Simulation of Traveling-Wave Tubes for Analysis

and Optimization in Modulated Back-Off

Vom Promotionsausschuss der

Technischen Universität Hamburg
zur Erlangung des akademischen Grades
Doktor-Ingenieur (Dr.-Ing.)

genehmigte Dissertation

von
Djamschid Safi

aus
Hamburg

2020
First Examiner: Prof. Dr.-Ing. Arne F. Jacob
Second Examiner: Prof. Dr. rer. nat. Dr. h. c. Manfred Thumm
Chair of Examination Board: Prof. Dr.-Ing. Gerhard Bauch

Date of Thesis Defense: June 22, 2020

Acknowledgment
In the following, I would like to take the opportunity to thank some of the many
people who made this work possible by their support.
First, I would like to thank Prof. Dr.-Ing. Arne F. Jacob, who gave me the
opportunity to explore an exciting and challenging field of study and has always
supported me both professionally and personally. Also, I would like to thank Prof.
Dr. rer. nat. Dr. h. c. Manfred Thumm for taking the time to prepare the second
expertise and Prof. Dr.-Ing. Gerhard Bauch for chairing the examination committee.
This work was performed in collaboration with the company Thales. Here, Dr.-Ing.
Philip Birtel in particular was not only an ally and colleague, but also a teacher,
without whom this work would not have been possible. I would also like express my
gratitude to Dr. Jürgen Wegener and Monsieur Jean-François David for all these
fruitful discussions at Thales.
Likewise, I would like to address a thank you to my colleagues and thematic comrades-
at-arms at the Hamburg University of Technology, especially Moritz Hägermann for
the regular exchange and thorough proofreading, and Dr.-Ing. Sascha Meyne for being
a fantastic mentor on top of that.
But, of course, I would also like to thank the students who accompanied me during
my doctorate and whose dedication has contributed much to this work: Till Pusch,
Lukas Buderath, and Jonathan Ritter within their bachelor’s theses and Michael Wulff
for his ideas and efforts as a student assistant and within his bachelor’s thesis.
I cannot forget to thank my friends and my family, who supported me at all times.
And, most importantly, thank you, Louise. I simply could not have done any of
this without you being there.
Contents

1 Introduction 1
1.1 Traveling-Wave Tubes in Modern Satellite Communications . . . . . . . 1
1.2 Outline of this Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Basics of Traveling-Wave Tubes 5

2.1 Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Electron Gun, Beam, and Magnetic Focusing . . . . . . . . . . . 5
2.1.2 Components of the Interaction Region . . . . . . . . . . . . . . 6
2.1.3 Multi-Stage Depressed Collector . . . . . . . . . . . . . . . . . . 8
2.2 Operation Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Nonlinearities and Nonlinear Effects . . . . . . . . . . . . . . . . . . . . 13
2.3.1 Causes of the Nonlinear Behavior . . . . . . . . . . . . . . . . . 13
2.3.2 Nonlinearity Compensation . . . . . . . . . . . . . . . . . . . . 16
2.4 Efficiency and Power Flow . . . . . . . . . . . . . . . . . . . . . . . . . 18

3 Simulation Models for Modulated Signals 21

3.1 Parametric Simulation with MVTRAD . . . . . . . . . . . . . . . . . . 22
3.1.1 Operating Principle . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.2 Multi-Tone Operation . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Full-Wave Particle-In-Cell Simulation . . . . . . . . . . . . . . . . . . . 24
3.2.1 Operation Principle and Motivation . . . . . . . . . . . . . . . . 24
3.2.2 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.3 Model Validation and Simulation Accuracy . . . . . . . . . . . . 32
3.3 Envelope Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3.1 Analytic Methods and Simple Two-Tone Estimates . . . . . . . 37
3.3.2 Static-Gain Curve Model . . . . . . . . . . . . . . . . . . . . . . 39
3.3.3 Frequency-Dependent Model Topologies . . . . . . . . . . . . . 46
3.3.4 Three-Box Model . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3.5 Quadrature Polynomial Model . . . . . . . . . . . . . . . . . . . 54
3.3.6 SIMBA and the Satellite Downlink Chain . . . . . . . . . . . . 60

i
Contents

3.4 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.4.1 Multi-Tone Simulation with Measurement-Fitted Models . . . . 63
3.4.2 Multi-Tone Simulation with MVTRAD-Fitted Models . . . . . . 65
3.4.3 Transients and Communication Signals . . . . . . . . . . . . . . 68
3.4.4 Folded-Waveguide Simulation . . . . . . . . . . . . . . . . . . . 77
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

4 Back-Off Analysis and Optimization 83

4.1 Analysis of Nonlinear Effects . . . . . . . . . . . . . . . . . . . . . . . . 84
4.1.1 Asymmetric Intermodulation . . . . . . . . . . . . . . . . . . . . 84
4.1.2 Impact on Communication Signals . . . . . . . . . . . . . . . . 87
4.2 Relating Single- and Multi-Tone Characteristics . . . . . . . . . . . . . 93
4.2.1 Low-Level Noise-Power Ratio Estimation . . . . . . . . . . . . . 94
4.2.2 Frequency-Dependent Intermodulation Ratio . . . . . . . . . . . 102
4.3 Automated Back-Off Optimization . . . . . . . . . . . . . . . . . . . . 105
4.3.1 Method and Workflow . . . . . . . . . . . . . . . . . . . . . . . 106
4.3.2 Target Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

5 Conclusion 113

List of Acronyms 118

List of Figures 119

List of Tables 125

Bibliography 127

ii
1 Introduction
Traveling-wave tubes (TWTs) are high-power, high-efficiency vacuum electronic am-
plifiers for radio-frequency (RF) applications. Besides their predominant application
in satellite communications, they also play an important role in airborne and surface
radars, as well as in electronic countermeasure applications. In TWTs, a propagating
RF wave is amplified by extracting kinetic energy from a synchronous electron beam,
reaching average RF powers in the range of watts to hundreds of kilowatts. Within
the group of microwave vacuum electronic devices, TWTs are the commercially second
most successful type1 [1], [2], as for many applications they are unmatched in terms of
combined power, bandwidth, and cost.

1.1 Traveling-Wave Tubes in Modern Satellite

Communications
Power amplifiers in satellite transponders need to fulfill a number of requirements, such
as high output power, high efficiency, large bandwidth, low mass, and long lifetime
of more than 15 years [3]. In most of these terms, today’s TWTs are hard to beat.
At C-Band, for instance, the available output power of commercial devices grew by
a factor of 40 compared to the first commercially used TWTs, while the weight was
reduced by a factor of 2-6 [3], [4]. At the same time, research on TWTs has resulted
in efficiencies starting from as little as 1 % to nowadays above 75 % at saturation [4].
This development has led to a continuing dominance of TWTs in the satellite sector
compared to solid-state power amplifiers (SSPAs). Considering, for instance, the
failure rate per 1 W as of 2004, TWTs have shown to be more than six times [5] up to
one order of magnitude more reliable [2]. While newer studies suggest that this gap
is slowly closing [6], [7], the data indicates that it will remain significant within the
foreseeable future. Also, the continuously improved performance of TWTs resulted in a
much smaller overall cost compared to solid-state devices, as impressively shown by the
numbers from 2003 given in [2]. There, the findings of a "major satellite manufacturer"
1
Number one is the microwave oven’s magnetron.

1
1 Introduction

are recited, stating they would require three times the number of satellites, resulting
in an increased cost of $ 1 billion when using SSPAs instead of TWTs.
This trend was enabled by constant improvement in two main areas. The rise of
multi-stage depressed collectors in the 50’s and 60’s up to the first three-stage depressed
collector in the 70’s led to a significant reduction of losses generated in the collector.
From there on, in parallel to further collector development, the beam efficiency has
rapidly been increased by delay-line velocity tapering. Nowadays, optimized TWTs
utilize four- or even five-stage depressed collectors and complex delay-line tapers.
In recent years, with the introduction of the DVB-S2 and the latest DVB-S2x
standards [8], [9], the operating mode of TWTs in such satellite links has changed.
In these modern applications, the modulation schemes shift from simple ones, like
constant-envelope quadrature phase-shift keying (QPSK), to more involved amplitude-
and phase-modulated signals. This is important, as – like any physical amplifier
– the TWT introduces strong nonlinear distortions at high power levels according
to the signal envelope. To avoid these distortions, the TWT typically is operated
far in the linear region below saturation, i.e., in back-off. While this diminishes
nonlinearity-induced modulation errors, it severely reduces the overall efficiency, as
the beam efficiency decreases quickly. As these more complex modulation schemes
typically require better linearity, this issue calls for increased attention [10].
Currently, when simulating TWTs, their performance is mainly characterized in
the near-saturation operating state based on their narrowband properties, such as the
output power or the nonlinear phase shift. Specialized TWT simulation tools, such as
MVTRAD [11] or CHRISTINE [12], are typically designed to calculate these frequency-
domain properties for saturation-based optimization. For the aforementioned modern
operating modes, this does not suffice anymore. The modulation of the input signal
results in continuous, more broadband frequency spectra. The consequential distortions
have a strong influence on the transmission quality, but can only rudimentally be
predicted by classical steady-state codes. Thus, these classical simulation routines do
not directly allow optimization of the TWT customized to modern applications.
This work aims at providing suitable solutions to this problem. The goal is to be
able to carry out extensive simulation of the downlink chain of a satellite, operated
with different modulation types. This requires the development of methods, which
enable the fast and accurate calculation of meaningful multi-tone and multi-carrier
characteristics.
From today’s point of view, there are three promising approaches to tackle this
specific task. General purpose simulation tools, such as CST Particle Studio’s [13]
Particle-In-Cell (PIC) [2] solver or MAGIC [14], operate directly in time-domain and

2
 1.2 Outline of this Thesis

inherently allow the simulation of arbitrary input signals. Thus, they can be used
to study, e.g., transients, memory effects, and stability issues. Unfortunately, this
comes at the cost of a high computational effort and a fairly involved modeling, which
is why their reliability requires thorough investigation. Frequency-domain codes, on
the other hand, are well-established for the simulation of TWTs and are known to
accurately predict the nonlinear interaction. As they rely on a number of simplifying
assumptions they are able to significantly reduce the computation times for single-
tone excitation compared to general purpose tools. The calculation of meaningful
multi-tone characteristics, on the other hand, demands for multi-tone input signals.
This can be achieved using a large number of harmonics, which requires adjusting the
numerical parameters. As a consequence, this strongly increases the numerical effort.
A third option is the usage of so-called Envelope methods. Assuming a slowly varying
modulation of the TWT’s excitation, these methods enable a fast calculation of the
output signal in time-domain, based on data provided, e.g., by a fast frequency-domain
code. Simple Envelope approaches have proven to be accurate for narrow carrier
spacings, while their limits for TWTs driven by signals with complex modulation
schemes are not clear a priori.
These three methods are complementary to each other. In this work, suitable
modeling approaches are developed and characterized regarding their computational
effort, their practicability, and the accuracy of the results, which is additionally verified
by measurement data. Suitable approaches enabling the analysis and optimization of
TWTs for modern operating modes need to be as fast as possible to allow their usage,
e.g., at the design stage, but also as accurate as necessary to yield reliable performance
indicators. The development of such methods is undoubtedly crucial to further improve
TWTs in back-off and ensure their competitiveness in future applications.

1.2 Outline of this Thesis

This work attempts to lay groundwork for tackling the challenges of using and improving
TWTs for modern communication satellite systems. For this purpose, an essential
component is the ability to obtain common communication characteristics of the TWT
with as little effort as possible, at best already at the design stage. Thus, after a brief
introduction to important basics of TWTs in Chapter 2, the following Chapter 3 covers
various simulation methods to predict the transmission behavior of TWTs for such
modulation schemes. This is a prerequisite for understanding the effect of nonlinearities

3
1 Introduction

on communication signals and consequently optimizing the TWT regarding its actual
operating conditions, as is discussed in Chapter 4.
For simplicity, comparability, and consistency, most of the investigations in this
work are carried out on a state-of-the-art 150 W commercial reference TWT for
communication in Ku-Band. In some cases, a distorted version of it is used to highlight
certain frequency-dependent effects, resulting in a more dispersive TWT simulation
model. Results from this "crooked" TWT are tagged accordingly whenever they are
considered. In addition, for validation purposes, Section 3.4.4 features the simulation
of a folded-waveguide (FW) TWT as a representative of future technology in the field
of satellite communications.

4
2 Basics of Traveling-Wave Tubes

2.1 Components
Accurate modeling, simulation, and optimization of TWTs require a solid understanding
of the main elements playing a role in the amplification process. Therefore, these
components will briefly be described in the following, starting from the electron gun
up to the collector. In Figure 2.1, a simplified schematic of a helix TWT is shown,
indicating all covered components.

2.1.1 Electron Gun, Beam, and Magnetic Focusing

The electron beam, which is required for the interaction with the traveling wave, is
generated in an electron gun. For this purpose, a cathode is heated by a filament
at typical temperatures of around 1000 K to 1500 K, such that electrons are emitted
following the thermionic emission mechanism [1], [3]. Lifetime and current density
considerations lead to a large, spherical cathode. The cathode is followed by a focus
electrode, which serves to bundle the beam towards the desired beam radius. One
or several anodes are used to further shape the beam, control the current flow, and
accelerate the electrons according to the applied anode voltages. Depending on the
application, the velocity of the electrons leaving the electron gun reaches around

PPM-stack (ring-magnets)
Support rods

N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N

N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N

Electron gun Collector

Delay line Sever
RF-in RF-out

Figure 2.1: Schematic view of a helix TWT.

5
2 Basics of Traveling-Wave Tubes

10 − 40 % of the speed of light. The electron beam is typically focused by a periodic

permanent magnet (PPM) configuration, since it allows a weight reduction of up
to two orders of magnitude compared to solenoidal magnets [1]. This is especially
important for satellite applications. According to Busch’s theorem [1], a magnetic flux
Bc through the cathode causes a rotation of the beam in azimuthal direction θ, with
angular velocity
!
∂θ η rc2
= B − Bc 2 , (2.1)
∂t 2 rb
where rc and rb are the cathode and beam radius, respectively, and η stands for the
electron charge to mass ratio. B denotes the axial magnetic field, which according to
the Lorentz force radially focuses the rotating beam, counteracting the space charge
forces and their diverging effect on the beam.

2.1.2 Components of the Interaction Region

The interaction region denotes the entire area between electron gun and collector,
where electron beam and traveling wave are coupled for the sake of energy transfer. The
principle behind this energy transfer is outlined in Section 2.2. The main ingredient of
the interaction region is the so-called delay line or slow-wave circuit. The amplifier’s
performance and properties strongly rely on the type of delay line and its implementa-
tion. The main purpose of the line is to slow down the electromagnetic traveling-wave
in axial direction by forcing it to take a defined detour along the slow-wave structure.
This is essential to enable coupling between the electron beam and the traveling wave.
Significant interaction occurs, if the axial phase velocity of the wave traveling on the
delay line is close to the velocity of the electrons in the beam. There are various types
of delay lines with different fields of application, the commercially most important
ones to be mentioned are the helix and the coupled-cavity line.
Helix TWTs offer very large bandwidths of up to several octaves, at the cost of a
comparatively moderate power capability. The circuit, as it is shown in Figure 2.2a,
consists of a wire which is wound up on a helical path around the beam axis. Especially
the pitch and the radius of the helix determine its properties, serving as set-screws to
adjust the phase velocity and the coupling to the beam. The helix wire is fixed inside
the TWT’s hull by means of ceramic support rods and therefore exhibits poor thermal
properties. Helical delay-lines are the dominant topology for satellite communication
applications at the currently most relevant frequency bands.
For high-power applications, such as Radar, coupled cavities out of bulk metal are
advantageous. There, the electron beam travels through a beam tunnel and interacts

6
 2.1 Components

Hull

Beam tunnel
Helix
Beam wire
tunnel Beam tunnel

Waveguide Waveguide
Support Pitch
rod Bulk metal Bulk metal

(a) Helix. (b) Folded-waveguide.

Figure 2.2: Two suitable delay-line types for satellite communications.

in gaps along the line with the electromagnetic wave in the cavities. While their
amplification is rather narrowband, their power capability is significantly larger due to
the thermal conductivity of the bulk metal. Somewhere in between these delay-line
types lies the FW geometry, shown in Figure 2.2b. It can be seen as a compromise
between the broadband helix and high-power coupled-cavity lines and is currently
investigated as a promising alternative for communication satellite applications at
Q-Band [15] and above. A rectangular waveguide is folded around a beam tunnel
to enable the required phase synchronicity. The winding can either be smooth in
a meander-like manner or with sharp 90◦ -corners, as depicted in Figure 2.2b. Like
for coupled-cavity delay-lines, the beam interacts with the traveling wave within the
discrete gaps at the crossing points between the beam tunnel and the folded waveguide.
Regardless of the delay-line type, typical TWTs exhibit high amplification. Today’s
helix TWTs can easily reach 60 dB in gain and more. Thus, stability against oscillation,
for instance due to mismatched delay-line couplers at the RF input and output, has
to be considered. For this purpose, one or several severs are introduced. They are
strongly attenuating and well-matched elements, which in practice, for instance, are
realized by means of thin layers of lossy material on the support rods. Effectively,
severs separate the interaction region into several sections, such that the loop gain
conditions for avoiding undesired oscillations is met everywhere. While the power of
the traveling wave is fully absorbed in the attenuator, the modulated electron beam
drifts through it and regenerates the signal afterwards.
The bandwidth and frequency dependency of the individual TWT types is a result
of the delay line’s dispersive property. According to Floquet’s theorem [16], the
electromagnetic field has the same periodicity as the delay-line except for a phase
factor ϕ. Thus, it is reproduced along the periodic structure with this phase factor
from one unit cell to another. This is often visualized by means of the dispersion

7
2 Basics of Traveling-Wave Tubes

diagram, which describes the phase advance on the delay line along one unit cell
pitch p versus frequency. In simulation, the diagram can be obtained by means of an
eigenmode analysis, where the unit cell of the delay line is modeled with phase-shifted
periodic boundary conditions in axial direction. Thus, a propagation constant
ϕ
β= (2.2)
p
can be obtained for the calculated eigenmode.
Due to the periodicity, the axial electric field can be expanded into a Fourier series
[17]
En In (γn r)e−jβn z e−jnθ ,
X
Ez (r, θ, z) = (2.3)
n

where In denotes the modified Bessel function of n-th order and first kind, with one
phase propagation constant
ϕ 2nπ 2nπ
βn = + = β0 + , (2.4)
p p p
and one phase velocity
ω
vph = (2.5)
βn
for each space harmonic n, but a common group velocity
∂ω
vgr = . (2.6)
∂β0
For a known field shape, each space harmonic with index n can be described by a
single scalar En , which is extracted from the eigenmode calculation. Typically, instead
of En , the coupling impedance [17]

|En |2
Zc,n = 2 (2.7)
2βn Pt
is used to characterize the coupling strength, with Pt describing the total transported
RF power.

2.1.3 Multi-Stage Depressed Collector

A main advantage of TWTs compared to SSPAs is their high total efficiency. A
significant contributor to this figure of merit is the collector, where the spent beam1
1
The beam is called "spent", when it has passed through the interaction region and the DC input
power has been used for interaction. Thus, the term denotes the beam at the interface between
delay line and collector. It is considered as spent, no matter which part of its energy is transferred.

8
 2.1 Components

at the end of the interaction region is collected to reuse as much of the remaining
beam power as possible. For this purpose, the focusing is removed to enable the
beam to diverge. The collector potential is depressed from the TWT’s body potential.
Otherwise, the electrons would hit the collector’s walls at full speed and their kinetic
energy would largely be converted to heat. A potential depression, where, relatively to
the cathode, a reduced voltage VColl is supplied to the collector, leads to a collector
current IColl , which provides more power from the beam to the DC power supply, while
the same current is processed. Therefore one can also consider this a DC input power
reduction by
PDC,out = VColl IColl , (2.8)

rather than a DC power recovering.

The electrons of the spent beam entering the collector exhibit a variety of velocities,
depending on, e.g., the delay-line taper, frequency content, and drive level. The spent
beam’s velocity spread is quantified in the electron-velocity spectrum (EVS). For
visualization purposes, the electrons are sorted according to their kinetic energy and
accumulated up to the total current entering the collector. Figure 2.3a shows an
exemplary spent beam resulting from a frequency-domain simulation for a large-signal
single-tone excitation close to saturation. The color represents the radial positions
of the particles, with yellow being those at the outer radii and dark blue the inner
ones. Due to its steady-state background, the resulting temporal beam distribution is
periodic according to the fundamental frequency. In Figure 2.3b, the shown period is
evaluated with respect to the kinetic distribution of the current and compared to a
small-signal result and the unmodulated case. The energy axes are normalized to the
initial kinetic energy Ekin,0 , and the current to the total beam current I0 . A strong
variation of the spectral properties with drive is apparent.
As the beam is slowed down by the collector depression, the maximum power that a
single-stage collector can recover is limited by the slowest expected electrons, arriving
min
at the collector with kinetic energy Ekin . The amount by which the potential of the
collector Vc is depressed against the helix potential VHelix must not fall below the
voltage at which the slowest electrons cannot reach the collector anymore. Thus, it is
limited by
min
q (VHelix − Vc ) < Ekin , (2.9)

with the elementary electron charge q. A breach of this condition would allow the
electrons to be reflected back into the interaction region and disturb the amplification
or even damage the device. To cover as much of the electron velocity distribution of
the spent beam as possible, modern collectors use several stages at different depressed

9
2 Basics of Traveling-Wave Tubes

1.2 rb 1.2

Cumulated current I/I0

Unmodulated
1 1

Radial position →
Energy Ekin /Ekin,0

0.8 0.8

0.6 0.6
Small-signal
0.4 0.4

0.2 0.2
Large-signal
0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 1.2
Time/RF-period Kinetic energy Ekin /Ekin,0

(a) Spent-beam energy in one RF period. (b) EVS from spent beams.

1.2
Vc1 Vc2 Vc3 Vc4
Cumulated current I/I0

1
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2
Kinetic energy Ekin /Ekin,0

(c) Energy recovering in the collector.

Figure 2.3: Large-signal spent beam, the corresponding EVS, and the recollection in
a multi-stage depressed collector.

10
 2.2 Operation Principle

potentials, resulting in a DC power reduction of

Vci Ici ,
X
PDC,out = (2.10)
i

where a collector current Ici is generated in the i-th collector stage at potential
Vci . Figure 2.3c shows the spent beam EVS from a large-signal simulation and
how it is collected in a perfectly sorting multi-stage depressed collector with stages
Vci , i = 1, . . . , 4. For each voltage, one can mark a point on the EVS curve, which
determines the collected current on the stage. Then, the sum of the four rectangular
areas framed in blue determines the power reduction. In consequence, the patterned
red area describes the power that is dissipated. Thus, the energy distribution of the
spent beam has a large influence on the possible DC power reduction, being optimal
when the EVS appears stair-like, according to the stage voltages.

2.2 Operation Principle

For the operation principle of TWTs, various figurative approaches can be found in
literature, such as the surfer, whose velocity has to match to the (water) wave’s phase
velocity for effective energy transfer2 , or a stream of cars moving over a hilly street
[3], where on the uphill side the traffic is decelerated forming a bunch towards the
hilltop and the opposite occurs on the downhill side. The need for such simplified and
more conceptual explanations hints at the general complexity behind the functional
principle of TWTs. A basic description of the helix TWT’s operation, as for instance
given in [1], is outlined in the following.
Applying an RF signal on the helical line, a charge distribution according to the
sinusoidal field distribution of the signal frequency can be observed on the helix,
moving along the line at a certain phase velocity. This can be seen in Figure 2.4.
In the shown configuration, one wavelength equals four turns of the delay line. A
force on the electrons in the electron beam is found following the electric field lines,
which are more concentrated inside the helix. Assuming an initial velocity from left
to right, alternating accelerating and decelerating areas can be determined. Thus, an
accelerating force acts on parts of the beam and a decelerating one on others, leading
to areas of bunched particles, moving from left to right. When the phase velocity
of the traveling wave is slightly smaller than the velocity of the beam, the electron
bunches move even further into the decelerating region.
2
Andrei Haeff, one of the inventors of the TWT, supposedly got the idea watching surfers, coining
the term "traveling wave". [18]

11
2 Basics of Traveling-Wave Tubes

−
−−−−− − − −−−−−
−−−−−− − −− −−−
− −−−
−−
− −
−−
− − −−−
−
− −−
− −−
−− Facc −−−
− −−−
− −−−−−
− − Facc −−
−
−
− −−−−−− −
−
−−−−
−−−
−
−−−
−
−−
−
− −− Fdec −
−− −−
−
−−
−
−
−
−−
−
−−−−−−
− Fdec −
−
−
−−−
−−
−
−−−−−
−−
−−−
−−−
− −− −−− −−−−
− −− −−
−−−
−

Figure 2.4: Charges on the circuit and resulting electric field for an RF signal applied
to a helical delay line. Facc and Fdec denote accelerating and decelerating
forces acting on the electrons, which, for reasons of clarity, are downscaled
in numbers and indicated by purple dots.

The bunches represent charge accumulations, which themselves act on the charges
on the helix wire. These charges on the line are pushed away from the bunches, leading
to a generally sharper charge distribution and thus an enhanced field on the helix. As
the charge on the helix moves away from the bunch in both directions, its velocity to
the left and the right of the bunch differs relatively to the beam’s DC velocity. As
a result, the electrons in front of the bunch accumulate, such that a phase shift of
90◦ occurs. The electron bunches are thus further decelerated and the bunches on the
beam and the accumulations on the line increase, leading to further amplification of
the signal.
This qualitative explanation is still fairly simplified and does not enable the quan-
titative characterization of the process. For this purpose, a small-signal theory of
the fundamental principle was developed by Pierce [17]. There, an equivalent slow-
wave circuit is considered, interfaced with the closely passing beam which serves as a
distributed current source to the circuit. In the synchronous case, where the beam
velocity equals the traveling-wave’s phase velocity, a determinantal equation with four
solutions for the different waves propagating on the circuit can be found. These have
different propagation characteristics, three traveling in forward direction and one in
backward direction. Of the forward waves, one shows exponentially growing behavior.
From this, the linear small-signal gain G for a homogeneous section can be estimated
to

G ≈ −9.54 + 47.3 CN dB, (2.11)

12
 2.3 Nonlinearities and Nonlinear Effects

where N denotes the length of the section divided by the signal wavelength, and C is
the Pierce gain parameter [1], [17]

KI0
C= , (2.12)
4V0

with K being the circuit impedance3 , I0 the beam current, and V0 the beam voltage.
For this estimate, a lossless, synchronous operation is assumed and large-signal effects
such as large-signal space charge considerations are neglected. Extended discussions of
the TWT’s operation theory can be found, e.g., in [1], [17].

2.3 Nonlinearities and Nonlinear Effects

The TWT is, as any physical amplifier, a nonlinear device. This can also be seen
in typical dynamic amplifier characteristics, such as the output power, which is
schematically shown in Figure 2.5 versus the RF input power of the TWT. The usable
small-signal region, where linear gain is observed as described in Section 2.2, is limited
at the lower end by the noise generated in the TWT, in which the low-power signal
then perishes. Above the noise floor, a change in input power leads to a proportional
change in output power, as the gain is constant. At some point, for increasing input
power, the gain decreases. This constitutes the beginning of the large-signal region.
Drawing from [2], the following shortly outlines the underlying mechanism leading to
this loss of linearity, as well as typical means to mitigate it.

2.3.1 Causes of the Nonlinear Behavior

While the delay line itself is linear, the coupled beam-wave system is inherently
nonlinear in nature. The electron beam can be described as a superposition of its
steady component and a periodic component, which increases in axial direction and
describes the beam modulation. We call these the DC and the AC part of the beam.
For the sake of simplicity, a steady-state excitation of a homogeneous delay line with
frequency ω is considered.
At the injection the beam only consists of a DC part. This changes with interaction,
as, due to the coupling, a velocity modulation is impressed on the beam, according to
the RF signal on the delay line. Thus, the electron velocity v has both a DC part v0
3
In [1] and others, the circuit impedance is denoted by Zc , sometimes leading to some confusion.
The circuit impedance sums up all information required on the circuit, such that no further
information on the actual representation is required.

13
2 Basics of Traveling-Wave Tubes

Saturation

Output power (log.)

Dynamic range
1dB -
Compression

ain
point

rg
Overdrive

ea
Lin
Noise floor

Input power (log.)

Figure 2.5: Amplifier characteristics of a TWT.

and an AC part v1 , with

v = v0 + v1 · exp (jωt) . (2.13)

As a result of this modulation, electrons travel at different velocities and the beam
starts to bunch. Consequently, the space-charge density ρ is modulated, too. It also
consists of both a DC component ρ0 and an AC component ρ1 , with

ρ = ρ0 + ρ1 · exp (jωt) . (2.14)

As to these descriptions, both AC components of v and ρ at first follow harmonic

functions.
For the current density J, their product

J =v·ρ (2.15)
= v0 ρ0 + (v1 ρ0 + v0 ρ1 + v1 ρ1 · exp (jωt)) · exp (jωt) . (2.16)
| {z } | {z }
=:J0 =:J1 (t)

is considered. The current density already shows harmonics of the drive frequency, and
its AC amplitude itself is time-dependent with the RF frequency. Even under purely
ballistic conditions, neglecting repelling forces between the particles and determining
their movement independently of each other, with time even higher harmonics can
occur, as the beam density is further peaked and more sharply bunched.
The harmonics of the beam current in turn effectuate an electric field at the harmonic
frequencies on the delay line, as the current Jind induced on the helix relates to the
beam current J with [1]
δJind δJ
∝ , (2.17)
δz δz

14
 2.3 Nonlinearities and Nonlinear Effects

which then results in a related electric field. This electric field again has a decelerating
and accelerating influence on the modulated beam.
We can consider the beam to be a nonlinear source. The signal on the line
now consists of both the fundamental frequency component and harmonics of it,
which again cause a – now more involved – velocity modulation and space-charge
density modulation. By multiplication, this introduces further harmonics and also
intermodulation products at sums and differences of multiples of the mixed frequencies.
Some of these generated intermodulation products introduce components at the
fundamental frequency. Thus, there, the power is determined by superposition of the
initial fields and the intermodulation products generated in the nonlinear beam.
The model of ballistic bunching is of course a simplification, which fits well far
in back-off, where the excitation is small and the bunching forces are weak. When
considering higher drive levels, the modulating forces and therefore also the mixing
becomes stronger. This requires taking space-charge forces into account. As the
electrons interact with each other, their repelling forces would, lacking an external
force, lead to a diffusion of the bunches. Thus, these can only be sustained, if the
delay line contributes with a sufficiently large axial electric field. This traveling wave
is amplified taking power from the beam in the coupled process and at the same time
sustaining the bunches of the beam.
On average, the kinetic energy of the electron beam is transferred to the electric
field, while the kinetic-energy spectrum of the beam is spread. Therefore, the beam
does not only slow down on average, but subsequently also loses synchronism with
the wave. As this synchronism is required for the traveling wave to counteract the
debunching, the velocity modulation spread leads to a relaxation of the space-charge
modulation. This effect depends on the drive level and causes the compression of
the gain curve. At some point, the bunches cannot be sustained anymore and no
more kinetic energy can be transferred from the beam to the wave. This limits the
amplification process as the amplifier is in saturation.
At the same time, the reduced average beam velocity influences the phase shift
observed in the RF wave. The superposition of the traveling wave with the intermod-
ulation products generated in the nonlinear beam already results in a phase shifted
oscillation. The phase shift increases with the amount of extracted power, as the slower
beam additionally pulls back the RF wave. Further increase of the RF excitation
above saturation even reduces the maximum power at the output, as energy partially
is transferred from the traveling wave back to the electron beam.
In the nonlinear region, two main effects occur. The first one is the gain reduction
or compression and is often denoted as amplitude-modulation to amplitude-modulation

15
2 Basics of Traveling-Wave Tubes

(AM-AM). The second one relates the phase change which is caused by a modulated
amplitude to the phase in small-signal operation. It is often called the nonlinear phase
shift or amplitude-modulation to phase-modulation (AM-PM). These two characteris-
tics are effective metrics to describe the nonlinearity and useful for predicting nonlinear
distortion or intermodulation, as will be shown in Chapter 3. Still, it is important to
note that they are not a cause of intermodulation, but rather caused by it, as discussed
in this section. Also, the AM-AM and AM-PM characteristics only summarize what
is seen from the outside of the TWT, possibly distorted by the coupler, while the
intermodulation itself in principle occurs throughout the device.

2.3.2 Nonlinearity Compensation

In the context of satellite communications, distortions as introduced by driving the

TWT in the nonlinear regime are to be avoided. In the following, two approaches to
reduce these distortions are briefly outlined.

Predistortion Linearization

Predistortion linearizers [19] are commonly used in satellite communications and

therefore an integral part of a holistic downlink chain investigation. They rely on
conditioning the input communication signal by means of a preceding nonlinear circuit,
such that the TWT’s output signal seems undistorted. This conditioning relates to
a controlled distortion of the signal according to the expected phase and amplitude
distortion in the TWT. Thus, the circuit is configured to "fix" the nonlinearity of
the AM-AM and AM-PM transfer characteristics, by offering, e.g., a gain expansion
where the TWT exhibits gain compression and analogously correcting the phase shift.
Integrating the gain expansion and compression into one transfer characteristic together
with the phase correction, the combination of linearizer and TWT at best yields an
ideal limiter. In reality, a non-ideal transition is found between the region of linear
amplification and the upper limit. This is shown in Figure 2.6. Spectrally, the signal
after predistortion shows the intermodulation products obtained by solely driving the
TWT, but at a phase shift of 180◦ (at best) such that destructive interference occurs.
There are many different types of both active and passive predistortion circuits and
linearization methods, allowing a trade-off between performance, cost, and complexity.
An extensive overview can be found in [19].

16
 2.3 Nonlinearities and Nonlinear Effects

Linearizer TWT

Pout , ∆ϕ

Pout , ∆ϕ

Pout , ∆ϕ
Pin Pin Pin

Figure 2.6: Schematic of a TWT with a preceding predistortion linearizer with

individual and combined transfer characteristics.

Power Combining and Velocity Tapering

Another way to achieve a certain higher level of linear power is to increase the margin to
saturation, for instance by increasing the available peak power in a controlled manner.
A simple way to do so is utilizing power combining. While it significantly increases the
power margin, using N TWTs does not necessarily result in an N -fold peak power. For
efficient combination of several TWTs, the phase at the single TWT’s output needs to
be preferably stable over time and taken into account for the superposition of signals.
Statistical variations in the output phase over drive, which are to be expected from, e.g.,
manufacturing tolerances, increase the power combining losses in addition to auxiliary
elements required for the combining, such as couplers and further transmission lines
that also introduce losses.
For the individual TWT, the peak power, linearity, and many other characteristics
have in the past significantly been improved by means of velocity tapering [1], [20]. In
modern helix TWTs the pitch profile of the delay line is inhomogeneously tapered to
control the phase velocity of the wave and thus the synchronism with the beam. This
can significantly increase the performance of the amplifier and strongly influence its
behavior. The profiles nowadays usually stem from numerical optimization and thus,
in principle, do not follow an analytic formulation.
The idea of tapering closely follows the operating principle of the TWT. As es-
tablished, the average beam velocity decreases along the interaction region due to
energy transfer towards the traveling wave. This relates to a change in synchronism
and, thus, an altered beam-wave coupling, possibly impeding an efficient amplification

17
2 Basics of Traveling-Wave Tubes

PDC,in Qloss ∝ Pout PDC,out

QC QSever QHelix

N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N

PB ≈ PDC,in PC

N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N N S S N

Electron gun Collector

Pf
Pin Pout PHarm.
Qdiss,coll

Figure 2.7: Schematic view of the power flow in a TWT.

process. Therefore, by velocity tapering, the circuit is varied along the beam axis,
such that the synchronism is adjusted according to the decelerated beam. Early taper
designs solely following this synchronism-adjustment approach focused on the area
near the output to improve beam efficiency, leading to a higher peak power, but also
to larger phase distortion. Later, more involved taper topologies have been utilized to
also compensate the nonlinear phase shift [20]. This has lead to immensely improved
TWT designs, both over drive and frequency, as well as against oscillations, and still
serves as a fruitful source of TWT performance progression.

2.4 Efficiency and Power Flow

Typically, the power amplifier is one of the most power hungry components in a
satellite. Therefore, its efficiency significantly contributes to the total efficiency of
the system and is crucial for the overall power and thermal budget. Additionally,
the efficiency plays a major role from an economic point of view, with a 1 % overall
reduction leading to launch cost savings of more than 400 ke per satellite4 . Thus, it is
important not only to identify the relevant loss contributors and mechanisms, but also
to mitigate their impact.
In Figure 2.7, the power flow in a helix TWT is sketched. The collector and electron
gun subsystems are each separated by a dotted line. All dissipated and harmonic
powers are marked in red, the interfacing powers between the subsystems are shown
in purple, and the RF- and DC-powers into and out of the TWT are marked in blue.
The heated cathode in the electron gun is powered by a heated filament. For this,
a power Pf is required, which usually is in the order of a few watts. Also, a DC power
supply is used to generate the current with an acceleration voltage V0 relative to the

4
This is assuming a typical satellite with 40 TWTs of 100 W output power each [4].

18
 2.4 Efficiency and Power Flow

cathode. A beam power

PB = V0 · I0 = PDC,in − Pf ≈ PDC,in (2.18)

leaving the electron gun subsystem enters the interaction region.

Here, the beam interacts with the RF input power Pin , which is amplified along the
delay line. In principle, the desired amplified signal and various harmonics arrive at
the output coupler. This coupler typically is optimized for the amplification frequency
band. Thus, large parts of the generated harmonics are reflected back into the device.
While this plays a role for instability considerations, most of the time the reflected
power will be absorbed in the sever without further consequences. The desired RF
output power Pout is coupled out with a remaining harmonic part PHarm. . The amount
of useful output power is related to the beam power by the beam efficiency
Pout
ηBeam = . (2.19)
PB
In addition, the total efficiency
Pout
η0 = (2.20)
PDC,in − PDC,out

covers the DC power regeneration in the collector, providing a more complete picture.
Along the delay line, various losses leading to an efficiency reduction are to be
considered. First of all, a loss QSever is produced in the severs. In addition, in the
helical delay line, the barrel, and the support rods, material losses QC are generated. In
parts, these are desired for stability reasons and depend on the RF power transported
along the delay line. The reflected power at the output coupler also contributes to
QC and QSever 5 . Finally, especially for a strongly bunched beam, some electrons hit
the delay line unchecked, locally generating dissipated power QHelix . Both QC and
QHelix are unevenly distributed along the beam axis, being strongest at the end, as
they depend on the beam modulation and local RF power. As the severs are usually
located at positions far from the output, the total dissipated power in the attenuator
is small compared to Pout , even though the total attenuation is at least in the order of
tens of dBs. Therefore, while the total interaction region losses

Qloss = QC + QSever + QHelix (2.21)

cannot exactly be calculated or accurately be predicted, in back-off they are roughly

proportional to the output power and can be estimated to around 15 % of Pout [2].
5
For the input coupler, this is negligible.

19
2 Basics of Traveling-Wave Tubes

100 100 100

100 90
80
100
70 90
Collector efficiency (%) 60

Total efficiency (%)

80 50 80 80

40
30
70

20

100
60 10 60

60

90
40 40

50

80
70
20 20

100 100
10

60
50
20
10

30

0 40 0
0 20 40 60 80 100
Beam efficiency (%)

Figure 2.8: Approximate total efficiency of a TWT as a function of the collector and
beam efficiency.

In modern designs, the collector losses

QColl = PB − Qloss − Pout − PDC,out (2.22)

reach similar dimensions. From this, a collector efficiency

PDC,out
ηColl = (2.23)
PB − Pout − Qloss
can be defined, relating the DC output power reduction to the remaining power in the
spent beam. With
ηBeam
η0 ≈ (2.24)
1 − ηColl (1 − ηBeam )
a simplified relationship between collector and beam efficiency can be found, where
filament and interaction region losses are neglected [3]. The impact of improving either
ηBeam or ηColl according to Equation (2.24) is shown in Figure 2.8.

20
3 Simulation Models for
Modulated Signals

This chapter presents a selection of possible ways to simulate multi-tone and com-
munication signals in of TWTs. In principle, there are various options to fulfill this
task.
From the TWT manufacturer perspective, the simulation of well-established cont-
inuous-wave (CW) characteristics like single-frequency saturated input and output
power or nonlinear phase shift have historically been most important for evaluating
the amplifier performance. As established in the introduction, today, increasingly
complex modulation schemes are used. Therefore, other performance metrics are on
the rise, which also changes the requirements for simulation methods. In general,
suitable methods for this task should be as fast as possible and as accurate as necessary.
Therefore, for each of the models the accuracy of the results, the required computation
time, and their practicability need to be considered.
Frequency-domain codes like MVTRAD [11] or CHRISTINE [12] are fast, reliable,
and well-validated for CW simulation, but are naturally limited for more general
time-varying signals. Still, in some cases, modified versions of such frequency-domain
codes can be used for multi-tone calculations. Anyway, the more intuitive choice are
time-domain based tools like CST Particle Studio [13] or DiMoHa [21], [22], which
allow for arbitrary input signals, and thus also allow the study of stability issues,
transients, or memory effects. For time-domain full-wave simulation tools, the biggest
drawbacks however are the quite involved modeling required for accurate results and
the very long computation times. Therefore, they are most suitable for validation and
investigation of special phenomena. Alternative proprietary time-domain tools like
DiMoHa, which are specialized on TWTs, are currently under development. While
they are potentially faster than general purpose tools, they still have to come to age.
Even then, a full simulation of, e.g., realistic QPSK modulated signals on multiple
carriers would take quite a long time. Within the possible options for modulated
signals, hybrid time- and frequency-domain envelope approaches are the middle ground,

21
3 Simulation Models for Modulated Signals

trying to combine the best of both worlds: a time-domain integration according to the
modulated signal with a frequency-domain backbone.
In the following sections, a discussion of the several simulation methods regarding
their performance, advantages, and disadvantages, as well as a comparison between
each other and to measurement results is given. Also, the relation between some
well-known CW quantities and multi-frequency quantities is briefly investigated.

3.1 Parametric Simulation with MVTRAD

MVTRAD, short for MouVemenT RADial, is a proprietary steady-state frequency-
domain code. Following the so-called parametric approach, the slow-wave circuit is
represented by cold parameters1 , such as the coupling impedance and the phase velocity.
To extract these parameters, the eigenmodes of a periodic unit cell are calculated
by full-wave simulation or analytic models, such as the tape-helix model [23], [24].
MVTRAD is specifically designed for the large-signal simulation of helix TWTs and,
for this case, well-verified and established. In the following, its operation principle is
briefly outlined, as it serves as a reference for many results and models in this chapter.

3.1.1 Operating Principle

The 2.5D PIC-code MVTRAD computes the beam-wave interaction by modeling the
electron beam as macro-particles in a 2D plane and describing the RF helix fields by a
number of harmonics. Rotational symmetry of both the electromagnetic fields and
the beam is assumed, while still taking the azimuthal movement of the particles into
account.
The circuit’s eigenfields are built by means of field expansion. For this purpose,
the harmonic expansion in Equation (2.3), and consequently the phase velocity from
Equation (2.4) and the coupling impedance from Equation (2.7) are used. In fact, inside
the beam region, in terms of beam-wave synchronism and field magnitude, mainly the
space harmonic with n = 0 is relevant for the amplifier simulation. Thus, for a uniform
helix geometry, the cold circuit properties can be represented by a single coupling
impedance value and a single phase shift per frequency point. The RF fields are
described by a number of harmonics according to the number of frequency components
that are to be considered in the simulation. The cold parameters describing the whole
1
"Cold" parameters are parameters of the structure without an electron beam. When an electron
beam is considered, one speaks of, e.g., "hot" parameters.

22
 3.1 Parametric Simulation with MVTRAD

slow-wave structure therefore need to cover all used frequencies and all geometric
configurations of the delay line.
In addition to the circuit fields, magnetic focusing fields and the space-charge
fields are taken into account. The latter are obtained numerically from the particle
distribution. In this implementation of the PIC approach, these particles are pushed in
time-domain, while the circuit fields are described in frequency-domain. In each time-
step of the simulation, particles are emitted into the interaction region. Following the
PIC-approach, their charges are distributed onto the numerical grid, representing the
space-charge density inside the beam. In contrast to the circuit fields, the space-charge
field is calculated at each time step, assuming it to be static during the time-step.
This enables the use of Poisson’s equation instead of the full description by Maxwell’s
equations, reducing the computational effort.
The particle trajectories resulting from the interaction with all considered fields are
computed iteratively, leading to an update of space-charge and circuit fields. This is
repeated until convergence is reached.
Any tool enabling meaningful simulation of modulated signals obviously needs to
include all significant nonlinear effects as described in Section 2.3. In MVTRAD, the
particle beam contains all necessary effects, as it is described as macro-particles and
their motion is integrated in time-domain. As the number of particles required is
significantly reduced by utilizing symmetries, for many cases MVTRAD is much faster
than general purpose PIC tools. Typically, large-signal CW computations can be
carried out in less than a minute per frequency and power setting on a typical desktop
computer.

3.1.2 Multi-Tone Operation

Although any nonlinear mixing in the electron beam can in principle be simulated in
MVTRAD, the electromagnetic fields are only considered in terms of known harmonics.
While this indeed simplifies the field calculation for determining typical amplifier design
characteristics from CW excitation, it complicates the multi-tone simulation and all the
more the simulation of arbitrarily modulated complex signals. Therefore, the offered
scope of applicability is limited. Still, MVTRAD can be used for multi-tone operation,
as the general approach behind it allows for not only the fundamental frequency,
but also harmonics of it. RF-harmonics are already required for CW excitation, as
they are generated in the electron beam and contribute to the amplification process.
This multi-harmonics capability can, in certain cases, be used for a multi-frequency
excitation, where the input signal consists of several tones inside the amplification

23
3 Simulation Models for Modulated Signals

band. In principle, such a signal can be calculated by taking the greatest common
divisor as the fundamental frequency with a sufficiently high number of included
harmonics. This requires a possibly exorbitant number of harmonics to be considered,
and as a result strongly increases the computation time. Also, the update frequency
∆fτ , as the inverse of the time step ∆τ in which fields and particles are updated, is
selected depending on the fundamental frequency f0 = T10 to

1 Ndiv
∆fτ = = = f0 Ndiv , (3.1)
∆τ T0

with an integer scaling Ndiv between them. Thus, of course, the number of time
steps per RF period for a multi-tone simulation needs further scaling to achieve a
similar time resolution as in the single-tone computation case. As the greatest common
divisor would generally be orders of magnitude smaller than the operating frequency
- and for arbitrarily modulated signals in principle infinitely small - the calculation
time would increase disproportionately. To counteract this, some adaptions have
been included in MVTRAD. While the number of harmonics is immense, many of
the harmonics are only dummies and not actually required for the field description.
Therefore, these are ignored for the field updates. Also, while a high number of time
steps per period is required, full particle injection into the calculation domain may
possibly be required less frequently without significant influence on the overall particle
density. Although these adaptions strongly reduce the computation time, the use of
codes like MVTRAD for multi-tone simulation still is only possible in special cases, as
is reported in Section 3.4.

3.2 Full-Wave Particle-In-Cell Simulation

Time-domain full-wave simulation tools allow arbitrary input signals and, thus, enable
the study of a variety of time-domain phenomena. The biggest drawbacks of these
geometry-driven solvers, however, are a significantly more involved modeling and long
computation times. In this section, at first the operation principle is shortly introduced.
Afterwards, some modeling approaches for one of the most prominent codes of this
category, CST Particle Studio’s PIC solver, are presented.

3.2.1 Operation Principle and Motivation

Full-wave PIC codes [2] follow the "first-principle"-approach. They directly solve
Maxwell’s equations in a geometry-driven strategy, fully in time-domain. Therefor,

24
 3.2 Full-Wave Particle-In-Cell Simulation

they combine many electrons to macro-particles and deposit the macro-particle charges
on a grid to reduce the number of inter-particle force calculations. The coupled
problem is then solved iteratively: From the particle velocities and positions, the
particle charges are projected on the numerical grid. Using Maxwell’s equations, the
electric and magnetic fields on the grid are updated and weighted back to the particle
positions. Then, using the Lorentz force law and the equation of motion, the particle
movement is calculated and the particle positions and velocities are updated. This is
repeated until the specified simulation time is reached.
CST Particle Studio’s PIC solver is a well-known example for such a general
purpose tool. It is regularly used in literature as a benchmark for new simulation
tools or to predict the performance of novel TWT topologies, for instance planar
delay-line topologies [25], [26] or FW-TWTs [27], [28]. In this work, CST’s time-
domain background is especially useful. It enables the study of transient processes
or complex modulation signals, where frequency-domain codes like MVTRAD are
naturally limited. While novel and promising time-domain codes based on similar
parametric information of the slow-wave circuit are under development [21], [29], they
are currently not yet fully validated, but similarly specialized as MVTRAD.
The usage of CST presents several challenges. Apart from the high demands on the
computation side, the tube modeling for accurate results is fairly complex. Having a
precise model and a reproducible modeling approach is a key requirement for using
CST for the simulation of TWTs under modulated excitation. The following section
considers some challenges and approaches and thus draws from the findings previously
reported in [30], [31], where complete analyses of realistic TWTs with CST’s PIC code
have been carried out.

3.2.2 Modeling

In the following, models for two of the most challenging problems are outlined:
the electron beam injection and the severs. The proposed approaches enable a
quick and mostly automated model generation. Therefor, the same input files and
model descriptions as for MVTRAD are used. To show the accuracy obtained with
the shown approaches, CST results for commercial medium power Ku-Band helix
TWTs for satellite downlink applications are compared to those of MVTRAD and to
measurements.

25
3 Simulation Models for Modulated Signals

Table 3.1: Currents from CST for simple geometries and realistic electron gun models,
compared to suitable references.
CST Reference Error (%)
Quadratic plates 65.9 A 65.4 A 0.76
(100 mm × 100 mm × 3 mm) (36 min) Child-Langmuir

Concentric spheres 27.0 A 27.5 A 1.81

(Radii 40 mm, 37 mm) (27 min) Langmuir-Blodgett

Concentric cylinders 0.595 mA 0.598 mA 0.50

(Radii 5 mm, 15 mm, length 10 mm) (30 min) Langmuir-Blodgett

Electron gun model A 83.9 mA 81.3 mA 3.19

(realistic geometry) (2:40 h) Gun2D (Thales)

Electron gun model B 74.1 mA 75.2 mA 1.46

(realistic geometry) (2:45 h) Gun2D (Thales)

Electron Gun Simulation and Direct Beam Calculation

In contrast to a direct computation of the electron gun as a part of the PIC simulation,
gun and delay line are generally considered separately. While this is far more efficient,
it requires the injection of the particles into the interaction region from an artificial
surface. As the operating principle of a TWT obviously relies on the electron beam,
its proper generation is crucial. There are two practicable approaches for this task.
To simulate the electron gun itself, the three-dimensional structure of the electron
gun is directly modeled in CST. Following a particle tracking simulation, this enables
the investigation of various beam properties. The particle emission is based on the
Child-Langmuir emission law, which is defined for the electron beam flow in parallel-
plane diodes [1]. To evaluate the usability of CST for electron guns, both simplified
structures and realistic gun geometries are investigated. For planar, parallel surfaces
and arrangements of concentric spheres, analytic solutions are readily available from
literature [1], [32]. In Table 3.1, these theoretical references are compared to the
simulated results from CST, showing good agreement. For the comparison of the CST
simulation of real gun geometries, the proprietary Thales code Gun2D is used as a
benchmark. This reference tool is much faster, as it takes advantage of the rotational

26
 3.2 Full-Wave Particle-In-Cell Simulation

symmetry of electron guns and runs in 2D, which CST is not capable of. The total
extracted beam current deviation is also reported in Table 3.1.
As an alternative, a direct injection of the beam according to specified beam
properties can offer the desired beam with less effort. In that case, the electron beam is
generated directly on an abstract particle emission surface. For rotating pencil beams,
concentric emission rings are defined. On each of them, the particles are emitted at
a given solid angle and with an initial energy. Alternatively, for full control over the
injected beam, precedingly calculated emission points and initial velocities, e.g., using
Equation (2.1), can be imported into CST. The injection points are distributed on
the beam cross-section such that the charge density is constant. This can be seen
in Figure 3.1a, where each injection point is also marked as the starting point of a
corresponding velocity vector. The base color indicates the initial velocity.
Both with the direct electron gun simulation and the beam calculation, the beam
data is to be recorded, e.g., in an ASCII-file. To include the recorded data into the
PIC computation domain, an adequate interface between gun and interaction region is
required. This is established in the following.

Particle Injection Interface

By default, the particle injection procedure in CST does not consider the static
potential of the emitted beam as a necessary boundary condition at the injection
plane. The macro-particles are simply pushed through an emission surface. Thus, the
missing beam potential alters the outgoing beam, leading to a different beam shape
and velocity distribution and ultimately a deviation in the trajectories.
To compensate for this, an additional boundary condition is included at the interface.
It stems from the assumption of an infinite beam extension in axial direction inside a
beam tunnel. Near the injection plane, the electron beam is not yet modulated. Thus,
the helix can be approximated by a hollow, perfectly conducting cylinder. Therefore,
a rotationally symmetric potential distribution
!
1 rh r2

+ ln − , r < rb ,

r2 
ρ0 b

V (r) = 2 rb 2rb2 (3.2)
20 ln rh ,

r ≥ rb ,
r
is defined in the transverse emission plane, resulting from direct integration of Poisson’s
equation [33], where rb denotes the beam radius and rh the helix radius.
The resulting electrostatic field is generated in a particle-free electrostatic simulation,
with a configuration of concentric rings in the emission plane as shown in Figure 3.1b.
Each ring is fixed at a specified potential, such that the equivalent boundary condition

27
3 Simulation Models for Modulated Signals

0V

Potential
vz
z

Vmin

(a) Distribution of injection points in the (b) Setup for the electrostatic simulation
injection plane with corresponding ini- with a defined number of potential
tial energy vectors. The color denotes rings in the injection plane. The outer
the initial velocity vz in z-direction. hull at zero potential is not shown.

Figure 3.1: Injection of particles and equivalent beam port with potential rings.

from Equation (3.2) is fulfilled. A PIC simulation without including this external field
from the electrostatic simulation would introduce a significant voltage depression in
axial direction. This causes an error in the synchronism between the traveling wave on
the delay line and the electron beam, which directly leads to an error in the predicted
behavior of the TWT. The resulting extended workflow representing the proposed
beam port is summarized in Figure 3.2.

Sever Modeling

Typical severs are, for instance, realized by nanometer-scale material layers applied to
the support rods. Realistic, physical modeling of such a structure is thus not feasible
in CST. Alternatively, equivalent lossy materials are defined to emulate the desired loss
properties. Therefor, in this work, a nonphysical "lossy vacuum" is placed outside of the
helix as the contributor of the virtual equivalent losses. In principle, the helical wire
or the support rods could also be chosen for this purpose. The lossy-vacuum material
is defined with a constant relative permittivity r = 1 to ensure minimal reflection,
but a loss tangent tan(δ) tapered in axial direction. This taper is implemented with a
large number of slices, each with constant tan(δ). This is sketched in Figure 3.3.

28
 3.2 Full-Wave Particle-In-Cell Simulation

PPM Stack
Configuration
Electron Gun Particle Tracking
3D Geometry
Poisson’s Equation Paraxial Approximation

Boundary Interface PIC: Interaction

Potential / Field Particle Injection Simulation
1
OR 0
r / rb

-1
−Vmax 0V
Electron Beam
Properties
Delay Line
Poisson’s Equation 3D Wire Model,
Couplers, Severs, ...
Particle Properties

Figure 3.2: Workflow for the beam port inside the PIC interaction simulation in CST
Particle Studio.

Lossy "vacuum"
Hull

Vacuum

Sever length

Support rods

Figure 3.3: Equivalent loss mechanism model. The transparency of the sever slices
represents the loss profile. The losses per slice decrease symmetrically
from the center of the sever towards the ends.

29
3 Simulation Models for Modulated Signals

The loss profile α(z) of the sever, e.g., from measurements, is translated into a
material distribution
(z) = 0 r [1 − j tan(δ(z))] , (3.3)
such that the absorption is substituted accurately. For this purpose, a perturbational
approach is adopted to find a relation between loss and material profile. Therefor,
several assumptions are made. Most fundamentally, one assumes that the introduc-
tion of losses does not relevantly alter the field distribution. This allows the initial
calculation of field properties without considering lossy materials, and significantly
reduces the computational effort. As a consequence, the result’s validity is limited to
a region around the initial field calculation at a reference value tan(δ0 ). The extent
of this region depends on the considered scenario, the degree of simplification in the
approach and on the required accuracy of the translation.
A similar fundamental approach can be found in literature [34], where the con-
ductivity losses of the helix wire instead of the losses in the sever are considered. It
assumes that introducing the lossy material does not affect the field distribution itself
and thus can be based on results of an unperturbed eigenmode simulation. In case
of [34], this assumption obviously holds, as the losses in general are small. Here, it
might be violated in the high-loss center of the sever. Fortunately, experience with
MVTRAD shows, that such inaccuracies are negligible compared to the accuracy at
the low-loss outskirts of the sever. For well-matched severs, very smooth transitions
are important. Thus, the transition from loss-less to highly lossy is rather long. Along
this transition, the fields do not change strongly.
In addition, at the sever positions, the beam is not yet strongly modulated, meaning
that the fill-factor is not much different from the one at the injection. Like in MVTRAD,
one can assume, that it suffices to consider the field properties on the beam axis. This,
of course, also holds for the lossy regions.
For the slices, the surface and volume losses, which can be calculated separately,
can be summarized by the total dissipated power PL . Surface losses, e.g., due to the
wire material, are assumed to be invariant in axial direction and small in comparison
to the volume losses, which stem from the complex permittivity (z). By means of PL ,
the total stored energy W can be related to the quality factor Q, which describes the
energy dissipated during one period t0 , through
!
2πW 2πf
Q(δ) = ⇒ W = W0 exp − t , (3.4)
PL (δ)t0 Q(δ)

weighted by an arbitrary initial energy scale W0 . In a homogeneous slow-wave structure

segment, the differential energy change along the axis can be related to the energy

30
 3.2 Full-Wave Particle-In-Cell Simulation

Scattering parameters (dB) 0 2

A
B 1

GSS /GSS,max (dB)

−20 C
0
−40
−1 A
−60 B
−2 C
D
−80 −3
0.9 0.95 1 1.05 1.1 0.9 0.95 1 1.05 1.1
Frequency f /f0 Frequency f /f0

(a) Performance of a sever with 75 dB total (b) Influence of the sever onset on the small-
attenuation from the perturbational ap- signal gain. (A): Reference TWT, (B):
proach. (A): reflection S11 , (B): transmis- steeper transition of the loss taper, (C):
sion S21 , (C): time-gated reflection S11 . smaller total attenuation, (D): larger to-
tal attenuation.

Figure 3.4: Sever performance and sensitivity to sever profile and total attenuation.

change over time and the energy transport velocity, which is equivalent to the group
velocity vg .
Assuming an exponential decay of the fields along the axis with exp (−αz) we can
define the attenuation constant α. Hence, with P ∝ |E|2 , the power decays according
to exp (−2αz). Together with Equation (3.4) and a group velocity based substitution
from time derivatives to spatial derivatives, a translation between material and loss
profile is
PL (δ) PL (δ) − PL (δ)
δ=0 δ=δ0 δ=0
α(z) = + tan(δ(z)), (3.5)
2W vg 2W vg tan(δ0 )

which can be applied in an eigenmode simulation as part of a pre-processor routine.

This expression is rather simple, reducing the possibly complex relationship between
α and  to a very manageable one. This is a direct consequence of the manifold
assumptions made. As discussed before, especially for the high-loss center, this might
at first seem worrisome, and thus requires thorough consideration. A particle-free
simulation reported in Figure 3.4a shows deviations of the total attenuation of around
3 dB for a sever with 75 dB loss, indicating that Equation (3.5) is still well suited. It
leads to a smooth transition between interaction region and sever, which yields input
reflections of less than −40 dB throughout the band.

31
3 Simulation Models for Modulated Signals

Although the total attenuation is represented more or less correctly following above
procedure, in the sever it is not its most important property. In comparison, Figure 3.4b
shows that the profile at the sever onset is far more critical, as it effectively changes
the electric lengths of the sections. Here, the small-signal gain of a reference TWT
including two severs (curve A) is compared to that of the same TWT with a different
loss profile on the second sever. This other loss profile relates to a typical simplification
of the shape as utilized in proprietary tools. The small-signal gain is normalized to the
maximum small-signal gain of the reference TWT. A steeper transition (curve B), and
thus a different profile at the onset, can easily result in an error as large as if the total
attenuation were chosen 50 dB smaller (curve C). This is mainly due to the changed
effective electric length of each section. A 50 dB larger total attenuation (curve D)
does not have a significant influence on the gain. As the losses are small close to the
onset, the perturbation approach is accurate enough there and good agreement in
terms of dissipated power can be expected.

3.2.3 Model Validation and Simulation Accuracy

To validate the presented modeling approaches, a variety of simulation runs and
comparisons have been carried out. In the following, simulation results for a Ku-Band
helix TWT are shown. A comparison of CST simulation with measurement results
and the proprietary code MVTRAD is carried out. The CST model is set up in such
a way, that it uses exactly the same input files describing setup and geometry as
MVTRAD. The 150 W Ku-Band TWT considered here has a tapered, multi-section
helical delay-line. Figure 3.5 shows the interaction region model as generated and
simulated in CST.
For validation against measurements, the measured TWT needs to be known in as
much detail as possible. In MVTRAD the phase velocity and the coupling impedance
can be fitted to measurements directly. Fast run times enable the identification of
unknown parameters by, for instance, parametric sweeps. In contrast, a model fitting
in CST requires a physical reasoning. Thus, geometric measurements, such as the wire
cross-section and shape, and also the contact between wire and support rods, are used
in the modeling procedure to minimize deviations.
By this, mechanical deformations and similar imperfections, which significantly
influence the phase velocity among other properties, are taken into account. While
the sever profile near the onset is measured and translated into a material profile to
be used in CST, the total attenuation is only an estimate, as the measurement of its
exact level in the high-loss center is quite involved.

32
 3.2 Full-Wave Particle-In-Cell Simulation

Figure 3.5: Full model of the interaction region with the tapered helix wire, two
severs, support rods, and discrete ports at in- and output.

Information on the measured magnetic focusing field is imported on the beam axis
and the magnetic field at the cathode is estimated from simulation. A large-signal
convergence study returns around 200 injection points for accurate discretization of
the electron beam. Also, the support rod permittivity is fine-tuned to adjust for other
uncertainties such as material tolerances and to correct the phase velocity, resulting
in a 2 % permittivity deviation from the nominal value. This is well within typical
material deviations for the support rods [35].
Finally, precise full-wave simulation requires a high mesh density. It is initially
investigated in particle-free simulation as an involved convergence study including
the electron beam is not feasible due to long computation times. The convergence is
assessed with respect to the electric field on the axis. Discrete ports are chosen to
couple energy into or out of the interaction region. They correspond to simple and
well localized current sources between the helix wire and the outer hull of the TWT.
In contrast, a realistic and well matched coupler geometry requires long structures
perpendicular to the axis. This results in a large number of additional cells in the
structured mesh over the whole length and therefore lowers the number of emission
points that can be used in the PIC simulation, due to the computation resource
competition of the particle and mesh-cell number.
Figure 3.6 shows a number of simulated properties from CST and MVTRAD and
the respective measurements. MVTRAD assumes perfectly matched couplers at input
and output. For Figure 3.6a, a chirp signal is used to reduce the computation time
from roughly half a day per frequency point to less than a day for the whole frequency
band. The gain is slightly overestimated by around 0.5 dB both in MVTRAD and CST,
which can be explained from the generally high sensitivity of the gain as outlined above.
A corrective shift of the gain curve shows that MVTRAD slightly underestimates the

33
3 Simulation Models for Modulated Signals

0 −2

Pout /Pout,max (dB)

G/Gmax (dB)

−4 Meas., 0.9f0
−2 CST, 0.9f0
−6 Meas., f0
MVTRAD CST, f0
CST (chirp) −8 Meas., 1.1f0
−4
Measurements CST, 1.1f0
−10
0.9 1 1.1 −15 −10 −5 0
f /f0 Pin /Pin,sat (dB)

(a) Small-signal gain versus frequency. (b) Output power versus input back-off.
Nonlinear phase shift (deg)

MVTRAD 1
Cumulated current I/I0

60
Measurements
CST 0.8 −15 dB
40 1.1f0 0.6

f0 0.4 −5 dB
20
0.2 MVTRAD
0.9f0 CST
0 0 0 dB
−15 −10 −5 0 0 0.2 0.4 0.6 0.8 1 1.2
Pin /Pin,sat (dB) Kinetic energy Ekin /Ekin,0

(c) Nonlinear phase versus input back-off. (d) EVS at three different back-off levels.

Figure 3.6: Comparison of different properties of the TWT from simulation in CST
and MVTRAD with measured results.

bandwidth, especially visible at the lower band edge. This is not the case for CST,
where the deviation is approximately constant.
As the small-signal gain computation exhibits a slight offset, the gain compression
and the nonlinear phase shift shown in Figure 3.6b and Figure 3.6c are normalized to
the respective saturated input powers. They are plotted versus input back-off at the
band edges and at the center frequency. The error in saturated output power is within
0.15 dB at the edges and significantly smaller in-between. Similar agreement can also
be seen for the nonlinear phase shift. The simulated values from MVTRAD and CST
are close to the measured ones, with a maximum error of below 3◦ .

34
 3.2 Full-Wave Particle-In-Cell Simulation

0.8 MVTRAD 0.8 MVTRAD

CST CST@t1 ..t4
0.6 0.6
r/rh

r/rh
0.4 0.4

0.2 0.2

0 0
z→ z→
(a) Zero-drive excitation. (b) Saturation.

Figure 3.7: Comparison of macro-particle trajectories at f0 from CST and MVTRAD.

Figure 3.6d compares the electron energy spectrum at the output of the interaction
region. Here, the accumulated current is plotted against the kinetic energy of the
macro-particles for small-signal excitation, for saturation, and at an intermediate power
level. Especially at high energies, the results from CST and MVTRAD are close. The
energy of the slowest particles, which typically determines the lowest collector stage
potential, is around 1 % to 3 % smaller in CST than in MVTRAD. As the electron
beam spectrum is not easy to measure directly, it is not clear whether the CST or
MVTRAD results are closer to reality. This is also the case for the particle trajectories
shown in Figure 3.7, which visualize the particle motion inside the interaction region
with a PPM-stack for focusing. For this purpose, the number of injection radii is
reduced to four. Figure 3.7a and Figure 3.7b show the trajectories at zero-drive and
at saturation, respectively. The trajectories are almost identical at zero-drive. For
the high-drive case, several snapshots of the particle distribution are taken within
one RF-period. They are compared to the beam envelope extracted from MVTRAD.
Except for some peaks visible from the CST results, the general shape of the beam is
represented accurately.
Overall, the CW interaction simulation results, i.e., gain compression, nonlinear
phase shift, and the EVS agree well with measurements and an MVTRAD model with
directly fitted parameters. The shown efficient approaches for the different parts of a
full-wave helix TWT simulation in CST lead to models which are usable for the study
of multi-frequency simulation in later sections of this work, especially Section 3.4.

35
3 Simulation Models for Modulated Signals

3.3 Envelope Models

In this work, methods, in which any output quantity is calculated based on the signal
envelope or related input signal characteristics, are called Envelope methods. TWT
models in which Envelope methods are utilized to estimate the amplifiers response
are so-called Envelope models, which enable the combination of frequency-domain
and time-domain approaches to simulate the behavior of the TWT when excited with
arbitrary time signals. The general idea behind Envelope methods is that for any small
time interval, the TWT can be considered operating in a short-term steady state. Thus,
for each of these intervals, one steady-state response can be found. This steady-state
response can, for instance, be extracted from a frequency-domain calculation or a CW
measurements. This way, the full signal can be integrated fast in time-domain.
In simplified terms, the envelope of a signal is a curve following the outlines of the
signal’s oscillation through its extremes. Regarding a modulated carrier

x(t) = A(t) cos (ω0 t + ϕ(t) + ϕ0 ) , (3.6)

with the carrier frequency ω0 , the non-negative modulated amplitude A(t), the phase
modulation ϕ(t), and an initial phase shift ϕ0 , the complex representation or complex
envelope form [36]
n o
x(t) = Re A(t)ejϕ(t) ejω0 t+ϕ0 (3.7)

describes the signal by separating carrier and modulation2 . Following this represen-
tation, the complex carrier ejω0 t+ϕ0 is modulated in time according to the complex
envelope function A(t)ejϕ(t) .
In simple Envelope method based approaches, a complex-valued transfer function
describes amplitude compression and phase distortions caused by the nonlinear device.
A set of amplifier transfer characteristics between input and output of the TWT, such
as the AM-AM and AM-PM curves, can most conveniently be used as a reference for
the model synthesis. This is well established and has led to various simple and fast
approximate TWT descriptions, which work reliably under certain constraints. The
most crucial requirement is that the time-domain input signal varies significantly slower
than any transient effects and time constants in the TWT. Otherwise, one could not
assume a steady-state response along the signal envelope. The TWT’s time constants
mainly stem from its electrical length and the retroactive length of the electron beam.
Thus, they are at most in the order of a few nanoseconds. Furthermore, the dispersive
properties of the TWT need to be negligible within the considered frequency range of
2
Although x(t) is still not a complex signal this is the typical terminology found in literature.

36
 3.3 Envelope Models

the signal. This is especially true for more simple descriptions of the Envelope method,
as these do not include any frequency-dependent properties. As helix TWTs developed
for communication systems are only marginally dispersive in practical frequency bands,
this complies in many cases.
In the following, an overview over different frequency-dependent and -independent
methods is presented, together with selected implementation details and comparisons.

3.3.1 Analytic Methods and Simple Two-Tone Estimates

Classically, the multi-tone performance of a TWT is described in terms of typical
two-tone characteristics. The spectral components at the excitation frequencies f1 and
f2 in the output signal relate to those that arise in the TWT with nf1 + mf2 > 0, for
(n, m) ∈ Z2 \ {(0, ·), (·, 0)}. These components are called intermodulation products and
are a result of the nonlinearity as described in Section 2.3. They can be interpreted as
distortions of the excitation and are desired to be small for TWTs used for instance in
satellite communications.
One well-established and simple characteristic is the intermodulation product ratio
i
D2n+1 . It can be calculated from
!
{1,2} P2 (f1,2 )
D2n+1 = 10 log10 , (3.8)
P2 (f1,2 + (−1){1,2} n ∆f )

where ∆f = f2 − f1 > 0 is the spacing between the inserted tones (and therefore
determines the intermodulation ratio) and P2 (fi ) is the output power at the frequency
fi . In this section, only narrowband scenarios are considered, such that the TWT’s
1 2
dispersion can be neglected. Thus, with symmetric excitation, D2n+1 = D2n+1 holds,
as the spectrum at the output then also is fully symmetric.
There are several ways to estimate the intermodulation product ratio from the CW
characteristics of the TWT. In the following, an operator M describing the TWT is
considered, such that the output signal is given by

y(t) = (M ◦ x)(t). (3.9)

For a two-tone input signal

xω0 ,ω0 +∆ω (t) = A cos(ω0 t) + A cos((ω0 + ∆ω)t + ϕ0 ) (3.10)

with equal amplitudes A, an arbitrary initial phase ϕ0 , and a frequency spacing of

∆ω  ω0 , in principle a common period T is not always available. Still, for the non-
dispersive case, one can assume that one can equivalently consider a signal xω̂0 ,ω̂0 +∆ω

37
3 Simulation Models for Modulated Signals

with ω̂0 ≈ ω0 and ω̂0 = p∆ω, p ∈ N. By this, the signal xω̂0 ,ω̂0 +∆ω is periodic with

2πp
T = (3.11)
ω̂0
and
P2,ω0 (ω0 + n∆ω) ≈ P2,ω̂0 (ω̂0 + n∆ω) (3.12)

holds, where P2,ωi (ωj ) is the output power at ωj after M , when the excitation is
xωi ,ωi +∆ω and n ∈ [−p; ∞[. Using the signal’s periodicity, the output signal can be
found by a Fourier series expansion

yωi ,ωi +∆ω (t) = (M ◦ xωi ,ωi +∆ω ) (t) (3.13)

∞
! X
= [an cos (n∆ωt) + bn sin (n∆ωt)] , (3.14)
n=1

with coefficients
T
2Z
an = (M ◦ xωi ,ωi +∆ω ) cos (n∆ωt) dt, (3.15)
T
0
ZT
2
bn = (M ◦ xωi ,ωi +∆ω ) sin (n∆ωt) dt. (3.16)
T
0

The respective intermodulation ratio results in

 
(i+1) a2p+i + b2p+i
D2n+1 = 10 log10  , (3.17)
a2p+i+(−1)i+1 n + b2p+i+(−1)i+1 n

with i = 0 for the lower and i = 1 for the upper intermodulation ratio.
While this description in principle also holds for any non-static operator, considering
a purely static gain enables describing M by simple series expansions. In such a
static gain description, the output signal y(t) instantly follows the input signal x(t).
Consequently, there is no phase at all, as phase can only be defined by frequency or
time evolution. This is a rather strong constraint, but it allows, for instance, to extend
M by terms of a Taylor series. That would lead to expressions for the intermodulation
products solely based on the parameters used to fit the Taylor series to the static gain
function and the amplitude of the input tones.
In [37], a non-static description is chosen for the two-tone estimation, expanding
around the operating point of the applied signal. In the case of small signal amplitudes
compared to saturation, this enables the approximate calculation of certain TWT
output quantities. Also, phase information is included in the model.

38
 3.3 Envelope Models

With a differential gain compression

Pin (f ) dPout (f )
c(f ) = 1 − , (3.18)
Pout (f ) dPin (f )
and the differential AM-PM conversion coefficient
dΘ(f )
kp (f ) = Pin (f ) , (3.19)
dPin (f )
with the nonlinear phase shift Θ, the third order intermodulation ratio for two equally
strong tones can be approximated by [37]
!
c2
D3 ≈ −10 log10 + kp2 . (3.20)
4
Another investigated quantity is the so-called phase-transfer or AM-PM transfer
factor kT . It characterizes the effect of changing the input power level of a strong
signal at one frequency on the phase of a second, weaker carrier at another frequency
according to
dΘ(f2 )
kT (Pin ) = Pin (f1 ) , Pin (f1 )  Pin (f2 ). (3.21)
dPin (f1 )
From [37], an approximation for kT can be found with
  
d   kp 
kT (Pin ) ≈ kp + Pin arcsin  r  . (3.22)
dPin   
c
2 
1− 2
+ kp2

3.3.2 Static-Gain Curve Model

In the past, straightforward two-tone approximations as given in Section 3.3.1 have
shown to be reliable compared to measurements [37], [38]. They can be used to
effortlessly characterize the degree of nonlinearity in limited frequency ranges, based
on typical quantities which are anyway considered in the design and measurement of a
TWT. Unfortunately, it is not clear how meaningful simple two-tone characteristics
really are for real modulated signals in narrow or broad frequency bands. Therefore, a
model enabling the direct integration of an arbitrary signal is desired.
To derive the two-tone expressions, the TWT is assumed to be frequency-independent
within a selected spectral range. Thus, a simple non-dispersive nonlinearity is consid-
ered for the amplifier model, which in principle can be used in a more generalized setting
with any excitation in time-domain, especially complexly modulated communication
signals. In literature there are several approaches using the frequency-independent,
nonlinear gain compression and phase shift to estimate the intermodulation products

39
3 Simulation Models for Modulated Signals

for multi-tone signals, e.g., the IMAL code [39]. In this thesis, the approach is used in a
slightly modified way to generate the output signal as well as specific TWT properties
like the EVS and various communication characteristics.

Model Realization

In the following, an implementation of this method is described, which also serves

as a base for the nonlinear description of the frequency-dependent models in the
later sections. Thus, the following description is also relevant for the methods in
Section 3.3.3, Section 3.3.4, and Section 3.3.5.
The goal is to include the developed simulation codes during the design process,
for instance for fast evaluation of, e.g., a helix pitch profile in a delay-line taper
optimization routine. Therefore, the model identification scheme needs to be highly
efficient. The nonlinear and frequency-dependent performance is obtained from steady-
state simulation codes such as MVTRAD, where the reference AM-AM and AM-PM
curves are generated within minutes. The CW gain and phase are calculated in
input-power intervals of ∆Pin and interpolated when needed.
In this method, the input signal x(t) is assumed to be a single-tone oscillation, the
amplitude of which changes slowly over time, in accordance with the constraints of the
Envelope method. This time-dependent amplitude can be extracted from the signal
itself by means of the Hilbert transform [40], [41]
∞
1 Z x(τ ) 1
H{x}(t) = dτ = x(t) ∗ , (3.23)
π t−τ πt
−∞

which is used to construct the so-called analytic signal

H{x}(t)
q !!
xinst (t) = x2 (t) + H{x}2 (t) · exp j arctan . (3.24)
x(t)

The Fourier transform of this description is purely single-sided. Thus, the imple-
mentation can be carried out by discrete-time Fourier transform, where the negative
frequency components are suppressed [42]. The analytic signal can be split into an
instantaneous phase term n o
ξ inst (t) = arg xinst (t) (3.25)
and an instantaneous amplitude

Ainst (t) = xinst (t) , (3.26)

which corresponds to the envelope of the original signal x(t).

40
 3.3 Envelope Models

The output signal y(t) can then be calculated by this static-gain curve (SGC) model
to
    
y(t) = gAM xinst (t) · exp jξ inst (t) − ΘPM xinst (t) , (3.27)

where gAM (A) 6= fct.(ω) describes the input-amplitude dependent output-amplitude

characteristics and ΘPM (A) 6= fct.(ω) the nonlinear phase shift of the amplifier.
Terminology-wise, naming this approach a static-gain curve model is not entirely
correct. In contrast to the static-gain description in Section 3.3.1, the usage of a
Hilbert transform and analytic signal description leads to an inherently memory-based
model, where the linear and nonlinear phase information is taken into account. Still,
for the sake of consistency with [43], [44], the name is maintained for this work.

Frequency Reference

The general idea of the SGC model is to assume non-dispersive behavior over a certain
time-interval. Thus, a single-frequency model can determine the result according to
Equation (3.27). A question that remains is the choice of the frequency which is best
for the considered interval. In principle, a variety of choices is conceivable for the
reference frequency. One obvious solution is the center frequency

fmax + fmin
f ctr = (3.28)
2
of the investigated signal, where fmax and fmin are the upper and the lower boundary of
the power-carrying frequency range. In a complexly modulated signal, this would also
be the carrier frequency, although there might be better choices, depending on where
the carrier is located in the amplification band. Alternatively, the reference frequency
could be some kind of dominant frequency, which does not necessarily need to be in
the center of the band. This could be useful when, e.g., two tones are considered with
strongly asymmetric power distribution, or when one tone is at a frequency where
significantly less gain is to be expected.
Equation (3.25) introduces the instantaneous phase. From its derivative, a quasi-
instantaneous frequency f inst

1 dξ inst (t)
f inst = (3.29)
2π dt
can be calculated, which represents a changing effective frequency over time.
A comparison of using this more involved frequency-identification process and
the simpler dominant-frequency reference has been carried out. Figure 3.8 shows
exemplary results. The used reference Ku-Band TWT was designed to be more

41
3 Simulation Models for Modulated Signals

3
f ctr , q = 1 : 9
f inst , q = 1 : 9
Error ∆Pout (dB)
f ctr , q = 3 : 7
2
f inst , q = 3 : 7
f ctr , q = 1 : 1
f inst , q = 1 : 1
1

0
−2 0 2
2(f − fc )/∆f

Figure 3.8: Comparison between the two reference frequency decision schemes for a
large-signal simulation at ∆f = 500 MHz around the amplification band
center fc . The error is compared to reference simulation carried out in
MVTRAD.

dispersive than a typical communication TWTs, to highlight the effects. The figure
shows three different constellations for two tones, which are ∆f = 500 MHz apart,
centered around the amplification band center at fc . Apart from a symmetric excitation,
two asymmetric cases are presented, with q describing the ratio of the powers at the
two input frequencies. The simulation is carried out at a large-signal operating point
near saturation of the band center. Thus, the presented scenarios are challenging for
any Envelope model, as they cover large frequency ranges at a critical power level.
For asymmetric excitation the usage of f inst shows advantages compared to the
center frequency. For q = 1 : 9, the picture is not completely clear, as the results for
the weaker main tone and the upper intermodulation product are slightly worsened by
the instantaneous frequency choice, but the lower intermodulation product, which was
far off before, is significantly improved. An improvement over the whole band can be
seen for q = 3 : 7, where for all considered frequencies the instantaneous frequency
selection shows a 0.1 − 0.3 dB lower error. However, for the symmetric excitation, no
significant difference can be identified. This can be explained by the progression of the
instantaneous frequency itself. For a two-tone signal with amplitudes A1 and A2 at
the frequencies f1 and f2 , the instantaneous frequency can be calculated to [45], [46]
1 A2 − A21
f inst = (f2 + f1 ) + (f2 − f1 ) 2 , (3.30)
2 Â2 (t)
with q
Â(t) = A21 + A22 + 2A1 A2 cos (2π(f2 − f1 )t). (3.31)

42
 3.3 Envelope Models

Thus, for symmetric excitation, the instantaneous frequency equals the center frequency,
such that no difference is observed.
Although the usage of f inst offers some advantages in certain cases, it needs to be
handled with care. For instance, when the instantaneous phase has discontinuities
or rapid variations, the changes in the resulting frequency could possibly be too fast
to comply with the fundamental constraint of slow signal variations. Also, for many
communication signals, the instantaneous frequency would oscillate closely around or
be approximately equal to a dominant frequency component, e.g., the center frequency.
Thus, for the following investigations in this thesis, when no further comment is given,
the center frequency approach is chosen as the reference for the model, as it is easy to
define for all scenarios and it matches the dominant frequency for equal-amplitude
two-tone characteristics.

Interpolation Schemes

The reference data, no matter whether extracted from measurements or simulation,

is necessarily limited to a number of sample points. A single-frequency SGC model
refers to sampled data in ∆Pin steps. In multi-frequency models, as the instantaneous
frequency SGC model, additional sampling dimensions must be taken into account.
For the model usage an interpolation scheme is required. Different interpolation
schemes have been tested. Linear interpolation is the most simple choice, but results
in non-differentiable curves and introduces numerical noise into the calculation [47]. In
fact, the error is proportional to the squared distance between sample points [48], so it
imposes more demanding conditions on the reference data. While this is manageable
for the single-frequency SGC model, already for the instantaneous frequency SGC
model much more data is required. This issue further increases with model complexity.
Thus, other, higher order interpolation schemes are beneficial. For oscillating curves
cubic spline interpolation [47] typically offers good results. Here, polynomials are fitted
in-between the sampling points, such that not only the resulting interpolated curve,
but also the first and second derivative are continuous. In contrast, using a monotone
piecewise cubic interpolation [49] does not result in a continuous second derivative.
While overshoots and unwanted oscillation might occur in the spline scheme, this
is not the case for the piecewise cubic interpolation, at the cost of possible second
derivative discontinuities and curvature steps. For the models in this thesis, in most
cases piecewise cubic interpolation is preferred and used for both drive- and frequency-
based interpolation, as the behavior of the interpolated curves can be considered

43
3 Simulation Models for Modulated Signals

non-oscillating, and overshoots, e.g., between frequency points, hurt far more than any
discontinuity in the second derivative.

Simulation of Effective Efficiencies and Collector Currents

In principle, various output quantities of the TWT can be estimated using Envelope
methods. Apart from the output waveform and related characteristics on the signal
side, the efficiency itself is of high interest, as its optimization is a key factor to improve
TWTs in back-off. As noted before, the envelope of a modulated signal changes over
time. Thus, the effective short-time operating point moves across the transfer curves
in time. It spans across a range of input power levels, for which an effective3 efficiency
is to be determined.
According to Section 2.4, the total efficiency of a TWT can be calculated from
the DC power inserted into the system, the beam efficiency, and the regained electric
power in the collector. While in the CW case, these quantities can be described by
static values over time, the effective values for modulated operation are weighted
time-averages. The DC input power as well as the collector stage potentials are fixed
known TWT parameters, and the effective beam efficiency can be calculated from the
output signal in an Envelope approach as previously described.
In contrast to what can sometimes be read in literature, neither effective collector
efficiency nor the effective total efficiency can be directly calculated by averaging the
respective steady-state efficiencies with the probability of occurrence of the respective
input power level over time. This is, because the collector efficiency relates the regained
electric power to the total power entering the collector. Both quantities in this ratio
change over time according to the input signal envelope distribution, which forbids
a direct envelope averaging scheme, e.g., using the SGC model, on (Pin , ηColl ) as the
reference data. Similarly, this holds for the total efficiency. Instead, a detour over
other quantities, for instance the currents in the individual collector stages, has to be
taken.
In Figure 3.9, the estimation of the effective total efficiency is compared to measure-
ments4 . In Figure 3.9a, the effective currents of a four-stage collector are calculated
by SGC models of the individual collector stages. The SGC models relate the effective
3
In the following, an effective metric describes the expected time-averaged metric at the output of
the TWT from a modulated input signal.
4
The definition of the term input back-off is not consistent in literature. In this work, the input
back-off levels are defined by Pin /Pin,sat . This may feel wrong from a linguistic point of view, but
enables presenting the results in the most intuitive way, with low powers to the left and high
powers to the right. The same holds for the output back-off.

44
 3.3 Envelope Models

0.8 A
Collected currents I/I0
60

Total efficiency (%)

B
C
0.6 C4
40
0.4
C3 A
20 B
0.2 C2 C
C1 D
0 0
−20 −15 −10 −5 0 5 −20 −15 −10 −5 0 5
Input back-off (dB) Input back-off (dB)

(a) Collector currents at the stages C1 to C4. (b) Total efficiency.

Figure 3.9: SGC simulation of collected currents and the total efficiency for 16QAM
signals based on measured characteristics. (A): CW measurements. (B):
16QAM measurements. (C): 16QAM simulation with SGC models of the
collector stages. (D): 16QAM simulation with an SGC model of the total
efficiency as the output characteristic.

currents to reference CW measurements of the collected currents per stage over a

range of input back-off levels, from small-signal to overdrive operation. For a random
16QAM signal, the resulting effective currents per stage are calculated and compared
to measured results for a 16QAM signal excitation. As can be seen, the agreement is
excellent over most of the considered power level range, with some deviation visible
on the lower end. On the right side, in Figure 3.9b, the total efficiency as calculated
from the currents in Figure 3.9a is compared to 16QAM measurements of the total
efficiency. It can be seen, that in contrast to a direct estimation based on the total
efficiency as a reference, the detour allows accurate estimation of the effective total
efficiency in modulated operation of the TWT.

Effective Electron Velocity Spectrum Simulation

In Section 3.3.2, the effective efficiency is estimated based on measured collector

currents. In simulation, the currents must be calculated first from the spent-beam
EVS. As can be seen from Figure 2.3, the spectral properties vary strongly with drive,
such that the collector stage potential choice significantly influences the collector
efficiency, and thus the total efficiency. Therefore, for fully optimizing a TWT in
modulated back-off, it is worthwhile knowing the effective EVS. In Figure 3.10, the

45
3 Simulation Models for Modulated Signals

estimation approach for this effective spectrum is shown for an exemplary two-tone
signal. For each drive level, the spent electron beam from steady-state simulation in
MVTRAD can be evaluated regarding its velocity spectrum, resulting in an array of
curves as shown in Figure 3.10c. Taking the full envelope spread into account, the array
can be used as a reference for a two-dimensional SGC model. With this, the weighted
time-averaged macro-particle energy distribution can be calculated, which can be seen
in Figure 3.10d. For comparison, the spent beam from a direct physics-based two-tone
solution in MVTRAD is added. It can be seen, that the SGC simulation based on
MVTRAD CW results enables accurate estimation of the modulated EVS.

Model Schematic and Signal Flow

The resulting hybrid SGC model is summarized in Figure 3.11. The input signal
x(t) is evaluated regarding its frequency components, which is used for the frequency
reference choice. The TWT’s steady-state information is generated in ∆Pin steps for
one or multiple frequencies, e.g., in MVTRAD. After interpolation, the respective
SGC models for the output signal, EVS or other desired properties, for instance the
i-th collector stage current Ici , are generated. From the Hilbert transform of the signal
x(t), the preprocessed analytic signal xinst (t) is calculated, which is inserted into the
SGC models to calculate the output quantities.

3.3.3 Frequency-Dependent Model Topologies

As previously established, the dispersive behavior of the TWT stems from that of the
delay line. By means of tapering techniques, geometric adaptions of the line, or even
special topologies, the "dispersiveness" can be mitigated, such that for classical, simple
modulations on small bandwidths (compared to the frequency-domain variations of
the amplifier), frequency-independent Envelope methods are expected to suffice. Still,
the limits of such methods are not inherently clear. Especially future topologies for
higher frequencies, e.g., FW lines, possibly impose higher requirements to tools for
estimating the broadband behavior of the TWT.
But already today, in the case of more broadband excitation, the fundamental
Envelope approaches have inherent limitations in terms of the usable bandwidth to
avoid dispersion-based distortions of the resulting signal. Thus, in the past, various
more involved approaches have been formulated [50]–[52]. They offer the inclusion
of the frequency-dependence, which is inherent to the beam-wave interaction, at the
cost of strongly differing model structures and more complex implementation and
usage. As the general idea behind this work is to generate models that can be used,

46
 3.3 Envelope Models

2.00 0.25

0.20
Amplitude (a.u.)

1.00

Probability
0.15
0.00
0.10
−1.00
0.05

−2.00 0.00
0 1 2 3 4 5 −30 −20 −10 0 10
Time (ns) Normalized envelope (dB)

(a) Modulated two-tone signal. (b) Amplitude distribution occurrence prob-

ability.

1.00 1.00
SGC
Cumulated current I/I0

Cumulated current I/I0

0.80 0.80 MVTRAD

0.60 0.60

0.40 Pin 0.40

0.20 0.20

0.00 0.00
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
Ekin /Ekin,0 Ekin /Ekin,0

(c) CW EVS for various power levels. (d) EVS from the two-tone signal.

Figure 3.10: Estimation of the velocity distribution of the spent-beam current.

47
3 Simulation Models for Modulated Signals

Frequency selection Geometry / settings

fref. , . . .

X(f )
e.g., MVTRAD - Parametric, ω-domain, PIC approach

Pout (n∆Pin ) e.g., Ici (n∆Pin )

ϕ(n∆Pin ) I(Ekin , n∆Pin )
Fourier / f inst (t)
x(t) Hilbert Interpolation: Piecewise cubic polynomials
transform
Pout (Pin )
e.g., Ici (Pin ) I(Ekin , Pin )
ϕ(Pin )
xinst (t)

SGC Miscellaneous SGC

waveform SGC electron spectra
model model model

y(t) i Ieff (Ekin )

e.g., Ic,eff , Pcoll,eff

Figure 3.11: Schematic of the hybrid SGC model.

48
 3.3 Envelope Models

Serial/Polar model

x(t) AM-PM / ϕ(t) AM-AM / g(t) y(t)

Figure 3.12: Serial model topology.

e.g., already at the design stage, possibly for optimization of a TWT, not all models
found in literature are suitable. Most models are developed with the aim to accurately
depict the behavior of a known amplifier, for instance for a system analysis containing
the fixed amplifier, or for the design of peripherical components, e.g., linearizers or
filters. For these applications, today, there basically are no significant limits to the
complexity of the fitting process or the reference data to generate the model. One
could for instance use measured, complexly modulated signals to ensure that the model
will precisely predict this reference data.
A selection of models from literature is described in [53]. The models can roughly be
sorted into three categories: models in serial structure (as shown in Section 3.3.2 and
Section 3.3.4), models in quadrature structure (as will be described in Section 3.3.5)
and models which do not fit into the two categories. The model type selection and
terminology is adapted from and inspired by [53], where approaches for the modeling
and simulation of nonlinear systems are outlined.
Serial or polar models are following the most intuitive order to describe the amplifier,
as shown in Figure 3.12. In principle, the nonlinearities of the transfer curve can be
considered as two seperate phenomena. The first one represents the phase distortion,
i.e., the AM-PM effect, characterizing the phase shift observed when driving the
amplifier to a certain operating point. The second one describes the amplitude
compression, i.e., the AM-AM effect. In a serial model, those two mechanisms are
directly applied in one way or another. One example is the frequency-independent
model as given in Section 3.3.2. There are certainly more complexly structured models
available in literature, following the same serial topology.
A well known but sometimes fairly involved group including frequency-dependent
components is the Multi-Box serial model family. In these models, individual effects
are summarized by separate functional blocks, which are connected to form the
full model. As most polar models, they inherently assume a shape- and therefore
functional similarity of the respective AM-AM and AM-PM transfer curves over
frequency. In many models, when merging the dispersive behavior of the TWT with
the drive-dependent characteristics into one representation, the two effects are still

49
3 Simulation Models for Modulated Signals

calculated separately and concatenated afterwards. As a result, with the analytic

signal representation x0 (t) and the corresponding instantaneous phase ξ0 (t) of the
input signal x(t) (as defined by Equations (3.24) to (3.26)), the output signal of the
TWT can typically be described as
y(t) = gAM (|x0 (t)|) · exp (jξ0 (t) − ΘPM (|x0 (t)|)) . (3.32)
with
(ω)
gAM := gAM (ω, A) (3.33)
and
(ω)
ΘPM := ΘPM (ω, A). (3.34)
(ω) (ω)
In contrast to Equation (3.27), here the amplitude and phase modifiers gAM and ΘPM
are functions of amplitude and frequency. As frequency itself is defined as the number
of cycles over time, we cannot determine a literal instantaneous frequency at any
instance. Any frequency-dependence therefore requires some sort of memory in the
model. This is why a model as described in Section 3.3.2 is sometimes referred to as
memoryless, even though this is not entirely correct in the presented implementation,
as any envelope-based approach inevitably is a memory-based approach in itself. In
frequency-dependent models, the memory serves to detect the spectral information of
the signal in addition to the envelope and phase as required by the SGC model. This
enables filtering of the signal according to the model properties and the short-time
frequency content of the signal. In general, filtering is conceivable both before and
after the nonlinearity, resulting in a now altered equation for the output signal
y(t) = hout (t) ∗ [gAM (|x1 (t)|) · exp (jξ0 (t) − ΘPM (|x1 (t)|))] , (3.35)
with
x1 (t) = (hin ∗ x0 ) (t). (3.36)
The two filters hin and hout represent different aspects of the TWT’s behavior and cannot
be merged or interchanged. Simplified, the first one predominantly incorporates drive
variations over frequency by considering the flatness of the input back-off, while the
second one in addition predominantly readjusts the spectral response in the nonlinear
regime. The actual filter properties need to be obtained in a possibly complicated
fitting process. This can be done either in frequency- or in time-domain, leading to a
natural description of the resulting filters in the respective domain. Oftentimes, as
arbitrarily modulated transient signals are used for fitting, a time-domain approach is
more intuitive and more typically applied. In our case, as the idea is to connect the
Envelope approach as a module to a hybrid time- and frequency-domain simulation
routine, a frequency-domain approach is more natural.

50
 3.3 Envelope Models

Lin. Filter Nonlinearity Lin. Filter

s1 (t) Hin (f ) G(A), ϕ(A) Hout (f ) s2 (t)

ω - Domain t - Domain ω - Domain

Figure 3.13: Typical chain of nonlinear and frequency-dependent models in a Three-

Box model.

3.3.4 Three-Box Model

A natural way of separating and concatenating amplitude- and frequency-dependent
parts of the model is the so-called Three-Box model (TBM). Its basic idea is to
use the similarity of the AM-AM and AM-PM curves over frequency. The TBM
is schematically shown in Figure 3.13. As the name suggests, the functionality is
summarized in three functional blocks. The nonlinearity box in the middle contains the
nonlinear phase and amplitude characteristics and is naturally applied in time-domain.
This is the actual Envelope simulation here. The filters before and after this part
representing the nonlinearity are purely linear and therefore do not need to live in
time-domain only. They are described and applied in frequency-domain. This requires
multiple discrete Fourier transforms to switch between time- and frequency domain,
which at carrier level might be rather time consuming.
The operation principle of this model assumes that a change in frequency results
in a simple shift of the nonlinear transfer curves both in domain and co-domain of
the nonlinear characteristic. Thus, it is assumed that the shape itself does not change
significantly over frequency.
This leads to a fairly simple and straightforward filter generation scheme. For an
AM-AM shape function gAM,ω0 (A) at a reference frequency ω0 , the filter HA,I at the
model input can be generated as
v
Pin, sat (ω)
u
u
HA,I (ω) := t , (3.37)
Pin, sat (ω0 )
with the saturated input power over frequency Pin, sat (ω), while the respective filter
HA,O at the model output is
v
Pout, sat (ω)
u
u
HA,O (ω) := t , (3.38)
Pout, sat (ω0 )

51
3 Simulation Models for Modulated Signals

with the saturated output power over frequency Pout, sat (ω). Similarly, for a AM-PM
shape function ΘPM,ω0 (A), the filters HP,I and HP,O at input and output can be built
as v
u Pin, ref (ω)
u
HP,I (ω) := t , (3.39)
Pin, ref (ω0 )
and

HP,O (ω) : = exp [j ΘPM,ω (Pin, ref (ω)) − j ΘPM,ω0 (Pin, ref (ω0 ))] (3.40)
= exp [j ϕout, ref (ω) − j ϕout, ref (ω0 )] . (3.41)

The tupel (Pin, ref (ω), ϕout, ref (ω)) denotes a reference point on the respective AM-PM
transfer curve. In principle, the chosen reference point only has the constraint that
the AM-PM transfer curve shapes should be as similar as possible around this point
for the desired frequency range. For

Pin, ref (ω) = Pin, sat (ω) (3.42)

or at least
δPin, sat
= 1, (3.43)
δPin, ref
the filters as described in Equation (3.37) to Equation (3.41) can directly be used in a
Three-Box scheme. Both the order in which the filters and nonlinearities are aligned
and the filter synthesis are schematically shown in Figure 3.14. The nonlinearity-block
containing the AM-AM and AM-PM transfer function can be implemented as shown
in Section 3.3.2. If Equation (3.43) does not hold, the filters need to be reassorted.
By this, the resulting scheme is identical to a more general Multi-Box model, which
is sometimes referred to as Poza-Sarkozy-Berger (PSB) model [52]. For this, the
nonlinearity block itself consists of an AM-PM nonlinearity, a following filter

HA,I (ω)
HPSB (ω) := , (3.44)
HP,I (ω)

and afterwards the AM-AM nonlinearity. This block is then framed by two linear filter
blocks. The preceding filter is

HPSB,in (ω) := HP,I (ω) · HP,O (ω), (3.45)

and the subsequent one is

HPSB,out (ω) := HA,O (ω). (3.46)

52
 3.3 Envelope Models

HA,I (f ) Pout HA,O (f )

f f
f1

f2 f3 Pin
f1 f2 f3 f1 f2 f3

HA,I (f ) AM/AM HA,O (f )

HP,I (f ) AM/PM HP,O (f )
HP,I (f ) ϕP M ϕ ⇒ HP,O (f ) = ejϕ(f )

f f1 f
f2
f1 f2 f3 f3 f1 f2 f3
Pin

Figure 3.14: Schematic view of the TBM, adapted from [53].

This representation offers more flexibility regarding the reference point than the TBM,
but is more complex and not modular with respect to the SGC model.
As noted before, the TBM assumes a static transfer-curve shape over frequency.
This is of course a rather limiting simplification, as only curves in a certain limited
back-off range can be seen as similar. The range shrinks even further, the larger the
frequency shift is. Therefore, only limited bandwidths can be represented accurately
in the model.
Still, in some cases the model improves the estimation compared to a simple
frequency-independent model, as is shown in [54]. Here, the average D3 for an equal
power two-tone excitation on a Ku-Band TWT is evaluated. The reference TWT
model used in [54] is more dispersive than typical commercial Ku-Band TWTs. Thus,
all frequency-dependent effects are emphasized. Results from a TBM according to
Equation (3.42) are compared to reference MVTRAD simulation. The findings show
that while for some narrow frequency spacings, the TBM improves the estimate
compared to the SGC models, it is still limited in bandwidth and in many cases
does not offer much improvement compared to the frequency-dynamic SGC model
representation.

53
3 Simulation Models for Modulated Signals

Inphase nonlinearity Sp (t)

x(t) Quadrature model + y(t)

π
ej 2 Quadrature nonlinearity Sq (t)

Figure 3.15: Typical structure of models in quadrature topology according to [53].

Thus, the TBM has both advantages and disadvantages compared to the SGC
approaches. To some extent, it mitigates the issue of the ambiguous frequency reference
choice that comes with the SGC models, adding some frequency information to the
estimate. Unfortunately, this comes at the cost of a limited back-off range that can be
considered. Also, the similarity of the AM-AM and AM-PM curves over frequency
might vary strongly depending on the TWT, such that the scope of usability of such a
model is strongly reduced. Especially when the TWT is strongly dispersive, it is more
important to use a model which is more broadband than the SGC model, but typically,
this also entails a more distorted curve shape. Thus, one needs to be very careful when
using either the TBM or also the PSB model, which has similar constraints.

3.3.5 Quadrature Polynomial Model

General Idea and Inspiration

While the TBM and similar serial models improve the estimation for broadband signals,
they have unfavorable properties, as discussed in the previous section. Especially the
shape-similarity constraint strongly limits the usability.
Thus, more involved models are considered. From the variety of models, there
are several ones with a high degree of flexibility. In principle, one could arbitrarily
concatenate blocks of nonlinearities and linear filters to render the TWT’s properties.
Then, a large variety of individual functional blocks leads to highly adaptable structures
with many degrees of freedom. The parameters corresponding to the considered device
then need to be fitted, if possible, to as many input signal constellations as possible,
as for instance required in a general Volterra [53] approach. Building such a model
based on an existing device, which can be characterized based on arbitrary signal
waveforms, one can expect the outlined approach to yield correct results at the cost of
being tedious and time-consuming.

54
 3.3 Envelope Models

Our problem is different. Instead of being able to reproduce the behavior of

the amplifier with some accuracy, the idea here is to predict it for multi-frequency
excitations, solely based on information effortlessly obtained by, e.g., frequency-domain
simulation results or classical measurements. This leads to a confinement of choices
from the pool of models. A subgroup of models can be based on sets of single-tone
characteristics. The frequency-dependent nonlinear amplifier model can be constructed
by splitting it into a number of concatenated blocks, following the description in
Section 3.3.3. The evaluation of the non-dispersive nonlinearities is carried out in time-
domain and the corresponding linear parts are represented by bandpass filters which are
evaluated in frequency-domain. A well known and rather simple type of model from the
usable subgroup is the so-called Saleh model [50]. It is best known as a single-frequency
approach, where the transfer curve is fitted by an expression with two parameters
for the amplitude and and two more for the phase information. Additionally, in [50],
a frequency-dependent model is proposed, where the characteristics are considered
in two distinct parts, an inphase branch and a quadrature branch with a 90◦ phase
shift. The idea of this separation stems from resorting Equation (3.32) by Euler’s
formula, resulting in the quadrature description [55], as can be seen in Figure 3.15.
Both the inphase and quadrature part are then fitted to two analytic expressions,
each with two parameters per considered frequency point. Thus, the model requires
only four degrees of freedom to describe the characteristics of the amplifier at a single
frequency. Therefore, the model identification is as simple as it gets. Unfortunately,
for the purpose defined in this work it has shown to be insufficient to predict the
nonlinearities and the dispersiveness of the TWT with the required accuracy.
A more versatile model, enabling a closer fit, is the so-called Bessel function model
(BFM) [51]. For this, the expressions for the inphase and the quadrature components
are generalized to a sum of Bessel functions. Instead of being limited to four, the
number of parameters is now a parameter itself. The adaptability comes at the cost of
a complex topology and fitting mechanism. In total, K selected gain and phase curves
over frequency are translated to the respective inphase and quadrature components SI
and SQ of the transfer characteristics, given by

cos(ΘPM (A, f )),

x = I,
Sx (A, f ) = gAM (A, f ) · (3.47)
sin(ΘPM (A, f )),

x = Q.

These components are then each fitted by N Bessel functions of first order and kind
N
nπA
 
! X
SI,Q (A, f ) = GI,Q (n, f ) · J1 , (3.48)
n=1 D

55
3 Simulation Models for Modulated Signals

 
1πA
J1 D
GI (1, f )
... +
 
N πA
J1 D
GI (N, f )

x(t) SI (A, f ) SQ (A, f ) + y(t)

π
 
1πA
ej 2 J1 D
GQ (1, f )

... +
 
N πA
J1 D
GQ (N, f )

Figure 3.16: Simplified block schematic of the BFM.

with a shape-controlling parameter D of the individual Bessel functions and 2 · N · K + 1

degrees of freedom. The quadrature and the inphase components contain N branches
each. These branches, in turn, are each composed of a Bessel function and a linear filter
according to Equation (3.47). The parameters GI,Q (n, f ) are sorted and reshaped into
2N linear filters, which are applied subsequently. A simplified schematic of the BFM
is shown in Figure 3.16. Simulation has shown that the number of Bessel branches is
required to be above 20 for reasonable results.
As has been noted, one drawback, amongst others, is the model complexity. Thus,
a simpler model inspired by the BFM and the other principles given in the previous
sections is derived. Again, the TWT is split into a quadrature and an inphase
component, as this has shown to be useful for mitigating the shape-similarity constraint
of polar models. Still, one could imagine a similar approach in a serial way. Thus, the
composition of the quadrature and the inphase components is the more crucial part.
From other models, e.g., general Volterra or Polyspectral approaches, the separation
of the transfer function into polynomial partial functions is well-known. Therefore,
inspired by these and in combination with the quadrature approach, each quadrature
branch is split into partial functions containing a part of the total nonlinear dispersive
behavior. A signal going through, for instance, the inphase branch then sees a different
part of the total nonlinearity in each subbranch. These partial nonlinearities are
constituted by partial polynomial terms. The nonlinear blocks are then evaluated in
time-domain with the envelope extracted using the Hilbert transform.

56
 3.3 Envelope Models

Structure and Workflow of the Quadrature Polynomial Model

The structure of this Quadrature Polynomial Model (QPM) is presented in Figure 3.17.
First, the reference data is generated in the MVTRAD part, based on a 2D grid of
input powers [Pin,min , (Pin,min + ∆Pin ), . . . , Pin,max ] and frequency points [fmin , (fmin +
∆f ), . . . , fmax ]. The values Pin,min , Pin,max , fmin , and fmax are the lowest and highest
considered input powers and frequencies, respectively. The grid points are equally
spaced with ∆Pin and ∆f . For each combination, an MVTRAD run is evaluated, such
that the output power and phase are obtained and stored. This 2D data is translated
to a two-dimensional piecewise cubic polynomial (see Section 3.3.2), which is then
considered in the fitting routine used for the model parameter identification. A further
explanation follows in the following section.
The Envelope model itself is split into two parts which are in quadrature. Each
block consists of N branches according to the chosen number of degrees of freedom.
Each branch contains a TBM according to Section 3.3.4, with Equation (3.43), where
the model parameters from the fitting routine describe the TWT behavior. The
individual TBM-blocks represent different parts of the transfer characteristics. Thus,
the TBM filters describe the impact of the partial nonlinearity considered in the block.
In principle, the nonlinear base functions used for the expansion can be chosen in
many ways. Here, monomial terms of varying order are used, as they have shown
to be well manageable. In principle, one can imagine several satisfactory solutions
based on the degrees of freedom and base function description. Similarly, different but
nevertheless still well-fitting parameters can be found without a frequency-dependence
of the output filter in each branch. This does not directly comply with the intuitive
description in Section 3.3.4, as the model is not based on a direct intuitive relationship
between physical properties and the model parameters. Nevertheless, an attempt to
find such a relationship in the QPM will fail anyway. Thus, for the sake of a reduced
model complexity, for the further study and simulation results in this work G2I,Q (n, f )
is set to one.
According to Figure 3.13, the filters are applied in frequency-domain, followed by
a time-domain evaluation of the partial nonlinearities. Therefor, the nonlinearities
are each considered in an individual SGC model with a Hilbert transform. To save
computation time, the Hilbert transform is merged with the inverse-FFT, as it is
required anyway. Also, as the Fourier transform is linear, an FFT is performed before
the signal enters the branch-points in front and behind the inphase and quadrature
components of the model, respectively. For the output signal y(t), the summed resulting
signals from the branches from both inphase and quadrature part are superimposed.

57
3 Simulation Models for Modulated Signals

Input power range

Frequency selection
Pin,min , (Pin,min + ∆Pin ), . . . , Pin,max

ax
fm
.
..
in
fm
MVTRAD - Parametric, ω-domain, PIC approach

Pout (n∆Pin , m∆f )

ϕ(n∆Pin , m∆f )

Interpolation: Piecewise cubic polynomials

ϕ(Pin , f )
Pout (Pin , f )

QPM model fitting

Model
parameters
·1
··
N IFFT SGC FFT
FFT G 1
1 Q (1, f ) IFFT H{·} SGC FFT G 2
2 Q (1, f )
G
1 Q (1, f ) IFFT H{·} SGC A 0
FFT G
2 Q (1, f)
G
1 Q (1, f ) IFFT H{·} SGC A 0
FFT G
2 Q (1, f )
G (1, f )
QTBM 1 IFFT H{·} SGC A 0
FFT 2 Q G (1, f )
G1I (N,
TBM f1 )
H{·} A0
N −1
GI (N, f )
TBM 1 A IFFT
TBM 1
TBM N
x(t) Inphase component SI (A, f ) y(t)
+
1 IFFT
· ·
N · IFFT SGC FFT
FFT
G 1
1 Q
(1, f ) IFFT H{·} SGC 0 FFT G 2
2 Q
(1, f )
e jπ G
1 Q (1, f ) IFFT H{·} SGC A FFT G
2 Q (1, f)
2
G
1 Q (1, f ) IFFT H{·} SGC A 0
FFT G
2 Q (1, f )
G (1, f )
QTBM 1 IFFT H{·} SGC
0
A FFT 2 Q G (1, f )
G1Q (N,
TBM 1
f) H{·} A0
N −1
GQ (N, f )
TBM 1 A
TBM 1
TBM N
Quadrature component SQ (A, f )

Figure 3.17: Full topology of the QPM and the model identification scheme.

58
 3.3 Envelope Models

Model Identification and Fitting Routines

For accurate simulation with the QPM, the model parameters need to be well-fitted to
the reference data as generated in MVTRAD. In the following, suitable fitting routines
are established. The general idea is, that
2
 
X N
X
min  SI,Q (A, f ) − GI,Q (n, f ) · qI,Q (n, A)  (3.49)
GI,Q
f,A n=1

holds, where qI,Q describes the base functions. In the special case described in this
section and Figure 3.17, these base functions relate to the base monomials An .
With a vector of input amplitudes A(fi ) = [A1 , A2 , ..., AM ], possibly interpolated
from the grid points, and the respective output inphase and quadrature components
SI,Q (fi ), the required parameters can be found as

GI,Q (fi ) = [M]+ (fi ) SI,Q (fi ), (3.50)

where [M]+ (fi ) is the Moore-Penrose pseudoinverse [56], [57] of the transfer matrix
h i
[M](fi ) = A0 (fi ), A1 (fi ), · · · AN −1 (fi ) . (3.51)

Here, An denotes the elementwise Hadamard product [58]

An = A ◦ An−1 . (3.52)

The Moore-Penrose pseudoinverse is calculated by estimating the singular value

decomposition of [M]
[M] = U [Σ] V ∗ , (3.53)
with the diagonal matrix
[Σ] = diag{σi } (3.54)
i

and σi denoting the singular values of [M]. For the numerical evaluation, singular values
below a threshold  are not considered, i.e., set to zero. Thus, with the dimension-wise
reduced matrix
[Σ ] = diag{σi } (3.55)
σi >

and the accordingly truncated matrices U and V , the pseudoinverse is

[M]+ ≈ V [Σ ]+ U∗ , (3.56)

with 
1

, for σi 6= 0,
[Σ ]+
i,i =
σi
(3.57)
0,

for σi = 0.

59
3 Simulation Models for Modulated Signals

Figure 3.18: SIMBA program written in MATLAB.

With M = N , the pseudoinverse equals the inverse and an explicit formulation is

available [59]. For M > N , there is a unique pseudoinverse, which returns the best
solution to the least-squares problem. Thus, the QPM inherently features an explicit
best fit for a specified number of degrees of freedom. Also, the fitted parameters have
shown to be less dynamic compared to the BFM, such that interpolation, e.g., between
frequency points is less critical.

3.3.6 SIMBA and the Satellite Downlink Chain

In order to use the developed models reproducibly, the so-called SIMBA (Simulation
of Inter-Modulations in Back-Off and Applications) program viewed in Figure 3.18
has been developed. Just as the Envelope models themselves, the interface is written
in an object-oriented manner in the MATLAB programming language and can be
extended modularly.
The amplifier models are integrated in a satellite downlink chain model including
modulation and demodulation of communication signals. The block schematic of the
signal chain as implemented in SIMBA is shown in Figure 3.19. For the TWT, both
the SGC model and the QPM are included. In addition, as it is typically required in
such a downlink chain, a linearizer can be inserted ahead of the TWT model. Here,
for the sake of simplicity, it is described in an SGC manner. In principle, a combined

60
 3.4 Comparison

linearizer and TWT model could be generated, e.g., in a QPM manner, but for some
applications it is worth considering them as two separate subsystems. This enables
visualizing the predistortion behind the linearizer. Also, when considering the total
efficiency, and especially the collector efficiency, the collected currents only depend
on the signal distribution at the input of the TWT, which obviously differs from the
one at the linearizer input. The collected currents, efficiencies, and electron spectrum
estimations are included as a part of the amplifier model blocks. Also, both in front of
and behind the TWT, white Gaussian noise can be added to the signal.
Supported signal types are single- and multi-tone signals, typical (single- or multi-
carrier) modulation schemes as defined in the DVB-S2 standards, and band-limited,
noise-like signals, as are for instance required for noise-power ratio (NPR, [60]) cal-
culations. Communication signals are generated from binary random data, which is
plugged into the modulator. Possible modulation schemes are quadrature phase-shift
keying (QPSK), M -ary quadrature amplitude modulation (M QAM), M -ary phase-
shift keying (M PSK) and M -ary amplitude- and phase-shift keying (M APSK), with
typical constellation radii ratios. The baseband signal generated by the modulator is
upsampled and scaled to the desired power level. A Root-Raised Cosine (RRC) filter
[61] is implemented for pulse-shaping. At the receiving end, a second Root-Raised
Cosine filter is used as a matched filter. Based on the demodulator following the
matched filter and the downsampling block, various communication system perfor-
mance characteristics are calculated, such as bit- and symbol-error rates and the
error-vector magnitude.
The models are applied at the carrier frequency, as this is the natural, physical
setting. In principle, one could stay at the baseband frequencies, but that would limit
the scope of investigations considered in this work.

3.4 Comparison
In the following, the presented QPM and SGC model are evaluated regarding their
performance in multi-tone operation. The BFM is also investigated as a benchmark,
as it follows similar principles with comparable complexity as the QPM. The material
for this section draws in parts on previously published studies [43], [44].
Here, the comparison is carried out in several steps. As the main objective of the
considered approaches is to enable the simulation in the design process, the models
are not supposed to be based on measured TWT characteristics, which is different to
many other approaches. Still, in the first step, measurements are used as a reference,

61
3 Simulation Models for Modulated Signals

Generate random data Data generation and comparison Receive data and compare

01110010100... 01110010100...

Modulation Demodulation

Modulation schemes:
QPSK
M QAM
M PSK
M APSK
...

Normalization
Upsampling

Scale to defined Downsampling

power level
1

Root-raised cosine Root-raised cosine

transmit filter receive matched filter
Pulseforming filter properties:
Filter span
Roll-off factor
Bandwidth
Data rate

Re{ T} Im{ T} Re{ R} Im{ R}

Oscillator/ Oscillator/
carrier carrier
/2 /2

cos( ) 2 cos( )

,…

2
Linearizer
TWT
(SGC)
SGC
QPM

Noise Noise
Channel and amplifier models

Figure 3.19: Schematic view of the signal chain from a single run with a communica-
tion signal in SIMBA.

62
 3.4 Comparison

as this mitigates other error sources, such as the numerical implementation or other
inaccuracies in the underlying interaction model.
Then, in a second step, the measured data is replaced by that from the steady-state
simulation tool MVTRAD. This defines the combined hybrid frequency- and time-
domain approach, which consists of the steady-state simulation of MVTRAD and the
presented Envelope codes.
In these two parts, both the MVTRAD- and the measurement-based approach
are evaluated with regards to two-tone excitation. Although TWTs are typically not
excited with simple two-tone signals, these can easily be understood and are routinely
measured by manufacturers. They represent the least challenging multi-frequency
scenario and yield meaningful figures of merit. There are even approximate formulas
readily available from literature, as was discussed in Section 3.3.1. Dual-tone signals are
thus best suited to verify the intermodulation performance predicted by the approaches
discussed here.
Afterwards, the broadband behavior is considered by means of communication
signals, transient steps and noise-like signals. Also, the models are compared for a
FW-TWT, which has a strongly dispersive behavior at a much higher frequency. In
the end of the section, a short conclusion summarizes the results.

3.4.1 Multi-Tone Simulation with Measurement-Fitted

Models
For the following comparisons, measurements are carried out on the reference commu-
nication Ku-Band TWT. The model parameters of the Envelope codes are extracted
from CW AM-AM and AM-PM measurements throughout the frequency band of
interest. In addition to the model base data, reference two-tone measurements with
different frequency spacings are carried out, such that a direct comparison between
Envelope simulation and measurement results is possible.
Therefore, a symmetric excitation is considered. The equal input power of the two
tones is simultaneously altered from small-signal excitation to saturation in steps of
2 dB. The simulated and measured intermodulation products for a constant central
frequency fc and a frequency spacing of ∆f = 100 MHz are shown in Figure 3.20a.
As the lower side-band is nearly indistinguishable from the upper side-band, it is not
shown. The intermodulation product ratio definition from Equation (3.8) is used. Both
D3 and D5 show near perfect agreement. This can be explained by the slowly changing
amplitude of the input signal and the almost constant amplifier characteristics in the
relevant band for this scenario.

63
3 Simulation Models for Modulated Signals

40 −4 −10

Pout /Pout,sat @(f2 + ∆f ) (dB)

SGC
D3 (dB) and D5 (dB)

Pout /Pout,sat @f2 (dB)

BFM
30 QPM −6 −15
(2)

Meas.
20 −8 −20
(2)
D5 SGC
10 (2) −10 BFM −25
D3 QPM
(2)

Meas.
0 −12 −30
−15 −10 −5 0 −15 −10 −5 0
Pin /Pin,sat (dB) Pin /Pin,sat (dB)

(a) Third- and fifth-order intermodula- (b) Main tone and intermodulation product at
tion product ratio at ∆f = 100 ∆f = 500 MHz spacing.
MHz spacing.

Figure 3.20: Output power at main- and side-tone and intermodulation ratios for
various measurement-based models compared to measurements.

The agreement still holds for a large frequency offset of ∆f = 500 MHz. The SGC
model shows a performance similar to the more involved methods, as can be seen in
Figure 3.20b. Here, the power levels at f2 and f2 + ∆f are reported. While the overall
agreement again is good, in this case all models underestimate the power at one of the
excitation frequencies by about 0.15 dB to 0.18 dB. This discrepancy might stem from
a ripple on the gain over frequency [62]. Such a ripple is typically mismatch-induced,
i.e., due to imperfect input and output couplers, and thus is not considered in the
models.
For communication systems, the phase distortion is likewise crucial. Therefore,
with the phase transfer factor kT as defined in Equation (3.21), a further metric is
investigated. It is shown for different ∆f versus input power in Figure 3.21. The
phase transfer factor is obtained with an asymmetric excitation, where Pin (f1 ) =
Pin (f2 ) − 15 dB. Then the phase difference at f1 is measured, while reducing the
power at the other tone Pin (f2 ) by 1 dB. For a better overview, the absolute error
∆kT between simulation and measurements is also shown. As the nonlinear phase is
strongly dispersive, the frequency dependent models yield improved predictions, as is
discussed in the following.
For two narrowly spaced tones at ∆f = 5 MHz, shown in Figures 3.21a and 3.21b, all
models accurately predict the measured results. In Figures 3.21c to 3.21f the advantage

64
 3.4 Comparison

of adding frequency information to the model is illustrated. This is especially obvious

near saturation where the phase error rapidly increases. While the error rises above
1◦ /dB for both SGC and BFM in the case of ∆f = 200 MHz, it stays below 0.5◦ /dB
for QPM. Likewise, for ∆f = 1 GHz the error is significantly larger for SGC and BFM,
exceeding 2.1◦ /dB, compared to the QPM, where it remains below 1.2◦ /dB. Thus, only
the QPM considerably benefits from the inclusion of dispersion information, although
the same information is present in the BFM. This is due to the strong fluctuation and
large range of the Bessel function fitting parameters over frequency, which yields a
highly sensitive model. The QPM parameters, in contrast, change far more gracefully,
such that, for instance, the required interpolation between fitting points is significantly
improved.

3.4.2 Multi-Tone Simulation with MVTRAD-Fitted Models

In the preceding section, only the Envelope methods have been evaluated. To extend
the study to the hybrid approach, an MVTRAD reference model has been generated
and fitted closely to the measured TWT. The measured results are adjusted by the
difference in saturated input power between the MVTRAD model and the measured
data to enable comparison with multi-tone measurements. In addition, the underlying
MVTRAD model is not only used inside the full simulation chain, but also for
standalone frequency-domain simulation with multi-tone signals in MVTRAD and
CST following Section 3.1.2 and Section 3.2, respectively. In the following, the legend
entry MVT denotes MVTRAD simulation results.
As before, different dual-tone signals with the TWT driven close to saturation are
measured and calculated with both the hybrid methods and MVTRAD alone. Therefor,
CW output power and absolute phase versus input power and frequency are simulated
in MVTRAD beforehand, to extract the parameters of the hybrid methods. The input
power is increased from small-signal excitation to 10 dB into overdrive in steps of 1 dB
and the frequency information is obtained in steps of 100 MHz throughout the band
of interest.
For Figure 3.22a and Figure 3.22b, the center frequency of the considered signals is
placed at the center of the amplification band, where the dispersion is naturally low
by design. In Figure 3.22a the power spectrum for a two-tone signal with frequency
spacing ∆f = 100 MHz at 0.5 dB input back-off is shown. All models closely match
both the measurements and the MVTRAD simulation results due to the small spacing
between the tones. Still, at a distance of 3∆f from the frequency center, a small error
can be seen, which is mainly caused by inaccuracies in the reference MVTRAD model.

65
3 Simulation Models for Modulated Signals

10 3.2
SGC SGC
8 BFM BFM
2.4

∆kT (deg/dB)
QPM QPM
kT (deg/dB)

6 Meas.
1.6
4
0.8
2

0 0
−15 −10 −5 0 −15 −10 −5 0
Pin /Pin,sat (dB) Pin /Pin,sat (dB)

(a) ∆f = 5 MHz: Phase transfer. (b) ∆f = 5 MHz: Error.

10 3.2
SGC SGC
8 BFM BFM
2.4
∆kT (deg/dB)

QPM QPM
kT (deg/dB)

6 Meas.
1.6
4
0.8
2

0 0
−15 −10 −5 0 −15 −10 −5 0
Pin /Pin,sat (dB) Pin /Pin,sat (dB)

(c) ∆f = 200 MHz: Phase transfer. (d) ∆f = 200 MHz: Error.

10 3.2
SGC SGC
8 BFM BFM
2.4
∆kT (deg/dB)

QPM QPM
kT (deg/dB)

6 Meas.
1.6
4
0.8
2

0 0
−15 −10 −5 0 −15 −10 −5 0
Pin /Pin,sat (dB) Pin /Pin,sat (dB)

(e) ∆f = 1 GHz: Phase transfer. (f) ∆f = 1 GHz: Error.

Figure 3.21: Phase transfer factor for different frequency spacings ∆f .

66
 3.4 Comparison

MVT Meas. SGC QPM BFM CST

0 0
Pout /Pout,sat,fc (dB)

Pout /Pout,sat,fc (dB)

−10 −5

−10
−20

−15
−30
−20
−7 −5 −3 −1 1 3 5 7 −3 −1 1 3
2 (f − fc )/∆f 2 (f − fc )/∆f

(a) ∆f = 100 MHz. (b) ∆f = 500 MHz.

Figure 3.22: Two-tone signals around fc with a frequency spacing of ∆f at 0.5 dB

input back-off. The legend on top is valid for both plots.

The results for a larger frequency spacing of ∆f = 500 MHz are shown in Figure 3.22b.
The output power at the upper main frequency Pout (f2 ) is slightly overestimated by up
to 0.2 dB not only by the hybrid models but also by MVTRAD. The error is even larger
at the other main tone with Pout (f1 ) differing by as much as 0.5 dB. This deviation
is not visible for the CST results, which hints at a model reference issue. Still, the
frequency-dependence helps improving the estimation, as can be seen at f1 − ∆f . With
the QPM the error shrinks to less than 0.5 dB compared to the SGC’s overestimation
by 1.2 dB. No further improvement is possible as the error is in a similar range in
MVTRAD. For higher intermodulation products, all models show similar deviations.
In addition to the signals at the band center, two more dispersive scenarios are
investigated in the following. For the results in Figure 3.23a, the center frequency
is shifted down by around 5 % and a large frequency spacing of ∆f = 500 MHz is
chosen. At the two main-tone frequencies, all models perform similarly well compared
to the measurements. On the low-frequency side of the output spectrum no reliable
measurements are available. This is due to the input and output coupler, and limited
calibration. Still, a comparison between Envelope methods, CST, and MVTRAD is
attempted. Figure 3.23a shows that the resulting power spectrum from MVTRAD is
slightly more asymmetric than in the previous cases, as the device is more dispersive.
This is confirmed by the CST simulation. Contrary to this, due to its underlying

67
3 Simulation Models for Modulated Signals

mechanism, the SGC model always results in a symmetric output spectrum for a
symmetric dual-tone input signal. Thus, it yields a large error of up to 1.5 dB. Both
QPM and BFM are significantly closer. For the upper side-band, all hybrid models
are similarly far away from either measurements, MVTRAD, or CST, with the error
being around 1 dB to 1.5 dB. Finally, Figure 3.23b reports the results for a center
frequency shifted further downwards to the lower band edge and a frequency spacing
of ∆f = 200 MHz. Again, both BFM and QPM yield some improvements at the
side-bands, reducing the error at f2 + ∆f to less than 0.3 dB compared to 0.7 dB
for SGC. At f1 − ∆f the results can only be compared to MVTRAD and CST. As
before, the dispersive models significantly improve the estimation compared to the
SGC, lowering the error from 1.2 dB to 0.5 dB.
In total, one can observe a benefit of using the QPM or BFM concerning their
applicability in a hybrid time- and frequency-domain TWT simulation. Still, another
outcome of this investigation is that for a real communications TWT - as is investigated
here -, at typical operating conditions and for practically relevant frequency spacings,
the basic SGC model is sufficient to predict the relevant intermodulation power levels.
On the other hand, when trying to predict the phase related properties, such as the
phase transfer coefficients, the more sophisticated models present their advantages
more clearly.

3.4.3 Transients and Communication Signals

Sets of two-tone excitations, while they are popular for amplifier characterization,
do of course not always provide a complete picture of TWTs in communication
systems. Such an approach is strongly limited, as the initial phase relations between
the two tones do not have an influence on the power spectrum. In addition, TWTs
for communication satellites are operated with complexly modulated communication
signals. Thus, communication system metrics such as constellation diagrams, eye-
diagrams, and others are required to fully characterize the tube in realistic operating
conditions.

Transients

The question of accurate frequency or phase representation is tightly connected to

the one of transients and how well the models follow signal transitions from one state
to another. Therefore, it is useful to consider various scenarios, where the signal is
stepped up from zero-drive with different signal bandwidths.

68
 3.4 Comparison

MVT Meas. SGC QPM BFM CST

0 0
Pout /Pout,sat,fc (dB)

Pout /Pout,sat,fc (dB)

−5 −5

−10 −10

−15 −15

−20 −20
−3 −1 1 3 −3 −1 1 3
2 (f − 0.95fc )/∆f 2 (f − 0.925fc )/∆f

(a) Tones around 0.95fc with ∆f = 500 MHz. (b) Tones around 0.925fc with ∆f = 200
MHz.

Figure 3.23: Two-tone signals with a frequency spacing of ∆f at 0.5 dB input back-off.
The legend on top is valid for both plots.

In CST, arbitrary signal excitation can be defined in time-domain, e.g., by inserting

it from an ASCII-coded input file. As the underlying model has a sound physical full-
wave basis, the results serve as a benchmark for the Envelope methods. In particular,
one can assume that a physical device cannot immediately follow a sharp ideal step in
the oscillation, but smoothens it according to its physical bandwidth. This is mainly a
result of the dispersive properties of the line.
In Figure 3.24 the output signal envelopes for several transitions simulated in CST,
SGC and QPM are shown. For all plots, the curves are shifted in time, as the absolute
delay is not taken into account and cannot accurately be matched between CST and
the hybrid approach. Here, an absolute shift enables a visual comparison without
corrupting the message.
Figure 3.24a reports the response for an ideal step, which is not filtered and therefore
inhibits an infinite spectrum. It directly drives the models from zero-drive to a large-
signal CW excitation. The physics-based simulation in CST shows a flattened response
compared to the input, as expected. The observed behavior from SGC and QPM shows
a major difference. The SGC model, not having any frequency-domain information
available, directly follows the input as a scaled version of it. In contrast, the QPM
model accurately reproduces the slope as predicted by CST. The waveform differs

69
3 Simulation Models for Modulated Signals

slightly, which is due to the limited spectrum considered in the QPM and the inherent
harmonic representation problem of Envelope models.
To investigate the frequency limits of the models, an RRC filter is used to limit
the input signal to a finite bandwidth. Even for a large bandwidth of 1 GHz, the
step is smooth enough for the SGC model to effortlessly follow it, which can be seen
in Figure 3.24b. Also, the ripple in the waveform, as seen for the QPM results in
Figure 3.24a, can be found in both CST and SGC for the filtered signal, which hints
at the aforementioned bandwidth issue of the QPM with the ideal step response. In
Figure 3.24c, two further transitions are considered. Both signals are RRC filtered
with a bandwidth of 1 GHz. The step from zero-drive to small-signal drive shows that
the sloping error is not a nonlinear effect but indeed caused by the dispersion of the
line. The same can be seen for a step from a steady-state large-signal oscillation to
a small-signal state. The small-signal level is not exactly reached, as the MVTRAD
and CST models do not fit perfectly for both saturation and small-signal input. As
mentioned in Section 3.4.2, the difference in saturated input power is compensated,
leaving room for a deviation in the small-signal gain.
In conclusion, for realistic bandwidths in the considered Ku-Band communication
scenario, even the SGC model accurately represents the transients. For higher band-
widths, where the SGC model is limited, the QPM is a viable alternative, behaving
very similarly to the physics based full-wave model.

QPSK and Eye-Diagram

As discussed before, a full characterization of a TWT in realistic operating conditions

requires communication signals, as for instance defined by the DVB-S2 standard. In
Figures 3.25a and 3.25b a constellation diagram of a large-signal QPSK sequence
from measurements is compared to simulation in SIMBA. As the signal bandwidth
is chosen as typical for the Ku-Band, the SGC model suffices for the calculation, as
was established in the previous sections. The roughly wedge-shaped point clouds
at each constellation point are notched at one end. These apertures, which are not
noise-related, but purely deterministic and a result of the different paths from one
constellation point to another, are accurately predicted in simulation.
A similar accuracy can be seen for the higher-order modulation case in Figures 3.25c
and 3.25d. Here, constellation diagrams of 16QAM signals are shown, again both for
measurements and simulation in SIMBA. The drive level is 5 dB above saturation,
strongly highlighting the nonlinear distortions.

70
 3.4 Comparison

25 25
CST
20 20 SGC
Envelope ( W )

Envelope ( W )
A
√

√
15 15 B

10 10
CST
5 SGC 5
QPM
0 0
−2 −1 0 1 2 −2 −1 0 1 2
Time (ns) Time (ns)

(a) Unfiltered ideally stepped signal from (b) Unfiltered ideal step (A), and 1 GHz RRC
zero-drive to large-signal excitation. filtered step (B), from zero-drive to large-
signal excitation.
25
CST: ZD → SS
20 SGC: ZD → SS
Envelope ( W )

CST: LS → SS
√

15 SGC: LS → SS

10

0
−2 0 2
Time (ns)

(c) RRC filtered signal steps with 1 GHz

bandwidth. (ZD): zero-drive, (SS): small-
signal, (LS): large-signal excitation.

Figure 3.24: Transient processes in SGC and QPM compared to CST simulation.

71
3 Simulation Models for Modulated Signals

(a) QPSK: SGC model in SIMBA. (b) QPSK: Measurements.

(c) 16QAM: SGC model in SIMBA. (d) 16QAM: Measurements.

Figure 3.25: Constellation diagrams from 800 symbols each, for a QPSK and a
16QAM signal at a drive level of 5 dB above CW saturation.

72
 3.4 Comparison

In Figure 3.26a the effective output power in the modulated case is shown in
comparison to simulation. Not only do the results align well with the measurements,
the behavior with higher-order modulation is also visible. A metric for the distortions
occurring in a communication system is the so-called error-vector magnitude (EVM).
It is defined as
v
u 1 P 
u  2  2 
ˆ
Ik − Ik + Qk − Q̂k
Perror
u
uN k
EVM = = u
1 Ph , (3.58)
Preference
u i
2 2
(Ik ) + (Qk )
t
N k

where Iˆk and Q̂k are the measured inphase and quadrature component of the k-th
symbol, Ik and Qk are those of the corresponding reference point, and N is the number
of considered symbols. When shown in logarithmic scale, it is sometimes referred to
as the modulation-error ratio (MER) and calculated with
Preference
 
MER = 10 log10 . (3.59)
Perror
In Figure 3.26b, the MER is shown for QPSK and 16QAM, comparing simulation and
measurements. The agreement is good from saturation downwards to typical back-off
values. For 16QAM, estimation in the overdrive region is particularly erroneous as
the peak-to-average power-ratio (PAPR) is high compared to the QPSK case. Far
in back-off, the deviations stem from the fact that the error vector there is small
compared to the reference power, such that it is subject to stronger fluctuation.
Looking at the spectrum of a modulated QPSK signal, as shown in Figure 3.27,
one can observe a fundamental limitation of the Envelope methods in the presented
form, compared to CST. While the information on the higher harmonics is inherently
connected to the considered transfer curves, components at these harmonics cannot be
found in the output signal. Thus, while the in-band components of the modulated
signal around a carrier frequency f0 are correct, the corresponding intermodulation
products around multiples of f0 are not present. When information on the harmonics
is required, physics-based simulation models, such as MVTRAD or CST, are therefore
better suited. Still, for many cases no components other than those around the carrier
f0 are required, as the output coupler typically is not well-matched for these frequencies.
Thus, the relatively low power generated there is reflected and absorbed in the sever.
While both the Envelope models and CST’s PIC solver can be excited with an
arbitrary communication signal, this is more difficult for MVTRAD. Anyway, a full
characterization with, e.g., an eye-diagram generation for a QPSK modulation requires
an extensive Monte-Carlo study. While this is possible in measurements or time-domain

73
3 Simulation Models for Modulated Signals

Measurements CW Measurements QPSK Measurements 16QAM

SGC CW SGC QPSK SGC 16QAM

0 40
Output back-off (dB)

-2

MER (dB)
30
-4
20
-6

-8 10
−15 −10 −5 0 5 −15 −10 −5 0 5
Input back-off (dB) Input back-off (dB)

(a) Total output power. (b) Distortions.

Figure 3.26: Output power and distortions of QPSK and 16QAM modulated signals
from measurements and simulation. The legend on top is valid for both
plots.

20 20
PSD, arbitrary norm. (dB)

PSD, arbitrary norm. (dB)

CST CST
SGC 0 SGC
10
−20
0
−40
−10
−60
−20 −80
0.9 1 1.1 0 1 2 3 4 5
Frequency f /f0 Frequency f /f0

(a) Spectrum around the carrier frequency. (b) Spectrum across several harmonics of the
carrier frequency.

Figure 3.27: Output spectrum of a modulated signal with neighboring QPSK channel
from CST and the SGC model.

74
 3.4 Comparison

simulation using Envelope codes, it would result in an infinite number of frequencies

in a frequency-domain simulation in MVTRAD or a very large simulation time in a
full-wave time-domain code as CST’s PIC-solver. Thus, these two cannot be employed
directly.
As the duration of a symbol is typically long, it is not feasable to simulate more
than a few symbols in a full-wave code. Therefore, CST alone will only yield a limited
picture. For MVTRAD, this is in principle even worse, as the continuous spectrum
would require an infinite number of harmonics. Thus, a symbol reduction scheme has
been conceived, limiting the number of simulation runs to a number of meaningful
periodic ones. The reduction is based on a derivation in a noise-free, linear case,
where the meaningful sequences are sufficient to depict the full eye-diagram. With
nonlinearities, this is of course not the case anymore, but the chosen sequences still
yield a good estimate of the general behavior. The reduction scheme and its underlying
idea are outlined in [44]. The resulting periodic sequences can then be simulated in
MVTRAD, and some comparisons to Envelope simulation for such reduced QPSK
symbol sequences are presented in the following.
Therefor, the data rate Rc , the filter symbol span ξ, and the roll-off factor α of the
pulse-forming filter are related to the fundamental frequency ∆f , which is equal to
the spacing between the discrete tones of the periodic signal, according to

Rc
∆f = . (3.60)
(1 + α) · (ξ + 1)

Here, ∆f is chosen to 50 MHz. The carrier frequency is selected such that all discrete
frequency components are multiples of ∆f . Thus, the total frequency band containing
significant information can be estimated to 350 MHz with a roll-off factor of 0.35. The
generated signal is periodic with T0 = (∆f )−1 .
A complex phasor description of the Fourier coefficients at multiples of ∆f rep-
resenting the operating band and possible intermodulation products, as well as the
corresponding second harmonic frequencies around 2f0 are computed and used as the
excitation in MVTRAD. The computation time lies around four days per sequence
as the number of harmonics and thus the number of required time steps within the
base period is large. Then, using the phasor Fourier sum of the resulting amplitudes
and phases, the resulting time-domain signal is reconstructed. After down-conversion
and receive-filtering, the eye-diagram can be recalculated and used for simulation in
the hybrid Envelope approach. In the following, an MVTRAD based SGC model is
compared to a full steady-state simulation with MVTRAD. The operating point is
chosen 2 dB below saturation to ensure strong intermodulation.

75
 3 Simulation Models for Modulated Signals

0 20
SGC
MVT
Pout /Pout,sat,f0 (dB)

Envelope ( W )
15
−20

√
10
−40 A
5
B
C
−60 0
0.9 0.95 1 1.05 1.1 0 0.2 0.4 0.6 0.8 1
Frequency f /f0 Time t/T0

(a) Output power spectrum of an examplary (b) (A): MVTRAD, with components around
QPSK sequence. f0 and 2f0 . (B): MVTRAD, only compo-
nents around f0 . (C): SGC model.
0 2
MVT SIMBA MVTRAD
SGC
Amplitude (a.u.)
Pout /Pout,sat (dB)

−20 1

0
0 dB−40
-6 dB
-12 dB −1
-18 dB
−60
−2
0.4 0.6 0.8 1 1.2 1.4 1.6 −0.4 −0.2 0 0.2 0.4
Frequency f /f0 Time t/T0

(c) Output power spectrum of another exam- (d) Quadrature eye-diagrams from Envelope
plary QPSK sequence. simulation and MVTRAD.

Figure 3.28: Exemplary periodic QPSK sequence simulated in MVTRAD and

SIMBA.

76
 3.4 Comparison

Figure 3.28a and Figure 3.28b exemplarily show the power spectrum and the signal
envelope of one period, generated using the Hilbert transform of the time-domain
signals for one sequence, from both MVTRAD and the combined SGC-MVTRAD
approach. Figure 3.28a reports an excellent overall agreement for the seven main input
frequencies and still very good results for the intermodulation products. At the lower
frequencies, the Envelope approach overestimates the power compared to MVTRAD.
Following a conversion of the output power and phase of the individual frequency
components from MVTRAD to a time-domain waveform, we see good agreement
between the signals from SGC and MVTRAD, when only frequency components
up to f0 + 40∆f are considered. Using the spectrum as calculated in MVTRAD,
with components around 2f0 , the modulus of the analytic signal representation looks
slightly different. This is due to the distortions introduced by these second harmonic
components, which are not seen from the Envelope approach. Anyway, as the signal
is low-pass filtered on the demodulation side, they ultimately are irrelevant for the
investigation. Figure 3.28c shows another sequence, highlighting that the spectral
distribution of each sequence differs strongly. A comparison of the hybrid MVTRAD
based SGC model with MVTRAD shows good agreement from small-signal excitation
at 18 dB below saturation to large-signal excitation at saturation. It can be seen, that
although the power spectrum is symmetric at the input, it can still be asymmetric at
the output, as the phase relations of a signal containing more than two tones influence
their respective power levels in a nonlinear device in an asymmetric manner.
Figure 3.28d finally compares the eye-diagrams obtained from MVTRAD and from
the combined QPM-MVTRAD approach. For the Envelope results, both the vertical
and the horizontal eye-opening are around 5 % smaller, as the eye is slightly more
blurred. Still, the overall agreement between both diagrams is good.

3.4.4 Folded-Waveguide Simulation

The SGC and QPM have been thoroughly tested for typical communication helix
TWTs. Here, there is no significant advantage of using the far more complex QPM
approach compared to a simple SGC. With increasing frequency, starting around the
Q-Band, more dispersive delay lines are needed, for instance the FW lines [15] as
introduced in Section 2.1.2. The validity of the Envelope method, especially the QPM
variant, is evaluated in the following. These FW-TWTs are strongly dispersive and
possess a lower cut-off frequency which is to be considered. Thus, they constitute a
more difficult problem, possibly reaching or illuminating the limits of the QPM.

77
3 Simulation Models for Modulated Signals

60 60
Output power (dBm)

Output power (dBm)

40 50

0d
Bm
40

∆=
20
MVTRAD

4d
30 CST (from [15])

B
0

−4
CST (beam port)

0d
20 KlysTOP
Bm
−20
41 42 43 44 45 41 42 43 44
Frequency (GHz) Frequency (GHz)

(a) Output power over frequency for small- to (b) Simulation results from CST, MVTRAD
large-signal input power levels, simulated and KlysTOP at 0 dBm input power.
in MVTRAD.

Figure 3.29: CW simulation results of the investigated FW-TWT.

In order to further investigate the model boundaries, a FW delay line structure

from [15], [63] is investigated. Figure 3.29a shows the frequency-dependence and
nonlinearity of the FW-TWT. Other quantities show similarly strong fluctuations over
frequency, e.g., the saturated input power shows a difference of up to 25 dB across a
1 GHz band in the lower half of the investigated frequency range.
In [63], the chosen FW-TWT is described and simulated using MVTRAD, which is
specialized for helical delay-lines. The FW’s cold parameters have to be preprocessed
to enable this. The simulation has been repeated using the specialized FW-TWT code
KlysTOP [64], where a good overall agreement between the codes is found, hinting
at the usability of MVTRAD for this task. In [15], simulation results obtained using
CST are added. They exhibit a frequency shift of around 250 MHz compared to the
other results.
With the same model as used for the CST results in [15], the beam injection is
improved using the more involved particle injection procedure from Section 3.2.2. The
beam setup file used in KlysTOP and MVTRAD is chosen as a reference for the beam
interface in CST. As reported in Figure 3.29b, this corrects the frequency offset, such
that the gain from the individual codes now agrees well.
In order to extract suitable envelope models from the CST or MVTRAD data, a
simulation of the saturation behavior at several frequencies within the amplification
band is carried out. Subsequently a sufficiently fine-meshed simulation lattice is defined
with respect to frequency and amplitude. For particularly dispersive structures, this

78
 3.4 Comparison

means that a large number of individual simulation runs is required, since the respective
saturation points are spread far apart. Also, it might be required for the amplitude
range to include both small-signal and overdrive region for several frequencies. In
addition, not only a high density over the frequency band is required, but also a wide
range of frequencies is to be considered. In particular, the required distance to the
lower cut-off around 39.5 GHz is unclear. There are therefore many, possibly lengthy
simulation runs to be carried out. In the course of the investigations, CST therefore
turned out to be unsuitable. Instead, the further investigations and the Envelope
models are based on MVTRAD.
Two-tone excitations are considered to characterize the model performance. Different
frequency spacings and center frequencies are selected in order to develop scenarios
with varying difficulty. The range of possible frequency spacings is mainly limited by
the model limits in MVTRAD. The calculation of the CW properties of the FW-TWT
requires a higher spatial resolution of the numerical grid than for the helical ones. In
addition, due to the narrow bandwidth of the structure, too large frequency distances
do not make sense, but too small frequency spacings are not possible due to the high
required numerical effort. Furthermore, the constraints given by the approach of the
multi-tone calculation in MVTRAD with the many harmonics have to be considered.
Therefore, in Figure 3.30, some scenarios from the amplification band edges to its
center are investigated with frequency spacings of 0.5 GHz. It can be seen in all cases
that the structure is too dispersive for accurate simulation with the SGC model. The
drive frequencies already differ strongly, especially at the edges of the desired operating
band. This is correctly predicted by the QPM. Still, at the lower end of the band,
the results from MVTRAD and QPM differ from each other. It is not clear, whether
any of the two is trustworthy there, as the lower cut-off frequency of the FW line is
already close to the lower main tone.
Figure 3.30f and Figure 3.30e highlight further differences between QPM and SGC
model. In Figure 3.30e, the variation of the power ratio between the two main tones
at the output from a symmetric two-tone excitation is plotted against the frequency
spacing ∆f , where the first tone is fixed at 43 GHz. In a small section around ∆f = 0,
the two output tones are close, such that the SGC model accurately predicts the
output power ratio. Still, for a slightly higher spacing of ∆f = 0.4 GHz the error
already reaches 1 dB, increasing rapidly for higher ∆f . The behavior is asymmetric,
indicating that the maximum predictable signal bandwidth strongly depends on the
position in the frequency band. In Figure 3.30f, the effect of a varying reference data
extent (with respect to the considered frequency range) is investigated. Exemplarily,
ranges of ±2.5∆f and ±4∆f around the center frequency are considered. The more

79
3 Simulation Models for Modulated Signals

100 80
MVT MVT
80 QPM QPM
SGC 60 SGC
Power (W)

Power (W)
60
40
40
20
20

0 0
41 42 43 44 45 46 41 42 43 44 45 46
Frequency (GHz) Frequency (GHz)

(a) Input tones at 42 and 42.5 GHz. (b) Input tones at 42.5 and 43 GHz.
50 5
MVT MVT
40 QPM 4 QPM
SGC SGC
Power (W)

Power (W)

30 3

20 2

10 1

0 0
41 42 43 44 45 46 41 42 43 44 45 46
Frequency (GHz) Frequency (GHz)

(c) Input tones at 43.5 and 44 GHz. (d) Input tones at 44 and 44.5 GHz.
2 60
MVT
Output power (dBm)

50 SGC
Power ratio (dB)

0 QPM/A
40 QPM/B

30
−2
SGC 20
QPM
−4 10
−0.6 −0.4 −0.2 0 0.2 0.4 0.6 42 43 44 45
Two-tone spacing ∆f (GHz) Frequency (GHz)

(e) Main-tone power ratio as a function of (f) Two-tone results. (A): narrowband,
frequency spacing, with f1 = 43 GHz. (B): broadband reference data.

Figure 3.30: Simulation of dual-tone signals using MVTRAD and Envelope methods.

80
 3.5 Conclusion

narrowband data range leads to a considerable deviation already in the neighboring

intermodulation frequencies and overall leads to an erroneous estimation.

3.5 Conclusion
In this chapter, various possibilities of evaluating the multi-frequency behavior of TWTs
have been assessed. The full-wave PIC approach in CST naturally enables the full
characterization. Although modeling procedures to accelerate and simplify the model
generation and simulation are presented, this still is a high-effort approach. It might
be suitable for analysis purposes, with calculation times of around half a day for a CW
excitation, but cannot be used efficiently at the design stage. In contrast, MVTRAD,
which is already well-established for the delay-line design, can fulfill this task. Still,
as the MVTRAD results are generated by calculating the greatest common divisor
of the tones and scaling the time resolution according to the number of harmonics,
this significantly increases the total interaction simulation time. Depending on the
common fundamental frequency, the latter may exceed half a day for the presented
scenarios. In principle, the MVTRAD multi-tone approach does not scale graciously,
limiting the applicability for scenarios with many harmonics. For the Ku-Band TWT
simulation, a comparison of the computation time in MVTRAD and in CST’s PIC
code for a number of two-tone excitations with varying frequency-spacing is listed in
Table 3.2. It can be seen that while large frequency spacings can easily be simulated
in MVTRAD and CST, the computation time rapidly rises, reducing their usability.
They are especially not helpful for modulated signals with continuous spectra, except
in some special, selected cases, as described in Section 3.4.3.
As an alternative, Envelope methods have been evaluated. The Envelope simulation
itself can be done within minutes, the computation time for the hybrid approach is
therefore mainly determined by the frequency-domain module. Employing parallel
computing and considering as few frequency points as possible would reduce the total
computation time to a few additional minutes. Two Envelope models have been
evaluated in detail. For typical communication TWTs with helical delay-lines, e.g., in
the Ku-Band, the basic SGC model has proven most practical, except for more involved
phase-related properties. Here, the more intricate QPM is at advantage, although it
requires more input data and thus in total the simulation takes considerably longer
than with the SGC model. Also, for modern applications in higher frequency bands,
for instance Q-Band, the QPM can surely play to its strengths, as more dispersive
topologies are used.

81
3 Simulation Models for Modulated Signals

Table 3.2: Comparison of approximate computation times in CST and MVTRAD.

In the fast mode, only frequencies in the range of (fi ± 5∆f ) and second
harmonics are considered.

MVTRAD CST PIC

∆f (MHz) Period (ns)
Harmonics Normal mode Fast mode (with GPU)

– (CW) ≈0.1 3 ≈ 10 s ≈ 10 s 12 h
500 2 60 2.5 h 20 min 16 h
250 4 120 11.5 h 50 min 18 h
125 8 240  1d 2.5 h 24 h
62.5 16 480 − 13 h 35 h
40 25 750 −  1d 2.5 d
31.25 32 960 − − 3d

Thus, the usability of the presented simulation tools depends on the considered
scenario. For continuous spectra and narrow bandwidths, the SGC model is the
method of choice. For a broader bandwidth, the QPM is more accurate. Both the
QPM and the SGC model are less helpful, the larger the tone distance is, as the
computation time in MVTRAD and CST reduces to more practical ranges and their
results are more trustworthy for larger frequency bands, due to their sound physical
background. Of course, CST has the largest range of applicability, but as it is the
most effort-expensive option, it mainly serves as a fall-back solution. Anyway, the
simulation of random complexly modulated signals with very large bandwidths is not
feasible in any of the tools, but also represents the least likely problem.

82
4 Back-Off Analysis and
Optimization
In Chapter 3, different tools to evaluate the behavior of TWTs in modulated operation
have been investigated. While the analysis of an already designed TWT is in itself an
interesting purpose for these tools, they obviously can also be used to improve the
device during the design process. This allows for the development of TWTs, which
are especially suited for modern applications where increasingly complex modulation
schemes are utilized. These are in principle less robust against distortions resulting
from the nonlinear nature of the TWT. Therefore, these distortions are required to be
as low as possible, which is ensured by driving the TWT far in back-off. Unfortunately,
the high linearity comes at the cost of low output power and, which is even worse,
a strongly diminished total efficiency. Comparing the total efficiency to the output
power, one can see that the drive level at their maximum values, as well as their
characteristics over drive, are slightly different. Therefore, in principle, a dedicated,
separate optimization of the two quantities is imaginable.
There are two evident set-screws for the optimization of the TWT. As the main
contributor of the total efficiency in back-off is the collector, a change in the collector
will surely lead to a different efficiency reduction in back-off. It is advantageous to
change the collector, since it is located at the end of the TWT, after the interaction
has ended. This means that no change in the other amplifier characteristics is to be
expected, but also that only little leeway is available. In contrast, the performance
of TWTs strongly relies on the pitch profile of the delay line, as has been established
in the previous chapters. Therefore, the design of this profile is essential for the
adaptation of the TWT to its planned application, offering a large range of possible
improvements of certain characteristics. Unfortunately, this usually happens at the
cost of one or more other, possibly similarly important characteristics.
Within the framework of the validation of the different tools in Chapter 3, data
was collected continuously. It has been shown that the dependencies between the
design of the TWT itself, the usual gain characteristics, and the considered effective
quantities in modulated operation cannot be read directly from the simulated and

83
4 Back-Off Analysis and Optimization

measured results. Instead, an initially quite complex, multidimensional problem with

many possible set-screws is revealed. Their influence cannot be clearly described due
to manifold interactions. Thus, it is difficult to make quantitative statements about
how, for example, the delay line would have to be adapted in order to carry out an
optimization in the back-off range for common and future modulation types.

4.1 Analysis of Nonlinear Effects

In the following, at first the impact of the nonlinearities on typical amplifier character-
istics is evaluated, to enhance the understanding of the relationship between classical
and effective modulated characteristics. Then, with these results in mind, ways to
tackle the problem of back-off optimization are investigated.

4.1.1 Asymmetric Intermodulation

Simulation and measurements of dual-tone intermodulation product ratios have shown
little variation over frequency in relevant bandwidths for typical communication
TWTs. Still, for large frequency spacings ∆f between the tones, a more slanted
power spectrum is found, as expected, as the signal then spans over frequencies
with significantly different transfer properties. The analysis of this slanting is rather
complicated when considering communication signals, as in principle it depends on the
actual modulation scheme. Also, the tilting of the spectral properties due to amplitude
and phase distortions cannot be separated and figures of merit can hardly be defined
unambiguously. Thus, as before, two-tone signals are taken into account.
Figure 4.1 shows QPM simulation results for a dispersive and a non-dispersive
scenario, comparing two-tone excitations to QPSK modulated signals at equal input
powers. For the sake of visual comparability, the power spectral density obtained from
the output waveforms by means of an FFT is shifted, such that the two-tone peaks at
the input frequencies are positioned at the average in-band value of the modulated
signal spectrum. It can be seen that the dual-tone response serves as a reasonable
indicator for the communication signal’s spectral behavior and the expected slanting.
Further experience gained throughout the verification study of the models in Chapter 3
leads to a similar conclusion. Therefore, the well-defined two-tone figures of merit are
considered to evaluate the spectral tilting.
In Figure 4.2, the intermodulation product ratio as simulated with MVTRAD is
investigated over frequency and drive, using the highly dispersive alternative reference
Ku-Band TWT, highlighting the asymmetry between the powers at the upper and

84
 4.1 Analysis of Nonlinear Effects

20 20
Shifted FFT (dB/Hz)

Shifted FFT (dB/Hz)

0 0

−20 −20

−40 −40

−60 −60

−80 −80
−2 0 2 −2 0 2
Frequency/Bandwidth Frequency/Bandwidth

(a) Non-dispersive case. (b) Dispersive case.

Figure 4.1: Simulation of equal-power two-tone and QPSK signals in SIMBA.

the lower intermodulation frequency. The lower tone is fixed at a frequency near the
center of the band and the higher one is shifted from it by ∆f .
The results for D3 and D5 show a similar behavior. While the average intermodula-
tion ratio obviously decreases strongly towards saturation, the curves do not vary much
with ∆f . More surprisingly, the asymmetry, expressed here by the difference between
upper and lower intermodulation ratio, is nearly constant over drive, especially in com-
parison to the variation of the intermodulation level with the input power. In contrast
to ∆f itself, which has a clearly visible impact, the degree of distortion seems to only
marginally influence the asymmetry. This has indirectly been seen in the Envelope
models, as this hints at the separability and chainability of frequency-dependence and
nonlinearity, allowing the application of the proposed models.
Another interesting property is the increase of the asymmetry for higher order
intermodulation products. This can be understood as a result of the larger impact
of the observed gain differences for higher order products and the corresponding
frequencies being further apart, typically resulting in a larger variation of the TWT’s
properties. Anyway, this finding is not of high relevance, as the intermodulation ratio
rapidly increases with its order and the asymmetry occurs around a roughly constant
average. Typically only the third order products are within a relevant power range in
back-off.
A similar study is conducted for the phase transfer. It is asymmetrically defined, as
one of the two tones is classically chosen significantly smaller than the other one. As
typically only small bandwidths are considered, it is often not even specified which one

85
4 Back-Off Analysis and Optimization

0.10 GHz 0.20 GHz 0.30 GHz 0.50 GHz 0.75 GHz 1.00 GHz

40 8

Intermod. asymmetry (dB)

Avg. intermod. ratio (dB)

6
30
4
20
2

10 0
−15 −10 −5 0 −15 −10 −5 0
Input back-off (dB) Input back-off (dB)

(a) Intermodulation product ratio. (b) Intermodulation product asymmetry.

Figure 4.2: Asymmetry of the intermodulation products over frequency and drive.
The legend on top is valid for both plots and denotes the frequency
spacing. The third order values are denoted by solid lines, the fifth order
values by dashed lines.

of the carriers is supposed to be weaker. Here, considering three different frequency

spacings from 0.2 GHz to 1 GHz, the asymmetry of kT is investigated. For this purpose,
using MVTRAD, first the lower tone is varied, recording the effect on the phase at
the higher frequency, and afterwards the opposite problem is simulated. As for D3
and D5 , the lower tone f1 is fixed at a frequency near the center of the band and the
higher one f2 is placed at f1 + ∆f .
Figure 4.3a shows the phase transfer in back-off for both cases. The difference
between the two increases with ∆f and input drive level. Their mean value does
not depend on ∆f , though. This again hints at the separability of the functional
blocks as mentioned in Section 3.3. A larger frequency spacing relates to a larger
asymmetry, which differs depending on whether the tone at f1 or the one at f2 is
altered. Additionally, it can be seen that the phase transfer is generally larger when
the lower frequency is driving the TWT, as the phase change is observed at a higher
frequency, and phase effects typically amplify in the direction of larger frequencies.
Surprisingly, when operating at f2 , the phase transfer further in back-off does not
seem to be a function of the spacing between the tones. This could be explained by a
similar back-off level at f2 for all three ∆f .

86
 4.1 Analysis of Nonlinear Effects

8 5
Phase transfer kT (deg/dB)

Phase transfer kT (deg/dB)

P (f1 ) > P (f2 ) (A)
P (f2 ) > P (f1 ) 4 (B)
6
3
4
2
0.2 GHz 0.2 GHz
2
0.5 GHz 1 0.5 GHz
1 GHz 1 GHz
0 0
−15 −10 −5 0 −15 −10 −5 0
Input back-off (dB) Input back-off (dB)

(a) Unequal input drive with 15 dB difference (b) Equal input drive. (A) Phase change at
between the two tones. f2 from drive change at f1 , and (B) vice-
versa.

Figure 4.3: Asymmetry of the phase transfer kT versus frequency spacing ∆f and
input drive for equal and unequal carriers, with f1 < f2 . Input back-off
here denotes the ratio between the total input power at f1 and f2 and
saturation at f1 .

Instead of an asymmetric excitation, a phase transfer can also be defined for equally
driven carriers. This is shown in Figure 4.3b, where the kT values are in general lower
than in Figure 4.3a. This complies with the theory in [37], where for low input powers
a factor of two is estimated. Again, the spread depends on the frequency spacing, but
it is more asymmetric now, as both tones drive the TWT. The mean value changes
little, as in the unequal case. For instance, at around 10 dB in back-off, the average
kT from the equal carrier case lies between 1.4 ◦ dB−1 and 1.6 ◦ dB−1 , and from the
unequal one it lies between 2.7 ◦ dB−1 and 3.2 ◦ dB−1 .

4.1.2 Impact on Communication Signals

As is seen, for instance, in Section 3.4.3, the interpretation of distortion metrics depends
on the considered signal itself. Differently modulated signals have different underlying
envelope distributions and encounter differently distributed distortions, even if their
average power is similar. Figure 4.4 shows the variation of typical characteristics
with various modulation schemes, i.e., from a simple QPSK to a complex 256QAM
signal. The simulation is carried out in SIMBA, using a frame of 8000 symbols and a
signal-to-noise ratio (SNR) of 20 dB.

87
4 Back-Off Analysis and Optimization

Figure 4.4b shows the effective bit energy per noise power spectral density (EbNo).
In contrast to the classical EbNo definition, where the bit energy is only related to the
power density from an added noise source, the effective EbNo is calculated from the
error vector due to the combination of added noise and distortions in the TWT:
1 Ph
 i 
(Ik )2 + (Qk )2

N k 
EbNoeff. = 10 log10  . (4.1)
 
 log2 (Ms ) P  2 
 2  
Ik − Iˆk

+ Qk − Q̂k
N k

Here, Ms denotes the number of points in the constellation diagram. The effective
EbNo additionally takes the signal distortions into account and therefore relates to an
effective noise power. As expected, a high symbol-error rate (Figure 4.4a) corresponds
to a low effective EbNo. This is in accordance with theory, from which it is known
that higher order modulation schemes are more sensitive to noise, as the constellation
points are effectively closer for equal average power.
Similarly to the intermodulation products from a two-tone excitation the distortions
for a modulated carrier yield a staircase-like spectrum to both sides. Thus, undesired
power is generated in neighboring bands. Larger distortions correspond to a larger
share of power found outside the carrier bandwidth. Therefore, it is helpful to consider
the ratio between in- and out-of-band power, as is shown in Figure 4.4c. For higher
drive levels, one can see that the QPSK and 16PSK signals with their low PAPRs show
significantly lower distortions than the other schemes. With increasing complexity, the
distortions increase, but converge towards a limiting curve. The same holds for the
MER, as shown in Figure 4.4d, which also increases for more complex modulation. Here,
a difference between 16QAM and 16APSK is visible due to the different distribution
of constellation points on two rings of equal power for 16APSK and three for 16QAM.
Finally, the effective efficiencies in Figure 4.4e and Figure 4.4f show that involved
multi-amplitude modulations flatten the behavior over drive, as they consist of a
broader range of amplitude levels over time, resulting in a weighted average of the
CW efficiencies. The flattening does not occur boundlessly, but does not change much
between the higher-order modulation schemes.
Typically, to reduce distortions at high output powers, TWTs are operated in
combination with a linearizer. As the linearizer is inserted in front of the TWT, the
now distorted signal envelope distribution at the input of the TWT might significantly
change compared to a setting without linearization. A comparison of the envelope
distributions from two exemplary modulation schemes, the 16APSK and the 64QAM, is
presented in Figure 4.5. Figure 4.5a and Figure 4.5b show the envelope distribution for
a 16APSK modulation at the input and output terminals of the TWT, both with and

88
 4.1 Analysis of Nonlinear Effects

QPSK 16 PSK 16 APSK 16 QAM 64 QAM 256 QAM

10−2 20

Effective EbNo (dB)

15
Symbol-error rate

10−3
10
−4
10
5
10−5
0

10−6 −5
−20 −15 −10 −5 0 5 −20 −15 −10 −5 0 5
Pin rel. to single-tone Pin,sat. (dB) Pin rel. to single-tone Pin,sat. (dB)

(a) Symbol-error rate. (b) Bit energy per noise power density.
30 0

25
−10
Power ratio (%)

MER (dB)

20
−20
15
−30
10

5 −40
−20 −15 −10 −5 0 5 −20 −15 −10 −5 0 5
Pin rel. to single-tone Pin,sat. (dB) Pin rel. to single-tone Pin,sat. (dB)

(c) Ratio between in- and out-of-band power. (d) Constellation distortions.
90 80
Collector efficiency (%)

Total efficiency (%)

60
88

40
86

20
84
−20 −15 −10 −5 0 5 −20 −15 −10 −5 0 5
Pin rel. to single-tone Pin,sat. (dB) Pin rel. to single-tone Pin,sat. (dB)

(e) Effective collector efficiency. (f) Effective total efficiency.

Figure 4.4: Nonlinear distortions and efficiency reduction over drive from SIMBA for
a non-linearized Ku-Band TWT. The legend on top is valid for all plots.

89
4 Back-Off Analysis and Optimization

without linearizer. To depict the behavior over drive, three different output back-off
levels are considered, starting in small-signal regime at 10 dB below saturated output
power, up to large-signal excitations at 3 dB and 0.5 dB below saturation. The same is
shown in Figure 4.5c and Figure 4.5d for 64QAM modulation. In both cases, one can
see that for small-signal excitation, as expected, no considerable changes in the input
or output distribution are observable. As the gain compression and nonlinear phase
shift are small, the linearizer does not need to alter the signal strongly. This can also
be observed in the effective EVS plots in Figure 4.5e and Figure 4.5f. In contrast to
this, for large-signal excitation, the envelope distribution shows a significant deviation
in front of the TWT. Instead of an upwards power-shifted distribution, it is modified
in a way that looks flattened and more spread. This increases closer to saturation.
Fortunately, as the spreading only distributes the curves further, the effect on the
EVS is small. As it is a result of averaging, the altered probabilities in parts cancel
each other out. For a 3 dB output back-off point, effectively no change in the collector
efficiency is to be expected between the linearized and the non-linearized setting.
Obviously, other characteristics, especially the linearity, are altered more significantly.
Thus, simply comparing the two settings at the same back-off operating point does not
provide a complete picture. Figure 4.6a shows the effective EbNo with and without
linearizer at different output power levels for different modulation schemes. A constant
SNR at the input is assumed, resulting in a base EbNo-level for small-signal excitation,
which depends on the size of the modulation alphabet1 . As soon as the nonlinearity-
based noise dominates the noise term, again roughly parallel lines can be observed,
while the offset is much different from the one from small-signal excitation. With
the linearizer, the same degree of distortion can be achieved at higher output power.
Following this it might, for instance, be of interest which change in drive is required
when changing from one modulation type to another and what this means for the
EVS. This is shown both with and without linearizer in Figure 4.6b. Here, for each
of the curves the output back-off value is chosen to reach an effective EbNo of 20 dB.
The change in drive together with the predistorted signal distribution results in a
significant change in the EVS and therefore not only in the beam efficiency, but also
in the collector efficiency. This relates to a considerably different optimal collector
stage voltage selection depending on the modulation scheme.
Thus, for the actual optimization of the TWT, it is beneficial to know whether it is
supposed to be linearized or not, and which modulation is chosen. Following from that,
1
As input SNR, bandwidth, and signal sampling are fixed, the ratio of symbol energy to the noise
power spectral density is the same for all signals. In that case, the EbNo base-level variation only
stems from the relationship between symbol energy and bit energy.

90
 4.1 Analysis of Nonlinear Effects

-10 dB back-off (NL) -3 dB back-off (NL) -0.5 dB back-off (NL)

-10 dB back-off (L) -3 dB back-off (L) -0.5 dB back-off (L)

10−1 10−1

10−2
10−2
Probability

Probability
10−3
10−3
−4
10

10−5 −7 10−4 1
10 10−6 10−5 10−4 10−3 10 101.5 102
Power (W) Power (W)

(a) 16APSK: Input envelope distribution. (b) 16APSK: Output envelope distribution.
10−1 10−1

10−2
10−2
Probability

Probability

10−3
10−3
−4
10
10−4
10−5

10−6 −6 −5 −4 −3 −2
10−5 1
10 10 10 10 10 10 101.5 102
Power (W) Power (W)

(c) 64QAM: Input envelope distribution. (d) 64QAM: Output envelope distribution.
1.2 1.2
Cumulated current I/I0

Cumulated current I/I0

1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
Kinetic energy Ekin /Ekin,0 Kinetic energy Ekin /Ekin,0

(e) 16APSK: Effective EVS. (f) 64QAM: Effective EVS.

Figure 4.5: Simulation with (L) and without linearizer (NL). The legend on top
shows output back-off values and is valid for all plots.
91
4 Back-Off Analysis and Optimization

64QAM (NL) 16APSK (NL) 16QAM (NL)

64QAM (L) 16APSK (L) 16QAM (L)

40 1.2

Cumulated current I/I0

Effective EbNo (dB)

1
30 0.8
0.6
20 0.4
0.2
10 0
−20 −15 −10 −5 0 0 0.2 0.4 0.6 0.8 1 1.2
Output back-off (dB) Kinetic energy Ekin /Ekin,0

(a) Distortions versus output power. (b) EVS with and without linearizer.

Figure 4.6: Equal-distortion based comparison with (L) and without linearizer (NL).
The legend on top is valid for both plots.

the choice of the operating point plays a major role. For the operating point, various
figures of merit can be taken into account. At maximum efficiency, typically near the
saturated output power, too much distortion is introduced. Too far from saturation,
distortion is not problematic anymore, but possibly power combining techniques, and
thus more amplifiers, are required to reach the required signal power. Somewhere in
between, one could expect some sort of sweet spot, where no characteristic is too bad.
This spot consequently depends on the application.
It is also important to note that the effective EbNo or EVM, while it is a useful
metric for the distortion, does not directly translate to a symbol-error rate, as typically
presumed in literature. This is, as the theoretical relationship between EbNo and the
symbol- or bit-error rate is usually calculated assuming distortions by an additive
white Gaussian noise channel. Figure 4.7a and Figure 4.7b report the error-vector
distribution with and without linearizer for which approximately the same effective
EbNo is achieved. Not only is the distribution different from what would be expected
from additive white Gaussian noise, the linearized and non-linearized error vectors
are also different from each other. Both effects are especially strong for the 64QAM.
Thus, while the effective MER will agree, the symbol-error rate might differ and is to
be considered separately.

92
 4.2 Relating Single- and Multi-Tone Characteristics

0.2 0.2
Quadrature error

Quadrature error
0.1 0.1

0 0

−0.1 −0.1

−0.2 −0.2
−0.2 −0.1 0 0.1 0.2 −0.2 −0.1 0 0.1 0.2
Inphase error Inphase error

(a) 16APSK. (b) 64QAM.

Figure 4.7: Equal-distortion based comparison of the error-vector distribution with

(blue) and without (red) linearizer.

Unfortunately, the linearizer itself is generated based on the TWT, and the combined
amplifier in reality is somewhat different from an ideal limiter. As a result, it is not so
easy to consider a correctly linearized amplifier in an optimization of the TWT, but
only to estimate the linearizer according to the TWT, for instance based on experience.
Thus, in contrast to this analysis, where actual characteristics from a real TWT and
the real, associated linearizer could be used, the optimization routine proposed in
Section 4.3 only considers a non-linearized TWT.

4.2 Relating Single- and Multi-Tone

Characteristics
To understand how to adapt the TWT’s geometry in the most promising way towards
its operation with complex modulation schemes, one could think of using a sufficiently
large number of simulated and measured TWT data and a more or less intelligent
pattern recognition, either human or computer-based. However, such an approach is not
only highly uncertain with regards to the gain in knowledge, but also disproportionately
time-consuming and resource-intensive, due to the required amount of data to be
generated, as previous experience has shown.
Thus, to allow some degree of understanding, the development of suitable expressions
to directly describe the relationships between - desirably - the structure, CW amplifier
quantities, and communication characteristics is investigated in the following.

93
4 Back-Off Analysis and Optimization

Classical approaches for partial problems known from literature (e.g., [17]) already
show that this is not a simple task either. Even the connection between cold line, beam
quantities, and classical amplifier characteristics can only be described analytically
for very simple cases. For example, the amplification in the small-signal range for a
homogeneous line is given in Equation (2.11). However, as soon as inhomogeneities
and nonlinearities are to be taken into account, finding such expressions is rather
challenging. Reasons for this can be found in the origin of the nonlinear effects from
the ballistic behavior of the electron beam and the complex interaction between the
coupled systems as described in Section 2.3. However, these difficulties are less grave in
the development of expressions for the relationship between single-tone and multi-tone
quantities. This is due to the fact that the physical causes and effects are already
inherent components of the usual amplifier characteristics and the task can therefore
be lifted from a purely physical to a model level. This is also the basis for the usability
of Envelope based simulation tools as investigated in Section 3.3.

4.2.1 Low-Level Noise-Power Ratio Estimation

During the previous investigations, it was reported that different types of modulation
have a different influence on common effective communication metrics. A characteriza-
tion of the amplifier therefore always depends on the respective excitation. However, it
has also been shown that, for a fixed TWT design, each of these metrics tends towards
a limit when increasing the complexity of modulation. This happens as the signals
then become more and more similar to noise from a spectral point of view. Therefore,
it makes sense to use such a "noise-like" excitation as a means of characterization.
Likewise approaches have been established for a long time. Therefor, a band-limited,
notch-filtered input signal on the TWT is used to calculate the NPR, which has
shown to be useful to evaluate the distortion performance of the amplifier [65]–[67].
At the output, the nonlinearity of the TWT introduces many additional frequency
components, e.g., in the notch. The intermodulation power content inside this notch
or test channel follows a stochastic process, as the excitation itself is not deterministic.
The NPR is a metric, which describes the distortions introduced by the TWT and is
defined as [60]
( )
BNotch
R /2
E S2 (ω − ω0 ) dω
−BNotch /2 E {P2,ω0 }
NPR = ( ) = , (4.2)
BNotch
R /2 E {P2,ωref }
E S2 (ω − ωref ) dω
−BNotch /2

94
 4.2 Relating Single- and Multi-Tone Characteristics

where E{X} is the expected value of a random variable X, P2,ω0 denotes the portion
of the output power inside the notch and P2,ωref the power in a reference band outside
the notch. This parameter represents a limit both for the distortions of complex
modulations and for the intermodulation of multi-tone signals. The NPR can be
calculated in SIMBA using both frequency-independent and frequency-dependent
Envelope models. The calculation duration for the latter can sometimes be rather
long. As described above, this can be helpful in the analysis, but offers hardly any
hints for a more evolved synthesis.
In order to remedy this situation, approximate expressions are developed in the
following, using approaches that are comparable to Envelope methods. These should
not only enable a fast calculation of the NPR, but also provide insight into the
dependencies of classical quantities. Since this happens at a model level, it is now
primarily a mathematical task. Even for a few tones in the excitation, the nonlinearity
of the TWT produces an infinite number of intermodulation products. A basically
continuous spectrum, which has to be considered as excitation for the NPR, is at first
much more complex. Nevertheless, the problem can be simplified by taking various
assumptions into account. The full derivation consists of many purely algebraic steps,
which do not add value to understanding. In the following, these steps are cut to
present the derivation in a condensed form.
The calculation is carried out in two parts. For the numerator, we need to distinguish
the individual frequency components at the output. Therefor, an approach using 2N
CW carriers given by

A0 X
x1 (t) = √ · cos(ω0 t + n∆ωt + ϕn ), U = [−N, N ] \ {0}, (4.3)
2N n∈U
N
! !
2A0 X ϕn + ϕ−n ϕn − ϕ−n
=√ · cos ω0 t + · cos n∆ωt + , (4.4)
2N n=1 | 2 }
{z | 2 }
{z
=:ϕn =:ϕ
n

with amplitude A0 , frequency ω0 + n∆ω and random phase distributions ϕi ∈

unif(0, 2π), is chosen for the input signal, where the center carrier is not excited.
This assumption has, for some purposes, shown to be a useful representation of the
band-limited, notch-filtered input signal for which the NPR is defined. With ∆ω  ω0 ,
the cosine terms at n∆ω change slowly compared to the one at ω0 , and can each be
considered as an amplitude A(t, n) which slowly changes over time. Thus, further

95
4 Back-Off Analysis and Optimization

sorting yields2
N
2A0 X 2A0
x1 (t) = √ · A(t, n) cos(ω0 t + ϕn ) = √ |z| · cos(ω0 t + arg(z)), (4.5)
2N n=1 2N
where z = |z| · exp(j arg(z)) can be expressed by
v
u !2 !2
u N N
u X   X  
|z| = t cos n∆ωt + ϕ cos(ϕ n n) + cos n∆ωt + ϕn sin(ϕn ) (4.6)
n=1 n=1

and  
N
P  
 cos n∆ωt + ϕn sin(ϕn ) 
 n=1
arg(z) = arctan  N  

 =: ϕz (t). (4.7)
P 
cos n∆ωt + ϕn cos(ϕn )
n=1
This helps forming the analytic description considering the slowness of the cosine terms
at n∆ω
x1 (t) = x̂1 (t) cos(ω0 t + ϕz (t)), (4.8)
such that, with Equation (4.4) and Equation (4.8), we can define the transforms
 
cos(ω0 t + ϕz (t)) N  cos (ω0 t + ϕ )
2A0 X n
· x̂ 1 (t) = √ cos(n∆ωt + ϕ n
) · (4.9)
sin(ω0 t + ϕz (t)) 2N n=1  sin (ω0 t + ϕn ) .
Using an expansion around the operating point [38]3 and analogously to other
applications in [37], [38], one can approximate the signal x2 at the output of the TWT,
in case the signal amplitude is small compared to the saturated amplitude, by
! !
λs x̂21 (t) x̂21 (t)
x2 (t) ≈ x̂1 (t)G 1 − · cos ω0 t + ϕz (t) + λϕ 2 , (4.10)
2x̂21,sat x̂1,sat
where G denotes the stationary gain at the operating point, and the time-varying
deviations of the gain and the nonlinear phase shift are considered by a drive-dependent
power proportionality factor λs and a drive-dependent phase proportionality factor λϕ .
Considering x̂1 (t)  x̂1,sat for a small-angle approximation and with Equation (4.9),
we can rewrite Equation (4.10) as
!" #
λs x̂21 (t) λϕ x̂21 (t)
x2 (t) ≈ x̂1 (t)G 1 − cos(ω0 t + ϕz (t)) − 2 sin(ω0 t + ϕz (t))
2x̂21,sat x̂1,sat
(4.11)
2A0 G
= √ · (A1 (t) − λϕ A2 (t) − λs A3 (t) + λϕ λs A4 (t)) . (4.12)
2N
2
Generally speaking, we can always express the superposition of multiple phase shifted sinusoidal
signals at frequency ω0 as another sinusoidal signal at frequency ω0 .
3
According to [38], the origin of the expansion is adapted from [68], which unfortunately is not
available.

96
 4.2 Relating Single- and Multi-Tone Characteristics

with
N
X
A1 (t) = cos (ω0 t + ϕn ) cos(n∆ωt + ϕn ) (4.13)
n=1

describing an exclusively scalar amplification of the initial frequency components of

x1 (t), not having any components at ω0 , and
N
X x̂21 (t)
A2 (t) = 2
sin (ω0 t + ϕn ) cos(n∆ωt + ϕn ), (4.14)
n=1 x̂1,sat
N
X x̂21 (t)
A3 (t) = 2
cos (ω0 t + ϕn ) cos(n∆ωt + ϕn ), and (4.15)
n=1 2x̂1,sat
N
X x̂41 (t)
A4 (t) = 4
sin (ω0 t + ϕn ) cos(n∆ωt + ϕn ), (4.16)
n=1 2x̂1,sat

including frequency components due to intermodulation. They arise from the multi-
plication of cos(n∆ωt + ϕn ) with x̂21 (t). As afterwards no further mixing is done, it
is from now on sufficient to consider the Ai,ω0 instead of the Ai . The Ai,ω0 denote all
components of the Ai at ω0 . This leads to

2A20 X

A2,ω0 (t) = sin (ω0 t + ϕn )
4N x̂21,sat X
 
· cos(ϕ|m| − ϕ|p| ) cos ϕn + (m) · ϕ|m| + (p) · ϕ|p| , (4.17)

where  
 n ∈ [1, N ], 
X := n+m+p=0 , (4.18)
 m, p ∈ [−N, N ]\{0} 

and
x
(x) := , for x 6= 0. (4.19)
|x|
An analytic signal description according to Equation (4.5)-Equation (4.7) yields

A20
A2,ω0 (t) = |z2 | · sin(ω0 t + arg(z2 )), (4.20)
2N x̂21,sat

where
 
r (Q)

(I) 2
 
(Q) 2
 Az,i
|zi | = Az,i + Az,i , ϕz,i := arg(zi ) = arctan  (I)
, (4.21)
Az,i

with
 
(I)
Az,2  X    cos(ϕn )
(Q)
:= cos(ϕ|m| − ϕ|p| ) cos ϕn + (m) · ϕ|m| + (p) · ϕ|p|  (4.22)
Az,2  X sin(ϕn ).

97
4 Back-Off Analysis and Optimization

Analogously for the other amplitude factors, one can find

A20
A3,ω0 (t) = |z3 | · cos(ω0 t + arg(z3 )), (4.23)
4N x̂21,sat

with z3 = z2 , and finally, for the last term,

A40
A4,ω0 (t) = |z4 | · sin(ω0 t + arg(z4 )) (4.24)
8N 2 x̂41,sat

with  
(I)  cos(ϕ
Az,4  X n ),
(Q)
:= B(n, m, p, q, r) ·  (4.25)
Az,4  Y sin(ϕn ),
where B(n, m, p, q, r) is a function of phase relations similar to (but more involved
than) the one in Equation (4.17), and
 
 n ∈ [1, N ], 
Y := n+m+p+q+r =0 . (4.26)
 m, p, q, r ∈ [−N, N ]\{0} 

(I,Q) (I,Q)
Az,2 and Az,4 only depend on the number of considered carriers 2N , A0 denotes
the operating point, and λϕ , λs , G and x̂1,sat (t) are partly operating-point dependent
characteristics of the TWT. The proportionality factors λs (power) and λϕ (phase)
can be approximated by the known TWT transfer-curve characteristics, i.e., the
compression c and kp [38]
A2
c ≈ λs 2 0 , (4.27)
2x̂1,sat
and
A20
kp ≈ λϕ , (4.28)
2x̂21,sat
such that Equation (4.17) can now be represented by TWT-independent constants
and the two known TWT characteristics. Thus, the power in the notch results in

c2 kp2 ckp2 kp2

!
A20 G2 c2 kp c2 ckp
P2,ω0 = · a1 5 + a2 4 + a3 3 + a4 4 + a5 3 + a6 3 . (4.29)
2 N N N N N N

For the numerator of Equation (4.2), we need to calculate the expected value
(I,Q)
of Equation (4.29). The six parameters E{aj } (j ∈ [1, 6]) all consist of Az,2 and
(I,Q)
Az,4 , which in turn consist of combinations of the initial 2N uniformly distributed
(I,Q) (I,Q)
random variables ϕi (i ∈ U). For large N , the Az,2 and Az,4 will follow zero-mean
normal distributions. They are not fully independent from each other, but only weakly

98
 4.2 Relating Single- and Multi-Tone Characteristics

(I,Q) (I,Q)
correlated, with Cov{Az,2 , Az,4 } ≈ 0. Therefore, one can simplify
1  (I) 2  (Q) 2 1 1
  n 
o n o
(I) (Q)
E{a1 } = E Az,4 + Az,4 = Var Az,4 + Var Az,4 , (4.30)
4 4 4
1 (I) (Q)
  
(Q) (I)
E{a2 } = E Az,2 Az,4 − Az,2 Az,4 ≈ 0, (4.31)
2
1 1 1
   
(I) 2 (Q) 2
  n o n o
(Q) (I)
E{a3 } = E Az,2 + Az,2 = Var Az,2 + Var Az,2 , (4.32)
n
4 o
4 4
(Q) (Q) (I) (I)
E{a4 } = E Az,2 Az,4 − Az,2 Az,4 ≈ 0, (4.33)
E{a5 } = 0, and (4.34)
n o n o
(Q) (I)
E{a6 } = 4E {a3 } = Var Az,2 + Var Az,2 . (4.35)

For the denominator of Equation (4.2), the signal x1 (t) is directly considered as a
stochastic signal

x1 (t) = R · cos (ω0 t + ΦR ) , (4.36)

where R is, e.g., a Rayleigh distributed process, while ΦR follows a uniform distribution
in unif(0, 2π)4 . Analogously to Equation (4.10), we can now estimate the output signal
! !
R2 R2
x2 (t) ≈ RG 1 − λs 2 · cos ω0 t + ΦR + λϕ 2 (4.37)
2x̂1,sat x̂1,sat
!v
R2 t1 + λ2 R
4
u
u
= RG 1 − λs 2 ϕ 4
2x̂1,sat x̂1,sat
2
  
sin(ΦR ) + λϕ x̂2R cos(ΦR )
1,sat
cos ω0 t + arctan   . (4.38)
  
2
cos(ΦR ) − λϕ x̂2R sin(ΦR )
1,sat

Then, finally, the expected value of the total power at the output of the TWT is
n o
E {P2 } = E |x2 |2 (4.39)
( ! !)
R2 G2 λs R 2 λ2 R4 R4
=E 1 − 2 2 + s4 1 + λ2ϕ . (4.40)
2 2x̂1,sat 4x̂1,sat x̂41,sat
Here, we can use the approximate relationships
R2 R2
c ≈ λs 2 , kp ≈ λϕ 2 , (4.41)
x̂1,sat x̂1,sat
which then yield
2 
c
 
2 2
E {P2 } ≈ σ G 1− 1 + kp2 , (4.42)
2
4
The calculation does not require knowledge of the full distribution, but only of some of its moments.

99
4 Back-Off Analysis and Optimization

where σ is the scale parameter of the distribution R. In accordance with Equation (4.3),
√
it is chosen to σ = A0 / 2. Thus, the NPR can be estimated from
 
E {P2,ω0 } 2N · E {P2,ω0 } b1 c2 kp2 + b2 c2 + 4kp2
NPR = = ≈   , (4.43)
E {P2,ωref } E {P2 } c 2

1− 1 + kp2
2
which only depends on the two (fixed) scalars
n o n o
(I) (Q)
Var Az,4 + Var Az,4
b1 := ≈ 50.97 (4.44)
4N 4
and n o n o
(I) (Q)
2Var Az,2 + 2Var Az,2
b2 := ≈ 0.358 (4.45)
N2
and the two drive-dependent amplifier parameters c and kp .
Again, considering a reference Ku-Band TWT, a comparison between the approxi-
mate solution and simulation in SIMBA as shown in Figure 4.8a reports good agreement
far in back-off. This is especially the case when no phase distortion is considered
(kp = 0), but also with kp 6= 0 the expression yields close agreement up to relevant
back-off regions of around 10 dB in input back-off. The variation far in back-off in
the case without phase distortion stems from a simplification in the simulation with
SIMBA. For this, the nonlinear gain compression is related to the gain as observed with
a small-signal reference simulation. Thus, in SIMBA it starts at zero for a small-signal
input power level, while the discussion in Section 2.3 hints at the occurrence of a
nonlinear gain already in a purely ballistic setting. Theoretically, the gain compression
drops continuously towards zero-drive excitation, causing intermodulation at any
non-zero drive level.
Comparing the expression in Equation (4.43) to the third order intermodulation
ratio approximation from Equation (3.20), a relationship
 
b1 c2 kp2 + b2 c2 + 4kp2 0.1D3

2 2

NPR ≈  2  ∝
 ∼ 10 ≈ c + 4kp (4.46)
c
1− 1 + kp2
2
is found for small c and kp . Thus, the D3 can be translated to the NPR in logarithmic
scale by a constant, TWT-independent shift, when considering small input powers. In
Figure 4.8b, this is shown for four different AM-AM and AM-PM curve sets, serving
as a reference for the Envelope simulation of D3 and NPR in SIMBA. For this purpose,
an MVTRAD CW simulation is carried out for a large range of input powers, starting
from input powers of around 50 dB below saturation. Also, a fine sampling along

100
 4.2 Relating Single- and Multi-Tone Characteristics

−10
SIMBA
−20

NPR and −D3 (dB)

Estimate −20
NPR (dB)

−30
−40
−40

−60 −50

−60
−40 −30 −20 −10 −40 −30 −20 −10
Pin (dBm) Pin (dB)

(a) Comparison with (solid line) and without (b) NPR with constant shift (dashed line)
(dashed line) phase distortion. compared to D3 (solid line) for four dif-
ferent transfer-curve sets.
−5 −10 −10
−10 −5
−15 −15
−15
−10 −20 −20
−10
−2
−25 −25 0
−20 NPR (dB)
−15
kp (dB)

−30 −30
−25

−35 −30
−15

−3
5
−20 −40
−45 −40
−30

−20
−40

−25 −5
0 −50
−5
−30 5
−30 −25 −20 −15 −10 −5
c (dB)

(c) Contour plot of the NPR as a function of c and kp .

Figure 4.8: NPR results from estimate expression.

101
4 Back-Off Analysis and Optimization

the input power range is chosen, such that the reference curves are accurate and
smooth and can be interpolated with close to no error. The corresponding results
from the same reference data are denoted by equal coloring. Far in back-off the slopes
correspond well. As expected, this agreement worsens, the closer to saturation the
operating point is chosen.
At last, Figure 4.8c shows the impact of kp and c on the expected NPR. It can be
seen that a large improvement on, e.g., c does not help much, when kp is already large,
and vice-versa.

4.2.2 Frequency-Dependent Intermodulation Ratio

In Section 3.3.1, a simple two-tone expression from literature has been reviewed,
approximating the third-order intermodulation ratio for two equally strong tones. It
was conceived assuming constant transfer behavior across the considered input tones
and intermodulation products. Such a simplified expression can serve as a quick
reference and help understanding the effect of the individual nonlinear effects. To
extend this to a more broadband expression, the derivation is incorporated into a
TBM-like approach as introduced in Section 3.3.4. For this purpose, a two-tone signal
ω∆ ω∆
     
x(t) = A1 sin ω0 − t + A2 sin ω0 + t (4.47)
2 2
with similar amplitudes A1 ≈ A2 and close frequency spacing ω∆  ω0 is consid-
ered. Constant phase terms, unnecessary for the resulting expression, are neglected
throughout this derivation. The TBM approach consists of two linear filters framing
the nonlinearities. At the end of the first filter, a signal
ω∆ ω∆ ω∆ ω∆
         
y1 (t) = A1 G∆ − sin ω0 − t + A2 G∆ sin ω0 + t
2 2 2 2
(4.48)
ω∆
  
= A1 A2 A∆ (t) sin ω0 − t + ϕ∆ (4.49)
2
is obtained, with
v
u 2  ω∆  ω∆
  
ω∆
  
t ∆ − 2
uG G2∆ 2
2G∆ 2
G∆ − ω2∆
A∆ (t) = + + cos(ω∆ t) (4.50)
A22 2
A1 A1 A2
and  
ω∆
 
A2 G∆ 2
sin(ω∆ t)
ϕ∆ (t) = arctan  
ω∆
   , (4.51)
A2 G∆ 2
cos(ω∆ t) + A1 G∆ − ω2∆
where G∆ (ω) is the impact of the input filter compensating for the different degree of
nonlinearity seen at a frequency ω compared to the assumed reference center frequency

102
 4.2 Relating Single- and Multi-Tone Characteristics

ω0 . Analogously to Equation (4.10), the nonlinear transfer of the reference curve at

ω0 is applied to y1 (t), resulting in a signal y2 (t) at the output of the nonlinear block.
Considering trigonometric identities, a small-angle approximation, and neglecting
small terms, the components of interest are be extracted, where
s
ω∆ ω∆ A21 λ2s
   
A2,(ω0 − 3 ω∆ ) ≈ GA2 G∆ G2∆ − + λ2ϕ (4.52)
2 2 2 A2sat 4
s
 2
ω∆ ω∆ c
   
≈ GG∆ G2∆ − A2 + kp2 (4.53)
2 2 2

is the amplitude at ω0 − 32 ω∆ and

s
ω∆ ω∆ A22 λ2s
   
A2,(ω0 + 3 ω∆ ) ≈ GA1 G∆ − G2∆ + λ2ϕ (4.54)
2 2 2 A2sat 4
s
 2
ω∆ ω∆ c
   
≈ GG∆ − G2∆ A1 + kp2 (4.55)
2 2 2

is the amplitude at ω0 + 32 ω∆ . Similarly to Section 4.2.1, G denotes the stationary gain

and the time-varying deviations of the gain and nonlinear phase shift are considered
(1,2)
by the proportionality factors λs (power) and λϕ (phase). For D3 , also the powers
at the input frequencies are required. For simplicity, it can be assumed that at these
frequencies the linear amplification outweighs higher order intermodulation products.
Assuming A1 =A2 , the third order intermodulation ratio is calculated from
     
ω∆
G∆ − ω2∆ H∆ − 3ω2∆

(1)
G∆ 2
D3 ≈ D3,0 − 20 log10    , (4.56)
H∆ − ω2∆

and      
ω∆
G∆ − ω2∆ H∆ 3ω∆

(2)
G∆ 2 2
D3 ≈ D3,0 − 20 log10  
ω∆
 , (4.57)
H∆ 2

with s 
 2
c
D3,0 = −20 log10  + kp2  , (4.58)
2

following the definition from Equation (3.8). H∆ (ω) is the equivalent filter at the
output, compensating for the difference in output power seen at ω compared to ω0
(1,2)
for the same drive level. It can be seen that for both D3 an additive term is to be
considered compared to the static expression in Equation (3.20). This additional term
compensates for the difference in back-off and gain. Also, as expected, the two ratios
are different, resulting in an asymmetric distribution.

103
4 Back-Off Analysis and Optimization

0 2
Estimate
Output back-off (dB)

Avg. D3 − D3,0 (dB)

MVTRAD
1.5
−1

1
f
−2
0.5

−3 0
−10 −5 0 5 −1 −0.5 0 0.5 1 1.5
Input back-off (dB) ∆f (GHz)

(a) Normalized transfer curves over fre- (b) Average D3 − D3,0 from the approximate
quency. expressions.

Figure 4.9: TWT data from MVTRAD and intermodulation product results.

Figure 4.9a shows the AM-AM curves generated in MVTRAD for a reference Ku-
Band TWT. The curves are shifted by their saturation point, such that the shape
similarity over frequency in the band of interest can be seen. The expressions from
Equations (4.56) to (4.58) are evaluated for this reference TWT in comparison to
two-tone MVTRAD simulation. Figure 4.9b shows the results, when one tone is
fixed at the band center, while the other is placed at a distance ∆f from it. As the
expression is not valid for drive levels close to saturation, a realistic operating point
3 dB in output back-off is chosen. It can be seen that the frequency-dependence of the
broadband dual-tone intermodulation is well matched in the approximate expression
even up to a large frequency difference of 1.5 GHz in positive direction. For negative
∆f , this does not hold similarly well, which can directly be understood from the shape
variations of the transfer curves, which are more pronounced in this frequency range.
Thus, to apply the derived expressions, the transfer curves themselves need to be
thoroughly evaluated.
This is also one reason why the derivation of similar expressions for the estimation
of kT is not feasible with reasonable effort. The variations of phase-related properties
over frequency are significantly larger, not only considering an absolute shift, but
also in the behavior over drive. This only allows for frequency-dependent solutions,
where the considered band is rather narrow, rendering the resulting expressions more
or less useless. Similarly, this also holds for broadband NPR estimations, which are
considerably more demanding in terms of bandwidth and linearity.

104
 4.3 Automated Back-Off Optimization

4.3 Automated Back-Off Optimization

Although simple expressions to connect CW characteristics with ones from modulated

operation have been found, the interpretation of these expressions with respect to the
previously mentioned multidimensional problem with its many trade-offs still presents
an ambitious endeavor. An alternative, heuristic approach is found analogously to
the practical design of a delay line in the CW optimization of the TWT. Here, a
satisfactory solution for the structure is iteratively calculated according to a previously
defined target function.
The amplifier performance strongly relies on the pitch profile of the delay line,
both for CW and modulated operation. As one cannot find a simultaneous optimum
in all characteristics and no analytic large-signal formulas are available, no obvious
solution exists. In this context, manual optimization is fairly time- and therefore
resource-consuming. Thus, in the past, various automated optimization routines have
been developed [69]–[71]. For this purpose, algorithms based on Simulated Annealing
(SA, [72]) have shown to be especially valuable and therefore have been used in various
taper optimization approaches. Classically, these approaches are based on single-tone
requirements, as can be obtained by frequency-domain codes. These are, among others,
the saturated input and output power, nonlinear phase, small-signal gain, compression,
and frequency flatness. Lately, optimization tools taking the device nonlinearity and
thus the expected signal distortion into account have been published [73]. While
fulfilling their aim of designs with improved linearity for modulated operation, they
do not directly consider the effective efficiency, although it is the other main cause
for the performance decline in back-off. There, especially the excess collector heat
deteriorates the effective total efficiency.
In this work, which draws from previously published results in [74], CW goals,
nonlinear distortions, and collector, as well as beam efficiencies in modulated operation
are combined to enable SA taper optimization specifically for communication signals.
The advantage is, of course, the possibility to achieve a context-dependent adaptation
without much understanding of the relationships between the variables to be optimized
and the adjustment screws of the structure. The underlying optimization routine
chosen for this is based on a previously published one [75], where the basics of the
SA approach and the CW optimization workflow, as well as the CW goal function are
discussed in detail.

105
4 Back-Off Analysis and Optimization

ηC,eff, ηB,eff
Adaptive
(modulated) Collector
Taper Simulated
profiles and NPR stage voltage
Annealing
optimization

Rayleigh
CW
(NPR)
characteristics Envelope
@ f1, ..., fN simulation

Electron
MVTRAD
beam Effective
CW EVS
vs. EVS
drive (Rayleigh)
Cold data

Continuous-wave optimization Modulated back-off optimization

Figure 4.10: Workflow of the extended SA optimizer, which considers the modulated
back-off.

4.3.1 Method and Workflow

In the following, the general approach is shortly outlined. In the classical helix-taper
optimization, as is schematically shown in the left half of Figure 4.10, an initial taper
is simulated, e.g., in MVTRAD, with regards to its CW characteristics. Therefor,
a wide set of cold parameters is pregenerated, which covers the whole domain of
possible tapers. Also, auxiliary information, such as the beam data, are predefined
and invariant. To evaluate the performance against specified goals and adapt to the
taper accordingly, an SA approach is used. The idea is that similarly to a cooling
process of a melted material, the initially not-optimal state gradually "cools down" to a
well-ordered, low-energy state, which in this case means an optimized taper. Therefor,
in [75], an exponential cooling schedule

Tn+1 = Tn · α = T0 · αn , α ∈ [0, 1[ (4.59)

was chosen, which determines the temperature reduction from one iteration to another.
According to the current temperature inside the SA code, the next taper then is
generated by random, defined alterations of the current taper.
Here, the CW optimization is extended by an additional module, taking effective
quantities into account. Therefor, the CW results are processed in an Envelope-based
routine. Also, the EVS’s are recorded for a range of drive levels. This way, effective
values are calculated, which serve as an additional source for the goal function defined
in the SA approach. This extended module is shown in Figure 4.10.

106
 4.3 Automated Back-Off Optimization

4.3.2 Target Functions

The optimization is realized by minimizing a total energy goal function
 
M N
1 X wj Xj
E(current taper) = 
2
H(XG,i,j − XT,j ) (XG,i,j − XT,j )2  , (4.60)
M j=1 Nj XT,j i=1

which consists of a weighted error between the individual j-th partial goal of in total
M goals, with weighting factors wj , target values XT,j , and the current performance
value XG,i,j . For each goal value, Nj frequency points are considered, which might
differ according to the individual goal. H(·) denotes the unit step function, ensuring
that a goal which is fulfilled with a large margin does not obscure a different, missed
goal.
This way, both CW and effective figures are integrated into a single goal function
for the SA approach to decide on the performance of the taper. Individual weighting
of the results enables highlighting certain aspects for the optimization. Details on the
goal functions are given in the following.

Continuous-Wave Goals

The partial goal function for the CW goals consists of the typical amplifier characteris-
tics for various frequencies, spread along the considered frequency band5 . For each of
these frequencies, the input and output power, the nonlinear phase shift at saturation,
the linear gain, and the compression are recorded and evaluated against predefined
target values. This is as defined in [75]. For frequency flatness and power symmetry
about the band center, the results are compared at several frequencies.

Modulation-Based Goals

The extension of the optimizer towards modulated operation is enabled by appending

an Envelope method to the MVTRAD simulation and including modulation-based
goals to the SA algorithm. While in principle the broadband performance is of interest,
and thus the QPM would be advantageous in this respect, it comes with a large cost in
terms of computation time. In a multi-dimensional, complex, automated optimization,
the time per iteration is highly critical. For quick evaluation of the taper, the SGC
model is chosen. Its major limit here is given by the frequency flatness of the TWT
characteristics. Luckily, the CW goal description ensures the validity of such a model
5
In this case, three frequencies are chosen: two at the boundaries of the band and one in its center.
While this is obvious for classical optimization, another choice could have advantages considering
effective metrics.

107
4 Back-Off Analysis and Optimization

for realistic bandwidths, enabling the usability of the SGC model for the proposed
purpose.
In this approach, both effective beam and collector efficiency are considered. To
obtain the effective collector efficiency in modulated operation, the EVS I(V ) is taken
into account. Using MVTRAD, each taper is evaluated from small-signal drive to a few
dB into overdrive to generate a set of CW ICW (V, Pin ) results. This data is then used
to transform an envelope distribution into the corresponding effective EVS Imod (V ),
which relates to a weighted average of the ICW (V, Pin ) and enables the estimation of
effective collector efficiencies, as was outlined in Section 3.3.2.
The envelope distribution may vary significantly even in the scope of communication
signals. Thus, the effective collector efficiency strongly depends on the chosen input
signal modulation, which is not always known. To optimize the taper independently
from the signal choice, a distribution representing a more general setting is desired.
This is also the idea behind the NPR. The total signal from multiple superimposed
modulated carriers is noise-like in both inphase and quadrature component, with the
envelope distribution following a Rayleigh distribution. Its scale parameter, which is
the only parameter in the distribution, is related to the average amplitude of the signal.
For a fixed input or output back-off level, the procedure from Section 3.3.2 returns the
effective Imod (V ) for such a Rayleigh distributed signal. Thus, hopefully, the optimizer
also yields an effective electron spectrum, which shows better collectability compared
to a CW approach.
In the end, the efficiency strongly depends on the choice of the collector stage
voltages. For meaningful comparison between two designs, the respective best collector
potentials need to be found for each of them. The first and last stage have a protective
function for the device and are therefore left unchanged. Instead, for a four-stage
design, the two remaining stages are optimized for the given effective Imod (V ). As
the range of possible solutions, the dynamics of the function to be optimized, and
the required potential accuracy are all rather low, even a simple brute-force approach
suffices for this purpose. From the optimized stages, an effective collector efficiency
ηColl,eff is extracted.
The same considerations regarding the envelope distribution also hold for the
effective beam efficiency ηBeam,eff . Again, for consistency, a Rayleigh distribution is
considered to estimate the effective output power, from which the beam efficiency
is calculated. Similarly, additional communication-related routines are implemented.
For instance, the NPR is calculated using the Envelope approach from a Rayleigh
distributed input signal. While the linearity can also be considered by means of the CW
characteristics, the usage of the NPR enables a more direct understanding regarding

108
 4.3 Automated Back-Off Optimization

the achieved degree of distortion reduction. As the Rayleigh distribution serves as the
basis for the effective values, designs coming from this extended optimization routine
are in the following referred to as a Rayleigh optimized designs.

4.3.3 Results
To evaluate the optimizer, a Ku-Band helix TWT from a pure CW optimization, in the
following denoted as design (A), serves as the starting point for the proposed Rayleigh
approach. Throughout an optimization run, a large set of data from a variety of tapers
is typically generated and stored. This allows for some statistical investigation. In
Figure 4.11, the effective collector and total efficiencies in back-off are shown for a
selection of 3000 taper evaluations, each dot representing a single evaluation, and the
color reporting the efficiency value at 3 dB in back-off. Some losses, such as coupler
and filament losses are not considered, also the collector currents are estimated from
the EVS using the described common approximation. For each taper, the saturated
output power and phase are also recorded, such that the effective efficiencies are related
to the CW goals at saturation. Small nonlinear phase shifts and high saturated output
powers are desirable. Thus, the further down to the right a point is, the better are
its CW properties. Unfortunately, the trade-off between these characteristics is easily
visible, such that it is difficult to find such an optimal point. Instead, the evaluated
points orient themselves roughly on a line from bottom left to the top right. Also the
results considering the coloring of the dots are mostly as expected. A lower saturated
output power or nonlinear phase shift typically means that the beam is not well spent.
The remaining beam power is large and the collectability is high, leading to a high
effective collector efficiency. As for the effective total efficiency, it includes the effective
beam efficiency and therefore increases to the top right. Still, for a fixed output
power, one can see an increased effective total efficiency in the direction of decreasing
nonlinear phase shift, although the general trend is the opposite. This is a result of a
better EVS, even though the same amount of power is extracted.
The Rayleigh optimization starting with design (A) leads to design (B), which is
optimized for the effective collector efficiency ηColl,eff and the effective total efficiency
η0,eff in back-off. Both η0,eff and the CW total efficiency η0,cw are estimated with ηColl,eff ,
ηBeam,eff , ηColl,cw and ηBeam,cw , neglecting for instance coupler and heater losses, which
are nearly identical for (A) and (B). Both designs fulfill all CW goals.
In Figure 4.12, the CW and effective efficiencies for both designs are shown. In
Figure 4.12a, the collector efficiency is shown versus back-off. For CW excitation, it can
be seen that although the saturated values are similar, but slightly higher for the CW

109
4 Back-Off Analysis and Optimization

Effective collector efficiency Effective total efficiency

0.87 0.88 0.89 0.90 0.91 0.58 0.60 0.62 0.64 0.66 0.68
Nonlinear phase at saturation (deg)

Nonlinear phase at saturation (deg)

70 70

60 60

50 50

40 40

30 30

20 20
100 120 140 160 180 100 120 140 160 180
Saturated output power (W) Saturated output power (W)

(a) Effective collector efficiency. (b) Effective total efficiency.

Figure 4.11: Effective back-off efficiencies for a selection of evaluated tapers versus
CW output power and nonlinear phase shift.

optimized design (A), the collector efficiency in back-off strongly drops. This comes
from the fact that the optimization is carried out, such that the effective efficiency
is improved, leading to a more robust design than the singular point optimization in
saturation. This can again be seen for the Rayleigh excited results, where the drop in
back-off is in comparison flattened, but similarly well visible. This improvement directly
impacts the total efficiency. As both designs fulfill the output power requirement, the
beam efficiency is in consequence similar. The improved collector efficiency in back-off
is clearly evident, increasing the total efficiency in simulation by a few percent. In
total, for the effective values, a far more robust behavior over the drive is shown, with
the efficiency-wise optimal operating point moved towards a linearity-wise preferred
back-off point.
The exact improvement to be expected in reality cannot be determined directly and
is most definitely lower than the shown differences, as it requires perfect knowledge
of the spent beam over drive and how it behaves in the actual collector. A shortage
in this knowledge leads to errors mainly in the collector efficiency, which is the main
contributor to the improvement shown in Figure 4.12. For instance, the actual EVS is
uncertain, e.g., due to inaccuracies in the magnetic focusing field and the lack of thermal
electrons in the simulation, and can only be estimated. Also, a large contributor to
uncertainty is the collector current calculation itself, as it relies on a common but

110
 4.3 Automated Back-Off Optimization

95

Est. total efficiency (%)

Collector efficiency (%)
80
90
60
85
(A): ηC,eff 40 (A): ηeff
(B): ηC,eff (B): ηeff
80
(A): ηC,CW (A): ηCW
20 (B): ηCW
(B): ηC,CW
75
−20 −15 −10 −5 0 5 −20 −15 −10 −5 0 5
Input back-off (dB) Input back-off (dB)

(a) Effective collector efficiency. (b) Effective total efficiency.

Figure 4.12: Estimated CW and effective efficiencies of design (A) and (B) for an ideal
four-stage collector, neglecting some loss mechanisms. CW excitation is
shown by dashed lines, Rayleigh excitation by solid lines.

simplified procedure, assuming a perfectly sorting, symmetric multiple-stage collector

without secondary electrons and other secondary effects [1]–[3]. While this has shown
to be a valid approximation, it introduces some uncertainty in the optimum collector
stage voltages, which contribute to all other results. In contrast, comparisons based
on the nonlinearity and distortions are more trustworthy, as it was shown that the
integral transfer properties are not strongly influenced by the expected typical beam
uncertainties. Also, in contrast to the collector current estimation, no further sharp
approximation is required.
To optimize against a distribution representing a generalized modulation setting, the
Rayleigh distribution was chosen. To see the effect of the Rayleigh-based optimization
on a known modulation scheme, Figure 4.13 shows the effective total efficiency and
MER for a QPSK and a 16QAM signal, both versus output back-off. The improvement
in back-off efficiency both in CW excitation and Rayleigh excitation also shows for
the communication signals, the improvement being similarly large throughout the
back-off. Luckily, the improved efficiency does not come at the cost of larger signal
distortions, which can be seen in the MER, representing the device nonlinearity. In
the relevant range around −3 dB to −5 dB output back-off, close to no change is seen
for the 16QAM signal, and only around 0.5 dB difference for the QPSK signal.
In Section 4.1.2, some issues regarding the operating point choice and change
according to the chosen modulation scheme were discussed. Here, a fixed output

111
4 Back-Off Analysis and Optimization

70 40
Effective total efficiency (%)

(A)QPSK
60 (B)QPSK
(A)16QAM 30

MER (dB)
50 (B)16QAM

40 20 (A)QPSK
(B)QPSK
30 (A)16QAM
10
(B)16QAM
20
−10 −8 −6 −4 −2 0 −10 −8 −6 −4 −2 0
Output back-off (dB) Output back-off (dB)

(a) Effective total efficiency. (b) Distortions.

Figure 4.13: Improvement of Rayleigh optimized design (B) compared to the CW

optimized design (A), evaluated with communication signals.

back-off point is considered for the modulation goals. Together with the CW output
power goal, this enables the comparison of the linearity throughout all evaluations.
Predistortion is not considered here. Considering a linearizer would increase the
complexity and uncertainty of the obtained results, as the real setup of linearizer
and TWT differs from an ideal limiter, and the linearizer itself is typically generated
based on TWT data. For instance, one might try to design the linearizer with the
aim of minimizing phase distortions, possibly at the cost of the linear gain, and
vice-versa. Depending on the operating point definition, collector stage number,
linearization, and other auxiliary constraints, fundamentally different results and
margins for improvement can be expected. Therefore, such a taper optimization is
most promising when the actual application is well-defined, which unfortunately rarely
is the case. Thus, the generalized Rayleigh approach, as it is considered here, helps
obtaining structures, which cover improvements over a wide range of applications.

112
5 Conclusion
Traveling-wave tubes are high-efficiency vacuum electronic power amplifiers, which are
predominantly used in satellite communications. Classical characterization is based
on continuous-wave characteristics, such as the saturated input and output power,
nonlinear phase, small-signal gain, compression, and frequency flatness. As a means
to evaluate the linearity, additionally narrowband quantities such as the noise-power
ratio, intermodulation product ratio, and phase transfer factor are evaluated.
Lately, modern satellite applications shift the modulation schemes from simple ones,
e.g., constant-envelope quadrature phase-shift keying, to more complex amplitude- and
phase-modulated signals. Thus, as more complex schemes typically are more sensitive
to nonlinear distortions, the amplifier is typically operated further in the linear back-off
region. This way, nonlinearity-induced modulation errors are diminished, whereas at
the same time the overall efficiency is typically severely reduced. Thus, both linearity
and efficiency in back-off require more attention than ever.
This thesis explores various means for enabling the analysis and optimization of the
traveling-wave tube for future satellite applications. For this purpose, in Chapter 2,
the main components of a traveling-wave tube and their role in the amplification
process are briefly described. A qualitative description of the operating principle is
given, followed by a brief explanation of the reasons behind the existence of nonlinear
distortions in traveling-wave tubes, and how to compensate for them. Important
terminology regarding the amplifier characterization is introduced, garnished with an
overview over common efficiency definitions.
In Chapter 3, methods and models for the simulation of multi-tone intermodulation
products and modulated communication signals are investigated. For this purpose,
time-, frequency-, and hybrid-domain codes are considered. Following a short intro-
duction into frequency-domain multi-frequency simulation in MVTRAD, approaches
for efficient models for different parts of full-wave helix traveling-wave tube simulation
in CST Particle studio are presented. Comparison with both MVTRAD and measure-
ment data is shown. The generated reference CST model is fitted by including cold
measurement data and tuning the support rod permittivity values. Continuous-wave
interaction simulation results, i.e., gain compression, nonlinear phase shift, and the

113
5 Conclusion

electron energy spectrum agree well with measurements and an MVTRAD model with
directly fitted parameters. Thus, the resulting models enable the study of transients,
multi-tone, and multi-carrier simulation, provided that time and computational effort
are not crucially limited.
As especially for optimization purposes fast solutions are required, Envelope meth-
ods are investigated as hybrid frequency- and time-domain approaches combining
advantages of both worlds. They are used to calculate the output behavior of the
traveling-wave tube based on input signal quantities and a set of reference data of
the traveling-wave tube. Various models from literature are evaluated and adapted
for accurate and fast estimation of typical output characteristics. The range of mod-
els reaches from simple two-tone approximations for the phase transfer factor and
the intermodulation product ratio, over frequency-independent models such as the
Static-Gain Curve model, up to the rather involved Quadrature-Polynomial Model.
Results from the validation of MVTRAD, CST Particle Studio, and different
Envelope codes are compared, revealing their respective advantages and disadvantages
based on the considered scenarios. For continuous-wave and far-spaced multi-tone
signals, both CST Particle Studio and MVTRAD have shown to be viable, trustworthy
options. But it is also apparent that their not-so-gracious scaling regarding computation
time for more intricate settings limits their applicability in many cases, especially
considering complex communication signals. Current, typical communication traveling-
wave tubes are so flat over frequency, that in typical bands the Static-Gain Curve
model suffices for the linearity characterization, especially considering the prediction of
powers at the intermodulation levels. The prediction of the phase transfer and faster
transients, on the other hand, can benefit from the more sophisticated Quadrature-
Polynomial Model. In the case of multi-tone signals with bandwidths, where the
Static-Gain Curve model starts to lose accuracy, MVTRAD and CST become more
useful. The simulation of future traveling-wave tube types, e.g., folded-waveguides,
has shown further areas where the Quadrature-Polynomial Model excels. These delay
lines are significantly more dispersive than classical helix lines and bring Envelope
models to the limit of their capabilities. Still, the Quadrature-Polynomial Model is
able to reliably predict the slope across the output spectrum.
For most simulation and measurement results shown here, a state-of-the-art 150 W
commercial reference TWT for communication in Ku-Band is used, such that the ex-
tracted statements are most meaningful for practical use. To enhance the occurrence of
certain frequency-dependent effects, for some accordingly tagged results, an artificially
distorted, more dispersive version of said traveling-wave tube is used in simulation.

114
 In addition to the models themselves, for reproducibly using them, the SIMBA
program was written. It incorporates different Envelope methods for the traveling-
wave tube with and without linearization in a satellite downlink-chain model with
modulation and demodulation of communication signals as defined in the DVB-S2
standard. Various communication characteristics are included to investigate the
back-off behavior of the amplifier.
Based on the developed models, together with the experience and data gained
throughout their validation, in Chapter 4 the impact of the nonlinearities on typical
communication characteristics is evaluated, with the purpose of enhancing the under-
standing of the relationship between classical and effective modulated characteristics.
After a briefly summarized study on the effects on communication signals, simple
expressions relating well-known transfer properties from continuous-wave simulation
to involved characteristics of multi-tone simulation are derived. Based on a third
order intermodulation ratio derivation from literature, a simple expression for the
estimation of the noise-power ratio is generated, which offers accurate results up to
typical back-off ranges. It has been shown that far in back-off, the noise-power ratio is
closely related to the third order intermodulation ratio. As a result, it is possible to
translate one to the other by a constant shift in logarithmic scale. Also, to extend
the mentioned intermodulation ratio expression from literature, it is combined with a
Three-Box Envelope approach.
As the interpretation of the expressions still remains a difficult task, an automated
optimization routine based on a previous continuous-wave optimization is developed.
For this, an optimizer based on Simulated Annealing is extended to include modulated
operation as a part of the goal function. It uses the electron velocity spectrum and
noise-power ratio to evaluate the effective collector and beam efficiencies and improve
the traveling-wave tube’s behavior in modulated back-off.
While this work provides some tools to analyze and optimize the traveling-wave
tube for modern communication satellite systems, there are many related topics left to
be investigated in future work. For instance, the pool of Envelope model topologies is
far from exhausted and element of ongoing research. Also, with DiMoHa, a proprietary,
specialized alternative in the time-domain field is under development and has recently
shown progress in terms of simulation time and validation. While potentially faster
than general purpose tools, its usability and reliability for the simulation of modulated
signals is to be evaluated in the future.
Within the scope of this thesis, the collector could only be considered in a reduced
manner. Further study on three-dimensional collector simulation with modulated
signals could be beneficial for better understanding of the collector’s behavior, although

115
5 Conclusion

the integrated currents have shown to be accurately and consistently predicted. Based
on the presented results, a study on collector adaptions for higher collector efficiency in
back-off could be of interest. Apart from using more stages to better follow the shape of
the electron velocity spectrum, literature offers interesting, more unconventional ideas,
e.g., using a time-dynamic, envelope following bias of the collector stages. Such methods
might be investigated following the method for calculating the effective electron velocity
spectrum. Likewise, this also holds for some other efficiency improvement methods,
such as certain waveform-altering signal-injection schemes.
As an alternative to adapting the collector, one could attempt shaping the electron
spectrum for better collectability, without sacrifices in terms of beam efficiency and
linearity. The many mechanisms leading to a system of trade-offs between most
important characteristics so far do not allow for such an ambitious endeavor. Still,
complimentary study could bring further light into the darkness of the traveling-wave
tube’s broadband back-off behavior.
Overall, one can summarize that the modulated operation of traveling-wave tubes
offers many challenging research topics. All this shows that even after all these
years the traveling-wave tube has not yet reached its limits. It offers plenty of room
for improvement to meet the ever rising demands of the satellite communications
community and its future applications.

116
117
List of Acronyms
APSK Amplitude- and Phase-Shift Keying
AM-AM Amplitude-Modulation to Amplitude-Modulation
AM-PM Amplitude-Modulation to Phase-Modulation
BFM Bessel-Function Model
CW Continuous-Wave
DVB-S Digital Video Broadcasting – Satellite
EbNo Effective Bit-Energy to Noise Power Spectral Density
EVS Electron-Velocity Spectrum
EVM Error-Vector Magnitude
FFT Fast Fourier Transform
FW Folded-Waveguide
MER Modulation-Error Ratio
MVTRAD MouVemenT RADial
MVT MVTRAD
NPR Noise-Power Ratio
PIC Particle-In-Cell
PAPR Peak-to-Average Power Ratio
PPM Periodic Permanent Magnet
PSK Phase-Shift Keying
PSB Poza-Sarkozy-Berger
QPSK Quadrature Phase-Shift Keying
QPM Quadrature Polynomial Model
QAM Quadrature-Amplitude Modulation
RF Radio-Frequency
RRC Root-Raised Cosine
SNR Signal-to-Noise Ratio
SA Simulated Annealing
SIMBA Simulation of Inter-Modulations in Back-Off and Applications
SSPA Solid-State Power Amplifier
SGC Static-Gain Curve
TBM Three-Box Model
TWT Traveling-Wave Tube

118
List of Figures

2.1 Schematic view of a helix TWT. . . . . . . . . . . . . . . . . . . . . . . 5

2.2 Two suitable delay-line types for satellite communications. . . . . . . . 7
a Helix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
b Folded-waveguide. . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Large-signal spent beam, the corresponding EVS, and the recollection
in a multi-stage depressed collector. . . . . . . . . . . . . . . . . . . . . 10
a Spent-beam energy in one RF period. . . . . . . . . . . . . . . . 10
b EVS from spent beams. . . . . . . . . . . . . . . . . . . . . . . 10
c Energy recovering in the collector. . . . . . . . . . . . . . . . . . 10
2.4 Charges on the circuit and resulting electric field for an RF signal
applied to a helical delay line. Facc and Fdec denote accelerating and
decelerating forces acting on the electrons, which, for reasons of clarity,
are downscaled in numbers and indicated by purple dots. . . . . . . . . 12
2.5 Amplifier characteristics of a TWT. . . . . . . . . . . . . . . . . . . . . 14
2.6 Schematic of a TWT with a preceding predistortion linearizer with
individual and combined transfer characteristics. . . . . . . . . . . . . . 17
2.7 Schematic view of the power flow in a TWT. . . . . . . . . . . . . . . . 18
2.8 Approximate total efficiency of a TWT as a function of the collector
and beam efficiency. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3.1 Injection of particles and equivalent beam port with potential rings. . . 28
a Distribution of injection points in the injection plane with cor-
responding initial energy vectors. The color denotes the initial
velocity vz in z-direction. . . . . . . . . . . . . . . . . . . . . . . 28
b Setup for the electrostatic simulation with a defined number of
potential rings in the injection plane. The outer hull at zero
potential is not shown. . . . . . . . . . . . . . . . . . . . . . . . 28
3.2 Workflow for the beam port inside the PIC interaction simulation in
CST Particle Studio. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

119
List of Figures

3.3 Equivalent loss mechanism model. The transparency of the sever slices
represents the loss profile. The losses per slice decrease symmetrically
from the center of the sever towards the ends. . . . . . . . . . . . . . . 29
3.4 Sever performance and sensitivity to sever profile and total attenuation. 31
a Performance of a sever with 75 dB total attenuation from the
perturbational approach. (A): reflection S11 , (B): transmission
S21 , (C): time-gated reflection S11 . . . . . . . . . . . . . . . . . 31
b Influence of the sever onset on the small-signal gain. (A): Refer-
ence TWT, (B): steeper transition of the loss taper, (C): smaller
total attenuation, (D): larger total attenuation. . . . . . . . . . 31
3.5 Full model of the interaction region with the tapered helix wire, two
severs, support rods, and discrete ports at in- and output. . . . . . . . 33
3.6 Comparison of different properties of the TWT from simulation in CST
and MVTRAD with measured results. . . . . . . . . . . . . . . . . . . 34
a Small-signal gain versus frequency. . . . . . . . . . . . . . . . . 34
b Output power versus input back-off. . . . . . . . . . . . . . . . . 34
c Nonlinear phase versus input back-off. . . . . . . . . . . . . . . 34
d EVS at three different back-off levels. . . . . . . . . . . . . . . . 34
3.7 Comparison of macro-particle trajectories at f0 from CST and MVTRAD. 35
a Zero-drive excitation. . . . . . . . . . . . . . . . . . . . . . . . . 35
b Saturation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.8 Comparison between the two reference frequency decision schemes for
a large-signal simulation at ∆f = 500 MHz around the amplification
band center fc . The error is compared to reference simulation carried
out in MVTRAD. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.9 SGC simulation of collected currents and the total efficiency for 16QAM
signals based on measured characteristics. (A): CW measurements.
(B): 16QAM measurements. (C): 16QAM simulation with SGC models
of the collector stages. (D): 16QAM simulation with an SGC model of
the total efficiency as the output characteristic. . . . . . . . . . . . . . 45
a Collector currents at the stages C1 to C4. . . . . . . . . . . . . 45
b Total efficiency. . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.10 Estimation of the velocity distribution of the spent-beam current. . . . 47
a Modulated two-tone signal. . . . . . . . . . . . . . . . . . . . . 47
b Amplitude distribution occurrence probability. . . . . . . . . . . 47
c CW EVS for various power levels. . . . . . . . . . . . . . . . . . 47
d EVS from the two-tone signal. . . . . . . . . . . . . . . . . . . . 47

120
 List of Figures

3.11 Schematic of the hybrid SGC model. . . . . . . . . . . . . . . . . . . . 48

3.12 Serial model topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.13 Typical chain of nonlinear and frequency-dependent models in a Three-
Box model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.14 Schematic view of the TBM, adapted from [53]. . . . . . . . . . . . . . 53
3.15 Typical structure of models in quadrature topology according to [53]. . 54
3.16 Simplified block schematic of the BFM. . . . . . . . . . . . . . . . . . . 56
3.17 Full topology of the QPM and the model identification scheme. . . . . . 58
3.18 SIMBA program written in MATLAB. . . . . . . . . . . . . . . . . . . 60
3.19 Schematic view of the signal chain from a single run with a communi-
cation signal in SIMBA. . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.20 Output power at main- and side-tone and intermodulation ratios for
various measurement-based models compared to measurements. . . . . 64
a Third- and fifth-order intermodulation product ratio at ∆f =
100 MHz spacing. . . . . . . . . . . . . . . . . . . . . . . . . . . 64
b Main tone and intermodulation product at ∆f = 500 MHz spacing. 64
3.21 Phase transfer factor for different frequency spacings ∆f . . . . . . . . . 66
a ∆f = 5 MHz: Phase transfer. . . . . . . . . . . . . . . . . . . . 66
b ∆f = 5 MHz: Error. . . . . . . . . . . . . . . . . . . . . . . . . 66
c ∆f = 200 MHz: Phase transfer. . . . . . . . . . . . . . . . . . . 66
d ∆f = 200 MHz: Error. . . . . . . . . . . . . . . . . . . . . . . . 66
e ∆f = 1 GHz: Phase transfer. . . . . . . . . . . . . . . . . . . . 66
f ∆f = 1 GHz: Error. . . . . . . . . . . . . . . . . . . . . . . . . 66
3.22 Two-tone signals around fc with a frequency spacing of ∆f at 0.5 dB
input back-off. The legend on top is valid for both plots. . . . . . . . . 67
a ∆f = 100 MHz. . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
b ∆f = 500 MHz. . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.23 Two-tone signals with a frequency spacing of ∆f at 0.5 dB input back-
off. The legend on top is valid for both plots. . . . . . . . . . . . . . . . 69
a Tones around 0.95fc with ∆f = 500 MHz. . . . . . . . . . . . . 69
b Tones around 0.925fc with ∆f = 200 MHz. . . . . . . . . . . . . 69
3.24 Transient processes in SGC and QPM compared to CST simulation. . . 71
a Unfiltered ideally stepped signal from zero-drive to large-signal
excitation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
b Unfiltered ideal step (A), and 1 GHz RRC filtered step (B), from
zero-drive to large-signal excitation. . . . . . . . . . . . . . . . . 71

121
List of Figures

c RRC filtered signal steps with 1 GHz bandwidth. (ZD): zero-

drive, (SS): small-signal, (LS): large-signal excitation. . . . . . . 71
3.25 Constellation diagrams from 800 symbols each, for a QPSK and a
16QAM signal at a drive level of 5 dB above CW saturation. . . . . . . 72
a QPSK: SGC model in SIMBA. . . . . . . . . . . . . . . . . . . 72
b QPSK: Measurements. . . . . . . . . . . . . . . . . . . . . . . . 72
c 16QAM: SGC model in SIMBA. . . . . . . . . . . . . . . . . . . 72
d 16QAM: Measurements. . . . . . . . . . . . . . . . . . . . . . . 72
3.26 Output power and distortions of QPSK and 16QAM modulated signals
from measurements and simulation. The legend on top is valid for both
plots. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
a Total output power. . . . . . . . . . . . . . . . . . . . . . . . . 74
b Distortions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.27 Output spectrum of a modulated signal with neighboring QPSK channel
from CST and the SGC model. . . . . . . . . . . . . . . . . . . . . . . 74
a Spectrum around the carrier frequency. . . . . . . . . . . . . . . 74
b Spectrum across several harmonics of the carrier frequency. . . . 74
3.28 Exemplary periodic QPSK sequence simulated in MVTRAD and SIMBA. 76
a Output power spectrum of an examplary QPSK sequence. . . . 76
b (A): MVTRAD, with components around f0 and 2f0 . (B):
MVTRAD, only components around f0 . (C): SGC model. . . . 76
c Output power spectrum of another examplary QPSK sequence. 76
d Quadrature eye-diagrams from Envelope simulation and MV-
TRAD. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.29 CW simulation results of the investigated FW-TWT. . . . . . . . . . . 78
a Output power over frequency for small- to large-signal input
power levels, simulated in MVTRAD. . . . . . . . . . . . . . . . 78
b Simulation results from CST, MVTRAD and KlysTOP at 0 dBm
input power. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.30 Simulation of dual-tone signals using MVTRAD and Envelope methods. 80
a Input tones at 42 and 42.5 GHz. . . . . . . . . . . . . . . . . . . 80
b Input tones at 42.5 and 43 GHz. . . . . . . . . . . . . . . . . . . 80
c Input tones at 43.5 and 44 GHz. . . . . . . . . . . . . . . . . . . 80
d Input tones at 44 and 44.5 GHz. . . . . . . . . . . . . . . . . . . 80
e Main-tone power ratio as a function of frequency spacing, with
f1 = 43 GHz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

122
 List of Figures

f Two-tone results. (A): narrowband, (B): broadband reference

data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

4.1 Simulation of equal-power two-tone and QPSK signals in SIMBA. . . . 85

a Non-dispersive case. . . . . . . . . . . . . . . . . . . . . . . . . 85
b Dispersive case. . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2 Asymmetry of the intermodulation products over frequency and drive.
The legend on top is valid for both plots and denotes the frequency
spacing. The third order values are denoted by solid lines, the fifth
order values by dashed lines. . . . . . . . . . . . . . . . . . . . . . . . . 86
a Intermodulation product ratio. . . . . . . . . . . . . . . . . . . 86
b Intermodulation product asymmetry. . . . . . . . . . . . . . . . 86
4.3 Asymmetry of the phase transfer kT versus frequency spacing ∆f and
input drive for equal and unequal carriers, with f1 < f2 . Input back-off
here denotes the ratio between the total input power at f1 and f2 and
saturation at f1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
a Unequal input drive with 15 dB difference between the two tones. 87
b Equal input drive. (A) Phase change at f2 from drive change at
f1 , and (B) vice-versa. . . . . . . . . . . . . . . . . . . . . . . . 87
4.4 Nonlinear distortions and efficiency reduction over drive from SIMBA
for a non-linearized Ku-Band TWT. The legend on top is valid for all
plots. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
a Symbol-error rate. . . . . . . . . . . . . . . . . . . . . . . . . . 89
b Bit energy per noise power density. . . . . . . . . . . . . . . . . 89
c Ratio between in- and out-of-band power. . . . . . . . . . . . . 89
d Constellation distortions. . . . . . . . . . . . . . . . . . . . . . . 89
e Effective collector efficiency. . . . . . . . . . . . . . . . . . . . . 89
f Effective total efficiency. . . . . . . . . . . . . . . . . . . . . . . 89
4.5 Simulation with (L) and without linearizer (NL). The legend on top
shows output back-off values and is valid for all plots. . . . . . . . . . . 91
a 16APSK: Input envelope distribution. . . . . . . . . . . . . . . . 91
b 16APSK: Output envelope distribution. . . . . . . . . . . . . . . 91
c 64QAM: Input envelope distribution. . . . . . . . . . . . . . . . 91
d 64QAM: Output envelope distribution. . . . . . . . . . . . . . . 91
e 16APSK: Effective EVS. . . . . . . . . . . . . . . . . . . . . . . 91
f 64QAM: Effective EVS. . . . . . . . . . . . . . . . . . . . . . . 91

123
4.6 Equal-distortion based comparison with (L) and without linearizer (NL).
The legend on top is valid for both plots. . . . . . . . . . . . . . . . . . 92
a Distortions versus output power. . . . . . . . . . . . . . . . . . 92
b EVS with and without linearizer. . . . . . . . . . . . . . . . . . 92
4.7 Equal-distortion based comparison of the error-vector distribution with
(blue) and without (red) linearizer. . . . . . . . . . . . . . . . . . . . . 93
a 16APSK. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
b 64QAM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.8 NPR results from estimate expression. . . . . . . . . . . . . . . . . . . 101
a Comparison with (solid line) and without (dashed line) phase
distortion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
b NPR with constant shift (dashed line) compared to D3 (solid
line) for four different transfer-curve sets. . . . . . . . . . . . . . 101
c Contour plot of the NPR as a function of c and kp . . . . . . . . 101
4.9 TWT data from MVTRAD and intermodulation product results. . . . 104
a Normalized transfer curves over frequency. . . . . . . . . . . . . 104
b Average D3 − D3,0 from the approximate expressions. . . . . . . 104
4.10 Workflow of the extended SA optimizer, which considers the modulated
back-off. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.11 Effective back-off efficiencies for a selection of evaluated tapers versus
CW output power and nonlinear phase shift. . . . . . . . . . . . . . . . 110
a Effective collector efficiency. . . . . . . . . . . . . . . . . . . . . 110
b Effective total efficiency. . . . . . . . . . . . . . . . . . . . . . . 110
4.12 Estimated CW and effective efficiencies of design (A) and (B) for an ideal
four-stage collector, neglecting some loss mechanisms. CW excitation is
shown by dashed lines, Rayleigh excitation by solid lines. . . . . . . . . 111
a Effective collector efficiency. . . . . . . . . . . . . . . . . . . . . 111
b Effective total efficiency. . . . . . . . . . . . . . . . . . . . . . . 111
4.13 Improvement of Rayleigh optimized design (B) compared to the CW
optimized design (A), evaluated with communication signals. . . . . . . 112
a Effective total efficiency. . . . . . . . . . . . . . . . . . . . . . . 112
b Distortions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
List of Tables
3.1 Currents from CST for simple geometries and realistic electron gun
models, compared to suitable references. . . . . . . . . . . . . . . . . . 26
3.2 Comparison of approximate computation times in CST and MVTRAD.
In the fast mode, only frequencies in the range of (fi ± 5∆f ) and second
harmonics are considered. . . . . . . . . . . . . . . . . . . . . . . . . . 82

125
Bibliography
[1] A. Gilmour, Principles of Traveling Wave Tubes, ser. Artech House radar library.
Artech House, 1994.
[2] R. J. Barker, N. C. Luhmann, J. H. Booske, and G. S. Nusinovich, Modern
Microwave and Millimeter-Wave Power Electronics. IEEE Press, Apr. 2005.
[3] J. A. Eichmeier and M. Thumm, Vacuum Electronics: Components and Devices.
Springer Berlin Heidelberg New York, 2008.
[4] G. Kornfeld and E. Bosch, “From History to Future of Satellite TWT Amplifiers”,
Frequenz, vol. 55, no. 9-10, pp. 258–262, 2001.
[5] J. M. Weekley and B. J. Mangus, “TWTA versus SSPA: A Comparison of
On-Orbit Reliability Data”, IEEE Transactions on Electron Devices, vol. 52,
no. 5, pp. 650–652, May 2005.
[6] E. F. Nicol, B. J. Mangus, J. R. Grebliunas, K. Woolrich, and J. R. Schirmer,
“TWTA Versus SSPA: A Comparison Update of the Boeing Satellite Fleet On-
orbit Reliability”, in 14th IEEE International Vacuum Electronics Conference
(IVEC), May 2013, pp. 1–2.
[7] E. F. Nicol and J. M. Robison, “TWTA On-Orbit Reliability for Satellite
Industry”, IEEE Transactions on Electron Devices, vol. 65, no. 6, pp. 2366–2370,
Jun. 2018.
[8] European Telecommunications Standards Institute, “Digital Video Broadcasting
(DVB); Second Generation Framing Structure, Channel Coding and Modula-
tion Systems for Broadcasting, Interactive Services, News Gathering and other
Broadband Satellite Applications”, en, European Telecommunications Standards
Institute, Standard ETSI EN 302 307-1, 2014.
[9] ——, “Digital Video Broadcasting (DVB); Second Generation Framing Structure,
Channel Coding and Modulation Systems for Broadcasting, Interactive Services,
News Gathering and other Broadband Satellite Applications; Part 2: DVB-S2
Extensions (DVB-S2X)”, en, European Telecommunications Standards Institute,
Standard ETSI EN 302 307-2, 2015.

127
Bibliography

[10] D. S. Komm, R. T. Benton, H. C. Limburg, W. L. Menninger, and X. Zhai,

“Advances in Space TWT Efficiencies”, IEEE Transactions on Electron Devices,
vol. 48, no. 1, pp. 174–176, Jan. 2001.
[11] P. Waller, “Modélisation numérique de l’interaction et diagnostic expérimental
du faisceau d’électrons dans un tube à ondes progressives spatial”, PhD thesis,
Université Paris 7, Paris, 1999.
[12] T. M. Antonsen Jr and B. Levush, “CHRISTINE: A Multifrequency Parametric
Simulation Code for Traveling Wave Tube Amplifiers”, Naval Research Lab
Washington DC Vacuum Electronics Branch, Tech. Rep., 1997.
[13] CST Particle Studio, https://www.cst.com/.
[14] MAGIC Tool Suite, https://www.orbitalatk.com/magic/.
[15] S. Meyne, “Simulation and Design of Traveling-Wave Tubes with Folded-
Waveguide Delay Lines”, PhD thesis, Technische Universität Hamburg-Harburg,
2017.
[16] R. E. Collin, Foundations for Microwave Engineering, ser. IEEE Press on Elec-
tromagnetic Wave Theory. John Wiley & Sons., 2001.
[17] J. R. Pierce, “Traveling-Wave Tubes”, The Bell System Technical Journal, vol. 29,
no. 2, pp. 189–250, 1950.
[18] J. Copeland and A. A. Haeff, “Andrei Haeff and the Amazing Microwave Ampli-
fier”, IEEE Spectrum, vol. 2015, no. 9, pp. 32–37, 2015.
[19] A. Katz, “Linearization: Reducing distortion in power amplifiers”, IEEE Mi-
crowave Magazine, vol. 2, no. 4, pp. 37–49, 2001.
[20] P. Thouvenin, D. Henry, and A. Pelletier, “New Helix Tapers Boost Space TWT
Efficiency to 55%, Broadband”, in 1987 International Electron Devices Meeting,
doi: 10.1109/IEDM.1987.191463, vol. 33, 1987, pp. 477–480.
[21] F. André, P. Bernardi, N. M. Ryskin, F. Doveil, and Y. Elskens, “Hamiltonian
Description of Self-Consistent Wave-Particle Dynamics in a Periodic Structure”,
EPL (Europhysics Letters), vol. 103, no. 2, 2013, doi: 10.1209/0295-5075/103/
28004.
[22] D. F. G. Minenna, Y. Elskens, F. André, A. Poyé, J. Puech, and F. Doveil,
“Dimoha: A time-domain algorithm for traveling-wave tube simulations”, IEEE
Transactions on Electron Devices, vol. 66, no. 9, pp. 4042–4047, Sep. 2019.
[23] S. Sensiper, “Electromagnetic Wave Propagation on Helical Conductors”, Mas-
sachusetts Institute of Technology Technical Report, 1951.

128
 Bibliography

[24] P. K. Tien, “Traveling-Wave Tube Helix Impedance”, Proceedings of the IRE,

vol. 41, no. 11, pp. 1617–1623, Nov. 1953.
[25] F. Shen, Y. Wei, H. Yin, Y. Gong, X. Xu, S. Wang, W. Wang, and J. Feng,
“A Novel V-Shaped Microstrip Meander-Line Slow-Wave Structure for W-band
MMPM”, IEEE Trans. on Plasma Science, vol. 40, no. 2, pp. 463–469, Feb. 2012.
[26] G. Ulisse and V. Krozer, “W -Band Traveling Wave Tube Amplifier Based on
Planar Slow Wave Structure”, IEEE Electron Device Letters, vol. 38, no. 1,
pp. 126–129, Jan. 2017.
[27] X. Bian, M. Miao, and Z. Li, “Design and Analysis of Micro-Machined Folded
Waveguide Slow-Wave Structure for 220 GHz Application”, in 2016 IEEE MTT-S
Int. Microwave Workshop Series on Advanced Materials and Processes for RF
and THz Applications (IMWS-AMP), doi: 10.1109/IMWS-AMP.2016.7588384,
2016.
[28] F. Li, T. Ma, and Y. Liu, “Design of W Band Three-Section Folded Waveguide
Traveling Wave Tube”, in 2016 IEEE Int. Vacuum Electronics Conf. (IVEC),
doi: 10.1109/IVEC.2016.7561897, Apr. 2016, pp. 1–2.
[29] D. F. G. Minenna, Y. Elskens, F. André, J. Puech, A. Poyé, F. Doveil, and
T. Pereira, “DIMOHA: Traveling-Wave Tube Simulations Including Band Edge
and Multiple-Carriers Operations”, in 2019 International Vacuum Electronics
Conference (IVEC), Apr. 2019, pp. 1–2.
[30] D. Safi, P. Birtel, S. Meyne, and A. F. Jacob, “Helix Traveling-Wave Tube
Interaction Simulation with CST Particle Studio”, in 2017 IEEE Int. Vacuum
Electronics Conf. (IVEC), 2017.
[31] ——, “A Traveling-Wave Tube Simulation Approach with CST Particle Studio”,
IEEE Transactions on Electron Devices, vol. 65, no. 6, pp. 2257–2263, 2018.
[32] I. Langmuir and K. B. Blodgett, “Currents Limited by Space Charge Between
Concentric Spheres”, Physical Review, vol. 24, no. 1, 1924, doi: 10.1103/physrev.
24.49.
[33] J. E. Rowe, Nonlinear Electron-Wave Interaction Phenomena. Academic Press,
2013, doi: 10.1016/c2013-0-08102-9.
[34] P. R. R. Rao and S. K. Datta, “Estimation of Conductivity Losses in a Helix
Slow-Wave Structure Using Eigen-Mode Solutions”, in 2008 IEEE International
Vacuum Electronics Conference, doi: 10 . 1109 / IVELEC . 2008 . 4556452, Apr.
2008, pp. 99–100.

129
Bibliography

[35] C. L. Kory and J. A. Dayton, “Investigation of the effect of support rod per-
mittivity on TWT performance”, in IEEE Conference Record - Abstracts. 1997
IEEE International Conference on Plasma Science, May 1997, p. 199.
[36] R. M. Gagliardi, Introduction to Communications Engineering, ser. A Wiley-
Interscience Publication. Wiley, 1978.
[37] J. Bretting and P. Treytl, “Größen zur Charakterisierung der Nichtlinearitäten
von Mikrowellen-Sendeverstärkern und deren Zusammenhänge”, NTZ-Kurier,
Berlin, 1972.
[38] J. Bretting, “Der Betriebswirkungsgrad von Wanderfeldröhren unter Berücksich-
tigung von nichtlinearen Verzerrungen”, PhD thesis, RWTH Aachen, 1971.
[39] G. Gatti, W. Bosch, S. C. Soriano, R. M. R. de Gopegui, and Y. Vichery, “IMAL-
2 a Simulation Software for Inter-Modulation Analysis of Power Amplifiers”,
XRM Tech. Notes, 1996.
[40] D. Gabor, “Theory of Communication. Part 1: The Analysis of Information”,
Journal of the Institution of Electrical Engineers-Part III: Radio and Communi-
cation Engineering, vol. 93, no. 26, pp. 429–441, 1946.
[41] B. Boashash and G. Jones, Instantaneous Frequency and Time-Frequency Distri-
butions. Longman Cheshire, 1992.
[42] L. Marple, “Computing the Discrete-Time "Analytic" Signal via FFT”, IEEE
Transactions on Signal Processing, vol. 47, no. 9, pp. 2600–2603, Sep. 1999.
[43] D. Safi, P. Birtel, S. Meyne, and A. F. Jacob, “Comparative Study of Envelope
Models for Multi-Tone Simulation of Traveling-Wave Tubes”, in 2017 47th
European Microwave Conference (EuMC), 2017, pp. 852–855.
[44] D. Safi, P. Birtel, S. Meyne, and A. F. Jacob, “Comparative Study of Envelope
Models for Hybrid Time- and Frequency-Domain Simulation of Traveling-Wave
Tubes”, International Journal of Microwave and Wireless Technologies, vol. 10,
no. 5-6, pp. 521–529, 2018.
[45] P. J. Loughlin and B. Tacer, “Comments on the Interpretation of Instantaneous
Frequency”, IEEE Signal Processing Letters, vol. 4, no. 5, pp. 123–125, 1997.
[46] L. Mandel, “Interpretation of Instantaneous Frequencies”, American Journal of
Physics, vol. 42, no. 10, pp. 840–846, 1974.
[47] G. Strang, M. Krieger, and K. Lippert, Wissenschaftliches Rechnen, ser. Springer-
Lehrbuch Masterclass Series. Springer Berlin Heidelberg, 2010.

130
 Bibliography

[48] M. Knorrenschild, Numerische Mathematik: Eine beispielorientierte Einführung.

Carl Hanser Verlag GmbH & Company KG, 2017.
[49] F. N. Fritsch and R. E. Carlson, “Monotone Piecewise Cubic Interpolation”,
SIAM Journal on Numerical Analysis, vol. 17, no. 2, pp. 238–246, 1980.
[50] A. A. Saleh, “Frequency-Independent and Frequency-Dependent Nonlinear Mod-
els of TWT Amplifiers”, IEEE Transactions on Communications, vol. 29, no. 11,
pp. 1715–1720, 1981.
[51] M. T. Abuelmaatti, “Frequency-Dependent Nonlinear Quadrature Model for
TWT Amplifiers”, IEEE transactions on communications, vol. 32, pp. 982–986,
1984.
[52] H. B. Poza, Z. A. Sarkozy, and H. L. Berger, “A Wideband Data Link Computer
Simulation Model”, in NAECON’75; Proceedings of the National Aerospace and
Electronics Conference, 1975, pp. 71–78.
[53] M. C. Jeruchim, P. Balaban, and K. S. Shanmugan, Simulation of Communication
Systems. New York, NY, USA: Plenum Press, 1992.
[54] L. Buderath, “Investigation of Nonlinear Multi-Box Models for Time-Domain
Simulation of Traveling-Wave Tubes”, Bachelor’s Thesis, TU Hamburg, Mar.
2017.
[55] P. Hetrakul and D. P. Taylor, “Nonlinear Quadrature Model for a Travelling-
Wave-Tube-Type Amplifier”, Electronics letters, vol. 11, no. 2, pp. 50–50, 1975.
[56] E. H. Moore, “On the Reciprocal of the General Algebraic Matrix”, Bull. Am.
Math. Soc., vol. 26, pp. 394–395, 1920.
[57] R. Penrose, “A Generalized Inverse for Matrices”, in Mathematical Proceedings
of the Cambridge Philosophical Society, Cambridge University Press, vol. 51,
1955, pp. 406–413.
[58] C. R. Johnson, Matrix Theory and Applications. American Mathematical Soc.,
1990, vol. 40.
[59] L. R. Turner, “Inverse of the Vandermonde Matrix with Applications”, NASA
Technical Note, 1966.
[60] R. W. Koch, “Random Signal Method of Nonlinear Amplitude Distortion Mea-
surement”, IEEE Transactions on Instrumentation and Measurement, vol. IM-20,
no. 2, pp. 95–99, May 1971.
[61] B. Evans, Satellite Communication Systems, ser. IET Telecommunications. In-
stitution of Engineering & Technology, 2011.

131
Bibliography

[62] S. O. Wallander, “Reflexions and Gain Ripple in TWT’s”, IEEE Transactions

on Electron Devices, vol. 19, no. 5, pp. 655–660, 1972.
[63] S. Meyne, P. Birtel, and A. F. Jacob, “Q-Band High-Power Traveling-Wave
Tube with Optimized Folded-Waveguide Delay Line”, in 2015 IEEE Int. Vacuum
Electronics Conf. (IVEC), doi: 10.1109/IVEC.2015.7223780, Apr. 2015, pp. 1–
2.
[64] S. Meyne, P. Bernadi, P. Birtel, J. F. David, and A. F. Jacob, “Large-Signal 2.5-D
Steady-State Beam-Wave Interaction Simulation of Folded-Waveguide Traveling-
Wave Tubes”, IEEE Trans. on Electron Devices, vol. 63, no. 12, pp. 4961–4967,
Dec. 2016, doi: 10.1109/TED.2016.2616948.
[65] J. Boyhan, “Normalized Power Spectra for FDM/FM Signals and Noise Power
Ratio Calculations”, IEEE Transactions on Communications, vol. 32, no. 1,
pp. 77–81, Jan. 1984.
[66] A. Darbandi and W. Zoccarato, “High Efficiency, Linearized Power Amplifier
for Communication Satellites”, in 1992 22nd European Microwave Conference,
vol. 2, Sep. 1992, pp. 1131–1136.
[67] C. Devieux, “Nonlinear Spectrum-Spreading Interference Effects Between Adja-
cent Satellite Transponders”, IEEE Transactions on Communications, vol. 23,
no. 11, pp. 1379–1385, Nov. 1975.
[68] O. Nilsson, Nonlinear Distortion In Traveling Wave Tubes, ser. Chalmers Univ of
Technology Goeteborg, Research Report (chalmers Tekniska Högskola. Forskn-
ingslaboratorium Foer Elektronik). Defense Technical Information Center, 1966.
[69] J. D. Wilson, “A Simulated Annealing Algorithm for Optimizing RF Power
Efficiency in Coupled-Cavity Traveling-Wave Tubes”, IEEE Transactions on
Electron Devices, vol. 44, no. 12, pp. 2295–2299, Dec. 1997.
[70] D. R. Whaley, L. K. Behnke, and J. A. McKay, “Application-Specific Multi-
Parameter TWT Optimization”, in 2007 IEEE International Vacuum Electronics
Conference, May 2007, pp. 1–2.
[71] J. A. David, C. L. Kory, H. T. Tran, R. L. Ives, and D. Chernin, “Enhanced
Features for Design of Traveling Wave Tubes Using CHRISTINE-1D”, IEEE
Transactions on Plasma Science, vol. 35, no. 4, pp. 1056–1064, Aug. 2007.
[72] S. Kirkpatrick, C. D. Gelatt, M. P. Vecchi, et al., “Optimization by Simulated
Annealing”, science, vol. 220, no. 4598, pp. 671–680, 1983.

132
 Bibliography

[73] J. X. Qiu, D. K. Abe, T. M. Antonsen, B. G. Danly, and B. Levush, “Traveling-

Wave Tube Amplifier Performance Evaluation and Design Optimization for
Applications in Digital Communications with Multilevel Modulations”, IEEE
Transactions on Microwave Theory and Techniques, vol. 51, no. 8, pp. 1911–1919,
Aug. 2003.
[74] D. Safi, P. Birtel, M. Wulff, S. Meyne, and A. F. Jacob, “Back-off Efficiency
Optimization Of Traveling-wave Tubes Using Simulated Annealing”, in 2018
IEEE International Vacuum Electronics Conference (IVEC), Apr. 2018, pp. 169–
170.
[75] S. Meyne, M. Krenz, and A. F. Jacob, “Taper Optimization for Helix Traveling-
Wave Tubes Using Adaptive Simulated Annealing”, in IEEE International Vac-
uum Electronics Conference (IVEC), Apr. 2014, pp. 143–144.

133