Everything you've always wanted to know about Hot-S22

(but we're afraid to ask)

Jan Verspecht
Jan Verspecht bvba
Gertrudeveld 15
1840 Steenhuffel
Belgium
email: contact@janverspecht.com
web: http://www.janverspecht.com
2002 Agilent Technologies - Used with Permission
Presented at the Workshop
Introducing New Concepts in Nonlinear Network Design
(International Microwave Symposium 2002)
1
Everything youve always
wanted to know about
Hot-S
22

(but were afraid to ask)

Jan Verspecht
Agilent Technologies
2
2
Purpose
Convince people of a better Hot S
22

Show that technology is fun (sometimes)

The purpose of this presentation is to convince people that one needs
to extend the classic concept of Hot S22 if one wants to describe
accurately how a device behaves under large-signal excitation with a
non perfect termination.
A second purpose is to show that, when the proper set of
experiments are performed, Hot S22 can be fun.
3
3
Outline
Introduction: What is Hot S
22
?
Getting and interpreting experimental data
Confront classic approaches with data
Derivation of the extended Hot S
22
theory
Confront extended Hot S
22
with data
Conclusion
Since there is a lot of confusion on Hot S22 I will first define what I
understand by it.
Next I will explain how to get and interpret practical experimental
data.
Then I will confront classical approaches to Hot S22 with this
experimental data and I will show that the classic approach is
inaccurate.
An accurate and extended Hot S22 is then derived from the
gathered experimental data.
I will show that the new Hot S22 model can indeed accurately
describe the measured data, in contrast with the classic approach.
I will finally draw conclusions.
4
4
What is Hot S
22
?
D.U.T. behavior is represented by
pseudo-waves (A
1
, B
1
, A
2
, B
2
)
Hot S
22
describes the relationship between
B
2
and A
2
Valid under Hot conditions (A
1
significant)
D.U.T.
A
1
A
2
B
1 B
2
So what do I mean by Hot S22?
Or a better question is what problem do we want Hot S22 to solve?
The answer is that one wants to describe how a device-under-test
(D.U.T.) behaves under variable load conditions at the output, while a
large-signal excitation is present.
Equivalent to classic linear S-parameters, Hot S22 should solve this
problem by describing the relationship between the B2 and the A2
traveling voltage (pseudo-)waves.
The important difference with classic S-parameters is that the
mathematical model should be valid while a large input signal (A1) is
present.
This large signal at the input causes the DUT to start behaving in a
nonlinear way, such that a classical S-parameter, which is based on
the superposition principle, is no longer valid.
For the same reason this kind of behavior can not be characterized
by a standard commercial network analyzer. This measurement
instrument does not only uses the superposition principle for its
experiment design (one-tone applied at the same time at one of the
DUT ports), but also for its advanced and sophisticated calibration
procedures.
5
5
Experimental investigation
Take a real life D.U.T. (CDMA RFIC amplifier)
Apply an A
1
signal
Apply a set of A
2
s
Look at the corresponding B
2
s
Mathematically describe the relationship
between the A
2
s and B
2
s
Repeat for different A
1
s
The approach used in this presentation is to come up with an
accurate Hot S22 measurement method and mathematical model
based upon observations of DUT behavior. In order to do this the
following experiment is performed.
First one takes a real life DUT. In our case this is an RFIC power
amplifier aimed for CDMA applications. Next one applies a large-
signal at the input, this implies the application of a significant A1
spectral component at the input of the RFIC. Next one applies an
extended set of A2 waves. These are generated by a second
synthesizer and are injected towards the RFIC output. Then one
records for all applied A2s the corresponding B2s and one tries to
discover the mathematical relationship between the A2s and the
B2s. Finally one repeats this process for different A1s (typically one
performs a power sweep). The mathematical relationship that is
discovered between the A2s and the B2s is what we call the Hot
S22-model of the DUT.
6
6
Experimental set-up
Large-Signal Network Analyzer
(LSNA, formerly known as NNMS)
RFIC
A
1
B
1
B
2
A
2
synth2
synth1
synth1 generates A
1
synth2 generates a set of A
2
s
LSNA measures all A
1
s, B
1
s, A
2
s, B
2
s
The experimental set-up we used looks as follows. The set-up can be
viewed in some sense as an extended vector network analyzer.
The set-up contains two synthesizers synth1 and synth2.
Synth1 generates the large-signal input, noted A1. This ensures that
all measurements can be performed under hot conditions.
Synth2 generates a set of A2s. For a good experiment design it is
very important that A2 can be applied independent from A1.
The two synthesizers can put the DUT in all experimental conditions
that are needed to investigate Hot S22.
Note that it is possible to use a tuner in stead of a Synth2. In our case
the use of a synthesizer makes the set-up more flexible. The
synthesizer allows, for example, to emulate any given load
impedance through active loadpull.
Next to the signal generation, there is of course the problem of the
accurate measurement of the A1s, the A2s and the B2s. A simple
vector network analyzer is of no use here since it can merely
measure ratios of waves (the S-parameters) rather than the absolute
waves.
For the purpose of absolutely measuring the waves we connect the
DUT to the test-set of a Large-Signal Network Analyzer, the
instrument formerly known as NNMS. This instrument allows to
accurately detect the incident and scattered (pseudo-)voltage waves
A1, B1, A2 and B2 which appear at both of the DUT signal ports.
7
7
Interpretation of the data
0.3 0.2 0.1 0 0.1 0.2 0.3
0.3
0.2
0.1
0
0.1
0.2
0.3
0.75 0.5 0.25 0 0.25 0.5 0.75
0.75
0.5
0.25
0
0.25
0.5
0.75
A
2
(V) B
2
(V)
IQ-plots of the A
2
s and B
2
s for a constant A
1
(x-axis = real part, y-axis = imaginary part)
Next we do the following experiment. We choose one value for A1.
Then we apply a set of different A2s and we look at the relationship
between the B2s and the A2s for the particular value of A1.
In the graph we have plotted the set of applied A2s and the
corresponding set of B2s in so-called IQ-plots. This means that we
plot the real part of each measured complex number on the x-axis
and the corresponding imaginary part on the y-axis. At a first glance it
is hard to interpret the data. In the past people have turned to
Volterra and other theories in order to get the mathematical
relationship between the two measured quantities A2 and B2. An
simple intuitive interpretation is possible, however, by introducing a
so-called phase normalization. The idea is to use the A1 wave as a
phase reference.
This is done as follows. Suppose that one has one measurement of
the quantities A1, B2 and A2. One will then apply a phase shift to all
of these quantities which equals the opposite of the phase of A1. As
a result one gets a new set of A1, A2 and B2 where A1 has a zero
phase. In what follows this kind of phase normalization is always
used for representing A2 and B2.
8
8
Phase normalization is good
Normalize the phases relative to A
1
P = e
j.arg(A1)
0.3 0.2 0.1 0 0.1 0.2 0.3
0.3
0.2
0.1
0
0.1
0.2
0.3
0.6 0.4 0.2 0
0
0.2
0.4
0.6
0.8
A
2
.P
-1
(V) B
2
.P
-1
(V)
When applied to the data previously shown, it is clear that the result
leans itself better for an intuitive understanding of the relationship
between the A2s and the B2s.
Note that the phase normalization is mathematically expressed as a
division by the complex phasor P, which has a phase equal to the
phase of A1 but which has a unity amplitude.
9
9
1.25 1 0.75 0.5 0.25 0
0
0.25
0.5
0.75
1
1.25
1.5
Varying the amplitude of A
1
B
2
.P
-1
(V)
A
1
i
n
c
r
e
a
s
e
s
Let us now do an experiment, using the same set of A2s, but varying
the amplitude of A1.
If we plot the set of corresponding B2s we note several things.
First of all the midpoint of the resulting B2s changes and gets a
larger amplitude with increasing A1. This makes a lot of sense. The
midpoint of the B2s can be interpreted as the output of the amplifier
when it is perfectly matched (A2 equal to zero). Applying the A2s can
be viewed as applying a deviation from the perfect match. We also
note that these midpoints are approximately located on a straight line.
This shows that the component that we are measuring has
insignificant AM-to-PM.
But the most striking observation is that the smiley gets distorted as
A1 increases. The question now is whether a classic Hot S22
predicts this kind of behavior, and, if this is not the case, whether it is
possible to come up with an alternative solution.
10
10
Varying the amplitude of A
2
Linear dependency versus A
2
0.5 0.25 0 0.25
0.6
0.4
0.2
0
0.2
0.4
0.5 0.25 0 0.25
0.5 0.25 0 0.25
0.6
0.4
0.2
0
0.2
0.4
0.5 0.25 0 0.25
0.5 0.25 0 0.25
0.6
0.4
0.2
0
0.2
0.4
0.5 0.25 0 0.25
1.3 1.2 1.1
1.4
1.45
1.5
1.55
1.6
1.65
1.3 1.2 1.1
1.3 1.2 1.1
1.4
1.45
1.5
1.55
1.6
1.65
1.3 1.2 1.1
1.3 1.2 1.1
1.4
1.45
1.5
1.55
1.6
1.65
1.3 1.2 1.1
B
2
.P
-1
(V)
A
2
.P
-1
(V)
Next we will do another experiment. We will now keep A1 constant
and we will vary the amplitude of A2.
From the plots above we can conclude that, in a good approximation
and despite the distortion, the B2 deviation from the midpoint is
proportional to A2. The brings us to the conclusion that the
relationship is almost linear.
11
11
Data interpretation as loadpull
Increasing amplitude of A
1
Just out of curiosity, the experimental data can also be viewed in a
loadpull mode.
The graph shown corresponds to the experiment with increasing
amplitude of A1 (the set of A2s is kept constant). For a loadpull
graph one plots the ratio of A2 and B2 on a Smith chart in order to
see the corresponding reflection coefficients and impedances. Note
that the area covered by the smiley decreases for an increasing level
of A1. This is because the corresponding B2 have a higher amplitude.
12
12
Classic S-parameter description
B
2
= S
21
(|A
1
|).A
1
+ S
22
.A
2
1.25 1 0.75 0.5 0.25 0 0.25
0
0.5
1
1.5
B
2
.P
-1
(V)
Let us know check how good existing approaches can explain the
experimental data corresponding to the increasing amplitude of A1.
As a first approach I consider a classic S22-parameter, combined
with a large-signal S21, which in fact corresponds to a compression
and AM-PM characteristic.
The mathematical relationship between the B2s and the A2s and
A1s in this case is given by a simple S-parameter relationship, where
the S21 has become a function of the amplitude of A1.
The experimental data is depicted in orange, the S22-model is
depicted in yellow.
As we expected the S22 is not capable of describing the fact that the
relationship between the A2s and B2s is clearly a function of the
amplitude of A1.
This S22-model is also not capable of describing the kind of distortion
that we see in the experimental data.
13
13
Simple Hot S
22
description
B
2
= S
21
(|A
1
|).A
1
+ S
22
(|A
1
|).A
2
1.25 1 0.75 0.5 0.25 0 0.25
0
0.5
1
1.5
B
2
.P
-1
(V)
In the past, people improved the accuracy of the previous model by
also making S22 a function of the amplitude of A1.
The resulting mathematical equation is shown above. This Hot S22
can actually be measured by modified vector network analyzer set-
ups. It is clear that this model is linear in A2 (what we want it to be),
and can handle a varying amplitude of A1. Unfortunately this kind of
model can never describe the kind of distortion that we see in the
experimental data. It still looks like something is missing.
14
14
Model linearity & squeezing
We look for a mathematical model which
is linear (superposition valid)
squeezes
Squeezing implies that the phase of A
2
.P
-1
matters
We need different coefficients for the real and
the imaginary part of A
2
.P
-1
More elegant expression results when using
A
2
.P
-1
and its conjugate
So we are looking for an alternative model which:
-Is linear in A2
-Describes the squeezing of the smiley
-Is a general function of the amplitude of A1
The fact that we see a flattening of the smiley implies that the phase
of A2 relative to A1 matters. This implies that we should use a
different coefficient for the real and imaginary part of the phase
normalized A2. This is in fact equivalent to using a different
coefficient for the phase normalized A2 and its conjugate (this results
in a more elegant expression).
15
15
Mathematical expression
B
2
.P
-1
= S
21
(|A
1
|).A
1
.P
-1
+
S
22
(|A
1
|).A
2
.P
-1
+
R
22
(|A
1
|).conjugate(A
2
.P
-1
)
B
2
= S
21
(|A
1
|).A
1
+
S
22
(|A
1
|).A
2
+
R
22
(|A
1
|).P
2
.conjugate(A
2
)
The resulting expression is shown above.
We start by writing a linear relationship between the phase
normalized quantities B2, A2, the conjugate of A2, and A1, where
each of the coefficients is a general function of the amplitude of A1.
Next we multiply both sides by the phasor P, which results in the
lower expression. This expression is actually an extension of the
previous Hot S22, where a linear term is added including the
conjugate of A2. Note that this variable is always multiplied by the
square of P. As such this term is not only a function of the amplitude
of A1 but also of the phase between A2 and A1.
16
16
Extended Hot S
22

B
2
= S
21
(|A
1
|).A
1
+ S
22
(|A
1
|).A
2
+ R
22
(|A
1
|).P
2
.conjugate(A
2
)
1.25 1 0.75 0.5 0.25 0 0.25
0
0.5
1
1.5
B
2
.P
-1
(V)
We will now confront the so-called extended Hot S22 with the
experimental data.
We immediately see that the introduction of the conjugate term is
precisely what is needed in order to describe the squeezing that
takes place.
Note that the resulting model is still linear in A2!
17
17
Quadratic Hot S
22

Further improvement is possible by using a

polynomial in A
2
and conj(A
2
)
E.g.: quadratic Hot S
22

B
2
= F.P + G.A
2
+ H.P
2
.conj(A
2
) +
K.P
-1
.A
2
2
+ L.P
3
.conj(A
2
)
2
+ M.P.A
2
.conj(A
2
)
Note the presence of the P factors
(theory of describing functions)
For those who want even more accuracy, especially for an increasing
amplitude of A2, the idea can easily be extended to describe
nonlinear dependencies on A2.
The result is shown above. Although not explicitly written for clarity,
all coefficients F, G, H, K, L and M are general functions of the
amplitude of A1. Also note the presence of the P factors raised to a
certain power (factors given by the theory of describing functions).
18
18
Comparison (highest A
1
amplitude)
1.4 1.3 1.2
1.4
1.5
1.6
1.7
1.4 1.3 1.2
1.4
1.5
1.6
1.7
1.4 1.3 1.2
1.4
1.5
1.6
1.7
1.4 1.3 1.2
1.4
1.5
1.6
1.7
1.4 1.3 1.2
1.4
1.5
1.6
1.7
1.4 1.3 1.2
1.4
1.5
1.6
1.7
1.4 1.3 1.2
1.4
1.5
1.6
1.7
1.4 1.3 1.2
1.4
1.5
1.6
1.7
B
2
.P
-1
(V)
Classic
S22
Simple
Hot S22
Extended
Hot S22
Quadratic
Hot S22
The figure above shows the performance of the 4 Hot S22
approaches when confronted with the experimental data.
A classic S22 is clearly inaccurate since it is not a function of A1.
The simple Hot S22 includes the dependency on the amplitude of
A1.
The extended Hot S22 also includes the dependency on the phase
of A2 versus A1 (the squeeze), resulting in a significantly better
match between experiment and model.
An even better match results with the quadratic Hot S22, which
describes a 2
nd
order nonlinear dependency on A2.
19
19
Residuals
25 20 15 10 5 0 5
50
45
40
35
30
Relative Error
(dB)
S
22
Simple
Hot S
22

Extended
Hot S
22

Quadratic
Hot S
22

Amplitude of A
1
(dBm)
A more qualitative measure for the performance of the models is
seen when looking at their residuals (relative to the total energy in the
signal) as a function of the amplitude of A1.
It is clear from the graph that the addition of the conjugate term
significantly reduces the value of the residual (up to 15 dB), except
for the lowest value of A1. This makes a lot of sense since for low
amplitudes of A1 the Hot S22 reduces to a perfectly linear S-
parameter model which contains no conjugate term (the same is true
for the quadratic Hot S22).
We also see that the quadratic Hot S22, for our measurements, only
results in a marginal improvement (3 dB) , and this only for the
highest input powers of A1.
Note that the relative residual becomes lower than 50 dB with the
proposed extended Hot S22, which approaches the dynamic range
of our measurement system (about 60 dB for this set of
measurements).
20
20
Conclusion
An accurate Hot S
22
exists
It has a coefficient for the conjugate of A
2
.P
-1
It can accurately be measured
It describes the relationship between A
2
and
B
2
under large-signal excitation
There exists a more accurate Hot S22 concept.
It contains a linear term in the conjugate of the phase normalized A2.
It can be measured with a Large-Signal Network Analyzer and
accurately describes the relationship between B2 and A2 under a
large signal excitation.
21
21
More information
More detailed information on this kind of
measuring and modeling techniques:
http://users.skynet.be/jan.verspecht
http://www.agilent.com/find/lsna