduction for RF/microwave ICs and systems, IEEE Trans. Microw. Theory
Tech., vol. 52, no. 9, part 2, pp. 22742284, Sept. 2004.
[5] D. E. Root, J. Wood, and N. Tullaro, New techniques for nonlinear behavioral modeling of
microwave/RF ICs from simulation and nonlinear microwave measurements, in 40th ACM/
IEEE Design Automation Conf. Proc., Anaheim, CA, June 2003, pp. 8590.
[6] J. Xu, M. M. Iwamoto, J. Horn, and D. E. Root, Large-signal FET model with multiple time
scale dynamics from nonlinear vector network analyzer data, in IEEE MTT-S Int. Microwave
Symp. Dig., Anaheim, CA, May 2010, pp. 417420.
[7] J. C. Pedro and S. A. Maas, A comparative overview of microwave and wireless power-
amplier behavioral modeling approaches, IEEE Trans. Microw. Theory Tech., vol. 53,
no. 4, Apr. 2005.
188 Memory

[8] A. Soury, E. Ngoya, and J. M. Nebus, A new behavioral model taking into account
nonlinear memory effects and transient behaviors in wideband SSPAs, in 2002 IEEE
MTT-S Conf. Dig., Seattle, WA, June 2002, pp. 853856.
[9] J. Verspecht, J. Horn, T. Nielsen, and D. E. Root, Extension of X-parameters to include
long-term memory effects, in IEEE ARFTG Conference, Clearwater Beach, FL, Dec. 2010.
[10] J. Wood and D. E. Root, Fundamentals of Nonlinear Behavioral Modeling for RF and
Microwave Design. Norwood, MA: Artech House, 2005, chap. 3.

Additional reading

P. Roblin, D. E. Root, J. Verspecht, Y. Ko, and J. P. Teyssier, New trends for the nonlinear
measurement and modeling of high-power RF transistors and ampliers with memory effects,
IEEE Trans. Microw. Theory Tech., vol. 60, no. 6, part 2, pp. 19641978, 2012.
D. E. Root, M. Iwamoto, and J. Wood, Device modeling for III-V semiconductors: an overview,
IEEE CSIC Symp., Monterey, CA, 2004, pp. 279282.
D. E. Root, D. Sharrit, and J. Verspecht, Nonlinear behavioral models with memory: formula-
tion, identication, and implementation, IEEE MTT-S Int. Microwave Symp. Workshop (WSL)
on Memory Effects in Power Ampliers, San Francisco, CA, June 2006.
J. Verspecht, D. Gunyan, J. Horn, J. Xu, A. Cognata, and D.E. Root, Multi-tone, multi-port, and
dynamic memory enhancements to PHD nonlinear behavioral models from large-signal meas-
urements and simulations, in IEEE MTT-S Int. Microwave Symp. Dig., Honolulu, HI, June
2007, pp. 969972.
 Appendix A: Notations and general
definitions

A.1 Sets

{x1, x2, x3} a set of elements (curly brackets)

N the set of natural numbers: {1, 2, . . .}
N* the set of natural numbers, including zero: {0, 1, 2, . . .}
N
N [ f 0g A:1
Z the set of integer numbers: {. . ., 2, 1, 0, 1, 2, . . .}
R the set of real numbers
R the set of real positive numbers
R the set of real negative numbers
C the set of complex numbers
z* complex conjugate

A.2 Vectors and matrices

A vector or a matrix is identied by a square bracket. The square bracket may be omitted
in the text when the ideas remain clear. Sometimes the omission of the square brackets
might be mentioned in the text, for clarity.
A vector of n elements (square brackets):
2 3
x1
6 x2 7
x 6 7
4 5: A:2
xn
A matrix of n  m elements (square brackets):
2 3
x11 x12 . . . x1m
6 x21 x22 . . . x2m 7
X  6
4
7: A:3
5
xn1 xn2 . . . xnm
190 Notations and general definitions

Transposed matrix:
2 3T 2 3
x11 x12 ... x1m x11 x21 . . . xn1
6 x21 x22 ... x2m 7 6 x12 x22 . . . xn2 7
X  6
T
4
7 6 7: A:4
5 4 5
xn1 xn2 ... xnm x1m x2m . . . xnm
Unity matrix:
2 3
1 0 ... 0
60 1 ... 0 7
6
I  4 7: A:5
5
0 0 ... 1
A.3 Signal representations

A.3.1 Time-domain representation (real signal)

st S r cos t ,
A:6
2f ,
where

S(r) (real) amplitude

angular frequency
f frequency
initial phase
t (instantaneous) phase

A.3.2 Complex representation (complex envelope signal)

st RefSe jt g,
p A:7
j 1,
where S is the complex amplitude:

S S r e j : A:8
It is common in the industry to use an identier for the complex amplitude (Sc, for
example). In this text the preference is to use an identier for the real amplitude (S(r))
because it is the complex amplitude that is used in the majority of cases presented.
The complex amplitude, S, is typically interpreted as a vector in the complex plane,
also known as a phasor. The RMS value of the complex amplitude, shown in (A.9), is
used often in engineering practice:
S
S RM S p : A:9
2
In this text, its vector interpretation is called the RMS vector or RMS phasor.
 A.4 Fourier analysis 191

A.4 Fourier analysis

Although other conventions exist and are in use, the convention presented here is used
in this book and is commonly used in the electrical engineering eld.
The Fourier series is expressed as shown in (A.10):
c0 X X
x t cn cos nt sn sin nt , A:10
2 n1 n1

where the Fourier coefcients are determined as follows:

T
2
c0 xt dt,
T
0
T
2
cn xt cos nt dt, A:11
T
0
T
2
sn xt sin nt dt:
T
0

A more compact form can be obtained using the notations in (A.12):

A 0 c0 ,
p
An c2n s2n ,
8 0 1 ( sn > 0 and cn  0
>
> s
>
> @ nA
>
> atan sn < 0 and cn  0
>
> cn
>
> s n 0 and cn > 0,
>
< A:12
0 1 ( sn < 0 and cn  0
n
>
> @ sn A
>
> atan sn > 0 and cn  0
>
> cn
>
> s n 0 and cn < 0,
>
>
>
:
0, sn 0 and cn 0:
The coefcient An is the amplitude and n is the phase of the spectral component on the
frequency n.
The Fourier series can then be expressed as shown in (A.13):

A0 X
x t An cos nt n : A:13
2 n1

The complex amplitude is dened using (A.14):

Anc An e jn : A:14
The complex form of the Fourier series using only positive frequencies is shown in (A.15):
A0 X  
x t Re Anc e jnt : A:15
2 n1
192 Notations and general definitions

The complex form of the Fourier series using both positive and negative frequencies is
shown in (A.16):

1 X
x t Anc e jnt : A:16
2 n

Using (A.12), the complex amplitude, Anc, is expressed as shown in (A.17):

Anc cn  jsn : A:17
Considering (A.11), the complex amplitude, Anc, can be determined directly from the
original function, x(t), as shown in (A.18):
T
2
Anc xt ejnt dt: A:18
T
0

The spectrum of a non-periodic signal is continuous and is determined by the Fourier

transform, F f:g, as shown in (A.19):

n o
X F xt xt ejt dt: A:19


The time-domain signal can be determined from its frequency-domain spectrum using
the inverse Fourier transform, as shown in (A.20):

n o
1
xt F 1 X X e jt d: A:20
2


A.5 Wave definitions

A.5.1 Generalized power waves

Given a network port, as shown in Figure A.1, the generalized power waves are dened
as a function of port voltage and port current.
The voltage, current, and waves in Figure A.1 are sinusoidal signals with the same
frequency, and they are represented in the frequency domain by their complex
amplitudes.

Figure A.1 A port of a network, with port voltage, current, and waves.
 A.5 Wave definitions 193

Considering a real and positive impedance for the port, Z0 2 R, the generalized
power waves are dened by (A.21) [1], where A is the incident wave and B is the
scattered wave:
V Z0I
A p
2 Z0
V  Z0I Z 0 2 R : A:21
B p
2 Z0
p
The unit of measurement for generalized power waves is watt, which results from (A.21).
The reference impedance is real and positive for the vast majority of practical
situations. Almost all modern network analyzers use transmission lines, of characteristic
impedance 50 , which set the reference impedances for the measurement process.
A more general notation is used in this text than in [1], in the sense that V and I in
(A.21) are not restricted to their RMS (complex) values, but rather they may have either
peak (complex amplitude) or RMS (complex) values depending on the purpose for
which the mathematical expression is used:
Apk
ARM S p 2 C: A:22
2
For example, when power calculations are conducted, the RMS value must be used, as
shown in (A.23) for the power of wave A:
V RM S Z 0 I RM S
ARM S p ,
2 Z0 A:23
 2
 
power of A-wave ARM S  :

When a time-domain representation is needed, the peak value must be used, as shown in
(A.24) for the incident wave, A:

V pk Z 0 I pk
Apk p ,
2 Z0 A:24
n o
at Re Apk e jt :

The time-domain representation of the generalized power waves is a relatively new

concept, and is needed for a better understanding of the cross-frequency phase relation-
ships between various sinusoidal signals used in the analysis of the behavior of non-
linear systems.
It can be easily proven that the time-domain waves are directly related to the time-
domain representation of the port voltage and port current, as shown in (A.25):
vt Z 0 it
a t p
2 Z0
Z 0 2 R : A:25
vt  Z 0 it
b t p
2 Z0
194 Notations and general definitions

The relationship between frequency and time domains, established by the Fourier
transform, is thus extended to the generalized power waves. The complex amplitudes
represent the coefcients of the Fourier series of a time-periodic wave, b(t), as shown in
(A.26):
T
pk 2 2
Bk bt ejkt dt, : A:26
T T
0

The reciprocal conversion shown in (A.27) determines the time-domain wave when its
frequency-domain representation is known (the complex amplitudes of its harmonics):
!
XN
pk jkt
bt Re Bk e : A:27
k1

For convenience of notation, the RMS identier is omitted in this text when possible.
As a result, unless otherwise noted, the upper-case letters represent RMS values of the
complex amplitudes. The pk identier for peak values is used explicitly, where needed.

A.5.2 Voltage waves

An alternative set of waves is sometimes used in the practice of RF/MW measurement
and design. These are voltage waves, which are dened as shown in (A.28), as a
function of port voltage and current using the reference impedance, which is considered
a real and positive number in this book:
V Z0I
Av
2
V  Z0I Z 0 2 R : A:28
Bv
2

The identier v is used here to distinguish clearly between the voltage waves A(v) and
B(v) and the generalized power waves A and B.
The unit of measurement for voltage waves is the volt, which results from (A.28).
All the other comments presented for generalized power waves are also valid for the
voltage waves.

A.6 Linear network matrix descriptions

A more detailed presentation of this topic can be found in [2].

An N-port network has N 1 pins. One of the pins is selected as reference, and each
of the remaining N pins forms a port with respect to the reference pin.
All port voltages, currents, incident waves, and scattered waves form corresponding
vectors, as shown in (A.29):
 A.6 Linear network matrix descriptions 195

2 3 2 3 2 3 2 3
V1 I1 A1 B1
6 V2 7 6 I2 7 6 A2 7 6 B2 7
V  6 7
4 5, I  6 7
4 5, A 6 7
4 5, B 6 7
4 5: A:29
VN IN AN BN
A.6.1 S-parameters
The S-parameters relate scattered waves to the incident waves, as shown in
B S A: A:30

A.6.2 Z-parameters
The Z-parameters relate voltages to currents, as shown in
V  Z I : A:31
A.6.3 Y-parameters
The Y-parameters relate currents to voltages, as shown in
I  Y V : A:32
References

[1] K. Kurokawa, Power waves and the scattering matrix, IEEE Trans. Microw. Theory Tech.,
vol. 13, no. 2, Mar. 1965.
[2] G. Gonzalez, Microwave Transistor Ampliers, 2nd edn. Englewood Cliffs, NJ: Prentice Hall,
1984.
 Appendix B: X-parameters
and Volterra theory

B.1 Introduction

This appendix describes the theoretical derivation of the properties of the X-parameter
functions for a small A1,1 amplitude. The theory is based on the assumption that the
modeled behavior can be described by the Volterra theory in the limit for the amplitude
of A1,1 going to zero. The theory shows that the frequency-domain version of the
Volterra theory can be considered as a McLaurin series polynomial approximation of
the X-parameters.

B.2 Mathematical notation and problem definition

For clarity, the mathematical notation that is used is recalled in the following:
Ap,h the spectral component with harmonic index h of the incident wave at port p
Bp,h the spectral component with harmonic index h of the scattered wave at port p
Y* refers to the complex conjugate of the phasor Y
P the unity length phasor having the same phase as A1,1, dened as


P exp jA1, 1 : B:1

The symbols X rFB S T

, m :, X r, m;s, n :, and X r, m;s, n : refer to the different kinds of
X-parameter functions. They are dened by the following X-parameter equation (for
simplicity it is assumed that the LSOP is dened solely by jA1,1j):

 m X 
 mn
 mn 
Br, m X rFB
, m jA1, 1 j P X rS, m;s

, n jA1, 1 j P As, n X rT, m;s

, n jA1, 1 j P A s, n :
s, n61, 1

B:2

Note that X rFB

, m : refers to the large-signal part of Br,m, which is a function of the
amplitude of A1,1 only, X rS, m;s

, n : refers to the coefcient associated with the con-
tribution to Br,m of the small-signal component As,n, and X rT, m;s
, n : refers to the
coefcient associated with the contribution to Br,m of the conjugate of the small-signal
component As,n. In order to simplify the analysis, the DC waves will rst be excluded.
The degenerate DC case will be analyzed in Section B.5.
 B.3 Application of the Volterra theory 197

The problem that is solved is the behavior of the functions X rFB S

, m :, X r, m;s, n :, and
X rT, m;s

, n : in the limit of the amplitude of A1,1 going to zero. In other words, the goal is
to nd the properties of the McLaurin series for these functions versus A1,1.

B.3 Application of the Volterra theory

In order to nd the characteristics of the scattering functions for small amplitudes of A1,1
it is assumed that the behavior can be described by the Volterra theory in the limit for the
amplitude of A1,1 going to zero. In the Volterra theory the output spectra are written as a
series of convolutions of the input spectra [1][2]. As a consequence, if the input spectra
are sets of discrete tones, where each tone is represented by an input phasor, the output
spectra will also comprise sets of discrete tones. The frequencies of the tones present in
the output are integer linear combinations of the frequencies of the input tones, and each
output phasor can be written as a polynomial function of the input phasors and the
conjugate of the input phasors. The order of the input phasors in each term of the
polynomial that represents an output phasor will be such that the time-invariance
principle holds. Expressed in the time domain, the time-invariance principle states that
the application of an arbitrary delay in the input signal has to result in exactly the same
delay for the output signal. Expressed in the frequency domain, the time-invariance
principle states that the application of any linear phase shift to the input phasors has to
result in the same linear phase shift to the output phasors. The above will be illustrated
by the following example.
Consider that there are two tones present as the input signal. The input signal is in this
case represented by a phasor x having a frequency fx and a phasor y having a frequency
fy. According to the Volterra theory, the output spectrum is discrete and is represented
by a set of phasors z[mn], with two integer indices m and n denoting the output
frequency, whereby the output phasors z[mn] have a frequency equal to mfx nfy.
Also according to the Volterra theory, z[mn] can be written as the following polynomial
in x, y, x*, and y*:
X
zmn H pqrs xp xq yr ys : B:3
8p, q, r, s  0
p  q  m f x r  s  n f y 0

In (B.3), the summation goes over a constrained set of positive integers p, q, r, and s
that satisfy the time-invariance principle. The coefcient Hpqrs is called a Volterra
kernel.
The principle of Volterra theory can be applied to the case where the input phasors are
the As,n and where the output phasors are the Br,m. This case is different from the above
two-tone case for three reasons. First, there is only one fundamental frequency present,
and as such the frequency of the phasors As,n and Br,m is indicated by their second
integer index. Secondly there is a plurality of signal ports, where the port is indicated by
the rst index of the phasors As,n and Br,m. The port association of a kernel will be
denoted by its rst two indices. The third difference is that the description is limited to
198 X-parameters and Volterra theory

terms that are at most of order 1 in any of the As,n and the conjugates of the As,n for all
indices (s,n) different from (1,1). The result is the following:
0 1
XB X X q C
Br , m B H rs, mn, pq10 Ap1, 1 A q
H rs, mn, pq01 Ap1, 1 A C
@ 1, 1 As, n 1 , 1 As , n A :
sn 8p,q  0 8p,q  0
pqnm 0 pqnm 0

B:4

In (B.4) the summation is not only constrained by the equations p  q n  m 0 and

p  q  n  m 0, respectively, but also to indices p and q that are both non-negative
integers.

B.4 Derivation of the McLaurin series

Next, (B.4) is rewritten by substituting A1,1 by jA1,1jP:

0 1
XB X X C
B H rs, mn, pq10 jA1, 1 jpq Ppq As, n H rs, mn, pq01 jA1, 1 jpq Ppq A C
Br , m @ s, n A :
sn 8p,q  0 8p,q  0
pqnm 0 pqnm 0

B:5

Because of the constraint on the ps and qs the exponent (p  q) can be substituted by

(m  n) and (m n), respectively:
0 1
XB X X C
B H rs, mn, pq10 jA1, 1 jpq Pmn As, n H rs, mn, pq01 jA1, 1 jpq Pmn A C
Br , m @ s, n A:
sn 8p,q  0 8p,q  0
pqnm 0 pqnm 0

B:6

Next, the above expression will be simplied further by rewriting the exponent
(p q).
Consider the rst term in (B.6). If (n  m) is positive, q can be substituted by
(p n  m) whereby the summation in p ranges from 0 to innity. If (n  m) is
negative, p can be substituted by (q  n m) whereby the summation in q ranges from
0 to innity. This is illustrated by the following equations:

X X

if n  m  0, H rs, mn, pq10 jA1, 1 jpq H rs, mn, ppnm10 jA1, 1 j2pnm ;
8p, q  0 p0
pqnm 0

B:7
 B.4 Derivation of the McLaurin series 199

X X

if n  m < 0, H rs, mn, pq10 jA1, 1 jpq H rs, mn, qnmq10 jA1, 1 j2qnm :
8p, q  0 q0
pqnm 0

B:8
Consider the second term in (B.6). The variable p can be substituted by (q n m)
whereby the summation in q ranges from 0 to innity:

X X

H rs, mn, pq10 jA1, 1 jpq H rs, mn, qnmq10 jA1, 1 j2qnm : B:9
8p, q  0 q0
pqnm0

To simplify the mathematical notation, the symbol q is replaced by the symbol p as the
iterator, and a new set of kernels is dened as follows:
rs, pmn H rs, mn, ppnm10 for n  m  0; B:10

rs, pmn H rs, mn, ppnm10 for n  m < 0; B:11

rs, pmn H rs, mn, pnmp10 : B:12

Using (B.7) to (B.12) and substituting the results in (B.6) results in the following
simplied expression:
!
X X
2pjmnj mn
X

2pmn mn
Br , m rs, pmn jA1, 1 j P As, n rs, pmn jA1, 1 j P As, n :
sn p0 p0

B:13

Comparing (B.13) with (B.2), the McLaurin series are identied for the functions
X rS, m;s
T
, n : and X r, m;s, n : as follows:

X

X rS, m;s

,n rs, pmn jA1, 1 j2pjmnj ; B:14
p0

X

X rT, m;s

,n rs, pmn jA1, 1 j2pmn : B:15
p0

The McLaurin series for the function X rFB

, m : is found by looking at the special case of
the summation in (B.13) where (s,n) (1,1):

 m X X

X rFB
,m

jA 1, 1 j P r1 , pm1 j A1 , 1 j2pjm1j m1
P A 1, 1 r1,pm1 jA1,1 j2pm1 Pm1 A
1, 1 :
p0 p0

B:16
Eliminating Pm on both sides of (B.16) and substituting A1,1 by jA1,1jP results in the
following:
200 X-parameters and Volterra theory

 X
X

X rFB
, m jA1, 1 j r1, pm1 jA1, 1 j2pjm1j1 r1, pm1 jA1, 1 j2pm2 : B:17
p0 p0

Excluding the DC output, which corresponds to m 0, all m are greater than or equal to
1. In that case, jm  1j equals m  1 and (B.17) can be rewritten as follows:

 X
X

X rFB
, m jA1, 1 j r1, pm1 jA1, 1 j2pm r1, pm1 jA1, 1 j2pm2 : B:18
p0 p0

Finally, (B.18) can be written as a McLaurin series as follows:

 
X 
X rFB
,m

jA 1, 1 j r1 , 0m1 jA 1, 1 j m
r1, p1m1 r1, pm1 jA1, 1 j
2pm2
: B:19
p0

B.5 McLaurin series for the DC output

The properties of the McLaurin series for the DC output equations can be found by
performing the above analysis for m 0. All equations up to (B.17) remain valid. The
result is the following:

 X  
X Y
r;s, n jA 1, 1 j rs, p0n rs , p0n jA1, 1 j2pn ; B:20
p0

 X  
X Z
r;s, n jA 1, 1 j rs, p0n rs, p0n jA1, 1 j
2pn
; B:21
p0

 
X 
X rFI jA1, 1 j r1, 001 r1, p101 r1, p01 jA1, 1 j2p2 : B:22
p0

B.6 Conclusions

Several terms are missing in the McLaurin series for the X rFB S
, m :, X r, m;s, n :, and
T
X r, m;s, n : functions. The results can be summarized as follows.
The lowest order of the jA1,1j term in the McLaurin series for X rFB
, m jA1, 1 j is m.
The lowest order of the jA1,1j term in the McLaurin series for X rS, m;s
, n jA1, 1 j
is jm  nj.
The lowest order of the jA1,1j term in the McLaurin series for X rT, m;s
, n jA1, 1 j
is m n.
For each of the functions the order of any consecutive jA1,1j term increases by 2.
The above conclusions also hold for the DC output functions (m 0).
 References 201

References

[1] J. C. Peyton Jones and S. A. Billings, Describing functions, Volterra series, and the analysis
of non-linear systems in the frequency domain, Int. J. Control, vol. 53, no. 4, pp. 871887,
1991.
[2] J. Verspecht, Describing functions can better model hard nonlinearities in the frequency
domain than the Volterra theory, Ph.D. thesis annex, Vrije Univ. Brussel, Belgium, Nov.
1995; available at http://www.janverspecht.com/skynet/Work/annex.pdf.
Appendix C: Parallel Hammerstein
symmetry

According to [1], among the best performing behavioral models are the parallel
Hammerstein (PH) models. This appendix shows that a PH model introduces a sym-
metry not consistent with realistic behavior in so-called complementary large-signal
step responses.
Consider the response, denoted BLS(A1, A2, t), of an amplier to a large-signal step
from an initial amplitude of A1 to an amplitude A2. The complementary large-signal step
response is just the response to a step input from amplitude A2 to an amplitude A1,
namely BLS(A2, A1, t). It will be shown in the following that for PH models the sum of
BLS(A1, A2, t) and BLS(A2, A1, t) is a constant, independent of time, that depends only on
A1 and A2.
In a general PH model the relationship between the input A(t) and the output B(t) is
given by

N

X    
Bt F i At  u e j Atu hi udu: C:1
i1
0

Expressed in words, the input signal goes through a set of N parallel static nonlinear
operators Fi(.), whereby each output is convolved with a specic linear lter impulse
response hi(u). The output is then equal to the sum of the results of the convolutions. It
is important to note that the PH model of (C.1) is actually an extension of the PH models
described in [1] and [2] in the sense that the nonlinear operators Fi(.) as described in [1]
and [2] are restricted to simple monomials rather than general nonlinear functions,
namely Fi(x) x2i1.
The complementary large-signal step responses are calculated as follows:
2t 3
X
N
BLS A1 , A2 , t 4 F i A2 hi udu F i A1 hi udu5; C:2
i1
0 t
2t 3
X
N
BLS A2 , A1 , t 4 F i A1 hi udu F i A2 hi udu5: C:3
i1
0 t

The sum of the complementary large-signal step responses is given by

 References 203

2 3
X
N
BLS A1 , A2 , t BLS A2 , A1 , t 4 F i A1 hi udu F i A2 hi udu5, C:4
i1
0 0

which can be written as

X
N
BLS A1 , A2 , t BLS A2 , A1 , t F i A1 F i A2 hi udu: C:5
i1
0

The right-hand side of (C.5) is independent of time, proving that the sum of comple-
mentary large-signal step responses generated by any PH model is constant in time, as
claimed.
A simple example demonstrates that this is not consistent with realistic behavior.
Consider the case whereby one switches from an amplitude equal to zero to a large
amplitude, and this for a system with signicant modulation-induced baseband memory.
The large-signal step response will show a transient with a duration that corresponds to
the time constants associated with the baseband memory effects. For any PH model, the
complementary large-signal step response, the response of the amplier when switching
from a high input amplitude to zero, will have the same transient response, but with an
opposite sign. As soon as the input amplitude is switched off, however, one would
expect the output signal to go rapidly to zero, in a time span that corresponds to the very
short time constants associated with the carrier memory effects, not to the long time
constants associated with the modulation-induced baseband effects. This nonsensical
relationship between complementary large-signal step responses is not present in the
dynamical X-parameter model.

References

[1] J. C. Pedro and S. A. Maas, A comparative overview of microwave and wireless power-
amplier behavioral modeling approaches, IEEE Trans. Microw. Theory Tech., vol. 53, no.4,
Apr. 2005.
[2] M. Isaksson, D. Wisell and Daniel Rnnow, A comparative analysis of behavioral models
for RF power ampliers, IEEE Trans. Microw. Theory Tech., vol. 54, no. 1, pp. 348359,
Jan. 2006.
Appendix D: Wide-band memory
approximation

In this appendix it is proved that

0 1

GjAtj, jAt  j, d  @ GjAtj, a, d Aada
0 0 0

under the following conditions:

(1) the input envelope is varying quickly compared to the time constants associated
with the function G(x,y,t);
(2) the statistical properties of the input envelope amplitude are stationary and are
described by the PDF (a).
The proof begins by splitting the time axis into small intervals, each of duration T.
The result is

 i1T

     X
    
   
G At , At  , d G At , At  u, u du: D:1
i0
0 iT

For any T that is small enough, the value of the function G(.,.,.) will not signicantly
depend on the time variable over each individual interval. This results in


i1T
     X
    
   
G At , At  , d  G At , At  u, iT du: D:2
i0
0 iT

Note that the value of jA(t  u)j may still vary signicantly over each interval in the
case of a fast varying envelope, which corresponds to a wideband modulation. The
remaining integrals on the right-hand side can be written as the limit of a Riemann sum:
    !
     X MX1  T 
G At, At  , d  lim G jAt j, A t  iT  j , iT T ,
M ! M  M
i0 j0
0
D:3
 Wide-band memory approximation 205

which can be written as

8M 1
X    !9
   
>
>
> G At , A t  iT  j
T , iT >
>
>
 >
<  >
=
     X
M
G At, At  , d 
j0
T lim :
M !>
> M >
>
i0 >
> >
>
0 : ;

D:4
Consider now the expression denoted by the curly brackets {}. This expression simply
represents the sample average of a large number of sampled values. The statistical law
of large numbers states that the value of this average will converge to the expected value
as the number of samples increases to innity. This is mathematically expressed by
8M 1
X      !9

>
>
>
 
G At , A t  iT  j
T , iT >
>
> 
>
< M  >
=   
j0  
lim G At , a, iT ada,
M ! >> M >
>
>
> >
>
: ; 0

D:5
whereby (a) represents the probability density function of A(t). Note that, in deriving
the above equation, one uses the fact that A(t) has stationary statistical properties.
Substitution of (D.5) in (D.4) results in
 
     X
  
   
G At , At  , d  T G At , a, iT ada: D:6
i0
0 0

The remaining sum corresponds to a Riemann sum. Taking the limit of this expression
for ever smaller time intervals T leads to another Riemann integral:
0 1
    X
   
   
G At, At, d  lim T G At ,a,iT ada @ G At,a, d Aada:
T!0
i0
0 0 0 0

D:7
Q.E.D.
Appendix E: Solutions to exercises

Exercise 1.1 Referring to Figure 1.9, Eq. (1.5) is used to write KCL at the internal
node according to (E.1):
1 2
it i2 t i1 t 0
1 1 2 2
E:1
a2 t  b2 t a1 t  b1 t:
The voltage at the internal node can be written using Eq. (1.5) according to (E.2):
1 1 2 2
vt a2 t b2 t a1 t b1 t: E:2
1
Solving (E.2) for a2 t yields
1 1 2 2
a2 t b2 t a1 t b1 t: E:3
1
Substituting for a2 t using (E.3) into (E.1) results in (E.4):
1 2
b2 t a1 t: E:4
Substituting (E.4) into (E.1) results in
1 2
a2 t b1 t: E:5
Transforming (E.4) and (E.5) into the frequency domain results in Eqs (1.17) and (1.18)
for the relationships between the complex amplitudes of the waves, reproduced here as
(E.6):
1 2
B2 A1 ,
2 1
E:6
B1 A2 :

Exercise 1.2 The objective is to reduce the problem to a 2  2 matrix dening the
1 2 1 2
linear dependence of B1 and B2 on the incident waves, A1 and A2 , since these are
the reected and incident waves of the composite structure.
Using the denitions of the S-parameter matrices of the two constituent systems, and
the solution of Exercise 1.1, the set of equations governing the composite system is
given by the sets of equations (E.7) and (E.6):
 Solutions to exercises 207

2 1
3 2 1 1
32 1
3
B1 S 11 S 12 0 0 A1
6 1 7 6 1 7 6 1 7
6B 7 6S 1
0 7 6 7
6 2 7 6 21 S 22 0 7 6 A2 7
6 2 7 6 7 6 7: E:7
6B 7 6 0 2
S 12 7
2 6 A2 7
4 1 5 4 0 S 11 54 1 5
2 2 2 2
B2 0 0 S 21 S 22 A2
1 2 2
Substituting for the incident waves A2 and A1 in terms of the scattered waves B1 and
1
B2 , respectively, using (E.6), (E.7) can now be written as (E.8):
2 1 3 2 1 1
3 2 1 3
B1 S 11 S 12 0 0 A
6 1 7 6 1 7 6 12 7
6B 7 6S 1
0 7 6 B1 7
7 6
6 2 7 6 21 S 22 0 7
6 2 7 6 76 7: E:8
6B 7 6 0 2 2 7 6 1 7
4 1 5 4 0 S 11 S 12 5 4 B2 5
2 2 2 2
B2 0 0 S 21 S 22 A2
2
The equations given by the second and third rows of (E.8) can be used to solve for B1
1 1 2
and B2 in terms of A1 and A2 and four of the eight S-parameters of the original
2 1
systems. Solving for B1 and B2 results in (E.9) and (E.10):

1 2 2
2 S 21 S 11 1 S 12 2
B1 2
A
1 1 2 1
A2 , E:9
1  S 11 S 22 1  S 11 S 22

1 2 1
1 S 21 1 S 12 S 22 2
B2 2
A
1 1 2 1
A2 : E:10
1 S 11 S 22 1 S 11 S 22
2 1
Equations (E.9) and (E.10) are used to substitute for B1 and B2 when evaluating the
1 2
rst and fourth rows of (E.8) for B1 and B2 , respectively. After collecting terms in
1 2 1
A1 and A2 , the resulting 2  2 matrix expressing the linear relationships of B1 and
2 1 2
B2 to A1 and A2 is the S-parameter matrix of the composite system. This is simply
expression (1.19), reproduced here as (E.11):
0 1 1 1 2 1
1 2 S 12 S 21 S 12 S 12
B S 11 S 11 1 2 1 2 C
B 1  S 22 S 11 1  S 22 S 11 C
B C
S composite
B 1 2 2 2 C: E:11
B S 21 S 21 S 12 S 21 2 C
@ 1
S 22 S A
1 2 1 2 22
1  S 22 S 11 1  S 22 S 11

Exercise 1.3 Start from the denition of the S-parameter matrix, given in terms of
the matrix relationship between the vector of scattered and incident waves, namely Eq.
(A.30):
B S A: E:12
208 Solutions to exercises

Substitute the vector relationships for the incident and scattered waves in terms of the
currents and voltages using Eq. (1.1). The result is given by (E.13):

I  Z 0 Y V  S I Z 0 Y V : E:13

Here I is the unit matrix and Y is the admittance matrix dened in Appendix A. This
must hold for all vectors, [V], so the matrix relationship (E.14) follows:

I  Z 0 Y  S I Z 0 Y : E:14

Assuming the matrix [I Z0Y] has an inverse, (E.14) is post-multiplied by [I Z0Y] 1

to obtain (E.15), which is equivalent to Eq. (1.31):

I  Z 0 Y I Z 0 Y 1 S  E:15

Exercise 1.4 Start from Eq. (1.30) for the model admittance matrix calculated in Eq.
(1.10). The relevant matrices to multiply, using (E.15), are given in (E.16) and (E.17),
where the bias-dependent arguments of the element values are not shown:
 
1  jcGS Z 0 0
I  Z 0 Y  , E:16
gm Z 0 1  gDS Z 0
 
1 1 gDS Z 0 0
I Z 0 Y 1 :
1 gDS jcGS Z 0 jcGS gDS Z 0 2 g m Z 0 1 jcGS Z 0
E:17

Matrix multiplying (E.16) and (E.17) produces, after algebraic simplication, Eq.
(1.32), reproduced here as (E.18):
0 1
1  jcGS Z 0
0
B 1 jcGS Z 0 C
B C
SB 2g m Z 0 1  gDS Z 0 C : E:18
@ A
1 gDS Z 0 1 jcGS Z 0 1 gDS Z 0

Exercise 2.1 Equation (2.51) is a time-invariant map provided that whenever

A1,1 ! A1,1e j then B2,1 ! B2,1e j for any angle .
Evaluating the proposed map (2.51) for A1,1e j yields (E.19):

GA1, 1 e j A1, 1 e j 3 GA1, 1 e j A31, 1 e3j

  E:19
6 GA1, 1 A31, 1 e j B2, 1 e j :

Therefore, Eq. (2.51) is not a time-invariant map and cannot be a correct model of a
time-invariant component.
 Solutions to exercises 209

Exercise 2.2 Evaluating the proposed map (2.52) for A1,1e j, (E.20) is obtained as
follows:
 2  2
   
GA1, 1 e j A1, 1 e j  A1, 1 e j GA1, 1 e j A1, 1  A1, 1 e j
  2  E:20
 
GA1, 1 A1, 1  A1, 1 e j B2, 1 e j :

Therefore, Eq. (2.52) denes a map that is time invariant.

Exercise 2.3 Equation (2.52) can be written in the following form:

   
B2, 1 G  jA1, 1 j2 A1, 1 G  jA1, 1 j2 jA1, 1 je jA1, 1 : E:21

Equation (E.21) can be written as (E.22):

B2, 1 A1, 1 X 2, 1 jA1, 1 j P, E:22

where
 
X 2, 1 jA1, 1 j G  jA1, 1 j2 jA1, 1 j E:23

and

P e jA1, 1 : E:24

Exercise 3.1 The response is determined by the X-parameter model:

FB S T
Bp,k X p:k ref LSOPS Pk X p,k;q,l ref LSOPS Aq,l Pkl X p,k;q,l ref LSOPS Aq,l Pkl :
E:25

A simpler notation is used in this exercise:

S X S
X p, k;q, l ref LSOPS Se j ,
T jX T
X p, k;q, l ref LSOPS T e , E:26
a
Aq, l ae j :
FB
The rst term in Bp,k, X p:k ref LSOPS Pk , does not depend on the small-signal
perturbation. It represents the response at the LSOP. It is thus sufcient to prove that
FB
the difference Bp, k  X p:k ref LSOPS Pk describes an ellipse when the phase of the
small-signal perturbation Aq,l varies.
FB
The difference Bp, k  X p:k ref LSOPS Pk formally shifts the origin of the orthog-
FB
onal system of coordinates to the tip of the vector X p:k ref LSOPS Pk .
210 Solutions to exercises

An additional notation is used for convenience in this exercise, as shown in (E.27):

S ! S
X p, k;q, l ref LSOPS Pkl S Se ,
j

T ! T
X p, k;q, l ref LSOPS Pkl T T e j ,
E:27
! a
Aq, l a ae j ,
FB !
Bp, k  X p:k ref LSOPS Pk z :

The notation P is related to the phase of the large-signal stimulus, A1, 1 jA1, 1 jejA1, 1 ,
as shown in its denition (E.28):

P e jA1, 1 : E:28
Using the denition (E.28), the phases (S)
and (T)
have the form shown in (E.29):

S X S k  l A1, 1 ,
E:29
T X T k l A1, 1 :

Using the above notation, (E.25) transforms into (E.30):

! !! ! !
z S a T a , E:30

z Sae j T ae j :
! S
a T
 a
E:31
!
The vector z has maximum amplitude when its two components are aligned. The
!
notation (M) designates the phase of z when it has maximum amplitude and aM
!
denotes the phase of a when this condition exists; (M) is determined by condition
(E.32):

M S aM T  aM 2k, k 2 : E:32
Equation (E.32) also determines the value of aM , as shown in (E.33):

T  S
aM k: E:33
2
The direction of the maximum is shown in (E.34):

T  S S T
M S aM S k k: E:34
2 2
Considering the periodicity of the phase, only two values represent distinct positions:
the ones obtained for k 0 and for k 1, as shown in (E.35):

M S T
1 M ,
2
E:35
S T
M
2 M :
2
The phase (M) can be expressed as a function of the X-parameter phases and the phase
of the original large-signal stimulus, as shown in (E.36), where (E.29) is used:
 Solutions to exercises 211

X S X T
M kA1, 1 : E:36
2
!
The two positions of the vectors z with maximum amplitude reside on the same line,
and the vectors extend into opposite directions from their origin. The distance between
the tips of these two vectors has a value of 2a(S T), as shown in (E.37):
 
 
aS T e j1  aS T e j1  2aS T :
M M

  E:37
!
In a similar manner, it can be proven that z has minimum amplitude when its two
components are counter-aligned:

m S am T  am 2k , k 2 : E:38
!
The notation designates the phase of z when it has minimum amplitude and the
(m)
!
notation am designates the phase of a when this condition exists; (m) is determined
by condition (E.39):
T  S
am 2k 1 : E:39
2 2
The direction of the minimum is shown in (E.40):

T S
m S am 2k 1 : E:40
2 2
Considering the periodicity of the phase, only two values represent distinct positions:
the ones obtained for k 0 and for k 1, as shown in (E.41):

m S T
1 m ,
2 2
E:41
m S T
2 3 m :
2 2

The phase (m) can be expressed as a function of the X-parameter phases and the phase
of the original large-signal stimulus, as shown in (E.42), where (E.29) is used:

X S X T
m kA1, 1 : E:42
2 2
!
The two positions of the vectors z with minimum amplitude reside on the same line,
and the vectors extend into opposite directions from their origin. The distance between
the tips of these two vectors has a value of 2a(S  T), as shown in (E.43):
 
 
aS  T e j1  aS  T e j1  2aS  T :
m m
E:43

It is also important to emphasize that the two directions determined so far (of the
maximum and minimum amplitudes) are perpendicular to each other, as shown in (E.44):

m  M : E:44
2
212 Solutions to exercises

It is thus conceivable that these two segments represent the two axes of an ellipse:
the major axis with a value of 2a(S T) and the minor axis with a value of
2a(S  T).
!
The foci need to be located on the major axis, symmetrically from the origin of z . Let
! !
us consider the two points identied as f 1 and f 2 , located on the major axis, symmetric-
ally from the center, at distances x and x, x  0. The sum of the distances from any of
the two ends of the major axis to these two points is shown in (E.45):
h i h i
S T a  x S T a x 2S T a: E:45

The sum of the distances from any of the two ends of the minor axis to these two points
is shown in (E.46):
q
2 S  T 2 a2 x 2 : E:46
! !
The two sums expressed in (E.45) and (E.46) must be the same if f 1 and f 2 are the foci
of an ellipse. The value of x determined from this condition is shown in (E.47):
p
x 2a ST : E:47
The two potential foci are thus identied by the following vectors:
! p M
f 1 2a ST e j ,
! p M
f 2a ST e j :
2 E:48
A rotation of the system of coordinates with (M) is needed for a simpler mathemat-
ical treatment. After this rotation, the x-axis is along the major axis. All vectors
maintain their magnitude, but their phases change with (M). After the rotation of
the coordinate system, all vectors are identied with the letter r in front of them, as
shown in (E.49):

Se j M
! ! jM S
rS S e ,
! !
T e j  ,
M T M
rT T ej E:49
ae j  ;
! ! j M a M
ra a e
! ! p M M p
2a ST e j  2a ST ,
M
rf f 1 ej
1
! ! p M M p p
2a ST e j  2a ST e j 2a ST :
M
rf f 2 ej E:50
2

After the coordinate system rotation, the vector ez can be expressed as in (E.51):

! !! ! !
rz rS ra rT ra
Se j  ae j  T e j  aej 
S M a M T M a M
E:51
Sae j T ae j :
S a M T a
2 
 Solutions to exercises 213

!
Considering the value of (M) from (E.35), rz can be expressed as in (E.52):

rz Sae j T
T aej T
! a a
: E:52
Using the notation in (E.53),

a  T , E:53
!
rz is reformulated as shown in (E.54):
!
rz Sae j T aej : E:54
! !
The distance from the tip of rz to the tip of rf 1 is shown in (E.55):
r
h
pi2 2
d1 aS T cos  2a ST aS  T sin  : E:55

After some algebraic transformations, d1 can be expressed as shown in (E.56):

p
d 1 aS T  2a ST cos : E:56
! !
In a similar fashion, the distance from the tip of rz to the tip of rf 2 is shown in (E.57):
r
h
pi2 2
d2 aS T cos 2a ST aS  T sin  : E:57

After some algebraic transformations, d2 can be expressed as shown in (E.58):

p
d 2 aS T 2a ST cos : E:58
The sum of the two distances is shown in (E.59); it is constant with respect to the
angle , and hence independent of the phase of the small-signal perturbation:
d d 1 d 2 2aS T : E:59
This proves that the contour described by the tip of the vector Bp,k is an ellipse.

Exercise 5.1 The signals can be written, according to Appendix A, as (E.60) and
(E.61):
 n o
a1 t Re A1, 1 e jt Re jA1, 1 je jA1, 1 t , E:60
 n o
a2 t Re A2, 1 e jt Re jA2, 1 je jA2, 1 t : E:61

At any instant in time, the phase difference between the signals, 2,1, is simply given by
(E.62):
2, 1 A2, 1  t  A1, 1  t A2, 1  A1, 1 : E:62
Clearly, from (E.62), 2,1 is independent of time (constant).
214 Solutions to exercises

Equation (E.63) then follows trivially from (E.62), for P e jA1, 1 :

e j2, 1 e jA2, 1 A1, 1  e jA2, 1 P1 : E:63

This is just Eq. (5.3).

Exercise 5.2 The number of ideal X-parameter measurements is the sum of the
number of measurements with one large incident tone and the number of measure-
ments with both large and small probe tones applied simultaneously.
In the former case, for each power and frequency, there are just N Z 1 measurements,
equal to the number of specied complex loads at the fundamental frequency.
For the cases where there are perturbation tones in addition to the large incident
signal, there are two measurements (one for each of the two distinct phases needed to
identify the X (S) and X (T) terms) at each of the second and third harmonics considered in
this exercise.
The sum of these numbers is given by (E.64), which is equivalent to Eq. (5.14):
N X -par N Z 1 22N Z 1 1 22N Z 1 5N Z 1 : E:64
For independent control of the fundamental, second, and third harmonic impedances,
the total number of measurements, per power per frequency, is just the product of the
number of complex loads independently presented at each harmonic, given by (E.65),
which is the result presented in Eq. (5.15):
N harm l-p N Z 1 N Z 2 N Z 3 : E:65

Exercise 5.3 Start from Eq. (5.27b), written in the manifestly time-invariant form
(E.66):
 
Bq, n, m Bq, n, m DCS, jAp1 , 1, 0 j, jAp2 , 0, 1 j, . . . , Ap0 , n, m Pn P m
1, 0 0, 1 P 1, 0 P 0, 1 :
n m

E:66
Expanding the non-analytic function (E.66) for small values of the arguments of all non-
trivial intermodulation products of the fundamental tones, Eq. (E.67) is obtained:
F
Bq, n, m X q, n,m ref LSOPS Pn1,0 Pm 0, 1
2 3
 
X 6 Bq, n, m  B 
 0 0 , n ,  0 0 7
Aq0 , n0 m0  Pn m
Aq0 ,n0 m0  Pn1,0 Pm
q m
4   1, 0 P 0, 1
  0, 1 5P1, 0 P 0, 1
n m
n0 m0  
Aq0 ,n0 m0  Pn1, 0 Pm
0 0
n0 , m0 , q0 Aq0 , n0 m0  P 1, 0 P 0, 1 ref LSOPS 0, 1 ref LSOPS

FB
X q, n, m ref LSOPS Pn1, 0 Pm
0, 1
X h S 0 T
i
mm0  nn0 mm0
X q, n, m;q0 ,n0 ,m0  ref LSOPS Aq0 ,n0 m0  Pnn
1, 0 P 0, 1 X q, n, m;q0 , n0 , m0  ref LSOPS Aq0 ,n0 m0  P 1, 0 P 0, 1 :
n0 , ,
m0 q0

E:67
 Solutions to exercises 215

In (E.67)
 
Bq, n, m  Bq, n, m 
S
 
0 0
X q, n, m;q0 , n0 , m0    Pn1, 0 Pm
0, 1
Pm  Aq0 , n0 , m0  
0 0
Aq0 , n0 m0  Pn
1, 0 0, 1 ref LSOPS ref LSOPS

E:68
and
 
Bq, n, m  Bq, n, m 
T
 
0 0
X q, n, m;q0 , n0 , m0     Pn m
1, 0 P 0, 1 :
Aq0 , n0 m0  Pn1, 0 Pm
0 0  Aq0 , n0 , m0  
0, 1 ref LSOPS ref LSOPS

E:69
Therefore (E.68) and (E.69) are equivalent to Eq. (5.32), and (E.67) is equivalent to
(5.29), for the refLSOPS given by (5.28), reproduced here as (E.70):
 
ref LSOPS DCS p , jAp1 , 1, 0 j, jAp2 , 0, 1 j, 0, . . . , 0 ; E:70

(E.71) has also been used:

 
F
X q, n, m ref LSOPS Bq, n, m DCS, jAp1 , 1, 0 j, jAp2 , 0, 1 j, 0, 0, . . . , 0 : E:71
Index

ACPR, 15961, 167 cross-frequency transfer functions, 62

adiabatic tracking, 152 current bias, 37
adiabatic variation, 172 cutoff frequency, 15
adjacent channel power ratio, see ACPR
admittance matrix, 14 DC behavior, 36
AM-to-AM, 136, 155, 162, 167, 183 DC equations, 29
AM-to-PM, 155, 167, 183 DC operating point, 45
analytic functions, 51 DC response, 23
analyticity, 46 DC stimuli, 23
auxiliary variables, 169 DC-bias stimuli, 23
DC-current bias, 23
bandwidth-limited noise, 162 DC-voltage bias, 23
bare transistors, 112 demodulation, 150
baseband, 150 differential equations, 30
baseband memory, 168, 203 digital pre-distortion, 167
behavioral model, 2, 36, 38 digitally modulated signal, 159
black-box, 16 distortion, 16, 160
hierarchical, 16 harmonic, 18
bias, 168 intermodulation, 18
bias current, 165 source harmonics, 120
bias modulation, 163, 165 distributed representation, 15
bias supply, 165 drain lag, 163
bias voltage, 165 dynamic compression characteristic, 177
biasing scheme, 24 dynamic load lines, 124
bias-line interaction, 152 dynamic range, 100
boundary conditions, 256 dynamic X-parameters, 150

cascadability, 11, 20 efciency, 167

cascade, 25, 27 eight-term error model, 96
chaotic system, 151 electro-thermal effects, 152
coherent cross-frequency phase measurement, 89 ellipse, 78, 84
commensurability condition, 138 envelope analysis, 159
commensurate frequency, 112, 151 envelope domain, 1501, 160
commensurate signal, 302, 34 envelope representation, 154
complementary large-signal step response, 202 envelope simulation, 152
complex amplitudes, 212, 24 equivalent circuit
complex map, 47 linearized elements, 13
composite design, 11 equivalent stimuli, 38
compression, 7, 20, 155, 177, 183 equivalent-time samplers, 92
constant envelope, 80 ergodic principle, 184
constitutive relation, 169 error vector magnitude, see EVM
convolution, 170, 197 EVM, 161
cross-frequency phase, 301, 34, 1401 experimental validation, 158, 169
 Index 217

feedback, 170 large-signal response, 56

fth-order intermodulation, see IM5 large-signal step response, 202
gure of merit, 159, 167 large-signal S-parameters, 68
Fourier analysis, 32 large-signal stimulus, 61
Fourier decomposition, 40 linear memory, 168
Fourier series, 32, 34, 84, 155 linear region of operation, 61
frequency domain, 30, 32 linearity, 167
frequency grid, 20, 22, 30 linearization, 169
functional relationship, 22 linearized map, 45, 50
fundamental frequency, 30, 345, 112, 138 load boundary conditions, 27
load match, 83
gain, 177 load mismatch, 40
generalized power waves, 24 load pull, 112
GSM amplier, 80 active, 112
active source injection, 11819
harmonic balance, 29 fundamental load, 119
harmonic frequency, 45 harmonic impedances, 11920
harmonic grid, 22, 29, 34, 46 hybrid, 112
harmonic indices, 27 independent harmonic impedance tuning, 134
harmonic load pull, 57 nonlinear validation, 120
harmonic superposition, 118, 125 passive, 116, 119
hidden variable, 16970 passive tuners, 112
hot S-parameters, 68 time-domain, 112
hot-S22, 69, 74 uncontrolled harmonic impedances, 119, 135
hysteresis, 179 local oscillator, 113
hysteretic behavior, 169 long-term-evolution signal, see LTE signal
long-term memory, 168
IM3, 157, 166, 186 lower sideband, 160, 162
IM5, 157, 186 LSOP, 489
IMD sweet spot, 179 multi-tone, 113
impulse response, 170, 172 LSOPR, 50
incident waves, 34 LSOPS, 50
incommensurate frequencies, 112, 138 LTE signal, 159
inductor, 165 lumped approximations, 15
input envelope, 160
input matching, 78 McLaurin series, 196, 1989
input reection coefcient, 78 matched device, 6
intermediate frequency matched load, 29
incommensurate, 113 measured data, 158
intermodulation, 139, 151, 157, 167, 177, 187 measurement errors, 105
internal node, 28 memory, 150, 161
interpolation, 118 memory effects, 1623
inter-stage interaction, 64 memory kernel, 168, 171, 174, 176
intrinsic behavior, 34 identication of, 171
IP3, 161 mismatch, 25, 28, 46, 55, 64, 113
iso-thermal characteristics, 183 harmonic, 123
large, 113
Kirchhoffs current law, see KCL multi-frequency, 120
Kirchhoffs voltage law, see KVL small, 45
KCL, 11, 25 mixer, 113, 144
kernel conversion gain, 147
bivariate, 187 isolation, 147
trivariate, 187 LO leakage, 147
KVL, 11, 25 mixer-based systems, 88
model validation, 175
large-signal input reection coefcient, 63 modulation, 150, 160, 165, 168, 177
large-signal operating point, see LSOP monomial, 202
218 Index

multi-harmonic, 34 power-added efciency, see PAE, 124

multi-harmonic spectral maps, 22 power amplier, 159, 167, 169
multi-port network, 38 AM-to-AM, 136
multi-port stimulus, 34, 38 auxiliary for Doherty, 131, 135
multi-tone map, 115 Doherty, 112, 119, 12930
multi-tone stimulus, 139 drain efciency, 137
multi-valuedness, 162 power dissipation, 163
power distribution, 29
narrow-band limit, 160 power transistor, unmatched, 114
noise, 162 power wave, 2
noise oor, 100, 185 probability density function, see PDF
non-analytic functions, 53
non-analytic maps, 46 quasi-static approximation, 151, 154, 158,
non-analytic spectrum, 48 161, 165
non-commensurate signals, 30
nonlinear equations, 27 radio-frequency integrated circuit, see RFIC
nonlinear functions, 35 radio-frequency signal, see RF signal
nonlinear map, 45 real-time samplers, 92
nonlinear mappings, 29 reference impedance, 2
nonlinear measurement calibration, 93 reference LSOP stimulus, 114
nonlinear operation, 64 reference response, 389, 60
nonlinear operators, 35 reference stimulus, 389
nonlinear process, 20 reection coefcient, 113, 116
nonlinear solver, 30 relative phase, 34
nonlinear spectral map, 20, 22, 44 resonances, 177
nonlinear vector network analyzer, response at DC, 23
see NVNA RF interaction terms, 51
NVNA, 91 RF signal, 113
RF solution, 29
offset frequency, 162 RFIC, 162
offset-frequency method, 102 RF-to-DC interaction terms, 51
offset-phase method, 104 Riemann integral, 205
operating point DC, 12 Riemann sum, 204
large-signal, see LSOP
output envelope, 160 sampler-based systems, 91
output power, 166 scalar loss correction, 94
output reection coefcient, 69 scattering parameters, see S-parameters
self-heating, 163, 168, 183
PAE, 124 sensitivity, 55
parallel Hammerstein model, 182, 202 harmonic, 124
partial derivatives, 51, 53 to mismatch, 55
passive load, 28 sensitivity functions, 117, 142
PDF, 168, 184, 204 harmonic load sensitivities, 120
peak-to-average ratio, 169 series expansion, 51, 53
perfect step, 163 short-term memory, 168
perturbation, 46, 51, 117 signal bandwidth, 160
perturbation theory, 45, 51 simulation-based extraction, 106
perturbation tones, 120 small signal, 16
phase reference, 97, 100 small-signal interactions, 52
golden device, 100 small-signal operating conditions, 168
repetitive pulse generator, 100 small-signal response, 56
phase-reference signal, 89, 97 small-signal transfer function, 56
phase shift, 39 source boundary conditions, 26
phasor, 35, 139 source harmonics, 68
physical interpretation, 39 source mismatch, 40, 68
physical meaning, 38, 55 source pull, 128
polynomial, 1967 S-parameter calibration, 94
 Index 219

S-parameter matrix, 4 transistor, 24

S-parameters, 1 above/below turn-on, 133
hot, 17 cutoff, 133
large-signal, 17 GaAs FET, 120
for linear component GaN, 123, 129
denition, 5 GaN HEMT, 123
for nonlinear component output impedance, 133
denition, 7 transistor models, compact, 120
spectral components, 29 transit time, 163
spectral Jacobian, 60 transportable model, 168
spectral linearization, 45, 60, 112, 142 trapping effects, 163, 168
spectral map, 7, 45 turn-on transient, 182
non-analytic, 117 two-stage amplier design, 64
spectral mapping, 20, 55 two-tone excitation, 161
spectral re-growth, 179 two-tone experiment, 158, 175
spectrum analyzer, 161 two-tone measurement, 160
state space, 36 two-tone signal, 181
steady state, 27, 36, 45, 152 two-tone simulation, 166
step excitation, 181
step response, 172, 184 upper sideband, 160, 162
step transition, 172
stochastic signal, 184 validation, 175
superposition, 8 vector network analyzer, see VNA
sweet spot, 179 vector signal analyzer, 168
symmetry conditions, 157 vector signal generator, 131
VNA, 88
temperature change, 163 voltage bias, 37
ten-term error model, 95 Volterra theory, 1967
test stimulus, 39
thermal capacitance, 163 wave variables, 2
thermal circuit, 163 WCDMA, 181
thermal relaxation, 163, 183 wide-band memory approximation, 204
thermal resistance, 163 wide-band modulation, 150, 182, 184, 204
thermal response, 183 wide-band X-parameters, 185
thermal transient, 172
third-order intercept, 161 .xnp les, 106
third-order intermodulation, see IM3 X-parameter correction, 106
time derivative, 172 X-parameter les, 106
time invariance, 10, 20, 24, 356, 197 X-parameter measurement, 101
time reference, 33 offset-frequency method, 102
time shift, 34, 39, 115 offset-phase method, 104
time domain, 29, 32 X-parameters
time-invariant components, 45 arbitrary load-dependent, 112
time-invariant map, 44 load-dependent, 119
time-invariant systems, 356 measurement hardware, 88
time-varying envelope, 152 type S, 52
trajectory, 72 type T, 52
transfer functions, 62
transient measurement, 168 Y-parameters, 14