15

CHAPTER 2
POWER AMPLIFIER CHARACTERISTICS &
MODELING
2.1 Introduction

Modeling of RF PA has been a subject of intense research for the last few years. But the
modeling task of PA will not be accomplished until the parameters on which its
performance can be evaluated are not known. This chapter starts with the basic insight into
the basic parameters of the PA like IMD Products, ACLR, Efficiency, CF, EVM and
PCDE, which can be used to evaluate its performance. This chapter also presents the
behavioral modeling approach to model PA, which does not involve a detailed low level
description of the PA. Non-linear Memory-less and Quasi Memory-less Models of PA like
Rapp Model, Saleh model, Ghorbani model, Hyperbolic Tangent Model has also been
discussed. Also non-linear Models of PA with Memory like Wiener Model, Hammerstein
Model, Parallel Hammerstein Model and Memory Polynomial Model most widely used in
literature has been presented in this chapter. All these models are based on Volterra series
which is used to represent the non-linear systems. Memory Polynomial is most suitable to
represent the actual behavior of PA with memory but is complex to implement and the
identification of coefficients of Memory Polynomial is a tedious task.
So, in this chapter a simplified and easy to implement Memory Polynomial model
of PA has been presented. Also, a simple and easy to implement method, to calculate
coefficients of Memory Polynomial using Matlab has been proposed. The effect of
memory depth and order of Memory Polynomial on PA performance has also been
evaluated. Although this simplified Memory Polynomial model can be employed to

16

characterize PAs with reasonable accuracy but there is no systematic way to verify if the
model structure chosen is truly appropriate to the PA under study. So, a novel approach,
called Black Box has been proposed to directly derive the PA from its circuit topology. In
this approach complex S- Parameters of PA have been derived for Motorola LDMOS PA
circuit available in Agilent library. Also the characteristics of the proposed model have
been evaluated and compared with the results of PA circuit topology.

2.2 Non-Linearity in Power Amplifier

The main characteristics of an amplifier are linearity, efficiency, output power, and signal
gain. An amplifier is said to be linear if the output voltage is simply a constant times the
input voltage

Vout (t ) G.Vin (t )

(2.1)

Fig. 2.1 Power Amplifier input-output diagram

As discussed in chapter 1, the RF PAs are typically the most power consuming devices in a
radio base station. They are non-linear because the transistors used in them have nonlinearities that make the output voltage a function of higher order terms of input voltage.

17

The output voltage Vout (t ) for such non-linear amplifier can be expressed mathematically
as a Taylor series (Pothecary, 1999).

Vout (t ) G1Vin (t ) G2Vin 2 (t ) G3Vin 3 (t )..... GnVin n (t )

(2.2)

Thus due to its non-linear characteristics, the PA is the main contributor for distortion
products in a transmitter (TX) chain. In this chapter PA characteristics and its modeling
techniques have been discussed.

With Multi carrier modulation schemes like W-CDMA even harmonic components
could be filtered out, but 3rd and 5th order distortion component will exist with the
fundamental channel. Thus equation 2.2 can be written as (Cripps, 2002)

Vout (t ) G1Vin (t ) G3Vin 3 (t ) G5Vin 5 (t )..... G2 n 1Vin 2 n 1 (t )

(2.3)

If a two tone signal Vin (t ) V1 cos 1t V2 cos 2t is applied to the PA, then 3rd order term
3

will produce G3 V1 cos 1t V2 cos 2t , which is equal to

2
2
G3 V13 cos3 1t V23 cos3 2t 3V12V2cos 21t.cos 2t 3VV
1 2 cos1t .cos 2t

Let us expand each term of equation 2.4 one by one

1 1

Now V13 cos3 1t V13 (cos 2 1t.cos 1t ) V13 cos 21t cos 1t

2 2

1
1
V13 cos 1t V13 cos 21t.cos 1t
2
2

1
1
V13 cos 1t V13 cos 21t 1t cos 21t 1t
2
4

(2.4)

18

1
1
3
1
V13 cos 1t V13 cos 1t cos 31t V13 cos 1t V13 cos 31t
2
4
4
2

(2.5)

3
1
Similarly V23 cos 3 2t V23 cos 2t V2 3 cos 32t
4
2

(2.6)

1 cos 21t
Also 3V12V2 cos 2 1t.cos 2t 3V12V2 cos 2t

3
3
V12V2 .cos 2t V12V2 .cos 2t.cos 21t
2
2

3
3
V12V2 .cos 2t V12V2 cos 21t 2t cos 21t 2t
2
4

(2.7)

Similarly

3
3
3V1V2 2 cos 1t cos 2 2t V1V2 2 .cos 1t V1V2 2 cos 22t 1t cos 22t 1t
2
4

(2.8)

From equations 2.5 to 2.8 we see that 3rd order term of equation 2.4 results in the frequency
components 3 f1 ,3 f 2 , 2 f1 f 2 , 2 f 2 f1 , 2 f1 f 2 , 2 f 2 f1 at the output of PA.

Fig.2.2 Two Tone Test Spectrum Results of a non-linear Power Amplifier

19

As seen from Fig. 2.2, the components 2 f 2 f1 and 2 f1 f 2 are in-band components,
which contributes for the in-band distortion at the output of PA.

2.3 Intermodulation Products

As seen in section 2.2, for two tone ( f1 , f 2 ) input, the 3rd order term has resulted in
2 f1 f 2 , 2 f 2 f1 , 2 f1 f 2 , 2 f 2 f1 harmonic components at the output of PA. So, in
general if we consider higher order terms also, we will get mf1 nf 2 harmonic components
at the output of PA, where m and n are integers. These harmonic components are also
known as IMD products. But out of these IMDs, the most serious is 3rd order IMD which
is very close to the desired frequency as shown in Fig. 2.2 and is not easy to filter out. This
product is usually characterized by the third order intercept point refereeing either to the
input or the output ( IIP3 or OIP3 ). This is best defined by looking at Fig. 2.3. It can be
shown that the slope of the linear gain for input and output powers in dBs is unity, likewise
the slope of the third gain of the third order IMD component is 3 (Pozar, 2000), the point
where the third order line intersects with the linear gain line is the third order intercept
point.

Fig. 2.3 Illustration of the first and third order intercepts points

20

Another figure of merit to characterize non-linearity is the 1dB compression point. For a
non-linear device, the 1dB compression point is defined as the point where the difference
between the non-linear devices output and the linear output is exactly 1 dB. 1dB
compression point can be calculated as (http://www.amplifonix.com/glossary.shtml: 1dB
Compression Point)

3
3

Y1dB G1V1 G3 V13 V1V2 2

2
4

(2.9)

The 1 dB compression point is typically 12 to 15 dB less than the 3rd order intercept point
assuming that they are referenced at the same point.

2.4 Adjacent Channel Leakage Ratio

As discussed in previous sections, PA produces unwanted signal emission to adjacent

channels, which are outlined in the related standards for W-CDMA and other modulation
schemes. The term ACLR is a measure of adjacent channel emission and is very often
faced as regulatory parameter (Digital cellular telecommunications system, 1999). For two
tone input, ACLR is defined as (Kenington, Linearity RF Amplifier Design)

ACLRdB IMR2 6 10 log10

n3

n 2
n
2n3 3n 2 2n mod 2 n mod 2


4

24
8
4

(2.10)

Where IMR2 is Two Tone Inter Modulation Ratio (dB), defined as (Kenington, Linearity
RF Amplifier Design)

21

3
G3V12V2
IMR2 4
G1V1

(2.11)

and n is number of tones for the band of interest.

The idea behind this formula is to estimate ACLR for multi carrier modulation
schemas, by expanding two tones inter modulation IMR to given number of in-band
carrier. Permutation of each tones interaction with all of the others will shape the spectral
growth depending on G3,G5 ....G2 n 1 odd order non-linearity coefficients. Also total power
due to IMD components can be calculated as (Kenington, Linearity RF Amplifier Design)

PIMD

3
2
4 G3V1V2

(2.12)

2.5 Efficiency

One important parameter of PA is its Power Efficiency (). Efficiency is a measure of how
effectively an amplifier converts power drawn from the dc supply to useful signal power
delivered to the output.

Pac
Pdc

(2.13)

Traditionally, PAs are categorized into different classes according to their historical
precedence. Different PA classes can be divided into two major groups: linear and nonlinear PAs. Class A, AB, B and C are some of the well-known linear PAs, which are
distinguished primarily by their bias condition. Linear PAs have the advantage of high
linearity that is important for variable envelope modulation schemes. However, linear

22

amplifiers suffer from poor maximum power efficiency which limits their applications in
low power devices. In practice, an efficiency of only below 20% can be achieved in those
systems. In contrast, non-linear PAs (also known as switched mode PAs) can achieve
better efficiency. As suggested by its name, non-linear PAs have poor linearity
performance. Nevertheless, it is still acceptable for constant envelope modulation schemes
(e.g. FSK). To overcome the problem of linearity for variable envelope systems, many
linearization techniques have been proposed for non-linear amplifiers (Kenington, 2000;
Pothecary, 1999). Therefore, due to their high efficiency and the development of
linearization techniques, non-linear PAs have received more attention over linear
topologies in mobile communication in the last decade. Class E and F are the most
common classes of non-linear PAs. In comparison, Class E PA requires fast switching
driver signal that is not required for Class F PA. Moreover, because of relatively large
switch stresses to active devices, Class E amplifiers do not scale gracefully with the trend
toward lower-power technology with lower breakdown voltage (Pothecary, 1999). For
these reasons, Class-F PA has drawn more attention for its easier implementation and
better integration with sub-micron CMOS technology. A Class F PA uses a output filter to
control the harmonic content of its drain voltage or drain-current waveforms, thereby
shaping them to reduce power dissipation by the transistor and thus to increase efficiency.

2.6 Crest Factor and its effect

CF is a measure of peak to average power ratio (PAPR) of the PA and is defined (Cripps,
2002; Kenington, 2000) as

max x 2

CF 10 log10
E x2

(2.14)

23

where max x 2 represents the maximum (peak) value of the signal x and E x 2 is
expected (average) value of x. In order to increase the linearity the PA can be backed off
from its saturation point. This back-off can be measured in terms of output back-off
(OBO), which is defined as (Kenington, 2000)

Psat

OBO 10 log10
E gnl ( x)2

(2.15)

where gnl is a non-linear function representing the non-linear gain response of the PA and
Psat is the maximum output power of the PA.

Fig. 2.4 Amplifier efficiency performances

As shown in Fig. 2.4, the OBO defines how many decibels less than the maximum
saturated output power the average output power is. If amplifier is assumed to be operating

24

in linear region, then as clear from Fig. 2.5

(2.16)

OBO CF

Fig. 2.5 Output Back-off Ratio

Thus CF has direct influence on OBO i.e. higher is the PAPR of the input signal more is
the back-off required to achieve the same linearity.

2.7 Error Vector Magnitude

The modulation accuracy of the W-CDMA signals is measured by EVM (EVM). EVM is a
measure for the difference between the theoretical waveform and modified version of the
measured waveform. EVM can be defined as the distance between the desired and actual
signal vectors (error vector), normalized to a fraction of the signal amplitude.

25

Fig. 2.6 Error Vector Magnitude

Mathematically, the error vector can be written as

e yx

Where

(2.17)

is the modified measured signal and

is the ideal transmitted signal.

The EVM can be defined as the square root of the ratio of the mean error vector
power to the mean reference signal power expressed as

EVM RMS

E e 2
E x 2

(2.18)

Also for each symbol k , EVM can be defined

EVM (k )

| E (k ) |
1 N
| S (k ) |2

N k 1

(2.19)

26

where E(k) is the error vector for symbol k, S(k) is the ideal signal vector of the symbol k
and N is the number of symbols. Root-mean-square (RMS) value of EVM for a number of
symbols is a widely used measure of system linearity and it can be defined as

N
2

| E (k ) |
EVM RMS

k 1

(2.20)

N
2

| S (k ) |
k 1

EVM is an in-band distortion causing high bit error rates during reception of the
transmitted data. Therefore EVM specifications must also be fulfilled in order to have
proper communication.

2.8 Peak Code Domain Error

The quality of a W-CDMA signal can also be measured by the PCDE. PCDE is computed
by projecting the power of the error vector onto the code domain at a specific spreading
factor (SF). The code domain error (CDE) for every code in the domain is defined as the
ratio of the mean power of the projection onto that code to the mean power of the
composite reference waveform. This ratio is expressed in dB. The PCDE is defined as the
maximum value for the CDE for all codes (Digital cellular telecommunications system,
1999). The projection of the error vector is calculated by despreading the error vector by all
codes. After the despreading operation, there are SF error signals ed , k . The PCDE is
calculated as

max( E e 2 )
d , k

PCDE 10 log10
2

Es

(2.21)

27

The relationship between the EVM and the PCDE can be written as (Vnnen, 2006)

PCDE 10 log10
EVM 2
SF

(2.22)

2.9 Power Amplifier Modeling

Modeling of RF PA has been a subject of intense research for the last few years. Modeling
can be performed from one of at least two main starting points, which includes equivalent
circuit modeling and behavioral modeling. The equivalent circuit modeling requires deep
knowledge and insight in the circuit layout, as well as analysis at component level. The
behavioral modeling approach does not involve a detailed low level description of the
system. A more practical point of view is utilized by trying to describe the relationship
between input and output signals in the PA. This is done by not caring about the actual
cause of the signal appearance, but simply trying to represent the dynamics
mathematically. Non-linear amplifier behavioral models can be divided into three types;
memory-less, quasi memory-less and models with memory. The causes of the memory
effects can be attributed to thermal constants of the active devices or components in the
biasing network that have frequency dependent behaviors. As a result, the current output of
the PA depends not only on the current input, but also on past input values (Taijun et al.,
2004). In other words, the PA becomes a non-linear system with memory. In the following
subsections some popular behavioral models of PA are presented and compared.

2.10 Non-linear Memory-less and Quasi Memory-less Models of Power Amplifier

The AM/AM conversion for a non-linear system is the relation between the amplitude of
the systems output and the amplitude of the systems input. The AM/PM conversion for a

28

non-linear system is the relation between the phase change of the systems input and
output, and the amplitude of the input signal. This is shown in Fig. 2.6.

Fig. 2.7 Amplitude-Phase non-linear model structure for a complex base band input
and output signals

Assuming the pass band input signal (Eun, 1995), the output of the AM/AM and AM/PM
model y (n) can be written as

y(n) g(r(n))cos(0n (n) f (r(n))

Where

(2.23)

g (n) represents AM/AM non-linearity and

f (n) represents AM/PM non-linearity.

2.10.1 The Rapp Model

The Rapp model uses three parameters, and models amplitude distortion but no phase
distortion. The general expression of the AM/AM conversions is as follows (Rapp, 1991)

29

g (r (n))

r ( n)

(2.24)

1/2 s

r ( n) 2 s
1

Osat

where Osat is a parameter that sets the output saturation level and s is a parameter that sets
the smoothness of the transition from linear to saturation states, the smaller s the smoother
the transition. The technique of this model is quite simple. It assumes linear performance
until the saturation point is approached. When the saturation point is approached, a
transition towards a constant saturated output is applied.

2.10.2 The Saleh Model

The Saleh model is a quasi memory-less model. It uses four parameters to fit the model to
measurement data. Its AM/AM and AM/PM conversion functions are described by the
following equations (Saleh, 1981)

a r (n)
g (r (n))
1 a r(n)2
r (n) 2
f ( r ( n))
1 r ( n) 2

(2.25)

(2.26)

Where a , a , and are the models parameters.

2.10.3 The Ghorbani Model

The Ghorbani model uses eight parameters to fit the model to the measurement data, this

30

model is quasi-memory-less, and its AM/AM and AM/PM conversions functions are
described by the following equations (Ghorbani and Sheikhan, 1991)

g ( r (n))

f (r (n))

x1r (n) x2
1 x 3 r ( n)

x2

y1r (n) y2
1 y 3 r ( n)

y2

x4 r (n)

y4 r (n)

(2.27)

(2.28)

x1 , x2 , x3 , x4 , y1, y2, y3, y 4 are the models parameters.

2.10.4 The Hyperbolic Tangent Model

The Hyperbolic Tangent model is quasi memory-less. It has five parameters these
are, IIP3 , linear gain G , upper power limit PU , lower power limit PL and the linear phase
gain GPM . One characteristic of this model is that its parameters are related to physical
attributes like IIP3 . Its AM.AM conversion function is described by the following equation

g ( r ( n )) tanh
r (n) G
IIP 3

(2.29)

Where IIP3 is the third order intercept point and G is linear gain.

The AM/PM conversion is linear and is specified by the slope of the AM/PM conversion
in degree/dB. This linearity is bounded by two parameters which are upper power limit PU
and lower power limit PL . If the power magnitude of the input is less than PL then no phase

31

distortion is added, and if the power magnitude of the input is greater than PU then a
constant phase shift of GPM ( PU PL ) is applied. This can be represented by the following
equation

G P M ( PU PL )

f ( r ( n )) r ( n ) G P M

r ( n ) PU
PL r ( n ) PU

(2.30)

PU r ( n )

2.11 Non-linear Models of Power Amplifier with Memory

In reality the output of the PA depends on previous inputs as well as the current input of
the amplifier. This phenomenon is called memory effect. These memory effects are due to
thermal effects and long time constants in DC bias circuits. It can be observed as
asymmetries in lower and upper sidebands and bandwidth dependent variations in the
magnitude of IMD products. For higher bandwidth applications, e.g. W-CDMA, the
memory effects becomes severe, and cannot be ignored. Hence, memory-less and quasi
memory-less models are not accurate enough. Therefore, a model which considers memory
effects should be used for such applications. In the following sub section sections some of
the most common models with memory are presented.

2.11.1 Volterra Series

Non-linear signal processing algorithms have been growing interest in recent years
(Bershad et al., 1999; Billings, 1980; Mathews and Sicuranza, 2000; Mathews,
1991).Numerous researchers have contributed to the development and understanding of
this field. To describe a polynomial non-linear system with memory, the Volterra series
expansion has been the most popular model in use for the last three decades. The Volterra
theory was first applied with non-linear resistor to a White Gaussian signal. In modern

32

DSP fields, the truncated Volterra series model is widely used for non-linear system
representations. The causal discrete-time Volterra filter is similarly based on the Volterra
series and can be shown (Mathews and Sicuranza, 2000; Mathews, 1991) as:

y(n) h0 h1 (k1 ) x(n k1 ) h2 (k1 , k2 ) x(n k1 ) x(n k2 )

k1 0

k1 0 k2 0

(2.31)

....... .... h2 (k1 ,....k p ) x(n k1 )....x(n k p ) ....

k1 0

kp 0

Where h0 is a constant and h j (k1 ,....k p ),1 j is the set of jth-order Volterra kernel
coefficients. Unlike the case of linear systems, it is difficult to characterize the non-linear
Volterra system by the systems unit impulse response. Also as the order of the polynomial
increases, the number of Volterra parameters increases rapidly, thus making the
computational complexity extremely high. For simplicity, the truncated Volterra series is
most often considered in literature.
The M-sample memory pth-order truncated Volterra Series expansion is expressed
as:

M 1

M 1 M 1

y ( n ) h0 h1 ( k1 ) x ( n k1 )
k1 0

M 1

h (k , k
2

) x ( n k1 ) x ( n k 2 )

k1 0 k 2 0

M 1

(2.32)

....... .... h2 ( k1 ,....k p ) x ( n k1 )....x ( n k p ) ....

k1 0

kp 0

There are several approaches to reduce the complexity of Volterra series. One approach is
the basis product approximation (Metzios, 1994; Newak and VanVeen, 1996; Paniker and
Mathews, 1996; Wiener, 1965), which represents the Volterra filter kernel as a linear
combination of the product of some basis vectors to attempt to reduce the implementation
and estimation complexity to that of the linear problem. Using the Volterra series, two

33

major models have been developed to perform non-linear signal processing. The first
model is the non-orthogonal model and is the most commonly used. It is directly based on
the Volterra series called the Volterra model. The advantage of the Volterra model is that
there is little or no pre-processing needed before the adaptation. But because of the
statistically non-orthogonal nature of the Volterra space spanned by the Volterra series
components, it is necessary to perform the Gram-Schmidt/modified Gram-Schmidt
procedure, which is crucial especially for the non-linear Least Mean Square (LMS) type
algorithms and also for the non-linear Recursive Least Squares (RLS) type adaptive
algorithms. The second model is the orthogonal model. In contrast to the Gram-Schmidt
procedure, the idea here is to use some orthonormal bases or orthogonal polynomials to
represent the Volterra series. The benefit of the orthogonal model is obvious when LMStype adaptive algorithms are applied. Due to difficulty in the measurement of Volterra
kernels and large eigen value spread issue, which implies slow convergence speed and
large misadjustment, especially for the LMS type adaptive algorithm, alternative
mathematical descriptions for non-linear systems with memory are of interest. One such
description is an alternative form of Volterra system called Memory Polynomial model
(Hyunchul and Kenney, 2003) consists of several delay taps and non-linear static functions.
The Memory Polynomial model is equivalent to the Parallel Hammerstein model (Ding,
2004). Obviously, the Parallel Hammerstein model includes the Hammerstein model as a
special case. Hammerstein and Wiener models are the most specialized with the least
number of coefficients, but are by no means the easiest to identify. The Memory
Polynomial model, however, offers a good compromise between generality and ease of
parameter estimation and implementation (Ding, 2004; Hyunchul and Kenney, 2003). In
the following sections Wiener, Hammerstein, Parallel Hammerstein and Memory
Polynomial models are presented.

34

2.11.2 Wiener Model

Wiener model consists of a LTI system followed by a non-linear model without memory as
shown in Fig. 2.7

Fig. 2.8 Wiener model

The output of LTI system is given as

Q

u (n) hq x (n q )

(2.33)

q 0

This acts as input to the non-linear model without memory. The output of non-linear
system can be written as

y(n) b2k 1u(n) u(n)

2( k 1)

(2.34)

q 0

Putting equation 2.33 in equation 2.34, we get

Q

y( n) hq b2 k 1 x(n q )
q0

k 0

2( k 1)

x (n q )

(2.35)

35

2.11.3 Hammerstein Model

Hammerstein model consists of a LTI system preceded by a non-linear model without

memory as shown in Fig. 2.8

Fig. 2.9 Hammerstein model

The output of non-linear system is given as

u (n) bk x(n) x(n)

k 1

(2.36)

k 1

This acts as input to the LTI system and the output of LTI system can be written as

y( n) b2 k 1u (n) u ( n)

2( k 1)

(2.37)

q 0

y (n) u ( n) h(q )

(2.38)

Putting equation 2.36 and equation 2.37 in equation 2.38, we get

y (n)

b h(q ) x (n q ) x (n q )
k

k 1 q 0

k 1

(2.39)

36

2.11.4 The Parallel Hammerstein Model

The Parallel Hammerstein is an extension of the standard Hammerstein model. The model
is illustrated in Fig. 2.10. The system in this case is modeled by (Ding, 2004; Isaksson et
al., 2006).

The output of Parallel Hammerstein model can be given as

2( k 1)

y ( n) H 2 k 1 (q) x( n)

x ( n)

(2.40)

k 1

x( n )
NL1

NL2

LTI0

LTI1
y ( n)

NL3

LTI2

NLQ

LTIQ

Fig 2.10 Parallel Hammerstein model

The main difference between the Parallel Hammerstein and the standard Hammerstein
models is that in the Parallel Hammerstein model, different static non-linear orders are
filtered with different LTI systems. For example, the term of the polynomial is filtered with
H1 (q ) , and the 2nd odd power term is filtered with H 3 (q ) and so on.

37

2.11.5 The Memory Polynomial Model

The Memory Polynomial model (Hyunchul and Kenney, 2003) consists of several delay
taps and non-linear static functions. This model is a truncation of the general Volterra
series, which consists of only the diagonal terms in the Volterra kernels. Thus, the number
of parameters is significantly reduced compared to general Volterra series. The Memory
Polynomial model can be described as

y (n)

2 k 1, q

x(n q )

2 ( k 1)

x(n q )

(2.41)

q 0 k 1

where

x(n) is the input complex base band signal

y (n) is the output complex base band signal
ck , q are complex valued parameters

Q is the memory depth

K is the order of the polynomial

As discussed in section 2.2 the even order terms are usually outside of the operational
bandwidth of the signal and can be easily filtered out. This model considers polynomials
with orders up to 2k 1 , where K is a design parameter. For simplicity of implementation,
the equation 2.41can be rewritten as follows

y (n )

(n q )

q0

F0 ( n ) F1 ( n 1) F 2 ( n 2) .... FQ ( n Q )

(2.42)

38

2 ( k 1)

Where Fq ( n )

2 k 1, q

x(n)

x (n )

k 1

In its expanded form above equation can be written as

2

Fq ( n ) c1, q x ( n ) c3, q x ( n ) x ( n ) c5, q x ( n ) x ( n )

.... c2 K 1, q x ( n )

2( K 1)

x(n)

(2.43)

equation 2.41 can be implemented as

Fig 2.11 Implementation of Memory Polynomial given by equation 2.41

2.12 The Proposed Power Amplifier Models

The term memory effects refer to the bandwidth-dependant non-linear effects often present
in PAs. These encompass envelope memory effects and frequency memory effects.
Envelope memory effects are primarily a result of thermal hysteresis and electrical
properties inherent to PAs. Frequency memory effects are due to the variations in the

39

frequency spacing of the transmitted signal (Kenney et al., 2001). To model a PA system
for DPD design, it is important to model the behavior of PA non-linearity and memory
effect accurately. In initial phase of the research PA was modeled with memory effects
based on Wiener model. Then a more realistic modeling has been proposed based on
simplified Memory Polynomial and at last Motorola PA is modeled using S-parameters.
All these modeling techniques are presented in the following subsections.

2.12.1 Power Amplifier Model based on Wiener Modeling

As discussed in sub section 2.11.2 PA with memory can be modeled by cascading LTI
system followed by non-linear system without memory. In this subsection a non-linear PA
with memory is implemented in Simulink as shown in Fig. 2.11

Fig. 2.12 Modeled memory non-linearity realized using Simulink

The output of memory subsystem y ' (n) is taken as (Ljung, 1999)

y ' (n) x(n) 0.7 x(n 1) 0.5 x(n 2) 0.3 x( n 3)

(2.44)

Equation 2.44 is implemented using LTI FIR-filter block. The non-linear system without
memory is taken as Saleh model given by equations 2.24 and 2.25. The value of parameters
is taken as (Ljung, 1999)

40

a 2.1587 ; a 1.1517 ; 4.0033 ; 9.104

A 64-QAM modulated test signal was used; the resulting power spectrum of input signal
and output spectrum after modeling are shown in Fig. 2.13.

Fig. 2.13 Spectrum of input and output signal

Fig. 2.13 shows that due to non-linear behavior of the PA, the spectral power re-grows and
results in out-of-band distortion. This out-of-band distortion results in increase in EVM,
which is a measure of non-linearity of the PA. Signal constellation at the input and output
is shown in Fig. 2.14 & Fig.2.15 respectively.

41

Fig. 2.14 Constellation of the input signal

Fig. 2.15 Constellation of the output signal

42

I Component of Input Signal

0.15

0.1

I Component

0.05

-0.05

-0.1

-0.15

-0.2

100

200

300

400
Time Samples

500

600

700

800

Fig. 2.16 I-component of the input signal

Q Component of Input Signal

0.15

Q Component

0.1

0.05

-0.05

-0.1

100

200

300

400
Time Samples

500

600

Fig. 2.17 Q-component of the input signal

700

800

43

I Component of Output Signal

0.6

0.4

I Component

0.2

-0.2

-0.4

-0.6

-0.8

100

200

300

400
Time Samples

500

600

700

800

Fig. 2.18 I-component of the output signal

Q Component of Output Signal

0.5
0.4
0.3

Q Component

0.2
0.1
0
-0.1
-0.2
-0.3
-0.4
-0.5

100

200

300

400
Time Samples

500

600

Fig. 2.19 Q-component of the output signal

700

800

44

It is concluded that due to non-linearity as well as memory effects the signal space points at
the output deviates from their positions. Fig. 2.16 & 2.17 shows I component of input and
output signal respectively, whereas Fig. 2.18 & 2.19 show Q component of input and
output signal respectively. From these plots effect of non-linearity of PA is clearly visible.

2.12.2 Memory Polynomial Modeling of Power Amplifier

In this section implementation of Memory Polynomial discussed in section 2.11.5 is used

to implement the PA. For PA identification the least square error (LSE) technique is also
presented.

2.12.2.1 Power Amplifier Identification

There are many different types of models for system identification of non-linear systems.
Some of the rich literature includes sources like (Billings, 1980; Diaz and Desrochers,
1988; Ibnkahla, 2003; Ljung, 1999; Nelles, 2000; Ogunnaike et al., 1994; Sjoberg, 1995).
A non-linear function of the parameter vector can be searched by using non-linear local
optimization techniques. Due its ability to find unique global optimum point and
computational easiness, the LSE technique (Hyunchul and Kenney, 2003) is used to find
the coefficients.

To identify the coefficients Y and H matrices are first defined as follows

Y [ y ( n)

y (n 1) ... y (n N 1)]T

and H [ H 0 ...H q ...H Q ]

where y (n) represents the output data samples.

(2.45)

(2.46)

45

N represents the number of output samples.

and

h3, q (n)
h1, q (n)
h (n 1)
h3, q (n 1)
1, q
Hq

h1,q (n N 1) h3,q (n N 1)

h2 k 1, q (n)

h2 k 1,q (n 1)

h2 k 1, q (n N 1)

(2.47)

Hence

h2 k 1, q (n) x (n q )

2( k 1)

x( n q)

(2.48)

If

c = c0 cq cQ

(2.49)

where

cq = c1,q

c 3,q c2k-1,q

(2.50)

Then the following matrix equation holds

Y = c

(2.51)

If a is the estimated parameter matrix then to have minimum root mean square (RMS)
error between the measured and simulated output a can be calculated from the following
equations.

Since
(HH* )(H* H) = I

(2.52)

46

H* is the conjugate transpose of H , also known as the Hermitian transpose or the adjoint
matrix.
Then equation 2.51 can be rewritten as

(HH* )(H* H)-1 Y = Hc

(2.53)

Then a can be approximated according to the least square criterion by a as follows

a = H* (H* H)-1 Y

(2.54)

Where

a = a 0 a q a Q

and H + = H* (H* H)-1

(2.55)

Where H + is the pseudo inverse matrix of H .

Then a = H + Y

(2.56)

The simulated output is calculated from the input and estimated parameters as

y (n)
Y

y (n 1) y (n N 1)

Ha H qa q

(2.57)

q 0

Then the error between measured and simulated data can be defined as

= e(n) e(n 1) e(n N 1T

E Y-Y

The least square estimate is constructed in such a way that E

(2.58)

is minimized.

47

Above identification algorithm is implemented in Matlab and the coefficients shown in

Table 2.1 are extracted for the PA. To evaluate the performance of the PA simulation set
up shown in Fig. 2.20 is used. A single carrier W-CDMA signal is applied at a chip rate of
3.84 MHz. keeping with memory depth, q 3 and polynomial of the 5th degree, k 3 .

Table 2.1 Estimated parameters for the Memory Polynomial model

Filter Coefficients

Value

c10

1.0513 + j0.0904

c30

-0.0542 + j0.2900

c50

-0.9657 - j0.7028

c11

-0.0680 - j0.0023

c31

0.2234 + j0.2317

c51

-0.2451 - j0.3735

c12

0.0289 - j0.0054

c32

-0.0621 - j0.0932

c52

0.1229 + j0.1508

48

Fig. 2.20 Simulink implementation of performance evaluation setup

Fig. 2.21 and 2.22 shows the input and output of the PA.

Fig. 2.21 Input signal applied to the Power Amplifier

49

Fig. 2.22 Output signal of the Power Amplifier

Calculating the data from Fig. 2.21 and 2.22, the AM/AM characteristics of the PA are
plotted as Fig. 2.23.

Fig. 2.23 AM-AM characteristics of the Power Amplifier model

50

Fig. 2.24 Spectral re-growth of the signal

Fig. 2.24 shows the power spectral density of the input signal. Spectral re-growth caused
by the PA non-linearity can be visualized Fig. 2.24.

Fig. 2.25 Effect of memory depth and order of the polynomial on Power Amplifier
performance

51

To investigate the effect of variation in memory depth and order of the polynomial, the set
up is simulated for three carrier W-CDMA signal, for various values of memory depth and
order of the polynomial. Fig. 2.25 show very little effect of non-linearity on PA
performance as the memory depth and order of the polynomial is increased.

2.13 S-Parameter Modeling of Power Amplifier

The Volterra series discussed in section 2.11.1has been used by several researchers to
describe the relationship between the input and the output of a PA with memory effects.
However, high computational complexity makes methods of this kind impractical in some
real cases, e.g., modeling a PA with strong non-linearities and/or with long-term memory
effects. This is because the number of coefficients to be estimated in the model increases
exponentially with the degree of non-linearity and with the memory length of the system.
To overcome the modeling complexity, various model-order reduction approaches have
been proposed to simplify the Volterra model structure (Metzios, 1994; Newak and
VanVeen, 1996; Paniker and Mathews, 1996; Wiener, 1965). Although these simplified
models have been employed to characterize PAs with reasonable accuracy under certain
conditions, there is no systematic way to verify if the model structure chosen is truly
appropriate to the PA under study. In this section the principle of a novel approach, called
Black Box has been presented. The Black Box model is directly derived from the
topology of the amplifier.

The variables used to describe the signals at both ports are the classical incident and
scattered voltage waves (Verspecht, 2003), typically defined in a characteristic impedance
of 50 , together with the dc current and voltage biasing parameters. The Black Box
model is derived directly from the circuit topology of the PA. The transistors can be
considered to be two port non-linear networks which can be modeled in terms of non-linear

52

Fig. 2.26 Parameter estimation of Black Box model

scattering parameters. If a1, a2 and b1,b2 represents the incident and reflected waves
respectively. Using first order Taylor series expansion, the scattering parameter model of
PA can be written as (Verspecht, 2003)

b1 S11 a1 S12 a1 a1 0 S12 a1 a1*

*

b2 S21 a1 S22 a1 a2 0 S 22 a1 a2

Sij ai

(2.59)

represents non-linear scattering parameters as a function of input waves. But in

real practice scattering parameters are also found to be function of PA bandwidth. So in

order to consider the effect of PA bandwidth also equation 2.59 is modified as (Estagerie et
al., 2007)

b1 S11 a1 .c11 S12 a1 .c12 a1 0 S12 a1 .c12 a1*

*

b
a
S
a
.
c

S
a
.
c

0
S
a
c

1 22 a2
2 21 1 21
22 1 22
22
2

(2.60)
where

cij

represents the normalized frequency function.

53

2.13.1 Valediction of the proposed model

The proposed model is implemented in Agilent Design Software (ADS) around 1950 MHz.
The input signal power is varied from 0 to 20 dB. The model is implemented for Motorola
LDMOS PA circuit available in Agilent library as shown in Fig. 2.27 (Courtesy Agilent
Technologies) and the measurement setup is shown in Fig. 2.28. The data file has been
extracted for the proposed model. For the valediction of the proposed model, the results are
compared with the results of PA circuit topology.

Fig. 2.27 Motorola Power Amplifier circuit topology used to validate the proposed
model

54

PARAMETER SWEEP

V_DC
SRC3
Vdc=BiasU

Var
Eqn

VAR
VAR1
BiasU=5.8
BiasL=2.0

Motorola_PA
X2
Lg ateline1=1425 mil
Lg ateline2=610 mil
Ldrainline1=540 mil
Ldrainline2=1161 mil

V_DC
SRC4
Vdc=BiasL

CIRCUIT BIAS SETUP

ParamSweep
Sweep_BiasU
SweepVar="BiasU"
SimInstanceName[1]="X1.HB1"
SimInstanceName[2]=
SimInstanceName[3]=
SimInstanceName[4]=
SimInstanceName[5]=
SimInstanceName[6]=
Start=5.3
Stop=6.3
Step=0.5

PARAMETER SWEEP

in

out

S2D
Setup
Agi l ent Tec hn ol ogi es

AmplifierS2D_Setup
X1
Filename="Motorola_PA.s2d"
Order=11
Freq _Start=1950 MHz
Freq _Stop=1950 MHz
Freq _Step=0.1 GHz
Pin_Start=0 _dBm
Pin_Stop=20 _dBm
Pin_Step=2 _dB
SSFreq_Start=-1.0 Hz
SSFreq_Stop=-1.0 Hz
SSFreq_Step=-1.0 Hz

ParamSweep
Sweep_BiasL
SweepVar="BiasL"
SimInstanceName[1]="Sweep_BiasU"
SimInstanceName[2]=
SimInstanceName[3]=
SimInstanceName[4]=
SimInstanceName[5]=
SimInstanceName[6]=
Start=1.9
Stop=2.1
Step=0.1

Fig. 2.28 Set-up used for measurement of parameters

Fig. 2.29 Comparison of gain compression (AM/AM characteristics)

55

Gain in dB and phase in degree is plotted against the input power in dBm as shown in Fig.
2.29 and 2.30. Results indicate the validity of the proposed model at the applied frequency.

Fig. 2.30 Comparison of phase compression (AM/PM characteristics)

The proposed model was also tested for measurements on a W-CDMA signal again
centered around 1950 MHz. Input and output spectrum of the input signal and the output
signal have been plotted as shown in Fig. 2.31 and 2.32 respectively.

Upper channel ACLR for the reference signal is calculated as -52.476 dB and for
the distorted signal is -34.525 dB. Also Lower channel ACLR for the reference signal is 52.717 dB and for the distorted signal is -34.608 dB.

56

Fig. 2.31 Spectrum of input signal

Fig. 2.32 Spectrum of output signal

57

Reference Constellation
1.5

1.0

Quadrature

0.5

0.0

-0.5

-1.0

-1.5
-1.5

-1.0

-0.5

0.0

0.5

1.0

1.5

In-Phase

Fig. 2.33 Constellations of the reference signal

Fig. 2.34 Constellations of the distorted signal

Constellations of the reference signal and the distorted signal have also plotted as shown in
Fig. 2.33 and 2.34 respectively. The value of EVM of the reference signal and distorted
signal are calculated as 35.13% and 53.21% respectively.

58

2.14 Conclusions

Modeling of Power Amplifier has been a subject of intense research the last few years. For
understanding the characteristics of PA; different parameters like IMD Products ACLR,
Efficiency, CF, EVM and PCDE can be used. Non-linear amplifier behavioral models can
be divided into three types; memory-less, quasi memory-less and models with memory.
The causes of the memory effects can be attributed to thermal constants of the active
devices or components in the biasing network that have frequency dependent behaviors.
For signals like WCDMA, which have large fluctuations in their envelope PA shows
memory effects also. So, for development of linearization technique for a non linear, the
PA should be modeled with memory effects. Majority of the memory models available in
literature are based on Volterra series, which is used to represent the non-linear systems.
Memory Polynomial, which is reduced form of Volterra series, is most suitable to represent
the actual behavior of PA with memory. But available Memory Polynomial models are not
easy to implement. The Memory Polynomial model proposed in this thesis is easy to
implement. It is also concluded that non-linearity does not show much effect as the
memory depth and the order of Memory Polynomial is increased. The novel approach,
called Black Box proposed in this chapter is used to directly derive the PA model from its
topology. In this approach complex S- Parameters of PA are derived for Motorola LDMOS
PA circuit available in Agilent library. Comparison of the proposed PA model, with the
actual PA circuit topology, show that the proposed model exactly models the behavior of
the actual PA.