Sunteți pe pagina 1din 254

WARSAW UNIVERSITY

OF TECHNOLOGY
Faculty of Electronics and Information
Systems

Ph.D. Thesis
Arkadiusz Lewandowski
Multi-frequency approach to vector-network-analyzer
scattering-parameter measurements

Supervisor
Professor Janusz Dobrowolski, Ph.D., D.Sc.

Warsaw, 2010

Abstract
Vector network analyzer (VNA) is the basic measurement instrument used in the characterization of microwave and millimeter-wave electronic circuits and systems. Much eort
has been put throughout the past three decades in improving the designs of VNA instrumentation and in establishing the principles of VNA calibration and uncertainty analysis
of VNA measurements. Modern VNAs are a culmination of this long standing research,
and are sophisticated, mature and reliable measurement instruments, commonly employed
in the industry and laboratories.
Recently, however, several new trends in the vector network-analysis started to emerge.
These new trends result from an increased interest in the application of millimeter- and
sub-millimeter-wave signals (frequencies up to 1 THz), rapid development of the nanotechnology, requiring characterization of structures with very large impedances (on the order of
100 k), and an increased demand for large-signal characterization of microwave circuits.
These new trends result, on one hand, in new concepts in the design of the VNA instrumentation, such as special VNA extension units, allowing the conventional VNAs to operate
up to 500 GHz, microwave scanning microscopes, or nonlinear vector network analyzers
(NVNA). On the other hand, these trends lead to new challenging demands regarding the
measurement accuracy and its reliable and complete evaluation.
The multi-frequency approach introduced in this work addresses this last issue. The
principle of this approach is to account for the relationships between scattering parameter
measurements at dierent frequencies. We show that this new approach allows to reduce by
several times the impact of errors in the description of calibration standards, resulting thus
in a signicant improvement of the VNA measurement accuracy. We further demonstrate
that the multi-frequency approach to the description of VNA instrumentation errors yields
better understanding of their physical origins, leading to their compact description based
on the stochastic modeling. We nally show that the multi-frequency representation of
the uncertainty in VNA scattering-parameter measurements is essential when using these
measurements in the calibration of time-domain measurement systems, such as high-speed
sampling oscilloscopes, or nonlinear vector network analyzers.

iii

Streszczenie
Wektorowy analizator obwodw (ang. Vector Network Analyzer - VNA) jest podstawowym urzdzeniem wykorzystywanym do charakteryzowania ukadw i systemw elektronicznych wielkiej czstotliwoci (w. cz). Konstrukcja wspczesnych analizatorw, jak
i metody wykorzystywane w ich kalibracji oraz analizie niepewnoci pomiaru, s owocem
wieloletnich prac badawczych oraz intensywnego rozwoju technologicznego. W konsekwencji nowoczesne wektorowe analizatory obwodw odznaczaj si niezwykle zaawansowanymi
i dojrzaymi rozwizaniami technicznymi oraz s z powodzeniem wykorzystywane w codziennej praktyce zarwno laboratoriw pomiarowych jak i przemysu ukadw w. cz.
W wektorowej analizie obwodw pojawiy si w ostatnim czasie nowe kierunki rozwoju, wynikajce z rosncego zainteresowania wykorzystaniem sygnaw w zakresie fal
milimetrowych i submilimetrowych (czstotliwoci blisko 1 THz), rozszerzania si zakresu
impedancji mierzonych struktur w. cz. (impedancje rzdu 100 k), zwizanego z intensywnym rozwojem nanotechnologii, oraz z zapotrzebowania na charakteryzowanie wielkosygnaowych wasnoci ukadw w. cz. Te nowe zastosowania wektorowej analizy obwodw
prowadz, z jednej strony, do nowych rozwiza konstrukcyjnych, jak na przykad gowice
powielajco-mieszajce rozszerzajce zakres pracy typowych analizatorw do czstotliwoci
rzdu 500 GHz, mikrofalowe mikroskopy skaningowe, czy te wielkosygnaowe wektorowe
analizatory obwodw (ang. Nonlinear Vector Network Analyzer-NVNA). Z drugiej strony,
stawiaj one zupenie nowe wyzwania, jeeli chodzi o dokadno pomiaru, oraz jej wiarygodne oszacowanie.
Przedstawione w niniejszej pracy nowatorskie wieloczstotliwociowe podejcie do pomiaru parametrw rozproszenia za pomoca wektorowego analizatora obwodw jest prb
odpowiedzi na te nowe wyzwania. Jego istot jest uwzgldnienie relacji midzy pomiarami
parametrw rozproszenia na rnych czstotliwociach. W pracy wykazano, e to nowe
ujcie pozwala kilkukrotnie zmniejszy wpyw bdw wynikajcych z niedokadnego opisu
wzorcw kalibracyjnych, a tym samym znaczco zwikszy dokadno pomiaru. Pokazano
rwnie, e wieloczstotliwociowy opisu bdw losowych w pomiarach analizatorem wektorowym pozwala lepiej wyjani ich zyczne przyczyny, prowadzc do prostego i spjnego
opisu tych bdw opartego na modelowaniu stochastyczym. W kocu, w pracy wykazano,
e uoglniony wieloczstotliwociowy opis niepewnoci pomiaru parametrw rozproszenia
jest niezbdny, gdy wykorzystuje si te pomiary w kalibracji urzdze dziaajcych w dziedzinie czasu, takich jak szybkie oscyloskopy prbkujce, albo wielkosygnaowe wektorowe
analizatory obwodw.
v

Wenn [meine] Arbeit einen Wert


hat, so besteht er [...] darin, dass in
ihr Gedanken ausgedrckt sind,
und dieser Wert wird umso grer
sein, je besser die Gedanken
ausgedrckt sind.
Ludwig Wittgenstein
Meine Resultate kenne ich lngst,
ich wei nur noch nicht,
wie ich zu ihnen gelangen soll.
Carl Friedrich Gauss

vii

Acknowledgment
This research project would not have been possible without the support of many people and institutions. First of all, I would like to thank to my advisor Dr. Dylan Williams
from the National Institute of Standards and Technology (NIST), Boulder, USA, for many
fruitful discussions and his continuous support during my ve years long stay at NIST. I
wish also to express gratitude to my supervisor at the Warsaw University of Technology,
Prof. Janusz Dobrowolski for his constant help and patience during the long period in which
this work was written. My gratitude is also due to Dr. Wojciech Wiatr for his encouragement and many invaluable advices without which this work have not been accomplished.
I would like also to acknowledge Denis LeGolvan of NIST, Boulder, USA, for introducing me into the world of coaxial connectors, and for his enormous help with the measurements. I would also like to convey thanks to Grzegorz Kdzierski and Karol Korsze of
the National Institue of Telecommunications, Warsaw, Poland, for performing the Type-N
measurements described in this work.
Special thanks is also due to all of my colleges in the Electromagnetics Division, NIST,
Boulder, and at the Institute of Electronic Systems, Warsaw, Poland, for their constant
support throughout the entire time in which this project was carried out.
My deepest gratitude is also due to my family for their love, patience, and understanding
without which nishing this work would not have been possible.
Last, but not least, I would like to acknowledge the Polish Ministry of Science and
Higher Education for the grant N N517 4394 33 from which this work was partially funded.

ix

Contents
Abstract

iii

Streszczenie

Acknowledgment

ix

Nomenclature

xx

1 Introduction

1.1

Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2

Previous research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.3

Objective and scope of this work

. . . . . . . . . . . . . . . . . . . . . . .

2 Principles of VNA S-parameter measurements


2.1

Denition of S -parameters . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1.1

Waveguide voltage, current and characteristic impedance . . . . . .

2.1.2

Wave amplitudes and scattering parameters . . . . . . . . . . . . .

11

2.1.3

Practical implications . . . . . . . . . . . . . . . . . . . . . . . . . .

16

2.2

VNA S -parameter measurement

. . . . . . . . . . . . . . . . . . . . . . .

17

2.3

Two-port VNA mathematical models . . . . . . . . . . . . . . . . . . . . .

20

2.3.1

Linear time-invariant two-port VNA . . . . . . . . . . . . . . . . .

20

2.3.2

Modeling VNA nonstationarity . . . . . . . . . . . . . . . . . . . .

28

Two-port VNA calibration techniques . . . . . . . . . . . . . . . . . . . . .

30

2.4.1

Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

32

2.4.2

Standards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

39

2.4.3

Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

42

2.4

xi

CONTENTS
3 Overview of uncertainty analysis for VNA S-parameter measurements

43

3.1

Sources of error in corrected VNA S -parameter measurements . . . . . . .

44

3.2

Statistical description of S -parameter measurement errors . . . . . . . . . .

45

3.2.1

Statistical model for S -parameter measurement . . . . . . . . . . .

45

3.2.2

Error description for a single S -parameter . . . . . . . . . . . . . .

46

3.2.3

Error description for a matrix of S -parameters . . . . . . . . . . . .

49

3.3

3.4

3.5

3.6

Statistical models for errors in corrected VNA S -parameter measurements

50

3.3.1

Systematic errors . . . . . . . . . . . . . . . . . . . . . . . . . . . .

51

3.3.2

Random errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

51

Representation of errors in corrected VNA S -parameter measurements . . .

53

3.4.1

Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

53

3.4.2

Errors in VNA calibration coecients . . . . . . . . . . . . . . . . .

54

3.4.3

Errors in VNA raw measurements . . . . . . . . . . . . . . . . . . .

58

Approximate uncertainty evaluation . . . . . . . . . . . . . . . . . . . . . .

61

3.5.1

Ripple analysis techniques . . . . . . . . . . . . . . . . . . . . . . .

62

3.5.2

Calibration comparison method . . . . . . . . . . . . . . . . . . . .

62

3.5.3

Statistical residual analysis . . . . . . . . . . . . . . . . . . . . . . .

63

Complete uncertainty evaluation . . . . . . . . . . . . . . . . . . . . . . . .

64

3.6.1

Linear error propagation . . . . . . . . . . . . . . . . . . . . . . . .

65

3.6.2

Monte-Carlo simulation

65

. . . . . . . . . . . . . . . . . . . . . . . .

4 Multi-frequency description of S-parameter measurement errors

67

4.1

Statistical model for the multi-frequency S-parameter measurement . . . .

68

4.2

The notion of a physical error mechanism . . . . . . . . . . . . . . . . . . .

68

4.3

Statistical properties of the multi-frequency measurement error . . . . . . .

71

4.3.1

Probability distribution function . . . . . . . . . . . . . . . . . . . .

71

4.3.2

Uncertainty reporting . . . . . . . . . . . . . . . . . . . . . . . . . .

74

4.3.3

Multi-frequency covariance-matrix structure . . . . . . . . . . . . .

75

Physical error mechanisms in VNA S -parameter measurements . . . . . . .

75

4.4.1

Calibration standard errors . . . . . . . . . . . . . . . . . . . . . .

75

4.4.2

VNA instrumentation errors . . . . . . . . . . . . . . . . . . . . . .

76

Practical implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

78

4.5.1

Time-domain waveform correction . . . . . . . . . . . . . . . . . . .

78

4.5.2

Device modeling based on S -parameter measurements . . . . . . . .

83

4.5.3

Error-mechanism-based VNA calibration . . . . . . . . . . . . . . .

84

4.4

4.5

xii

CONTENTS
4.6

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5 Generalized multi-frequency VNA calibration

87
89

5.1

Formulation of the VNA calibration problem . . . . . . . . . . . . . . . . .

90

5.2

Coaxial multi-line VNA calibration . . . . . . . . . . . . . . . . . . . . . .

92

5.2.1

Classical multi-line VNA calibration . . . . . . . . . . . . . . . . .

93

5.2.2

Coaxial air-dielectric line as a calibration standard . . . . . . . . .

93

5.2.3

Computational aspects . . . . . . . . . . . . . . . . . . . . . . . . .

94

Errors in the coaxial multi-line VNA calibration . . . . . . . . . . . . . . .

95

5.3.1

Variation of connector-interface electrical parameters . . . . . . . .

95

5.3.2

Variation of lines characteristic impedance and propagation constant 100

5.3.3

Line length error . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

5.3.4

Reect asymmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

5.3

5.4

5.5

Calibration standard models . . . . . . . . . . . . . . . . . . . . . . . . . . 106


5.4.1

Lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

5.4.2

Reect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

5.4.3

Thru . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

Solution uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111


5.5.1

Mathematical framework . . . . . . . . . . . . . . . . . . . . . . . . 111

5.5.2

Relationships between estimated parameters . . . . . . . . . . . . . 112

5.5.3

Optimal constraint choice . . . . . . . . . . . . . . . . . . . . . . . 114

5.6

Numerical solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

5.7

Residual analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

5.8

Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

5.9

5.8.1

Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

5.8.2

Type-N coaxial connector . . . . . . . . . . . . . . . . . . . . . . . 119

5.8.3

1.85 mm coaxial connector . . . . . . . . . . . . . . . . . . . . . . . 126

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

6 Multi-frequency stochastic modeling of VNA nonstationarity errors

137

6.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

6.2

Generic physical model for the VNA nonstationarity . . . . . . . . . . . . . 139

6.3

Stochastic model for connector nonrepeatability and cable instability . . . 144


6.3.1

Statistical properties of circuit parameters . . . . . . . . . . . . . . 144

6.3.2

Estimation of the covariance matrix of circuit parameters . . . . . . 145


xiii

CONTENTS
6.3.3
6.4

6.5

Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

Stochastic model for VNA test-set drift . . . . . . . . . . . . . . . . . . . . 157


6.4.1

Drift as the multidimensional random walk . . . . . . . . . . . . . . 157

6.4.2

Estimation of the process covariance matrix . . . . . . . . . . . . . 159

6.4.3

Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

7 Conclusions

167

A Real-valued representation of complex vectors and matrices

173

B Maximum likelihood approach to system identication

175

B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175


B.2 Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
B.2.1 Errors in system responses . . . . . . . . . . . . . . . . . . . . . . . 177
B.2.2 Errors in system responses and excitations . . . . . . . . . . . . . . 182
B.3 Covariance matrix of the estimates . . . . . . . . . . . . . . . . . . . . . . 186
B.3.1 Errors in system responses . . . . . . . . . . . . . . . . . . . . . . . 186
B.3.2 Errors in system responses and excitations . . . . . . . . . . . . . . 188
B.4 Numerical solution techniques . . . . . . . . . . . . . . . . . . . . . . . . . 189
B.4.1 Errors in system responses . . . . . . . . . . . . . . . . . . . . . . . 189
B.4.2 Errors in system responses and excitations . . . . . . . . . . . . . . 191
B.5 Solution uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
B.6 Systems with complex-valued inputs and outputs . . . . . . . . . . . . . . 191
C Air-dielectric coaxial transmission line

195

C.1 Innite metal conductivity . . . . . . . . . . . . . . . . . . . . . . . . . . . 195


C.2 Finite metal conductivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
D Center-conductor gap impedance

199

D.1 Innite metal conductivity . . . . . . . . . . . . . . . . . . . . . . . . . . . 199


D.2 Finite metal conductivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
D.3 Finger eect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
E Slightly nonuniform coaxial transmission line

205

F Small changes of two-ports scattering parameters

209

xiv

CONTENTS
G Estimation of VNA nonstationarity model parameters

213

H Vector stochastic Wiener process

217

Bibliography

221

xv

CONTENTS

xvi

Nomenclature
a, b, c
A, B, C
{ai }N
i=1
T
A
AH
AB
vec(A)
Re x
Im x
a

vectors
matrices
set of vectors a1 , . . . , aN
transpose of the matrix A
conjugate transpose (Hermitian transpose) of the matrix A
Kronecker product of the matrices A and B
vector representation [a11 , a21 , a31 , . . . , a12 , a22 , a32 , . . .]T of the matrix A
real part of x
imaginary part of x
real-valued representation [Re a1 , Im a1 , Re a2 , Im a2 , . . .]T of the complex-valued
vector a
x
estimate of x

x
true value of x
x
measurement of x
x
error in the measurement of x
E(x)
expectation value of x
Var(x)
variance of x
Cov(x, y) covariance of x and y

attenuation constant, i.e., Re

phase constant, i.e., Im

sought parameters of the VNA and calibration standards


0
free-space phase constant
c
vector of calibration-standard unknown parameters
C
capacitance
c
speed of light in vacuum
c0
vector of calibration-standard known parameters
xvii

CONTENTS
C
D
d
dp
DUT
e
EDF
EDF
EDF
EDF
EDF
EDF
0
r

f
g

In
J
K
k
L
L
Lg
l
li
lo
l
l
M
0
N

normalized capacitance
diameter of the outer conductor in the coaxial transmission line
diameter of the inner conductor in the coaxial transmission line
center-conductor-pin diameter in the coaxial transmission line
device under test
eccentricity of the inner conductor in the coaxial transmission line
forward directivity
reverse directivity
forward tracking
reverse tracking
forward source match
reverse source match
dielectric permittivity of vacuum
relative dielectric permittivity
physical error mechanism
frequency; probability density function
center-conductor-gap width in the coaxial transmission line
complex propagation constant
reection coecient
identity matrix of size n n
Jacobian matrix
number of frequencies
frequency index
inductance
normalized inductance
normalized gap inductance per-unit-length
transmission line length
inner conductor length
outer conductor length
misalignment of outer and inner conductor symmetry axes
loss correction factor
number of mechanisms
magnetic permeability of vacuum
number of calibration standards

xviii

CONTENTS
n

p
PDF
r
R
R
0
rDF
rDF
rDF
rDF
rDF
rDF
S
s
sm

SOLT
SOLT
T
TRL
v
VNA
V
w
Y
Z
Z0
Z00
Y

calibration standard index


angular frequency
vector of VNA calibration coecients
probability density function
vector of residuals
resistance
normalized resistance
characteristic-impedance correction factor
eective forward directivity
eective reverse directivity
eective forward tracking
eective reverse tracking
eective forward source match
eective reverse source match
scattering matrix
vector representation vec(s) of the scattering matrix S; vector of calibrationstandard S-parameter-denitions
vector of raw measurements of calibration standards
covariance matrix
conductivity, standard deviation
short open load through
Singular Value Decomposition
transmission matrix
through reect line
phase velocity
vector network analyzer
weight matrix in the VNA calibration
width of the in-cut between the connector socket ngers
admittance
impedance
characteristic impedance for the TEM mode in a lossy coaxial transmission line
characteristic impedance for the TEM mode in a lossless coaxial transmission
line
normalized admittance

xix

NOMENCLATURE
Z
Zref

normalized impedance
reference impedance

xx

Chapter 1
Introduction
One day when Pooh Bear had nothing else to
do, he thought he would do something [...].
A. A. Milne, House at the Pooh corner

1.1

Motivation

Vector network analyzer (VNA) is the basic measurement instrument for characterization of microwave and millimeter-wave electronic circuits. The VNA measures scattering
parameters (S-parameters) which constitute a complete description of small-signal deterministic properties of an electronic circuit [1]. This measurement is typically performed in
a broad frequency range, starting from tens of kHz and reaching even hundreds of GHz [2
5]. The VNA measured S-parameters, along with noise parameters, are then traditionally
used in the design and testing of both single components and complex systems working at
microwave and millimeter-wave frequencies.
Much eort has been put throughout the past 30 years in establishing the principles of
vector-network analysis and improving the VNA instrumentation. A good review of this
research can be found in [68]. Modern VNAs, such as [35], are a culmination of this long
standing research and are very mature and reliable measurement instruments, commonly
employed in the industry and laboratories.
Recently, however, several new trends in the vector network-analysis started to emerge.
These trends push the boundaries of the conventional VNAs with respect to the maxi1

1. INTRODUCTION
mum measurement frequency, impedance level of the device under test (DUT), and the
assumption of the DUT linearity.
The trend to extend the frequency range of modern VNAs results from an increased
interest in the use of signals with frequencies in the millimeter- and sub-millimeter-wave
range. Examples are high-capacity data transmission systems [9], millimeter-wave radar
system [10, 11], radiotelescopes [12, 13], or the broad range of terahertz applications [14].
Eorts to extend the maximum VNA measurement frequency have recently brought about
special VNA extension units, allowing the conventional VNAs to operate up to 500 GHz
with rectangular waveguide connectors [2]. Extension up to 1 THz is likely to happen in
the nearest future [15, 16].
Accurate VNA S-parameter measurement at millimeter- and sub-millimeter-wave frequencies, however, is very challenging due to some specic error sources negligible at lower
frequencies. Due to small wavelength, these measurements require the use of waveguides
with aperture size below 1 mm in order to avoid overmoding. While it is possible to manufacture such waveguides with quite a high precision, some irregularities, such as rounding
of the waveguide corners or erosion of the leading edges of the waveguide apertures, are
unavoidable and may lead to signicant systematic errors in the VNA calibration [17].
Furthermore, the connection of two waveguide anges with such small apertures requires
very precise alignment. Although some special alignment solutions, involving the use of
multiple alignment pins, have been devised, the random errors due to ange misalignment
can still signicantly deteriorate the measurement accuracy [18]. Finally, the noise uctuations of the VNA test-signals are another important source of errors in VNA S-parameter
measurement at sub-millimeter-wave frequencies. These uctuations, caused by the thermal and phase noise originating in the frequency multiplication and sub-harmonic mixing
circuitry of the VNA extension units, signicantly reduce the dynamic range and increase
the short-term instability (also referred to as the trace jitter [19]), as compared with
VNAs operating at lower frequencies.
While the VNA calibration techniques used at lower frequencies can be adapted to
work with millimeter and sub-millimeter-wave VNAs, due to those specic errors, their
accuracy is often not satisfactory [18, 20]. Thus, new more accurate calibration techniques,
less sensitive to those specic error sources, need to be devised.
Similar challenges regarding the measurement accuracy are encountered in VNA S-parameter measurements of devices whose impedance diers signicantly from the typical
VNA impedance level of 50 . Examples of such devices are nanotubes, nanowires or
2

1.1. MOTIVATION
metamaterials, which may exhibit impedances in the range of tens and hundreds of ks,
or fractions of an [21]. Since the VNA test-ports are typically built based on 50 transmission-line components, the energy coupling to devices with impedance level close
to 50 is very good. Thus, such devices are measured with the highest accuracy. However,
when the DUT impedance is much smaller or much larger than this value, only a little
signal is coupled and most of the signal is reected back to the VNA. While some special
VNA architectures (e.g. [22]), aiming at increasing the VNA receiver resolution, may help
to improve the measurement accuracy in such cases, new VNA calibration techniques, less
sensitive to the VNA measurement errors (e.g., [23]), are needed.
The last trend in the modern vector-network-analysis results from an increased demand
for the characterization of active high-frequency circuits in the large-signal regime. Accurate large-signal characterization of such circuits is essential in the design and testing of
various applications. Examples are portable data transmission systems where the high
power eciency (hence, long battery life) needs to be combined with a minimal nonlinear distortion to the transmitted signal, or active high-frequency circuits such as signal
generators, mixers or frequency multipliers, that are by nature operating in the nonlinear
regime.
Characterization of large-signal properties of high-frequency circuits poses multiple difcult problems. It requires specialized instrumentation, such as nonlinear vector-networkanalyzer (NVNA), also referred to as large-signal network analyzer (LSNA) [24, 25]. The
NVNA characterizes the nonlinear DUT properties in terms of either voltages and currents,
or wave quantities, such as X-parameters [26, 27] or S-functions [25, 28]. The characterization is performed at the principal frequency and its harmonics. The measurement is
then either directly used, for example, in the circuit simulator, or converted into the time
domain in order to analyze the shape of the voltage or current waveforms.
Accurate NVNA measurements require a specialized calibration procedure. This procedure, apart from the traditional linear VNA calibration, involves also power and phase
calibration. The power calibration is required so as to enable the measurement of absolute
quantities (voltages, currents or wave quantities). The phase calibration is necessary because in the NVNA measurement one is not only interested in magnitudes and phases of
voltages and currents (or wave quantities) at each frequency, but also in the phase relationships between those quantities. These relationships are essential, for example, when reconstructing the time-domain voltage and current waveforms from the NVNA measurements.
As a result, the accuracy assessment of NVNA measurements requires new uncertainty
3

1. INTRODUCTION
analysis approaches that account not only for uncertainties at a single frequency, but also
for the statistical correlations between uncertainties at dierent frequencies.
Consequently, the new trends in the vector network analysis discussed above, while
stimulating the development of new hardware solutions, lead also to new, more stringent
demands as to the measurement accuracy and its reliable and complete evaluation. The
multi-frequency approach presented in this work addresses this issue.

1.2

Previous research

Enhancement of VNA measurement accuracy and its more reliable and complete evaluation have always been stimulating the development of VNA measurement techniques. A
detailed review of this development can be found in [68, 19]. Here we shall indicate the
most important turning points in this development, which will allow us to better understand the origins of the multi-frequency approach proposed in this work.
The rst turning point was the invention of the self-calibration methods. The idea
of self-calibration in the two-port VNA calibration problem had rst been employed in
Engens TRL method [29] and was then generalized by Eul and Schiek [30]. The concept
of self-calibration in one-port VNA calibration methods appears also in papers by Wiatr
[3133] and Bianco [34]. The principle of self-calibration is to use calibration standards
that are only partially known and to determine their complete S-parameter description
along with the VNA calibration coecients. For example, in the TRL method, the transmission line is used with known length and unknown propagation constant. Consequently,
the contribution of systematic errors in calibration standard denition can be reduced,
since instead of the specic numerical values of calibration standard parameters, which are
inevitably subject to measurement errors, the information as to the relationships between
these parameters is used.
Another turning point was the application of statistical methods in VNA calibration
problem. This approach was initiated in the case of one-port VNA calibration in [31, 32, 35]
and in the case of two-port VNA calibration methods in [36]. The application of statistical
methods in VNA calibration is based on the use of redundant calibration standards and
statistical processing of the resulting overdetermined set of equations. Consequently, the
contribution of random measurement errors can be signicantly reduced.
Another important paradigm change in the development of VNA calibration methods
was initiated in [37] and [38]. In these references, for the rst time, the relationships
4

1.3. OBJECTIVE AND SCOPE OF THIS WORK


between the calibration standard parameters at dierent frequencies are exploited in the
VNA calibration. The experimental result presented in [37] and [38] indicate that this
improves the accuracy and reliability of the VNA calibration.
The multi-frequency description of S-parameter measurement errors naturally complements a VNA calibration approach that accounts for the relationships between S-parameter
measurements at dierent frequencies. Such a description was introduced in [39], and is
based on the covariance-matrix representation which had already been used in the uncertainty analysis of single-frequency S-parameter measurements in dual six-port measurement systems [40, 41], and then recently rediscovered in the context of uncertainty evaluation in VNA S-parameter measurements [42, 43]. The generalized covariance-matrix
description proposed in [39] uses additional terms in the covariance matrix in order to
account for statistical correlations between uncertainties at dierent frequencies. These
correlations have been shown to be essential when applying the VNA S-parameter measurements in the calibration of time-domain measurement systems [39, 44].

1.3

Objective and scope of this work

As pointed out above, accounting for the relationships between VNA S-parameter measurements at dierent frequencies can be benecial in terms of increased measurement accuracy (see [37, 38]) and its more complete evaluation (see [39, 44]). The objective of this
work is to generalize these results by developing a comprehensive multi-frequency approach
to VNA S-parameter measurements.
We shall attain this objective in two step. In the rst step, we will develop a mathematical description of the relationships between VNA measurement at dierent frequencies
which unies the descriptions used in the calibration approaches of [37, 38] and in the uncertainty analysis of [39]. With the use of this generalized description, in the second step,
we will investigate the benets which could be gained by accounting for those relationships
at various stages of the VNA measurement procedure. A particular emphasis will be put
here on the VNA calibration.
This organization of this work is as follows. In the introductory part (Chapter 2 and
Chapter 3) we review the foundations of VNA S-parameter measurements and uncertainty
analysis. This part serves as the theoretical background for the discussion presented in the
main part of this work.
The main part of this work consists of three chapters. In Chapter 4 we develop a uniform
5

1. INTRODUCTION
framework for the representation of relationships between VNA S-parameter measurements
at dierent frequencies. We further review the practical applications for which accounting
for these relationships is important. These applications include the correction of timedomain measurements, measurement-based device modeling and the VNA calibration on
which we focus in the this work. We show that a statistically sound description of the VNA
calibration problem should be done in terms of the error mechanisms underlying the calibration standard and VNA instrumentation errors. As a consequence, the VNA calibration
should be performed jointly at all measurement frequencies so as to account for the simultaneous contribution of those error mechanisms to S-parameter measurements at dierent
frequencies. We refer to this approach as the error-mechanism-based VNA calibration and
in the remainder of this work we develop the necessary tools for the implementation of
such a calibration approach. These tools include the generalized multi-frequency VNA
calibration (see Chapter 5) and the framework for error-mechanism-based description of
the VNA nonstationarity errors (see Chapter 6).
In the last part of this work (see Chapter 7) we present conclusions and discuss possible
directions of further research.

Chapter 2
Principles of VNA S-parameter
measurements
All models are wrong, some are useful.
George Box

In this chapter, we review the principles of S-parameter measurements with the vector
network analyzer (VNA). We begin with a brief review of the S-parameter denition.
Following on that, we discuss the two-port VNA S-parameter measurements, and analyze
the imperfections of a typical two-port VNA measurement setup. The errors caused by
these imperfections are systematic as they are very stable in the course of typical VNA
measurement. Therefore, they can be characterized in a calibration procedure and then
removed from the actual S-parameter measurements in the correction procedure. Both
procedures assume a mathematical model of these VNA. In the calibration procedure, a
set of devices with some known characteristics is measured and the parameters of the
VNA model are determined. Then, in the correction procedure, the model obtained in the
calibration is used to correct for the imperfections of the VNA setup. We discuss dierent
mathematical VNA models and VNA calibration techniques in the last two sections of this
chapter.
7

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS

2.1

Denition of S-parameters

Scattering parameters form a description of an electronic circuit in terms of complex


amplitudes1 of electromagnetic waves interacting with the circuit. This description is
used in the situation when the dimensions of the circuit are become comparable with the
wavelength, which typically takes place at microwave and millimeter-wave frequencies. In
this situation, the physical phenomena occurring in the circuit are of a wave nature, and
the conventional circuit description in terms of terminal voltages and currents looses its
physical correspondence.
Scattering parameter description uses the concept of the circuit port instead of the
circuit terminal. A circuit port is dened as section of a uniform arbitrary waveguide
through which the electromagnetic wave may enter and exit the electronic circuit. In
order to ensure the uniqueness of the scattering parameter description, we require that
these ports are electromagnetically separated and that a given set of circuit ports (with
waveguide modes propagating through them) encompasses all of the possible means by
which electromagnetic energy can enter and leave the circuit. This means that we need to
account not only for all of the physical ports through which the electromagnetic waves are
interacting with the circuit, but also for all of the waveguide modes propagating through
the circuit ports. In order to simplify the scattering parameter description, we typically
assume single mode propagation through circuit ports, however, extension to the case of
multiple modes is possible.
Scattering parameters describe the relationships between the amplitudes of the waves
propagating through the circuit ports. These amplitudes are dened with the use of a
simple normalization such that the wave with a unit root-mean-square amplitude, in the
absence of the wave propagating in the opposite direction, carries unit power. Although
the principle of this normalization is very simple, its systematic derivation requires some
consideration. In the following, we briey review the origins of this normalization (for
more details refer to, e.g., [1, 45]). We rst introduce the concepts of waveguide voltage,
current and characteristic impedance. These concepts, although not required for the definition of normalized waves and, consequently, of S-parameters, allow one the relate the
scattering parameter description to the methods of the transmission line theory. We then
present two dierent types of normalized waves. The rst type, referred to as the traveling
waves, originates in the physics of wave propagation in the waveguide. Thus, properties of
1

We assume time-harmonic dependence of the elds.

2.1. DEFINITION OF S -PARAMETERS


traveling waves closely reect the physical properties of the actual waves propagating in
the waveguide. In some cases, however, these properties lead to results that are surprising
in the context of the transmission line theory. Consequently, another type of normalized
waves, referred to as pseudo-waves, is introduced, which leads to more intuitive results in
the framework of transmission line theory, at a cost, however, of not as close correspondence to the physics of wave propagation in the waveguide. We conclude with a discussion
of practical implications of the dierent normalization schemes for scattering parameter
measurements.

2.1.1

Waveguide voltage, current and characteristic impedance

Electromagnetic waves traveling in a waveguide are described in terms of modes which


are solutions to the Maxwell equations in the waveguide cross-section. For time-harmonic
eld dependence, these solutions can be characterized by the normalized transverse electric
and transverse magnetic eld distributions, et (x, y) and ht (x, y), respectively, and the
complex-valued propagation constant . In the case of lossless transmission lines, the eld
distributions are real-valued, and the propagation constant is imaginary = j. In the
case of transmission lines with losses, the eld distributions are in general complex-valued
and the propagation constant has also a real part, that is = + j.
With the use of the normalized eld distributions and the propagation constant, we can
write the complex peak amplitudes of the elds at any point in the waveguide (propagation
occurs along the z axis) in a normalized way as
V (z)
et (x, y),
V0
I(z)
=
ht (x, y)
I0

Et (x, y, z) = C + et (x, y)ez + C et (x, y)e+z =

(2.1)

Ht (x, y, z) = C + ht (x, y)ez + C ht (x, y)e+z

(2.2)

where V0 and I0 are normalization constants with the dimension of voltage and current,
respectively, C + and C are unitless constants specifying the amplitude of the forward
and backward propagating wave at z = 0, respectively, and V (z) and I(z) are dened as
waveguide voltage and current.
The unitless constants C + and C depend on the normalization used for et (x, y) and
ht (x, y). The waveguide voltage and current, however, are independent of this normalization due to the use of normalization constants V 0 and I 0 . These constants have units of
voltage and current, respectively, hence in the following we refer to them as the normal9

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


ization voltage and current, respectively.
The choice of normalization voltage V0 and current I 0 is not arbitrary and we require
that
1
1

P 0 = V0 I0 =
et ht dS
(2.3)
2
2 S
where the superscript indicates the complex conjugate, and S denotes the cross-section
of the waveguide. From (2.3) it follows that the net power ow in the waveguide is
1
1
P (z) =
Et (x, y, z) Ht (x, y, z)dS = V (z)I (z)
2 S
2


S

et ht dS
1
= V (z)I (z). (2.4)

V0 I0
2

Consequently, by imposing the condition (2.3), we require that the net power owing
through the cross-section of the waveguide can be determined applying conventional circuittheory denition to the waveguide voltage and current. Note that the magnitude of P0
depends on the normalizations used for et (x, y) and ht (x, y), however, its phase is independent of this normalizations and is an inherent property of the mode.
Due to the power constraint (2.3), only one of the constants V 0 and I 0 can be arbitrarily
chosen. For example, for the voltage V 0 , we may choose to use the path integral along
some arbitrarily chosen path P in the cross-section of the waveguide


V0 =

et (x, y)dl,

(2.5)

with an obvious constraint V0 = 0, and then determine I0 from (2.3) (voltage-power


normalization). Alternatively, we can also x the current I0 based the loop integral around
a closed loop L in the cross-section of the waveguide


I0 =

ht (x, y)dl,

(2.6)

with a similar constraint I0 = 0 and then determine V0 from the constraint (2.3) (currentpower normalization). Other normalizations are also possible, e.g., [1] or [46].
Based on the denition of the normalization voltage V0 and current I0 , we dene the
characteristic impedance of the mode
Z0 =

V0
.
I0

(2.7)

The magnitude of Z0 depends, in general, on the eld normalizations and the chosen
strategy for setting the constants V0 and I0 . The phase of Z0 can be easily determined in
10

2.1. DEFINITION OF S -PARAMETERS


terms of the mode power P0 . Indeed, after some simple transformations, we obtain
ImZ0
ImP0
=
ReZ0
ReP0

(2.8)

Phase of Z0 is therefore the inherent property of the mode and does not depend on the
eld normalizations and the constants V0 and I0 .
For TEM modes, the integral (2.5) depends only on the end points of the path and it is
natural to choose these point to lie on dierent conductors. Denition (2.7) becomes then
the conventional denition of the characteristic impedance for TEM modes.
The above denition of the waveguide voltage, current and characteristic impedance
are the fundamental concepts of the transmission line theory. This theory extends the
conventional circuit theory by allowing voltages and currents to depend also on the location.
The main tool of this theory is a set of dierential equations, referred to as Telegraphic
equations, which describe the wave propagation in terms time- and location-dependent
voltages and currents. For details refer to, e.g., [1] or [47].

2.1.2

Wave amplitudes and scattering parameters

So far, we have presented two dierent means of representing elds in the waveguide.
The voltage-current description is independent of the eld normalizations used in et (x, y)
and ht (x, y) and allows us to use the methods of the transmission line theory. However,
this description does not represent well the underlying physics of wave propagation phenomenon which is best both analyzed and experimentally observed in terms of forward and
backward propagating waves rather than voltages and currents. The other description we
discussed, resulting directly from the solution of Maxwell equations, uses unitless constants
C + and C . These constants have straightforward physical interpretation and describe the
amplitudes of the forward and backward propagating wave. However, they are dicult to
both interpret and measure sue to the dependence on the normalization of the mode elds
et (x, y) and ht (x, y).
Solution to that problems is the description in terms of traveling-wave amplitudes. This
description arises from the following normalization of the constants C + and C


a0 (z) =

+ z

2ReP0 C e

, and b0 (z) =

2ReP0 C e+z .

By use of this normalization, we readily obtain the elds in the waveguide as


11

(2.9)

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS

a0 (z) + b0 (z)

et (x, y),
2ReP0
a0 (z) b0 (z)
=
ht (x, y),
2ReP0

Et (x, y, z) = C + et (x, y)ez + C et (x, y)e+z =

(2.10)

Ht (x, y, z) = C + ht (x, y)ez + C ht (x, y)e+z

(2.11)

where the quantities a0 (z) and b0 (z) are referred to as the forward and backward traveling
wave amplitudes.
The adjective traveling indicates that amplitudes a0 (z) and b0 (z) describe the actual
forward and backward propagating waves in the waveguide. Similarly to waveguide voltage
and current, the traveling wave amplitudes are independent of the normalization used in
the modal elds et (x, y) and ht (x, y). The traveling wave amplitudes have also a close
correspondence to the real power carried by the propagating mode. Indeed, we can show
that, in the absence of the backward propagating wave, the forward wave carries the real


power ReP (z) = 12 Re S Et (x, y, z)H


(x, y, z)dS = 12 |a0 |2 . Hence, a traveling wave with a
 t
 
unit root-mean-square amplitude,  a02  = 1 carries a unit real power. Similar relationships
holds for the backward propagating wave.
It is important to note that the denition of traveling wave amplitudes does not use
the waveguide voltage, current, nor the characteristic impedance. These amplitudes are
therefore directly related to the modal solutions of Maxwell equations in the waveguide.
However, the traveling wave amplitudes a0 (z) and b0 (z) can easily be related to the voltagecurrent description. Indeed, by comparing (2.1) and (2.2) with (2.10) and (2.11), we obtain

1 ReP0
a0 (z) =
[V (z) + I(z)Z0 ] ,
2 V0

1 ReP0
b0 (z) =
[V (z) I(z)Z0 ] ,
2 V0

(2.12)
(2.13)

and thus
V0
1
[a0 (z) + b0 (z)] ,
V (z) =
2 ReP0
I0
1
[a0 (z) b0 (z)] .
I(z) =
2 ReP0

(2.14)
(2.15)

The direct correspondence of the traveling wave amplitudes to the forward and backward propagating modes leads to some surprising results in the case when there is a phase
12

2.1. DEFINITION OF S -PARAMETERS


shift between the electric and magnetic elds, that is, when arg P0 = 0. This phase shift is
a consequence of power loss in the waveguide, which is commonly encountered in practice
due to nite conductivity of real conductors and losses in dielectrics.
When the waveguide is lossy, the forward and backward propagating modes are not
orthogonal. Therefore, the real power owing through a given cross section is not equal
to the sum of powers carried by the modes, which is usually assumed in the classical
transmission line theory [48, 49]. This result can easily be conrmed by writing the real
power owing through the waveguide cross-section with the use of the traveling wave
amplitudes

1
1
Imp0
1
1
ImZ0
ReP (z) = |a0 |2 |b0 |2 + Im(a0 b0 )
= |a0 |2 |b0 |2 + Im(a0 b0 )
.
2
2
Rep0
2
2
ReZ0

(2.16)

Indeed, for arg P0 = 0 we obtain an additional term related to the phase of P0 . Note that,
according to (2.8), this phase is a property of the mode and does not depend on the choice
of eld normalizations. It is related to the characteristic impedance of the mode through
(2.8), hence arg P0 = 0 implies that characteristic impedance Z0 is complex.
Another property, surprising in the context of transmission line theory, that results from
the loss in the waveguide, is that the ratio of real powers incident at and reected from
an discontinuity in the waveguide is not equal to ||2 where is the reection coecient
= b0 /a0 [48]. Therefore, in some cases, magnitude of may exceed one which is also
unusual for the classical transmission-line theory. This result can also be easily obtained
with the use of traveling wave amplitudes [45].
Therefore, for practical reasons, it is sometimes desirable to have an alternative normalization which would lead to more intuitive results in the context of the transmission-line
theory. Also, when characteristic impedance Z0 exhibits a signicant frequency dependence, it is more convenient to have a xed relationship between the wave amplitudes and
waveguide voltages and currents that be independent of the frequency dependence of Z0 .
A normalization that has these properties is proposed in [45] and has form


|V0 | ReZref
[V (z) + I(z)Zref ] ,
a(z) =
V0 2|Zref |

(2.17)

|V0 | ReZref
[V (z) I(z)Zref ] ,
V0 2|Zref |

(2.18)

b(z) =

13

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


where Zref is an arbitrary parameter with ReZref 0, and a(z) and b(z) are referred to
as the pseudo-wave amplitudes. This normalization has the property that the real power
ow in the waveguide is described by the expression as (2.16), however with Z0 replaced by
Zref . Consequently, for a real Zref , the additional term in expression (2.16) vanishes and
we obtain a description that is more intuitive in the context of transmission line theory.
It can be also shown that for Zref = Z0 , the pseudo-wave amplitudes (2.17) and (2.18)
become traveling-wave amplitudes (2.12) and (2.13).
It is important to note that the pseudo-wave amplitudes do not directly correspond
to the traveling wave amplitudes. Indeed, we can readily show that a(z) and b(z) are
linear combinations of a0 (z) and b0 (z). Therefore pseudo wave amplitudes are more of a
mathematical artifact than a physical representation of wave propagation in the waveguide.
This results in a well know property that for an innite waveguide stimulated by a traveling
wave with |a0 (z)| = 0 we have |b0 (z)| = 0, however, after conversion to the pseudo-wave
amplitudes, we obtain |b(z)| =
 0 [45]. This again conrms that the pseudo-wave amplitudes
do not reect the physics of wave propagation in the waveguide.
It is sometimes desirable to convert from on set of pseudo-wave amplitudes, a(z) and
b(z), with a reference impedance Zref to another set, a(z) and b(z) with a dierent

reference impedance Zref
. The relationship between the two sets can be easily determined
from (2.17) and (2.18) as [45]

a(z)
a(z)

= N
,

b(z)
b(z)
where
N=
and



1 jImZref /ReZref



1 jImZref
/ReZref

(2.19)

1
1

,
2
1
1


Zref
Zref
= 
.
Zref + Zref

(2.20)

(2.21)


It can be shown that for real reference impedances Zref and Zref
, matrix (2.20) becomes
the transmission matrix of an ideal impedance transformer [1].

Having discussed dierent denitions of wave amplitudes, we nally introduce the scattering parameters. For a circuit with N ports, we group the wave amplitudes (traveling14

2.1. DEFINITION OF S -PARAMETERS


wave amplitudes of pseudo-wave amplitudes) at the circuit ports into two vectors

a
1

..
a = . , b =

b1
..

. ,

(2.22)

bN

aN

and dene a linear relationship between the two vectors with a matrix S
b = Sa.

(2.23)

When matrix (2.23) is dened in terms of traveling-wave amplitudes, we refer to its elements
as scattering parameters. In the case of pseudo-wave amplitude, we refer to the elements
of S as pseudo-scattering parameters. However, since the pseudo-wave amplitudes become
traveling wave amplitudes for Zref = Z0 , we often talk briey about scattering parameters
dened with reference to a certain impedance Zref .
In the case of two-port devices2 , it is sometimes more convenient to represent the
relationship between the wave amplitudes with the use of transmission matrix dened as

b
a
1 = T 2 .
a1
b2

(2.24)

This description has the useful property that the transmission matrix of the cascade connection of two-port networks described with transmission matrices Ti , for i = 1, . . . , N , is
given by a product
T=

N


Ti .

(2.25)

i=1

In the common case of a two-port device, we give the relationship between the two
representations explicitly as they are used very often throughout this work. For a two-port
network with the scattering parameters given by

S11 S12
,
S=
S21 S22

(2.26)

The transmission matrix representation can easily be extend to the case of multiport devices with an
even number of ports, see [47].

15

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


and the transmission parameters

T11 T12
,
T=
T21 T22

(2.27)

the relationships between the elements of (2.26) and (2.27) is [1]

1 S12 S21 S11 S22 S11


T=
,
S21
S22
1
and

2.1.3

(2.28)

1 T12 T11 T22 T12 T21


S=
.
T22
1
T21

(2.29)

Practical implications

In practice, the general denition of S-parameters presented in the previous section,


can often be simplied. In most practical cases, the normalization voltage V 0 is real3 ,
hence we have |V0 |/V0 = 1. We obtain then the following relations between the waveguide
voltages and currents, and the traveling-wave amplitudes

ReZ0
[V (z) + I(z)Z0 ] ,
2|Z0 |

ReZ0
[V (z) I(z)Z0 ] ,
b0 (z) =
2|Z0 |

a0 (z) =

(2.30)
(2.31)

while for the pseudo-wave amplitudes we have




a(z) =

2|Zref |

b(z) =

ReZref

ReZref

2|Zref |

[V (z) + I(z)Zref ] ,

(2.32)

[V (z) I(z)Zref ] ,

(2.33)

Voltage V 0 becomes complex if the plane of a constant phase velocity it not perpendicular to the
direction of propagation. This occurs in waveguides with dielectrics that are anisotropic or inhomogeneous
in the waveguide cross-section.

16

2.2. VNA S -PARAMETER MEASUREMENT


and thus
1
|Zref | [a0 (z) + b0 (z)] ,
V (z) = 
ReZref

(2.34)

|Zref |
1
[a0 (z) b0 (z)] .
I(z) = 
ReZref Zref

(2.35)

For a real reference impedance, we further obtain the familiar expression known from the
circuit theory [5052]

1
[V (z) + I(z)Zref ] ,
a(z) = 
2 Zref
1
[V (z) I(z)Zref ] ,
b(z) = 
2 Zref

(2.36)
(2.37)

and
V (z) =

Zref [a0 (z) + b0 (z)] ,


1
[a0 (z) b0 (z)] .
I(z) = 
Zref

(2.38)
(2.39)

In the context of the VNA S-parameter measurements, it is important to note that the
VNA measures S-parameters with respect to some unknown reference impedance. Hence
an important aspect of the VNA calibration is the determination of this impedance. This
will be discussed in more detail in Paragraph 2.4.2-C.

2.2

VNA S-parameter measurement

In the previous section, we demonstrated that the scattering parameters, as captured in


S-matrix dened by (2.23), describe relationships between normalized guided electromagnetic waves incident at and reected from the ports of an electronic circuit. This suggests
an intuitive method for their measurement, namely through an observation of these waves
in some controlled conditions, such as when only one of the device ports is excited at a
time. This observation should disturb the waves as little as possible (this is analogous to
the condition in low-frequency oscilloscope measurements that the probe has high input
17

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS






Fig. 2.1: VNA block diagram (switch is shown in the forward position).

impedance such that it does not disturb voltages and currents in the circuit), and in order
to avoid any interferences, waves emerging from the device-under-test (DUT) should be
absorbed at some place far enough from the DUT (this is analogous to the measurement
of impedance or admittance parameters when we require, respectively, low impedance or
high impedance termination of the circuit terminals).
This simple idea is the operational principle of the vector network analyzer (VNA).
A simplied diagram of a typical VNA, dedicated to the measurement of devices with
two or less ports (in short, a two-port VNA), is shown in Fig. 2.1. On either side of the
DUT, there is a set of two directional couplers with detectors. We refer to each set as
a reectometer. The function of the reectometer is to measure the complex amplitudes
of the wave incident at and reected from the DUT. This process is realized by coupling
part of each wave out the detection circuit and converting it to a lower frequency at which
the analog to digital (A/D) converters can be used. We show the detection circuit as a
single mixer excited by the local oscillator (LO), however, in the reality the signal may
undergo multiple frequency conversions. The A/D conversion may take place either at
some intermediate frequency (IF) or in the baseband.
18

2.2. VNA S -PARAMETER MEASUREMENT


In the forward position of the switch (this is the position shown in Fig. 2.1), signal
from the RF source is sent to the rst port of the device under test while the other port
is terminated with a matched termination. The function of this termination is to absorb
the wave emerging from the non-excited port of the DUT. Complex voltages a1m and
b1m , which are approximately proportional to the amplitudes of the waves incident at and
reected from port one, a1 and b1 , respectively, are then measured in the left reectometer.
Similarly, signal b2m , approximately proportional to the wave b2 transmitted through the
DUT, is measured in the right reectometer. From these measurements, we approximate
scattering parameters S11 and S22 of the DUT as
S11m =

b1m
b2m
, and S21m =
,
a1m
a1m

(2.40)

respectively. A similar description holds for the reverse position of the switch and we
obtain the following approximations of S22 and S12

S22m =

b2m
b1m
, and S12m =
.
a2m
a2m

(2.41)

We refer these approximations as raw, measured or uncorrected S-parameters.


The practical implementation of the VNA is far more complicated then the diagram
in Fig. 2.1. For detailed discussion of dierent architectures see for example [53]. The
main objective of the VNA construction is to provide wideband operation (for example
from 70 kHz up to 70 GHz [4]) while maintaining the error of approximations (2.40) and
(2.41) reasonably small. This objective is very hard to attain in practice, therefore the
errors of approximations are (2.40) and (2.41) usually not acceptable, even for approximate
assessment of the DUT S-parameters.
These errors result from various imperfections of the VNA construction. The most
important ones are the nite directivity of directional couplers, impedance mismatches in
the VNA (such as between the generator and the adjacent coupler, or between the other
coupler and the matched termination), discontinuities in the transmission lines guiding the
measured signals, phase shift and attenuation introduced by these lines, parasitic coupling
between the VNA ports (e.g. through the LO circuitry), and dierences between the load
impedance in the forward and reverse position of the switch.
19

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS

2.3

Two-port VNA mathematical models

In this section, we discuss mathematical models for two-port VNA measurements.


These models describe the relationship between the measured and actual S-parameters
of a DUT as a function of a set of model parameters describing the systematic errors
introduced by the VNA. These models form the foundation for the correction of VNA
measurements and the development of VNA calibration algorithms.
The basic premise in the formulation of VNA models is that the relationships between
the measured and actual wave amplitudes are linear and time-invariant, thus the VNA
is assumed to be a linear time-invariant (LTI) system. We begin our discussion with the
description of VNA models based on this premise. Then we discuss situations when the
time-invariance assumption is violated. We do not discuss the case when the VNA becomes
non-linear as it is beyond the scope of this work.

2.3.1

Linear time-invariant two-port VNA

A. 16-term model. The most general model of a


linear time-invariant two-port VNA is the 16-term
model [6, 5456]. This model, shown schematically


in Fig. 2.2, results directly from the VNA diagram
in Fig. 2.1. In the 16-term model model, the linear
relationship between the measured and actual waves
as a four-port liner network, denoted with E Fig. 2.2,
Fig. 2.2: 16-term model of a twoand dened as
port VNA.

b1m
b2m
a1
a2

Se

a1m
a2m
b1
b2

e11 e12 e13 e14


e21 e22 e23 e24

S11e S12e
where Se =
=

e31 e32 e33 e34


S21e S22e

e41 e42 e43 e44

(2.42)

Model parameters contained in the matrix Se encompass all possible transmission and
reection paths in the VNA. Referring to Fig. 2.2 and Fig. 2.1, we note that
the diagonal terms of S11e , S12e , S21e , and S22e describe the systematic error introduced by the VNA reectometers themselves;
20

2.3. TWO-PORT VNA MATHEMATICAL MODELS


anti-diagonal terms of S11e (that is e12 and e21 ) correspond to the internal crosstalk between the VNA reectometers; this cross-talk results from the nite isolation
between dierent ports of the switch;
the anti-diagonal terms of Se (that is e14 , e23 , e32 and e41 ) correspond to the coupling
between VNA reectometers, for example, through the IF circuitry; in modern VNAs
this coupling is negligible, therefore one typically assumes e14 = e23 = e32 = e41 = 0;
the anti-diagonal terms of S22e (that is e34 and e43 ) correspond to the external crosstalk between the VNA reectometers, that is, to the direct cross-talk between the
VNA measurement ports; when measuring open-waveguide structures (e.g., miscrostrip lines or coplanar waveguides), this cross-talk can be signicant, however,
in the case of enclosed waveguides (e.g., coaxial lines or rectangular waveguides) this
cross-talk does not occur.
We determine the relationship between the raw and actual S-parameters in the 16-term
model by applying their denitions

b
a
b
a
1m = Sm 1m , and 1 = S 1 ,
b2m
a2m
b2
a2

(2.43)

respectively, to (2.42) and solving the resulting set of linear equations. This yields


Sm = S11e + S12e S (I S22e S)1 S21e = S11e + S12e S1 S22e


and

S = S21e (Sm S11e )1 S12e + S22e

1

1

S21e ,

(2.44)

(2.45)

By exploiting the structure of (2.44) and (2.45), we note that, although the model (2.2)
has 16 terms, only 15 terms need to be known to solve (2.44) and (2.45). Indeed, if we
multiply all elements of S21e by an arbitrary constant and divide all elements S12e by the
same constant, relationships (2.44) and (2.45) do not change. Therefore one of the elements
in S12e or S21e can be arbitrarily chosen, for example xed to one.
We further observe that relationships between the actual and measured S-parameters,
S and Sm , respectively, given by (2.44) and (2.45), are nonlinear functions of the model
parameters contained in the matrix Se , dened in (2.42). Therefore, the 16-term model is
often expressed in an alternative form which uses a dierent set of parameters for which
21

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


the relationships between S and Sm become linear. This alternative form is dened by
[6, 55, 56]

b1m
b2m
a1m
a2m

Te

b1
b2
a1
a2

t11 t12 t13 t14


t21 t22 t23 t24

T11e T12e
where Te =
=

t31 t32 t33 t34


T21e T22e

t41 t42 t43 t44

(2.46)

After applying (2.43) to (2.46) and solving the resulting set of linear equations we obtain
T11e S + T12e Sm T21e S Sm T22e = 0,

(2.47)

where the model parameters contained in the submatrices of Te can be expressed in terms
of the original parameters of the 16-term model as
T11e = S12e S11e S1
21e S22e ,

(2.48)

T12e = S11e S1
21e ,

(2.49)

T21e = S1
21e S22e ,

(2.50)

T22e = S21e .

(2.51)

The relationship (2.47) can further be brought to a very convenient form with the use of
matrix vectorization operator [57]. For a matrix X represented as X=[x1 , . . . , xN ], where
x1 , . . . , xN are the columns of X, the vectorization operator is dened as [57]

x
1
..
vec(X) = . .

(2.52)

xN
Consequently, the vector (2.52) consists of stacked columns of matrix X. Applying this


operator to (2.47) and with the use of the identity vec (ABC) = CT A vec (B), where
is the Kronecker product (see [57]), we obtain


ST I22 t11e + t12e ST Sm t21e (I22 Sm ) t22e = 0,

22

(2.53)

2.3. TWO-PORT VNA MATHEMATICAL MODELS


where tije = vec (Tije ), for i, j = 1, 2. This can be further transformed to

ST I2 I4 ST Sm

I2 Sm

t11e
t12e
t21e
t22e

= 0,

(2.54)

where I2 and I4 are identify matrices of size 2 2 and 4 4, respectively. Equation (2.54)
forms a foundation for the 16-term VNA model identication [55, 56].

B. 8-term model. For modern VNAs, the internal cross-talk and coupling are usually
very small. Also, when performing VNA measurements with closed waveguides, such as
coaxial transmission line or rectangular waveguide, the external cross-talk is negligible
4
. In this case, the VNA reectometers are electrically separated and the model (2.42)
simplies to the 8-term model [6, 29], referred to also as error-box model (see Fig. 2.3). In
the 8-term model, the VNA reectometers are represented as two linear two-port networks
A and B, referred to as the error boxes. We shall rst formulate denitions of these
networks, following a similar convention to that used in model (2.46), and then show
another formulation which stems from the basic form (2.42) of the 16-term model.
As the networks A and B are electrically separated, we can rewrite (2.46) as two sets
of independent equations

b
b
a
a
1m = TA 1 , and 2 = TB 2m ,
a1m
a1
b2
b2m

(2.55)

and represent the measured and actual S-parameters, Sm and S, as the transmission
parameters, Tm and T, respectively, dened by

b
a
a
a
1 = T 2 , and 1m = Tm 2m ,
a1
bb
b1m
b2m

(2.56)

which immediately yields


Tm = TA TTB ,
4

(2.57)

In the case of VNA measurements involving open waveguides, such as in the case of on-wafer measurements or measurements employing xtures with microstrip lines, the external cross-talk may become
signicant.

23

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS



Fig. 2.3: 8-term model of a two-port VNA.


and consequently
1
T = T1
A Tm TB .

(2.58)

Equations (2.57) and (2.58) constitute probably the most commonly used formulation
of the 8-term model. Following a similar reasoning as in the case of the 16-term model, we
can show that out of total 8 complex terms in TA and TB , only 7 are necessary to describe
the relationship between Tm and T. Therefore one of the terms in the 8-term model can
be arbitrarily chosen, for example xed to one.
Parameters of the eight term model can be chosen in dierent ways. A simple choice
is to directly use the coecient of matrices TA and TB . Another choice is to relate the
parameters of the eight term model to the actual sources if systematic error in the VNA
measurement. This can be done with the use of the ow graph in Fig. 2.4 [58]. The
terms in the graph correspond to the dierent sources of systematic errors in the VNA
reectometers. The second letter in the subscript, F or R, denotes position of the
switch, forward or reverse, respectively, in which the signal source is connected to the
reectometer. The individual terms correspond to the systematic errors resulting from
nite directivity of the reectometers (EDF and EDR ),
mismatch at the reectometer input (ESF and ESR ),
reection tracking of the reectometers (ERF and ERR ).
The additional terms and describe the asymmetry in the parameters of both reectometers. However, since the 8-term model has 7 independent terms, only the ratio /
appears in the equations (2.57) and (2.58). Therefore the ow graph for the 8-term model
can also be represented in an alternative form, by lumping the non-reciprocity of both error
boxes into one of them. In Fig. 2.5, we show such an alternative representation where the
non-reciprocity is lumped into the error box representing port two of the VNA. Similar
graph can also be obtained by lumping the nonreciprocity into error box representing port
one of the VNA.
24

2.3. TWO-PORT VNA MATHEMATICAL MODELS

Fig. 2.4: Flow graph for the 8-term VNA model.

Fig. 2.5: Alternative form of the ow graph for the 8-term VNA model.
With the use of the terms shown in Fig. 2.4, we can rewrite (2.57) as
Tm =

1
EA TEB ,
Et

(2.59)

where the transmission matrices Tm and T can be derived from the measured and actual
DUT S-parameters with the use of (2.28) , while

ERF EDF ESF EDF


ERR EDR ESR ESR
EA =
, EB =
,
ESF
1
EDR
1
and
Et =

ERR .

(2.60)

(2.61)

The appealing simplicity of (2.57) and (2.58) allows to describe the VNA calibration
problem in a very concise and elegant way which has lead to numerous interesting results
(see for example [30, 36]). However, formulation (2.57) and (2.58), has also an important
disadvantage. After examining (2.28), we note that transmission matrix T cannot be
dened for a DUT that does not have a forward transmission, that is, when S21 = 0.
Indeed, in such a case, pairs of variables a1 , b1 and a2 , b2 are unrelated and the transmission
matrix T does not exist. Hence, matrix formulation (2.57) and (2.58) cannot provide a
uniform description for the VNA measurements of both two-port and one-port devices.
In order to describe the measurement of a one-port device with the use of matrices TA
25

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


and TB , we apply the denitions of reection coecient measurement on port A and B
mA =
to obtain
mA = EDF +

b1m
b2m
, and mB =
a1m
a2m

(2.62)

ERF A
ERR B
, and mB = EDR +
,
1 ESF A
1 ESR B

(2.63)

where A and B are the reection coecients of the one-port device connected to the
VNA port A and B, respectively. Equations (2.63) can be easily inverted to obtain the
correction formulas.
An alternative formulation of the eight-term model that provides a uniform description
of both one-port and two-port measurements, can be derived from the basic form (2.42)
of the 16-term model. Taking into account that the VNA reectometers are electrically
separated and including the denitions in Fig. 2.4, we can write submatrices of Se as

Se =

S11e S12e
=
S21e S22e

EDF

ERF

EDR

ESF
1

ERR

(2.64)

ESR

Applying then (2.64) to formulas (2.44) and (2.45), we obtain a uniform description of
both one-port and two-port measurements
ERF
ERF ESR S

,
D
D
ERR ERR ESF S

,
= EDR + S22
D
D
ERR
S21m = S21
,
D
ERF
S12m = S12
,
D

S11m = EDF + S11

(2.65)

S22m

(2.66)
(2.67)
(2.68)

where
S = S11 S22 S21 S12 ,

(2.69)

D = 1 ESF S11 ESR S22 + ESF ESR S .

(2.70)

and

26

2.3. TWO-PORT VNA MATHEMATICAL MODELS


C. Model parametrization choice. We showed in the previous section that the 16-term
and the 8-term model can be represented in terms of dierent sets of parameters. Although
this dierent parametrizations are equivalent, special attention needs to be paid to their
properties in the context of their application in VNA calibration algorithms. Important
property in this context is the uniqueness of the parametrization. Parametrizations having
this property allow one to avoid the so called root choice problem in analytical VNA calibration methods, and improves the robustness of the iterative VNA calibration techniques.
We shall discuss this issue in more detail for the 8-term model.
The primary parametrization we use for the 8-term model, which we refer to as the
base parametrization, results directly from (2.64). In this parametrization, we write the
vector of VNA-model parameters as


p = EDF , ESF , ERF , EDR , ESR , ERR ,

T

(2.71)

By expanding equations (2.44) and (2.45) in terms of parameters in p, we can easily show
that these parameters describe a unique solution to (2.44) and (2.45). In other words, if
some p solves equations (2.44) and (2.45), there is no other p = p that also solves these
equations.
Parametrization (2.71), however, is not the common one encountered in the literature.
Two other parametrizations that are often used are the reciprocal parametrization (see [36,
59]) and the transmission parametrization (see [29, 30]). In the reciprocal parametrization,
write the vector of VNA parameters is written as


pR = EDF , ESF , ERF , EDR , ESR , ERR ,

ERR
ERF

T

= [EDF , ESF , etF , EDR , ESR , etR , k]T . (2.72)


This parametrization has a very convenient property that the joint eect of the nonreciproc
ERR
.
ity of both VNA error boxes is lumped into a single non-reciprocity factor k = E
RF
Thus the VNA error boxes A and B are represented as reciprocal two-port linear networks
with S21A = S12A = etF and S21B = S12B = etR , respectively. For reciprocal error boxes
(that is when = ERF / and = ERR /), we have k = 1, otherwise |k| =
 1. Consequently, adding a reciprocal linear network between the VNA error box and the DUT does
not change k. This is not the case of the base parametrization, for which adding such a
27

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


linear network aects, in general, all of the parameters contained in (2.71).
Parametrization (2.72) is, however, not unique. We can demonstrate that by investigating the conversion of the base parametrization (2.71) into (2.72). To this end, we write
etF and etR as


+
+
etF = sRF ERF , and etR = sRR ERR ,
(2.73)
+
where x indicates one of the square roots of the complex number x (e.g., with the positive
real part), and sRF = 1 and sRR = 1. Based on that, we rewrite (2.72) as

pR (sRF , sRR ) = EDF , ESF , sRF ERF , EDR , ESR , sRR ERR

+ T
sRR ERR
,
. (2.74)

sRF ERF +

Consequently, we see that there are four dierent vectors pR (sRF , sRR ) which lead to the
same vector p, depending on which square root we choose.
The transmission parametrization is based on the transmission matrix representation
(2.59) of the 8-term model. In this parametrization, the vector of VNA parameters is
written as


pT = EDF , ESF , EDF ESF

ERF , EDR , ESR , EDR ESR ERR , ERR

T

= [EDF , ESF , F , EDR , ESR , R , r]T . (2.75)


We can easily show that (2.75) can be uniquely derived from the base parametrization
(2.71).

2.3.2

Modeling VNA nonstationarity

The 16-term and eight term model presented in the previous section capture the primary systematic errors introduced in the VNA measurements. These models assume that
the VNA does not change with time, which is essential for reliable identication of and
correction for VNA systematic errors.
In practice, however, this assumption is not met. The primary reason for the VNA
nonstationarity is the very manner in which the VNA operates, that is, by repeatedly
switching the source generator and the matched termination between the two reectometers
(see Fig. 2.1). Due to the asymmetry of the switch, the VNA reectometers see a slightly
dierent impedance of the source generator and of the matched termination in each position
28

2.3. TWO-PORT VNA MATHEMATICAL MODELS

(a)

(b)

Fig. 2.6: Modeling changes of the matched termination impedance: (a) forward measurement, (b) reverse measurement.

of the switch. The dierence


positions of the switch does
captured in the 16-term or
However, the change of the
[6, 58].

between the source generator impedance in the two dierent


not lead to an error since the reectometer parameters, as
eight term model, do not depend on this impedance [60].
matched termination impedance leads to systematic errors

Other sources of the VNA nonstationarity are the nonrepeatability of the switch, the
test-set drift, imperfect connector repeatability and errors due to cable exure. Apart from
the test-set drift which is strongly dependent on the temperature and humidity changes,
these errors are of a random nature and without any further knowledge about their character they cannot be corrected for. We discuss those sources of VNA nonstationarity in
more detail in Subsection 3.3.2.
As to the changes of the matched-termination impedance, there exist two common
approaches for modeling their impact of VNA-model parameters. In either approach, a
dierent models is used for the VNA operating in the forward and reverse direction. In
the 12-term model [6], additional terms are used to described the eect of the matchedtermination variation whose values, however, are not directly related to the value of the
matched-termination impedance seen through the switch. Another approach, which is
illustrated in Fig. 2.6, extends the VNA models presented in the previous section by adding
some additional terms related directly to the impedance of the matched termination. In
the forward and reverse positions of the switch, the matched termination presents dierent
reection coecient, F and R , respectively. This reection coecients are referred to as
the switch terms. Consequently, the measured parameters can be expressed as a function
of the time-invariant VNA measurement Sm and the switch terms. In particular, when the
29

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


switch is in the forward direction, the VNA measures
F
S11m
=

b1F m
b2F m
F
and S21m
=
,
a1F m
a1F m

(2.76)

and when the switch is in the reverse direction, the VNA measures
R
S22m
=

b2Rm
b1Rm
R
and S12m
=
.
a2Rm
a2Rm

(2.77)

With the use of these measurements, one can solve the linear equations resulting from the
ow graphs in Fig. 2.6 to obtain [58]

R
F
R
F
R
S F S12m
S21m
F S12m
S11m
S12m
R
1
11m
Sm =
.
R
F
F
R
F
R
R
F
1 S12m S21m R F S21m S22m S21m F S22m S12m S21m R

(2.78)

Matrix Sm obtained in that way is then used with the time-invariant VNA models presented
in Section 2.3.1.
The reection coecients F and R are typically measured in a separate step [58].
The VNA ports are then directly connected so that the matched termination is excited
from the opposite port. This impedance is typically very stable [58], therefore repeated
measurement of the reection coecients F and R are not necessary.

2.4

Two-port VNA calibration techniques

VNA calibration procedures have been extensively studied in the literature and there
exists several good and detailed reviews, such as [6], or more recently [8]. Based on those
reviews, one concludes that these procedures dier from each other in many aspects, such
as the number and the type of calibration standards used, model of the VNA systematic
errors, mathematical formulation of the calibration problem, or the numerical method
used to solve it. Thus, VNA calibration procedures seem to form an inhomogeneous realm
with the only commonality being the high complexity of the mathematical description.
Therefore, in this overview, instead of going deep into mathematical details, we shall take
a higher-level look at the VNA calibration procedures and try to draw some similarities
between them.
The VNA can be thought of as a system whose input and output are the actual scatter30

2.4. TWO-PORT VNA CALIBRATION TECHNIQUES


ing parameters of a device and their raw measurement, respectively (see Fig. 2.7). Denoting
the vector representation of the measurement and the actual S-parameters of the device
under test (DUT) with the vectors sm and s, respectively, and the VNA model parameters
(calibration coecients) with the vector p, we may model the VNA operation at a given
frequency as a vector function
sm = f (s, p) ,
(2.79)
where the particular form of this function depends on the VNA mathematical model and
the parametrization choice (see Section 2.3).
The objective of the VNA calibration procedure is
now to determine the parameters p of the VNA based on
the measurement of a number of devicesreferred to as
calibration standardswith some known characteristics. Fig. 2.7: Representation of the
The VNA calibration problem falls therefore into the gen- VNA operation.
eral class of system identication problems [61, 62], also
referred to as inverse problems [63] or nonlinear regression problems [64, 65].
In order to formulate this problem
 

in a more precise way, consider Fig. 2.8.


  
and
s
denote
the
measurement
Let sm
n
n
and the actual parameters of the n-th
calibration standard, for n = 1, . . . , N ,
Fig. 2.8: VNA measurement of a calibration stanwhere N is the number of calibration
dard.
standards. Let further the actual Sparameters of the n-th calibration standard be represented by the function
sn = gn (c0n , cn ) ,

(2.80)

where c0n and cn are known and unknown parameters of the n-th calibration standard,
respectively. The particular form of the function gn depends on the physical model of the
calibration standard, and the choice of the known and unknown parameters.
The VNA calibration procedure can now be thought of as a system identication problem in which one determines the vector p of VNA parameters and the unknown calibraN
tion standard parameters {cn }N
n=1 based on the denitions {c0n }n=1 and raw measurements
N
{sm
n }n=1 of the calibration standards. This description will serve as a basis for the overview
of VNA calibration techniques. In this overview, we consider three dierent aspects of VNA
31

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


calibration techniques: mathematical formulation of VNA calibration problem, types of
calibration standards, and numerical methods used to solve the calibration problem.

2.4.1

Formulation

A. Deterministic and statistical approach. In order to solve the VNA calibration


problem, we may use either a deterministic or a statistical approach. The deterministic
approach is based on the following assumptions:
the measurements and denitions of the calibration standards are accurate,
the number of independent relationships describing calibration standard measurements is equal to the total number of estimated parameters (i.e. parameters contained in the vector p and the set of vectors {cn }N
n=1 ).
Consequently, in the deterministic approach, the parameters p and {cn }N
n=1 are obtained
by solving a set ofin general non-linearequations. This equations are formed by (2.79)
and (2.80), written for all of the calibration standards, that is,

sm
1 = f (g1 (c01 , c1 ), p)
..
.

(2.81)

sm
N = f (gN (c0N , cN ), p)

Since the number of equations is assumed to be equal to the number of sought parameters,
the calibration problem has either a unique or a nite number of solutions. We generally
strive to pose the VNA calibration problem such that it has a unique solution, however,
due to its non-linear character, some ambiguity often cannot be avoided (e.g., the so-called
root-choice problems [29, 30]).
The deterministic approach had long been the primary direction along which the development of VNA calibration procedures was taking place. The ag examples of deterministic VNA calibration procedures are the Short-Open-Load (SOL) procedure for one-port
VNAs, and the Short-Open-Load-Thru (SOLT)5 procedure for two-port VNAs [66]. In
both procedures, Short, Open and Load stand for a one-port device realized as a
transmission line section terminated with a short circuit, an open circuit, and a matched
5

In the SOLT procedure, the number of equations is actually larger then the number of sought VNA
parameters. These equations, however, are solved using deterministic methods (see [6, 66]), by neglecting
some pieces of information.

32

2.4. TWO-PORT VNA CALIBRATION TECHNIQUES

Fig. 2.9: VNA measurement of a calibration standard with errors.


termination, respectively. The Thru denotes either the direct or indirect (e.g., with a
known transmission line or an adapter) connection of the VNA ports.
The main advantage of the deterministic approach is that is allows to better understand
the relationships between the solution to the calibration problem and the number and
type of calibration standards. The rigorous analysis of equations (2.81) has therefore lead
to many interesting results, such as the self-calibration methods (see [29, 30, 67, 68]).
The Thru-Reect-Line (TRL) procedure [29, 67] and the unknown-thru procedure [69],
discussed in more detail in Paragraph 2.4.2-A, are the most representative examples of
these methods.
The main disadvantage of the deterministic calibration methods is their sensitivity to
measurement errors. Indeed, since the denitions and measurements of calibration standards are assumed to be error free, any error in those values is directly mapped into an
error of the VNA calibration coecients. Therefore, when calibrating the VNA with deterministic procedures, it is a good practice to perform an additional verication procedure, in
order to ensure that no serious measurement error occurred during the calibration. In the
verication procedure, one either remeasures the calibration standards and compares their
corrected measurements with the denitions (e.g., in order to exclude a serious connector
repeatability error), or, preferably, measures a set of check standards whose S-parameters
dier signicantly from calibration-standard S-parameters. By comparing the measurement of these standards with their denitions, both systematic and random errors in the
VNA calibration procedure can be detected.
Statistical approach the VNA calibration problem allows to avoid the above problems.
This approach is based on the assumptions that:
the measurements and denitions of the calibration standards are corrupted with
errors (see Fig. 2.9),
the number of independent relationships describing calibration standard measurements is larger than the total number of estimated parameters (i.e. parameters
33

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


contained in vector p and the set of vectors {cn }N
n=1 ).
Consequently, in the statistical approach, we obtain an overdetermined (redundant) set of
equations describing the calibration standard measurements, that is,

m
sm
1 = f (g1 (c01 , c1 ) + s1 , p) + s1
..
.

(2.82)

m
sm
N = f (gN (c0N , cN ) + sN , p) + sN

where sm
n and sn denote the error in the measurement and denition of the n-th calibration standard, respectively. Since the measurements and denitions of the calibration
standards are both corrupted with errors, a deterministic solution to this set of equations
does not exist (see Section B.2.2). In other words, we cannot nd a set of VNA model
parameters p and unknown parameters of the standards {cn }N
n=1 that would simultaneously solve all of the calibration equations, under the assumption that sm
n = sn = 0,
for n = 1, . . . , N . Hence, instead of looking for a deterministic solution, we need to look
for a solution that is in some sense optimal.
In the system identication, the process of nding such an optimal solution if referred
to as the estimation of system parameters, and the optimal solution is referred to as the
estimator of system parameters [61, 62, 64, 65]. The criteria for the solution optimality are
formulated based on the statistical model of the measurement errors. The most commonly
used estimation technique is based on the maximum likelihood criterion [64, 65]. The
fundamental concepts of this technique in the context of the system identication problems
are reviewed in Appendix B. Examples of other estimation techniques are the method of
moments or the Bayesian estimation [64, 65, 70].
The key concept in the estimation based on the maximum likelihood criterion is the
likelihood function. This function is dened based on the statistical properties of the
measurement errors (see Appendix B). In the context of the VNA calibration, two common
cases are when either only the calibration standard denitions or both the calibration
standard denitions and the raw measurements are aected by errors. In both cases, the
sought parameters of the VNA and calibration standards can be represented as a vector

p
c1
..
.
cN

34

(2.83)

2.4. TWO-PORT VNA CALIBRATION TECHNIQUES


where p denotes the VNA calibration coecients, and {cn }N
n=1 are the vectors of unknown
parameters of the calibration standards.
The case when only the calibration standard denitions are aected by errors is often
encountered in practice. Indeed, most modern VNAs have a very low noise oor, and oer
additionally various means of reducing the raw measurement noise, such as averaging or
narrowing the IF bandwidth. Consequently, the dominant error sources are the connector
nonrepeatability, cable instability and the test-set drift. Although these error sources are
considered as instrumentation errors, their impact can be thought of as a disturbance
added to the calibration standard denitions. Considering that, we can rewrite (2.82) as

f 1 (sm
1 , p) = g1 (c01 , c1 ) + s1
..
.

(2.84)

f 1 (sm
N , p) = gN (c0N , cN ) + sN

These equations describe a variation of the system (B.6), in which both the system outputs
and inputs depend on the estimated parameters. We can still, however, explicitly dene
the residual errors as

r1 () = f 1 (sm
1 , p) g1 (c01 , c1 )
..
.

(2.85)

rN () = f 1 (sm
N , p) gN (c0N , cN )

Assuming now that the errors sn , for n = 1, . . . N , have a normally probability density
function (PDF), we can write the estimator of system parameters (see Section B.2.1) as
= arg min

where

N

n=1

T
1
rn ()T 1
sn rn ()=arg min r () s r () ,

(2.86)

r ()

s1
..

...
, s =
r () =
.

(2.87)

sN

rN ()

the matrix sn is the covariance matrix of the errors in calibration standard denitions,
and the underline denotes the convention for real-valued representation of complex vectors
described in Appendix A. We often assume that the covariance matrix is known up to a
scaling factor, that is, s = 2 Vs , where Vs is a known matrix and 2 is the unknown
35

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


can then be
residual variance (square of the residual standard deviation). The estimate
obtained by replacing s with Vs in (2.86) (see Section B.2.1).
In the case when also the raw measurement errors need to be accounted for, the VNA
calibration problem is solved with the errors-in-independent-variables approach (see Subsection B.2.2). The equations describing the measurements take on the form which is a
variation of (B.5), that is,

m
f 1 (sm
1 s1 , p) = g1 (c01 , c1 ) + s1
..
.

(2.88)

m
f 1 (sm
N sN , p) = gN (c0N , cN ) + sN

Along with the parameters (2.83), we estimate then also errors in raw measurements, thus
the vector of sought parameters becomes

sm
1
..
.

sm
N

(2.89)

and the residuals are

m
r1 () = f 1 (sm
1 s1 , p) g1 (c01 , c1 )
..
.

rN () = f

(sm
N

sm
N , p)

(2.90)

gN (c0N , cN )

Following (B.51), we can then write the maximum likelihood estimate of (2.89) as
N


= arg min

n=1

rn ()T 1
sn rn () +

N

n=1

1
sm
sm
n sm
n =
n

1
m T
m
= arg min r ()T 1
s r () + (s ) sm s , (2.91)

where

r ()
sm
1
1

..

..
m
, s = .
r () =
.

sm
rN ()
N

sm =

sm
1

...
sm
N

36

(2.92)

2.4. TWO-PORT VNA CALIBRATION TECHNIQUES


Again, we often assume that both covariance matrices are known up to a scaling factor,
that is, s = 2 Vs and sm = 2 Vsm , where Vs and Vsm are known matrices
and 2 is the unknown residual variance (square of the residual standard deviation). The
solution is then obtained by replacing s with Vs and sm with Vsm in (2.91) (see
Subsection B.2.2).
The statistical approach to the VNA calibration has many advantages. The most important one is that the use of redundant calibration standards reduces the uncertainty due
to random VNA measurement errors. Thus, unlike the deterministic approaches, statistical
VNA calibrations methods are less sensitive to serious measurement errors. Moreover, such
errors can easily be detected through the residual analysis. The principle of this analysis
is to determine how well the optimal solution to the VNA calibration problem describes
the individual measurements of calibration standards. This is discussed in more detail in
Section 3.5.3.
The drawback of statistical approaches is that in general they require more computer
resources and are more time consuming than deterministic approaches. This drawback,
however, looses its signicance considering the constantly increasing computational power
of personal computers.
Development of statistical approach to the VNA calibration has been initiated by some
early papers on the calibration of one-port VNAs [33, 71, 72] and six-port reectometers
[35, 73]. Newer contributions on the statistical calibration of one-port VNAs include [38,
7477]. In all of these approaches, apart from [75], the least-squares formulation (2.86) is
used.
Statistical approach in the two-port VNA calibrations was started in [36, 55]. In the
excellent paper of Marks [36] a rigorous uncertainty analysis of the TRL calibration is
shown, and based on that analysis a statistical technique for including measurements of
multiple transmission lines with dierent lengths is proposed. Newer publications, such
as [56, 59], generalize this approach to the case of arbitrary two-port VNA calibration
algorithms. The approach of [59] uses the general formulation (2.91).

B. Single-frequency and multi-frequency approach. A typical formulation of the


VNA calibration problem, as described in the previous section, is developed at a single
frequency. Such a formulation, however, becomes inadequate when some of the parameters
estimated in the VNA calibration are frequency independent. Examples include the sliding
load calibration of [37] with unknown positions of the sliding load, or the multi-reect
37

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


through calibration of [38, 78] in which the lengths of the oset terminations are unknown.
In these approaches, the multi-frequency formulation of the VNA calibration problem is
employed. In the following, we demonstrate the principle of this formulation for the general
model (2.88).
Let the model parameters be dened by the vector

f1
..
.

fK

(2.93)

where fk are the parameters (2.89) at the frequency fk , for k = 1, . . . , K, and is a


vector of frequency independent model parameters (those parameters are removed from
{ fk }K
k=1 ). Assuming that the errors are normally distributed, we estimate the parameters
(2.93) by considering the a likelihood function dened jointly at all frequencies. In the case
when the statistical correlations between the errors at dierent frequencies are neglected,
this leads to the minimization of a weighted sum of squared residuals at all measurement
frequencies (see B.2.2)

= arg min

K

k=1

rfk ( k , )T 1
sf rfk ( k , ) +
k

K 

k=1

sm
fk

T

m
1
sm sfk ,
fk

(2.94)

are dened in (2.87) and


where the single frequency covariance matrices sf and sm
fk
k
(2.92), respectively. If the errors in raw measurement are neglected, (2.94) reduces further
to
K

= arg min

rfk ( k , )T 1
(2.95)
sf rfk ( k , ) ,

k=1

which is the formulation whose slightly modied versions are employed in [37, 38]. If there
are no frequency independent model parameters in the vector , solving (2.94) and (2.95) is
obviously equivalent to solving the problem (2.91) and (2.86), respectively, independently
at each frequency.
Reference [37] is, to the authors knowledge, the earliest contribution on the statistical
VNA calibration employing the multi-frequency formulation. In this work, a sliding load
calibration (see [6, 32]) is considered for which the positions of the load are unknown.
These positions are then determined along with the VNA calibration coecients based on
38

2.4. TWO-PORT VNA CALIBRATION TECHNIQUES


the model for the frequency-dependence of the sliding-loads phase. Thus, the accuracy
and robustness of the calibration is improved as the positions of the load need not to be
precisely controlled.
Reference [78] presents for the rst time, to the authors knowledge, a complete VNA
calibration based on the multi-frequency formulation. In this work, the multiple osetshort calibration (see [32, 34, 79]) is described in which the lengths of the oset shorts are
unknown. Based on the model for the frequency-dependence of the oset-short reection
coecient, these lengths are then determined along with the VNA calibration coecients.
Thus, the calibration accuracy is signicantly increased as the systematic errors in the
length determination of the oset shorts are corrected for.

2.4.2

Standards

A. Partially unknown standards. The principle of the VNA calibration is to determine the VNA model parameters based on measurements of a number of calibration
standards with known characteristics. The term known characteristics is intuitively
understood as known numerically, and this understanding underlies the VNA calibration
methods based on fully known standards, such as the SOLT method [6]. However, this term
can also be understood as known in terms of some relationships governing the standards
characteristics. This alternative way of dening calibration standards is the foundation of
VNA calibration methods based on the self-calibration of standards.
The principle of self-calibration is to use the relationships between calibration-standard
S-parameters as an additional piece of information in the VNA calibration. Then, as a
side result of the VNA calibration procedure, some missing parameters of the calibration
standards, encapsulated in the set {cn }N
n=1 , are determined. From these parameters the
numerical value of these characteristics, as dened by {sn }N
n=1 , can be further derived.
A ag VNA calibration method employing the principle of self-calibration is the TRL
method [29, 67]. In this method, three dierent types of standards are used: a fully known
direct thru connection, a transmission line with known length and unknown propagation
constant, and an unknown reective standard for which we only assume that it is presents
identical reection coecient on both VNA ports. Then, in the course of calibration, along
with VNA model parameters, the unknown propagation constant of the line standard and
the unknown reection coecient of the reect standard are determined. As a result, with
the use of TRL calibration, one can calibrate the two-port VNA to measure S-parameters with reference to an unknown characteristic impedance of the transmission line. This
39

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


impedance is then determined based on the propagation constant and lines cross-sectional
parameters (see Paragraph 2.4.2-C).
A dual calibration with respect to the multiline TRL method is the multi-reect thru
calibration of [38], in which at least three oset terminations and a thru connection are
used. In this calibration the thru connection is also assumed to be fully known. As for the
oset terminations, we assume that their lengths are known and that these sections are
formed out of a transmission line with the same propagation constant and characteristic
impedance. The reection coecient of the termination at the end of each oset section is
assumed to be unknown but identical for all oset sections. Similarly to the TRL method,
in the course of calibration, along with VNA model parameters, the unknown propagation
constant of the oset sections and the the reection coecient of the termination are
determined. Also, an additional step is needed to determine the characteristic impedance
of the oset sections.
A generalization of the self-calibration principle for the two-port VNA calibration methods is discussed in [30], where other self-calibration methods, such as the Line-ReectMatch (TRM) are proposed. In this method a fully known matched-termination standard
is used instead of a line standard. The extension of the TRM method is the LRM methods
(thru replaced with a transmission line) and the LRRM method (some parameters of the
match standard are determined in the process of self-calibration) [80]. Also a very interesting implementation of the self-calibration principle is the unknown thru method [69]
which uses as a calibration standard a reciprocal but otherwise unknown two-port, referred
to as the unknown thru.
The main advantage of the self-calibration methods is that they require less information
about the calibration standards. Also, since this information is expressed in terms of some
physical relationships between the standards characteristics instead of numerical values,
self calibration methods are, in general, less prone to measurement errors than the methods
employing fully dened calibration standards. Consequently, self-calibration methods in
combination with the statistical approach (e.g. [3638]), lead to the most accurate VNA
calibration techniques.

B. Primary and transfer calibration standards. Primary calibration standards are


dened through their physical models and some fundamental parameters, such as geometrical dimensions. Transfer calibration standards are dened directly in terms of their
S-parameters which are obtained either from measurements or electromagnetic simulations.
40

2.4. TWO-PORT VNA CALIBRATION TECHNIQUES


Examples of primary calibration standards are the transmission line in the TRL calibration [29, 67] or an oset termination in the multi-reect thru calibration [38]. Another
interesting primary standard is the direct thru connection of VNA ports. This standard is
only applicable in connectorized measurements, such as with the rectangular waveguides or
coaxial transmission lines. An example of a transfer standard is the matched termination
for which we usually do not have a physical model.
C. Calibration reference impedance. The objective of VNA calibration is to determine VNA model parameters such that the corrected DUT S-parameters are expressed
with respect to some known reference impedance Zref , usually 50 . In the VNA calibration methods which use fully dened calibration standards, this reference impedance
is set by the reference impedance with respect to which S-parameters of the calibration
standards are given. Clearly, this reference impedance needs to be identical for all of the
calibration standards, otherwise a systematic error is introduced.
However, when using self-calibration methods, this reference impedance is often unknown. This is the case of self-calibration methods which use sections of a transmission
lines with an unknown characteristic impedance Z0 as a part of calibration standards (e.g.,
multiline TRL of [36] and multi-reect thru method of [38]). We can easily show that
the VNA calibrated with these methods measures DUT S-parameters with respect to an
unknown characteristic impedance Z0 of the transmission line. Therefore an additional
procedure is required for determining Z0 so that DUT S-parameters can be expressed with
respect to a typical reference impedance, such as 50 .
Much has been written on the issue of measuring the characteristic impedance of transmission lines, in particular in the case of on-wafer calibrations [81, 82]. For low-loss transmission lines, which we deal with here, there exist two simple methods. In these methods,
one assumes the line has small dielectric losses and its capacitance per unit length does
not signicantly change with frequency. The characteristic impedance can then be approximately written as [81]

Z0 =
,
(2.96)
jC0
where is the propagation constant of the lines determined in the calibration, and C0
is the quasi-static capacitance per-unit-length of the lines. This capacitance can then
be determined either based on dimensional parameters of the line, such as in the case of
precision coaxial air-dielectric lines [83], or based on the reection coecients measurement
of a resistor with known DC resistance [82]. With the use of the relationship (2.19), the
41

2. PRINCIPLES OF VNA S -PARAMETER MEASUREMENTS


reference impedance of corrected DUT S-parameters can then be reset from Z0 to a desired
value.

2.4.3

Solution

VNA calibration methods can also be classied with respect to the mathematical methods used to solve the VNA calibration problem. All of the deterministic methods and some
of the statistical methods (such as [36]) use a close-form analytical solutions. Most of the
statistical methods, such as [33, 37, 38, 56, 59], use iterative algorithms which are usually
dierent variations of the nonlinear least-squares minimization.
The main disadvantage of the iterative solution of VNA calibration problem, as with
all optimization-based methods, is that it might not nd a global minimum of the cost
function. Therefore special attention needs to be paid to a robust problem parametrization
(see Section 2.3.1-C), and a good choice of the starting point.

42

Chapter 3
Overview of uncertainty analysis for
VNA S-parameter measurements
The word chance then expresses only our
ignorance of the causes of the phenomena
that we observe to occur and to succeed
one another in no apparent order. Probability is relative in part to this ignorance,
and in part to our knowledge.
Pierre Simon de Laplace

Although the VNA calibration and correction procedure allows to remove the major
systematic errors in the VNA measurement, corrected S-parameter measurements are still
deteriorated with some residual measurement errors. These errors are caused by the imperfection of the VNA calibration, and by errors in raw VNA S-parameter measurements
(see Section 3.1).
The measurement error is in general quantied with the measurement uncertainty. In
the case of scalar measurements, there exist well established methods for representation
and evaluation of the measurement uncertainty [84, 85]. However, these methods cannot
be directly applied to complex- and matrix-valued measurand, such as the S-parameters.
Therefore, multiple conventions for representing the measurement uncertainty in S-parameters, extending the methods presented in [84, 85], are used. We review these conventions
in Section 3.2.
43

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS


When evaluating the measurement uncertainty in corrected VNA S-parameter measurements, we typically proceed in three steps. We rst develop statistical models for
phenomena responsible for the imperfection of the VNA calibration and for the errors in
raw VNA measurements (see Section 3.3). With the use of dierent uncertainty analysis
methods (see Section 3.5 and Section 3.6), these errors are the mapped into equivalent
errors in the VNA calibration coecients. These equivalent errors, typically represented
with VNA residual-error model (see Section 3.4), are then propagated into the resulting
errors in corrected VNA S-parameter measurements.

3.1

Sources of error in corrected VNA S-parameter


measurements

The VNA calibration procedure allows to characterize and subsequently remove the
systematic errors of the VNA test-set. The resulting corrected VNA S-parameter measurements are, however, still subject to some measurement errors. These errors results
from inaccurate calibration-standard denitions and from VNA instrumentation errors
that occur during the measurement of both the calibration standards and the device under
test (DUT).
Errors in the denitions of calibration standards can be classied into two groups:
numerical errors and inconsistency errors. Numerical errors are pertinent in the case of
calibration standards that are treated as fully known. These errors result from the limited
accuracy with which the values of some parameters of the calibration standard were determined. Examples here are uncertainties in the length measurement of transmission lines
in the TRL calibration, or the non-zero reection coecient of the matched termination
standard in the SOLT calibration
Inconsistency errors stem from the violation of some assumptions made as to the characteristics of one or more calibration standards. These errors occur in the case of calibration
standards to which we apply the self-calibration principle (see Paragraph 2.4.2-A), that
is, calibration standards that are partially unknown. Examples here are dierences in the
characteristic impedance of the lines or a variation in the connector interface discontinuity
in the coaxial multi-line TRL calibration.
VNA instrumentation errors can be divided into two groups: nonstationarity errors
and receiver errors. The nonstationarity errors manifest themselves as changes of VNA
electrical parameters with time. These changes are caused by the VNA test-set drift
44

3.2. STATISTICAL DESCRIPTION OF S -PARAMETER MEASUREMENT ERRORS


(primarily due to temperature and humidity changes), bending of the connecting cables,
and connector nonrepeatability.
The VNA receiver errors result from the noise and nonlinear eects in the measurement
of the waves a1m , b1m , a2m , and b2m (see Fig. 2.1). The noise in these measurements is
usually classied into low-level noise and high-level noise (also referred to as the trace
jitter) [19]. The low-level noise results from the wideband thermal noise in the VNA
receivers. The high-level noise is caused by the narrowband thermal noise in the IF part
of the receiver, and by the phase-noise of the LO.
The nonlinearity errors are caused by the nonlinearities of the VNA receivers. In modern VNAs, the impact of these nonlinearities is minimized through some corrections and
adaptive signal level adjustment. Therefore, in typical operating conditions, the nonlinearity errors can be neglected [19, 86].

3.2

Statistical description of S-parameter measurement errors

Scattering-parameters measurement forms a complex-valued matrix. Uncertainty evaluation for a measurement represented in such a form is very inconvenient, since standard
statistical methods operate on real-valued vector random variables [87]. Therefore, in the
context of the uncertainty analysis, we typically represent S-parameter measurements as
a real-valued vectors and describe their uncertainty with a covariance matrix [40, 42]. In
the following, we review this representation starting with a general statistical model for
S-parameter measurements, and following with the details of this model for the case of
measurement of a single S-parameter and a matrix of S-parameters.

3.2.1

Statistical model for S-parameter measurement

Our statistical model for S-parameter measurement is based on the measurement model
for vector-valued measurands (see [88]). We represent the S-parameter measurement as
vector by simply stacking the S-parameters in some arbitrary order. One preferred choice
of this order follows from the vectorization operator (2.52), and will be discussed in more
detail in Subsection 3.2.3.
We dene a single frequency measurement s of S-parameters of N -port device as a sum
45

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS


of the true value
s of the S-parameters and the measurement error s
s =
s + s,

(3.1)

where the underline denotes the real-valued vector representation of a complex-valued


vector dened by (A.2). We assume further that the true value
s is constant, and that the
2
measurement error s is a vector of 2N random variables
s = [s1 , . . . , s2N 2 ]T .

(3.2)

The statistical properties of the joint probability density function (PDF) fs (x) of the
measurement error s are discussed in the following sections. For now we only assume
that the measurement error has zero expectation value, that is, E (s) = 0, and hence
E (s) =
s.

3.2.2

Error description for a single S-parameter

For a single scattering parameter Sij , the model (3.1) reduces to


Sij = Sij + S ij ,
where

(3.3)

ReSij
S ij =
,
ImSij

(3.4)

is the real/imaginary part representation of the measurement error. This representation


is the most fundamental description of errors in single S-parameter measurements. It
usually very well corresponds to the underlying physical origins of the measurement error
and leads to an intuitive statistical description [42]. However, this representation does
not have a straightforward physical interpretation. Indeed, when interpreting S-parameter
measurement, instead of the real and imaginary part, we rather talk about the magnitude
and phase which correspond to the simple physical concepts of the attenuation and phase
shift. Therefore, the uncertainty in S-parameters is often alternatively represented in
terms of errors in magnitude and phase. However, the unapparently complicated statistical
properties of this representation may lead to incorrect uncertainty estimates [42]. We
discuss these properties below and show that the in-phase/quadrature error representation
oers a combination of the statistical simplicity of the real/imaginary part representation
46

3.2. STATISTICAL DESCRIPTION OF S-PARAMETER MEASUREMENT ERRORS


and straightforward physical interpretation of the magnitude/phase representation.
The statistical properties of all the representation can
be explained based on the graph shown in the Fig. 3.1. In
this graph we show a complex-valued measurand z that
is perturbed with a measurement error z. We assume
that z has a two-dimensional PDF dened in terms
of its real and imaginary part with Ez = 0, and the
covariance matrix is

Var (Rez)
Cov (Rez, Imz)
RI =
.
Cov (Rez, Imz)
Var (Imz)
(3.5) Fig. 3.1: Error representation
As a result of the perturbation z, we obtain a new value for complex measurands.
z  = z + z. A change in the magnitude and phase of z
can be easily derived as


|z| = |z  | |z| = |z| + zI + jzQ  |z|,


and
= arg z  arg z = arctan

zQ
,
|z| + zI

(3.6)

(3.7)

where zI and zQ and are the in-phase and quadrature error components, respectively,
dened as
zI = cos Rez + sin Imz,
(3.8)
and
zQ = sin Rez + cos Imz.

(3.9)

We see that the transformation between the real/imaginary and magnitude/phase representation is in general nonlinear. In order to better understand statistical properties of this
transformation, we shall investigate two limit cases, that is, when |z|  |z| and when
|z| |z|.
The case when |z|  |z| corresponds the situation when a measurement of a large
reection coecient (e.g., short or open), or a large transmission (e.g., line or thru) is
subject to small measurement errors. Equations (3.6) and (3.7) can then approximately
be written as
|z| zI ,
(3.10)
47

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS


and

zQ
.
|z|

(3.11)

In this situation, the relationship between all the error representations is linear. We can
then write the covariance matrix of vector zM A = [|z|, ]T as
M A = M(|z|)R()RI R()T M(|z|)T ,
where

cos sin
R() =
,
sin cos
and

(3.12)

1
M(|z|) =

(3.13)

1/|z|

(3.14)

The case |z| |z| corresponds to the situation when measuring devices with a small
reection coecient (e.g. matched termination), or small transmission (e.g. attenuator)
with large measurement errors. In this case, equations (3.6) and (3.7) take on a highly
nonlinear form. The statistical distribution of |z| and has then complicated properties,
hence statistical inference based on |z| and for |z| |z| requires special care [42].
To illustrate that, consider the case when |z| 0. We obtain then
|z| = |z  | |z| |zI + jzQ | ,

(3.15)

zQ
.
zI

(3.16)

and
= arg z  arg z arctan

Now, if RI corresponds to a Gaussian distribution with E (z) = 0, then |z| has a


Rayleigh distribution [42]. This distribution is dened only for positive values, therefore we
obtain E (|z|) > 0 for E (z) = 0. Consequently, for example, when averaging magnitude
of matched termination measurements pertubed with noise, we would counterintuitively
obtain a non-zero value.
As an alternative to the magnitude/phase representation, the in-phase/quadrature representation has been proposed [89]. This representation is dened by a xed linear transformation (3.8) and (3.9) of the real/imaginary representation. The covariance matrix for
48

3.2. STATISTICAL DESCRIPTION OF S-PARAMETER MEASUREMENT ERRORS


vector zIQ = [zI , zQ ]T can be written as
IQ = R()RI R()T .

(3.17)

As equations (3.10) and (3.11) demonstrate, for |z|  |z| the in-phase/quadrature representation has a similar interpretation as the magnitude/phase representation. The in-phase
component corresponds then directly to the magnitude change, and the quadrature component is the length of an arc corresponding the rotation of vector with magnitude
|z|. As opposed to the magnitude/phase representation, however, the in-phase/quadrature
representation maintains its linear relationship to the real/imaginary representation for
|z| |z|.

3.2.3

Error description for a matrix of S-parameters

In the case of matrix of S-parameters, the measurement uncertainty is represented with


a covariance matrix containing 2N 2 real elements where N is the number of devices ports.
This matrix consists of the variances of the uncertainty components (i.e. real and imaginary
part, magnitude and phase, or in-phase and quadrature part) of individual S-parameters
and covariances of all possible pairs of these components [4042].
The particular structure of this covariance matrix depends on the ordering of real and
imaginary part of S-parameters in the vector representation of S-parameters. The most
common convention, which we adopted in this work, is based on the vectorization operator
vec () [57]. To illustrate this convention, we apply the denition (2.52) of this operator to
a two-port scattering matrix S of size 2 2, to obtain a vector with four complex elements

s = vec (S) = vec

S11 S12
=
S21 S22

S11
S21
S12
S22

(3.18)

The real-valued representation of complex variables in s (such as with the real/imaginary or


magnitude/phase components) can be obtained in multiple ways. One common convention,
dened by (A.2), expands each element of the complex vector into its real and imaginary
part, that is, for (3.18) we obtain
s=

ReS11 ImS11 ReS21 ImS21 ReS22 ImS22


49

T

(3.19)

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS


With the use of (3.19), we can apply the denition of the covariance matrix (see [87]) to
the measurement error s in the vector s, obtaining

Var (ReS11 )
Cov (ReS11 , ImS11 ) Cov (ReS11 , ImS22 )

Var (ImS11 )
Cov (ImS11 , ImS22 )
Cov (ImS11 , ReS11 )

.
s =
.
..
..

.
.
.

.
.
.
.

Cov (ImS22 , ReS11 ) Cov (ImS22 , ImS11 )


Var (ImS22 )
(3.20)
Conversion to the in-phase/quadrature representation can easily be obtained as

R(11 )

IQ
s =

R(11 )

R(21 )
R(12 )

R(21 )

R(22 )

R(12 )

R(22 )
(3.21)

where matrices R(ij ) are dened by (3.13) and ij = arg Sij .

3.3

Statistical models for errors in corrected VNA Sparameter measurements

In this section we briey review statistical models used to describe errors in VNA
S-parameter measurements. Such models are used in the statistical methods for VNA
calibration to determine the appropriate weighting functions for the residual errors (see
Paragraph 2.4.1-A) and are a prerequisite for the evaluation of uncertainties in corrected
VNA S-parameter measurements.
In Section 3.1, we divided the errors in corrected VNA S-parameter measurements
according to their origin, that is, into the calibration standard errors and VNA instrumentation errors. In the context of the statistical description of these errors, however,
it is more convenient to talk about the systematic and random errors. Systematic errors
do not change between experiments and result from the limited accuracy with which we
one can determine parameters of the measurement set up. Random errors change between
experiments and result from random variations of some parameters of the measurement
set up. In the case of VNA S-parameter measurements, systematic errors are primarily related to inaccuracy of the calibration-standard denitions or their inconsistency (see
50

3.3. STATISTICS OF ERRORS IN CORRECTED VNA MEASUREMENTS


Subsection 3.1). The random errors result mainly from VNA instrumentation errors, such
as connector nonrepeatability, cable exure, and the test-set drift 1 . In the evaluation
of random errors, we typically use statistical methods based on repeated measurements
(Type A uncertainty evaluation [84]) to determine these properties, while in the case of
systematic errors, other methods are used (Type B uncertainty evaluation [84]).

3.3.1

Systematic errors

As already mentioned, systematic errors in corrected VNA S-parameter measurements


are primarily related to inaccuracies of the calibration-standard denitions. In the case of
primary calibration standards (see Paragraph 2.4.2-B), their denitions are typically developed based on the physical description of the calibration standards and their dimensional
and material parameters [90]. The uncertainties of these parameters are then propagated
into the uncertainties of the S-parameter denitions with the use of either the linear error
propagation (e.g. [36, 41] ) or the Monte Carlo approach (e.g. [77]). In the case of transfer
standards, the systematic errors in their S-parameter denitions are derived directly from
the uncertainties of the measurement technique used to characterize them. Often times, in
particular when specifying the accuracy of the calibration standards, these uncertainties
are expressed in a simplied fashion, for example, by use of the magnitude/phase uncertainty representation, by neglecting the statistical correlations between uncertainties, or by
approximating their statistical properties with the circular-normal PDF (see Section B.6).

3.3.2

Random errors

A. Connector repeatability and cable instability. The impact of connector nonrepeatability and cable instability errors is usually described jointly with a single statistical
models. This model typically assumes the circular-normal PDF and statistical independence of the errors at dierent frequencies [56, 71, 76]. It is also typically assumed that
the connector repeatability and cable instability errors aect independently each of the
scattering parameters [56].
This typical approach, however, has no justication in the physics of the connector
repeatability and cable instability errors. Reference [91] shows that the connector repeatability errors exhibit a very regular frequency dependence, therefore errors at dierent
1

Position of the center conductor in coaxial air-lines, or the plunger reection coecient in the sliding
load and sliding short are exceptions here: these errors are random error but aect the calibration standard
denition.

51

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS


frequencies cannot be assumed statistically independent. Reference [92] develops a simple equivalent circuit describing the connector repeatability errors in terms the electrical
parameters of the connector discontinuity. Subsequently, it shows that the connector repeatability errors strongly depend on the reection coecient the DUT. Reference [93]
demonstrates also that the PDF of the connector repeatability errors, as determined from
repeated reection-coecient measurements, has typically a PDF with one of the variances
much smaller then the other. Consequently, some calibration approaches (e.g., [75]) used
repeated measurements to determine the full covariance matrix of the connector repeatability errors.
Another description of the connector repeatability errors, applicable primarily when determining the weights in the statistical VNA calibration methods, has been proposed in [36].
This approach approximates the connector interface variation with a linear two-port with
random S-parameters described by the circular-normal PDF. The connector repeatability
errors in the actual measurements of a two-port device are then determined by attaching
such a connector-error two-port at either end of the device. Thus, the resulting error in
the DUT S-parameters due to the connector nonrepeatability depends also on the DUT
S-parameters themselves, which is often observed in the actual measurements [91]. Consequently, the approach of [36] allows also to capture the relationships between the connector
repeatability errors in dierent calibration standards, and thus to more accurately specify
the weighting matrices in the statistical calibration methods (see Paragraph 2.4.1-A).

B. Test-set drift. The impact of the VNA test-set drift has rarely been treated in
the literature. The approach of Reference [94] (see Subsection 3.5.2) is often used to
approximately quantify the impact of the VNA test-set drift. Another method, suggested
by Reference [95], uses repeated measurements of dierent verication standards. Based
on those measurements, the worst case estimate of errors due to the test-set drift is derived.

C. Receiver noise and nonlinearity. Similarly to the VNA test-set drift, the VNA
receiver errors are rarely considered. Reference [56] suggests that the VNA receiver noise
can be approximated with the circular-normal PDF which seems to well correspond to the
actual operation of the VNA receivers. Reference [56] proposes also a simple model for the
dependence of the the VNA receiver errors on the raw S-parameters. This dependence is
the consequence of the fact, that the raw S-parameters are obtained as ratios of complex
voltages measured by the VNA receivers. Following that reasoning, Reference [96] suggests
52

3.3. REPRESENTATION OF ERRORS IN CORRECTED VNA MEASUREMENTS


that the VNA receiver errors should be characterized directly in terms of VNA receiver
voltages instead of raw S-parameter measurements.
Regarding the receiver nonlinearity errors, Reference [86], based on an experimental
study, develops an approximate method to predict the impact of receiver nonlinearities.
It also points out that special attention needs to be paid to the linear operation of VNA
receivers as the nonlinear eects can nonintuitively alter the corrected S-parameter measurements.

3.4

Representation of errors in corrected VNA Sparameter measurements

In this section, we present a mathematical model for errors in corrected VNA S-parameter measurements. This model consists of two components: description of errors caused
by the imperfection of the VNA calibration (errors in VNA calibration coecients) and
description of errors due to noise and nonlinearities of VNA receivers (errors in raw VNA
measurements). We begin with an overview of how the errors discussed in Section 3.1 aect
the VNA S-parameter measurements (see Subsection 3.4.1). Following on that, we present
a detailed analysis of how errors in VNA calibration coecients (Subsection 3.4.2) and in
VNA raw measurements (Subsection 3.4.3) propagate into the corrected VNA S-parameter
measurements.

3.4.1

Overview

We illustrate sources of error in corrected VNA S-parameter measurements in Fig. 3.2.


In this gure we show the ow graph of the 8-term VNA model (compare Fig. 2.5) with
additional terms related to errors discussed in Section 3.1. We assume that errors related
to the switch operation (see Subsection 2.3.2) and cross-talk between VNA ports are either
not present or have been perfectly corrected for. This is a viable approximation, since the
errors introduced by the switch are typically very stable (see Subsection 2.3.2), while the
cross talk between the VNA ports is neglegible in modern VNAs (see Paragraph 2.3.1-B).
In Fig. 3.2 the error sources are represented with two sets of parameters. The rst
set describes the errors in the determination of VNA calibration coecients and can be
53

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS

Fig. 3.2: Basic error model for corrected two-port VNA S-parameter measurements.
written as a vector of perturbations


! "T

p = EDF , ESF , ERF , EDR , ESR , ERR ,

(3.22)

added to the nominal set of VNA calibration coecients (2.71). The second set describes
errors in VNA raw measurements and is represented as an additional matrix

11 12
,
=
21 22

(3.23)

added to the raw measurement (2.44).

3.4.2

Errors in VNA calibration coecients

For the eight term model (see Subsection 2.3.1-B) with the basic parametrization dened by (2.71), we can determine the eect of these errors by analyzing the sensitivity of
(2.45) to changes described by (3.22). This is, however, inconvenient, as it requires taking
derivatives of (2.45) with respect to submatrices dened by (2.64). Therefore we follow a
dierent approach, based on the residual error-box representation (see [6]).
In this representation, one uses a set of additional linear two-port networks, referred
to as residual error-boxes, that are attached to the actual error boxes representing VNA
calibration coecients. The residual error-boxes can be thought of as a set of VNA calibration coecients that one would obtain in an error-free calibration, performed after the
actual calibration. Thus, we referr to the parameters of the residual error-boxes as the
eective VNA parameters.
With the use of residual error-boxes, one can then easily determine the errors in both
corrected VNA S-parameter measurement and VNA calibration coecients, by attaching
54

3.4. REPRESENTATION OF ERRORS IN CORRECTED VNA MEASUREMENTS

(a)

(b)

Fig. 3.3: Error models for corrected one-port VNA S-parameter measurements: (a) basic
model, (b) residual error-box representation.

the residual error-boxes to the network representing the DUT or the actual VNA errorboxes, respectively. We rst discuss the residual-error box representation for the case of
one-port VNA measurements and then show its extension to the the two-port case.

A. One-port measurement. Flow graph for a one-port VNA with errors in calibration
coecients and raw measurement and the equivalent residual error-box representation are
shown in Fig. 3.3. Since the reection coecient measurement depends only on the product
of forward and reverse transmission through the error box (ERF and ERR in Fig. 2.4), we
lump this product into a single parameter ER .
In the ow graph in Fig. 3.3a, parameters ED , ER , and ES denote errors in the
determination of VNA calibration coecients and corresponds to the overall error in
one-port-VNA raw measurement. The residual error-box representation in Fig. 3.3b allows
to easily determine the error in corrected reection coecient measurements as
(1 + rR )2
= rD +
rD + 2rR + rS 2 .
1 rS

(3.24)

The parameters rD , rS , and rR are referred to as the eective directivity, eective


source-match, and eective tracking of the calibrated one-port VNA. In order to determine
these parameters, we compare the two ow graphs in Fig. 3.3. We assume the that errors
ED , ER , and ES are small which, after applying the standard methods for ow-graph
55

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS

Fig. 3.4: Residual error-box representation of the error model for corrected two-port VNA
S-parameter measurement.
analysis, leads to a rst-order approximation
ED ER rD ,

(3.25)

ES rS + 2ES rR + ES2 rD ,

(3.26)

ER 2ER rR + 2ER ES rD ,

(3.27)

which after some straightforward manipulations yields


ED
,
ER
1 ER
ES

ED ,
rR
2 ER
ER
ES
ES2
ER +
ED .
rS ES
ER
ER

rD

(3.28)
(3.29)
(3.30)

In the case when VNA systematic errors are small, that is, |ES | 0 and |ES /ER | 0,
we can further simplify these expressions to obtain
ED
,
ER
1 ER
rR
,
2 ER

rD

rS ES .

(3.31)
(3.32)
(3.33)

This last set of expression shows that in the case when the VNA systematic errors are
small, the changes in VNA calibration coecients can be represented as an additional
linear reciprocal two-port, whose parameters depends only on those changes and the single
VNA calibration coecient ER .

56

3.4. REPRESENTATION OF ERRORS IN CORRECTED VNA MEASUREMENTS


B. Two-port measurement. Flow graph of the residual error-box representation of
the error model from Fig. 3.2 is shown in Fig. 3.4. This representation is dened by
a forward (parameters rDF , rSF , rRF ) and reverse (parameters rDR , rSR , rRR , and rK )
residual error-box. The parameters rDF , rSF , rRF are referred to as the forward eective
directivity, forward eective source-match, and forward eective tracking, respectively,
while the parameters rDR , rSR , rRR , are the reverse eective directivity, reverse eective
source-match, and reverse eective tracking, respectively. The parameter rK is referred
to as the eective non-reciprocity factor rK . We decided to lump the this factor into the
reverse residual error-box, however, it can be also added to the forward residual error-box.
Based on the ow graph in Fig. 3.4 and under the assumption that the errors are small,
we can determine errors in corrected S-parameter measurements as

2
S11 rDF + 2rRF S11 + rSF S11
+ rSR S21 S12 ,

(3.34)

2
S22 rDR + 2rRR S22 + rSR S22
+ rSF S21 S12 ,

(3.35)

S21 (rRF + rRR + rK + rSF S11 + rSR S22 ) S21 ,

(3.36)

S12 (rRF + rRR rK + rSF S11 + rSR S22 ) S12 .

(3.37)

In order to obtain parameters of the residual error-boxes, we compare the ow graphs


in Fig. 3.2 and Fig. 3.4. We assume the that errors in (3.22) are small which, after applying
the standard methods for ow-graph analysis, leads to rst-order approximations
EDF
,
ERF
1 ERF
ESF

EDF ,
2 ERF
ERF
ESF
E2
ESF
ERF + SF EDF ,
ERF
ERF

rDF

(3.38)

rRF

(3.39)

rSF

(3.40)

and
EDR
,
ERR
1 ERR
ESR

EDR ,
2 ERR
ERR
ESR
E2
ESR
ERR + SR EDR ,
ERR
ERR

rDR

(3.41)

rRR

(3.42)

rSR

57

(3.43)

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS


with
! "

1 ERR ESR
rK =
+
+
EDR .

2 ERR
ERR

(3.44)

In the case when VNA systematic errors are relatively small, that is, |ESF | 0, |ESR |
0, |ESF /ERF | 0, and ESR /ERR | 0, we can further simplify these expressions to obtain
EDF
,
ERF
1 ERF

,
2 ERF

rDF

(3.45)

rRF

(3.46)

rSF ESF .

(3.47)

and
EDR
,
ERR
1 ERR

,
2 ERR

rDR

(3.48)

rRR

(3.49)

rSR ESR ,

(3.50)

with
! "

1 ERR
rK =
+
.

2 ERR

(3.51)

Similarly to the one-port VNA case, this last set of expression shows that when the VNA
systematic errors are small, changes in VNA calibration coecients can be represented as
additional linear reciprocal two-ports, whose parameters depend only on those changes and
on the calibration coecients ERF and ERR respectively.

3.4.3

Errors in VNA raw measurements

The description of VNA receiver errors is based on a set of perturbations added to the
raw VNA S-parameter measurement. These perturbations form a joint description of VNA
receiver noise and nonlinearities. These errors cannot be accounted for with the residual
error-box representation, hence we need to directly determine derivatives of (2.45). To this
end, we rst represent the dierentials in DUT S-parameters and their measurement, S
and Sm , respectively, in a vector form by use of the vectorization operator [57], dened
58

3.4. REPRESENTATION OF ERRORS IN CORRECTED VNA MEASUREMENTS


by (2.52). We then obtain

vec (S) =

S11
S21
S12
S22

and vec (Sm ) =

11
21
12
22

(3.52)

By applying the rules of matrix calculus we then obtain

S11
S21
S12
S22

11
21
12
22

(3.53)

where
J=





vec (S)
T
1
T
T
=
S

S
S

S
(S

S
)

(S

S
)
,
21e
m
11e
m
11e
12e
vec (Sm )T
T

(3.54)


where denotes the Kronecker product [57] and AT is a short form for (A1 ) = AT
We can further simplify the second term in (3.54) by inserting (2.44), that is,


1
S22e S1
(Sm S11e )1 = S1
21e S
12e ,

and

1

(3.55)

T
ST22e ST
(Sm S11e )T = ST
12e S
21e .

(3.56)

With the use of identities (AB) (CD) = (A C) (B D) and (A B)1 = A1 B1


we can further simplify this term
(Sm S11e )T (Sm S11e )1 =


1
= ST
12e S21e

 #

= ST12e S21e

ST ST22e ST
21e

1 



S1 S22e S1
12e

ST ST22e S1 S22e

 

$

ST21e S12e

1

, (3.57)

which after insertion into (3.54) gives




 

J = ST S

ST ST22e S1 S22e
59

 

ST21e S12e

1

(3.58)

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS


and after expansion simplies to




J = I4 ST ST22e I2 I2 (SS22e ) + ST S

ST22e S22e

 

ST21e S12e

1

(3.59)
where I2 and I4 are identity matrice of size 2 2 and 4 4, respectively.
Expression (3.59) is general and can be applied to both the 16-term and the 8-term
model. In the case of the 8-term model, under the assumption that the VNA has small
reections, that is, |ESF | 0, |ESR | 0, |ESF /ERF | 0, and |ESR /ERR | 0 and that
the DUT has small transmissions (S21 0 and S12 0), we obtain
2
2
1 2ESF S11 + ESF
S11
,
ERF
2
2
1 2ESR S11 + ESR
S11
22
,
ERR
1 ESF S11 ESR S22 + ESF ESR S11 S22
21
,

ERR
1 ESF S11 ESR S22 + ESF ESR S11 S22
12
.

ERF

S11 11

(3.60)

S22

(3.61)

S21
S12

(3.62)
(3.63)

Comparison with (3.38) through (3.43) shows that errors in reection coecient measurement due to errors 11 and 22 in VNA raw measurements can then be described with an
equivalent eective directivity equal to the VNA receiver error, that is, rDF = 11 and
rDR = 22 .
In the case when the DUT is well matched (S11 0 and S22 0) and the VNA has
small reections, we obtain
11
,
ERF
22

,
ERR
21

,
ERR
12

.
ERF

S11

(3.64)

S22

(3.65)

S21
S12

(3.66)
(3.67)

In this case, we see that the resulting error in reection coecient measurement can be
seen as VNA receiver errors 11 and 22 scaled by inverse of the round-trip transmission
though the relevant errors boxes, ERF and ERR . The error in transmission measurements
60

3.5. APPROXIMATE UNCERTAINTY EVALUATION


can be interpreted as VNA receiver error 21 and 12 scaled by the inverse of the rst-order
overall transmission through both error boxes.

3.5

Approximate uncertainty evaluation

The approximate uncertainty analysis techniques allow to determine parameters of the


residual error-box model (see Subsection 3.4.2) for a previously calibrated VNA. These
techniques use either an experimental or analytical approach to estimate how the actual
random and systematic errors in the calibration standards and VNA raw measurements
transform into the parameters of the residual error-box model. Thus, approximate uncertainty evaluation techniques cannot be used to perform a complete uncertainty analysis.
In such a analysis, instead of one particular realization of systematic and random errors,
we are rather interested in the statistical distribution of measurement errors, induced by
multiple realizations of underlying errors in the calibration standards and the raw VNA
measurement. In other words, we are interested in a distribution of these errors that would
be observe in a hypothetical experiment when a large number of VNA calibrations were
performed with dierent realizations of the same set of calibration standards. Therefore,
the approximate uncertainty evaluation techniques, although oering a very useful information about the error of one particular VNA calibration, cannot provide the information
about the statistical distribution of this error. They cannot also be used to estimate the
uncertainty due the errors occurring during corrected VNA measurements, such as the
errors due to imperfect DUT connector repeatability or cable exure.
The other disadvantage of the approximate techniques is that are not traceable. Traceability requires in general that the uncertainties determined in an uncertainty analysis of
a measurement procedure be traceable back to uncertainties of some fundamental physical
standards, such as length or impedance standards [97, 98]. In other words, it requires
that these measurement uncertainties be expressed as a function of the uncertainties of the
fundamental standards and the model of the measurement procedure. For example, in the
case of transmission lines used as calibration standards in the TRL calibration [29], the
uncertainty in the length measurement of this lines should be traceable to the uncertainty
of some fundamental length standards. The residual error-box parameters determined with
the approximate techniques originate either from a direct measurement or mathematical
treatment of the calibration problem and cannot be traced back to the uncertainties of
some fundamental standards, such as length or impedance standard.
61

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS

3.5.1

Ripple analysis techniques

Ripple analysis technique is probably the earliest method for an approximate evaluation
of VNA calibration accuracy [99]. It is primarily used in the analysis for one-port VNA
calibrations [95]. Extensions to the two-port case were also proposed [100], however, we
will not discuss them here.
Ripple analysis technique is an experimental method. It uses a long section of a uniform
transmission line, terminated either with a well-matched or a highly-reective termination.
A calibrated measurement of such a verication device reveals a standing-wave pattern
(ripples), parameters of which can be transformed into the equivalent parameters of
the residual error-box model [95]. This transformation can be done in dierent ways. In
the simplest case from the ripple amplitude one can determine the magnitude of residual
directivity (line terminated with a matched termination) and residual source match (line
terminated with a highly reective termination). More complex analysis techniques allow
to determine the complete residual error-box model [101, 102].

3.5.2

Calibration comparison method

Calibration comparison method [94] was originally developed as a tool for the uncertainty evaluation in on-wafer VNA calibrations. Subsequently, however, it has also been
applied to the uncertainty evaluation of VNA calibration in other environments (e.g.,[103]).
The principle of this method is to analyze the dierence between two calibrationsa benchmark calibration and a calibration corrupted with errorsand then estimate the resulting
worst-case error in the DUT measurement.
The dierence between the two calibration in this method is expressed with the use
of the residual error-box representation (see Subsection 3.4.2). The worst-case estimate of
error in the DUT S-parameters is then determined as follows. Consider a DUT described
with S-parameters Sij , for i, j = 1, 2 and cascaded with the residual error-boxes to obtain
error-corrupted S-parameters Sij , for i, j = 1, 2. We dene the worst-case error as [94]
= max |Sij Sij |.
|Sij |1
i,j=1,2

(3.68)

Assuming that the dierence between the calibrations is small, rst order approximation
for (3.68) can be derived [94].
62

3.5. APPROXIMATE UNCERTAINTY EVALUATION


The worst case estimate (3.68) can be thought of as single-number measure of the
deviation captured in the residual error-boxes. It has been reported, however, that this
estimate is too conservative when applied to the measurement of actual passive DUTs [104].
This can be attributed to the fact, the range of DUTs for which the worst case error is
determined, specied by the constraints in (3.68), includes also non-physical devices, e.g.,
when S11 = S21 = 1.

3.5.3

Statistical residual analysis

The principle of the statistical residual analysis is to assess how well the measurement
modelthat is the VNA calibration coecients along with the denitions of the calibration
standardsdescribes the actual measurements of the calibration standards.The inadequacy
of the measurement model is then transformed into the equivalent uncertainties of the VNA
calibration coecients. These uncertainties can futher be used to predict the accuracy of
corrected VNA S-parameter measurements.
The inadequacy of the VNA measurement model can only be observed in statistical
VNA calibration methods, that is, when a redundant number of calibration standards
is measured (see Paragraph 2.4.1-A). The most common formulation of statistical VNA
calibration is based on the least-squares estimation. In the formulation (2.86) with s =
2 Vs , the inadequacy of the VNA measurement model (2.84) is quantied with the sum
1
r () ,
S () = r ()T Vs

(3.69)

where r () are the residual errors of the VNA calibration model, given by (2.85), are
VNA calibration model parameters dened by (2.83), s = 2 Vs is the covariance
matrix of errors in calibration standard denitions given in (2.87), Vs is a known matrix,
and 2 is the unknown scaling factor, referred to as the residual variance. The estimate
of VNA calibration model parameters (which contains also the unknown parameters of

calibration standards if the calibration method employs the principle of self-calibration;


see Paragraph 2.4.2-A) is obtained from (2.86) and the estimate of the residual variance is
determined as (see Subsection B.2.1)

2 =

1
S(),

(3.70)

where is the number of degrees of freedom, that is, the number of equations less the
63

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS


number of estimated parameters. With the use of above denition, we can derive the
caused by the inadequacy of the measurement
estimate of covariance matrix of errors in
model. This estimate can be written as (see see Subsection B.3.1)


1

T 1 Jr ()

2 Jr ()
r

(3.71)

is the Jacobian of r () calculated at ,


that is
and Jr ()



= r () 
Jr ()
.
T =

(3.72)

Statistical properties of errors in VNA calibration coecients, as captured in covariance


matrix (3.71), can be further transformed into the residual error-box representation with
the use of relationships derived in 3.4.2. With this representation we can then estimate
the resulting error in corrected VNA S-parameter measurements.
Statistical residual analysis is the most rigorous technique to estimate the contribution
of systematic and random errors in a particular set of calibration standards and raw VNA
measurements. It is typically built into the calibration software based on statistical methods [56, 59], which allows to quickly verify the quality of the VNA calibration. However,
similarly to other approximate methods, statistical residual analysis does not allow for
complete uncertainty evaluation and is not traceable.

3.6

Complete uncertainty evaluation

Complete uncertainty analysis approaches are based on a rigorous analysis of how the
calibration standard errors and VNA instrumentation errors propagate into the corrected
VNA measurement. These approaches can be divided into two groups: techniques based
on the linear error propagation [19, 73, 105109], and techniques employing the numerical
Monte Carlo simulation [110, 111]. These two groups of techniques are also referred to as
the uncertainty propagation and distribution propagation, respectively. This nomenclature
reects the fact that in the linear error propagation one analyzes only the rst two moments
of the input and output error distributions, while in the Monte Carlo simulation the entire
probability density function of the input and output error is determined.
64

3.6. COMPLETE UNCERTAINTY EVALUATION

3.6.1

Linear error propagation

The techniques based on linear error propagation usually employ analytical relationship between the errors in corrected VNA measurements and the underlying calibrationstandard errors and VNA instrumentation errors. These relationships are based on the
rst-order Taylor expansion of the function describing the VNA measurement process.
This expansion has usually a complicated form, closely bound to a specic VNA calibration procedure and a VNA model. Therefore, whenever the model of the VNA measurement changes, a signicant eort needs to be put into deriving these relationships anew
(see [36, 108, 109, 112, 113]). This makes these techniques very cumbersome in practical
applications.
In order to determine the uncertainties in the denitions of the calibration standards,
these techniques usually employ a physical model of the standard (see [6, 90]). Based on
that model and uncertainties of its parameters, such as lengths, diameters, or material
parameters, the uncertainty in the S-parameters of the standards is then calculated. This
provides a traceability path for the resulting uncertainties [97]. In order to quantify the
VNA instrumentation errors, usually a statistical analysis based on a large number of
repeated measurements is used [41, 95, 114].
Probably the most advanced uncertainty analysis approach based on the analytical
techniques has been developed at NIST [40, 41]. This approach has been developed for the
TRL calibration [29] in the coaxial 7 mm standard applied to a 2 18 GHz dual six-port
measurement system [115]. Apart from developing analytical expressions for the VNA
measurement uncertainties, this approach features a very rigorous statistical treatment of
the measurement results based on the covariance matrix description.

3.6.2

Monte-Carlo simulation

Monte Carlo simulation is a well-established numerical technique for uncertainty estimation [116]. The principle of this technique is to treat the measurement problem as
an arbitrary function transforming input quantities (in the case of VNA measurements
calibration standard denitions, their raw measurements, and the raw measurements of
the DUT) into output quantities (in the case of VNA measurementscorrected DUT Sparameters). The input quantities are then randomly varied according to the statistical
distribution of their errors, and the respective responses of the output quantities are collected. Based on that, statistical distribution of the output quantities is determined, from
65

3. OVERVIEW OF UNCERTAINTY ANALYSIS FOR VNA MEASUREMENTS


which dierent uncertainty measures can be derived.
The key advantages of numerical approaches based on the Monte Carlo simulation
are the statistical rigorousness and exibility. Indeed, in the Monte Carlo simulation no
assumptions need to be made as to the linearity of error propagation, and an arbitrary
measurement model (in the case of VNA measurementscombination of the VNA model
and VNA calibration procedure) can easily be analyzed. This makes the Monte Carlo
approach particularly attractive for complicated problems for which deriving analytical
equations would be infeasible.
The main disadvantage of the Monte Carlo simulation is that it is very time consuming
and requires huge computer resources. The time consumption becomes particularly critical
when estimating covariance matrices, as the number of Monte Carlo iterations needs to
signicantly exceed the number of estimated parameters [87, 116]. Therefore, the Monte
Carlo approaches are often used to verify uncertainty analyzes based on the linear error
propagation.

66

Chapter 4
Multi-frequency description of
S-parameter measurement errors
Logic, after all, is a trick devised by the human
mind to solve certain types of problems.
Richard Bellman

In this chapter, we introduce the multi-frequency description of errors in VNA Sparameter measurements. This description generalizes the conventional single-frequency
covariance matrix representation (see Subsection 3.2.3) by accounting for the statistical
correlations between errors at dierent frequencies. These statistical correlations result
from the fact that the VNA S-parameter measurement errors at dierent frequencies have
some common physical causes.
The discussion we present in this chapter has two objectives. First of all, we shall
develop the mathematical and physical foundations for the multi-frequency description of
errors in S-parameter measurements. Secondly, we shall identify the implications of such
an extended description for practical applications. A more detailed analysis of some of
those applications will then be given in the remainder of this work.
We start out by introducing the statistical model for the multi-frequency VNA S-parameter measurement. We then review the notion of the physical error mechanism and discuss
statistical properties of the S-parameter measurement error dened jointly for multiple
frequencies. Subsequently, we briey review the aspects of uncertainty reporting specic
to the multi-frequency error description. Following on that, we discuss the physical error
67

4. MULTI-FREQUENCY ERROR DESCRIPTION IN VNA MEASUREMENTS


mechanisms in the VNA S-parameter measurements and nally we analyze the practical
implications of the generalized multi-frequency error description we introduced.

4.1

Statistical model for the multi-frequency Sparameter measurement

Our statistical model for the multi-frequency S-parameter measurement is an extension


of the model introduced in Subsection 3.2.1. Following (3.1), we dene measurement s of
S-parameters of N -port device for a set of frequencies {fk }K
k=1 as a sum of the true value

s of the S-parameters and the measurement error s, that is,


s =
s + s,
where

s
1
..
s = . ,

(4.1)

(4.2)

sK
the underline denotes the real-valued vector representation of a complex-valued vector
dened by (A.2), and sk , for k = 1, . . . , K is the real-valued vector representation of
S-parameters at the frequency fk . The vector s has thus Q = 2N 2 K elements.
We assume further that the true value
s is constant, and that the measurement error
s is a vector of Q random variables
s = [s1 , . . . , sQ ]T .

(4.3)

The statistical properties of the joint PDF fs (x) of the measurement error s are discussed in the following sections. For now, we only assume that the measurement error has
zero expectation value, that is, E (s) = 0, thus E (s) =
s.

4.2

The notion of a physical error mechanism

The notion of physical error mechanism is essential for our multi-frequency description
of errors in S-parameter measurements. It reects the intuition that the overall measurement error is caused by some common fundamental error mechanisms. The fundamental
68

4.2. THE NOTION OF A PHYSICAL ERROR MECHANISM


character of these mechanisms manifests itself in the fact that they give the simplest physical explanation of the underlying causes of the measurement error. For example, in the
case of VNA S-parameter measurements, the fundamental error mechanisms correspond
to both systematic measurement errors (e.g., uncertainties on dimensional and material
parameters of the calibration standards) and random measurement errors (e.g., bending
of the cables, misalignment of inner and outer conductors or displacements of the inner
conductor ngers in coaxial connectors, VNA test-set drift, or VNA receiver noise and nonlinearities). Consequently, we see that the fundamental error mechanisms characterizing
the causes of the measurement error, have physical character, and are statistically uncorrelated and frequency independent. In the following, we referr to the error mechanisms
fullling these three conditions as the physical error mechanisms.
We dene a single physical error mechanism as a scalar random variable that describes
the variability of an underlying physical parameter characterizing the mechanism, and a
function
s = m () ,
(4.4)
which represents a physical model that relates the parameter and the corresponding
error s in the S-parameter measurement. Since the physical error mechanism characterize
changes, we assume that E () = 0, m (0) = 0. We further dene the variance Var () = 2 .
Typically, the random variable has a Gaussian or uniform probability density function.
The function m () is in general nonlinear. However, assuming that the variance 2 is
small, we can approximate m () with the rst order Taylor expansion around = 0, that
is,
s j,
(4.5)
where

m () 

j=
.
=0

(4.6)

The vector (4.6) can be thought of as a nominal response of the measurement s to the
error mechanism .
We further represent a set of M physical error mechanisms with the vector
= [1 , . . . , M ]T ,
69

(4.7)

4. MULTI-FREQUENCY ERROR DESCRIPTION IN VNA MEASUREMENTS


the covariance matrix

21

..

.
2M

(4.8)

and the Jacobian matrix


M = [j1 , . . . , jM ] .

(4.9)

The matrix (4.8) is diagonal due to the assumption of statistical independence of the
mechanisms in the vector (4.7). With the use (4.9), we can eventually approximate the
error in S-parameter measurement as
s = M.

(4.10)

The properties of the matrix M require some additional remarks. Based on the discussion at the beginning of this chapter, a natural assumption would be that the columns
of the matrix M are linearly independent, or in other words, that we can distinguish between the contributions of dierent physical error mechanisms to the measurement error.
In practice, however, we often encounter the situation when dierent physical mechanism
lead to indistinguishable contributions. For example, in the case of a coaxial transmission
line, we cannot distinguish in the frequency-dependent S-parameters of the line between
the eect of the same relative change in the inner and outer conductor diameter.
When the contributions of some of the physical error mechanisms to the measurement
error s are indistinguishable, the columns of the matrix M become linearly dependent,
thus, the matrix M is rank decient. In order to avoid that, we introduce the concept of
electrically-equivalent physical error mechanisms. We write = Q  , where  is the set of
electrically-equivalent physical error mechanisms, and dene the measurement error as
s = MQ  ,

(4.11)

where we require the matrix MQ to be full rank. Columns of the matrix Q constitute the
basis for the range of the matrix M and can be easily determined based on the physical
considerations. For example, coming back to the example of the coaxial transmission
line, the eect of a change in the inner and outer conductor diameter can be equivalently
captured in terms of a change of the characteristic impedance of the line.
In the following, unless otherwise noted, we assume that the physical error mechanisms
70

4.2. STATISTICS OF THE MULTI-FREQUENCY MEASUREMENT ERROR


describing the measurement error are distinguishable. We always denote the error mechanisms with the vector which, depending on the context, refers either to the actual or
electrically-equivalent physical error mechanisms.

4.3

Statistical properties of the multi-frequency measurement error

We now investigate the statistical properties of the measurement error as dened by


(4.10). We rst rewrite (4.10) with the use of real-valued vectors. To this end, we use the
conventions (A.2) and (A.6) to write
s = M,

(4.12)

and consequently, we write the covariance matrix of (4.12) as




s = E M (M)T = M MT ,

(4.13)

where the matrix s has form

s =

Cov(s1 , s2 ) Cov(s1 , sQ )
Var(s2 )
Cov(s2 , sQ )
..
..
...
.
.
Cov(sQ , s1 ) Cov(sQ , s2 )
Var(sQ )
Var(s1 )
Cov(s2 , s1 )
..
.

(4.14)

where Cov(si , sj ) = Cov(sj , i), for i = j and i, j = 1, . . . , Q. Now, we note that the
length Q of the measurement error s, in most cases, is much smaller than the number
M of physical error mechanisms. Consequently, although the matrix (4.8) has size Q Q,
its rank is only M . Hence, the matrix (4.8) is singular. In the following we discuss
the consequences of this fact related to the form of the PDF for s and reporting the
uncertainties described with s , and nally we take a closer look at the structure of the
matrix s .

4.3.1

Probability distribution function

Assuming that the measurement error has the normal probability distribution, we cannot write the PDF of s in the typical form (see [87]) due to singularity of s . The
71

4. MULTI-FREQUENCY ERROR DESCRIPTION IN VNA MEASUREMENTS


probability distribution function for a normal vector random variable with a mean
and a singular covariance matrix has the form [117]
1

f (x) =
(2)

M
2

M
%
i=1

1
2

e 2 (x)

(x)T

(4.15)

where {m }M
m=1 are the positive eigenvalues of and denotes any generalized inverse of
the matrix . The distribution (4.15) is also referred to as the singular normal distribution
[118, 119]. The generalized inverse A of a matrix A is dened as a matrix fullling the
condition (see [120])
AA A = A.
(4.16)

A well known generalized matrix inverse is the Moore-Penrose inverse A+ , also referred
to as the matrix pseudo-inverse, which, additionally to (4.16), fullls the conditions



AA+

T
T

A+ A

= AA+ ,

(4.17)

= A+ A,

(4.18)

A+ AA+ = A.

(4.19)

Unlike the generalized inverse dened by (4.16), the Moore-Penrose inverse is uniquely
dened for a given matrix A [120]. For a full rank matrix A we can easily show that
A+ = A1 by simply inserting A1 into the conditions (4.16) through (4.19) 1 .
In the case of the random variable s dened by (4.12), the probability distribution
function takes on the form
fs (x) =

1
(2)

M
2

e 2 (M x)
1

1
2

det( ) det(MT M) 2

1
M x

(4.20)

where M is any generalized inverse of M. By comparing (4.20) and (4.15) we see, that in
%
order to justify the form (4.20), we have to prove two conjectures: that the product M
i=1 i


T
of the positive eigenvalues of s is det ( ) det M M , and that any generalized inverse


of the matrix s can be written as


s = M

T

1
M .

Another, more intuitive, denition of the Moore-Penrose inverse is A+ = arg minX AX I F , where
F is the Frobenius matrix norm (square root of the sum of all of the matrix elements squared) and I is
the identity matrix [121].

72

4.3. STATISTICS OF THE MULTI-FREQUENCY MEASUREMENT ERROR


Consider the singular-value decomposition of the matrix M, that is [121],
M = USVT ,

(4.21)

where U RQQ and V RM M are orthogonal square matrices, and S RQM is a


rectangular matrix containing on the diagonal the M singular values of the matrix M, that
is,

s1

...

,
sM

S=

(4.22)

where is an empty matrix. The singular values are non-zeros as the matrix M is assumed
to be full rank [121]. We can easily show that the eigenvalues of the matrix MT M are




equal to squared singular values in S, consequently det MT M = det ST S . Inserting
(4.21) into (4.13) and postmultiplying by US we obtain


s US = US VT VST S = USP,

(4.23)

where P = VT VST S. Now, consider the spectral decomposition of the matrix P


P = XX1 ,

(4.24)

where X1 = XT since the matrix P is symmetric. Inserting (4.24) into (4.23) and postmultiplying by X we obtain
s USX = USX,
(4.25)
which expresses the condition for the eigenvalue decomposition (with only positive eigenvalues) of s : the eigenvalues are contained in the diagonal matrix and the corresponding
eigenvectors are the columns of the product USX. The eigenvalues of P are all positive
and real since P is full rank and symmetric. These eigenvalues constitute all of the positive eigenvalues of s since rank s = rank P = M . The product of the eigenvalues






%
T
T
T
contained in is M

=
det
(P)
=
det
V

V
det
S
S
=
det
(
)
det
M
M
,
i

m=1
which proves the rst conjecture.
As to the second conjecture, based on the denition of the generalized inverse (4.16),
73

4. MULTI-FREQUENCY ERROR DESCRIPTION IN VNA MEASUREMENTS


we need to show that


M M M

T

T
T
1
M M M = M M ,

(4.26)

for any generalized inverse M . This requires that M M = IM for any generalized inverse
M which is only true if the matrix M is full rank. Indeed, taking the singular value
decomposition of the matrix M, we can easily show that
M = VS UT

(4.27)

fullls the condition (4.16). With the use of (4.27), we can write
M M = VS SVT .

(4.28)

Now, considering the diagonal structure of the matrix S, as shown in (4.22), and accounting
for the assumption that singular values are non-zero (i.e., the matrix M is full rank), we
see that the condition (4.16) is equivalent to S S = IM , thus M M = IM .

4.3.2

Uncertainty reporting

Second important consequence of the singularity of the matrix s concerns estimating


condence regions. The typical methods for the condence region estimation (e.g., [88])
require the inversion of the covariance matrix, hence cannot be used with s . Consequently, some other methods for reporting measurement uncertainty need to be devised,
based on projecting the uncertainty regions in the space of the physical error mechanisms
onto the space of the actual measurement errors (e.g. see [122, 123]). This is, however,
beyond the scope of this work.
In practice, however, we are often interested in the uncertainties of only single elements
of the vector s. Examples include the real and imaginary part of an S-parameter at
a particular frequency. The condence region then reduces to a typical one-dimensional
interval and can be determined with the use of classical uncertainty evaluation methods
(see [88]).
74

4.3. PHYSICAL ERROR MECHANISMS IN VNA MEASUREMENTS

4.3.3

Multi-frequency covariance-matrix structure

We shall now look in more detail at the structure of (4.14). Matrix (4.14) contains
the variances and covariances for all components of the vector (4.3), that is, for real and
imaginary parts of all of the S-parameters of the N -port device for frequencies fk , for
k = 1, . . . , K. We can rewrite it as a block matrix

s =

f1 ,f1
f2 ,f1
..
.

f1 ,f2
f2 ,f2
..
.

fK ,f1 fK ,f2

. . . f1 ,fK
. . . f2 ,fK
..
...
.
. . . fK ,fK

(4.29)

where all of the sub-matrices have size 2N 2 2N 2 . The diagonal sub-matrices fk ,fk are
covariance-matrices for the single-frequency measurement errors sk , and have the same
structure as the single-frequency covariance matrix dened in Subsection 3.2.3. In the
following, we referr to those matrices in short as fk . The o-diagonal sub-matrices fk ,fl
dene the correlations between measurement error in S-parameters at dierent frequencies,
that is, for vectors sk and sl , for k = l. Note that if there is no statistical correlations
between S-parameter measurement errors at dierent frequencies, the o-diagonal matrices
fk ,fl have all entries equal zero.

4.4

Physical error mechanisms in VNA S-parameter


measurements

The source of errors in corrected VNA S-parameter measurements and their statistical
models have been discussed in Section 3.1 and Section 3.3, respectively. In the following,
we shall take another look at these errors, however, from the perspective of the underlying
physical error mechanisms.

4.4.1

Calibration standard errors

The error mechanisms in the calibration standards are identied based on physical
models of the standards. These models express S-parameters of calibration standards in
terms of some fundamental dimensional and material parameters. For example, in the
case of the coaxial transmission lines, these parameters are the lengths and diameters of
75

4. MULTI-FREQUENCY ERROR DESCRIPTION IN VNA MEASUREMENTS


the conductors. The uncertainties of the measurements of these fundamental parameters
become then the estimates for standard deviations of the error mechanisms , while the
relationships resulting from the calibration standard physical models give the functions
m (). We discuss this in more detail in Chapter 5.

4.4.2

VNA instrumentation errors

A. Nonstationarity. A physical-mechanism-based description of the VNA nonstationarity errors poses a complex problem. In practice, it is dicult to characterize the VNA
nonstationarity errors with analytical models derived from fundamental mechanical and
electrical parameters of the VNA. For example, in the case of connector repeatability errors, this would require careful mechanical characterization of all of the possible mechanical
displacements in the connector interface, and then electro-mechanical modeling of their inuence on the interface S-parameters. Although such an approach has been applied to
simplied connector models (e.g. see [124126]), it is extremely dicult to model real connector structures. The situation is even more dicult with the cable instability or VNA
test-set driftanalytical modeling of electrical parameters of such complex structures is
beyond our capacity.
On the other hand, as demonstrated in [91], the connector repeatability errors exhibit
a very regular frequency dependence which can be modeled with a simple lumped-element
equivalent circuit. This observation is a foundation of our approach to the description
of the VNA nonstationarity which is presented in Chapter 6. We show there that the
VNA nonstationarity errors can be characterized in terms of a very small set of frequency
independent error mechanisms and xed functions capturing the frequency dependence
of this disturbance errors. Consequently, we represent the VNA nonstationarity errors as
sk = m (sk , fk , )

(4.30)

which, under the assumption that errors are small, can be approximately written as
sk M (sk , fk ) ,
where the matrix M (sk , fk ) is dened by (4.9).

76

(4.31)

4.4. PHYSICAL ERROR MECHANISMS IN VNA MEASUREMENTS


B. Receiver noise. As discussed in Section 3.1, the VNA receiver noise is caused by the
thermal noise and phase noise in the receivers. Since the VNA performs the measurement
at one frequency at-a-time, the noise realizations at dierent measurement frequencies
can be considered statistically independent2 . Consequently, the VNA receiver noise can
be described with a separate set of physical error mechanisms m (fk ) at each frequency
fk . The statistical properties of these mechanisms depend on the raw S-parameters (or
VNA receiver voltages), which need to be accounted for in the physical model of the error
mechanism. Consequently, at a given measurement frequency fk , we postulate the following
description of raw S-parameter measurement error
m
m
m
sm
k = m (sk , fk , k )

(4.32)

which, under the assumption that error are small, can approximately be written as
m
m
m
sm
k M (sk , fk ) k ,

(4.33)

where the matrix M (sm


k , fk ) is dened by (4.9).

C. Receiver nonlinearity. The nonlinearity of VNA receivers introduces errors of a


systematic nature as their impact is clearly repeatable for the same measurement conditions. Following a similar reasoning as for the VNA receiver noise, we can assume there
are no statistical correlations between the VNA receiver nonlinearity errors at dierent frequencies. Provided that a model for the nonlinear behavior of VNA receivers is available,
these errors could be quantied and eventually corrected for. A preliminary analysis of
[86] could be a starting point for the development of such a model. However, as the VNA
receiver nonlinearity errors are negligible under typical VNA operating conditions, we will
not consider them in more detail in this work.

The exception here is the measurement in the so-called ramp mode, available in some VNAs. In
this mode, the source is swept continuously without stopping at each measurement frequency. Hence,
we can expect the noise for adjacent measurement frequencies to be statistically correlated. However, in
the laboratory-precision VNA measurements, we avoid operation in the ramp mode due to the loss of
accuracy caused by the increased measurement speed.

77

4. MULTI-FREQUENCY ERROR DESCRIPTION IN VNA MEASUREMENTS

4.5

Practical implications

In this section, we discuss the practical implications of the multi-frequency approach to


the description of S-parameter measurement errors. These implications include the errormechanism-based approach to the VNA calibration, uncertainty analysis in time-domain
waveform correction, and device modeling based on S-parameter measurements.

4.5.1

Time-domain waveform correction

VNA S-parameter measurements are often used in the calibration of time-domain measurement systems such as high speed oscilloscopes [127] or electro-optic sampling systems
[44]. In the calibration process of such systems, systematic errors of some of the system
components, such as adapters, pulse sources, oscilloscopes, or on-wafer probes are characterized with S-parameters. These errors are then removed from the actual time-domain
measurement in a correction procedure. In the following, we present the details of timedomain waveform correction and demonstrate the importance of the multi-frequency error
description is the uncertainty analysis of this procedure. The exposition presented below
generalizes the discussion presented in [39, 128]. Reference [39] shows also experimental
result that illustrate the role of the multi-frequency error description in the uncertainty
analysis of the time-domain waveform correction.
A. Introduction. A typical situation of time-domain waveform correction is shown in
Fig. 4.1. A waveform source, for example, a pulse generator, is connected through a
two-port device to a load, for example, an oscilloscope. In the frequency domain, the
source is characterized with the wave amplitude bS (j) and the input reection coecient
S (j). The corresponding voltage delivered by the source to a matched termination

can be determined from (2.36) and (2.37) as VS (j) = bS (j) Zref (j) where Zref (j)
is the real reference impedance, typically 50 . The load is characterized with a load
reection coecient L (j). We measure the voltage VL (j) at the load which is related

to the forward wave amplitude aL (j) through VL (j) = aL (j) Zref (j) 3 . Finally,
the two-port device is characterized with the scattering parameters.
In the time-domain waveform correction problem, we are interested in determining the
time-domain waveform vS (t) based on the measurement of other time-domain waveform
3

In the following, for the sake of notational simplicity, we skip the explicit indication of the frequency
dependence.

78

4.5. PRACTICAL IMPLICATIONS

Fig. 4.1: Time-domain waveform source connected to a load through a two-port device.
vL (t), and the measurements of the two-port S-parameters and the reection coecients
of the source and load, S and L , respectively. Based on the analysis of the ow graph
in Fig. 4.1, one can easily show that the relationships between these parameters in the
frequency domain has the form [128]
VL = HVS ,

(4.34)

where the factor H is dened by


H=

1 S S11 L S22 L S [S21 S12 S11 S22 ]


.
S21

(4.35)

In the case without the two-port device (that is S11 = S22 = 0 and S21 = S12 = 1), this
factor reduces
H = 1 L S ,
(4.36)
which corresponds to mismatch correction.
In order to convert (4.34) into the time domain, we rst introduce some notational
conventions. We represent the frequency-domain measurements of VL and VS , and the
correction factor H as the vectors

VL [0]

..

vL =

VL [N ]

,

VL [N ]

..

VL [1]

VS [0]

..

vS =

VS [N ]

,

VS [N ]

..

VS [1]

H [0]

..

and h =

H [N ]

,

H [N ]

..

(4.37)

H [1]

where VL [n] = VL (jn), VS [n] = VS (jn), and H [n] = H(jn), for n = 0, . . . , N ,


= 2f it the angular frequency step, the star denotes the complex conjugate, and
79

4. MULTI-FREQUENCY ERROR DESCRIPTION IN VNA MEASUREMENTS


N + 1 is the number of positive frequencies (including the DC). The conversion into the
time-domain requires the knowledge of the spectrum for both the positive and negative frequencies. However, since all of the waveforms considered here are real-valued, the spectrum
for negative frequencies is the complex conjugate of the spectrum for positive frequencies
[129]. This is explicitly shown in the vectors in (4.37) which have M = 2N + 1 elements.
We further represent the corresponding time-domain waveforms as vectors

vL [0]
..
.

vL =

vL [M 1]

vS =

vS [0]
..
.
vS [M 1]

and h =

h [0]
..
.
h [M 1]

(4.38)

where vL [n] = vL (nt), vS [n] = vS (nt) , and h [n] = h (nt) , for n = 0, . . . , M 1, and
.
t = M2

The relationship between the frequency-domain spectrum and time-domain waveform


is now determined by the Fourier series. For a frequency domain spectrum X (jn) and
the time domain waveform x (nt), this relationship has the form [129]
x (nt) =

X (jn) ejnkt .

(4.39)

k=

In the case of truncated frequency data, as described by the vectors (4.37), equation (4.39)
reduces to the inverse discrete Fourier transform [129]
x [n] =

M


X [k] ejnk M ,

(4.40)

k=0

which is typically evaluated with the use of the Fast Fourier Transform (FFT) algorithm
[129].
Now, in order to write the relationship between the time-domain waveforms (4.38), we
apply (4.40) to (4.34). With the use of the convolution theorem, we obtain [129]
vL [n] =

M
1


h [k] vS [(n k) mod M ] =

k=0

M
1


vs [k] h [(n k) mod M ]

(4.41)

k=0

which describes the circular convolution of discrete series vS [n] and h [n], and the operator
x mod y is the remainder of the division of two integer numbers x and y.

80

4.5. PRACTICAL IMPLICATIONS


B. Matrix formulation. The equation (4.41) can be used directly to perform the timedomain correction based on the measured waveform samples vs [n] and the time-domain
representation h [n] of the frequency-domain correction factor (4.35), as obtained through
(4.40). However, in order to perform the uncertainty analysis of this procedure, it is more
convenient to write (4.41) in a matrix notation as
vL = T (vS ) h,
where

T (vS ) =

(4.42)

vS [M 1] vS [M 2]
vS [0]
vS [M 1]
vS [1]
vS [0]
..
..
.
.
vS [M 1] vS [M 2] vS [M 3]
vS [0]
vS [1]
vS [2]
..
.

. . . vS [1]

. . . vS [2]

. . . vS [3]
,
..

...
.

. . . vS [0]

(4.43)

is the Toeplitz matrix (see [121]) constructed from the vector vS . In a similar manner, we
can obtain the relationship
vL = T (h) vS ,
(4.44)
where the Toeplitz matrix T (h) is dened analogously to (4.43).
We can further represent the inverse discrete Fourier transform (4.40) with the use of
matrix notation to obtain
(4.45)
h = Fh = FCM h ,
where F is the Vandermonde matrix [121]

F=

1
1
1
..
.

1
w
w2
..
.

1
w2
w4
..
.

...
...
...
...

1
wM 1
w2(M 1)
..
.

1 wM 1 w2(M 1) . . . w(M 1)(M 1)

(4.46)

w = ej M , CM is dened by (A.3), and the underline denotes the real-valued convention


for representation of complex vector dened by (A.2). Now, combining (4.42) and (4.45),
we eventually obtain
vL = T (vS ) FCM h .
(4.47)

81

4. MULTI-FREQUENCY ERROR DESCRIPTION IN VNA MEASUREMENTS


C. Uncertainty analysis. Equation (4.47) is an important results. It describes explicitly the relationship between the samples vL [n] of the de-embedded waveform and samples
H [n] of the correction factor spectrum. The correction factor spectrum is a function of
VNA S-parameter measurements of the two-port device and the source and load reection coecients. Uncertainty of this factor can be easily derived from the uncertainties
of S-parameter measurements with the use of linear error propagation. We denote this
uncertainty with a multi-frequency covariance matrix h .
Now, the uncertainty of the waveform samples vL [n], denoted with a covariance matrix
vL , can be determined directly from (4.47). However, we derive this covariance matrix
in two steps in order to better demonstrate the role of statistical correlation between
uncertainties at dierent frequencies in h . In the rst step, we write the relationship
between uncertainties in vL [n] and h [h]. This relationship can be easily derived from
(4.42) as [87]
vL = T (vS ) h T (vS )T .

(4.48)

The matrix T (vS ) depends on the waveform vS [n]. In the simplest case of vS [n] = [n],
this matrix becomes diagonal and then the uncertainty for each sample of vL [n] depends
only on the uncertainty of h [n] at the same time point. In a general case, however, when
vS [n] contains non-zero elements for n > 0, the uncertainty for each sample of vL [n]
depends on the uncertainties and statistical correlations between them for all samples in
h [n].
Now, the covariance matrix h can be derived from (4.45) as
h = FCM h (FCM )T .

(4.49)

The matrix F, as dened by (4.46), is dense. The matrix CM is block diagonal and represents only the transformation between the two dierent representations of the waveform
spectrum. Consequently, the product FCM is also a dense matrix. This property has an
important consequence. Examining closer (4.49), we note that the uncertainty of each time
domain sample h [n], thus also of each time domain sample vL [n], depends not only on
the single frequency variances in h but also on all of the statistical correlations between
them, including the correlations between uncertainties at dierent frequencies. Experimental results illustrating the importance of these correlations for the correctness of the
uncertainty analysis are given in [39].
82

4.5. PRACTICAL IMPLICATIONS

4.5.2

Device modeling based on S-parameter measurements

In the device modeling based on S-parameter measurements, we identify the parameters


of an electrical model of the device based on the measurement of the devices S-parameters.
Typically, these measurements are performed at multiple frequencies. Examples include
modeling of active devices, such as microwave transistors, and passive devices, such as
transmission line discontinuities.
The approaches used in the measurement-based device modeling fall into the category
of the system identication methods. The most common approach used in the system
identication is based on the maximum-likelihood estimation principle. In Appendix B,
we review the main aspects of the use of this principle in the system identication problems.
In this section, we briey demonstrate how to account for the multi-freqency description
of errors in S-parameter measurements in the maximum-likelihood system identication.
Consider a nonlinear system identication problem of the class described in Section B.2.1. The equations describing the measurements are given by

s1 = f (x1 , ) + s1
..
.

(4.50)

sN = f (xN , ) + sN

where {xn }N
sn }N
n=1 are the known system inputs, {
n=1 are multi-frequency S-parameter
measurements in a vector form, and {sn }N
n=1 are the measurement errors. We further
assume that the measurement errors are normally distributed with E (si ) = 0, and the


covariance matrix E sn sTn = sn which may be singular.
Following the methodology presented in Appendix B and by use of the denition (4.20),
we can write the log-likelihood function as
Mn
N
N 
N 
T

1
1
1

Mn
ln n,m
M
1
ln L () = ln (2)
n rsn ()
n Mn rsn (),
2
2
2
n=1
n=1 m=1
n=1
(4.51)
with the residuals dened by

rs1 () = s1 f (x1 , )
..
.
rsN () = sN f (xN , )
83

(4.52)

4. MULTI-FREQUENCY ERROR DESCRIPTION IN VNA MEASUREMENTS


which readily leads to the following least-squares optimization problem
= arg min

N 


T

M
n rsn ()

n=1

1
n Mn rsn (),

(4.53)

where M
n is an arbitrary generalized inverse of the matrix Mn . Now, consider the use
of the Moore-Penrose generalized inverse M+
of the matrix Mn . Since matrices Mn , for
n
1
T
n = 1, . . . , N are full rank, we have M+
=
M
M
MTn (see [121]), which yields
n
n
n
= arg min

N &


MTn Mn

1

'T

MTn rsn ()

n=1

T
1
n M n Mn

1

MTn rsn ().

(4.54)

Now, this result is dierent from the typical maximum-likelihood estimators discussed
in Appendix B which simply use the sum of squared residuals weighted by the inverse of the

1
covariance matrix. The term MTn Mn
MTn rsn () in (4.54) is the orthogonal projection
of the residual error rsn () onto the column space of the matrix Mn . In other words, this
term contains the coecients of the representation of the frequency-dependent residual
error in terms of the error mechanisms captured in the covariance matrix sn . Hence, the
minimizes the weighted sum of residuals expressed in terms of error mechanisms
estimator
rather then frequency-dependent S-parameters. This result can be easily extend to more
complex system identication models discussed in Appendix B.

4.5.3

Error-mechanism-based VNA calibration

In the error-mechanism-based VNA calibration, we recognize the fact that the measurement errors are caused by a set of physical error mechanisms. Consequently, the inconsistency of the calibration equations describing the measurement of a redundant number
of calibration standards is completely characterized by a realization of these mechanisms,
instead of some derived statistics expressed in terms of S-parameters (see Section 3.3).
This is shown in Fig. 4.2 which summarizes the inuence of dierent classes of VNA error
mechanisms discussed in Section 4.4 on the measurements of calibration standards. With
the use of the Fig. 4.2, we can rewrite the equations (2.82) describing the measurements
of calibration standards as
84

4.5. PRACTICAL IMPLICATIONS

Fig. 4.2: Physical error mechanisms in the VNA calibration.

m
m
sm
sm
sm
nk )
nk =
nk + mn ( nk ,
p
m
p

snk = f (snk + mn ( n , snk , fk ) , pk )

c
s = g (
nk
n c0n + n , cnk , fk )

, for n = 1, . . . , N, and k = 1, . . . , K.

(4.55)

where N is the number of calibration standards, K is the number of frequencies, the


indices n and k, refer to the calibration standard number and the measurement frequency
fk , respectively, and cn , pn and m
nk are the physical error mechanisms, to be explained
below. The unknown calibration parameters are dened by

.
= .. ,
K

(4.56)

where k denotes the solution vector to the single-frequency VNA calibration problem at
the frequency fk , for k = 1, . . . , K, given by (2.83) and repeated here for the sake of clarity

k =

pk
c1k
..
.
cN k

85

(4.57)

4. MULTI-FREQUENCY ERROR DESCRIPTION IN VNA MEASUREMENTS


Errors in the calibration standard denitions and their raw measurements, as dened in
Paragraph 2.4.1-A, can be expressed in terms of the physical error mechanism as
m
m
sm
snk = mnp ( pn , snk , fk ) , and sm
nk = mn ( nk ,
nk ) .

(4.58)

The physical error mechanisms are dened as a vector

c
p
m

(4.59)

where the subvectors c , p , and m contain the physical error mechanisms corresponding
to the calibration standard errors, VNA nonstationarity errors, and VNA receiver errors,
respectively, and are dened as

p1
m

..
1K
m

m
,
and

=

.
N1
.
pN
.
.

m
NK

c
1
. p

.. , =
c =
cN

m
11
.
..

(4.60)

The joint PDF of the error mechanism is f (). As the mechanisms are assumed to be
statistically independent (see Section 4.2), we have E () = 0 and the covariance matrix
of is given by (4.8). We further explicitly recognize the fact that the physical errors
mechanisms modeling the VNA receiver errors at dierent frequencies are statistically
independent (see Paragraph 4.4.2-B). Consequently, the vector m is made up of a set of
mechanisms for each measurement frequency.

Now, in order to formulate the maximum-likelihood solution of the set of equations


(4.55), we need to dene the residual errors (see Appendix B). The residual errors are
estimates of the error mechanisms, dened by (4.59) and (4.60), and are thus dened as

r=

rc
rp
rm

86

(4.61)

4.6. SUMMARY
and

rc =

rc1

..

p
,
r
=

rp1
..

m
. , and r =

rcN

rpN

rm
11

m
r1K
.

rm
N1
..

..
.

(4.62)

rm
NK

Now, the residuals need to satisfy the equations (4.55), from which we obtain the relationships

m
m
sm
sm
sm
nk =
nk + mn (rnk ,
nk )
m
p
p

s = f (snk + mn (rn , snk , fk ) , pk )


nk

c
s = g (
nk
n c0n + rn , cnk , fk )

, for n = 1, . . . , N, and k = 1, . . . , K..

(4.63)

With these denitions, we can write the maximum likelihood estimate of the unknown
residual errors r and the calibration parameters as
arg max f (r) ,
,r

(4.64)

under the equality constraints (4.63).


The remainder of this work (Chapter 5 and Chapter 6) is devoted to developing the necessary means for implementing the VNA calibration based on the formulation (4.64).

4.6

Summary

In this chapter, we developed the theoretical and physical foundations of the multifrequncy description of errors in S-parameter measurements. The key concepts we introduced are the physical error mechanism and the multi-frequency S-parameter measurement
error. As pointed out in [39], the main mathematical diculty with the multi-frequency
description of errors in S-parameter measurements is the singularity of the multi-frequency
covariance matrix. This singularity makes it impossible to use the traditional form of the
multivariate normal PDF. We showed that this PDF can be dened by use of the physical
error mechanisms rather then directly in terms of the S-parameter measurement them87

4. MULTI-FREQUENCY ERROR DESCRIPTION IN VNA MEASUREMENTS


selves.
We further discussed the practical implications of the multi-frequency description of
S-parameter measurement errors for calibrated time-domain measurements. Examples of
such measurements are oscilloscope measurement with the correction for the impedance
mistmatch or for the adapter parameters. We performed a detailed uncertainty analysis of a
generalized form of a time-domain-waveform correction procedures that employs S-parameter measurements. This analysis revealed that the statistical correlations between S-parameter measurement errors at dierent frequencies are essential for the correct evaluation of
uncertainties in such waveform correction procedures. Measurement examples illustrating
the importance these correlations are given in [39].
Subsequently, we analyzed the use of the multi-frequency description of errors in S-parameter measurements in the device modeling. We focused on the most common approach
based on the maximum-likelihood estimation. Our analysis yielded an intersting results
showing that when using the multi-frequency error description in the device modeling, the
mist between the model and the measurement (quantied with residuals) needs to be
written in terms of the physical error mechanisms rather the S-parameters.
Finally, we showed that the statistically sound description of the VNA calibration
problem should be done in terms of the error mechanisms underlying the calibration standard and VNA instrumentation errors. As a consequence, the VNA calibration should be
performed jointly at all measurement frequencies so as to account for the simultaneous
contribution of the error mechanisms to S-parameter measurements at dierent frequencies. We referr to this approach as the error-mechanism-based VNA calibration and in the
remainder of this work we develop some necessary tools for the implementation of such a
calibration approach. These tools include the generalized multi-frequency VNA calibration
(see Chapter 5) and the error-mechanism-based description of the VNA nonstationarity
errors (see Chapter 6) which are the primary source of VNA instrumentation errors.

88

Chapter 5
Generalized multi-frequency VNA
calibration
Pooh looked at his two paws. He knew one of
them was right, and he knew that when you had
decided which one of them was the right, then
the other one was left. But he never could remember how to begin.
A. A. Milne, House at the Pooh corner

In this chapter, we develop a generalized multi-frequency VNA calibration approach


and demonstrate it in the context of the multi-line TRL calibration method [36] in the
coaxial environment. In the multi-frequency approach to the VNA calibration (see Paragraph 2.4.1-B), the relationships between the calibration standard S-parameters at dierent frequencies are exploited as an additional piece of information in the VNA calibration.
Here we generalize this way of looking at the VNA calibration problem by using the concept of the physical error mechanism (see Section 4.2) to characterize these relationships.
We focus only on the mechanisms aecting the calibration standard denitions and use an
approximate description of the VNA instrumentation errors (see Section 3.3.2). However,
by use of the framework for the description of the VNA nonstationarity errors presented in
Chapter 6, the results presented here can be extended to establish a calibration approach
based entirely on the concept of the physical error mechanism (see Subsection 4.5.3).
We start out with the generalized multi-frequency formulation of the VNA calibration
problem and then proceed with a brief overview the coaxial multi-line TRL calibration.
89

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


Following on that, we discuss sources of errors in calibration standards used in the coaxial multi-line VNA calibration and based on that we establish the calibration standard
models. Compared to the classical multi-line TRL calibration, these models contain some
additional parameters which characterize the impact of the physical error mechanisms on
the calibration-standard S-parameters. These parameters, such as inner and outer conductor length errors or inner conductor displacement, are then identied along with the
VNA calibration coecients in the course of generalized multi-frequency VNA calibration.
In the last section, we present experimental verication of our calibration approach for
multi-line TRL calibration with Type-N and 1.85 mm coaxial transmission lines.

5.1

Formulation of the VNA calibration problem

Our generalized multi-frequency VNA calibration approach is a simplied variant of the


error-mechanisms-based calibration discussed in Subsection 4.5.3. The simplications we
make concern the description of the VNA instrumentation errors, and statistical properties
of the physical error mechanisms aecting the calibration standards.
As for the VNA instrumentation errors, we pointed out in Subsection 4.4.2 that identication of the physical error mechanisms responsible for these errors is fairly dicult.
Therefore, simplied statistical description of the VNA instrumentation error is often used,
such as the models discussed in Section 3.3. In our calibration approach, we adapt these
models to the multi-frequency formulation of the VNA calibration. In particular, we neglect
the VNA receiver errors, assume that the VNA nonstationarity errors at a single frequency
fk are normally distributed, and neglect the statistical correlations between those errors at
dierent frequencies. Following the notation in Fig. 4.2, we write the probability density
function (PDF) of the VNA nonstationarity errors as
N W/2

fsk (sk ) = (2)

1/2

|sk |

1
exp sTk 1
sk sk ,
2

(5.1)

where N is the number of calibration standards, and W is the number of real-valued


measurements obtained for each calibration standard. For the VNA with P ports, we
have W = 2P 2 . We further assume that the covariance matrix sk is known up to a
multiplicative factor, that is,
sk = k2 Vsk ,
(5.2)
where Vsk is a known matrix and k2 is the residual variance (square of the residual
90

5.1. FORMULATION OF THE VNA CALIBRATION PROBLEM


standard deviation).

Concerning the calibration standard errors, we employ the physical error mechanism
description discussed in Subsection 4.4.1. We assume that the physical error mechanisms
uniform PDF, that is, within the range of the parameters c we have the PDF fc ( c ) = C,
where C is a constant. Consequently, we can write the original problem (4.64) as
arg

K


max
C
fsk (rk |k2 ),
, rc , rp1 , . . . , rpK k=1
2
12 , . . . , K

(5.3)

under the equality constraints

s
sm
nk = f (snk + rnk , pk )
, for n = 1, . . . , N, and k = 1, . . . , K.
c
s = g (
nk
n c0n + rn , cn , fk )

(5.4)

Taking now the natural logarithm of the likelihood function, we can write this optimization
problem equivalently as

arg

2
max
ln L(, rc , rs1 , . . . , rsK , 12 , . . . , K
)
c s
s
, r , r1 , . . . , rK
2
12 , . . . , K

(5.5)

under the equality constraints (5.4), where the log-likelihood function, after dropping the
constant terms, is given by
2
)=
ln L(, rc , rs1 , . . . , rsK , 12 , . . . , K

K
K
1 sT 1 s
NW 
1
ln k2
r V r .
2 k=1
2 k=1 k2 k sk k

(5.6)

We further incorporate the constraints (5.4) into the log-likelihood function, which eventually yields the following problem
arg

2
ln L(, rc , 12 , . . . , K
),
max
c
2
2
, r , 1 , . . . , K

91

(5.7)

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


where the likelihood function is given by
2
)=
ln L(, rc , 12 , . . . , K

K
K
1 s
NW 
1
1 s
ln k2
r (, rc )T Vs
r (, rc ),
k k
2 k=1
2 k=1 k2 k

(5.8)

the underline denotes the convention (A.2), and the residual vector is dened as

rs (, rc )
1k

..

,
rsk (, rc ) =
.
rsN k (, rc )

(5.9)

and
c0n + rcn , cn , fk ) , for n = 1, . . . , N, and k = 1, . . . , K.
rsnk (, rc ) = f 1 (sm
nk , pk ) gn (
(5.10)
With the formulation (5.7), we generalize the multi-frequency approach to VNA calibration initiated in [37, 38] (see Paragraph (2.4.1-B)). Our formulation (5.7) diers from
the methods presented in References [37, 38] in that it does not require the complete knowledge of statistical properties of the errors sk . Instead, only the relationship between these
errors, as captured by the matrix Vsk , need to be known, and the absolute value of these
errors is determined from the residual variances k2 . In this manner, we follow the typical
approach used in the single-frequency statistical VNA calibration (see [58, 59] and the
discussion in Paragraph 2.82).
Furthermore, the formulation (5.7) along with the numerical technique described in
(5.6), is generic and can be easily adapted to any type of VNA calibration. This, again,
diers our approach from the methods presented in References [37, 38] which are dedicated
to a particular VNA calibration technique.

5.2

Coaxial multi-line VNA calibration

In this section, we briey review the coaxial multi-line TRL calibration. This calibration
serves as a workhorse for the demonstration of the generalized multi-frequency approach to
VNA calibration. We begin with a brief overview of the classical multi-line TRL method
of [36], and then discuss the implementation of the multi-line TRL method in the coaxial
environment. Following on that, we briey review some computational aspects of the
multi-line TRL method.
92

5.2. COAXIAL MULTI-LINE VNA CALIBRATION

5.2.1

Classical multi-line VNA calibration

The multi-line TRL calibration method [36] is an extension of the classical TRL method
[29, 67]. In the TRL method, the following calibration standards are used:
a section of a transmission line with known length and unknown characteristic
impedance and propagation constant,
a reect standard which is assumed to have the same but otherwise unknown reection coecient on both VNA ports,
a thru standard which is typically realized as a direct connection of the VNA ports1 .
In the course of the calibration, the unknown propagation constant of the line is determined
along with the VNA calibration coecients. Based on this propagation constant and the
capacitance per unit length of the transmission line, one can then determine the characteristic impedance of the line, and thus reset the reference impedance of the calibration to
an arbitrary value, typically 50 (see Paragraph 2.4.2-C).
Also, the unknown reection coecient of the reect standard is determined along with
the VNA calibration coecients. The value of this reection coecient is not necessary
to complete the calibration, however, it can be used to verify the calibration correctness,
provided that an priori estimate of this reection coecient is available.
In the multi-line TRL method, instead of a single transmission line, a set of multiple
transmission lines with dierent lengths is used. This allows to overcome the inherent
bandwidth limitation of the classical TRL method [29] and provides higher accuracy due to
redundant measurements [36]. Also, the use of redundant number of calibration standards
allows to employ the residual analysis (see Subsection 3.5.3) to detect random connection
errors or calibration-standard-denition errors.

5.2.2

Coaxial air-dielectric line as a calibration standard

The implementation of the multi-line TRL calibration in the coaxial environment is


based on transmission lines realized as air-dielectric lines (commonly referred to as airlines)
and the reect standard realized as an oset termination, that is, a termination with a
1

If such a connection cannot be realized (e.g., in on-wafer environments, or a non-insertable conguration of VNA test ports in coaxial connector environment), a short section of transmission line with a
dierent length than the base line is used instead. This modication of the TRL method is referred to as
LRL method [130].

93

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


section of an airline. The center conductor in airlines is not supported by plastic beads,
which are used in other coaxial devices, such as adapters or attenuators. The lack of beads
allows to better predict the S-parameters of the air-lines, because electrical parameters
of the beads cannot be well predicted nor controlled [131]. Connecting the air-lines with
an unsupported center conductor is, however, dicult and requires great operator skill,
particularly at smaller connector sizes.
As for the oset terminations, their center conductor is supported by the termination
itself, therefore connecting them is much simpler and less prone to accidental damaging.
Consequently, calibration methods based on multiple reect standards, such as the osetshort calibration (see [33, 38, 79, 132]), are often preferred at smaller connector sizes [133].
Coaxial air lines are manufactured in either insertable or noninsertable conguration.
In the insertable conguration, line ends have connectors with opposite sexes, while in the
noninsertable conguration, the line connectors are of the same sex, typically male. The
noninsertable conguration is preferred as far as the calibration accuracy is concerned since
the same reect standard can be used on both VNA ports; thus the error due to its asymmetry is reduced (see Subsection 5.3.4). However, two-port DUTs are typically insertable,
therefore an additional adapter needs to be rst characterized, in order to measure such
DUTs. The insertable conguration eliminates this problem, however, at the cost of two
reect standards with dierent connector sexes which inevitably have slightly dierent reection coecients. In this work, we focus on the coaxial multi-line TRL calibration in
the insertable calibration.

5.2.3

Computational aspects

The original implementation of the multi-line TRL method [36] is based on an approximate error analysis of the classical TRL method. This analysis yields the uncertainties
for individual TRL calibrations that can be constructed out of a given set of transmission
lines. Based on these uncertainties, at each frequency, a set of optimal TRL calibrations
is constructed whose results are then statistically averaged.
The uncertainty analysis of the classical TRL method by [36] is based on the linear
error propagation, hence the statistical averaging is performed with the use of linear leastsquares techniques. Consequently, the resulting algorithm is very robust as it does not
require any starting point. However, it does not provide a true least-squares solution to
the calibration problem due to some simplifying assumptions made in the error analysis.
Also, the resulting algorithm is quite complicated. Therefore, some other implementations
94

5.3. ERRORS IN THE COAXIAL MULTI-LINE VNA CALIBRATION


of the multiline TRL method have been proposed [56, 59] which use a true least-squares
formulation of the calibration problem and employ nonliner optimization techniques to
solve it. These approaches are conceptually much simpler than that of [36], however, they
are more time consuming and require a good starting point.

5.3

Errors in the coaxial multi-line VNA calibration

In this section, we describe the sources of errors in the implementation of the multi-line
TRL calibration in the coaxial environment. We discuss the physical origins of these errors
and develop models that characterize their impact on calibration-standard S-parameter
denitions. The discussion in this section serves as background for the development of
models for the coaxial calibration standards (see Subsection 5.4) that employ the concept
of the physical error mechanism (see Section 4.2).

5.3.1

Variation of connector-interface electrical parameters

One assumption commonly made in



 

the VNA measurements, and in partic  
 
  
ular in the multi-line TRL calibration,
is that the connector interface param

eters, that is, electrical parameters of
the discontinuity at the VNA test port,
are identical for all calibration standards and DUTs. This is illustrated
in Fig. 5.1. Connector discontinuity D Fig. 5.1: Connector discontinuity and the electriphysically occupies some space around cal reference plane.
the connector mating plane. Denitions of the calibration standards, however, do not account for this discontinuity and
are established with respect to the electrical reference plane. Consequently, as long as the
discontinuity D is the same for all measured device, it becomes part of the measurement
system and the electrical reference plane becomes the calibration reference plane.
In practice, however, this assumption is often not met. As a result, some common part
of the connector interface discontinuity gets lumped in to the VNA calibration coecients,
however, the dierences with respect to that common part contribute to the measurement
error. The main reason for the variation of the connector interface electrical parameters,
95

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


apart from imperfect mechanical repeatability of the connection, are the width variation
of the center conductor gap and variation of the socket ngers bending due to dimensional
tolerances on the connector pin and thickness of socket ngers, and variation of the center
conductor eccentricity.


   

  



 

Fig. 5.2: Simplied schematic of the coaxial


connector interface (not in scale).
guration of its ngers vary when connecting
on the thickness of the socket ngers lead to
the interface.

The origin of these two errors can be


better understood based on the Fig. 5.2.
This gure shows a simplied longitudinal cross section of the slotted pin-socket
(sexed) connector interface [134]. The connection between the inner conductors in
this interface occurs through the ngers of
the socket (usually four or six ngers) that
contact the pin. These ngers are exible
so as to reliably contact pins that have different diameters or are slightly eccentricity.
Consequently, for the same socket, the condierent pins. Also, dimensional tolerances
the variation of the electrical parameter of

As to the center conductor gap, its presence is necessary so as to allow the connection
between the pin and the socket. Consequently, both female and male center conductors are
set back by a small distance of around few tens of micrometers. This distance is referred to
as the pin depth. Although the pin depth limits can be well controlled in the manufacturing
process, its particular value for a given device cannot be predicted. The pin depth may
also vary due to multiple reconnections or mishandling of the connectors2 .
In the case of the air-dielectric coaxial transmission lines the situation is even more
complicated. The center conductor of such lines is unsupported, thus the center conductor
gap is dierent each time the line is connected. This variation can be minimized through
a proper operator technique, however, it cannot be completely eliminated.

A. Center conductor gap. The gap forms a radial in-cut in the center conductor whose
width g is much smaller than the depth, determined by the inner conductor and pin
2

Calibration standard manufactures typically provide the users with special gages allowing for a periodic
control of the pin-depth and thus eliminating devices with protruding center conductors.

96

5.3. ERRORS IN THE COAXIAL MULTI-LINE VNA CALIBRATION


diameters, d and dp , respectively (see Fig. D.1). Dimensions of this in-cut for typical
connector sizes can be determined based on the values given in Tab. 5.1.
Electrical properties of the
Connector Type-N 3.5 mm 2.4 mm 1.85 mm
center conductor gap can be
d [m]
3040
1520
1042
804
modeled with a short secdp [m]
1651
927
511
511
tion of a radial waveguide tergmax [m]
13
13
13
13
minated with an ideal short
circuit [92, 136].
In Ap- Tab. 5.1: Inner conductor and pin diameter, d and dp , rependix D, we show that the spectively, and maximal width of the gap gmax for typical
input impedance of such a coaxial-connector types [135].
section can be approximated
as
!
"
!
"
*
0
d N w
0 1
d N w
Z = j g ln
ln
+ (1 + j)
.
(5.11)
2
dp N w
2
dp N w
where N is the number of ngers, w is the width of the space between the ngers, and
is the metal conductivity. Equation (5.11) shows that the gap impedance is a sum of
an inductive component, depending on gap width g and a surface-impedance component
which is independent of g. This surface-impedance component results from the skin-depth
eect in the gap walls and the connector pin. Since the ratio of the gap width to the
2g
 1, we neglect the skin-depth eect in the connector
gap depth is small, that is, dd
p
pin. Reference [136] gives, without derivation, a similar expression for the gap inductance
which, however, does not account for losses. Also the formula given in [92] for the 7 mm
connector can be derived from (5.11) by substituting proper dimensions for d and dp (see
[135]), N = 0, and neglecting the conductor loss.
Scattering parameters of the series impedance representing the center conductor gap
can be written as [1]

1 Z 2
,
(5.12)
Sgap = 
Z + 2 2 Z
where
Z =

Z
,
Z0

(5.13)

is the normalized impedance, and Z0 is the characteristic impedance of the coaxial transmission line given by (C.7). Typically, the correction due to conductor losses is very small,
97

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


that is,

vRc
2Z00

 1, hence we can approximate (5.13) as




Z =

vRc
Z00 1 + (1 j) 2Z
00

Z

= jL + (1 + j)R
,
Z00
0

(5.14)

where the normalized gap inductance L and surface resistance R are functions of the gap
geometry and are given by
!

"

d N w
0
ln
,
L =g
2Z00
dp N w


and

1
R =
Z00


0 0 1
d N w
ln
2
dp N w

(5.15)
"

(5.16)

where 0 is a xed reference frequency, typically 0 = 2109 rad/s. Consequently, to rst


order, the frequency dependence of the normalized gap impedance does not depend on
the frequency variation of the characteristic impedance Z0 . Taking further into account
that the gap impedance is small, that is |Z  |  1, we can eventually approximate its
S-parameters as

1 
1 
Z
1

Z
2
.
Sgap 2 1 
(5.17)
1 
1 2Z
Z
2
In order to quantitatively illustrate the eect of the gap, we calculated the magnitude
of S11 in (5.17) based on model (5.11). We used the the dimensions corresponding to the
1.85 mm coaxial standard taken from Tab. 5.1 and assumed the conductivity of copper
( = 5.7 107 S/m). For this connector we have four ngers, that is, N = 4, and from
observations under the microscope we estimated w 100 m. Inserting those values into
(5.15) and (5.16), we obtain the normalized inductance per micrometer of gaps width
L1.85 mm = 2.28 1015 [1/m],

(5.18)

and the normalized surface resistance at 1 GHz



4
R1.85
mm @ 1 GHz = 1 10 .

(5.19)

Magnitude of the gaps reection coecient for the frequency range 0 67 GHz is shown in
Fig. 5.3. We see that this reection coecient increases linearly with frequency. The maximum value we expect based on the model (5.11) for the gap width within the specications
98

5.3. ERRORS IN THE COAXIAL MULTI-LINE VNA CALIBRATION


given in Tab. 5.1 is around 0.006 which corresponds to 44 dB. The surface-impedance
component of this reection coecient is of order 4 104 at 67 GHz, hence it can be
neglected compared to the inductive component described by (5.15).
Measurements and electromagnetic sim
ulations, in general, conrm that the pri
mary electrical properties of the gap can

be modeled as a series inductance [124,
126, 135]. In measurement results reported

in [135], reection coecients larger by

roughly up to 30 % than the prediction of

       
the model (5.11) are obtained. However,



no uncertainties for those measurements are
reported and some obvious systematic er- Fig. 5.3: Magnitude of the gap reection coefrors (ripples) can be observed in those mea- cient for the 1.85 mm coaxial connector stansurements. Reference [124] simulates a sim- dard for some typical gap widths.
plied model of the gap which does not include the ngers. Reection coecients of the gap obtained in [124] for 3.5 mm, 2.4 mm,
and 1.0 mm coaxial connectors agree well with the model (5.11) when setting N = 0.
Reference [126] simulates a more com
plex model of the gap which accounts for



the socket ngers and other geometrical

 

details of the gap structure. In Fig. 5.4
we show a comparison of results reported

in [126] with predictions of the analytical
model (5.11) at the frequency 67 GHz, for





 
some typical gap widths. We see that reection coecients reported in [126] are in
general much larger than those predicted by Fig. 5.4: Magnitude of the gap reection coefmodel (5.11). This is due to the fact that cient at 67 GHz for some typical gap widths,
analysis of [126] accounts also for the re- based on the model (5.11) and electromagection caused by the socket ngers. This netic simulations [126].
reection is independent of the gap width,
and from Fig. 5.4 it can be estimated to be around 0.01 at 67 GHz. If we add this reection to the prediction of (5.11), both sets of results agree very well. This conrms that
99

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


the analytical model (5.11) is capable of estimating the reection coecient caused by the
gap itself.
Reference [126] reports also that for extremely small gaps, a resonance behavior can
be observed. We believe, however, that such a situation can be avoided through a proper
operator technique.
B. Connector socket ngers. The errors due to nonreproducibility of the connectorsocket nger bending have not received much attention in the literature. Reference [135]
compares results of various electromagnetic analyzes with the prediction of analytical models. From these results, one concludes that these errors may be comparable with the errors
due to nonreproducibility of the center conductor gap. However, no further details are
given in Reference [135] as to the description of the socket nger used in the electromagnetic analyzes nor the assumptions made in the analytical models. In this work, we decided
to neglect this source of errors.

5.3.2

Variation of lines characteristic impedance and propagation constant

In the multiline TRL calibration method, we assume the all of the lines have the same
characteristic impedance Z0 and propagation constant . Nominal S-parameters of a single
line are therefore given by

l
0
e
,
S = l
(5.20)
e
0
where is the complex propagation constant and l is the line length. Scattering parameters
in (5.20) are normalized with reference to the characteristic impedance Z0 of the line, thus
the reection coecients are equal to zero.
In practice, however, we observe a slightly dierent value of Z0 and for each line. This
variation results from manufacturing tolerances of line diameters and metal conductivity,
and changes in the eccentricity due to dierent alignment of each lines center conductor
with respect to the test-ports. Another important factor is the nonuniformity of line diameters, and variation of the eccentricity and conductivity along the line. This nonuniformity
is primarily attributed to the mechanical vibrations of the tools used to manufacture the
inner and outer conductor.
In the following, we analyze deviations in S-parameters of a line caused by the above
factors. We rst consider the case of a line with uniform but perturbed diameters and
100

5.3. ERRORS IN THE COAXIAL MULTI-LINE VNA CALIBRATION


conductivity. We then proceed with the analysis of a slightly nonuniform line and show
that such a line can be approximately treated as a uniform line with some equivalent
perturbations of the propagation constant and characteristic impedance.

A. Uniform line. Characteristic impedance and propagation constant are linked with
the diameters, eccentricity and metal conductivity through the relationship (C.7) given in
Appendix C. Since the dependence on the eccentricity e is quadratic, small changes in this
parameter can be neglected. As for the other parameters, we can readily determine the
relationship between changes in the inner and outer diameter, d and D, respectively,
and the metal conductivity and the resulting changes Z0 and in the characteristic
impedance and propagation constant, respectively, as
Z0
Z00
Rc
Z00
Z0

+ (1 j)

,
Z0
Z00
Z00
0 Rc
Z00
and

where

and

"

Rc Z00
(1 + j)

,
Rc
Z00
Z00
1
= D
Z00
ln d

(5.21)

(5.22)

"

D d

,
D
d

d D
D d
1
Rc

,
=
Rc
2
D+d D
D+d d

(5.23)

(5.24)

and = Re is given by (C.9), Z00 is the lossless characteristic impedance of the line given
by (C.1), and Rc is the resistance per unit length of the transmission line given by (C.10).
In (5.21) we neglected the contribution of errors in Rc to the change of characteristic
impedance, since for low loss coaxial airlines 0 1.
Equation (5.21) indicates that the relative changes in the characteristic impedance
are to rst order frequency independent. On the other hand, from (5.22), we see that
changes of the propagation constant follow the frequency dependence of the real part of
the propagation constant.
The impact of errors in the lines characteristic impedance on its S-parameters can
be determined with the use of the relationship (2.19). Assuming that the characteristic
impedance changes from Z0 to Z0 = Z0 + Z0 , scattering parameters of the two port
network described by (2.19), which correspond to a circuit description of a transformer,
101

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


can be written as [45]

Q
1

,
Str = 1
Q
1 2

where

Z0 Z0
,
=
Z0 + Z0

and
Q=

(5.25)

(5.26)



1 jImZ0 /ReZ0

.

(5.27)

1 jImZ0 /ReZ0

is a factor accounting for the complex-valued character of the characteristic impedance.


For the coaxial transmission line with small conductor losses, that is, when 0  1, we can
write the rst-order approximations

and

1
Q1j
2 0

1 Z00
,
2 Z00
Rc Z00

Rc
Z00

(5.28)
"

1.

(5.29)

Assuming further that the changes of the characteristic impedance are small, we can nally
approximate (5.25) as

1 Z00
1
.
(5.30)
Str 2 Z00
00
1
12 Z
Z00
By adding the network (5.30) on either end of the line described by (5.20) and accounting for the error in propagation constant (5.22), we obtain a rst-order approximation of
lines S-parameters

1 Z00 
1 e2L ,
S11 = S22
2 Z00

!
" 
R
Z
00
c
l
1 (1 + j)

l .
S21 = S12 e
Rc
Z00

(5.31)
(5.32)

c
00
In Tab. 5.2 we summarize the worst case values of Z
and R
due to typical variations
Z00
Rc
of diameters obtained from [135]. Based on that, we estimated that for the 1.85 mm
standard the maximum error in magnitude of the reection coecient due to dimensional
variation is around 0.0023, that is 53 dB. As to the transmission coecient, we estimate
that for a line with conductivity of copper ( = 5.7 107 S/m), at 67 GHz, an error of

102

5.3. ERRORS IN THE COAXIAL MULTI-LINE VNA CALIBRATION


1 % in Rc or Z00 leads to 0.4 % of error in which corresponds to a relative error in
the transmission coecients of 0.006 % per 1 cm of line length. Consequently, we see
that dimensional variations aect mainly the reection coecients, while the transmission
coecients are much less sensitive to the tolerances on lines diameters.
Regarding the error in
metal conductivity, the situConnector
Type-N 3.5 mm 2.4 mm 1.85 mm
ation is more complex. In
D [m]
7000
3500
2400
1850
order to lower the resistivd [m]
3040
1520
1042
804
ity, air transmission lines are
D [m]
5
2.5
2.5
2.5
typically plated with silver
d [m]
2.6
2
2
2


or gold. The thickness of
Z00
[%]
0.19
0.24
0.35
0.46
the plating layers and their  Z00 M AX
Rc
[%]
0.08
0.11
0.17
0.21
Rc M AX
roughness are in general not
well controlled [136, 137], Tab. 5.2: Diameters and their tolerance for typical connechence the eective conductiv- tor types [135], and the resulting maximum relative changes
ity of the plating may undergo in the lossless characteristic impedance Z00 and resistance
much larger relative varia- per-unit-length Rc .
tions than dimensional parameters. From (5.24), we see that 1% of error in conductivity leads to 0.003 % of error
in transmission coecients per 1 cm of line length. Therefore variation of the conductivity
may be the most important contributor to the error in the propagation constant.

B. Slightly nonuniform line. Another factor causing the variation of the characteristic impedance is the nonuniformity of the diameters, conductivity and eccentricity along
the line. In Appendix E, we show that S-parameters of a slightly nonuniform coaxial
transmission line with small conductor losses can be written as

S11

l
0

S22 e2l

Z00 (x) 2x
e
dx,
Z00
l
0

Z00 (x) 2x
e dx,
Z00

and
103

(5.33)

(5.34)

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION

S21 = S12 = el [1 (1 + j) l ] ,
where

l 

l =
0

(5.35)

Rc (x) Z00 (x)

dx,
Rc
Z00

(5.36)

In the derivation made in Appendix E, we neglect the eect of conductivity variation on


the reection coecients, and the eccentricity variation as they lead to second order terms.
The frequency-independent parameter l characterizes the change in line losses due to
the nonuniformity of the diameters, conductivity and eccentricity along the line. Rewriting
(5.35) in a slightly dierent form
S21 = S12 e(1+j)(l+l ) ejl ,

(5.37)

we see that this parameter can be interpreted as an equivalent lengthening of the line
for the losses.
The change in the reection coecients, as given by (5.33) and (5.34), cannot in general
be characterized by a single frequency independent parameter. Furthermore, the frequency
dependence of (5.33) and (5.34) is determined be the actual prole of the diameters. Since,
in general, we do not know this prole, we cannot also predict these dependence.
In the context of multiline TRL calibration, a slightly nonuniform line can be seen as a
uniform line with an equivalent perturbation of the propagation constant and characteristic
impedance and some residual reections resulting from the actual nonuniformity prole.
This can be prooven in the following manner. In the calibration algorithm, we assume that
the line is uniform, and optimize the VNA calibration coecients and unknown parameters
of the line so as to minimize the mist between the line model prediction and the lines
S-parameters determined from raw measurements. In the case of a nonuniform line, this
mist is thus characterized by the integrals in (5.33) and (5.34). We can show that these
integrals become minimal when we subtract the mean value of the impedance prole, that
is,

104

5.3. ERRORS IN THE COAXIAL MULTI-LINE VNA CALIBRATION

l

Z00 (x) Z00 2x
1 Z00 
2l

1e
+

e
dx,
S11
2 Z00
Z00
Z00
0

(5.38)

l


Z
Z00 (x) Z00 2x
1
00
2l
2L
S22
1e
+ e

e dx,
2 Z00
Z00
Z00

(5.39)

Z00
1  Z00 (x)
=
dx.
Z00
l
Z00

(5.40)

where

Consequently, a nonuniform line is seen in the multiline TRL calibration as a uniform line
with the characteristic impedance error dened by (5.40), error in the propagation constant
described by (5.35), and residual reections given by the integrals in (5.38) and (5.39).

5.3.3

Line length error

The only numerical parameters passed to the multiline TRL calibration method are the
line lengths. For the coaxial airlines, these lengths are typically dened as the lengths of the
outer conductor and are measured with mechanical blocks [90]. The value stemming from
this measurement will be in general dierent than the actual length of the line due to the
measurement error and the compression of the line in the actual measurement environment
due to the torque applied to the connector interface [138].
Scattering parameters of a line with perturbed length can be readily obtained from
(5.20) as
S11 = S22 = 0,
(5.41)
and

S21 = S12 el (1 l) .

(5.42)

From (5.42), we can derive the upper bound for relative errors in the transmission coecients as


 S
21 



 S21 



 S
12 



 S12 

105

|0 |l.

(5.43)

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


For 1.85 mm precision air lines, typically the error in length determination is smaller
than 5 m. This leads to the maximum relative error in transmission coecients of 0.7 %
at 67 GHz.

5.3.4

Reect asymmetry

In the classical TRL [29] and multi-line TRL [36] calibrations, it is assumed that the
reect standard presents the same reection coecient on both VNA ports. Thus, this
standard can be seen as a symmetrical two-port network without transmission.
Dierence in the reection coecient of the reect standard presented on both VNA
ports, which we refer to as the asymmetry of the reect standard, leads to errors in both
the classical TRL [29] and the multiline TRL [36] calibrations. In the case of the coaxial
implementation of the multiline VNA calibration, sources of the reect asymmetry are
dierent in the noninsertable and insertable congurations. In the case of the noninsertable
conguration, we use the same reect standards on both VNA ports. Consequently, the
asymmetry results only from the nonrepeatability of the connector interface parameters
which is of a random nature. However, in the case of the of the insertable conguration
the situation is dierent. In this conguration, we use in fact two reect standards with
dierent connector sexes. Although these standards are designed so as to exhibit the same
reection coecient, the nonreproducibility of the connector interface (see Subsection 5.3.1)
and the manufacturing tolerances always lead to some small dierence in the reection
coecient of both devices.
In the case of a single reect standard, its asymmetry gets lumped into the VNA
calibration coecients and thus cannot be detected. This asymmetry, however, could be
detected with the use of multiple reect standards with dierent reection coecients. The
resulting calibration would use both multiple lines and multiple pairs of reect standards.
Implementation of such a calibration method, however, is beyond the scope of this work.

5.4

Calibration standard models

In this section, we present extendend models for the coaxial calibration standards used
in the multi-line VNA calibration. These models account for the error mechanisms discussed in Section 5.3. Contribution of these mechanisms is described with additional
frequency-independent parameters which can be thought of as the electrically-equivalent
106

5.4. CALIBRATION STANDARD MODELS








 

Fig. 5.5: Coaxial airline in the multi-frequency multiline TRL calibration: (a) overview,
(b) electrical model.
error mechanisms (see Section 4.2). These parameters are then identied, along with the
VNA calibration coecients, in the course of the multi-frequency VNA calibration.

5.4.1

Lines

As discussed earlier on, a set of coaxial airline exhibits characteristics that slightly dier
from the nominal description (5.20). We illustrate that in Fig. 5.5a where we show dierent
physical error mechanisms aecting a real coaxial airline. These mechanisms include (see
Section 5.3):
errors in the lengths of the inner and outer conductor, lin and lon , with respect
to the nominal lengths, li0n and lo0n , respectively;
107

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


nonreproducibility of the center conductor gap on both line ends due to the random
displacement ln of the horizontal symmetry axis of the inner conductor with respect
to the horizontal symmetry axis of the outer conductor;
nonuniformity and nonreproducibility of the diameters of the inner and outer conductor, Dn (x) and dn (x), dened with respect to the nominal diameters, d0 and
D0 , respectively;
nonuniformity and nonreproducibility of the conductor losses, characterized by the
conductivity error (x) dened with respect to the nominal conductivity 0 .
The impact of these physical error mechanisms on the line S-parameters can be equivalently
described with the circuit shown in Fig. 5.5b. This circuit consists of:
an ideal transmission line with the length li0n + li0n , propagation constant + (1 +
j)ln and the nominal characteristic impedance Z0 ; the frequency-independent
parameter ln , referred as the loss correction factor, is dened by (5.36), and characterizes the impact of diameter errors and conductivity error, and the nonuniformity
of the line on the propagation constant;
two transformers describing the change of the lines characteristic impedance with
respect to the nominal impedance Z0 ; these transformers are characterized by the
frequency-independent characteristic impedance correction Z00n , dened by (5.40),
which captures the impact of diameter errors and the nonuniformity on the lines
characteristic impedance; in the calibration algorithm we use the normalized characteristic impedance correction 0n = Z00n /Zref , where Zref = 50 .
two inductances modeling the nonreproducibility of the center conductor gap; these
inductances are determined from the model (5.15) and are function of the inner and
outer conductor length errors, lin and lon , respectively, and the center conductor
displacement ln .
Assuming that the errors are small, scattering parameters of the line standard shown in
Fig. 5.5b can to rst order be written as

SLn = SL0n

T
wT
eli0n w12
Ln
+ l 11 LT n
,
T
e i0n w21 Ln
w22
Ln

108

(5.44)

5.4. CALIBRATION STANDARD MODELS


where the electrically-equivalent error mechanisms (see Section 4.2) are given by
TLn = [ln , 0n , ln , lin , lon ] ,

(5.45)

the frequency-dependent weighting functions, resulting from models developed in in Subsections 5.3.1 through 5.3.3, are captured in vectors

w11 =

1
2

0
1 e2li0n
jLg (1 e2li0n )
j 12 Lg (1 + e2li0n )
j 12 Lg (1 + e2li0n )

w22 =

and

w21 = w12 =

1
2

0
1 e2li0n
jLg (1 e2li0n )
j 12 Lg (1 + e2li0n )
j 12 Lg (1 + e2li0n )

(5.46)

(1 + j)
0
0
j 12 Lg
j 12 Lg

(5.47)

and SL0n are the nominal S-parameters of the air-line. These parameters are determined
under the assumptions that Ln = 0, that is, for a uniform airline with the nominal
conductivity 0 and characteristic impedance Z0 , the nominal diameters, d0 and D0 , the
nominal conductor lengths, li0n and lo0n , and equal center-conductor-gap widths on both
line ends. Consequently, nominal S-parameters of a coaxial airline can be written as

SL0n =

j 12 Lg 1 + e2li0n g0


eli0n 1 jLg g0

eli0n 1 jLg g0


j 12 Lg 1 + e2li0n g0

(5.48)

where Lg is the normalized inductance per-unit-length of the center conductor gap, dened
in (5.15), and g0n = (lo0n li0n )/2 is the nominal center conductor gap.

109

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


Consequently, a set of NL real lines is characterized by the following parameters
l = [l1 , . . . , lNL ]T

(5.49)

0 = [01 , . . . , 0NL ]T ,

(5.50)

l = [l1 , . . . , lNL ]T

(5.51)
T

li = [li1 , . . . , liNL ] ,

(5.52)

lo = [lo1 , . . . , loNL ]T .

(5.53)

which are determined, along with the VNA calibration coecients, in the course of the
multi-frequency VNA calibration. However, not all of the parameters in vectors (5.49)
through (5.53) can be uniquely determined. This will be discussed in Section 5.5.

5.4.2

Reect

A nominal description of the reect standard in the classical multi-line TRL calibration
is given by

0
R
,
SR =
(5.54)
0 R
where R is the unknown reection coecient of the reect, determined during the calibration. In our approach we do not account for the possible asymmetry of the reect standard
(see Subsection 5.3.4) and use the nominal description (5.54).

5.4.3

Thru

In the noninsertable calibration, the thru standard is realized as a line section and
described by the model from Subsection 5.4.1. In the case of the insertable conguration,
considered in this work, a direct thru connection of VNA ports is typically used. We
assume that its S-parameters are

0 1
ST =
.
1 0
110

(5.55)

5.5. SOLUTION UNIQUENESS

5.5

Solution uniqueness

In this section, we discuss this issue of assuring the uniqueness of the solution to the
calibration problem (5.7). We rst review the mathematical conditions for the solution
uniqueness and show that these conditions are related to the linear independence of the
estimated parameters in the vicinity of the solution. We then analyze the relationships
between the parameters estimated in the multi-line VNA calibration and identify those
that are linearly dependent. Following on that, we propose linear constraints for these
parameters so as to provide their unique identication. These constraints are related
to some intuitive statistical properties of the error mechanisms aecting the calibration
standards.

5.5.1

Mathematical framework

Conditions for the solution uniqueness in the maximum-likelihood system identication


are reviewed in Section B.5. We show there that the solution is unique if the Jacobian
matrix of the residuals determined at the solution is full rank. In the case of the multifrequency VNA calibration problem (2.94), with residuals dened by (5.10), this Jacobian
matrix is given by3

rs1 (, rc ) rs1 (, rc )

(rc )T
T

..
..

.
J=
(5.56)
.
.

rs (, rc ) rs (, rc )

K
K
T
c
T
(r )

Now, direct investigation of the rank of the matrix (5.56) is complicated, hence we
follow here a dierent approach. We rst note the the matrix (5.56) has more rows then
columns as there are more equations than estimated parameters . Thus, if the matrix J is
rank decient, some of its columns are linear combinations of other columns. This in turn
reects the fact that there exist some relationships between the estimated parameters
and rc . Hence, when verifying the uniqueness of the solution to (5.7), instead of analyzing
the rank of (5.56), we look for relationships between the identied parameters.
The existence of such relationships indicates, that the description of the estimation
problem uses more parameters than the actual number of degrees of freedom. The most
3

2
The dependence of the likelihood function on the residual variances 12 , . . . K
has been eliminated by
use of the concentrated likelihood approach (see Subsection 5.6).

111

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


common reason for such a situation, apart from the error in the mathematical description
of the problem, is the inherent unidentiability of some parameters from a given set of
measurements. An example of such a situation is the identication of the 8-term VNA
model (see Section 2.3.1-B). Although the physical description of this model uses eight
parameters, only seven independent terms can be identied in the calibration since some of
the physical model parameters appear always as products in the equations (2.65) through
(2.68). Consequently, we need to either accept this ambiguity (which does not introduce
any error in the case of the 8-term VNA model, since both the calibration and correction
equations use only seven independent terms describing the VNA), or make some additional
assumptions as to the relationships between parameters that cannot be identied.
One way to write such assumptions is through some additional constraints imposed on
the optimization problem (5.7). In the simplest case, these constraints take on the form of
linear equality constraints, that is,
A = 0,

(5.57)

where matrix A consists of M linearly independent rows and rank (A) = M where M
is the dimension of the null space of the Jacobian matrix J. The constrained optimization problem formed by (5.7) and (5.57) can then be solved through the direct variable
elimination [139].

5.5.2

Relationships between estimated parameters

The parameters estimated in the multi-frequency multiline TRL calibration are contained in the vectors and rc . The elements of the vector rc are the estimates of the
airline parameters grouped in the vectors (5.49) through (5.53), while the vector , dened
by (2.89), contains the VNA calibration coecients along with the unknown propagation
constant of the lines and the unknown reection coecient of the reect at the frequency
fk , for k = 1, . . . , K.
The potential non-uniqueness of the solution to (5.7), given by estimates of and
rc , may result either from relationships between the frequency independent parameters
grouped in the vector rc or relationships between those parameters and the parameters
grouped in the vector . Indeed, the columns of the Jacobian matrix of the residual
function (5.10), corresponding to the parameters , are obviously linearly independent,
since the VNA model does not account for the relationships between VNA calibration
112

5.5. SOLUTION UNIQUENESS


coecients at dierent frequencies.
Now, in the classical multiline TRL calibration, the identication of the propagation
constant requires the knowledge of the line lengths. In the course of the TRL calibration
the electrical length of the line is identied which, based on the physical line length, allows
to determine the propagation constant. In the multi-line TRL calibration, the algorithm
for calculating the propagation constant is more complicated and uses t averaging of the
propagation constants estimated in individual TRL calibrations. As a consequence, the
unique identication of both line length corrections captured in the vector (5.52) and (5.53),
and the propagation constant is impossible. Indeed, we can easily show that for a given
solution to the calibration problem (5.7), another solution with the same maximum value
of the likelihood function can be derived by simply adding an arbitrary length to all of the
elements of the vectors (5.52) and (5.53). This does not aect the center conductor gap
and the error resulting from this additional length gets compensated by the adjustment
the complex propagation constant determined in the calibration. Consequently, we need
to make some assumption as to the length corrections (5.52) and (5.53) in the form of the
constraints

a
p
li li = 0.
(5.58)
alo
pl0
A similar reasoning leads to the conclusion that another assumption in the form
T
p
= 0,
al
l

(5.59)

needs to be made as to the loss correction factors in the vector (5.51). Indeed, a constant
value l added to all of the elements of the vector (5.51) gets also compensated by the
adjustment of the propagation constant.
Regarding the correction of the characteristic impedance captured in the vector (5.50),
we also note that the unique identication of those parameters and the VNA model parameters is not possible. Indeed, for a given solution to the calibration problem (5.7), adding
an arbitrary impedance to the parameters in the vector (5.50) leads to another solution
with the same value of the likelihood function. The transmission parameters of the lines in
(5.44) do not change due to lack of dependence on 0n . The resulting error in the reection
coecients of the lines can be compensated with additional transformers with opposite
impedance ratios added to the VNA error boxes. These transformers do not change the
measurement of the through standard and get absorbed in the parameters of the reect
standard (and into the calibration coecient determined based on the measurement of the
113

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


reect standard). Hence, in order to uniquely identify the elements of the vector (5.50), we
need to make an assumption as to the values of these elements. We write this assumption
in the form of the constraint
aT0 p0 = 0.
(5.60)
Finally, as for the width of the center conductor gap, we rst note that this width
is determined by the parameters contained in the vectors (5.51), (5.52) and (5.53). We
already noticed that some assumptions need to be made regarding the parameters (5.52)
and (5.53). As far as the the parameters (5.51) are concerned, consider perturbing them
by l . This is equivalent to adding the inductance Lg l to the inductances on the line
end which is attached to VNA port 1, and subtracting this same inductance from the
inductances on the line end which is attached to VNA port 2 (see Fig. 5.5b). We can
easily show that this leads to another solution of the VNA calibration problem (5.7) with
the same value of the likelihood function. Indeed, the transmission coecient of the lines
in (5.44) remain the same as they do not depend on the parameter ln . The resulting
error in the reection coecients of the lines is further compensated by adding impedances
jLg l and jLg l in series to the VNA ports. These impedances do not change the
measurement of the through standard, and get absorbed in the parameters of the reect
standard. Consequently, we see the parameters in the vector (5.49) cannot be uniquely
determined and we also need to make an assumption as to their values. We write this
assumption in the form
T
al
pl = 0.

5.5.3

(5.61)

Optimal constraint choice

The assumptions given by equations (5.58) through (5.61) can be made in various ways.
The simplest choice is to select one of the lines as a reference and assume that it is not
aected by the errors captured in the vectors (5.49) through (5.53). Through this choice,
however, one assumes that the dimensional measurements for the reference line are perfect.
This is obviously not true, and introduces an additional error.
The optimal choice for the constraints (5.58) through (5.61) can be deduced based on the
statistical properties of the errors captured in the vectors (5.49) through (5.53). We assume
that the dimensional measurements and manufacturing tolerances are not corrupted with a
systematic error and the variance of each error type is the same for all lines. Furthermore,
it is plausible to assume that the VNA ports are equally likely to pull the center conductor
114

5.6. NUMERICAL SOLUTION


towards itself, therefore the expectation value of the displacement l should be zero.
Hence, we can write assumptions that express these observations as
T
= [1, . . . , 1] ,
al
i

(5.62)

T
al
= [1, . . . , 1] ,
0

(5.63)

T
= [1, . . . , 1] ,
al

(5.64)

aT0 = [1, . . . , 1] ,

(5.65)

T
al
= [1, . . . , 1] .

(5.66)

and

These assumptions are met with a much smaller error than when taking one of the lines as
a reference. Indeed, we can show that the standard deviation of this error will be smaller
by 1N where N is the number of lines.

5.6

Numerical solution

In this section, we review the key points of the algorithm for solving the optimization
problem (5.7). This problem has a large number of optimized variables, hence its direct
solution is infeasible. Therefore we apply dierent techniques to reduce the dimensionality
of (5.7).
2
We rst separate the determination of the residual variances 12 , . . . , K
and the parameters and rc by use of the stagewise minimization approach, referred to as the concentrated likelihood approach (see Section B.2.1). To this end, we determine the values
2
12 (, rc ), . . . K
(, rc ) of the residual variances which maximize the likelihood function
(5.8) for a given value of the parameters and rc . This values are then inserted into (5.8)
to yield the concentrated likelihood function which is further maximized with the use of
numerical techniques.
2
In order to determine 12 (, rc ), . . . K
(, rc ), we take the derivative of (5.8) with respect
to k2 to obtain
2
)
ln L(, rc , 12 , . . . , K
NW 1
1 1 s
1 s
=
+
r (, rc )T Vs
r (, rc ),
2
2
k k
k
2 k 2 (k2 )2 k

115

(5.67)

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


which after setting to zero yields
k2 (, rc ) =

1 s
1 s
rk (, rc )T Vs
rk (, rc ).
k
NW

(5.68)

Inserting these solutions into the original likelihood function (5.8) and dropping the constant terms, gives the concentrated likelihood function
K
NW 
1 s
ln rsk (, rc )T Vs
rk (, rc ).
ln L(, r ) =
k
2 k=1
c

(5.69)

Consequently, we can rewrite (5.7) as a two step optimization process. In the rst step,
we obtain the estimates and rc from
arg max ln L(, rc ),
, rc

(5.70)

and then calculate the maximum-likelihood estimates of residual variances as


2
2 c

M
r ).
LE,k = k (,

(5.71)

2
We can further show, that the estimates
M
LE,k are biased. The unbiased estimates of the
residual variances can be written as

k2 =

NW

2
,
N W R/K M LE,k

(5.72)

where R is the length of the vector rc and K is the number of frequencies.

Now, the maximization problem (5.70) can be equivalently written as


arg min
, rc

K

k=1

1 s
ln rsk (, rc )T Vs
r (, rc ).
k k

(5.73)

In order to solve this problem, we use an iterative numerical approach based on the
Levenberg-Marquardt algorithm (see SubsectionB.4.1-B). We can show that after expanding the residuals into the Taylor series and neglecting the nonlinear terms, the equations
for the corrections to the estimates (q) and rc(q) (without the Levenberg-Marquardt reg116

5.6. NUMERICAL SOLUTION


ularization) are given by
X(q)T V 1 X(q) (q) = X(q)T V 1 r(q) ,
where

(q)

X(q) =

1
D

and
r(q)

and

(q)
=
,
rc(q)

rs1 (, rc ) 

T =(q) ,rc =rc(q)
..
.

rsK (, rc ) 

T


(5.74)

= (q) ,rc =rc(q)

(5.75)


rs1 (, rc ) 

(rc )T =(q) ,rc =rc(q)
..
.


rsK (, rc ) 

(rc )T 

(5.76)

= (q) ,rc =rc(q)

rs ( (q) , rc(q) )
1

..

,
= D1
.

rsK ( (q) , rc(q) )

(5.77)

Vs1

...

V=

(5.78)

VsK
and

D=

1 s
rs1 ( (q) , rc(q) )T Vs
r ( (q) , rc(q) )
1 1

...
1
rsK ( (q) , rc(q) )T Vs
rs ( (q) , rc(q) )
K K

IN W ,

(5.79)
where the underline denotes the convention (A.2), IN W is the identity matrix of size N W
N W , and denotes the Kronecker product [57]. In order to solve the normal equations
(5.74), we further exploit the sparsity of the matrix X(q) to reduce the dimensionality of
the problem (5.74). We use here a similar approach to [140].

117

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION

5.7

Residual analysis

In order to assess uncertainties of the the estimates and rc , we perform the residual
analysis (see Subsection 3.5.3). We can show that the unbiased estimate of the covariance
matrix of errors in these estimates is given by (see Section B.3)
=


1
1
XT V 1 X
,
N W R/K

(5.80)

where the matrix X is given by (5.76) evaluated at and rc . When evaluating (5.80), we
are often interested in only covariance matrices of rc or parts of the vector corresponding
to one frequency. This submatrices of can be easily determined by accounting for the
sparsity of the matrix X and applying the formulas for the inversion of block matrices (see
[121]).

5.8

Experimental results

In order to verify our multi-frequency multiline TRL algorithm, we compare it with


the classical multiline TRL method [36]. To this end, for a given set of check devices and
multiline-TRL calibration standards, we measure their S-parameters as raw quantities
(i.e., without applying any error correction in the VNA). Subsequently, we process the
measurement data o-line with the two calibration algorithms and compare the results.
In this section, we present results of this verication for the multiline TRL calibration in
the Type-N and 1.85 mm coaxial connector standard. Other results, obtained for coaxial
transmissions lines with the LPC-7 connector standard, can be found in [141].

5.8.1

Overview

We implemented our multi-frequency multiline TRL algorithm in the MATLAB environment [142]. We used the 8-term VNA model (see Paragraph 2.3.1-B) with the base
parametrization (2.71). We determined the switch terms (see Subsection 2.3.2) from the
measurement of the through connection and removed them from the raw measurements by
use of the relationship (2.78). In the absence of any detailed information as to the nature
of the VNA instrumentation errors, we assumed that these errors have a circular-normal
118

5.8. EXPERIMENTAL RESULTS

Fig. 5.6: Measurement setup.


PDF (see Paragraph B.6) and are uncorrelated. Thus, the matrix Vsk , as introduced in
(5.2), is the identity matrix.
The classical multiline TRL calibration was performed with the use of the software
package MultiCAL [143, 144]. As a result of the calibration, this package yields a set of
VNA calibration coecients which we then use in the MATLAB environment to correct
the DUT measurements. The MultiCAL package uses the 8-term VNA model with the
reciprocal parametrization (2.72). Thus, the VNA calibration coecients obtained from
the MultiCAL package were converted to the base parametrization (2.71). The switch
term correction in the MultiCAL package is performed with the methods described in
Subsection 2.3.2.

5.8.2

Type-N coaxial connector

The measurement setup is schematically depicted in Fig. 5.6. Measurements were


performed with the Agilent E8363 VNA in the frequency range 0.1 18 GHz (359 points).
In order to reduce the impact of VNA receiver errors, the IF bandwidth was reduced down
to 50 Hz. In the setup two exible 3.5 mm cables with additional 3.5 mm-Type-N adapters
were used.
As calibration standards we used six transmissions lines, a pair of reect standards, and
a direct through connection of the VNA ports. The lengths of the airline outer conductors
specied along with the measurement error, as specied by the manufacturer [145], are
given in Tab. 5.3. The lengths of the inner conductors were not specied by the manufacturer, thus we assumed them to be unknown and modied the condition (5.58) to account
119

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


only for the outer-conductor length corrections.
As the reect standard we used a pair of female and male oset shorts from the SOLT
calibration kit [146]. According to the manufacturer, the two oset shorts were designed
so as to have the same electrical length, corresponding to the phase shift of a 9.9 mm long
coaxial airline.
In Tab. 5.3 we put the corrected lengths of the inner and outer conductor of the airlines
determined in our multi-frequency multiline TRL calibration. The corrected lengths of the
outer conductor dier from the nominal ones by up to 33 m, and thus are well within the
100 m measurement uncertainty bounds specied by the manufacturer.
In Tab. 5.3 we also give the standard deviation of the corrected lengths determined
in the residual analysis (see Section 5.7). This standard deviation can be thought of as
the precision with which the corrected line lengths are determined4 , and for all of the
lines is 0.2 m, which indicates that our calibration method is capable of a very precise
determination of those lengths.
In Tab. 5.4 we put together the center conductor displacements l, corrections to the
characteristic impedance Z0 = 0 Zref , and the loss correction factors l determined
in our multi-frequency multiline TRL calibration, along with the standard uncertainties
determined in the residual analysis. We see that these uncertainties are typically on the
order of several percent which shows that our calibration algorithm can very precisely
determine those corrections. The center conductor displacements are less then 35 m.
The characteristic impedance corrections are less then 0.1 % and are smaller then the
worst-case value of 0.19 %, given in Tab. 5.2 and derived from the specications [135].
Finally, the relative loss corrections (i.e., l /li0 ), except for the line L2, are less than
10 %. For the line L2 this correction is on the order of 20 %.
In Fig. 5.7 we compare the real and imaginary part of the propagation constant determined with the use of the classical multiline TRL calibration (solid gray), and with the use
of our multi-frequency multiline TRL calibration (solid black). We see that our method,
by using more detailed models for the coaxial airlines, leads to a slightly dierent estimate
of both the real and imaginary part of the propagation constant, as compared with the
classical multiline TRL calibration. The relative dierence in the phase constant is larger
than in the attenuation constant, which indicates that the errors of a reactive character
(i.e., nonrepeatability of the connector interface, nonuniformity of the diameters) have a
4

In order to determine the accuracy of these corrections, we would have to perform full uncertainty
analysis, accounting for the error of the constraints discussed in Subsection 5.5.3. This will not be discussed
here.

120

5.8. EXPERIMENTAL RESULTS

Line
L1
L2
L3
L4
L5
L6

Outer conductor
length
Measured Corrected
[mm]
[mm]
29.97
30.001
49.94
49.958
59.93
59.934
74.93
74.913
99.88
99.876
149.83
149.797

Inner conductor
length
Corrected
[mm]
29.978
49.933
59.913
74.904
99.856
149.772

Measurement
uncertainty
[mm]
0.1
0.1
0.1
0.1
0.1
0.1

Residual
standard
uncertainty
[m]
0.2
0.2
0.2
0.2
0.2
0.2

Tab. 5.3: Outer conductor lengths with measurement uncertainties, as specied by the
manufacturer [145], and outer and inner conductor lengths obtained in our multi-frequency
multiline TRL calibration, along with standard uncertainty from the residual analysis.

Line

Value

L1
L2
L3
L4
L5
L6

[m]
8.7
12.7
35.1
6.1
5.8
1.8

l
Standard
uncertainty
[m]
0.1
0.1
0.1
0.1
0.1
0.1

Z0
Standard
Value
uncertainty
[m]
[m]
6
1
51
1
45
1
7
1
20
1
27
1

l
Standard
Value
uncertainty
[mm]
[mm]
2.3
0.2
9.4
0.2
3.0
0.2
0.9
0.2
1.8
0.2
3.2
0.2

Tab. 5.4: Estimates of the center conductor displacement l, characteristic impedance


correction Z0 , and loss correction factor l obtained in our multi-frequency multiline
TRL calibration, along with standard uncertainties from the residual analysis.

dominant inuence on the determination of the propagation constant.


In order to assess the accuracy of our multi-frequency multiline TRL calibration, we
compared the residual standard deviation obtained in our approach (see Section 5.1) and
in the classical multiline TRL method (see Paragraph 2.4.1-A). This comparison is shown
in Fig. 5.8 as a function of frequency. We see that the residual standard deviation yielded
by our methods is at least four times smaller than for the classical multiline TRL method.
This indicates that our description of the calibration standards (i.e., denitions of the
calibration standards along with the VNA model) ts better the actual measurements of
the standards than the classical multiline TRL method.
121

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION







 !"#











  !"

         


(a)

         


(b)

Fig. 5.7: Propagation constant for the multiline TRL calibration in the Type-N coaxial
connector standard: (a) imaginary part normalized to the free-space propagation constant
0 , (b) real part; as obtained from the classical multiline TRL (solid gray) [36], and from
our multi-frequency multiline TRL calibration (solid black).



We further performed the residual analysis and calculated the residual standard


  !"
deviations of the VNA calibration coe
cients (see Section 5.7). In Figure 5.9 we
show the residual standard deviation for the

eective source match, eective directiv
ity and eective tracking, respectively (see
         


Paragraph 3.4.2-B), as obtained from the
classical multiline TRL calibration (solid
Fig. 5.8: Residual standard deviation for the
gray) and from our multi-frequency multimutiline TRL calibration in the TypeN coaxline TRL calibration (solid black). We obial connector standard, as obtained from the
serve that our method consistently yields
classical multiline TRL (solid gray) [36], and
much smaller uncertainties of VNA calibrafrom our multi-frequency multiline TRL caltion coecients than the classical multiline
ibration (solid black).
TRL method. Our uncertainties are typically four times smaller than in the classical
multiline TRL, and at lower frequencies (i.e., below 2 GHz) the uncertainties we provide
are even of an order of magnitude smaller than in the classical multiline TRL method.
In order to verify that the VNA calibrated with our approach yields correct measurements, we measured some check devices: a pair of oset-open terminations, a pair of
matched terminations, and a 10 dB attenuator. In Fig. 5.10 and Fig. 5.11, we show mag122

5.8. EXPERIMENTAL RESULTS

















  !
         


(a)









  !
         


(c)




  !
         















         


(b)









  !


  !
         


(e)



(d)


  !
         


(f)

Fig. 5.9: Residual standard deviation of the eective directivity of (a) the VNA port 1 and
(b) the VNA port 2, eective source match of (c) the VNA port 1 and (d) the VNA port
2, and eective tracking of (a) the VNA port 1 and (b) the VNA port 2, as obtained from
the classical multiline TRL (solid gray) [36], and from our multi-frequency multiline TRL
calibration (solid black).

123

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION






 !"#












         






(a)

         






  !







 !"#



(b)


  !




         


(c)

         


(d)

Fig. 5.10: Reection coecient magnitude for a coaxial Type-N (a) female and (b) male
oset-open termination, along with the residual standard uncertainty (c,d), as obtained
from the classical multiline TRL (solid gray) [36], and from our multi-frequency multiline
TRL calibration (solid black).

nitude and phase (with the linear part removed), respectively, of the corrected reection
coecient of the male and female oset-open termination, as determined with the classical multiline TRL calibration (solid gray), and with the use of our method (solid black),
along with the residual standard uncertainties. We see that the agreement between both
methods is very good, and our method yields residual standard uncertainties that are at
least four time smaller than for the classical multiline TRL method.
We also note the for both methods the reection coecient magnitude at some frequencies exceeds one. This non-physical result can be attributed to the asymmetry of the reect
standard which may have a signicant impact on measurements of highly reective devices
[109]. This asymmetry cannot be accounted for in the approximate uncertainty assessment
provided by the residual analysis, since the errors in the reect standard get completely
124

5.8. EXPERIMENTAL RESULTS


















  !
         







(a)

         






  !"








  !

(b)


  !"




         


(c)

         


(d)

Fig. 5.11: Reection coecient phase (with linear part removed) for a coaxial Type-N (a)
female and (b) male oset-open termination, along with the residual standard uncertainty
(c,d), as obtained from the classical multiline TRL (solid gray) [36], and from our multifrequency multiline TRL calibration (solid black).

absorbed into the VNA calibration coecients, and thus do not manifest themselves as the
mist between the calibration standard measurements and their model.
In Fig. 5.12, we show the reection coecient magnitude of a male and female matchedtermination, along with the in-phase residual standard uncertainties (see Section 3.2), as
determined with the classical multiline TRL calibration (solid gray), and with the use of
our method (solid black). We again see a very good agreement between the two methods
with much smaller uncertainties (at least four times) provided by our method.
In Fig. 5.13 we show the magnitude and phase (with the linear part removed) of the
transmission through the 10 dB attenuator, as determined with the classical multiline TRL
calibration (solid gray), and with the use of our method (solid black), along with residual standard uncertainties. While the agreement between the two calibration methods is
125

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION





 !"#$






 !"#$





         






(a)

(b)




  !"






         




  !"




         



         



(c)

(d)

Fig. 5.12: Reection coecient for a coaxial Type-N (a) female and (b) male matched
termination, along with the in-phase residual standard uncertainty (c,d), as obtained from
the classical multiline TRL (solid gray) [36], and from our multi-frequency multiline TRL
calibration (solid black).
excellent, our method again yields much smaller uncertainties.

5.8.3

1.85 mm coaxial connector

The measurement setup is schematically depicted in Fig. 5.14. Measurements were


performed with the Agilent E83631 VNA in the frequency range 0.2 67 GHz (335 points).
In order to reduce the impact of VNA receiver errors, the IF bandwidth was lowered down
to 100 Hz, and an average of four sweeps was taken. We used only one cable, connected
to the VNA port 2, in order to minimize errors due to the cable instability. Both the
VNA port 1 and the cable end on the DUT side were equipped with additional adapters
in order to minimize the wear on the VNA-port and cable connectors as those elements
cannot easily be replaced. These adapters were carefully selected to have a small pin depth
126

5.8. EXPERIMENTAL RESULTS







 !"#$








  !





         







(a)




  !



(b)


  !"






         



         


(c)

         


(d)

Fig. 5.13: Transmission for a coaxial TypeN 10 dB attenuator: (a) magnitude, (b) phase
(with the linear part removed), along with the (c) magnitude and (d) phase residual standard uncertainty, as obtained from the classical multiline TRL (solid gray) [36], and from
our multi-frequency multiline TRL calibration (solid black).

( 10 m), in order to reduce the range within which the center conductor gap may vary.
The calibration kit consisted of ve transmissions lines, a pair of reect standards, and
a direct through connection of the VNA ports. The lengths of the air-line outer and inner
conductors were measured with a mechanical block and are given, along with the standard
uncertainty of the measurement, in Tab. 5.5. As the reect standard we used a pair of
female and male oset shorts. According to the manufacturer, the two oset shorts were
designed so as to have the same electrical length, corresponding to the phase shift of a
5.4 mm long coaxial airline.
In Tab. 5.5 we compare the nominal lengths of the inner and outer conductor of the
airlines with the corrected lengths obtained from our multi-frequency multiline TRL calibration. The corrected lengths dier from the nominal lengths by up to 13 m which is
127

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION

Fig. 5.14: Measurement setup.


within the 99 % condence interval (15 m), determined from the dimensional measurement uncertainty.
In Tab. 5.5 we also show the standard deviation of the corrected lengths determined in
the residual analysis. This standard deviation quanties the precision with which we can
determine the corrected line lengths, and for all of the lines is 0.2 m, which indicates that
we can very precisely determine those lengths.
In Tab. 5.6 we put together the center conductor displacements l, corrections to the
characteristic impedance Z0 = 0 Zref , and the loss correction factors l determined
in our multi-frequency multiline TRL calibration, along with the standard uncertainties
from the residual analysis. We see that the precision with which we determine most of the
corrections is very high, typically of on the order of a few percent. The center conductor
displacements are very small and less than 4 m. The characteristic impedance corrections
are less then 0.3 % and are smaller then the value of 0.46 %, given in Tab. 5.2 and derived
from the specications [135]. Finally, the relative loss corrections are less than 15 %.
Comparing to the results for the Type-N connector given in Tab. 5.4, we note the the
characteristic impedance and loss corrections for the 1.85 mm coaxial airlines are in general
larger. This may be attributed to the larger impact of manufacturing tolerances for smaller
connector sizes.
In Fig. 5.15 we compare the real and imaginary part of the propagation constant determined in the classical multiline TRL calibration (solid gray), and with the use of our
multi-frequency multiline TRL method (solid black). Similarly to results for the Type-N
connectors, we see that our calibration algorithm yields slightly dierent estimate of both
the real and imaginary part of the propagation constant, as compared with the classical
128

5.8. EXPERIMENTAL RESULTS

Line
L1
L2
L3
L4
L5

Outer conductor
length
Measured Corrected
[mm]
[mm]
14.999
15.012
16.337
16.344
18.573
18.580
23.066
23.076
29.981
29.982

Inner conductor
length
Measured Corrected
[mm]
[mm]
14.991
14.990
16.331
16.326
18.572
18.567
23.060
23.051
29.977
29.967

Measurement
standard
uncertainty
[m]
5
5
5
5
5

Residual
standard
uncertainty
[m]
0.2
0.2
0.2
0.2
0.2

Tab. 5.5: Outer and inner conductor lengths, as measured with mechanical blocks and
determined in our multi-frequency multiline TRL calibration, along with the measurement
standard uncertainty and standard uncertainty from the residual analysis.

Line

Value

L1
L2
L3
L4
L5

[m]
1.7
2.4
4.0
3.5
0.2

l
Standard
uncertainty
[m]
0.1
0.1
0.1
0.1
0.1

Z0
Standard
Value
uncertainty
[m]
[m]
78
2
27
2
167
2
1
2
63
2

l
Standard
Value
uncertainty
[mm]
[mm]
2.25
0.05
1.90
0.05
2.70
0.05
0.60
0.05
2.05
0.05

Tab. 5.6: Estimates of the center conductor displacement l, characteristic impedance


correction Z0 , and loss correction factor l obtained in our multi-frequency multiline
TRL calibration, along with standard uncertainties from the residual analysis.

multiline TRL calibration. The relative dierences in the real part are larger than for the
multiline TRL calibration in the coaxial Type-N calibration standard. We believe that this
may be attributed to a larger impact of surface roughness on the loss variation for smaller
connector diameters and at higher frequencies.
In Fig. 5.16 we show a plot of the residual standard deviation as a function of frequency,
as obtained in the classical multiline TRL (solid gray) and in our multi-frequency multiline
TRL method (solid black). We see that our method delivers the residual standard deviation
which for most frequencies is at least two times smaller than in the classical multiline
TRL method. This indicates that our description of the calibration standards (i.e., the
denitions of the calibration standards along with the VNA model) ts better the actual
measurements than the classical multiline TRL method. It is also important to note that
the improvement in the residual standard deviation is not as large as for the multiline
129

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION





 !"#









 !"#$




      


(a)

      


(b)

Fig. 5.15: Propagation constant for the 1.85 mm coaxial multiline TRL calibration: (a)
imaginary part normalized to the free-space propagation constant 0 , (b) real part; as
obtained from the classical multiline TRL (solid gray) [36], and from our multi-frequency
multiline TRL calibration (solid black).

TRL calibration in the coaxial Type-N calibration standard. We think that this can be
explained with an increased impact of connector nonrepeatability errors (i.e., bending on
the connector socket ngers, misalignment of the conductors) for smaller connector sizes.



 !"#





      



Fig. 5.16: Residual standard deviation for


the mutiline TRL calibration in the 1.85 mm
coaxial connector standard, as obtained from
the classical multiline TRL (solid gray) [36],
and from our multi-frequency multiline TRL
calibration (solid black).
method.
130

In Figure 5.17 we compare the residual


standard deviation of the eective source
match, eective directivity and the eective
tracking, respectively, as obtained from the
classical multiline TRL calibration (solid
gray) and from our multi-frequency multiline TRL calibration (solid black). We observe that our method consistently yields
much smaller uncertainties in the VNA calibration coecients than the classical multiline TRL method. Our uncertainties are
typically two times smaller than from the
classical multiline TRL, and at lower frequencies (i.e., below 1 GHz), the uncertainties we provide are by up to four times
smaller than in the classical multiline TRL

5.8. EXPERIMENTAL RESULTS






 !"#






 !"#





      


(a)



(b)




 !"#


 !"#





      







      


(c)




 !"#

      


(d)








 !"#





      


(e)

      


(f)

Fig. 5.17: Residual standard deviation of the eective directivity of (a) the VNA port 1
and (b) the VNA port 2, eective source match of (c) the VNA port 1 and (d) the VNA
port 2, and eective tracking of (a) the VNA port 1 and (b) the VNA port 2, as obtained
from the classical multiline TRL (solid gray) [36], and from our multi-frequency multiline
TRL calibration (solid black).

131

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION





! "#$%








! "#$%








      


(a)






 !"#$






      


(b)




 !"#$






      


(c)

      


(d)

Fig. 5.18: Reection coecient magnitude for a 1.85 mm coaxial (a) female and (b) male
at-short termination, along with the residual standard uncertainty (c,d), as obtained from
the classical multiline TRL (solid gray) [36], and from our muli-frequency multiline TRL
calibration (solid black).

In order to verify that the VNA calibrated with our approach yields correct measurements, we measured three verication standards: a pair of at-short terminations, a pair
of matched terminations, and a 10 dB attenuator. In Fig. 5.18 and Fig. 5.19 we show the
magnitude and phase, respectively, of the corrected reection coecient of the male and female at-short termination, along with the residual standard uncertainties, as determined
with the classical multiline TRL calibration (solid gray), and with the use of our method
(solid black). We see that the agreement between both methods is very good, however,
our method yields residual standard uncertainties that are at least two time smaller than
for the classical multiline TRL method.
We also note the for both methods the reection coecient magnitude at some frequencies is larger than one. Similarly to the measurement of the Type-N oset-open
132

5.8. EXPERIMENTAL RESULTS





 !"#









 !"#






      


(a)










 !"#





      


(c)

      


(b)




 !"#



      


(d)

Fig. 5.19: Reection coecient phase (with linear part removed) for a 1.85 mm coaxial (a)
female and (b) male at-short termination, along with the residual standard uncertainty
(c,d), as obtained from the classical multiline TRL (solid gray) [36], and from our mulifrequency multiline TRL calibration (solid black).

terminations, this non-physical result may be attributed to the asymmetry of the reect
standard. Another reason for this behavior is that when connecting the at-short termination, a dierent connector interface is formed than for the calibration standards, as there
is no center conductor gap on the DUT side. This dierence in the connector interface
contributes to a systematic error in the measurement of the at short.
In Fig. 5.20 we show the reection coecient magnitude of a male and female matchedtermination, along with the in-phase residual standard uncertainty (see Section 3.2), as
determined with the classical multiline TRL calibration (solid gray), and with the use
of our multi-frequency multiline TRL method (solid black). We again see a very good
agreement between the two methods with much smaller uncertainties (at least four times)
provided by our method.
133

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION






 !"#









 !"#






      


(a)





 !"#









(b)






      






 !"#



      


(c)

      


(d)

Fig. 5.20: Reection coecient for a 1.85 mm coaxial (a) female and (b) male matched
termination, along with the in-phase residual standard uncertainty (c,d), as obtained from
the classical multiline TRL (solid gray) [36], and from our muli-frequency multiline TRL
calibration (solid black).
In Fig. 5.21 we show the magnitude and phase (with the linear part removed) of the
transmission through the 10 dB attenuator, as determined with the classical multiline TRL
calibration (solid gray), and with the use of our method (solid black), along with residual standard uncertainties. While the agreement between the two calibration methods is
excellent, our methods again yields much smaller uncertainties.

5.9

Summary

We presented a generalization of the multi-frequency approach to VNA calibration (see


Paragraph 2.4.1-B). The principal idea of our generalization is to use the concept of the
physical error mechanism (see Section 4.2) to capture the relationships between calibration
134

5.9. SUMMARY





 !
" #$%&














 !"#



      


(a)




 !"#





      


(b)




 !"#






      


(c)

      


(d)

Fig. 5.21: Transmission for a 1.85 mm coaxial 10 dB attenuator: (a) magnitude, (b) phase
(with the linear part removed), along with the (c) magnitude and (d) phase residual standard uncertainty, as obtained from the classical multiline TRL (solid gray) [36], and from
our muli-frequency multiline TRL calibration (solid black).

standard S-parameters at dierent frequencies. We further developed a generic numerical


framework for the solution of such a multi-frequency VNA calibration problem and demonstrated it in the context of the multiline TRL calibration in the coaxial environment. The
experimental results for the multiline TRL calibration with Type-N and 1.85 mm coaxial
transmission lines clearly show that by accounting for the relationships between calibration
standard S-parameters at dierent frequencies, we can signicantly (by a few times) reduce
the uncertainty in the corrected VNA S-parameter measurements.
The approach we presented is the rst step towards the development of the VNA
calibration approach based entirely on the description of measurement errors in terms of the
physical error mechanisms (see Subsection 4.5.3). In order to implement such an approach,
a description of VNA instrumentation errors in terms of physical error mechanisms is
135

5. GENERALIZED MULTI-FREQUENCY VNA CALIBRATION


needed. In the following chapter, we propose a description of the VNA nonstationarity
errors, which are the primary source of VNA instrumentation errors, employing the concept
of the physical error mechanism.

136

Chapter 6
Multi-frequency stochastic modeling
of VNA nonstationarity errors
Entities should not be multiplied
beyond necessity.
William of Ockham

In this chapter, we present a framework for a multi-frequency description of VNA nonstationarity errors. These errors, caused by the temperature and humidity drift, cable
instability, and connector nonrepeatability, are the primary source of the VNA instrumentation errors (see Section 3.1). Our framework is based on a stochastic model whose
parameters are identied from repeated S-parameter measurements. The core of this model
is a generic description of the VNA nonstationarity which uses a set of lumped elements
added to the two-port network capturing the VNA calibration coecients. The stochastic
model is then constructed by allowing the parameters of the lumped elements to randomly vary. As a result, we describe the complicated frequency-dependence of the VNA
nonstationarity errors with the use of a very small set of frequency-independent random
variables.

6.1

Introduction

In Chapter 4, we introduced the multi-frequency description of errors in S-parameter measurements and discussed the origins of the statistical correlations between these
errors at dierent frequencies. We showed that these correlations result from the fact the
137

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY


S-parameter measurement errors are caused by some common fundamental physical error
mechanisms (see Subsection 4.2).
We further pointed out that the number of these mechanism is usually much smaller
than the number of variables captured in the S-parameter measurement, thus the information contained in the covariance matrix for the multi-frequency S-parameter-measurement
error is redundant. Consequently, when developing the multi-frequency error description, it
is more ecient to follow the bottom-up approach, that is, to rst identify and characterize
the fundamental error mechanisms, and from that to construct the covariance matrix.
This bottom-up approach is easily followed in the case of calibration standard errors.
Indeed, these errors result from some well-dened physical error mechanisms, related primarily to the tolerances on the dimensional and material parameters of the standards.
Consequently, physical models for errors in S-parameters of calibration standards follow
directly from the physical models of the standards themselves.
In the case of VNA nonstationarity errors, however, application of the bottom-up approach is much more dicult. The nonstationarity of VNA electrical parameters, manifesting itself as the connector nonrepeatability, cable instability, and the test-set drift, is
caused by a number of dierent complicated physical mechanism and is commonly regarded
as dicult to both understand and model [19]. In the case of the connector nonrepeatability, some interesting eorts have been made towards understanding of these mechanism
based on electromagnetic modeling of simplied connector structures [125, 126]. These
eorts provide valuable guidelines as to the connector interface design, however, it seems
that the complexity of real connector and cable structures is far beyond the capacity of
direct electromagnetic modeling.
Also, some attempts have been made in the literature to statistically model the drift of
measurement instruments [147149]. These models, however, developed for simple scalarmeasurement instruments, such as power meters, are not easily extendable to complicated
measurement systems such as the VNA.
Because of this complexity, VNA instrumentation errors are traditionally characterized
with the use of either an approximate statistical analysis or some heuristic methods. For
example, in the case of connector nonrepeatability or cable instability, the standard uncertainty of these errors is assessed based on repeated measurements of devices with dierent
S-parameters, such as matched and highly reective terminations [93, 95, 150, 151]. In
the case of the VNA test-set drift, the heuristic method of [94] is often used (see Subsection 3.5.2). In this method, the dierence between two sets of VNA calibration coecients,
138

6.2. GENERIC PHYSICAL MODEL FOR THE VNA NONSTATIONARITY


obtained before and after an experiment, is used as an estimate of the VNA test-set drift
during the experiment.
The above methods, however, have two important drawbacks. First of all, these methods do not allow to reliably estimate the statistical correlations between measurement
errors at dierent frequencies. Indeed, due to the large size of the covariance matrix for
the multi-frequency S-parameter measurement error, direct estimation of these correlations
based on repeated measurements would require an impractically large number of measurements [87]. Another important drawback of these methods, except for [94], is that instead
of considering the statistical properties of the VNA nonstationarity itself, they analyze only
the impact of this nonstationarity on the S-parameter measurement of some individual devices. As a results, the impact of the VNA nonstationarity errors on the measurements of
devices with other S-parameters is dicult to predicted.
Hence, our description of VNA nonstationarity errors is dierent. Instead of focusing
on the impact of the VNA nonstationarity on measurements, we begin with a generic
physical model for the VNA nonstationarity itself, proposed in [152] and discussed in
more detail in Section 6.2. This model describes the VNA nonstationarity with a set of
lumped elements added to the two-port network describing the VNA calibration coecients
(commonly referred to as the VNA error boxsee Paragraph 2.3.1-B). Based on that
description, we build a stochastic model for the VNA nonstationarity by allowing the
frequency-independent parameters of the lumped elements to randomly vary. We then
identify statistical properties of these parameters based on repeated measurements of a
single highly-reective oset termination. As a results, our model can be used to predict
the complicated frequency-depedendence of VNA nonstationarity errors for devices with
arbitrary S-parameters.

6.2

Generic physical model for the VNA nonstationarity

VNA nonstationarity errors often exhibit a regular frequency-dependence [91, 93, 150,
151, 153]. Also, a common experience of a microwave metrologist is that any instability in
the VNA setup, for example, due to cable bending or a loose connection, leads to periodic
ripples in corrected S-parameter measurements.
This type of a regular frequency-dependent behavior can easily be explained from a
physical point of view. Consider a small discontinuity that occurs within the VNA setup
139

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY












Fig. 6.1: Small change of one-port-VNA electrical parameters as a set of N + 1 perturbations.


after the calibration has been performed. Correction with the original VNA calibration
coecients does not account for this new discontinuity, hence the reections o this discontinuity aect the corrected VNA S-parameter measurements. Now, if one measures
a device that is not perfectly matched, will ripples are observed which correspond to the
distance between this new discontinuity and the plane at which the reection occurs in the
device.
Reference [91] attempts to quantitatively describe this behavior in the case of connector
repeatability errors. It uses a connector-interface equivalent circuit of [92] which is formed
by a series inductance, a series resistance, and a shunt capacitance inserted between the
DUT and the VNA. The elements of this circuit model the eects responsible for the
connector repeatability errors, such as the misalignment the center and outer conductors
(the shunt capacitance) and variation of the joint impedance (the series inductance and
resistance). A simple procedure for the identication of model parameters is then proposed
which is based on repeated measurement of a highly-reective oset termination.
The model for VNA nonstationarity errors we propose here (see [152]) generalizes the
approach of [91, 92]. We observe that in many cases a single discontinuity cannot explain
the observed ripple behavior in corrected S-parameter measurements. This is especially
true for connector repeatability errors for which changes of connector interface often occur
not only at the connector joint plane. Examples are bending of the center conductor ngers
or exing of the beads supporting the center conductor due to the mechanical strain applied
to this conductor.
Our equivalent circuit accounts therefore an arbitrary but otherwise xed number of
discontinuities, located at dierent distances from the VNA reference plane. This circuit
140

6.2. GENERIC PHYSICAL MODEL FOR THE VNA NONSTATIONARITY


 






Fig. 6.2: Lumped-element equivalent circuit for a single perturbation.
is shown in Fig. 6.1. It describes a small change of one-port-VNA calibration coecients
with a set of N + 1 small perturbations. These perturbations are added at dierent xed
distances dn , for n = 0, . . . , N , to the linear two-port network representing the VNA
calibration coecients. For the perturbation P0 which occurs at the VNA reference plane
we have d0 = 0.
We describe each of the perturbations with a lumped-element circuit shown in Fig. 6.2.
Compared to [91, 92], our circuit has an additional transformer which accounts for changes
of the characteristic impedance. Such changes may result, for example, from eccentricity of the inner conductor, variation of the connector socket diameter, or changes of the
transmission line diameters due to the temperature drift. We further describe the series
impedance perturbation as


Zn = Rn + jXn = RDC,n + (1 + j)RRF 0,n

f
+ jLn ,
f0

(6.1)

where f0 is a xed reference frequency. Such a description allows to account for both the
change of DC resistance and the change in the skin-depth resistance and inductance.
The circuit in Fig. 6.2 is fairly general and accounts for typical electrical characteristics
that a small discontinuity in a coaxial line might exhibit. In the case when additional
information on the electrical character of the perturbations is available, this circuit can be
accordingly modied. However, in this work, we do not strive to obtain nor to account for
such information. Instead, we show that a statistically and physically sound description
of the VNA nonstationarity errors can be developed with the use of the general model
presented in Fig. 6.1, and thus without any need to directly model the design details of
the connector interface, cable, or the VNA itself.
We represent the joint contribution of the perturbations Pn , for n = 0, . . . , N to the
141

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY








 














Fig. 6.3: Equivalent representation of the VNA nonstationarity.
error in corrected DUT S-parameter with an equivalent two-port network E inserted between the VNA and the DUT (see Fig. 6.3). In order to determine parameters of this
network, we assume that the perturbations are small, which allows us to neglect multiple
reections and superimpose the contributions of individual perturbations. The scattering
matrix of this equivalent network is given in compact form as (see Appendix F)

S=

N
+
j0 dn
n=0 wR1 (f )pR,n e

1 wT (f )pT

1 wT (f )T pT
N
+
n=0

j0 dn

wR2 (f )pR,n e

(6.2)

where real-valued model parameters are dened by


pT = [

N
+
n=0

pR,n =


RDC,n

N
+
n=0

R RF 0,n

N
+
n=0

(Ln + Cn ) ]T ,


RDC,n R RF 0,n Ln Cn Z0,n

T

(6.3)
(6.4)

and
RDC,n
,
Zref
RRF 0,n

,
RRF
0,n =
Zref
Ln
,
Ln =
Zref
RDC,n =

Cn = Cn Zref ,
Z0,n

=
,
Z0,n
Zref
142

(6.5)
(6.6)
(6.7)
(6.8)
(6.9)

6.2. GENERIC PHYSICAL MODEL FOR THE VNA NONSTATIONARITY


where Zref is the reference impedance, typically 50 . The quantities wT (f ), wR1 (f ), and
wR2 (f ) are the frequency dependent weighting functions dened as

wT (f ) =

f
f0

(6.10)

j2f
and

wR1 (f ) =

f0

j2f

f0

j2f

wR2 (f ) =

(6.11)

where 0 = /c is the propagation constant in the air, c is the speed of light in vacuum.
We neglect the conductor loss in the transmission lines since the distances dn are small.
Also, we assume that Z0,n is real since small dimensional changes are, to rst order, aecting
only the magnitude of lines characteristic impedance (see Subsection 5.3.2).
The overall error due to the perturbations Pn , as described by (6.2), has important
properties. To rst order, this error is independent of the VNA calibration coecients
and is determined only by the electrical parameters of the perturbations, pT and the pR,n ,
and the distances dn of these perturbations from the VNA reference plane. Consequently,
equation (6.2) can be applied to describe the connector repeatability and cable instability
errors which are independent of the individual VNA setup and are rather a property of a
particular connector interface or cable.
The distances dn , for n = 0, . . . , N are xed and do not undergo any statistical variation.
For example, for the connector repeatability errors, these distances correspond to some
design-xed dimensions (e.g. position of the bead, length of connector socket ngers,
length of the connector pin) of the connector interface. The same reasoning applies to the
cable instability errors or the test-set drift errors. Hence, the only parameters that vary
are pT and the pR,n , which we thus group into a single vector

p=

pT
pR,0
..
.
pR,N

143

(6.12)

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY


The models for VNA nonstationarity errors we describe in the following sections are then
constructed by assuming dierent statistical properties of the parameters (6.12).
Finally, it is important to note that the frequency dependence of (6.2) is set by xed
weighting functions (6.10) and (6.11), the distances dn , and a small set of frequency independent parameters (6.12). Consequently, once the distances dn and statistical properties
of (6.12) are known, the impact of the VNA nonstationarity can easily be determined any
set of measurement frequencies.

6.3

Stochastic model for connector nonrepeatability


and cable instability

In this section, we describe the stochastic models for the connector nonrepeatability and
cable instability. We describe these models in the same section because the assumptions
underlying them are identical. However, it is important to note that both models are
identied and used independently of each other. In these models, we use the description
(6.2) for VNA nonstationarity with the parameters (6.12) replaced by random variables.
In the following, we rst discuss the assumptions made as to the statistical properties of
these variables. We then proceed with the procedure for identication of those properties,
and nally present experimental verication of our models. The discussion we present here
is an extension of [154].

6.3.1

Statistical properties of circuit parameters

We capture the electrical parameters of the connector interface or the cable which
undergo random variations when reconnecting the interface in the vector (6.12). Since
these parameters describe the changes of some characteristics, we assume that E (p) = 0.
We further assume that the probability density function (PDF) of the vector p is normal.
Consequently, the statistical properties of the vector p are completely described with the
covariance matrix


p = E ppT .
(6.13)
The assumption that the vector p has a normal PDF can easily be justied from
the physical point of view. Parameters in the vector p represent the impact of mechanical
changes of the connector interface or the cable in terms of some electrical parameters. Since
these changes are small, parameters in the vector p are linear combinations of changes in
144

6.3. STOCHASTIC MODEL FOR CONNECTOR AND CABLE ERRORS


some geometrical parameters of the connector interface or the cable, such as displacements
of the conductors, bending of the socket ngers, exing of the center conductor beads or
change of the cable length. It is reasonable to assume that all of those mechanical changes
are normally distributed, hence the vector p has also a normal probability distribution
function.

6.3.2

Estimation of the covariance matrix of circuit parameters

We estimate the covariance matrix (6.13) from repeated corrected VNA-measurements


of an oset termination. In the case of the connector-nonrepeatability model, we perform repeated measurement of the termination on the VNA port without the cable while
reconnecting and rotating the device before each measurements. When identifying the
cable-instability model, we perform measurements of the termination on the VNA port
extended with the cable, while randomly bending the cable before each measurement. Preferred devices for the model identication are highly-reective oset-terminations, due to
the larger impact of VNA nonstationarity on the their measurements [91]. However, our
estimation method can be used for oset terminations with arbitrary non-zero reection
coecients.
Let
k , denote the unknown true value of the one-port device reection coecient at
the frequency fk , for k = 1, . . . , K, and let mk be the m-th corrected measurement of this
reection coecient in a series of repeated measurements, where m = 1, . . . , M . We can
model this measurement by transforming
k through the linear two-port network given by
(6.2). Considering only rst order terms we obtain the approximation
k + xk (d,
k )T pm + mk ,
mk =

(6.14)

where the parameters pm characterize the change of VNA calibration coecients in the
m-th measurement, mk is the VNA measurement noise,

k ) =
xk (d,

2
k wT (fk )
ej20 d0 wR1 (fk ) + ej20 d0
2k wR2 (fk )
ej20 d1 wR1 (fk ) + ej20 d1
2k wR2 (fk )
..
.
ej20 dN wR1 (fk ) + ej20 dN
2k wR2 (fk )
145

(6.15)

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY


and
d = [d0 , . . . , dN ]T .

(6.16)

m of the connector interface parameters pm ,


Our goal now is to determine the estimates p
p of their covariance matrix, the estimates dn of the perturbation locations
the estimate
k of the unknown reection coecient
dn , for n = 0, . . . , N , and the estimates
k .
Generic solution to that problem requires the use of statistical methods for the analysis
mixed-eect models for repeated measurements [155]. In these methods, the likelihood
function for all of the estimated parameters is written and the sought parameters are determined by maximizing this function. In the case of the nonlinear model (6.14), this leads
to a dicult constrained nonlinear optimization problem (the estimate of the covariance
matrix (6.13) needs to be positive denite), therefore in the following we proceed in a
simplied manner.
k of
We rst note that, since the VNA nonstationarity errors are small, the estimate
the reection coecient at the frequency fk is well approximated by the mean value
M

k = k = 1
mk .

M m=1

(6.17)

m , dn , and the covariAfter inserting this estimate into (6.14), we only have to determine p
p . We further assume that this matrix can be approximated as the sample
ance matrix
m , that is [87]
covariance of the estimates p
p =

M

1
mp
Tm .
p
M 1 m=1

(6.18)

The error of this approximation will depend on the accuracy with which we determine
m . This accuracy depends on the variance of the VNA measurement noise
the estimates p
mk , the number of measurement frequencies, and the error of the model (6.2). In order
to increase this accuracy, we perform the measurements for a large number of frequencies,
and reduce the VNA receiver errors by narrowing the IF bandwidth and averaging multiple
results.
Equipped with the above assumptions, we can pose the original problem as the task of
m , for m = 1 . . . , M , and dn , for n = 0, . . . , N under the condition (6.17).
determining p
To this end, we write the measurements (6.14) in a more compact form as
ym = y + X (d) pm + m ,
146

(6.19)

6.3. STOCHASTIC MODEL FOR CONNECTOR AND CABLE ERRORS


where

m1

ym =
.. , y =

m1
..
.
, m = ..
.

mK
K

mK
and

x1 (d, 1 )T

..

.
X (d) =
.
xK (d, K )T

(6.20)

(6.21)

We further assume that the VNA receiver errors have the same circular-normal PDF at each
frequency (see Subsection B.6), and that the errors at dierent frequencies are statistically
uncorrelated (see Subsection 4.4.2). Thus, the VNA measurement noise mk has also the
circular-normal PDF whose variance can be determined from (3.60) as
E(mk mk ) =



2 2

2
 1 2ESk k + ESk

k 



 E( )
ERk



= V2k 2 ,

(6.22)

where ESk and ERk are the source match and tracking at the frequency fk , determined in
the VNA calibration preceeding the experiment, and E( ) = 2 is the unknown variance
of the VNA receiver errors. Consequently, we can write the covariance matrix for the
vector m as

V21

...
2
)
=

(6.23)
E(m H

= 2 V .
m

V2K

m
Now, we use the maximum likelihood approach (see Appendix B) and determine the p

and the dn as the solution to the nonlinear least squares problem


= arg min

where the parameter vector is

M


rm (d, pm )H V1 rm (d, pm ) ,

(6.24)

m=1

p
1
..
.

pM

147

(6.25)

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY


and the residuals are
rm (d, pm ) = ym y X (d) pm .

(6.26)

We nd the solution to the problem (6.24) with the use of the Levenberg-Marquardt
algorithm (see Subsection B.4.1-B). We exploit the linear dependence of (6.26) on the
parameters pm to reduce the dimensionality of the optimization problem. Details are
given in Appendix G.
m are determined, we calculate their sample covariance matrix
Once the estimates p
(6.31). We further reduce the dimensionality of this covariance matrix with the use of the
principal component analysis [87]. That is, we determine the approximation
p U UT ,

(6.27)

where is a diagonal matrix of size M M with non-zero diagonal elements, referred to


as principal components, U is a full column-rank rectangular matrix of size P M with
orthogonal columns, P is the length of the vector p, and M < P is the number of principal
components. We choose the number of principal components in the approximation (6.27)
so as to capture 99 % of the total variance contained in the original matrix (6.31), which
corresponds to the error of this approximation (in the sense of the Frobenius matrix norm
[121]) less than 1 %. This results typically in two or three principal components.
The principal components obtained through the approximation (6.27) can be considered
as the electrically-equivalent physical error mechanisms (see Section 4.2) characterizing the
connector interface or cable instability errors. Indeed, by rescaling the weighting functions
(6.10) and (6.11) by the columns of the matrix U, we can completely characterize the
connector nonrepeatability or cable instability, as described by the matrix (6.2), with a set
of M independent variables with variances determined by the diagonal elements of .

6.3.3

Experiments

A. Connector repeatability errors. We illustrate our approach with experimental


results for the 1.85 mm coaxial-connector interface. We used two measurement setups
which are schematically shown in Fig. 6.4. In each case, measurements were performed
with the Agilent E83631 VNA in the frequency range 0.2 67 GHz and with the frequency
spacing of 50 MHz. In order to reduce the inuence of the VNA receiver errors, we narrowed
the IF bandwidth to 10 Hz and averaged four sweeps. This resulted in a relatively long
measurement time of around 2 min for a single reection-coecient measurement. Both the
148

6.3. STOCHASTIC MODEL FOR CONNECTOR AND CABLE ERRORS

(a)

(b)

Fig. 6.4: Measurement setup for (a) female and (b) male one-port DUTs.
VNA port 1 and the cable end on the DUT side were equipped with additional adapters,
in order to minimize the wear on the VNA-port and cable connectors as those elements
cannot easily be replaced.
In the setup shown in Fig. 6.4a, we rst calibrated the VNA with the SOLT calibration
employing four oset shorts with dierent lengths, an oset open, a matched termination
and a direct thru connection. Then we performed repeated measurements of dierent
one-port female DUTs on the VNA port 1. Before each measurement, the DUT was
disconnected, rotated by 45 , and connected again. A similar experiment was performed
in the setup show in Fig. 6.4b, however, after the calibration, we measured dierent oneport male DUTs.
Fig. 6.5 shows the in-phase and quadrature component (see Subsection 3.2.2) of the
standard deviation of 16 repeated reection-coecient measurements of 5.4 mm and
7.6 mm long, female and male oset shorts, and 5.4 mm long female and male oset opens,
along with the standard deviation predicted from our stochastic model. In the model,
we used three perturbations, corresponding to the nonrepeatability of the connector joint
(perturbation located at the connector joint plane), exing of the bead supporting the
center conductor (perturbation located within the adapter, 3 mm away from the connector
joint plane), and bending of the connector socket ngers (perturbation located within the
DUT for female DUTs, and within the adapter for male DUTs, in each case 1.1 mm away
from the connector joint plane). We estimated the locations of the perturbations based
on observations under the microscope and the specications [135], and kept them xed in
the analysis.
The stochastic models we used employed two random variables in the case of the 5.4 mm
149




 ! 
"



#  









  

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY

$ %&' !


      
 




! 
"



#




$%&'!



 !  

 "




#  



$
%&' !




(b)

      
 
 


 


 

(a)








 

 !





" 

# $%& 

      
 
 
(d)


  



  


(c)



      


      
 
(e)




 

 !
" 






# $%& 
      
 
(f)

Fig. 6.5: In-phase and quadrature component of the standard deviation of 16 repeated
reection coecient measurements of 1.85 mm coaxial oset terminations : (a) 5.4 mm
long female short, (b) 5.4 mm long male oset-short, (c) 7.6 mm long female oset-short,
(d) 7.6 mm long male oset-short, (e) 5.4 mm long female oset-open, (f) 5.4 mm long male
oset-open; measurement (gray) and stochastic model prediction (black).

150

6.3. STOCHASTIC MODEL FOR CONNECTOR AND CABLE ERRORS


long oset shorts, three random variables in the case of the 7.6 mm long short, two random
variables for the 5.4 mm long female oset open, and only one random variable for the
5.4 mm long male oset open. We chose the number of variables such that the error of the
approaximation (6.27) is less than 1%. We observed that for each termination, when using
only one random variable, the error of the approximation (6.27) was less then 10% (in the
case of the 5.4 mm long male oset open it was alread less than 1%), and increasing the
number of variables yielded only a small improvement.
In Fig. 6.5 we see that the overall agreement between the stochastic model prediction
and the measurements for the quadrature (phase) errors is very good except for a small
discrepancy in the frequency range below 4 GHz. We think that the discrepancy is caused
by an increased VNA test-set drift in this frequency range (the 16 repeated measurements
took over 30 min), since we observe it in the measurements of all of the devices.
The agreement for the in-phase (magnitude) errors is not as good as for the quadrature
errors and, in general, our model underestimates the in-phase errors. We think that this
is due to the VNA receiver errors which are not accounted for by our model. On the
other hand, the in-phase errors are of an order of magnitude smaller than the quadrature
errors, and therefore less important. Consequently, our stochastic model is capable of
representing the dominant component of the connector repeatability errors, as observed in
the measurements of the three dierent oset terminations.
We further note that the connector repeatability errors for the 5.4 mm long male open
are much larger than for other terminations. We think that this might be caused by an
excessive eccentricity of the center conductor or a loose connection between the adapter and
the VNA ports during the measurment of the open. In order to verify those hypotheses,
we plan on inspecting the oset open under a microscope, and if necessary, repeating the
experiment.
In the measurments of all of the terminations, except for the 5.4 mm long female open,
we observe that the connector repeatability errors are comparable for each pair of male
and female oset terminations with the same length. This can be justied by the fact
that both male and female terminations have the same connector interface, and that the
mechnical forces put on this interface (e.g., due to the misalignment of the inner and outer
conductors) are approximately the same since both termiantions have the same length.
We also note that the connector nonrepeatability errors are smaller for longer osetterminations. We think that this can be explained by the fact that for longer terminations
the displacement of the inner conductor at the connector joint causes a more gradual
151





" # $ %




 


 

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY






      
   !

! " # $









(b)


  


 

(a)

" # $ %






     & 
   !



 ! 







     " 
  
(d)

 ! "#

     $ 
   

     !


  

(c)










     % 
  

(e)









"#$ %

     & 



(f)

Fig. 6.6: Connector repeatability errors in 16 repeated reection coecient measurements


of 1.85 mm coaxial one-port devices along with the 95 % uncertainty bounds predicted from
the stochastic model developed from the measurements of a 5.4 mm long female osetshort: (a) in-phase error for a 6.3 mm long female oset-short, (b) quadrature error for the
6.3 mm long female oset-short, (c) in-phase error for a 5.4 mm long female oset-open,
(d) quadrature error for the 5.4 mm long female oset-open, (e) real part of the error for
a matched termination, (f) imaginary part of the error for the matched termination.

152

6.3. STOCHASTIC MODEL FOR CONNECTOR AND CABLE ERRORS


bending of the center conductor along the termination.
In the next step, we investigated whether the model we developed based on the measurements of an oset short can be used to predict the connector repeatability errors in the
measurements of other one-port devices. Fig. 6.6 compares the distribution of 16 repeated
measurements of a 6.3 mm long oset short, a 5.4 mm long oset open, and a matched
termination, respectively, with the 95 % uncertainty bounds predicted from the stochastic
model developed based on the measurements of the 5.4 mm long oset short. In Fig. 6.6a
and Fig. 6.6b, we compare the in-phase and quadrature component of the deviation of each
measurement from the mean, while in Fig. 6.6c we show the real and imaginary part of this
deviation. In the case of the 6.3 mm long oset short and the 5.4 mm long oset open, our
stochastic model accurately predicts the uncertainty bounds for the quadrature component.
For the in-phase component, we underestimate the observed variability which, however, is
of an order of magnitude smaller then the variability in the quadrature component.
In the case of the matched termination our prediction of the uncertainty bounds is not
as good as for the oset terminations. We think that this may be attributed to the design
of the matched termination. This termination is constructed with the use of a planar
resistor inserted into the coaxial transmission line at some distance from the connector
joint. The center conductor in such a construction is less rigid than in an oset short,
hence the mechanical strain put on this conductor aects not only the connector interface,
but also the planar structure forming the matched termination.

B. Cable instability errors. In order to verify our model, we made an experiment


with a high-quality 66 cm long microwave cable. The measurements were performed in the
setup shown in Fig. 6.4b, with the same VNA settings as in the experiment described in the
previous paragraph. We performed repeated reection-coecient measurements of dierent
one-port devices attached to VNA port 2 while randomly bending the cable between the
measurements. We kept the cable positions within the range typically observed in practice.
We determined the initial estimates of the perturbation locations for the optimization
procedure (6.24) based on a time-domain representation of the measurments. We obtained
this represenation by transforming each reection coecient measurement into the time
domain (see [156, 157]), and then taking the standard deviation of the resulting waveforms
at each time point. The maxima of such a representation correspond therefore to the
locations within the cable (and the adapter attached to it) at which changes occur while
the cable is being bent.
153



 






6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY







             

 














     


(a)

(b)

Fig. 6.7: Standard deviation of the time-domain representation of 16 repeated reection


coecient measurements of a 5.4 mm long 1.85 mm coaxial male oset-short, with random
cable exure: (a) from 1.4 m to the VNA reference plane (b) zoom into the range from
80 mm to the VNA reference plane.
Fig. 6.7 shows the time-domain represenation we obtained for 16 repeated measurements
of a 5.4 mm long 1.85 mm coaxial male oset short. The time points were then transformed
into equivalent distances by assuming that the wave is propagating in the free space. The
VNA reference plane is located at the distance 0 and negative distances correspond to the
locations away from the VNA reference plane towards the VNA. In Fig. 6.7a we see that
the standard deviation has the largest values at around 50 mm, 700 mm, and 800 mm.
These distances correspond to both ends of the cable1 . The changes around 50 mm are
dominant, therefore we decided to disregard the perturbations at 700 mm and 800 mm.
In Fig. 6.7b, we further expand the time-domain representation for the range from
80 mm to 0 mm. The largest variations are occuring in the range between 60 mm
and 40 mm, and between 30 0 mm and 0 mm, that is, within the adapter attached
to the cable. The change withint the adapeter are smaller then the perturbations at
700 mm and 800 mm, hence we disregerd them in our model. As a result, we use
two perturbations related to two peaks with maximal amplitude which occurr in the range
between 60 mm and 40 mm. Based on the Fig. 6.7a, we estimated the locations of
these peaks to 53.7 mm and 49.2 mm, which were further rened in the estimation
procedure to 54.2 mm and 48.7 mm, respectively.
In Fig. 6.8 we show the in-phase and quadrature component of the standard deviation of
1

Since the cable is lled with a dielectric, the dierence between the location of the rst and last
perturbation is not equal to the physical length of the cable.

154

  


6.3. STOCHASTIC MODEL FOR CONNECTOR AND CABLE ERRORS






!  
%



! "#$
%

&







      




Fig. 6.8: Standard deviation of 16 repeated reection coecient measurements of a 5.4 mm


long 1.85 mm male coaxial oset-short with random cable exure, represented as the inphase (blue) and quadrature (red) component, along with the stochastic model prediction
(black).

the repeated measurements for the oset short along with the standard deviation predicted
from our stochastic model. The model we used employed only two random variables
determined from on the approximation (6.27). The agreement between our model and the
measurements is very good for both the in-phase and quadrature component. We note
that in-phase errors are comparable with the quadrature errors which indicates that the
instability of the cable aects both its phase shift and attenution.
We further used the stochastic model for cable instability errors developed for the
5.4 mm long 1.85 mm coaxial male oset short to predict these errors in measurements
of other one-port devices, that is, a 7.6 mm long male oset-short, a 5.4 mm long male
oset-open and a coaxial male matched termination. The results of those measurements
are shown in Fig. 6.9. In both gures, we show the deviations of the measurements from
their mean along with the 95 % uncertainty bounds predicted from our stochastic model.
In the case of the oset open, we show separately the in-phase and quadrature component
of the deviations, while in the case of the matched termination, we compare the real and
and imaginary part of the deviations. In both cases, we see that the measurements lie
within the uncertainty bounds predicted from the model. The prediction of these bounds
for the matched termination is slightly worse than for the open. This may be attributed
to the VNA receiver errors whose relative contribution is larger for devices with smaller
reection coecients.
155









"#$  %









 


 


6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY

      !
 


!"#

     
  
(d)


"#$  %






       !
 


 


 


(c)









!"#

     
  

 
 

(b)

    !

(a)


!"#$






      


 

(e)









#$% &

      "



(f)

Fig. 6.9: Cable instability errors in 16 repeated reection coecient measurements of


1.85 mm coaxial one-port devices along with the 95 % uncertainty bounds predicted from
the stochastic model developed from the measurements of a 5.4 mm long male oset-short:
(a) in-phase error for a 7.6 mm long male oset-short, (b) quadrature error for the 7.6 mm
long male oset-short, (c) in-phase error for a 5.4 mm long male oset-open, (d) quadrature error for the 5.4 mm long male oset-open, (e) real part of the error for a matched
termination, (f) imaginary part of the error for the matched termination.

156

6.4. STOCHASTIC MODEL FOR VNA TEST-SET DRIFT

6.4

Stochastic model for VNA test-set drift

In the following, we discuss the stochastic model for the VNA test-set drift. In this
model, we use the description (6.2) of VNA nonstationarity errors with the vector (6.12)
replaced by a set of stochastic processes. We employ here the stochastic Wiener process
which forms the mathematical description of the random walk pheonomenon. We rst
discuss the properties of the scalar stochastic Wiener process, and then show how to adopt
this concept to a more complex situation of the VNA test-set drift. Then we discuss
the estimation of the process covariance matrix and present experimental verication our
model.

6.4.1

Drift as the multidimensional random walk

The drift manifests itself as a measurement bias varying with time [147]. In the case of
VNA S-parameter measurements, this bias results from a continuous change of VNA electrical parameters in time due to aging, and temperature and humidity changes. Therefore,
at any time point after the calibration, the VNA calibration coecients slightly dier from
those determined during the calibration procedure. We further observe that the maximum
value of this dierence increases with time. However, it is also possible that, after some
period of time, the VNA calibration coecients take on again the values close to those
determined during the calibration.
The above description of the VNA test-set drift process resembles closely the random
motion of a particle in a uid, commonly referred to as the Brownian motion [158]. At
any time after the motion begins, the particle, due to a series of random collisions, is away
from the starting point. Although we expect the worst case value of the distance of the
particle from the starting point to increase with time, it is also possible the particle comes
back to the vicinity of the starting point.
The mathematical description of the one-dimensional Brownian motion is given by the
stochastic Wiener process [158]. Considering the similarities between the Brownian motion
and the drift process, Reference [147] suggests applying the stochastic Wiener process to
model the drift of a simple scalar measurement instrument, such as power meters. The
experimental results presented in [147] conrm, that the statistical properties of the drift
process agree well with those of the stochastic Wiener process.
In the context of scalar measurements, we can summarize the properties of the stochastic Wiener process w(t) as follows [147, 158]:
157

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY


a) w(t) = 0, that is, at the time t = 0, right after the calibration has been performed,
the measurement instrument has no bias;
b) w(t) N (0, t), that is, w(t) has a normal PDF with E[w(t)] = 0 and E[w(t)2 ] = t,
where > 0 is the process parameter;
c) w(t2 ) w(t1 ) N (0, (t2 t1 )), that is, dierences in bias for non-overlapping time
periods are statistically independent and have the normal PDF with zero expectation
value and the variance proportional to the length of the time period;
d) E[w(t)|w(t1 ); t1 t] = w(t1 ), that is, if the measurement bias w(t1 ) at the time t1
were subtracted from the measurements for time t t1 , the expectation value of the
process would again be zero, as for t1 = 0.
In the case of the VNA test-set drift, we have to do with complex system characterized by a
large number of frequency-dependent parameters. However, as pointed out in Section 6.2,
small changes of VNA calibration coecients at multiple frequencies can be approximated
with the model (6.2). This model employs a set of frequency-independent parameters captured in the vector (6.12) and some xed functions characterizing the frequency-dependence
of the change in the VNA calibration coecients. Consequently, we can describe the
VNA test-set drift by letting the parameters (6.12) vary according to the vector stochastic
Wiener process. This process has analogous properties to the scalar process, that is (see
Appendix H):
a) w(t) = 0,
b) w(t) N (0, t),
c) w(t2 ) w(t1 ) N (0, (t2 t1 )),
d) E[w(t)|w(t1 ); t1 t2 ] = w(t1 ),
where is positive-denite symmetric full-rank matrix characterizing the process. Consequently, we construct the description of the VNA test-set drift with the model (6.2) where
statistical properties of the model parameters p are described by the vector stochastic
Wiener process.
158

6.4. STOCHASTIC MODEL FOR VNA TEST-SET DRIFT

6.4.2

Estimation of the process covariance matrix

We estimate the process covariance matrix based on repeated corrected VNA measurements of a one port device, performed over a long period of time. During the measurements, the VNA setup is kept unchanged, that is, the device is not reconnected and the
position of the cables is not changed. Similarly to the estimation of the connector interface
model, the preferred devices are highly reective terminations, due to the larger impact of
the changes in VNA calibration coecients on the their measurements.
The measurements are described by a model similar to (6.14), that is
mk =
0k + xk (d,
0k )T pm + mk ,

(6.28)

where pm is the change of VNA parameters in the m-th measurement with reference to
the state at the beginning of the experiment,
0k is the value of the reection coecient
measured with the drift-free VNA, mk is the VNA receiver noise, d is given by (6.16) and

0k ) =
xk (d,

2m,k wT (fk )
j20 d0
e
wR1 (fk ) + ej20 d0
20k wR2 (fk )
ej20 d1 wR1 (fk ) + ej20 d1
20k wR2 (fk )
..
.
ej20 dN wR1 (fk ) + ej20 dN
20k wR2 (fk )

(6.29)

m of the VNA parameters changes pm , for


Our goal now is to determine the estimates p
p of the process
m = 1 . . . , M where M is the number of measurements, the estimate
covariance matrix, the estimates dn of the perturbation locations dn , for n = 0, . . . , N ,
0k of the unknown reection coecient
and the estimates
0k , for frequencies fk , where
k = 1, . . . , K.
We use a similar approximate approach as in the case of the connector interface errors.
We rst note that the true value
0k of the reection coecient can be well approximated
by the measurement taken right after the calibration, that is,
0k = 1k .

(6.30)

p . We
m , dn , and
After inserting this estimate into (6.28), we only need to determine p
p with the use of the p
m and based on (H.13). That
further approximate the estimate
159

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY


is,
p =

M
1

1

pm
pTm ,
(M 1) T m=1

(6.31)

m+1 p
m , for m = 1, . . . , M 1, and T is the time increment between the
where
pm = p
consecutive measurements. The error of this approximation will depend on the accuracy
m . In order to increase the accuracy of the
with which we determine the estimates p
approximation (6.31), we perform the measurements for a large number of frequencies,
and reduce the VNA measurement noise by narrowing the IF bandwidth and averaging
multiple results.
By use of the above assumptions, we can now pose the original problem as the task of
m , for m = 1 . . . , M , and dn , for n = 0, . . . , N under the condition (6.30).
determining p
To this end we write the measurements (6.28) in a more compact form as
ym = y + X (d) pm + m ,
where

m,1
11
..

..
ym = . , y = .

m,K
1K
and

(6.32)

m1

..
, m = . ,
mK

x1 (d, 11 )T

..

X (d) =

.
xK (d, 1K )T

(6.33)

(6.34)

m and the dn as the solution to a nonlinear least squares


Now, we can determine the p
problem analogous to (6.24). We solve this problem with the same methods as in the case
of the connector interface errors (see Appendix G).
m are determined, we calculate the estimate (6.31) of the process
Once the estimates p
covariance matrix. Similarly to the connector repeatability and cable instability model, we
reduce the dimensionality of this covariance matrix with the use of the principal component
analysis [87]. That is, we determine the approximation
p U UT ,

(6.35)

where is a diagonal matrix of size M M with non-zero diagonal elements, referred to


as principal components, U is a full column-rank rectangular matrix of size P M with
160

6.4. STOCHASTIC MODEL FOR VNA TEST-SET DRIFT









       


  



 



 










    


     


(a)

(b)

Fig. 6.10: Standard deviation of the time-domain representation of changes in the reection
coecient of a 5.4 mm long 1.85 mm male coaxial oset-short within 2 min long periods,
as repeatedly measured over the period of 12 hours: (a) from 1.4 m to the VNA reference
plane (b) zoom into the range from 100 mm to the VNA reference plane.
orthogonal columns, P is the length of the vector p, and M is the number of principal
components. We choose the number of principal components in the approximation (6.35) so
as to capture 99 % of the total variance contained in the original matrix (6.31). This results
typically in two or three principal components. Analogously to the connector interface
model, these principal components can be treated as the electrically-equivalent physical
error mechanisms (see Section 4.2), characterizing the VNA test-set drift.

6.4.3

Experiments

We veried our approach by making repeated reection-coecient measurements of


dierent one-port devices over a long period of time. The measurements were performed
in the setup shown in Fig. 6.4b, with the same setting as for the connector repeatability
measurements (see Paragraph 6.3.3-A). The one-port devices were measured on the VNA
port 1.
Fig. 6.7 shows the time-domain representation of the results we obtained for a 5.4 mm
long 1.85 mm coaxial female oset short. The measurements were performed every 2 min
over the period of 12 hours. For each pair of consecutive measurement, we calculated the
change in the reection coecients, transformed those changes into the time domain in
the same manner as in the case of the cable instability errors (see Paragraph 6.3.3-B), and
then calculated the standard deviation at each time point. The peaks in this representation
161

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY


correspond therefore to locations within the VNA at which changes due to the drift are
occuring.
The standard deviations shown in Fig. 6.7 are much smaller (almost by an order of
magnitude) than for the cable instability errors. This is because we are observing changes
due to the drift within short 2 min long periods of time. For the same reason, the impact
of the VNA receiver errors is more visible in Fig. 6.7, and manifests itself by a high noise
oor at the level of around 2 106 .
In Fig. 6.7a we see further, that the dominant changes are occuring in the range between
100 mm and the VNA reference plane. This range is expanded in Fig. 6.7b, where we
see peaks in the range between 70 mm and 50 mm, and between 10 mm and 0 mm.
Based on that, we decided to use in our model four perturbations. In Tab. 6.1 we put
together the initial locations of those perturbation, approximated based on the Fig. 6.7b,
along with the rened estimates determined in the modeling procedure.
In Fig. 6.8 we present results
of this modeling procedure. In
Locations [mm]
this gure, we show the in-phase
Initial
0
-4.5
-31.3 -63.8 -66.0
and quadrature component of the
Estimated
0
-4.9
-30.9 -63.5 -66.3
standard deviation of the changes
Tab. 6.1: Locations of the perturbations in the calculated in the frequency dostochastic model for the VNA test-set drift developed main, along with the prediction
based on the measurements of a 5.4 mm long 1.85 mm of the stochastic model we develcoaxial female oset-short.
oped. The model we used employed two random variables. The
agreement between the stochastic model prediction and measurement for the quadrature
(phase) errors is very good except for the discrepancy in the frequency range below 4 GHz.
We also noticed this discrepancy is in other experiments (see Subsection 6.3.3-A). We think
the this discrepancy is related to the fact that the VNA uses a dierent set of of couplers
and mixers in below 4 GHz. Another conrmation for this hypothesis can be seen in the
in-phase component of the standard deviation, determined based on the measurements,
which abruptly increases below 4 GHz.
The agreement for the in-phase (magnitude) errors is not as good and, in general,
our model underestimates the in-phase component of the standard deviation. We think
that this component is larger in the actual measurements due to the presence of the VNA
receiver noise not accounted for by our model. On the other hand, the in-phase errors
162

  


6.4. STOCHASTIC MODEL FOR VNA TEST-SET DRIFT




  
$



  !"#
$

%







      




Fig. 6.11: Standard deviation in the reection coecient of a 5.4 mm long 1.85 mm male
coaxial oset-short within 2 min long periods, as repeatedly measured over the period of
12 hours, represented as the in-phase (blue) and quadrature (red) component, along with
the stochastic model prediction (black).

are much smaller and, therefore, less important than the quadrature errors. Consequently,
our stochastic model is capable of adequately representing the VNA test-set drift errors as
observed in the measurement of the short.
We further used the stochastic model for VNA test-set-drift we developed for the 5.4 mm
long 1.85 mm coaxial female oset short to predict the errors related to the drift in other
measurements performed on the same VNA. To this end, we remeasured the same short,
and then measured also other one-port devices, that is, a 5.4 mm long 1.85 mm coaxial oset
female open and a coaxial female matched termination. In each case, the measurement
were performed every 1/2 hour for the period of 12 hours.
The results of those measurements are shown in Fig. 6.12. In all gures, we show
the deviations of the measurements from the initial measurement performed right after
the calibration, along with the 95 % uncertainty bounds predicted from our stochastic
model. In the case of the oset open and oset short, we show separately the in-phase and
quadrature component of the deviations, while in the case of the matched termination, we
compare the real and and imaginary part of the deviations.
In the case of the oset short, the variation of the quadrature component of the deviations is well predicted by the uncertainty bounds predicted from the model. As for the
oset open, our model slightly underestimates the deviations of the quadrature component
of the deviations. We think that this may be caused by a larger variation of the temperature in the laboratory during the measurements performed for the open, as compared with
the variation that occurred during the measurements used to establish the model. Also,
163

 
 

"#$ %

      !
  

(a)

"#$ %






       !
  
(a)
 
!"#$
 
 

 
 
 
     


 
 









 
   

 


 


 

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY


















!" 

      
 

(b)

!" 









      
 

(b)

#$%
&

      "

  !

(a)
(b)
Fig. 6.12: VNA test-set drift errors in repeated reection-coecient measurements of
1.85 mm coaxial one-port devices, taken every 1/2 hour over the period of 12 hours, along
with the 95 % uncertainty bounds predicted from the stochastic model developed based on
the measurements of the 5.4 mm long female oset-short: (a) in-phase error for a 5.4 mm
female long oset-short, (b) quadrature error for the 5.4 mm female long oset-short, (c)
in-phase error for a 5.4 mm female long oset-open, (d) quadrature error for the 5.4 mm
female long oset-open, (e) real part of the error for a matched termination, (b) imaginary
part of the error for the matched termination.

164

6.5. SUMMARY
in both cases, our model underestimates the in-phase component of the deviations, which
may be again attributed to the VNA receiver errors which are not accounted for by our
model. This component, however, is much smaller than the quadrature component, consequently, our stochastic model is capable of adequately predicting the dominant element
of the VNA test-set drift errors, as observed in the measurements of the oset short and
oset open.
The prediction of the uncertainty bounds due to the VNA test-set drift for the matched
termination is slightly worse than for other one-port standards. This may be again attributed to VNA receiver errors whose relative contribution is larger for devices with smaller
reection coecients.

6.5

Summary

We proposed a new approach to the description of the VNA nonstationarity errors. Our
approach is based on a stochastic model constructed by adding randomly varying lumped
elements to the two-port describing the VNA calibration coecients. In the case of the
connector nonrepeatability and cable instability, we describe the statistical properties of
those elements with a multivariate Gaussian PDF, while in the case of the VNA test-set
drift, we employ the vector stochastic Wiener process.
Parameters of the stochastic model are then identied based on a simple experiments
involving repeated measurements of a single highly-reective oset terminations. The
simplicity of this measurement allows us to independently characterize dierent VNA nonstationarity sources, that is connector nonrepeatability, cable instability, and the test-set
drift.
The preliminary experimental results demonstrate that our models are capable of modeling the complicated frequency-dependent characteristics of the VNA nonstationarity errors in the measurements of coaxial one-port devices by use of only a few statistically uncorrelated and frequency inpdendent random variables. We further show that our models,
once establihed, can fairly accurately predict the uncertainty bounds for VNA nonstationarity errors in the future measurements of devices with arbitrary reection coecients.
This indicates that the models we developed consistently describe the properties of a given
connector interface, cable, or VNA, rather than just accurately reproducing the measurements obtained in one experiment. Hence, the random parameters that underlie our models
are can be treated as the electrically-equivalent error mechanisms responsible for the VNA
165

6. MULTI-FREQUENCY STOCHASTIC MODELING OF VNA NONSTATIONARITY


nonstationarity errors (see Section 4.2).
The stochastic models for we developed here have already been applied to assess the
uncertainties due the VNA nonstationarity errors in the uncertainty analysis for VNA measurements [39] which accounts for the statistical correlations between VNA measurement
uncertainties at dierent frequencies. These models, once fully veried and extended to
the measurements two-port devices, can also be incorporated into the error-mechanismbased VNA calibration (see Subsection 4.5.3), in order to further increase the calibration
accuracy.

166

Chapter 7
Conclusions
Good tests kill awed theories;
we remain alive to guess again.
Karl Popper

In this work, we developed theoretical foundations and presented the most important
practical applications of the multi-frequency approach to VNA S-parameter measurements.
The principle of this approach is to account for the relationships between S-parameter
measurements at dierent frequencies. Thus, the multi-frequency approach breaks with
the traditional paradigm in VNA measurements, according to which the measurement
procedure is carried out independently at each frequency under consideration.
The multi-frequency approach stems from the observation that S-parametermeasurement errors at dierent frequencies are related to each other. These relationships
can be intuitively explained by the fact that all of the measurement errors, both at the
same and at dierent frequencies, have some common fundamental physical causes. These
causes correspond to errors in the denitions of calibration standards (e.g., tolerances of
their dimensional and material parameters) and to errors in the VNA instrumentation
(e.g., instability of the cables, misalignment of the inner and outer conductors or displacements of the inner conductor ngers in coaxial connectors, test-set drift, or receiver noise
and nonlinearities).
The simplest mathematical representation for the relationships between measurement
errors at dierent frequencies is given by the covariance matrix [39]. The diagonal terms
of this matrix describe the variances of the measurement errors while the o-diagonal
terms capture the statistical correlations between those errors, including the correlations
167

7. CONCLUSIONS
between errors at dierent frequencies. The vector underlying the covariance matrix is
comprised of the real and imaginary parts of the measurement errors at all frequencies.
Thus, we refer to this vector as the multi-frequency S-parameter-measurement error. Such
a representation generalizes the approach, employed in some early contributions on six-port
S-parameter measurement systems, and recently rediscovered for the VNA measurements
[42], in which the single-frequency S-parameter measurement errors are also characterized
with a covariance matrix.
The use of the covariance matrix for the multi-frequency S-parameter measurement
error, although seemingly intuitive, leads to some diculties. In most practical cases, the
columns and rows of this matrix are linearly dependent due to the fact that it characterizes the variability of a large number of random variables (real and imaginary parts
of S-parameters at all frequencies), dependent on a much smaller number of other independent random variables, corresponding to the fundamental causes of the measurement
errors. Consequently, the covariance matrix is rank decient and cannot be inverted, hence
the conventional form of the multivariate Gaussian PDF (see [87]) cannot be used with
the multi-frequency S-parameter-measurement error. As a result, applications where the
knowledge of the PDF is required, such as statistical VNA calibration methods or statistical procedures for the measurement-based device modeling, cannot be directly extended
to use the covariance matrix for the multi-frequency S-parameter-measurement error.
Therefore, in order to remedy this mathematical diculty, we followed a dierent approach in this work. Instead of directly employing the covariance matrix of the multifrequency S-parameter measurement error, we focused in our analysis on the fundamental
causes of the measurement errors which determine the structure of this matrix. A mathematical model for these causes is given by the notion of the physical error mechanism.
We dene a single physical error mechanism as a scalar random variable, corresponding to a physical parameter which is aected by errors, and a corresponding frequencydependent function which characterizes the relationships between this parameter and the
multi-frequency S-parameter measurement error. By applying this concept and with the
use of the generalized matrix-pseudo-inverse (see [117]), we then construct the multivariate
Gaussian PDF for the multi-frequency S-parameter-measurement error. This PDF turns
out to have an intuitive form in which the probability for given region in the domain of Sparameters is expressed in terms of the corresponding region in the domain of the physical
error mechanisms. This result is an important and unique contribution of this work.
We further analyze how to adapt the VNA measurement procedure to account for a
168

more complete description of measurement errors oered by the concepts of the physical
error mechanism and the multi-frequency S-parameter-measurement error. We focus on
the VNA calibration and on the use of the multi-frequency S-parameter-measurement error
in the uncertainty analysis and in the measurement-based device modeling.
We rst reformulate the VNA calibration problem in terms of the physical error mechanisms underlying the measurement errors. As these mechanisms contribute simultaneously
to the measurement errors at all frequencies, such a generalized VNA calibration problem
needs to be solved jointly at all measurement frequencies. We refer to this generalized
physics-based formulation of the VNA calibration problem as the error-mechanism-based
VNA calibration. This novel formulation is a unique contribution of this work and, to the
authors knowledge, has not been considered in the literature before.
Practical implementation of the error-mechanisms-based VNA calibration leads to some
diculties. It requires, on one hand, detailed modeling and characterization of all of the
physical error mechanisms responsible for the VNA measurement errors. On the other
hand, the error-mechanism-based VNA calibration results in an optimization task which
is dicult to solve due to its nonlinear and ill-posed character, and large scale. In the
main part of this work, we propose solutions to those problems in the context of VNA
calibration in the coaxial environment.
The development of an error-mechanism-based description of VNA measurement errors
requires modeling of the error mechanisms aecting the calibration standards and the VNA
itself. In the case of the coaxial transmission lines, employed as calibration standards in
this work, we model those mechanism based on a detailed analysis of possible errors in
the denitions of the standards. This analysis results in extended models of the lines that
account for the phenomena such as errors in the determination of the inner and outer
conductor length, variation of conductor diameters and conductor loss among the lines,
nonuniformity of the conductor diameters, and nonreproducibility of the center conductor
gap.
Regarding the modeling of error mechanisms aecting the VNA itself, we focus on
the VNA nonstationarity errors, that is, the connector nonrepeatability, cable instability
and test-set drift, which are the primary sources of VNA instrumentation errors. We
propose a novel approach that uses a stochastic model whose parameters are identied from
repeated S-parameter measurements. The core of this model is a generic description for the
frequency dependence of the VNA nonstationarity errors which uses a set of lumped element
perturbations, added at xed distances to the two-port describing the VNA calibration
169

7. CONCLUSIONS
coecients. The stochastic models for the VNA nonstationarity errors are then established
by allowing the parameters of the lumped elements to vary randomly.
In the case of the connector repeatability and cable instability errors, we assume that
the parameters of these elements vary according to the multivariate Gaussian PDF. This
assumption has a simple physical justication. The lumped elements model the impact
of changes in some dimensional parameters on the electrical properties of the connector
interface or the cable, and we can reasonably assume those changes to be small and to follow
the multivariate Gaussian PDF. The electrical parameters are therefore, approximately,
linear combinations of the dimensional parameters, and have also the multivariate Gaussian
PDF.
As for the test-set drift, we follow the approach used in the drift modeling of scalar
measurement instruments, and assume that it can be described as the random walk phenomenon, referred to also as the Brownian motion [147, 148]. The mathematical description
of this phenomenon in the scalar case is given by the stochastic Wiener process [158]. We
generalize this description to the multidimensional case and then apply it to model the
variability of the lumped elements in the perturbations.
We further present methods to identify parameters of the stochastic models. These
methods involve repeated reection-coecient measurements of only a single highlyreective oset termination. Preliminary experimental results for the coaxial connector
standard show that we are able to model and predict the complicated frequency dependence of the connector nonrepeatability, cable instability and test-set drift errors,
in the measurements of one-port devices, by use of only a few statistically uncorrelated
and frequency-independent random variables. Establishing the validity bounds for the
stochastic models we developed and their extension to the two-port measurements are the
topics of our ongoing research.
Our description of VNA nonstationarity errors is a signicant and unique contribution of
this work. To the authors knowledge, such as simple and yet comprehensive model for this
important source of errors in VNA measurements has not been proposed in the literature
so far. The variables used in this description are frequency independent and statistically
uncorrelated and can therefore be thought of as the electrically-equivalent error mechanisms
responsible for the VNA nonstationarity errors. Consequently, our description, once fully
established and veried, can be used in the error-mechanism-based VNA calibration to
further improve the measurement accuracy. This is the topic of our further research.
Regarding the numerical algorithm for the error-mechanism-based VNA calibration, we
170

develop it in the context of the coaxial multiline TRL calibration [36]. In our implementation, we account only for the error mechanisms aecting the calibration standards and
use a conventional description of the VNA instrumentation errors. The resulting algorithm
is generic and can easily be adopted to other redundant calibration schemes, such those
considered in [37, 38]. This novel algorithm is one of the main contributions of this work.
Two main problems we faced with the implementation of this algorithm are the large
scale of the resulting nonlinear optimization task and its ill-posed character. The large
scale of the optimization task results from the fact the VNA calibration coecients are
sought simultaneously at all measurement frequencies. Consequently, direct solution of
the error-mechanism-based VNA calibration problem is very time consuming. Hence, we
developed a robust iterative numerical approach which exploits the relationships between
the sought parameters in order to reduce the dimensionality of the optimization problem.
Our approach is based on a modied version of the classical Levenberg-Marquardt algorithm (see [159]) in which we account for the sparse structure of the Jacobians of the goal
function in a similar manner to [140].
The ill-posed character of the error-mechanism-based VNA calibration, manifesting itself with diculties in obtaining a unique solution, results from the fact that some of the
estimated parameters are related to each other. We analyze the origins of those relationships and devise a general methodology for assuring the identiability of the solution. This
methodology relies on restricting the space of possible solutions with a set of linear equality
constraints, based on some intuitive statistical properties of the physical error mechanisms.
We test the resulting multi-frequency calibration algorithm for a coaxial multiline TRL
calibration with the 1.85 mm and 7.0 mm transmission-line standards. The extensive experimental results we present indicate that by exploiting the relationships between errors
in calibration standard S-parameters at dierent frequencies, we can reduce the residual
standard uncertainty in corrected S-parameter measurements by a few times. This result clearly demonstrates the benet of the multi-frequency approach to VNA calibration.
Adaptation of the error-mechanism-based VNA calibration to other calibration schemes,
such as the oset-short calibration [79], and to measurement environments with other
waveguiding structures, such as the planar (microstrip line, coplanar waveguide) or rectangular waveguides, is the topic of our further research.
Equipped with the error-mechanism-based representation of multi-frequency S-parameter-measurement errors, we further analyze the application of such a description in the
uncertainty analysis and in the measurement-based device modeling. We rst show that
171

7. CONCLUSIONS
the statistical correlations between measurement errors at dierent frequencies are essential
when evaluating the uncertainties in calibrated time-domain measurements that employ
the VNA S-parameter measurements. Examples of such time-domain measurements are
the correction for the impedance mismatch or removing the errors due to an adapter in
oscilloscope measurements. The importance of these correlations has already been pointed
out in [44], and in this work we present a detailed derivation of the uncertainty analysis
for a generalized procedure for calibrated time-domain measurements.
We further reformulate the standard methods for measurement-based modeling of electronic devices (e.g. transistors or transmission-line discontinuities), based on the maximum
likelihood principle, to account for the relationships between S-parameter measurement errors at dierent frequencies. The key conclusion of our analysis concerns the quantication
of the mist between the measurement and the model. In standard methods, which do
not account for the statistical correlations between S-parameter measurement errors at
dierent frequencies, this mist is quantied in terms of S-parameters. We showed that
when accounting for these correlations, the mist needs to be expressed in terms of the
physical error mechanisms. This new result is an important contribution of this work that,
to the authors knowledge, has not been published before.
Summarizing, the multi-frequency approach presented in this work, by accounting
for the relationships between measurements at dierent frequencies, introduces a new
paradigm into the VNA S-parameter measurements. We show that this new paradigm
leads to a signicant improvement of the VNA S-parameter measurement accuracy and
its more complete description. We hope that the multi-frequency approach to VNA Sparameter measurements may become an answer to the new demands put on the VNA
measurements accuracy by the terahertz applications, nanotechnology, and nonliner measurements.

172

Appendix A
Real-valued representation of
complex vectors and matrices
In this appendix, we briey review the convention used in this work for a real-valued
representation of complex-valued vectors and matrices. Let the vector x CM where
x = [x1 , . . . , xM ]T .

(A.1)

This convention (see also [40, 42, 160]) expands each element of the complex vector into its
real and imaginary part, that is, the real-valued representation, denoted with underline, is
x = [Rex1 , Imx1 , Rex2 , Imx2 , . . . , RexN , ImxN ]T .

(A.2)

The inverse relationship between the two representations is readily written as

x = CM x =

1 j

x,

1 j
...

(A.3)

1 j
where the size of the matrix CM is M 2M . Let further the vector y Cn , and the matrix
A Cnm where

a11 a1M

..
.

...

(A.4)
A = ..
.

aN 1 aN M
173

A. REAL-VALUED REPRESENTATION OF COMPLEX VECTORS AND MATRICES


and y = Ax. With the use of the convention (A.2), we can now write the relationship
between the two vector as
y = A x.
(A.5)
We can easily show that

R (a11 ) R (a1M )

..
..

...
A=
.
.

R (aN 1 ) R (aN M )

(A.6)

where the operator R (x) is dened by

Rex Imx
R (x) =
.
Imx Rex

174

(A.7)

Appendix B
Maximum likelihood approach to
system identication
In this appendix, we discuss the key concepts of the application of the maximumlikelihood approach to the nonlinear system identication. Our discussion is based on
References [64, 65].

B.1

Introduction

A generic nonlinear static system can be described with a multivariate vector-valued


function
y = f (x, ) ,
(B.1)
where x RP is a vector of system inputs (excitations), y RQ is a vector of system
outputs (responses), RR is a vector of system parameters. The objective of system
identication is to determine the system parameters based on a set of measured responses
N
{yn }N
n=1 , obtained for a set of applied excitations {xn }n=1 . In the error-free case, these
measurements are described by a consistent set of equations

y1 = f (x1 , )
..
,
.

y = f (x , )
N
N

(B.2)

which, provided that N Q R, can readily be solved for by selecting any R dierent
equations out of (B.2).
175

B. MAXIMUM LIKELIHOOD APPROACH TO SYSTEM IDENTIFICATION

Fig. B.1: System with errors in responses and excitations.


In the reality, however, both system excitations and responses are aected by errors,
that is, we observe the system response with some measurement error
n =
yn + yn ,
y

(B.3)

n + xn ,

xn = x

(B.4)

under a disturbed excitation

where the vectors yn and xn are the error in the measurement of the system response
n and
yn and the error in the application of the system excitation xn , respectively, y
yn
are the measured value and the unobservable true value of the system response yn , and
xn
n are the unobservable true value and the estimated value of the system excitation
and x
xn , respectively. In the following, we refer to the errors yn and xn in short as the
measurement errors and excitation errors, respectively.
The measurements we obtain in the real case are thus described by a set of equations

1 = f x
+ x1 ,
+ y1
y
..
.


(B.5)

N = f x
N + xN ,
y
+ yN

N
where the measurement and excitation errors, {yn }N
n=1 and {xn }n=1 , respectively, are
additional unknowns, and
denotes the unknown true value of the system parameters.
N
This set of equations (B.5) cannot be uniquely solved for
, {yn }N
n=1 and {xn }n=1 ,
as the number of unknowns is larger the the number of equations in (B.5). Hence, some
additional criteria need to be used to obtain a solution. In the following, we describe the
application of the maximum-likelihood approach for solving the set (B.5).

The principal idea of the maximum-likelihood approach is to assign a metric to dierent


N
solutions of (B.5), based on the statistical properties of the errors {yn }N
n=1 and {xn }n=1 .
This metric is referred to as the likelihood function. Equations (B.5) are then solved by
selecting the solution for which the likelihood function attains the maximum.
176

B.2. FORMULATION
In the following, we rst consider dierent formulations of the maximum-likelihood
solution to the problem (B.5), that result from dierent assumptions as to the measurement
and excitation errors. Then we discuss the error analysis of the solution to (B.5) and
briey review numerical techniques used for obtaining the maximum-likelihood estimators.
Finally, we discuss the conditions for the identiability of the solutions to (B.5) and discuss
the extension of our results to the case of systems with complex-valued inputs and outputs.

B.2

Formulation

The two common cases we consider when solving the set (B.5) are the simplied case
when the system excitations are assumed to be error-free, and the general case when both
the system responses and excitations are aected by errors. In the following, we rst
discuss in detail the simplied case and then extend our results to the general case.

B.2.1

Errors in system responses

In the case when only system responses are aected by errors, the set (B.5) reduces to

1 = f x1 ,
+ y1
y
..
.


(B.6)

N = f xN ,
y
+ yN

and we are seeking the vector


and the set of vectors {yn }N
n=1 that satisfy the set
(B.6). In order to form the maximum-likelihood solution to this problem, we need to know
the statistical properties of the measurement errors {yn }N
n=1 . For each measurement
yn , these properties are captured in the probability density function fyn (yn ). For
normally distributed errors with E (yn ) = 0 and with a full-rank covariance matrix


E yn ynT = yn , this PDF is given by
(

1
|yn |
exp ynT 1
(B.7)
fyn (yn ) = (2)
yn yn .
2
Assuming further that errors for the consecutive measurements are statistically indepen

dent, that is, E yk ylT = 0 for k = l, we can write the joint PDF for the vector
Q/2

1/2

y = [y1 , . . . , yN ]T ,
177

(B.8)

B. MAXIMUM LIKELIHOOD APPROACH TO SYSTEM IDENTIFICATION


as
fy (y) =

N


N Q/2

fyn (yn ) = (2)

1/2

|y |

exp

n=1

1
yT 1
y y ,
2

(B.9)

where the covariance matrix for the joint PDF is

y1

y =

..

(B.10)

yN
Now, when formulating the maximum-likelihood solution to the system identication
problem, two concepts of errors and residuals play an important role[64]. Both concepts
refer to the variables used in the description of the system that are subject to some random
disturbances. For a given variable x, the error is dened as the dierence between the
observed value x and the true (unobservable) value
x of this variable, that is, ex = x
x.
The residual is an estimate rx of the error ex . The likelihood function is then formed by
substituting the residuals into the joint PDF of the errors.
For the measurements (B.6) performed on the system shown in Fig. B.1, under the
assumption of error- free excitations, the errors are dened by the vector (B.8). The
residuals
ry = [ry1 , . . . , ryN ]T
(B.11)
are the estimates of the errors (B.8). For a given sample {
yn }N
n=1 of system response meaN
surements and system excitations {xn }n=1 , and for a given estimate of system parameters
, the residuals need to satisfy the equations (B.6), which yields

1 = f (x1 , ) + ry1 ,
y
..
.

(B.12)

N = f (xN , ) + ryN
y

The likelihood function is now formed by substituting the residuals ry into the PDF of the
errors y, which immediately yields
N Q/2

) = (2)
L (, ry |x, y

1/2

|y |

&

'

1
exp ry T 1
y ry .
2

(B.13)

where
N ]T ,
= [
y
y1 , . . . , y
178

(B.14)

B.2. FORMULATION
and
x = [x1 , . . . , xN ]T .

(B.15)

in order to simplify the notation.


In the following we drop the dependence on x and y

and y
The maximum likelihood estimates of the system parameters and the errors,
n
are then obtained by maximizing the likelihood function (B.13) accounting for the equality
constraints (B.12), that is
arg max L(, ry ),

(B.16)

under the constraints (B.12).

(B.17)

,ry

By directly incorporating the constraints we obtain an unconstrained problem


= arg max L () ,

(B.18)

where

&

'

1
L () = (2)N Q/2 |y |1/2 exp ry ()T 1
y ry () ,
2

(B.19)

ry () = [ry1 (), . . . , ryN ()]T ,

(B.20)

with

and

1 f (x1 , )
ry1 () = y
..
.

(B.21)

N f (xN , )
ryN () = y

of the measurement error


From we can obtain the maximum likelihood estimate y
n

yn by evaluating (B.21) at , that is, yn = ryn (), for n = 1, . . . , N .


The maximum likelihood estimate (B.18) has an intuitive interpretation [64]. Let
. Then the probabildenote a region in the hyperplane around the vector of measurements y
ity obtaining the measurements in a region around the actually observed measurements
is L () . Consequently, for all possible values of , the probability of obtaining measurements within of the actual ones attains the maximum for the estimate (B.18).
Now, when determining the estimate (B.18), both numerically and analytically, it is
typically much easier to maximize the logarithm of L (), referred to as the log-likelihood
[64]. This readily leads to the same result owing to the monotonicity of the logarithm.
179

B. MAXIMUM LIKELIHOOD APPROACH TO SYSTEM IDENTIFICATION


That is,
= arg max L () = arg max ln L () .

(B.22)

In the case of (B.13), we the log-likelihood function takes on the form


N
N
1
1
1
ln L () = N Q ln (2)
ln |yn |
ry ()T 1
yn ryn ()=
2
2 n=1
2 n=1 n
1
1
1
= N Q ln (2) ln |y | ry ()T 1
y ry . (B.23)
2
2
2

Now, when solving the problem (B.18), we encounter mainly two cases: when the
covariance matrices yn of errors in the system responses are fully known and when the
matrices yn are known up to a scaling factor, that is,
yn = 2 Vyn ,

(B.24)

where the Vyn are known matrices, and the scaling factor 2 is referred to as the residual variance (square of the residual standard deviation). In the case of the fully known
covariance matrices yn , maximizing the log-likelihood function (B.23) readily leads to
the minimization of a weighted sum of squared residuals, that is,
= arg min

N

n=1

T 1
ryn ()T 1
yn ryn ()=arg min ry () y ry (),

(B.25)

Thus, in this case, the maximum likelihood estimation is equivalent to the familiar weighted
least squares problem [64, 65].
In the case of when the covariance matrices are given by (B.24), the log-likelihood
function takes on the form
N
N
1
1
1
1 
1
2
ln L(, ) = N Q ln (2)
ln |Vyn | N Q ln 2
ryn ()T Vy
r (),
n yn
2
2 n=1
2
2 n=1
2

(B.26)
and, in order to obtain the estimates of and 2 , the function (B.28) needs to be maximized
in terms of both parameters. A common approach here is the stagewise maximization of
the likelihood function, referred to as the concentrated likelihood approach [65]. In this
180

B.2. FORMULATION
approach, we rst assume that is xed and, from the condition for the stationary point
N
NQ
1 
ln L(, 2 )
1
=

+
ryn ()T Vy
r () = 0,
2
n yn
2
2
2 2
2 ( ) n=1

we obtain the estimate


2 () as
N
1 
1

() =
ryn ()T Vy
r ().
n yn
N Q n=1
2

(B.27)

Now, after inserting this estimate into the original likelihood function (B.26), we obtain
N
1
1
ln L(,
2 ()) = N Q[ln (2) + 1 ln (N Q)]
ln |Vyn |+
2
2 n=1
N

1
1
N Q ln
ryn ()T Vy
r (), (B.28)
n yn
2
n=1

from which we obtain the estimate of as


2 ()).
= arg max ln L(,

(B.29)

Due to the monotonicity of the logarithm, this estimate (B.29) can be equivalently determined as

= arg min

where

N

n=1

1
1
ryn ()T Vy
r ()=arg min ry ()T Vy
ry ().
n yn

Vy =

Vy1

(B.30)

...

(B.31)

VyN
The estimate inserted into (B.27) yields eventually the estimate of the residual variance
2
=

M
2 ()
LE =

N
1 
T V1 ryn ()
= 1 ry ()
T V1 ry ().

ryn ()
yn
y
N Q n=1
NQ

(B.32)

This estimate can be shown to be biased [64]. The unbiased estimate of the residual
181

B. MAXIMUM LIKELIHOOD APPROACH TO SYSTEM IDENTIFICATION


variance is given by [64]

2 =

NQ

2 ,
N Q R M LE

(B.33)

where R is the number of elements in the vector .

B.2.2

Errors in system responses and excitations

When both system responses and excitations are subject to disturbances, the solution
to (B.5) is constituted by not only the estimates of system parameters and the errors
N
{yn }N
n=1 in the system responses, but also by the estimates of the errors {xn }n=1 in
the system excitations . The resulting problem and its variants are known in the literature
under dierent names, such as: errors-in-independent-variables regression (EIV) [64, 65],
orthogonal-distance (ODR) regression [140], total least-squares (TLS) regression [161], full
least-squares regression [162], or generalized-distance regression [75]. In this work we follow
the nomenclature of [64, 65].
In order to write the likelihood function for the problem (B.5), similarly to the problem
N
(B.6), we need to know the statistical properties of the errors {yn }N
n=1 and {xn }n=1 .
Regarding the errors in system responses we follow the assumptions made in Section B.2.1.
As for the errors {xn }N
n=1 in the system excitations, for each excitation we describe
statistical properties of these errors with the probability density function fxn (xn ). For
normally distributed errors with E (xn ) = 0 and with a full-rank covariance matrix


E xn xnT = xn this PDF can be written as
(

1
fxn (xn ) = (2)
|xn |
exp xnT 1
(B.34)
xn xn .
2
Assuming further that these errors for the consecutive measurements are statistically in

dependent, that is E xk xlT = 0 for k = l, we can write the joint PDF for the vector
P/2

1/2

x = [x1 , . . . , xN ]T

(B.35)

as
fx (x) = (2)N Q/2

N

n=1

N
1
xnT 1
xn xn =
2 n=1
(
)
1
= (2)N Q/2 |x |1/2 exp xT 1
x
, (B.36)
x
2

|xn |1/2 exp

182

B.2. FORMULATION
where the covariance matrix for the joint PDF is

x =

x1

..

(B.37)

xN
We also assume that the errors in the excitations and errors in the responses are statistically


independent, that is that is E xk ylT = 0 for k, l = 1, . . . , N .

Now, for the measurements (B.5) are dened by the vectors (B.8) and (B.36). As those
errors are statistically independent, their joint PDF is given by
f (x, y) = fx (x) fy (y) .

(B.38)

rx = [rx1 , . . . , rxN ]T

(B.39)

ry = [ry1 , . . . , ryN ]T

(B.40)

The residuals

and
are the estimates of the errors x and y, respectively. For a given sample {
yn }N
n=1 of
N
system response measurements and system excitations {
xn }n=1 , and for a given estimate
of system parameters , the residuals need to satisfy the equations (B.5), which yields

1 = f (
x + rx1 , ) + ry1
y
..
.

(B.41)

N = f (
y
xN + rxN , ) + ryN

The likelihood function is now formed by inserting the residuals into the joint PDF (B.38)
of the errors. This yields
) =
x, y
L (, rx , ry |
N Q/2N P/2

= (2)

1/2

|y |

1/2

|x |

&

'

1
1 T 1
exp ry T 1
y ry rx x rx , (B.42)
2
2

where
N ]T ,
= [
y
y1 , . . . , y
183

(B.43)

B. MAXIMUM LIKELIHOOD APPROACH TO SYSTEM IDENTIFICATION


and
= [
N ]T .
x
x1 , . . . , x

(B.44)

and y
, in order to simplify the notation.
In the following, we drop the dependence on x
and the errors in reThe maximum likelihood estimates of the system parameters ,
and x
n , respectively, is then obtained by maximizing the
sponses and excitations, y
n
likelihood function (B.42) while accounting for the equality constraints (B.41), that is,
arg max L(, ry , rx ),

(B.45)

under the constraints (B.41).

(B.46)

,ry ,rx

We can further directly incorporate the constraints into (B.45), which yields an equivalent
problem

=
= arg max L(),

(B.47)

x
where


=
,
rx

(B.48)

and

L () = (2)N Q/2N P/2 |y |1/2 |x |1/2


(

exp
with

1
1 T 1
ry ()T 1
y ry () rx x rx , (B.49)
2
2

1 f (
ry1 () = y
x1 + rx1 , )
..
.

(B.50)

N f (
xN + rxN , )
ryN () = y

is obtained by evaluating
The maximum likelihood estimate of measurement errors y
n

(B.50) at , that is, yn = ryn (), for n = 1, . . . , N .


Similarly to the problem (B.6), we can show that with fully known covariance matrices
of errors in the system excitations and measurements, (B.48) is equivalent to the weighted
184

B.2. FORMULATION
least squares problem, that is,
= arg min

N

n=1

ryn ()T 1
yn ryn () +

N

n=1

rTxn 1
xn rxn =

T 1
= arg min ry ()T 1
y ry ()+rx x rx , (B.51)

Another case we often consider in practice is when the covariance matrices are known up
to a scaling factor, that is,
yn = 2 Vyn , and xn = 2 Vxn ,

(B.52)

whereVxn and Vyn , for n = 1, . . . , N are known matrices, and 2 is the unknown residual
variance. In this case, we can also perform a stagewise optimization of the log-likelihood
function to obtain (see [64])
= arg min

N

n=1

1
ryn ()T Vy
r () +
n yn

N

n=1

rTxn 1
xn rxn =

1
ry () + rTx 1
= arg min ry ()T Vy
x rx , (B.53)

and
2

M
LE =

& N
'
N

1 
1
T
T V1 ryn ()
+

ryn ()

r
r
xn xn xn =
yn
N n n=1
n=1
&
'
T 1
1
T 1

=
ry () Vy ry () + rx x rx . (B.54)
Nn

Similarly to (B.32), the estimator (B.54) is biased. The unbiased estimator of the residual
variance is given by [64]
NQ
2

M
.
(B.55)

2 =
N Q N P R LE

We shall now make one additional remark about the formulation (B.51) and (B.53).
Both problems can be equivalently written in a form similar to (B.25) and (B.30), respectively. That is, in the case of (B.51), we can write
= arg min r ()T 1 r () ,

185

(B.56)

B. MAXIMUM LIKELIHOOD APPROACH TO SYSTEM IDENTIFICATION


where

ry ()
r () =
,
rx
and the matrix is given by

(B.57)

y
=

(B.58)

In the case of the formulation (B.53), we replace in (B.56) with

Vy
V=

Vx

(B.59)

and in the development


This formulation is helpful in the error analysis of the estimate
of the numerical methods for solving (B.51) and (B.53).

B.3

Covariance matrix of the estimates

n of the
Once we have the estimate of the system parameters (and the estimates x
disturbances in the system excitations), we are often interested in the uncertainty of these
estimates. These uncertainties are determined in the residual analysis and are obtained
based on sensitivities of the maximum-likelihood solution to changes in system responses
and excitations.

B.3.1

Errors in system responses

We rst analyze the problem (B.25) when the covariance matrices of errors in system
responses are fully known, and the extend our results to the case (B.30) when the covariance
matrices are given by (B.24).
In order to determine the sensitivity of the solution (B.25), we consider the perturbed
+
due to the errors in system responses y with E(y ) = 0 and with a
solution
n
n

T
known covariance matrix E(yn yn ) = yn , for n = 1, . . . N . That is,
+
= arg min

N

n=1

[
yn + yn f (xn , )] 1
yn + yn f (xn , )] .
yn [
T

(B.60)

By expanding the function (B.1) in the Taylor series around the unperturbed estimate
186

B.3. COVARIANCE MATRIX OF THE ESTIMATES


and neglecting the nonliner terms, we obtain
+
= arg min

N 


T

+ y + J()(

ryn ()
)
n

n=1

where

1
yn ryn () + yn + J()( ) , (B.61)

f (xM , ) 
n = ryn () = f (x1 , ) 

J()

,

T
T =
T
=

(B.62)

is the Jacobian matrix of the residuals. Now, expanding (B.61), removing the constant
terms and accounting for the fact that from the stationarity of the unperturbed solution
T ryn ()=0,

we have J()
we readily obtain a linear least squares problem
= arg min

N 


yn + Jn ()

T

n=1

1
yn yn + Jn () ,

(B.63)

T 1 y ,
J()
y

(B.64)

which has the solution


where = ,


1

T 1 J()

= J()

y
where

1
J()
y1

..
..

, y =
J()
.
.


N
J()
yN

and y =

y1

...

(B.65)

yN

we obtain
Now, writing the covariance matrix of ,

T ) =
= E(


1

T 1 J()

= J()
y

1

1

T 1 E(y yT )1 J()


J()
T 1 J()

J()
y
y
y

T 1 J()

= J()
y

=
. (B.66)

In the case when the covariance matrices are known up to the scaling factor, we replace

187

B. MAXIMUM LIKELIHOOD APPROACH TO SYSTEM IDENTIFICATION


y in (B.66) with the estimate

y =
2

Vy1

...

=
2 Vy ,

(B.67)

VyN
which yields

1


T ) =
T V1 J()

= E(
2 J()
y

B.3.2

(B.68)

Errors in system responses and excitations

In order to determine the covariance matrix of the estimates , we use the formulation
(B.56) and follow the similar reasoning as above. For the problem (B.51) we then obtain


1


T ) = J()
T 1 J()

= E(

(B.69)

while for the problem (B.53) we have




1


T) =
T V1 J()

2 J()
= E(
with

1
J ()

..

=
J()

where

and

(B.70)

1
Jx ()

..

(B.71)

N Jx ()
N

J ()

0
IP


rn () 
n = rn () 


,
J
(
)
=
J ()
x
n
xT =
T =

...

and V =

V1

...
VN

and IP is the identity matrix of the size P P .


188

(B.72)

(B.73)

B.4. NUMERICAL SOLUTION TECHNIQUES

B.4

Numerical solution techniques

Numerical techniques for solving the problems (B.25), (B.30) and (B.51), (B.53) are in
general iterative. We rst discuss classical algorithms for solution of the problems (B.25)
and (B.30), and then briey review the possible strategies for solving (B.51) and (B.53).

B.4.1

Errors in system responses

The principle of the Gauss-Newton and Levenberg-Marquardt algorithm is to solve


(B.25) in a sequence of quadratic approximations. In the q-th iteration of this sequence,
(q) . From a quadratic approximation of the goal function around
the solution estimate is
this solution estimate, we obtain a correction (q) which is then used to update the
solution estimate with the formula
(q) .
(q) +
(q+1) =

(B.74)

This procedure is repeated until some convergence criteria are met. The two algorithms
dier in the way the correction (q) is computed.

A. Gauss-Newton algorithm. In the Gauss-Newton algorithm, the quadratic approximation of the goal function in (B.25) is obtained by approximating (B.1) with its rst-order
Taylor series expansion, that is,
(q)

f (xn , ) f (xn , (q) ) + Jf (xn )( (q) ),


(q)

where the matrix Jf

(B.75)

is the Jacobian of the function at (q) , that is,




(q)
Jf (xn )

f (xn , ) 

.
=
T =(q)

(B.76)

The correction (q) can then be obtained from a classical linear weighted least squares
problem, that is,
(q) = arg min

N 


(q) n=1

(q)

(q)
r(q)
yn Jf (xn )

189

T

(q)
(q)
1
,
yn ryn Jf (xn )
(q)

(B.77)

B. MAXIMUM LIKELIHOOD APPROACH TO SYSTEM IDENTIFICATION


where
n f (xn , (q) ), for n = 1, . . . , N.
r(q)
yn = y

(B.78)

Solution to (B.77) can then be obtained from the normal equations


(q)
(q)
(q)
= X(q)T 1
X(q)T 1
y X
y ry

(B.79)

and

r(q)
y

r(q)

y1
y1
..

...
= . , y =

r(q)
yN

and X(q)

(q)

J (x1 )
f

..

.
=
.

yN

(q)

(B.80)

Jf (xN )

Solution to the normal equations (B.79) can be written as




(q)
(q) = X(q)T 1
y X

1

(q)
X(q)T 1
y ry ,

(B.81)

however, equation (B.81) is not recommended to compute the correction (q) due to
poor numerical properties. The singular value decomposition (SVD) method (see [121]) is
typically used to solve the normal equations (B.79).
B. Levenberg-Marquardt algorithm. The problem with the Gauss-Newton algorithm
is that the updated solution computed from (B.74) and (B.82) may lay outside the region
in which the approximation (B.75) holds. The Levenberg-Marquardt algorithm (see [159])
addresses this problem by introducing an adaptively controlled damping in the computation
of the step (q) . This damping is implemented through a following modication of the
normal equations


(q)
(q)
X(q)T 1
X
+
A
(q) = X(q)T 1
(B.82)
y
y ry ,
(q)
(q)
where is a damping factor and matrix A contains the main diagonal of X(q)T 1
y X .
The factor may be chosen according to dierent strategies (see [64, 65, 139, 159]). The
general principle here is to start with a small number, such as = 103 . We then
compute the step (q) and check if the goal function at (q+1) has decreased. If not,
then the damping factor is increased (say by 10) and the step is recomputed. If yes,
then the new solution (q+1) is accepted and the damping factor is decreased (say by
10). In the recomputing of the step (q) for dierent values of the damping factor ,
the Sherman-Morrison-Woodbury identity for matrix inversion (see [121]) may be used to

190

B.5. SOLUTION UNIQUENESS


speed up the computations.

B.4.2

Errors in system responses and excitations

The numerical solution to the problems (B.51) and (B.53) is, in general, obtained by
transforming these problem to the formulation (B.56) and then by applying the methods
described in Subsection B.4.1. The particular structure of the Jacobian matrix (B.76), as
given by (B.71), can be exploited to reduce the dimensionality of the optimization problem
[140].

B.5

Solution uniqueness

When solving the system identication problems, we are often interested in verifying
whether the estimate of the system parameters we obtained is unique. We can show that
for the solution to class of problem described in Section B.2.1 and Section B.2.2 to be
unique, the Jacobian matrices (B.76) determined at the solution need to be full rank.
Indeed, consider the normal equations (B.79) written at the solution, that is,
T 1
XT 1
y X = X y ry ,

(B.83)

where we dropped the iteration index to signify that the residuals (B.78) and the Jacobian
At the solution, we have = 0 and consequently
(B.76) are evaluated at the solution .
XT 1
y ry = 0. Now, if the matrix X is not full rank, the dimension of its null space is
larger than one. Consequently, from the denition of the null space, there exist an innite
number of vectors z such that z = 0 and Xz = 0. For a vector z null (X), we then
readily have
T 1
XT 1
(B.84)
y Xz = X y ry = 0,
+ x is also a solution. As a result, instead of a unique solution, we have a
hence =
and the null space of X.
space of solutions dened by

B.6

Systems with complex-valued inputs and outputs

In practice, we often encounter the case when the system has complex-valued inputs
and/or outputs. In this case, we can easily adopt the methods described in this appendix
191

B. MAXIMUM LIKELIHOOD APPROACH TO SYSTEM IDENTIFICATION


by representing the complex-valued vectors as real-valued vectors, comprised of real and
imaginary parts of relevent complex quantities. Typical convention for such a representation is reviewed in the Appendix A.
In the case, however, when the complex random variables characterizing the errors
have the circular-normal PDF, we can use a simpler approach [62, 163, 164]. A vector of
complex random variables x is said to have a circular PDF if [62]


E (x E (x)) (x E (x))T = 0.

(B.85)

A complex random variable with such a property is also sometimes referred to as a proper
complex random variable [163]. Expanding (B.85), we can easily show that it is equivalent
to the following conditions
Cov (Re x, Re x) = Cov (Im x, Im x) ,

(B.86)

Cov (Re x, Im x) = Cov (Re x, Im x)T .

(B.87)

In the case of a scalar random variable x, these conditions imply Var (Re x) = Var (Im x) =
x2 and Cov (Re x, Im x) = 0, that is,

2
x = x

x2

(B.88)

where the underline denotes the convention (A.2) for real-valued representation of complex
random variables.
Now, let Cov (Re x, Re x) = Cov (Im x, Im x) = RR and Cov (Re x, Im x) = RI . We
can then show that


E (x E (x)) (x E (x))H = 2RR + j2RI = x ,

(B.89)

where the superscript H denotes the conjugate transpose (Hermitian transpose). Consequently, the statistical properties the vector of complex random variables with a circular
PDF are characterized with a single complex matrix and can be obtained in the analogous
fashion to the real-valued case. We can further show that
1
x = x ,
2
192

(B.90)

B.6. SYSTEMS WITH COMPLEX-VALUED INPUTS AND OUTPUTS


where the underline denotes the convention (A.6) for the real valued representation of
complex matrices.
We assume now that the vector x has a normal PDF with a non-singular covariance
matrix x , fullling the conditions (B.86) and (B.87). We can then show that its PDF
can be written with the use of (B.89) as [62]
fx (x) = fx (x) =

1
N

(2) |x |

e(x)x

(x)

(B.91)

This property greatly simplies the formulation and solution of least-squares problems
involving vectors of complex normal random variables, as the PDF given by B.91 has a
similar form to the PDF for real-valued normally distributed vectors. We can show that
for a system with complex inputs and outputs the estimates of system parameters can
be obtained from equations given in Subsection B.2.1 and Subsection B.2.2, by employing
the complex covariance matrices B.89 and replacing the transpose T with the conjugate
transpose H.

193

B. MAXIMUM LIKELIHOOD APPROACH TO SYSTEM IDENTIFICATION

194

Appendix C
Air-dielectric coaxial transmission
line
In this appendix we summarize characteristics of the TEM mode in the coaxial transmission lines with air dielectric. We rst consider the case of line without conductor losses
and then discuss the case with small conductor losses.

C.1

Innite metal conductivity

Let the diameters of the inner and outer conductor of the line d and D, respectively. For
both conductors we assume the innite conductivity. Let the center of the inner conductor
be oset from the center of the outer conductor by e. Characteristic impedance of the
TEM mode in the line is given by [137]

Z00

d2 + D2 4e2 +

ln
=
2

where

(D2 d2 + 4e2 )2 (4De)2


,
2dD

0
,
r 0

(C.1)

(C.2)

and 0 = 4107 H/m and 0 8.854 107 F/m1 are the magnetic permeability and dielectric permittivity of vacuum, respectively, and r is the relative dielectric permittivity of
-1
,
Dielectric permittivity of vacuum is dened as 0 = 0 c2
, where c = 299792458 m/s is the speed
of light in vacuum.
1

195

C. AIR-DIELECTRIC COAXIAL TRANSMISSION LINE


the air. The propagation constant is imaginary and is given by

= j0 = j r 0 0 = j/v,

(C.3)

c
v= ,
r

(C.4)

and the phase velocity v is

where c = 1/ 0 0 is speed of light in vacuum.

For concentric conductors, that is, when e = 0, equation C.1 reduces to [165]
Z00 =
For slightly eccentric lines, that is, when
approximated as [131]


Z00

C.2

ln .
2
d

2e
Dd

4e2
ln

1 2
2
d
D d2

(C.5)

0 , characteristic impedance C.1 can be

"

"

4e2
D

.
ln 2
2
d
D d2

(C.6)

Finite metal conductivity

The characteristic impedance and propagation constant of the quasi-TEM mode in a


coaxial transmission line with small conductor losses are given by [131, 165]


Z0 = Z00 1 + (1 j)

&

'

vRc
vRc
Z00 1 + (1 j)
,
Z00
2Z00

(C.7)

and


= j0

&

'

vRc
vRc
1 + (1 j)
j0 1 + (1 j)
= + j (0 + ) ,
Z00
2Z00

with
= 0

vRc
Rc
=
,
2Z00
2Z00

(C.8)

(C.9)

where Rc is the conductor resistance per-unit-length. For concentric conductors, this resistance is given by
(
)
1
1
1
+
,
(C.10)
Rc =
D d
196

C.2. FINITE METAL CONDUCTIVITY


and is the metal conductivity and is the skin depth given by


=
which lead to
Z0 Z00
and
j0

2
,
0

&

v
1 + (1 j)
2Z00

&

v
1 + (1 j)
2Z00

(C.11)
(

0
2

0
2

1
1
+
D d

1
1
+
D d

)'

(C.12)

(C.13)

)'

In the case of eccentric conductors, the surface resistance is given by [137]


1
Rc =

where
d = *

"

1
1
+
,
D D d d

(C.14)

D2 d2 4e2
D2 (d + 2e)2

and
D = *



D2 (d 2e)2

,

(C.15)

.

(C.16)

D2 d2 + 4e2
2

(D + 2e)

d2



(D 2e)

d2

2e
For small values of eccentricity, that is, when Dd
0, these expressions can be approximated as
2d2 e2
d 1 +
,
(C.17)
(D2 d2 )2

and
D 1 +

2D2 e2
.
(D2 d2 )2

197

(C.18)

C. AIR-DIELECTRIC COAXIAL TRANSMISSION LINE

198

Appendix D
Center-conductor gap impedance

D.1

Innite metal conductivity

A simplied schematic of the center conductor gap is shown in Fig. D.1. For gap
widths g much smaller the depth of the gap
(d dp ) /2, electrical properties of the center conductor gap can be modeled as that
of a radial waveguide stub [92, 136]. For the
fundamental TEM mode in this waveguide,
Fig. D.1: Cross-section of the center conducthe admittance transformation is given by
tor gap (not in scale).
[48]
j + y(r0 )(0 r, 0 r0 )ct(0 r, 0 r0 )
y(r) =
,
(D.1)
Ct(0 r, 0 r0 ) + jy(r0 )(0 r, 0 r0 )
where
Z0 (r) =

g
1
=
,
Y0 (r)
2r

(D.2)

is the characteristic impedance at r, y(r) and y(r0 ) are the normalized input admittances
at r and r0 , where r > r0 , 0 is the propagation constant of the air dened by (C.3), is
199

D. CENTER-CONDUCTOR GAP IMPEDANCE


the wave impedance of the air given by (C.2) and [48]
J1 (x)N0 (y) N1 (x)J0 (y)
,
J0 (x)N0 (y) N0 (x)J0 (y)
J1 (y)N0 (x) N1 (y)J0 (x)
Ct(x, y) =
,
J1 (x)N1 (y) N1 (x)J1 (y)
J0 (x)N0 (y) N0 (x)J0 (y)
(x, y) =
,
J1 (x)N1 (y) N1 (x)J1 (y)
ct(x, y) =

(D.3)
(D.4)
(D.5)

where Jm (t) is the Bessel function of the rst kind of order m, and Nm (t) is the Bessel
function of the second kind (also referred to as the Neumann function [166]) of order m.
In the case shown in Fig. D.1, this stub transforms a short circuit at r = dp /2, that is,
y(dp /2) = , to some admittance y(d/2) at r = d/2. Inserting y(dp /2) = into (D.17),
we readily obtain
Y (d/2)
= jct(0 d/2, 0 dp /2).
y(d/2) =
(D.6)
Y0 (d/2)
For 0 d/2  1 and 0 dp /2  1, we can approximate (D.21) with the use of approximate
forms of Bessel functions for t  1 [1]
J0 (t) 1,
t
J1 (t) ,
2
2
N0 (t) ln t,

21
N1 (t) ,
t
to obtain

1
Y (d/2)
j
,
Y0 (d/2)
0 d/2 ln ddp

(D.7)
(D.8)
(D.9)
(D.10)

(D.11)

which, after inserting (D.2), gives the gap impedance


Z(d/2) j

g
d
d
0
0 d/2 ln
= j g ln .
d
dp
2
dp

200

(D.12)

D.2. FINITE METAL CONDUCTIVITY

D.2

Finite metal conductivity

For conductors with small losses, the propagation constant and characteristic
impedance become complex and can be approximated as
j (0 + ) + ,
and

(D.13)


1 + (1 j)
,
Z0 (r) =
2r
0

where
=

Rc
,
2Z0

(D.14)

(D.15)

and Rc is the conductor resistance per unit length. For the radial waveguide, this resistance
depends on r and can be easily calculated as
1
1 1
=
=
Rc = Rs
r
r

0 1
,
2 r

(D.16)

where Rs is the conductor surface resistance, is the metal conductivity, and is the skin
depth given by (C.11). Inserting (D.16) into (D.15) gives
*

0 1
.
2 g

In order to account for the losses in the admittance transformation, we need to solve
the Telegraphic equations for the radial line, accounting for the complex character of the
propagation constant. This solution can be shown to have the form
y(r) =

1 + y(r0 )h(0 r, 0 r0 )cth(0 r, 0 r0 )


,
Cth(0 r, 0 r0 ) + y(r0 )h(0 r, 0 r0 )

(D.17)

where
I1 (x)K 0 (y) + K 1 (x)I0 (y)
,
I0 (y)K0 (x) K0 (y)I0 (x)
I1 (y)K0 (x) + K1 (y)I0 (x)
Cth(x, y) =
,
I1 (y)K1 (x) K1 (y)I1 (x)
I0 (y)K0 (x) K0 (y)I0 (x)
h(x, y) =
,
I1 (y)K1 (x) K1 (y)I1 (x)
cth(x, y) =

201

(D.18)
(D.19)
(D.20)

D. CENTER-CONDUCTOR GAP IMPEDANCE


and Im (t) is the modied Bessel function of the rst kind of order m, and Km (t) is the
modied Bessel function of the second kind of order m. Inserting y(dp /2) = into (D.17),
we readily obtain
Y (d/2)
= cth(0 d/2, 0 dp /2).
(D.21)
y(d/2) =
Y0 (d/2)
For 0 d/2  1 and 0 dp /2  1, we can approximate (D.21) with the use of approximate
forms of the modied Bessel functions for |z|  1 [1] as
I0 (z) 1,
z
I1 (z) ,
2
K0 (z) ln z,
1
K1 (z) ,
z

(D.22)
(D.23)
(D.24)
(D.25)

After inserting the above equations into (D.21), we obtain the rst order approximation
0
d
Z(d/2) j g ln
+ (1 + j)
2
dp

0 1
d
ln .
2 dp

(D.26)

We note the surface impedance in (D.26) does not depend on the gap width. This is due
to the fact that this impedance results from the skin-depth eect in the gap walls. Indeed,
the total surface impedance of the gap walls can be calculated as
d/2

Zs = 2
dp /2

D.3

Rs
Rs
d
dr =
ln
= (1 + j)
2r

dp

0 1
d
ln .
2 dp

(D.27)

Finger eect

The ngers of the connector socket are formed by in-cuts along the center conductor.
The in-cuts are very narrow, that is, we have w  d, where w is the in-cut width and
d is the diameter of the inner conductor. Due to the manufacturing method, the in-cuts
have the same width at both outer and inner side of the ngers. We can approximately
substitute the in-cuts with magnetic walls and treat the center conductor gap as a set of
N sectorial radial waveguides, with the angle of each waveguide given approximately by
/2

1 d N w
1
w
1
=

,
d/2 2N
N
d
N
202

D.3. FINGER EFFECT


where N is the number of ngers. Each of the waveguides supports a T EM mode with
the characteristic impedance given by
Z0 (r) =

g
2
1
1
= =
Z0 (r) Z0 (r),
Y0 (r)
r

(D.28)

and impedance transformation given by (D.21), with Y0 determined from (D.28). In order
to determine the diameters that need to be used in (D.21), we dene the eective radius
for a given arc length L and angle as
ref f =

L
.

(D.29)

For the outer diameter, we have then


def f
(d N w) /N
d Nw
=
=
,
2
2/N
2
2

(D.30)

and similarly for the pin diameter


dp N w
dp,ef f
=

.
2
2
2

(D.31)

After inserting those eective diameters into (D.32), we obtain


0
d N w
Z(d/2) j g ln
+ (1 + j)
2
dp N w

0 1
d N w
ln
.
2 dp N w

A similar expression for the eect of ngers is given in [136].

203

(D.32)

D. CENTER-CONDUCTOR GAP IMPEDANCE

204

Appendix E
Slightly nonuniform coaxial
transmission line

Consider a nonuniform transmission line with length l and characterized by the characteristic impedance Z0 (x) and complex propagation constant (x), for x [0, l]. In order
to determine S-parameters of such a line, we dene the port one at x = 0 with reference impedance Zref 1 = Z(0), and the port two at x = l and with reference impedance
Zref 2 = Z0 (l). The rst order solution to S-parameters (accounting for the power normalization introduced in Chapter 2) of such a line is [167]
S11 =

l

N (x) e

S22 =

(x )dx

dx,

l

x

N (x) e

x

(E.1)

 (x )dx

dx,

(E.2)

and

l

S21 = e

S12 = e

l
0

205

(x)dx

(E.3)

(x)dx

(E.4)

E. SLIGHTLY NONUNIFORM COAXIAL TRANSMISSION LINE


where N  (x) = N (l x),  (x) = (l x), and
N (x) =

1 1 dZ0 (x)
.
2 Z0 (x) dx

(E.5)

Writing the characteristic impedance as


Z0 (x) = Z0 + Z0 (x),

(E.6)

and assuming that |Z0 (x) |  |Z0 |, we can write approximately




1 d Z0 (x)
1 1 dZ0 (x)
=
N (x)
.
2 Z0
dx
2 dx
Z0

(E.7)

Writing further the propagation constant as


(x) = + (x),
we obtain

l

(E.8)

l

(x)dx = l +
0

(x)dx,

(E.9)

which gives us an approximation for the transmission coecients

S21 = S12 = el 1

l

(x)dx .

(E.10)

Inserting further (E.7) in (E.1), we obtain an approximation

S11 =

l

2x2

N (x) e

x

(x )dx

dx

l

N (x) 1 2

l

N (x) e2x dx + 2

x

(x )dx e2x dx

l

N
0


(x) (x )dx e2x dx.

(E.11)

For a slightly nonuniform coaxial transmission line, we can write with the use of (5.21)
206

and (5.22)


1 d Z00 (x)
N (x)
,
2 dx
Z00


Rc (x) Z00 (x)
(x) (1 + j)

.
Rc
Z00

(E.12)
(E.13)

Inserting the last expression into the transmission coecient equation, we readily obtain

S21 = S12 el 1 (1 + j)

l 
0

Rc (x) Z00 (x)

dx ,
Rc
Z00

(E.14)

which leads to a phase shift of the transmission coecients.


As to the reection coecient the situation is more complicated. The rst integral can
be simplied after integrating by parts to the form

l


Z00 (x) 2x
1 Z00 (0) 1 Z00 (0) 2l
dx =

e
+
e
dx.
2 Z00
2 Z00
Z00
l

2x

N (x) e
0

(E.15)

Regarding the second integral, after integrating by parts, we obtain approximately


l

N
0


(x) (x )dx e2x dx

l
0

Z00 (x)
2x
2x


(x)e

e
(x
)dx
dx,

Z00
0

which to rst order can be neglected. Hence we obtain the approximation for the reection
coecient

Z00 (x) 2x
1 Z00 (0) 1 Z00 (L) 2l

+
e
+
e
dx.
2 Z00
2 Z00
Z00
l

S11

(E.16)

In a similar manner, we can derive



Z00 (x) 2x
1 Z00 (l) 1 Z00 (0) 2l
2l

+
e
+ e
e dx.
2 Z00
2 Z00
Z00
l

S22

(E.17)

We further change the reference impedance at both ports to Z0 , which nally leads to
207

E. SLIGHTLY NONUNIFORM COAXIAL TRANSMISSION LINE


approximations
S11

l
0

S22 e

2l

Z00 (x) 2x
e
dx,
Z00
l
0

Z00 (x) 2x
e dx,
Z00

(E.18)

(E.19)

while S21 to rst order remains the same and is determined by (E.14).
We can now show that setting Z00 (x) = Z00 and Rc (x) = Rc reduces (E.18),
(E.19), and (E.14) to

Z00 
1 e2l ,
Z00

!
" 
R
Z
c
00
= el 1 (1 + j)

l ,
Rc
Z00

S11 = S22 =

S21 = S12

(E.20)
(E.21)

which agrees with the description (5.32) of a slightly mismatched uniform transmission
line.
If we now assume that the line is lossless, that is, = j0 , expressions (E.16), (E.17),
and (E.14) reduce to
S21 = S12 ej0 l ,
(E.22)
and
S11


l 
1  D (x) d (x) j20 x
e
dx,

j0 D
D
d
ln d

(E.23)

j20 l

S22 j0 e


l 
1  D (x) d (x) j20 x
e
dx.

D
d
ln Dd
0

Similar results are reported in [168].

208

(E.24)

Appendix F
Small changes of two-ports
scattering parameters

In this appendix, we give a detailed derivation of the model (6.2). Consider a calibrated
one-port VNA, with an error-box described by a transmission matrix T. Due to some random uctuations, the VNA error-box changes to a new transmission matrix T . We assume
that the change is due to a small perturbation that has occurred at some location inside
the VNA (see Fig. 6.1). We describe the perturbation with a transmission matrix Tn
(scattering matrix Sn ), where n is the index of the perturbation. For small perturbations
Cn , Ln , Rn , and Z0,n (see Fig. 6.2), we may write approximately

1 T,n
R1,n
R1,n
1 + T,n
Tn
, and Sn
,
R2,n 1 + T,n
1 + T,n
R2,n

(F.1)

where R1,n , R2,n , and T,n are small numbers dened as


1 
(Z Yn n ) ,
2 n
1
R1,n = (Z n Yn + n ) ,
2
1
T,n = (Zn + Yn ) ,
2
R1,n =

209

(F.2)
(F.3)
(F.4)

F. SMALL CHANGES OF TWO-PORTS SCATTERING PARAMETERS


with
Zn

RDC,n + RRF 0,n

f
f0

+ jLn

Zref


RDC,n


RRF
0,n

f
+ jLn ,
f0

Yn = jCn Zref = jCn ,


Z0,n

n =
= Z0,n
.
Zref

(F.5)
(F.6)
(F.7)

In order to express T with Tn , we split the error box T into two parts. The part
between the raw measurement plane and the plane, where the perturbation is located, is
described with a matrix T1 . The remaining part of the error-box is represented by a matrix
T2 . Thus we can write
T = T1 T2 , and T = T1 Tn T2 .
(F.8)
We then represent the error in corrected VNA S-parameter measurements due to the perturbation Tn by cascading a transmission matrix Te,n with the error-box transmission
matrix T. Thus we can write
Te,n = T1 T = (T1 T2 )1 T1 Tn T2 = T1
2 Tn T2 .

(F.9)

In order to expand this last expression, we assume that the error-box can be seen as a set
of electrically lumped discontinuities connected with low-loss transmission-line sections.
Hence, we can write approximately
T2 =

N


T2,n ,

(F.10)

n=1

where

T2,n

(1 T,n ) ej0 ln
R1,n
,

R2,n
(1 + T,n ) ej0 ln

(F.11)

and R1,n , R2,n , and T,n are small numbers, j0 is the propagation constants, ln is the
physical length of the n-th section, and N is the number of sections. By inserting (F.10)
and (F.11) in (F.9) and considering only rst-order terms, we obtain

Te,n
where dn =

+n

n=1 ln

1 T,n
R1,i ej0 dn

,
R2,n ej0 dn 1 + T,n

(F.12)

is the distance of the perturbation from the reference plane. Finally,


210

since we assumed that the perturbation are small, we can neglect multiple reections and
write the overall error as the linear combination

Te

N
+
n=0

N
+
n=0

T,n

R2,n ej0 dn

N
+
n=0

j0 dn

R1,n e

1+

N
+
n=0

T,n

(F.13)

where N + 1 is the number of perturbations. We further expand (F.13) by inserting


equations (F.2) through (F.7), which, after conversion to scattering parameters, yields the
expressions (6.2) through (6.11).

211

F. SMALL CHANGES OF TWO-PORTS SCATTERING PARAMETERS

212

Appendix G
Estimation of VNA nonstationarity
model parameters

We rewrite (6.24) in a more compact form as


= arg min r ()H V1 r () ,

where

r1 (d, p1 )

..

,
r () =
.

(G.1)

(G.2)

rM (d, pM )
and

V=

...

(G.3)

V
contains M copies of the matrix V on its diagonal. We solve the nonlinear optimization
problem (G.1) with the use of the iterative numerical Levenberg-Marquardt algorithm. In
a single iteration, we determine a new estimate of sought parameters from the previous
estimate 0 as = 0 + , where the correction is obtained as a solution of the
normal equations


JH V1 J + D = JH V1 r ( 0 ) ,
(G.4)
213

G. ESTIMATION OF VNA NONSTATIONARITY MODEL PARAMETERS


where the Jacobian matrix J is dened as


r () 

,
J=
T =0

(G.5)

D is a xed positive denite diagonal matrix (typically the matrix constructed out of the
main diagonal of JH V1 J [64]), and is chosen according to the strategy presented in [159].
In the following, we rst consider the case when = 0 (the Gauss-Newton algorithm). We
then show how to derive the solution to (G.4) with = 0 based on the solution for = 0.

The Jacobian J can be written as a block matrix

X (d0 )

1 (d0 )

..

,
.

...

J=

(G.6)

X (d0 ) M (d0 )
where X(d0 ) is given by (6.21) in the case of the connector interface model, and by (6.34)
in the case of the test-set drift model, and


(X (d) pm ) 

.
m (d0 ) =

dT
d=d0

(G.7)

Taking into account (6.21) and (6.15), we can show that

x1 (d0 , 1 )T

..

= j20 X (d0 ) ,

m (d0 ) = j20
.

where

xk (d0 , k )


xK
(d0 , K )

(G.8)

T

ej20 d1 wR1 (fk ) ej20 d1 2k wR2 (fk )


..
.

T

ej20 dN wR1 (fk ) ej20 dN 2k wR2 (fk )

pR,1

(G.9)

pR,N

and k is given by k in the case of the connector interface model, and by 1k in the case
of the test-set drift model.

With the use of (G.6), we can rewrite the normal equations (G.4) for = 0 as a set of
214

equations (we omit the dependence on d0 for the sake of notational simplicity)
XH V1 Xp1 + XH V1 1 d = XH V1 r1
..
.
H

X
and

M


V1 XpM

1
H
m V Xpm

+X

! M


m=1

V1 M d

=X

"
1
H
m V m

d =

m=1

(G.10)

V1 rM
M


1
H
m V rm .

(G.11)

m=1

Solution to (G.10) and (G.11) can be obtained in two steps. We rst assume that d is
xed, and write the solutions to (G.10) as


1

pm = XH V1 X

XH V1 (rm m d) ,

(G.12)

for m = 1, . . . , M . These solutions are the inserted into (G.11), which yields
! M


"

H
m Hm

d =

m=1

M


H
m Hrm ,

(G.13)

m=1

where

1

H = V1 V1 X XH V1 X

XH V1 .

(G.14)

Consequently, instead of solving the original optimization problem (G.4), we solve the
problem (G.13) which has smaller dimensionality.
In order to account for = 0, a simple modication need only be made to (G.10) and
(G.11). Writing the matrix D as

D=

DX

...
DX

(G.15)

D


where DX = diag XH V1 X , and D = diag


(G.13) and (G.14) as
!

D +

M


+

M
m=1

"

H
m H m

m=1

d =

M

m=1

215

1
H
m V m . We can the rewrite


H
m H rm ,

(G.16)

G. ESTIMATION OF VNA NONSTATIONARITY MODEL PARAMETERS


and

H = V1 V1 X XH V1 X + DX

216

1

XH V1 .

(G.17)

Appendix H
Vector stochastic Wiener process
Consider an ensemble of N independent scalar stochastic Wiener processes
w (t) = [w1 (t) , . . . , wN (t)]T

(H.1)

Each of these processes is characterized with the parameter n > 0, for n = 1, . . . , N .


Thus we can write
w(t) N (0, Dt)
(H.2)
and
w(t2 ) w(t1 ) N (0, D(t2 t1 ))
where

D=

(H.3)

...

(H.4)

N
We now dene a general N -dimensional Wiener process, that is, a process whose underlying one-dimensional processes can be statistically correlated. Let A be a full rank
N N matrix, and let (t) be a new process dened as
(t) = Aw(t).

(H.5)

Since the transformation is linear, (t) is also normally distributed. We calculate the rst
217

H. VECTOR STOCHASTIC WIENER PROCESS


two moments of this distribution as
E [(t)] = E [Aw(t)] = AE [w(t)] = 0,

(H.6)

and

E (t)(t)T = E Aw(t) (Aw(t))T = AE w(t)w(t)T AT = tADAT = t, (H.7)


hence
(t) N (0, t).

(H.8)

Using (H.8), we can readily show that

(t2 ) (t1 ) N (0, (t2 t1 )) .

(H.9)

Matrix = t = tADAT is the covariance matrix of the new process. We can readily
show that it is a valid covariance matrix, that is, a symmetric and positive semi-denite
matrix. Indeed,


T = t ADAT

T

= t AT

T

DT AT = tADAT = ,

(H.10)

and since for every x such that |x| > 0 we have


xT tDx 0,

(H.11)

we also see that




xT tx = xT tADAT x = AT x


T

tD AT x = yT tDy 0,

(H.12)

since |y| = AT x 0 for |x| 0.


In order to estimate the parameter , we use the maximum likelihood estimation
theory. For a given realization y(t) for t = 0 < t1 < . . . < tN of the process (t), we form
the dierences yn = y(tn ) y(tn1 ). It can be then easily shown that the maximum
218

of the process parameters has form


likelihood estimator
N

yn ynT
= 1
.

N n=1 tn tn1

(H.13)

is not biased. Indeed, for each n we have


Estimator
yn
N (0, ),
tn tn1
hence

N

yn ynT
= 1
E
E ()
N n1
tn tn1

219

N
1 
= .
N n1

(H.14)

(H.15)

H. VECTOR STOCHASTIC WIENER PROCESS

220

Bibliography
[1] D. M. Kerns and R. W. Beatty, Basic theory of waveguide junctions and introductory
microwave network analysis. Pergamon Press, 1967.
[2] A. Fung, L. Samoska, G. Chattopadhyay, T. Gaier, P. Kangaslahti, D. Pukala,
C. Oleson, A. Denning, and Y. Lau, Two-port vector network analyzer measurements up to 508 GHz, IEEE Trans. Instrum. Meas., vol. 57, no. 6, pp. 11661170,
June 2008.
[3] PNA Millimeter-Wave Network Analyzers Technical Overview, Agilent Technologies,
2008. [Online]. Available: http://www.agilent.com
[4] VectorStar MS4640A Series VNA Brochure, Anritsu. [Online]. Available: www.
anritsu.com
[5] R&S ZVA Vector Network Analyzer, Rohde & Schwarz, 2008. [Online]. Available:
www.rohde-schwarz.com
[6] D. K. Rytting, Network analyzer error models and calibration methods, 1998, white
Paper.
[7] D. Rytting, Network analyzers - 50 years on, in 71th ARTFG Conf. Digest, 2008.
[8] A. Rumiantsev and N. Ridler, VNA calibration, IEEE Microw. Mag., vol. 9, no. 3,
pp. 8699, 2008.
[9] G. Raybon and P. J. Winzer, 100 Gb/s challenges and solutions, in Proc. Conference on Optical Fiber communication/National Fiber Optic Engineers Conference
OFC/NFOEC 2008, 2428 Feb. 2008, pp. 135.
[10] J. Wenger, Automotive radar - status and perspectives, in Proc. IEEE Compound
Semiconductor Integrated Circuit Symposium CSIC 05, 30 Oct.2 Nov. 2005, p. 4pp.
221

BIBLIOGRAPHY
[11] M. Skolnik, Ed., Radar handbook, 3rd ed. McGraw-Hill, 2008.
[12] R. T. Schilizzi, P. E. F. Dewdney, and T. J. W. Lazio, The square kilometre array,
L. M. Stepp and R. Gilmozzi, Eds., vol. 7012, no. 1. SPIE, 2008, p. 70121I.
[13] Atacama Large Millimeter/submillimeter Array. [Online]. Available:
//www.almaobservatory.org

http:

[14] P. de Maagt, Terahertz technology for space and earth applications, in Proc. International Workshop on Antenna Technology: Small and Smart Antennas Metamaterials and Applications IWAT 07, 2123 March 2007, pp. 111115.
[15] T. W. Crowe, W. L. Bishop, D. W. Portereld, J. L. Hesler, and I. Weikle, R. M.,
Opening the terahertz window with integrated diode circuits, IEEE J. Solid-State
Circuits, vol. 40, no. 10, pp. 21042110, Oct. 2005.
[16] D. W. Portereld, High-eciency terahertz frequency triplers, in Proc.
IEEE/MTT-S International Microwave Symposium, 38 June 2007, pp. 337340.
[17] C. Oleson and A. Denning, Millimeter wave vector analysis calibration and measurement problems caused by common waveguide irregularities, in 56th ARFTG
Conference Digest, vol. 38, 2000, pp. 19.
[18] L. B. Lok, S. Singh, A. Wilson, and K. Elgaid, Impact of waveguide aperture dimensions and misalignment on the calibrated performance of a network analyzer from
140 to 325 GHz, in 73th ARFTG Conference Digest, Jun. 1212, 2009, pp. 14.
[19] D. K. Rytting, Network analyzer accuracy overview, in 58th ARTFG Conf. Digest,
vol. 40, November 2001, pp. 113.
[20] Z. Liu and R. Weikle, A reectometer calibration method resistant to waveguide
ange misalignment, IEEE Trans. Microw. Theory Tech., vol. 54, no. 6, pp. 2447
2452, June 2006.
[21] P. Kabos et al., High frequency metrology at nanoscale: calibration, model validations and device characterization, in IEEE MTT-S 2009 Workshop, 2009.
[22] M. Randus and K. Homann, A simple method for extreme impedances measurement, in Proc. 71nd ARFTG Microwave Measurement Symposium, 2008.
[23] A. Lewandowski, D. LeGolvan, R. A. Ginley, T. M. Wallis, A. Imtiaz, and P. Kabos,
Wideband measurement of extreme impedances with a multistate reectometer, in
222

BIBLIOGRAPHY
Proc. 72nd ARFTG Microwave Measurement Symposium, 912 Dec. 2008, pp. 4549.
[24] J. Verspecht, Large-signal network analysis, IEEE Microw. Mag., vol. 6, no. 4, pp.
8292, Dec 2005.
[25] W. Van Moer and L. Gomm, NVNA versus LSNA: enemies or friends? IEEE
Microw. Mag., vol. 11, no. 4, pp. 97103, Dec 2010.
[26] J. Verspecht, D. F. Williams, D. Schreurs, K. A. Remley, and M. D. McKinley,
Linearization of large-signal scattering functions, IEEE Trans. Microw. Theory
Tech., vol. 53, no. 4, pp. 13691376, April 2005.
[27] Verspecht and D. E. Root, Polyharmonic distortion modeling, IEEE Microw. Mag.,
vol. 7, no. 3, pp. 4457, June 2006.
[28] F. Verbeyst and V. Bossche, VIOMAP, the S-parameter equivalent for weakly nonlinear RF and microwave devices, IEEE Trans. Microw. Theory Tech., vol. 42,
no. 12, pp. 25312535, 1994.
[29] G. Engen and C. Hoer, Thru-reect-line: An improved technique for calibrating
the dual six-port automatic network analyzer, IEEE Trans. Microw. Theory Tech.,
vol. 27, no. 12, pp. 987993, 1979.
[30] H.-J. Eul and B. Schiek, A generalized theory and new calibration procedures for
network analyzer self-calibration, IEEE Trans. Microw. Theory Tech., vol. 39, no. 4,
pp. 724731, April 1991.
[31] I. Kasa, Closed-form mathematical solutions to some network analyzer calibration
equations, Instrumentation and Measurement, IEEE Transactions on, vol. 23, no. 4,
pp. 399 402, Dec. 1974.
[32] G. F. Engen, Calibrating the six-port reectometer by means of sliding terminations, IEEE Trans. Microw. Theory Tech., vol. 26, no. 12, pp. 951957, Dec 1978.
[33] W. Wiatr, A method for embedding network characterization with application to
low-loss measurements, IEEE Trans. Instrum. Meas., vol. 36, pp. 487490, June
1987.
[34] B. Bianco and M. Parodi, Measurement of the eective relative permittivities of
microstrip, Electronics Letters, vol. 11, no. 3, pp. 7172, February 6 1975.
[35] L. C. Oldeld, J. P. Ide, and E. J. Grin, A multistate reectometer, IEEE Trans.
223

BIBLIOGRAPHY
Instrum. Meas., vol. 34, no. 2, pp. 198201, June 1985.
[36] R. Marks, A multiline method of network analyzer calibration, IEEE Trans. Microw. Theory Tech., vol. 39, no. 7, pp. 12051215, July 1991.
[37] G. Vandersteen, Y. Rolain, J. Schoukens, and A. Verschueren, An improved slidingload calibration procedure using a semiparametric circle-tting procedure, IEEE
Trans. Microw. Theory Tech., vol. 45, no. 7, pp. 10271033, July 1997.
[38] W. Wiatr and A. Lewandowski, Multiple reect technique for wideband one-port
VNA calibration, in Proc. of 16th Int. Conf. on Microwaves, Radar and Wireless
Communications, 2224 May 2006, pp. 3740.
[39] A. Lewandowski, D. F. Williams, P. D. Hale, J. C. M. Wang, and A. Dinestfrey,
Covariance-matrix vector-network-analyzer uncertainty analysis for time and frequency domain measurements, accepted for publication in the IEEE Trans. Microw.
Theory Tech.
[40] R. M. Judish and G. F. Engen, On-line accuracy assessment for the dual six-port
ANA: Statistical methods for random errors, IEEE Trans. Instrum. Meas., vol. 36,
pp. 507513, 1987.
[41] C. A. Hoer, On-line accuracy assessment for the dual six-port ANA: Treatment of
systematic errors, IEEE Trans. Instrum. Meas., vol. 36, pp. 514523, 1987.
[42] N. M. Ridler and M. J. Salter, An approach to the treatment of uncertainty in
complex S-parameter measurements, Metrologia, vol. 39, no. 3, pp. 295302, 2002.
[43] , Propagating S-parameter uncertainties to other measurement quantities, in
58th ARTFG Conf. Digest, vol. 40, 2001, pp. 119.
[44] D. F. Williams, A. Lewandowski, T. S. Clement, J. C. M. Wang, P. D. Hale, J. M.
Morgan, D. A. Keenan, and A. Dienstfrey, Covariance-based uncertainty analysis of
the NIST electrooptic sampling system, IEEE Trans. Microw. Theory Tech., vol. 54,
no. 1, pp. 481491, 2006.
[45] R. B. Marks and D. F. Williams, A general waveguide circuit theory, National
Institute of Standards and Technology Journal of Research, vol. 97, pp. 533562,
Oct. 1992.
[46] W. K. Gwarek and M. Celuch-Marcysiak, Wide-band s-parameter extraction from
224

BIBLIOGRAPHY
FD-TD simulations for propagating and evanescent modes in inhomogeneous guides,
IEEE Trans. Microw. Theory Tech., vol. 51, no. 8, pp. 19201928, Aug. 2003.
[47] R. Collin, Foundations of microwave engineering. Wiley-IEEE Press, 2000.
[48] N. Marcuwitz, Waveguide handbook. New York: McGraw-Hill, 1951.
[49] A. A. Barybin, Modal expansions and orthogonal complements in the theory of
complex media waveguide excitation by external sources for isotropic, anisotropic,
and bianisotropic media, Progress in Electromagnetic Research, vol. 19, pp. 241
300, 1998.
[50] E. W. Matthews, The use of scattering matrices in microwave circuits, IRE Transactions on Microwave Theory and Techniques, vol. 3, no. 3, pp. 2126, April 1955.
[51] H. Carlin, The scattering matrix in network theory, IRE Transactions on Circuit
Theory, vol. 3, no. 2, pp. 8897, Jun 1956.
[52] K. Kurokawa, Power waves and the scattering matrix, IEEE Trans. Microw. Theory Tech., vol. 13, no. 2, pp. 194202, Mar 1965.
[53] Exploring the architectures of network analyzers, Agilent Technologies, Application
note AN 1287-2, 2000.
[54] R. A. Speciale, A generalization of the TSD network-analyzer calibration procedure, covering n-port scattering-parameter measurements, aected by leakage errors, IEEE Trans. Microw. Theory Tech., vol. 25, no. 12, pp. 11001115, Dec 1977.
[55] J. V. Butler, D. K. Rytting, M. F. Iskander, R. D. Pollard, and M. Vanden Bossche,
16-term error model and calibration procedure for on-wafer network analysis measurements, IEEE Trans. Microw. Theory Tech., vol. 39, no. 12, pp. 22112217, Dec
1991.
[56] H. Van Hamme and M. Vanden Bossche, Flexible vector network analyzer calibration with accuracy bounds using an 8-term or a 16-term error correction model,
IEEE Trans. Microw. Theory Tech., vol. 42, no. 6, pp. 976987, June 1994.
[57] J. Brewer, Kronecker products and matrix calculus in system theory, IEEE Trans.
Circuits Syst., vol. 25, no. 9, pp. 772781, Sep 1978.
[58] R. B. Marks, Formulations of the basic vector network analyzer error model including switch-terms, in 50th ARFTG Conference Digest, vol. 32, 1997, pp. 115126.
225

BIBLIOGRAPHY
[59] D. Williams, J. Wang, and U. Arz, An optimal vector-network-analyzer calibration
algorithm, IEEE Trans. Microw. Theory Tech., vol. 51, no. 12, pp. 23912401, Dec.
2003.
[60] G. F. Engen, Microwave circuit theory and foundations of microwave metrology. London: Peter Peregrinus Ltd., 1992.
[61] L. Ljung, System identication: theory for the user.
River, New Jersey, 1999.

Prentice Hall, Upper Saddle

[62] R. Pintelon and J. Schoukens, System identication: a frequency domain approach.


IEEE Press, 2001.
[63] R. Aster, B. Borchers, and C. Thurber, Parameter Estimation and Inverse Problems.
Academic Press, 2005.
[64] Y. Bard, Nonlinear parameter estimation. Academic Press, New York, 1974.
[65] G. A. F. Seber and C. J. Wild, Nonlinear regression. Wiley-Interscience, 2003.
[66] W. Kruppa and K. F. Sodomsky, An explicit solution for the scattering parameters
of a linear two-port measured with an imperfect test set (correspondence), IEEE
Trans. Microw. Theory Tech., vol. 19, no. 1, pp. 122123, Jan 1971.
[67] B. Bianco, M. Parodi, S. Ridella, and F. Selvaggi, Launcher and microstrip characterization, IEEE Trans. Instrum. Meas., vol. 25, no. 3, pp. 320323, February 6
1976.
[68] K. J. Silvonen, A general approach to network analyzer calibration, IEEE Trans.
Microw. Theory Tech., vol. 40, no. 4, pp. 754759, April 1992.
[69] A. Ferrero and U. Pisani, Two-port network analyzer calibration using an unknown
thru, IEEE Microw. Guided Wave Lett., vol. 2, no. 12, pp. 505507, Dec. 1992.
[70] S. M. Kay, Fundamentals of statistical processing: Estimation theory. Prentice Hall,
Upper Saddle River, New Jersey, 1993.
[71] D. Kajfez, Numerical determination of two-port parameters from measured unrestricted data, IEEE Trans. Instrum. Meas., vol. 24, no. 1, pp. 411, March 1975.
[72] I. Kasa, A circle-tting procedure and its error analysis, IEEE Trans. Instrum.
Meas., vol. 25, pp. 814, 1976.
226

BIBLIOGRAPHY
[73] C. A. Hoer, R. M. Judish, J. R. Juroshek, and G. F. Engen, Theory, uncertainty
analysis, and statistical control for the NIST 2-18 GHz dual 6-port automatic network
analyzer, NIST, Tech. Rep., 1980, NIST Special Publication 250 [unpublished].
[74] W. Wiatr, A broadband technique for one-port VNA calibration and characterization of low-loss two-ports, in Proc. Conference on Precision Electromagnetic Measurements Digest, 610 July 1998, pp. 432433.
[75] M. J. Salter, N. M. Ridler, and P. M. Harris, Over-determined calibration schemes
for RF network analysers employing generalised distance regression, in Proc. 62nd
ARFTG Microwave Measurements Conference Fall 2003, 45 Dec. 2003, pp. 127142.
[76] D. Blackham, Application of weighted least squares to OSL vector error correction,
in Proc. 61st ARFTG Conference Digest Spring 2003, 13 June 2003, pp. 1121.
[77] K. Wong, Uncertainty analysis of the weighted least squares VNA calibration,
ARFTG Microwave Measurements Conference, Fall 2004. 64th, pp. 2331, Dec. 2004.
[78] W. Wiatr, Statistical VNA calibration technique using thru and multiple reect
terminations, in Proc. of 17th Int. Conf. on Microwaves, Radar and Wireless Communications, May 1921, 2008, pp. 12.
[79] G. J. Scalzi, J. Slobodnik, A. J., and G. A. Roberts, Network analyzer calibration
using oset shorts, IEEE Trans. Microw. Theory Tech., vol. 36, no. 6, pp. 1097
1100, June 1988.
[80] F. Purroy and L. Pradell, New theoretical analysis of the LRRM calibration technique for vector network analyzers, IEEE Trans. Instrum. Meas., vol. 50, no. 5, pp.
13071314, Oct. 2001.
[81] R. B. Marks and D. F. Williams, Characteristic impedance determination using
propagation constant measurement, IEEE Microw. Guided Wave Lett., vol. 1, no. 6,
pp. 141143, June 1991.
[82] D. F. Williams and R. B. Marks, Transmission line capacitance measurement,
IEEE Microw. Guided Wave Lett., vol. 1, no. 9, pp. 243245, Sept. 1991.
[83] J. R. Juroshek and G. M. Free, Measurements of the characteristic impedance of
coaxial air line standards, IEEE Trans. Microw. Theory Tech., vol. 42, no. 2, pp.
186191, Feb. 1994.

227

BIBLIOGRAPHY
[84] Guide to the Expression of Uncertainty in Measurement, International Organization
for Standarization, Genvea, 1993.
[85] B. N. Taylor and C. E. Kuyatt, Guidelines for evaluating and expressing the uncertainty of NIST measurement results, Tech. Rep., 1994, nIST Tech. Note 1297.
[86] J. Martens, On quantifying the eects of receiver linearity on VNA calibrations,
70th ARTFG Conf. Digest, pp. 6265, 2007.
[87] D. F. Morrison, Multivariate statistical methods. McGraw-Hill, 1967.
[88] R. Willink and B. D. Hall, A classical method for uncertainty analysis with multidimensional data, Metrologia, vol. 39, no. 4, pp. 361369, 2002.
[89] D. F. Williams, C. F. Wang, and U. Arz, In-phase/quadrature covariance-matrix
representation of the uncertainty of vectors and complex numbers, 68th ARTFG
Conf. Digest, pp. 6265, 2006.
[90] K. H. Wong, Characterization of calibration standards by physical measurements,
in 39th ARTFG Conf. Digest, vol. 21, June 1992, pp. 5362.
[91] J. Juroshek, A study of measurements of connector repeatability using highly reecting loads (short paper), IEEE Trans. Microw. Theory Tech., vol. 35, no. 4, pp.
457460, 1987.
[92] W. C. Daywitt, A simple technique for investigating defects in coaxial connectors,
IEEE Trans. Microw. Theory Tech., vol. 35, no. 4, pp. 460464, 1987.
[93] P. R. Young, Analysing connector repeatability of microwave vector measurements,
in Proc. IEE Colloquium on Interconnections from DC to Microwaves (Ref. No.
1999/019), Feb. 18, 1999, pp. 8/18/5.
[94] D. F. Williams, R. B. Marks, and A. Davidson, Comparison of on-wafer calibrations, in Proc. the ARFTG Conference Digest, R. B. Marks, Ed., vol. 20, 1991, pp.
6881.
[95] Guidelines on the evaluation of Vector Network Analyzers (VNA), July 2007, Doc.
EURAMET/cg-12/v.01.
[96] Y. Rolain, W. Van Moer, and D. DeGroot, A rst step towards a wave-based
stochastic calibration for multi-port vectorial network analyzers, in Proc. 63rd
ARFTG Conference Digest Spring, June 11, 2004, pp. 151156.
228

BIBLIOGRAPHY
[97] B. C. Belanger, Traceability: An evolving concept. ASTM Standarization News,
vol. 22, no. 8, 1980.
[98] International vocabulary of metrology Basic and general concepts and associated
terms, 3rd ed., International Organization for Standarization, Genvea, 2008, JCGM
200:2008.
[99] F. L. Warner, Microwave attenuation measurement. Peter Peregrinus Ltd., 1977.
[100] J. Stenarson and K. Yhland, A new assessment method for the residual errors in
SOLT and SOLR calibrated VNAs, in 69th ARTFG Conf. Digest, Broomeld, CO,
June 2007.
[101] P. Persson, An algortihm for the evaluation of the residual directivity ripple trace,
ANAMET, Tech. Rep., May 2002, ANAMET Report 034.
[102] G. Wbbeler, C. Elster, T. Reichel, and R. Judaschke, Determination of complex
residual error parameters of a calibrated vector network analyser, in 69th ARTFG
Conf. Digest, 2007.
[103] D. F. Williams, A. Lewandowski, D. LeGolvan, and R. Ginley, Electronic vectornetwork-analyzer verication, Microwave, vol. 10, Oct 2009, [to be published].
[104] M. Wojnowski, M. Engl, V. Issakov, G. Sommer, and R. Weigel, Accurate broadband RLCG-parameter extraction with TRL calibration, in Proc. IEEE MTT-S
International Microwave Symposium Digest, 1520 June 2008, pp. 4146.
[105] B. Donecker, Determining the measurement accuracy of the HP 8510 microwave
network analyzer, in Proc. rd ARFTG ARFTG Conference Digest-Spring, vol. 5,
June 1984, pp. 5184.
[106] R. F. Kaiser and D. F. Williams, Sources of error in coplanar-waveguide TRL calibrations, in Proc. 54th ARFTG Conference Digest-Spring, vol. 36, Dec. 2000, pp.
16.
[107] R. D. Pollard, Verication of system specications of a high performance network
analyzer, in Proc. 23rd ARFTG ARFTG Conference Digest-Spring, vol. 5, June
1984, pp. 3850.
[108] U. Stumper, Inuence of TMSO calibration standards uncertainties on VNA Sparameter measurements, IEEE Trans. Instrum. Meas., vol. 52, no. 2, pp. 311315,
229

BIBLIOGRAPHY
April 2003.
[109] , Uncertainty of VNA S-parameter measurement due to nonideal TRL calibration items, IEEE Trans. Instrum. Meas., vol. 54, no. 2, pp. 676679, Apr 2005.
[110] J. P. Homann, P. Leuchtmann, and R. Vahldieck, Computing uncertainties of Sparameters by means of Monte Carlo simulation, in 69th ARTFG Conf. Digest, June
2007.
[111] J. Leinhos and U. Arz, Monte-Carlo analysis of measurement uncertainties for onwafer thru-reect-line calibrations, in Proc. 2008 IEEE MTT-S International Microwave Symposium Digest, 1520 June 2008, pp. 3336.
[112] B. Bianco, A. Corana, S. Ridella, and C. Simicich, Evaluation of errors in calibration
procedures for measurements of reection coecient, IEEE Trans. Instrum. Meas.,
vol. 27, no. 4, pp. 354358, Dec. 1978.
[113] U. Stumper, Uncertainties of VNA S-parameter measurements applying the TAN
self-calibration method, IEEE Trans. Instrum. Meas., vol. 56, no. 2, pp. 597600,
April 2007.
[114] J. Juroshek, C. M. Wang, and G. P. McCabe, Statistical analysis of network analyzer
measurements, International Journal of Metrology, vol. 35, pp. 2633, MayJune
1998.
[115] G. A. Engen, On-line accuracy assessment for the dual six-port ANA: Background
and theory, IEEE Trans. Instrum. Meas., vol. 36, pp. 501506, 1987.
[116] Evaluation of measurement data Supplement 1 to the Guide to the expression of
uncertainty in measurement Propagation of distributions using a Monte Carlo
method, 1st ed., Joint Committe for Guides in Metrology, 2008.
[117] C. R. Rao and S. K. Mitra, Generalized inverse of a matrix and its applications, in
Proc. Sixth Berkeley Symp. on Math. Statist. and Prob., vol. 1, 1972, pp. 601620.
[118] K. S. Miller, Multidimensional Gaussian Distributions. John Wiley & Sons, 1964.
[119] T. W. Andreson, An introduction to multivariate statistical analysis, 3rd ed.
Wiley & Sons, 2003.

John

[120] C. R. Rao, A note on a generalized inverse of a matrix with applications to problems in mathematical statistics, Journal of the Royal Statistical Society. Series B
230

BIBLIOGRAPHY
(Methodological), vol. 24, no. 1, pp. 152158, 1962.
[121] G. H. Golub and C. F. Van Loan, Matrix computations.
versity Press, 1996.

The Johns Hopkins Uni-

[122] K.-S. Kwong and B. Iglewicz, On singular multivariate normal distribution and its
applications, Computational Statistics & Data Analysis, vol. 22, no. 3, pp. 271285,
July 1996.
[123] A. Genz and K. Koon-Shlng, Numerical evaluation of singular multivariate normal
distributions, Journal of Statistical Computation and Simulation, vol. 68, pp. 121,
2000.
[124] B. B. Szendrenyi, Eects of pin depth in LPC 3.5 mm, 2.4 mm, and 1.0 mm connectors, in Proc. IEEE MTT-S International Microwave Symposium Digest, vol. 3,
1116 June 2000, pp. 18591862.
[125] J. Miall and K. Lees, Modeling the repeatability of Type-N connectors using Microwave Studio, in Proc. 19th ANAMET Meeting, 2003.
[126] J. P. Homann, P. Leuchtmann, and R. Vahldieck, Pin gap investigations for the
1.85 mm coaxial connector, in Proc. European Microwave Conference, 912 Oct.
2007, pp. 388391.
[127] T. S. Clement, P. D. Hale, D. F. Williams, C. M. Wang, A. Dienstfrey, and D. A.
Keenan, Calibration of sampling oscilloscopes with high-speed photodiodes, IEEE
Trans. Microw. Theory Tech., vol. 54, no. 8, pp. 31733181, Aug. 2006.
[128] D. F. Williams, T. S. Clement, P. D. Hale, and A. Dienstfrey, Terminology for
high-speed sampling-oscilloscope calibration, in 68th ARTFG Conf. Digest, 2006.
[129] A. V. Oppenheim and R. W. Schafer, Digital signal processing. Prentice Hall, 1975.
[130] C. A. Hoer and G. F. Engen, Calibrating a dual six-port or four-port for measuring
two-ports with any connectors, in Proc. MTT-S International Microwave Symposium Digest, vol. 86, no. 1, 2 Jun 1986, pp. 665668.
[131] N. Sladek, Fundamental considerations in the design and application of high precision coaxial connectors, in Proc. IRE International Convention Record, vol. 13,
Mar 1965, pp. 182189.
[132] W. Sigg and J. Simon, Reectometer calibration using load, short and oset shorts
231

BIBLIOGRAPHY
with unknown phase, Electronics Letters, vol. 27, no. 18, pp. 16501651, 29 Aug.
1991.
[133] K. Howell and K. Wong, DC to 110 GHz measurements in coax using the 1 mm
connector, Microwave Journal, vol. 42, pp. 2234, 1999.
[134] B. Oldeld, The connector interface and its eect on calibration accuracy, Microwave Journal, pp. 106114, March 1996.
[135] IEEE standard for precision coaxial connectors (DC to 110 GHz), IEEE Std 2872007 (Revision of IEEE Std 287-1968), pp. C1119, 21 2007.
[136] T. E. MacKenzie and A. E. Sanderson, Some fundamental design principles for the
development of precision coaxial standards and components, IEEE Trans. Microw.
Theory Tech., vol. 14, no. 1, pp. 2939, Jan 1966.
[137] P. Leuchtmann and J. Rufenacht, On the calculation of the electrical properties of
precision coaxial lines, IEEE Trans. Instrum. Meas., vol. 53, no. 2, pp. 392397,
April 2004.
[138] M. Horibe, M. Shida, and K. Komiyama, S-parameters of standard airlines whose
connector is tightened with specied torque, IEEE Trans. Instrum. Meas., vol. 56,
no. 2, pp. 401405, April 2007.
[139] R. Fletcher, Practical methods of optimization. Wiley, 2008.
[140] P. T. Boggs, R. H. Byrd, and R. B. Schnabel, A stable and ecient algorithm for
nonlinear orthogonal distance regression, SIAM J. Sci. Stat. Comput., vol. 8, no. 6,
pp. 10521078, 1987.
[141] A. Lewandowski and W. Wiatr, Correction for line-length errors and centerconductor-gap variation in the coxial multiline through-reect-line calibration, in
74th ARTFG Conf. Digest, 2009.
[142] MATLAB 2009b Users Reference, The Mathworks. [Online]. Available:
//www.mathworks.com

http:

[143] National Institute of Standards and Technology De-embedding Software Program


MultiCAL, 1995, Revision 1.00.
[144] D. DeGroot, J. Jargon, and R. Marks, Multiline TRL revealed, in Proc. 60th
ARFTG Conference Digest Fall 2002, 2002, pp. 131155.
232

BIBLIOGRAPHY
[145] Maury Microwave 2553K Calibration KitUsers Guide.
[146] Maury Microwave 8850C Calibration KitUsers Guide.
[147] B. Stuckman, C. Perttunen, J. Usher, and B. McLaughlin, Stochastic modeling of
calibration drift in electrical meters, Instrumentation and Measurement Technology
Conference, 1991. IMTC-91. Conference Record., 8th IEEE, pp. 530536, 14-16 May
1991.
[148] A. Bobbio, P. Tavella, A. Montefusco, and S. Costamagna, Monitoring the calibration status of a measuring instrument by a stochastic model, IEEE Trans. Instrum.
Meas., vol. 46, no. 4, pp. 747751, Aug. 1997.
[149] K.-H. Lin and B.-D. Liu, A gray system modeling approach to the prediction of
calibration intervals, IEEE Trans. Instrum. Meas., vol. 54, no. 1, pp. 297304, Feb.
2005.
[150] J. Furrer, Die Tcke steckt im Stecker, metINFO, vol. 10, pp. 1016, 2003.
[151] Y.-S. Lee, Re-visiting repeatability issues of Type-N connectors, in 68th ARFTG
Conference Digest-Fall, 2007.
[152] A. Lewandowski and D. F. Williams, Characterization and modeling of random
vector-network-analyzer measurement errors, in Proc. of 17th Int. Conf. on Microwaves, Radar and Wireless Communications, Wroclaw, Poland, May 19-21 2008.
[153] J. Furrer, Type-N and APC-7 connector repeatability: experiences during power
sensor calibration, in Proc. 17th ANAMET Meeting, 2002.
[154] A. Lewandowski and D. F. Williams, Stochastic modeling of coaxial-connector repeatability errors, in 74th ARTFG Conf. Digest, 2009.
[155] E. Vonesh and V. M. Chinchilli, Linear and nonlinear models for the analysis of
repeated measurements. Marcel Dekker, Inc., 1997.
[156] S. Hines, M. E. and H. E. Stinehelfer, Time-domain oscillographic microwave network analysis using frequency-domain data, IEEE Trans. Microw. Theory Tech.,
vol. 22, no. 3, pp. 276282, Mar. 1974.
[157] B. Ulriksson, A time domain reectometer using a semiautomatic network analyzer
and the fast fourier transform, IEEE Trans. Microw. Theory Tech., vol. 29, no. 2,
pp. 172174, Feb. 1981.
233

BIBLIOGRAPHY
[158] A. Papoulis and S. U. Pillai, Probability, random variables, and stochastic processes,
4th ed. McGraw-Hill, 2001.
[159] D. W. Marquardt, An algorithm for least-squares estimation of nonlinear parameters, Journal of the Society for Industrial and Applied Mathematics, vol. 11, no. 2,
pp. 431441, 1963.
[160] B. D. Hall, On the propagation of uncertainty in complex-valued quantities,
Metrologia, vol. 41, no. 3, pp. 173177, 2004.
[161] I. Markovsky and S. V. Huel, Overview of total least-squares methods, Signal
Processing, vol. 87, no. 10, pp. 2283 2302, 2007.
[162] G. DAntona, The full least-squares method, IEEE Trans. Instrum. Meas., vol. 52,
no. 1, pp. 189196, Feb. 2003.
[163] F. D. Neeser and J. L. Massey, Proper complex random processes with applications
to information theory, IEEE Trans. Inf. Theory, vol. 39, no. 4, pp. 12931302, July
1993.
[164] K. Yhland and J. Stenarson, A simplied treatment of uncertainties in complex
quantities, in Proc. Conference on Precision Electromagnetic Measurements Digest,
June 2004, pp. 652653.
[165] W. C. Daywitt, First-order symmetric modes for a slightly lossy coaxial transmission
line, IEEE Trans. Microw. Theory Tech., vol. 38, no. 11, pp. 16441650, Nov. 1990.
[166] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables, 9th ed. New York: Dover, 1964.
[167] A. Bergquist, Wave propagation on nonuniform transmission lines (short papers),
IEEE Trans. Microw. Theory Tech., vol. 20, no. 8, pp. 557558, Aug 1972.
[168] D. A. Hill, Reection coecient of a waveguide with slightly uneven walls, IEEE
Trans. Microw. Theory Tech., vol. 37, no. 1, pp. 244252, Jan. 1989.

234

S-ar putea să vă placă și