Professional Documents
Culture Documents
by
Kevin Gard
Committee in charge:
2003
Copyright
microfilm:
Chair
2003
iii
DEDICATION
To my loving family:
To my parents, Bill and Dorothy, for many years of love and support.
iv
Table of Contents
Signature Page......................................................................................................iii
Dedication ............................................................................................................iv
List of Tables........................................................................................................xi
Acknowledgements..............................................................................................xii
I. Introduction ...................................................................................................... 1
v
I.5.3 Envelope Simulation of Spectral Regrowth ..........................................21
II.8 Summary...................................................................................................80
vi
III.6 Spectral Results .......................................................................................99
V. Conclusions ..................................................................................................130
References.........................................................................................................145
vii
LIST OF FIGURES
Figure I-2: Typical operation of mobile stations within a cellular network. .................. 4
Figure I-5: Amplitude PDF for CDMA and Gaussian modulation signals. ..................12
Figure II-3: Real part of autocorrelation for OQPSK and QPSK IS-95 signals. ..........43
Figure II-4: Real part of autocorrelation for real and complex Gaussian signals..........43
Figure II-15: Carrier gain characteristic of power series limiter models. .....................68
viii
Figure II-17: Spectrum components from autocorrelation analysis. ............................72
Figure II-18: Total output power spectrum at 6 dBm for each limiter model. .............73
Figure II-19: Distortion spectrum at 6dBm output power for each limiter model........73
Figure II-20: Adjacent channel power at 885 kHz offset for limiter models. ...............78
Figure II-21: Adjacent channel power at 1.98 MHz offset for limiter models..............78
Figure III-1: Power spectrum at 2 dBm with complex Gaussian input signal. ...........101
Figure III-2: Distortion spectrum at 2 dBm with complex Gaussian input signal.......101
Figure III-3: Spectrum from Gaussian moment and autocorrelation methods. ..........102
Figure IV-6: Measured AM-PM response of 900 MHz CDMA amplifier. ................118
Figure IV-7: Block diagrams of RFIC transmitter devices. .......................................120
Figure IV-9: Cell band AM-AM AM-PM for superheterodyne RFIC. ......................123
ix
Figure IV-10: PCS band AM-AM AM-PM for superheterodyne RFIC.....................123
Figure IV-11: Cell band AM-AM AM-PM for direct conversion RFIC. ...................124
Figure IV-12: PCS band AM-AM AM-PM for direct conversion RFIC. ..................124
Figure IV-13: Modeled and measured AM-AM/AM-PM for CDMA amplifier. ........126
Figure IV-14: Measured and calculated ACPR for CDMA reverse link signal. .........127
Figure IV-15: Measured and calculated ACPR for complex Gaussian input signal....128
x
LIST OF TABLES
Table I-1: Adjacent channel emissions limits for CDMA and WCDMA mobile
transmitters......................................................................................................... 5
Table I-2: Decibel peak to average ratio for CDMA and Gaussian signals. .................12
Table IV-1: Complex power series coefficients for 900 MHz CDMA amplifier. .......125
xi
ACKNOWLEDGEMENTS
I would like to take a moment to graciously thank all the folks who made it possible
Professor Larry Larson and Professor Michael Steer of North Carolina State
encouragement, and fortitude to pursue and successfully complete this body of work.
I would also like to thank my Ph.D. committee members, Peter Asbeck, Paul Yu,
Bill Hodgkiss, and Chung-Kuan Cheng for their valuable comments and
A special round of thanks goes to my colleagues in the RF, PA, and RFIC design
fields whose conversations and debates about the origins of spectral regrowth
Vladimir Aparin, John Sevic, Steve Kenney, Paul Draxler, Brett Walker, Hector
throughout this long endeavor. And to my son, William, for his ability to make me
smile at anytime and for making me realize that reaching for your childhood dreams is
what life should be about. Last, but not least, I thank my parents for supporting me
always and teaching me that I should strive to be all that I can be.
of the material as it appears in our published papers or as it has been submitted for
xii
of the IEEE International Microwave Symposium, and Proceedings of the IEEE
Custom Integrated Circuits Conference. The dissertation author was the primary author
listed in these publications directed and supervised the research which forms the basis
xiii
VITA
PUBLICATIONS
K. Gard, L.E. Larson, M.B. Steer, “AM-AM and AM-PM Measurement of Baseband to
RF Integrated Circuits for ACPR Calculations,” 2003 IEEE Radio and Wireless
Conference, pp. 273-276.
K. Gard, L.E. Larson, M.B. Steer, “Generalized Autocorrelation Spectral Regrowth From
Bandpass Nonlinear Circuits, 2001 IEEE Int. Microwave Symposium,” vol. 1, pp. 9-12.
K. Gard, L.E. Larson, M.B. Steer, “Autocorrelation Analysis of Distortion Generated From
Bandpass Nonlinear Circuits,” 2001 IEEE Custom Integrated Circuits Conference, pp. 345-
348.
K. Gard, L.E. Larson, M.B. Steer, “Bandpass Techniques for Modeling and Analyzing Spectral
Regrowth,” 2000 Santa Clara Valley MTT-S Workshop, pp. 153-179.
xiv
K. Gard, M.B. Steer, “Efficient Simulation of Spectral Regrowth Using Nonlinear
Transformation of Signal Statistics,” 1999 IEEE Topical Workshop on Power Amplifiers for
Wireless Communications.
K. Gard, H. Gutierrez, M.B. Steer, “A Statistical Relationship for Spectral Regrowth in Digital
Cellular Radio, 1998 IEEE Int. Microwave Symposium,” vol. 2, pp. 989-992.
FIELDS OF STUDY
by
Kevin Gard
Amplitude variations in the modulated waveforms give rise to distortion products when
handsets are limited to prevent system degradation and interference to users in adjacent
cells. However, there is an inherent tradeoff between transmitter efficiency and the
expression for the output power spectrum when the nonlinearity is modeled a complex
versus predicted ACPR values for a CDMA amplifier for CDMA and Gaussian signals.
radio system. The technique yields an analytical expression for the autocorrelation
function of the output signal as a function of the statistics of the quadrature input signal
Cellular phone technology has transitioned from bulky suitcase sized phones to palm
sized data phones with full color displays and internet surfing capabilities. This
revolution occurred, in part, from the implementation of digital cellular standards and
the development of high volume low cost components for developing digital cellular
phones. A key to success was the subsidizing of handset costs by the carriers in order
to raise the number of subscribers. The demand for low cost handsets pressures
handset manufactures to produce the lowest cost handset while still meeting all
performance tends to be stretched near the limits of the system requirement in order to
component is mainly concerned with maximizing the yield of one or more components
used in the system. Likewise, the phone manufacture is concerned with maximizing the
yields of phones. However, each component used must have specifications such that
the end product, a phone, is capable of being manufactured with high yields. This leads
question the need for conservative component specifications in an effort to gain yield
margin for the components. Iterations continue until the cost and yield issues are
strength that the handset transmitter is permitted to radiate outside of the transmission
1
2
nonlinear distortion, mixing products, and spurious signals generated by phase locked
loops (PLL). These products can degrade the signal to noise ratio (SNR) of other
stations operating on the same frequency. Therefore, wireless cellular standards specify
maximum limits on the levels of undesired products allowed to keep the system
operational.
gain amplifier (VGA), upconverter mixer (UPC), driver amplifier (DA), band select
filter, and power amplifier (PA) as shown in Figure I-1. Any one of these system
blocks is a source of nonlinear distortion; however, most emphasis is placed on the PA.
The power amplifier consumes the most amount of power from the battery, in the case
of a handset, or from the power line, in the case of a base station. Power efficiency of
the PA is a significant factor in determining the available "talk time" or battery life of a
handset or the monthly power bill cost of operating a basestation transmitter. High
efficiency operation of power amplifiers occurs when the amplifier is operated at its
maximum output power capacity; however, this requires driving the amplifier into a
highly nonlinear region of operation. Thus, the most difficult part of linear PA design is
achieving the highest efficiency possible without exceeding the linearity requirements.
Likewise, the phone designer should select the lowest cost power amplifier that
RX LO PLL PLL
PA Digital
i (t )
Modem
q (t )
IC
SAW DA VGA UPC
Duplexor
Receiver Chain
Figure I-1: CDMA handset transmitter block diagram.
The demand for low cost, high efficiency, and high linearity power amplifiers and
transmitter designers require specialized tools for predicting linearity and efficiencies
from their designs. Time to market requirements demand quick design cycles.
Consequently, only tools that yield accurate results in a short amount of time are useful
in the design process. Tools that yield higher accuracy, but take a long amount of
simulation time are not useful in the timely development of new products. Thus,
tradeoffs are made when selecting an engineering tool to assist in the assessment of
the resulting power spectrum when a modulated carrier envelope is passed through the
nonlinear model. The end result is an analysis tool that provides quick assessment of
nonlinear performance with a minimal amount of information from the circuit. The tool
is useful for power amplifier, transmitter component, and radio frequency integrated
WCDMA systems are full duplex systems where the transmitter and receiver operate
engaged in a call within a cell site operate simultaneously along with their respective
receivers as is shown in Figure I-2. If one base station operates on the adjacent channel
of the other, then nonlinear distortion generated by a mobile user in one cell will
spillover into the adjacent channel and degrade the SNR of users operating in the
neighboring cell as is shown in Figure I-3. Therefore digital cellular standards restrict
the maximum amount of emissions permitted in the adjacent and alternate (two
channels away from the operating channel) channels. A listing of adjacent and alternate
channel emissions limits for mobile stations operating in CDMA and WCDMA bands is
Basestations
Cell
Mobile
Coverage
Users
DESIRED
SIGNAL
ADJACENT
CHANNEL
SIGNAL
ADJACENT
CHANNEL
INTERFERENCE
Table I-1: Adjacent channel emissions limits for CDMA and WCDMA mobile
transmitters.
Adjacent channel power arises from spectrum regeneration, the process by which a
portion of the band-limited spectrum to leak into adjacent frequency bands due to
the various wireless standards, the main difference being the way in which adjacent
channel power affects the performance of another wireless receiver for which the
defined as:
f4 ~
∫f 3 S gg ( f )df
ACPRUPPER = f2 ~ (I.1)
∫f1 S gg ( f )df
where frequencies f1 and f2 are the frequency limits of the main channel; and f3, and f4
are the limits of the lower adjacent channel. The denominator represents the power in
10
MAIN CHANNEL POWER
-10
SPECTRUM (dB)
-20 ADJACENT
CHANNEL
POWER
-30
f1 f2
-40 f3
f4
-50
-2.45 -1.96 -1.47 -0.98 -0.49 0.00 0.49 0.98 1.47 1.96 2.45
FREQUENCY (MHz)
spectral regrowth observed in the power spectrum of a signal that is passed through a
nonlinear circuit. The distortion generated depends on the amplitude variations of the
modulated signal and the nonlinear input/output characteristic of the nonlinear circuit.
The input and output signals can be analyzed in the time, frequency, or the statistical
domains. For time domain analysis, a time domain input signal is mapped through a
nonlinear input/output response and the power spectrum is observed using a Fourier
transform of the time domain solution. Frequency domain analysis requires the input
sinusoidal sources. The output spectrum is obtained by convolving the input spectrum
with itself as necessary to generate the equivalent output power spectrum for the
signal. The statistical domain approach transforms the statistical properties of the input
8
signal through the nonlinearity to yield, or predict, the output statistical properties.
The power spectrum of the transformed statistical signal is obtained through the
designed to maximize the number of users that can access the system at any given time.
Two high level systems are used; code division multiple access (CDMA) and time
division multiple access (TDMA). CDMA systems permit multiple users to access the
user. IS-95, CDMA2000, WCDMA, UMTS, and TD-SCDMA are digital cellular
standards for CDMA systems. TDMA systems permit users to access the channel one
at a time for a short period of time. GSM, DCS, NADC, and Tetra are digital cellular
standards for TDMA systems. Currently both CDMA and TDMA technologies are
used throughout the world. Carriers in the United States use CDMA, TDMA, and
analog FM systems. Carriers in Europe use GSM and the emerging WCDMA system,
while the rest of the world is a mixture of CDMA, WCDMA, and TDMA systems.
However, it should be noted that almost all third generation (3G) systems, including
Both CDMA and TDMA systems use a variety of digital modulation formats. Data
from each digital system is encoded as 1's and 0's, but the encoded information bits
must be imparted on a radio frequency (RF) carrier before the signal can be transmitted
over a wireless channel. Channel bandwidth and signal-to-noise ratio (SNR) are two
important tradeoffs when selecting a modulation scheme to use for a wireless system.
The available bandwidth for any wireless standard is limited by the system design to
For the purpose of this dissertation, modulation formats are classified as either
Gaussian mean shift keying (GMSK), phase shift keying (PSK), and other techniques
phase shift keying (QPSK), offset QPSK (OQPSK), differential QPSK (DQPSK),
transmission for many modulation schemes with envelope amplitude variation. For
instance, a QPSK signal consists of two digital data streams, equal in amplitude,
envelope; however, the occupied bandwidth is quite large and the first sidelobe of the
sin(x)/x, or sinc, spectrum will only be 13dB down from the carrier in the middle of the
adjacent channel. Typically, a low-pass filter is applied to each digital data stream to
minimize or limit the out of band spectrum of the signal. The filters impart some finite
memory on the data stream which results in amplitude variations as the ringing energy
from a previous data pulse add with the current data pulse.
the average power requirements and the peak amplitude excursions without generating
excessive out of band distortion. However, it is possible for a signal with a higher PAR
10
to exhibit less nonlinear distortion than a signal with lower PAR [3]. The reason for
the inconsistency is because the signal peak is a singular point measurement with,
maximum and minimum variation along with the relative probability of occurrence of
amplitudes within the variation. The APDF is typically estimated from a histogram of
N
f ( A) = (I.2)
∆A * N c
where N is the number of counts per bin, ∆A is the bin amplitude width, and Nc is the
total number of samples. The shape of the amplitude density between the mean and
peak amplitude influences the sensitivity of a particular signal to spectral regrowth due
to nonlinear gain compression or expansion. For example, Figure I-5 shows the APDF
for a CDMA mobile transmitter using OQPSK modulation, the same signal using
QPSK modulation, a real Gaussian signal, and a complex Gaussian QPSK signal where
the average power of each signal is set to 0 dBm. The OQPSK signal has the
quadrature data stream offset in time by half the symbol rate while the inphase and
quadrature data for the QPSK signal are clocked together. The real Gaussian signal is
a carrier modulated samples of a Gaussian process passed though the IS-95 reverse link
baseband transmitter filter while the complex Gaussian signal is the quadrature sum of
two independent samples of a filtered Gaussian process. The PAR for each signal is
shown in Table I-2. The shape of the amplitude density after the mean differs for both
signals where a significant portion of the QPSK amplitude above the mean resides close
to the mean while the OQPSK amplitude density is more linear after the mean. Thus it
11
is difficult to determine, a priori, which signal will be more sensitive to nonlinear gain
compression or expansion even though the QPSK has a higher PAR than OQPSK.
The Gaussian signals are interesting because of the difference between the real and
complex Gaussian signals. The complex Gaussian signal is the quadrature sum of two
real Gaussian processes, so an intuitive guess would suggest that the amplitude
distribution should be wider for the complex signal. Exactly the opposite is true; the
POR of the complex Gaussian signal is 1.7 dB less than the real Gaussian signal. An
explanation for this is that the average power of each real Gaussian input signal is
scaled down by 3 dB to yield the correct complex Gaussian power level; however, it is
an unlikely event that two peaks from each of the real Gaussian signals will occur at the
same time leading to a peak signal that is less than 3 dB plus the peak Gaussian
amplitude. Thus the POR is reduced since the peak distributions do not add in power.
Again, from just the POR, it is difficult to determine which signal will yield the least
amount of distortion for the same output power level.
12
7
Amplitude Probability Density
4
CDMA OQPSK
CDMA QPSK
3
Complex Gaussian
2 Real Gaussian
-1
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2
Envelope Magnitude (V)
Figure I-5: Amplitude PDF for CDMA and Gaussian modulation signals.
Table I-2: Decibel peak to average ratio for CDMA and Gaussian signals.
that distortion terms from other tones and harmonics related to the carrier do not
overlap in the resulting output spectrum. A carrier signal with amplitude and phase
modulation is expressed as
where A(t) and θ(t) are the respective amplitude and phase components of the
modulation. The carrier modulation is often referred to as the complex envelope and is
expressed either in polar form
~
z (t ) = A(t )e jθ (t ) (I.4)
or rectangular form
~
z (t ) = i (t ) + jq(t ) (I.5)
where i(t) and q(t) represent the inphase and quadrature components of the baseband
input signal. The modulated carrier expressed in terms of the complex envelope is
1~ 1 *
w(t ) = z (t )e jωc t + ~z (t )e − jω ct . (I.6)
2 2
G[~
z (t )] = F [ A(t )]e j {θ (t ) + Φ [ A(t ) ]}
~
(I.7)
where F [ A(t ) ] and Φ [ A(t )] are the AM-AM and AM-PM response functions
∞ 2
j 2π ft
S( f ) = ∫ zɶ(t )e dt . (I.8)
−∞
estimation of the output power spectrum. While this method is straightforward in its
application, it provides little insight into how the nonlinear model and signal interact
together to produce the output spectrum. At best, different parameters can be altered,
extract a new model, run the signal through the model, observer the resulting spectrum,
and deduce the sensitivities.
The indirect method is an alternative approach to calculate the power spectrum by
first calculating the autocorrelation function of the signal at the output of the model
then compute the Fourier transform of the autocorrelation function. The
autocorrelation function is the convolution of a signal with its complex conjugate
∞
~
R (τ ) = ∫ ~ z * (t + τ )dt .
z (t ) ~ (I.9)
−∞
Once the autocorrelation function is calculated the power spectrum is estimated from
the Weiner-Khinchine theorem [10]
∞
~ ~
S(f ) = ∫ R (τ )e jωτ dτ . (I.10)
−∞
where ω = 2π f . If ~
z (t ) is a series expansion, then the autocorrelation function is a
summation of products of all combinations of terms in the series, and the resulting
16
intrinsic device characteristics, can impact ACPR in the Japanese TDMA digital
system. Chen, Panton and Gilmore [13] developed a method to predict ACPR and
NPR based on a time domain analysis technique and bandpass nonlinearity theory. AM-
AM and AM-PM transfer characteristics are used to directly predict samples of the
output complex envelope based on samples of an input complex envelope and the
algebraic expression for a bandpass nonlinearity given by the describing function and
corresponding nonlinear phase and amplitude.
10 × f max
N tstep ≈ (I.11)
RSBW
where fmax is the maximum frequency source in the circuit and RSBW is the desired
resolution bandwidth needed to resolve the spectrum of the modulation. For example,
transient simulation of a 2 GHz carrier frequency with 1 kHz resolution requires a
minimum 25 psec time step and 1 msec simulation duration time resulting in a solution
of forty million time points. Transient simulation of a RFIC signal path of moderate
complexity would require a week or more of CPU time to complete the simulation. An
18
Time
t
f (v (t ), t ) = i (v(t )) + i (qɺ (v(t ))) + ∫ y (t − τ )v(τ )dτ + u (t ) = 0 (I.12)
−∞
circuit currents are expressed in terms of conductance, voltage dependent charge, linear
dispersive, and input sources. The KCL equation transformed to the frequency domain
is
F (V ) = I (V ) + ΩQ(V ) + YV + U = 0 (I.13)
IFFT
FFT
~
V1 (ω ) v1 (t )
~
I1 (ω ) i1 (t )
~
V2 (ω ) v2 (t )
~
I 2 (ω ) i2 (t )
~
Linear V3 (ω ) v3 (t ) Nonlinear
~
Circuit I3 (ω ) i3 (t ) Circuit
~
Elements V4 (ω ) v4 (t ) Elements
~
I 4 (ω ) i4 (t )
~
VN (ω ) vN (t )
~
I N (ω ) iN (t )
[19]. Harmonic balance simulation has been demonstrated on large nonlinear circuits
with sinusoidal modulation [20] and on small circuits with Fourier coefficients of
21
digitally modulated waveforms [21, 22]; however, harmonic balance simulation of large
RFIC circuits with digital modulated waveforms applied is not yet practical.
(
dQɶ Vɶ (t ) ) + ΩQɶ Vɶ (t ) + Iɶ Vɶ (t ) + Uɶ (t ) = 0
Fɶ (
Vɶ (t ), t )= dt
( ) ( ) (I.14)
where the extra time varying charge term comes from the envelope Vɶ (t ) and Ω is a
diagonal matrix with j 2π kf on the k th diagonal. Differentiation of the envelope charge
term can be approximated with a finite difference formulation such as backwards Euler
method
( ) (
Qɶ Vɶ (tm ) − Qɶ Vɶ (tm−1 ) ) + ΩQɶ Vɶ (t ) + Iɶ Vɶ (t ) + Uɶ (t
Fɶ (
Vɶ (t ), t )= tm − tm −1
( ) ( )m m m ) = 0 . (I.15)
Harmonic balance frequency domain models can be used for the carrier steady state
response; however, additional time domain descriptions for each component are
required for the transient envelope solution. Recently, time domain steady-state
22
Newton shooting methods [26, 27] have been applied to commercial transient envelope
simulators [23, 28].
The required number of samples of the envelope depends on the input signal length,
desired resolution bandwidth, and the simulator accuracy and time step control
parameters. Thus the number of simulation points is greatly reduced over the transient
and harmonic balance solutions since the solution requires approximately
BW
N≈ (I.16)
RSBW
steady-state solutions of the carrier signal where BW is the bandwidth the signal and
the spectral regrowth. Compared to the earlier transient example, envelope simulation
requires 10,000 single carrier solutions to provide 1 kHz resolution of a signal
spectrum with 10 MHz bandwidth. An example plot showing the sampled envelope
and steady-state carrier solutions versus time are shown in Figure I-8. For the special
case of two tone modulation, the RSBW can be set to the tone spacing and the BW to
ten to twenty times the RSBW to capture the resulting intermodulation distortion
spectrum thus the number of required steady-state solutions of the carrier is reduced to
ten or twenty. However, for wide bandwidth digital modulated signals the required
number of samples increase linearly with the bandwidth for a fixed RSBW.
23
Amplitude
Time
communication systems. Kaye [7] formulated the complex gain as a quadrature sum of
two independent memoryless nonlinearities, modeled as Chebyshev transforms, to
account for AM-PM effects. Previous analysis was either based on analytical models
not derived from measurements [8] or on AM-AM characterization only.
Many early models used Chebyshev or Bessel series expansions as behavioral
models of the channel nonlinearity because the formulations included integrals of
sinusoids with functions of time included in the argument. Koch [29] proposed using a
Taylor series expansion of a nonlinear AM-AM transfer function to calculate distortion
generated when a white noise signal is passed through a memoryless nonlinearity.
Later Kuo [30] proposed improvements and corrections to the Taylor series derivation;
although, the analysis was still limited to AM-AM only nonlinearity. Hieter [31] added
independent time delays to each nonlinear variable in a power series expansion to
account for AM-PM effects. Steer and Kahn [32] generalized the power series
representation by adding complex coefficients. Later the generalized power series
analysis (GPSA) approach was applied to a variety of microwave device modeling and
simulation problems including MESFET amplifiers [5], bivariate power series modeling
of amplifiers [33], and spectral balance simulation techniques [34, 35].
Lajoinie et al. [36] developed a modification of the method described by Chen et
al. [13] that extends it to the prediction of NPR in amplifiers exhibiting nonlinear low
frequency dispersion (memory effects) such as satellite transponders. Sevic and
Staudinger [37] presented a comparison of the behavioral model approach and a
commercial, envelope simulation technique. The envelope simulation technique does
take into account nonlinear circuit memory effects, but has the disadvantages of
requiring accurate circuit component models and prohibitive simulation run time.
25
several assumptions in the model and derivation. First, the authors use an AM-AM only
model asserting that AM-PM effects cannot be represented by a Taylor series
expansion. However, a complex Taylor series expansion can account for both AM-AM
and AM-PM effects of a memoryless nonlinearity [31, 50, 51]. IP3 specifications are
typically available for amplifiers, but IP5 and higher order terms are not. Thus a
designer must make IP5 measurements to obtain better accuracy in the ACPR estimate
using the method described in [49]. Moreover, both AM-AM and AM-PM effects are
easily extracted from single tone complex gain measurements swept over input power
[31]. Second, their derivation represents the CDMA waveform as a single real
Gaussian random variable when the modulation signal is actually a complex sum of two
random variables (the I and Q data channels). Third, the authors assume the power
spectrum is flat over the modulation bandwidth when the actual signal spectrum shape
is determined by a specific baseband filter response defined by the IS-95 CDMA
standard [2]. Gutierrez, Gard, and Steer [52] have addressed most of these limitations
by developing a closed from expression for the autocorrelation of the amplifier output
based on a moment theorem for complex Gaussian processes and an nth order complex
power series model of the nonlinear amplifier. The output power spectrum is
calculated from the Fourier transform of the output autocorrelation expression in terms
of the input signal autocorrelation and thus does not make assumptions about the shape
of the resulting output power spectrum. This allows convenient evaluation of the
output spectrum for a variety of modulation formats by using estimates of the input
autocorrelation function of each modulation format.
I.5.7 Summary
The previous sections reviewed the current state of the art for simulation,
behavioral modeling, and analytical analysis of spectral regrowth generated by
nonlinear wireless circuits. Circuit simulators struggle with efficiently resolving the
28
order problems. A comparison of the relative properties of the different methods for
analyzing spectral regrowth is provided in Table I-3.
analyze the distortion of wireless integrated circuits. Future areas of work are
discussed as a followup to both the behavioral models approaches and the analysis
techniques presented in this thesis.
II. Autocorrelation Analysis of Bandpass Nonlinearities
Analysis of spectral regrowth involves spectral analysis of a signal that has been
passed through a nonlinear channel. The nonlinear channel is represented by a
behavioral model based upon measurement data, simulation data, analytical models, or
look up table models. The simplest analysis involves passing a signal through the
model and observing the spectrum at the output using a fast Fourier transform (FFT);
however, this analysis does not provide insight into how components of the nonlinear
model interact with the signal to generate the observed distortion and gain compression
or expansion characteristics. To gain additional insight requires an analysis technique
that describes the resulting power spectrum in terms of elements of the underlying
nonlinear process, such that relationships between characteristics of the output
spectrum and the model parameters can be determined.
The power spectrum of an output signal can be obtained by either taking the
magnitude squared of the Fourier transformation of the output signal or by first
computing the autocorrelation of the output signal then take the Fourier transformation
of the autocorrelation function. Many engineers simply elect to take the Fourier
transform of the output signal; however, there are benefits to exploring the
autocorrelation approach to obtain the power spectrum. The autocorrelation function
of the output signal involves taking the convolution of the output signal with a time
shifted version of the complex complement of the output signal. The resulting
formulation contains a summation of power terms that are described by the nonlinear
model and individual autocorrelation terms. Finally the Fourier transformation of each
of the terms is taken and summed to describe the output power spectrum. The
resulting power spectrum can be investigated in a variety of ways by singling out
individual terms, groups of relevant terms such as gain compression/expansion or
32
33
distortion terms, or summed together to view the complete output power spectrum.
Such flexibility provides a means to gain additional insight into the relationship between
the nonlinear model and the distortion products generated when a nonlinear circuit or
system component is modulated waveform with amplitude variation.
Irregardless of the spectral analysis method used, a behavioral model of the
nonlinear circuit or system component must be selected to assist the analysis. There
are many models to choose from; however, a power series representation is one of the
simplest models to deal with when working with analytical expressions involving
signals and nonlinearities. The simplicity of the model adds clarity to the signal analysis
techniques presented in this and the following chapter; moreover, the analysis presented
can always be expanded to more complicated models including Volterra series
descriptions.
This chapter presents an overview of bandpass nonlinearity analysis of a modulated
carrier signal passed through a complex power series behavioral model of a nonlinear
wireless circuit. The bandpass nonlinearity analysis yields a description of the transfer
function of the input signal to the first harmonic response of the nonlinear circuit. An
autocorrelation analysis of the output of the nonlinear model is formulated and a
Fourier transform performed to yield the output power spectrum in terms of the power
series model and correlation terms of the input signal. Power sweeps of the input
signal are performed with power series representations of several different analytical
limiting amplifier models and the power spectrum and ACPR are measured at the
output versus the output power. The results are compared to transient envelope
simulation results using the analytical limiter models.
with two analog input signals representing the inphase and quadrature symbol
components of the data. The mixed signals are summed together to form the
quadrature modulated carrier signal.
i (t )
90°
Σ Gɶ
ωc
q (t )
where
zɶ (t ) = i(t ) + jq (t ) ,
A(t ) = i 2 (t ) + q 2 (t ) ,
q (t )
θ (t ) = tan −1 ,
i (t )
zɶ(t ) is the complex envelope, i(t) and q(t) are the inphase the quadrature components
envelope representation of narrow band signals was presented early in the development
The modulated carrier signal is applied to a nonlinear circuit with a nonlinear gain
characteristic, G [w(t )] . The nonlinear gain characteristic is assumed to be a bandpass
~
amplitude changes from the modulated carrier signal. It is important to note that the
AM-AM and AM-PM response represents the transfer characteristic of the input to the
desired output frequency. A complex power series expansion is used to model the
∑
( N −1) / 2 (II.1)
= aɶ2 n+1w2n +1 (t ).
n= 0
The use of complex coefficients in the power series provides the necessary degrees of
freedom to represent both the AM-AM and AM-PM properties of a nonlinear gain
characteristic [32].
In general, the power series contains all powers of the input signal; however, only
the odd order terms generate output components at the fundamental frequency. To see
why this is, consider the case where the input single is an unmodulated carrier signal
w2 (t ) = A2 cos 2 (ωc t )
A2 (II.3)
= [1 + cos(2ωct )]
2
36
thus even order terms generate components at baseband and at even order harmonics of
the carrier frequency. Now consider the first even order term, above the fundamental,
of the power series
w3 (t ) = A3 cos3 (ωct )
A3
= [1 + cos(2ωc t )] cos(ωc t ) (II.4)
2
3 A3 A3
= cos(ωct ) + cos(3ωct )
4 4
thus even order series terms generate components at the fundamental and at odd order
harmonics of the carrier frequency. Therefore only an odd order series, or function,
can describe the nonlinear transfer function of an input carrier signal to the output
components which end up at the carrier frequency.
To simplify the analysis, a binomial expansion is used to compute the mth power of
w(t) yielding [50, 55]
∑ mk [~z (t )] [~z (t )]
m
1 m −k
w m (t ) = e jωc ( 2 k − m ) t .
k *
(II.5)
2m k =0
Consider now only the terms centered at the carrier frequency (this is usually
referred to as the first zonal filter output of the nonlinearity). This implies 2k − m = ±1
for odd m only. Substituting k = (m + 1) / 2 , the terms of (II.5) around the carrier
frequency are
m m +1 m −1
1
wωmc (t ) = m −1 m + 1 [ zɶ (t ) ] 2 zɶ * (t ) 2 e± jωct . (II.6)
2 2
m=2n+1
37
1 2n + 1 n ± jω t
[ zɶ (t ) ] zɶ (t ) e
n +1
wω2cn +1 (t ) = 2 n
* c
. (II.7)
2 n +1
∑
( N −1) / 2
Gɶωc [ w(t ) ] = aɶ2n +1wω2cn+1 (t ) e ± jωct . (II.8)
n =0
The fundamental response can also be defined as a function of the complex envelope
where
N −1
∑ a2 2n + 1
2 ɶ
Gɶωc [ zɶ (t ) ] = n +1 n
2 n +1 *
zɶ (t ) zɶ (t ) . (II.10)
n +1
2n
n= 0
This expression describes the complex envelope of the first harmonic of a modulated
power series.
So far the analysis assumes that the complex power series coefficients are obtained
directly from the underlying nonlinear function. The binomial and 1/ 22 n terms from
(II.10) account for the effect of a sinusoidal carrier signal being passed through a
nonlinear function is known [56]. Describing functions are used in control theory to
simplify the analysis of a nonlinear block with feedback by just considering the
characteristic directly either by simulation or measurement. For this case, the carrier
effects are already included in the measurement data in which case the complex
N −1
∑b
2
Gɶωc [ zɶ (t ) ] =
n
ɶ
2 n +1 zɶ(t )n +1 zɶ* (t ) (II.11)
n =0
where
aɶ2n +1 2n + 1
bɶ2n +1 = .
22 n n + 1
The bɶ2n +1 coefficients can be obtained from a least squares fit of the measured or
the power spectrum because of its simplicity and easy of computation; however,
fundamentally the power spectrum is related to the time domain signal via the
autocorrelation function. Therefore it is worthwhile to understand the relationship
between the autocorrelation function, the time domain signals it represents, and the
power spectrum of the signal. Starting with a band limited time domain signal
zɶ (t ) (II.12)
2
Sɶ ( f ) = Zɶ ( f )
(II.13)
= Zɶ ( f ) Zɶ * ( f )
39
*
where denotes the complex conjugate of the signal and X ( f ) is the Fourier
transform of zɶ (t )
∞
Zɶ ( f )= ∫ zɶ (t )e − j 2π ft dt . (II.14)
−∞
One useful property of Fourier transform pairs is multiplication in the frequency domain
∞ ∞
jωτ f
∫ Zɶ ( f ) Zɶ * ( f )e df = ∫ zɶ (t ) zɶ* (t − τ )dt . (II.15)
−∞ −∞
∞
ℜ
ɶ
zz (τ ) = ∫ zɶ (t ) zɶ * (t − τ )dt (II.16)
−∞
where the subscript zz indicates what signals the autocorrelation is operating on. The
∞ ∞
Sɶ ( f ) = ∫ ∫ z (t ) z * (t − τ )e − j 2π f τ dtdτ
−∞ −∞
∞
(II.17)
= ɶ (τ )e − j 2π f τ dτ .
ℜ
∫ zz
−∞
ɶ (τ ) = ℜ
ℜ ɶ * ( −τ ) (II.19)
zz zz
thus the real part of autocorrelation function is an even function while the imaginary
part is odd. The autocorrelation function is related directly to the power of the signal
when τ = 0
T
1
Pz = lim
2
∫ zɶ (t ) dt
T →∞ 2T (II.20)
−T
=ℜ
ɶ (0) .
zz
ˆ (m) =
ℜ zz k =0
K
∑
1 K − m −1
zɶ (k ) zɶ * (k + m), m > 0
(II.21)
ℜ
zz (− m), m < 0
ˆ*
where K is the length of the sequence zɶ (k ) , m defines the mth lag of the
autocorrelation estimate, and the normalization factor 1/ K denotes that this is a biased
41
estimation of the autocorrelation function [9]. Signal construction for a discrete time
CDMA signal is outlined in Figure II-2.
−n
z2
zɶ (t )
Σ
OQPSK
HALF CHIP A Rɶ zz (τ )
DELAY
AUTO
POWER
j SCALING
CORRELATION
ESTIMATION
The signal is constructed by taking hard limited samples of a zero mean Gaussian
number generator and passing them through the baseband filter response as specified in
the IS-95 CDMA standard [2]. The inphase and quadrature filtered data streams are
added in phase quadrature and scaled, in power, to a convenient reference level of 0
dBm. The real and imaginary parts of an autocorrelation function estimate for both
OQPSK and QPSK modulation of an eight times oversampled IS-95 CDMA reverse
link signal using (II.21) with K = 216 samples and m = 211 lags are shown in Figure II-3
and Figure II-5 respectively. The OQPSK signal is the solid trace overlaid by the
QPSK signal as a dashed trace. Similarly the real and imaginary parts of an
autocorrelation function estimate for a real and complex Gaussian modulated QPSK
signal are shown in Figure II-4 and Figure II-6 respectively. The complex Gaussian
signal is a solid trace overlaid by the real Gaussian signal as a dashed trace. Note that
the real part of the autocorrelation function is an even function about τ = 0 and the
imaginary part is an odd function. Also, the imaginary part is less than two orders of
magnitude smaller than the real part for all signals. Interestingly there are only subtle
differences in the ringing between the real parts for all the signals shown. The
42
0.05
0.04
0.03
Re{Rzz (τ )}
0.01
-0.01
-20 -15 -10 -5 0 5 10 15 20
µsec)
Time Shift (µ
Figure II-3: Real part of autocorrelation for OQPSK and QPSK IS-95 signals.
0.05
0.04
0.03
Re{Rzz (τ )}
0.01
-0.01
-20 -15 -10 -5 0 5 10 15 20
µsec)
Time Shift (µ
Figure II-4: Real part of autocorrelation for real and complex Gaussian signals.
44
4.0E-04
3.0E-04
2.0E-04
1.0E-04
Im{Rzz (τ )}
0.0E+00
-1.0E-04
-2.0E-04
-4.0E-04
-20 -15 -10 -5 0 5 10 15 20
µsec)
Time Shift (µ
4.0E-04
3.0E-04
2.0E-04
1.0E-04
Im{Rzz (τ )}
0.0E+00
-1.0E-04
-2.0E-04
The output autocorrelation function of the nonlinear model is computed from (II.10)
and (II.18) as
T
1
ɶ (τ ) = lim
ℜ gg ∫ Gɶωc [ zɶ (t )] Gɶω*c [ zɶ(t + τ )] dt (II.22)
T →∞ 2T
−T
N −1 N −1
∑ ∑ a2 2n + 1 2m + 1 n +1 *
*
( ) ( zɶ )
2 2 ɶ ɶ
a * m +1 m
Gɶωc [ zɶ (t ) ] Gɶω*c [ zɶ (t + τ ) ] =
n
2 n +1 2 m +1
2( n + m ) zɶ1 zɶ1 z , (II.23)
ɶ2
n +1 m +1
2
n = 0 m =0
zɶ1 = zɶ ( t ) , and
zɶ2 = zɶ ( t + τ ) .
ɶ (τ ) = a ɶ (τ ) +
3
ℜ 13 (τ ) + a1 a3ℜ31 (τ ) +
ɶ1 ℜ aɶ1aɶ3 ℜ
2 *ɶ
ɶ ɶ ɶ
*
4
gg 11
9 5
ɶ (τ ) + a
aɶ3 ℜ 15 (τ ) + a1 a5ℜ51 (τ ) +
ɶ5 ℜ
2 *ɶ
ɶ ɶ ɶ
*
ɶ1a (II.24)
8
33
16
15 ɶ (τ ) + a ɶ (τ ) +
25 2 ɶ
aɶ3 aɶ5*ℜ
35
*
ɶ5 ℜ
ɶ3 a 53
aɶ5 ℜ55 (τ ) + …
32 64
where
The outut power spectrum is obtained from the Fourier transform of the output
~ 2 ~ 3
[ ~ ~
S gg ( f ) = a~1 S11 ( f ) + a~1a~3* S13 ( f ) + a~1* a~3 S 31 ( f ) +
4
]
9 ~ 2~
16
5
8
[ ~ ~
a3 S 33 ( f ) + a~1 a~5* S15 ( f ) + a~1* a~5 S 51 ( f ) + ] (II.25)
32
[
15 ~ ~ * ~ ~
a3 a5 S 35 ( f ) + a~3* a~5 S 53 ( f ) + ]
25 ~ 2 ~
64
a 5 S 55 ( f ) + …
where
∞
− jωτ
Sɶ f)= ℜ z2 n +1z2 m+1 (τ ) e dτ .
ɶ
( 2 n +1)( 2 m +1) ( ∫
−∞
Therefore, the output spectrum is a sum of the Fourier transforms from each
expansion for an Nth odd order power series expansion. For a particular modulation
input signal, the individual autocorrelation and spectrum terms are computed only once
and stored in a file. At run time, the spectral components are read, then scaled by the
power series coefficients and input power level, and summed to yield the output
spectrum.
Summing the weighted output spectral terms is considerably faster than performing
an FFT of the time domain waveform passed through the bandpass nonlinear model.
Moreover, the output spectrum can be separated according to the order of distortion or
by relevance to the input signal. For instance, the spectral terms correlated to the input
signal, or the linear term, represent the gain expansion or compression of the desired
signal at the output while all other terms represent the uncorrelated nonlinear distortion
about the carrier. The gain compression or expansion terms from (II.25) are
47
2 3
SɶggGain ( f ) = aɶ1 Sɶ11 ( f ) + aɶ1aɶ3* Sɶ13 ( f ) + aɶ1* aɶ3 Sɶ31 ( f ) +
4
(II.26)
5 35
aɶ1aɶ5 S15 ( f ) + aɶ1 aɶ5 S51 ( f ) +
*ɶ * ɶ aɶ aɶ S ( f ) + aɶ1 aɶ7 S71 ( f ) + …
* ɶ * ɶ
8 64 1 7 17
while the nonlinear distortion terms from (II.25) representing the spectral regrowth are
9 15
SɶggDistortion ( f ) = aɶ3 Sɶ33 ( f ) + aɶ3aɶ5* Sɶ35 ( f ) + aɶ3*aɶ5 Sɶ53 ( f ) +
2
16 32
105 25 2 ɶ
aɶ3 aɶ7* Sɶ37 ( f ) + aɶ3* aɶ7 Sɶ73 ( f ) + aɶ5 S55 ( f ) + (II.27)
256 64
175 315
aɶ5 aɶ7* Sɶ57 ( f ) + aɶ5* aɶ7 Sɶ75 ( f ) + aɶ5 aɶ9* Sɶ59 ( f ) + aɶ5* aɶ9 Sɶ95 ( f ) + … .
512 1024
-40
-50
-60
-70
-80
-3.0 -2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0
FREQUENCY (MHz)
POGain = ℜ ggGain (τ = 0)
ɶ
3
= aɶ1 ℜ11 (0) + aɶ aɶ *ℜ (0) + aɶ1* aɶ3ℜ31 (0) +
2
1 3 13
4
5 35
aɶ1aɶ5*ℜ15 (0) + aɶ1* aɶ5ℜ51 (0) + aɶ1aɶ7*ℜ17 (0) + aɶ1* aɶ7 ℜ 71 (0) + … (II.28)
8 64
N −1
∑ 2m + 1
∫
2 ɶ ɶ* T
1 aa 2( m +1)
= 1 2 m +1
zɶ(t ) dt
m + 1 −T
2m
2T m= 0 2
where ℜ(2 n +1)(2 m +1) (0) is denoted as a real value because the imaginary part of
ℜ(2 n +1)(2 m +1) (0) is zero. Similarly the total distortion power about the carrier is
calculated from the autocorrelation function terms associated with the distortion
PODistortion = ℜ ggDistortion (τ = 0)
9 2 15
= aɶ3 ℜ33 (0) + aɶ3aɶ5*ℜ35 (0) + aɶ3*aɶ5ℜ53 (0) +
16 32
105 25 2
aɶ3 aɶ7*ℜ37 (0) + aɶ3*aɶ7 ℜ73 (0) + aɶ5 ℜ55 (0) + (II.29)
256 64
175 315
aɶ5 aɶ7 ℜ57 (0) + aɶ5 aɶ7 ℜ75 (0) + aɶ5 aɶ9 ℜ59 (0) + aɶ5 aɶ9ℜ95 (0) + …
* * * *
512 1024
N −1 N −1
∑∑ a2 n+1aɶ2*m+1 2n + 1 2m + 1
∫
T
2( n + m +1)
2 2 ɶ
1
= 2( n + m ) zɶ (t ) dt .
2T n =1 m =1 2 n + 1 m + 1 −T
Only the uncorrelated distortion will degrade system SNR and from the spectral
plot it is not clear what portions of the distortion are correlated and uncorrelated with
the input signal. To determine the correlation, the crosscorrelation function between
the input signal and output signals needs to be calculated. The time average
crosscorrelation function between the input and output is defined as
T
1
ℜ
ɶ
zg (τ ) = lim ∫ zɶ (t )Gɶ * [ zɶ (t + τ ) ] dt . (II.30)
T →∞ 2T
−T
50
N −1
∑ 2n + 1
∫
2 ɶ* T
1 a n +1
ɶ (τ ) = lim
zɶ (t ) zɶ (t + τ ) [
zɶ(t + τ ) ]
n
ℜ 2 n +1
*
dt . (II.31)
n + 1 −T
zg 2n
T →∞ 2T 2
n =0
Computation of (II.31) represents the portion of the output that is correlated with
the input signal. Rho, ρ , is a CDMA waveform quality metric related to the
croscorrelation coefficient between the input and output divided by the product of the
2
ℜ
ɶ (0)
ρ=
zg
ℜ
ɶ (0)ℜ
zz
ɶ
gg (0)
2
N −1
∑ a2 2n + 1
∫
2 ɶ* T
2 n +1 2( n +1)
zɶ (t ) dt (II.32)
n + 1 −T
2n
n =0
= N −1 N −1
.
∫ z(t) dt ∑ ∑ a 2 2n + 1 2m + 1
∫
T 2 2 ɶ * T
2 ɶ
a
2 n +1 2 m +1 2( n + m +1)
ɶ zɶ (t ) dt
2( n + m )
−T n =0 m =0 n + 1 m + 1 −T
through a nonlinearity leading to additional distortion components that are related to all
permutations of products of the different input signals. A practical example of this is
when a narrowband interferer signal is present at the input of a CDMA receiver that is
operating near the outskirts of cell coverage where the receiver is at maximum
sensitivity and the transmitter is near maximum output power [42]. The interferer and
transmitter signals are both present at the input to the low noise amplifier (LNA) at the
front end of the receiver. The two signals will generate crossmodulation distortion as
the signals pass through the nonlinear characteristics of the LNA. The receiver
51
receive channel that the crossmodulation distortion overlaps part or all of the receive
channel.
Consider the case when two quadrature modulated carriers are applied to a
bandpass nonlinearity
∑ a~ w (t)
N
G[w(t )] =
~ n
n
n=0
where
wn (t ) = { A(t ) cos [ω1t + θ A (t ) ] + B(t ) cos [ω2t + θ B (t )]} .
n
1 jω t − jω t jω t − jω t n
= uɶ (t )e 1 + uɶ (t )e 1 + vɶ (t )e 2 + vɶ (t )e 2
n
* *
2
∑
1 n n k
= n uɶ (t )e jω1t + uɶ * (t )e − jω1t vɶ(t )e jω2t + vɶ* (t )e − jω2t
2 k =0 k
( n −k )
= ∑ ∑ ∑
1 n k
n
k (k − x ) n −k
n−k ( n− k − y )
[u(t )] u (t ) j ( 2 x − k )ω1t
[ ] j ( 2 y − n + k )ω 2t
x * y *
ɶ ɶ e vɶ (t ) vɶ (t ) e
n
k x
2 k =0 x =0 y =0 y
= ∑∑∑
1 n k n−k
n k n− k
(k − x) ( n −k − y )
[ u (t ) ] [vɶ(t )] j ( 2 x − k )ω1 + ( 2 y − n + k )ω2 t
x * y *
ɶ uɶ (t ) vɶ (t ) e .
n
2 k x
k = 0 x =0y y =0
The crossmodulation terms that end up around the first carrier frequency, ω1 , occur
k n − k
∑
k ±1 k ∓1 n− k
1 n
n
wω1 (t ) = n
n
k ±1 n − k
[
uɶ (t )
]2 uɶ
*
(t ) 2 vɶ(t )vɶ
*
(t ) 2 e ± jω1t . (II.33)
2 k
k =0
2 2
Similarly, the crossmodulation terms that end up around the second carrier frequency,
ω 2 , occur when 2 x − k = 0 and 2 y − n + k = ±1 which implies x = k / 2 and
y = (n − k ± 1) / 2
52
k n − k
∑
k n − k ±1 n − k ∓1
1 n
n
wω2 (t ) = n
n
k n − k ±1 [
uɶ (t )uɶ* (t ) 2 vɶ (t )
] 2 ɶ
v
*
(t ) 2 e ± jω2t . (II.34)
2 k =0 k
2 2
Other frequencies of interest are the crossmodulation terms that end up at either
2ω1 − ω2 or 2ω2 − ω1 because these terms could potentially end up in the desired
receive band or in some other nearby protected band. One hypothetical case is a full
duplex system where two transmitter signals are sharing a power amplifier. The out of
band crossmodulation terms could end up somewhere within the receive band if the
transmitter channel separation is wider than the separation between the receive and
transmitter bands. The US cellular transmit band is 25 MHz wide, 824 MHz to 849
MHz, with a 45 MHz duplex frequency, so the lowest receive channel is 869 MHz. If a
transmitter signal exists at 849 MHz and the other at 824 MHz then the 2ω2 − ω1 term
ends up at 874 MHz which is inside the cellular receive band. Terms that end up at
2ω1 − ω2 occur when 2 x − k = ±2 and 2 y − n + k = ∓1 which implies x = (k ± 2) / 2 and
y = (n − k ∓ 1) / 2
w2nω1 −ω2 (t ) =
k n − k
∑
k ±2 k ∓2 n − k ∓1 n − k ±1 (II.35)
1 n
n
k
± 2 n − k ∓1[
] [ ]
uɶ (t ) 2 uɶ* (t ) 2 vɶ (t ) 2 vɶ * (t ) 2 e ± j ( 2ω1 −ω2 )t .
2n k =0 k
2 2
w2nω2 −ω1 (t ) =
k n − k
∑
k ∓1 k ±1 n− k ± 2 n −k ∓ 2 (II.36)
1 n
n
k ∓ 1
n − k ± 2
[ ] [
uɶ (t ) 2 uɶ * (t ) 2 vɶ (t )
] 2 vɶ
*
(t ) 2 e
± j ( 2ω2 −ω1 )t
.
2n k =0 k
2 2
53
2
1 1
b2 w (t ) = b2
2
2
zɶ (t )e jωct + zɶ* (t )e− jωc t
2 (II.37)
b
{ 2
}
= 2 zɶ 2 (t )e j 2ωct + zɶ* (t ) e − j 2ωc t + zɶ (t ) zɶ* (t ) + zɶ* (t ) zɶ(t )
4
where b2 is the second order power series coefficient of the input nonlinearity. The
harmonic of the input signal. The input distortion components add with the input signal
u (t ) = b2 w2 (t ) + w(t )
=
b2 2
4 { 2
}
zɶ (t )e j 2ωc t + zɶ* (t ) e − j 2ωc t + zɶ(t ) zɶ* (t ) + zɶ* (t ) zɶ (t ) + (II.38)
1
2
{
zɶ (t )e jωc t + zɶ* (t )e− jωc t . }
The new input signal is then applied to the transconductance nonlinearity which also
contains a second order distortion. Consider just the second-order nonlinearity
component of the transconductance
54
2
aɶ2u 2 (t ) = aɶ2 b2 w2 (t ) + w(t )
(II.39)
= aɶ2 b22 w4 (t ) + 2b2 w3 (t ) + w2 (t ) .
A new third-order distortion component at the output is generated from the product of
the input signal and the second-order distortion term at the input, i.e.
2aɶ2b2 w3 (t ) . (II.40)
Additional even order components are also generated; however, they do not generate
distortion components about the carrier frequency.
g
vo = L tanh vin (II.41)
L
where g is the linear gain and L is the limit value of the output signal. One drawback of
the hyperbolic tangent function is that the sharpness of the transition from linear to
limiting is fixed in relation to the gain and can not be adjusted without introducing
additional parameters. The drain current expressions for several gallium arsenide
(GaAs) metal semiconductor field effect transistor (MESFET) SPICE models are based
upon a parameterized hyperbolic tangent function [64, 65].
55
gvin
vo = 1
(II.42)
s
g s
1 + v
L in
where g is the linear gain, L is the limit value of the output signal, and s controls the
sharpness of the transition from linear to limiting. Another widely cited limiter model
nearly identical to the Cann model is reported by Rapp [67]. A plot of the large signal
hyperbolic tangent and Cann limiter models are shown in Figure II-8 and Figure II-9.
The carrier transfer characteristic is the first coefficient of a Fourier series expansion
which limits to 4 L / π as can be seen in the plots from Figure II-9. Also, the
compression characteristic for the fundamental is more gradual and has a softer knee
than the large signal transfer characteristic. The Cann limiter model, while convenient
and flexible, exhibits derivative behavior for different values of s which leads to
nonphysical behavior of the intermodulation products [68]. Care must be used when
using the Cann model to represent a physical limiting amplifier to ensure the
intermodulation products behave as expected for the amplifier being modeled. Despite
the odd derivative and intermodulation product properties described in [68], the Cann
model is widely used in system simulations to model the effects of a limiting amplifier
on the spectrum and bit error rate of wireless communication signals [69-75].
56
1.4
1.2
1.0
0.8
0.6
0.4
Output (V)
0.2
Cann s=2
0.0
Tanh
-0.2 Cann s=4
Multi-Tanh
-0.4 Cann s=10
Cann s=100
-0.6
-0.8
-1.0
-1.2
-1.4
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Input (V)
1.4
1.2
1.0
0.8
0.6
0.4
Output (V)
0.2
Cann s=2
0.0
Tanh
-0.2 Cann s=4
Multi-Tanh
-0.4 Cann s=10
Cann s=100
-0.6
-0.8
-1.0
-1.2
-1.4
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Input (V)
transistor has an exponential collector current dependency on the input voltage across
VBE ( t )
I c (t ) = I s e nF vT
(II.43)
where
VBE is the DC value of the base emitter voltage, Vbe (t ) is the time varying component
of the input signal, vT = kT / q is the thermal voltage, I s is the collector saturation
current, and nF is the nonideality factor [76]. There is an additional collector voltage
nonlinearity due to the Early voltage effect that is not considered here. In addition, the
base current is also exponentially dependent on the base emitter voltage
VBE ( t )
Is
I B (t ) = e nF vT
(II.44)
βF
where β F is the forward collector to base current gain. The base current nonlinearity
∑
∞
1 VBE (t )
I c (t ) = I s . (II.45)
k =0 k ! vT
Separating out the DC bias component from the input signal leads to a simple
behavioral model for a bipolar or heterojunction common emitter amplifier
58
k
∑
VBE ∞
1 Vbe (t )
I c (t ) = I s e nF vT
. (II.46)
k = 0 k ! vT
However, common emitter amplifiers exhibit gain compression at high output signal
whereas (II.46) is an exponential function that grows without bound. Gain
compression for a common emitter amplifier arises from either a limitation in the base
current drive supplied by the previous stage, voltage swing at the collector, or by local
feedback.
The common emitter differential pair amplifier shown in is a circuit widely used in
RFIC and analog IC designs. The large signal transfer function of the bipolar polar
differential pair is a hyperbolic tangent function of the input signal
−Vbe (t )
vo (t ) = RI E tanh (II.47)
2vT
with limiting value of RI E and a gain of ( RI E ) /(2vT ) . The linearity of the circuit can
be improved at the expense of lower gain by adding either shunt feedback resistance
from the output to input or by adding series feedback resistor in the emitters; however,
the noise performance suffers significantly since the gain is reduced but the noise
contribution of devices is not.
59
Vcc
R R
vo
Q1 Q2
vin
IE
∑I −Vbe (t ) + V j
N
vo (t ) = R Ej tanh . (II.48)
j =1 2vT
where V j is an offset voltage. One practical form of the multi-tanh circuit is the so
called triplet, named for the number of differential pairs, is shown in Figure II-11
consisting of one differential pair in parallel with two differential pairs with scaled
emitter areas to generate ± offset voltages in the hyperbolic tangent transfer functions
60
where the offset voltage Vos = vT ln( N ) and N is the area ratio for the offset transistors
[81]. The total sum of tail current for the doublet, I E ( K + 1 + 1) , should be scaled to
match the tail current of a single differential pair for a fair comparison of the linearity
improvement. The gain of the triplet is less than a standard differential pair by a factor
of
1 8N
+ K = 0.466 (II.50)
2+ K (1 + N )
2
for the same bias current with K=0.75 and N=13; however, the noise is also decreased
almost by the same amount resulting in similar noise figure performance to the
differential pair. The improvement in dynamic range for the multi-tanh circuit is not
apparent until the compression and linearity characteristics of the differential pair and
the multi-tanh triplet are compared later in this chapter.
61
Vcc
R R
vo
Q5 Q3 Q1 Q2 Q4 Q6
vin vin
N N
IE KI E IE
The carrier gain compression characteristic of the hyperbolic tangent, Cann limiter,
and the multi-tanh models are shown in Figure II-12. The input power independent
variable is defined as the power when the shunt input impedance is 50 Ω . The linear
gain for each model was set to 1/(2vT ) which is the same as the differential pair
amplifier with a RI E product of unity. The gain of the multi-tanh triplet model from
(II.49) is 6.6 dB lower than a differential pair for the same bias current; however, the
limiting output value is the same. The input referred one dB gain compression point
for each of the models is shown in Table II-1. The input compression point for the
Cann model is higher for models with sharper transition factors because the signal does
not compress until it approaches the knee value whereas the softer transition models
start to compress well before the knee value. The multi-tanh triplet response benefits
from both an increase in the sharpness and reduction in gain in comparison to the
62
26
24
22
Gain (dB)
20
18
Cann s=2
16 Tanh
Cann s=4
Cann s=10
14
Cann s=100
Multi-Tanh
12
10
-35 -30 -25 -20 -15 -10 -5 0 5 10
Input Power (dBm)
Multi-tanh Triplet -8
Tanh -19
S=2 -20
S=4 -16
S=10 -14
S=100 -14
63
{ } { }
(II.51)
= Re F [ A(t )] e− jΦ[ A(t )] vɶin + j Im F [ A(t )] e − jΦ[ A(t )] vɶin
where vɶin and vɶo are the complex envelopes of input and output signals respectively.
The rectangular form is also known as a quadrature model of the bandpass nonlinearity
[7, 51]. A graphical representation of a quadrature model is shown in Figure II-13.
{
Re F [ A(t ) ]e − jΦ [A( t ) ] }
v~in ∑ v~o
{
Im F [ A(t )]e Φ [ A(t ) ] } j
The real and imaginary components are separated and each fitted to a real power series
expansion using a least mean squared error fit to the data
64
An odd power series is used because only odd terms contribute to the fundamental
characteristic; however, there are other methods by which a power series can be
analytic models or for measurement data with smooth numerical derivatives. The
nth order Taylor series expansion of a function, f ( x ) , about a point a, is defined as
f ′′(a )( x − a) 2
⋯+ f (a )( x − a) n
(n)
f ( x ) = f (a) + f ′(a )( x − a) + + . (II.53)
2! n!
For example, the Taylor series expansion for the hyperbolic tangent model is
(2 )
2 k −1
− 1 B2 k g
∑
2k 2k
g
∞ 2
π
L tanh vin (t ) = L vin (t ) ; vin (t ) < (II.54)
L k =1 ( 2k ) ! L 2
where B2k are Bernoulli numbers [82]. Notice that the hyperbolic tangent expansion
generates only odd order power series coefficients. The Cann model from (II.42) was
shown to be more flexible by providing independent model parameters for gain, limiting
value, and sharpness of the limiting transition; however, analytically the Cann model
has problems with particular derivatives depending on the value s, the sharpness
parameter [68]. In particular, the third-order derivative does not exist when s is an
even integer and, for s > 2 , the third-order term of the Taylor series is zero. Thus for
a two-tone input signal the decibel slope of the third-order intermodulation distortion
term will not be 3:1 in the small signal region of operation whereas many physical
limiting amplifiers exhibit a clear 3:1 ratio at low signal levels. However, there are
65
notable exceptions to the 3:1 ratio such as the multi-tanh circuit [80] which exhibits
nulls in the third harmonic distortion term.
Another issue with the Taylor series expansion is the minimum order required to
accurately represent an amplifier operating in the gain compression region. The range
of validity of a model is defined as the peak signal level for which the model agrees
with the measured or analytical data. The valid power range is defined by the peak to
average of the input signal level where the maximum input power is the maximum peak
signal minus the peak to average of the signal. The model needs to be valid well into
the gain compression region, at least several decibels of compression, to accurately
model high level distortion products. A plot of the Taylor series expansions from
(II.54), of the hyperbolic tangent function (II.47), for odd orders 23, 49, and 99 are
shown in Figure II-14. Notice that none of the Taylor series expansions adequately
represent the hyperbolic tangent function in the fully limited region of the nonlinear
characteristic. Moreover, there is only an incremental improvement in the range for a
doubling of the order of the expansion. This implies that a Taylor series expansion of
very high order is necessary to represent the hyperbolic tangent function in the limiting
region of operation. Recall that the number of spectral terms required for calculation
of the output power spectrum (II.25) is the square of the number of series coefficients
in the model [( N − 1) / 2]2 . Thus it is not efficient to use high order series expansions
{ }
Re vɶo1 = aRe1vɶin1 + aRe3vɶin31 + aRe5vɶin51 + … + aRe m vɶinm1
Re {vɶ } = a
o2 vɶ + aRe3vɶin3 2 + aRe5vɶin5 2 + … + aRe m vɶinm2
Re1 in2
(II.55)
⋮ ⋮ ⋮ ⋮ ⋮ ⋮
{ }
Re vɶoi = aRe1vɶini + aRe3vɶin3 i + aRe5vɶin5 i + … + aRe m vɶinmi
where i denotes the ith input/output sample pair. Separating the approximation
equations into matrix form
{ } vɶ
Re vɶ
o1
in1 vɶin31 ⋯ vɶinm1 a
1
{ } = vɶ
Re vɶ
o2 in2
3
ɶin
v 2
⋯
m
ɶin
v a3
2
(II.56)
⋮
⋮ ⋮ ⋮ ⋮
{ }
Re vˆo
vɶ
ini
vɶin3 i ⋯ vɶinmi am
i
or in vector notation
v o = Vin a . (II.57)
We want to find a solution where the error vector between the approximated and actual
output values is orthogonal to the actual data
Vin e = 0 (II.58)
where
e = v o - Vin a .
Solving (II.58) for a yields the least mean squared error approximation of the output
voltage by the power series model provided the columns of the input voltage matrix are
independent
( )
-1
a = VinT Vin VinT v o (II.59)
67
where ( VinT Vin ) VinT is also know as the pseudo inverse of the overdetermined system
-1
of equations [83].
A plot of the least squares fit of the hyperbolic tangent function for odd order 23 is
shown in Figure II-14. The least squares model fits well into the limiting region of the
nonlinear gain characteristic with just twelve terms of the expansion. A least squares fit
to odd order 39, twenty coefficients, is enough to represent the full transfer
characteristic of the hyperbolic tangent function; however, functions with sharper
transitions to limiting require higher order fits to represent the function far into limiting.
A least squares fit of odd order twelve was applied to each of the nonlinear carrier
transfer functions shown in Figure II-12. An order of twelve was selected as a tradeoff
between goodness of fit to the nonlinear characteristic in the limiting region and the
resulting number of spectral terms needed to calculate the output power spectrum from
(II.25). A plot of the carrier gain characteristics from each of the least squares power
series models is shown in Figure II-15. The full output range is shown to highlight the
valid range of operation for each of the models. In general, the valid range of
operation is less for models with sharper nonlinear transition into compression because
additional series coefficients are needed to accurately fit the sharper transition over a
wider range of signal levels.
68
1.2
1.0
0.8
0.6
Output Voltage (V)
0.4
0.2
0.0
-0.2
-0.4
-0.6
TANH
-0.8 LS 23
Taylor 23
-1.0 Taylor 49
Taylor 99
-1.2
-0.25 -0.20 -0.15 -0.10 -0.05 0.00 0.05 0.10 0.15 0.20 0.25
Input Voltage (V)
28
26
24
22
Gain (dB)
20
18
Cann s=2
Tanh
16 Cann s=4
Multi-Tanh
Cann s=10
14
Cann s=100
12
-30 -25 -20 -15 -10 -5 0
Input Power (dBm)
autocorrelation estimates using (II.24) and (II.21) then taking the FFT of each
correlation term and finally saving all the spectral terms in a single file. A spectral file
only needs to be generated once for each signal of interest. The power of the input
signal used to generate the spectral file is normalized to a convenient value, in this case
0 dBm or 1 mW, such that the power of each spectral term is simply scaled by the ratio
of the desired input power to the normalized input power raised to the appropriate
power of the spectral term
2( n + m +1)
pin
( 2 n +1)( 2 m +1) =
Sɶ Sɶnorm(2 n+1)(2 m+1) (II.60)
pnorm
where the subscript norm denotes the normalized input power of the spectral term.
The output spectrum is calculated by summing the product of the scaled spectral terms
from (II.60) by the corresponding power series coefficient aɶ2 n +1aɶ2*m +1 . A flow chart of
the spectral calculation is shown in Figure II-16. A sweep of the input power is easily
performed by sweeping the pin variable over the power range of interest and
recalculating the sum of spectral terms.
70
Construct Input
Signal
Compute
Input AM-AM and
Autocorrelation
AM-PM Data
Terms
Save Spectral
Save Coefficients
Terms to a Binary
to Text File
File
Input Files
Measure
ACPR
power spectrums for the hyperbolic tangent limiter model are shown in Figure II-17.
The power spectrum for the gain compression/expansion terms from (II.26) is nearly
identical to the spectrum of the input signal. The out-of-band spectrum is limited by
the finite rejection of the CDMA baseband FIR filter. Notice that the out of band
distortion of the composite output power spectrum is also limited by the finite rejection
of the baseband filter response. However, the separated distortion spectrum clearly
shows the distortion components below the limiting value of the input signal.
A composite plot of the output power spectrum for each of the limiter models with
a CDMA IS-95 reverse link input signal applied and an output power of 6 dBm is
shown in Figure II-18. For the same output power, the out of band distortion is
highest for the softer limiter models like the Cann s=2 and hyperbolic tangent models,
and lowest for models with a sharper nonlinear transition. Close in distortion spectrum
for the triplet multi-tanh model is roughly 10.5 dB lower than the hyperbolic tangent at
6 dBm output power. The spectrum of the input signal limits the measurable output
power spectrum at far offsets to the carrier which is why the far out spectrums
shown in Figure II-19 for each of the limiter models. Now the out of band distortion is
limited by the nonlinear transformation of the input signal resulting in a significantly
lower power spectrum than the input signal itself. Of greater interest is the inband
distortion which is hidden by the desired signal when viewing the composite power
spectrum. Again the inband distortion is higher for models with softer limiting
characteristic and correspondingly lower for models with a sharper transition region.
72
However, there is a greater difference between the inband distortion compared to the
out of band distortion between the hyperbolic tangent and triplet multi-tanh models.
This may indicate that the signal waveform quality factor, ρ, will degrade more rapidly
than the increase in the out of band distortion.
-20
-30
Power Spectrum (dBm/2.4 kHz)
-40
Composite Spectrum
-50 Gain Comp./Exp. Spectrum
Distortion Spectrum
-60
-70
-80
-90
-100
-110
-120
-5 -4 -3 -2 -1 0 1 2 3 4 5
Frequency (MHz)
-10
-20
Power Spectrum (dBm/2.4 kHz)
-30
-40
-50
-60
-70
Figure II-18: Total output power spectrum at 6 dBm for each limiter model.
-20
-30
Power Spectrum (dBm/2.4 kHz)
-40
-50
-60
-70
-80
Figure II-19: Distortion spectrum at 6dBm output power for each limiter model.
74
possible for the ACPR performance to degrade with decreasing drive level and the
ACPR specifications apply at reduced output power levels. An input power sweep was
performed from -50 dBm to -5 dBm in 0.5 dB steps to obtain the ACPR versus output
power characteristic for each of the models. The separated distortion spectrum was
used to measure ACPR to eliminate limiting the measurement by the baseband filter
response of the input signal. The ACPR versus output power plots for each of the
limiter models is shown in Figure II-20 and Figure II-21 for distortion offsets of 885
kHz and 1.98 MHz respectively. These two offsets correspond to those specified for
the adjacent and alternate channel ACPR measurement in the IS-95 CDMA
specification [2]. For these plots, the distortion spectrum was compared against the
main channel power to eliminate the effects of the baseband filter response limiting the
There are striking differences between ACPR produced by the different models.
The hyperbolic tangent and Cann s=2 models both exhibit an ACPR to output power
ratio of 2:1 which is expected if the slope of the spectral regrowth versus input power
is 3:1 and the desired signal versus input power is 1:1. A plot of the ACPR slope
versus output power is shown in Figure II-22. A slope of 2:1 is expected for most
class A or class AB amplifiers at lower output power levels where the third order term
of the nonlinearity should dominate the distortion term. The triplet multi-tanh, Cann
s=10, and Cann s=100 models exhibit a 2:1 slope at low output power level; however,
there are noticeable notches in the ACPR response in the 0 dBm to 6 dBm output
power range. The Cann s=4 model has a notch at lower output power levels, around -
75
18 dBm, and the slope appears to be asymptotically approaching 2:1 at lower output
power levels. The multi-tanh exhibits one notch while the Cann s=10 and s=100
models exhibit two notches back to back which is also evident in the slope plots where
there are one and two inversions in the slope. It is interesting that the slope of the
Cann model approaches 2:1 despite the odd behavior reported about the derivates [68].
There are also differences in the ACPR performance of the second offset at 1.98
MHz. The second offset is generally thought to be dominated by fifth order or higher
nonlinearities because the offset falls outside of the spectrum where the third order
test where the input tone spacing is ± 495 kHz and the third order intermodulation
terms would appear at ± 990 kHz and the fifth order terms at ± 1.98 MHz. This is true
for the hyperbolic tangent and Cann s=2 models at higher output power levels where
the ACPR slope visually appears to be 4:1; however, at lower output power levels the
slope decreases to 2:1 indicating a third order dominant term. The other models reveal
a more complicated response at the second offset; however, all but the Cann s=4
models indicate a third order dominance at low output power levels. It should also be
noted that the second offset distortion contributions are low enough, at lower output
power levels, to have little significance to system performance.
Differences in ACPR between the triplet multi-tanh versus the hyperbolic tangent
models are of interest since each represents at physical circuit response. At lower
level. Near the notch of the triplet multi-tanh response the ACPR improvement is as
much as 35 dB. After the notch the distortion rises and approaches that of the
hyperbolic tangent function at high output power levels. This result shows that the
linearity performance for the same output power and bias current of a differential pair
circuit.
driving the nonlinear model [84]. The CDMA waveform has a significantly different
gain compression characteristics should show some differences. The CDMA gain
output power for the desired CDMA signal channel. A plot of CDMA gain
results from Figure II-12, the CDMA gain compression is more significant for models
comparison the CDMA input referred one dB compression point to the sinusoidal
compression results for the nonlinear models is shown in Table II-2. As expected the
wider amplitude variations of the CDMA signal contribute more to the gain
-20
Adjacent Channel Power (dBc)
-40
-60
-80
-100
Cann s=2
Tanh
-120 Cann s=4
Multi-Tanh
Cann s=10
-140
Cann s=100
-160
-22 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12
Output Power (dBm)
Figure II-20: Adjacent channel power at 885 kHz offset for limiter models.
-20
Adjacent Channel Power (dBc)
-40
-60
-80
-100
Cann s=2
Tanh
-120 Cann s=4
Multi-Tanh
Cann s=10
-140 Cann s=100
-160
-22 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12
Output Power (dBm)
Figure II-21: Adjacent channel power at 1.98 MHz offset for limiter models.
79
12
ACPR to Output Power Slope (dBc/dBm)
10
0 Cann s=2
Tanh
Cann s=4
-2 Multi-Tanh
Cann s=10
Cann s=100
-4
-24 -22 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12
Output Power (dBm)
30
29
28
27
26
25
CDMA Gain (dB)
24 Cann s=2
Tanh
23
Cann s=4
22 Multi-Tanh
21 Cann s=10
Cann s=100
20
19
18
17
16
15
-40 -38 -36 -34 -32 -30 -28 -26 -24 -22 -20 -18 -16 -14 -12 -10 -8 -6 -4
Input Power (dBm)
II.8 Summary
This chapter presented an approach to modeling the AM-AM and AM-PM carrier
characteristics of nonlinear circuit using a complex power series behavioral model using
the measured or simulated AM-AM and AM-PM response of the circuit. Power series
coefficients are obtained from a least squared error fit of the coefficients to the
rectangular representation of the AM-AM and AM-PM response of the circuit. Least
squared error formulation was shown to greatly enhance the dynamic range of the
behavioral model compared to a Taylor series expansion of the same or even higher
order. The properties of several limiting amplifier models were introduced with
Power series models were fit to each of the models to use in later analysis.
complex power series model was presented. Distortion components that are centered
about the carrier frequency are separated from other components using a binomial
applied to a bandpass nonlinearity. The output power spectrum is obtained from the
associated with the carrier signal, and distortion components. Autocorrelation analysis
is applied to power series representations of each limiter model and the spectral and
ACPR results are presented. The analysis shows how characteristics of each model
influence the power spectrum results. Models with softer limiter characteristics
generally have worse distortion performance than those with sharper nonlinear
81
transition from linear to limiting operation. Even at reduced output power levels the
The text of this chapter, in part, is a reprint of the material as it appears in our
published papers in IEEE Conferences [CICC01 and IMS01] and publication in IEEE
modulation schemes to increase capacity of the system and raise the available
bandwidth for data users. The first CDMA systems only utilized a single signal or
orthogonal Walsh code channel from the handset to the base station. However newer
CDMA2000 1x, EvDo, and DV standards support synchronous reverse links with a full
time pilot signal and a multitude of control, paging, and traffic Walsh code channels.
The forward link always supported multiple code channels for each of the users and a
pilot code channel [85]. In general, the more code channels that are added the wider
the amplitude variation of the signal. This is a result of a statistical law of random
signals known as the central limit theorem which states that the sum of independent
variables [61]. The statistical properties of most practical communication signals are
through the central limit theorem when several independent signals are summed
together. This opens the door to the use of statistical mathematics to calculate the
The previous section presented the general case of a carrier signal with arbitrary
amplitude and phase modulation passed through a nonlinear behavioral model. The
generality of the analysis lead to a solution with [(N − 1) / 2]2 spectral terms for a
statistical properties of the signal provided the statistical properties are known and are
applicable to the problem at hand. Statistical signal analysis often leads to evaluation of
82
83
include closed form and reduced order solutions over the general autocorrelation
formulation presented in the previous chapter. The benefit for nonlinear circuit analysis
is reduced order equations which yield clearer insight into how the signal and circuits
This section presents a method for predicting spectral regrowth based on the
nonlinear transformation of the statistics of the input signal. A scheme that is simple to
calculate, and therefore more readily available to the practicing engineer seeking to
estimate the nonlinear transfer response, is developed. The proposed formulation seeks
to provide design insight into how the nonlinearity affects the output spectrum, by
input statistics through the modulation scheme and the nonlinearity itself. The output
complex power series obtained from the measured or simulated AM-AM and AM-PM
of statistical and time averaged moments for signals that are ergodic in regards to the
of random process converge to the time average of one realization of the process, in
Rɶ zz (τ ) = E zɶ (t ) zɶ * (t + τ )
T
1
= lim ∫ zɶ (t ) zɶ
*
(t + τ )dt (III.1)
T →∞ 2T
−T
ɶ (τ ) .
=ℜ zz
This is a very useful, and remarkable, property that permits substitution of time
averages for statistical averages. Moments for certain random processes can be
formulated from known properties of the process, but evaluation of the moments from
the statistical description may be mathematically difficult. In those cases, the moments
formulation leads to fewer autocorrelation terms than the general time average
(III.1) to be valid. Strict sense stationarity is a property of a random process where all
the moments of the process are dependent only on the time difference between samples
of the process or put another way the process is invariant with the absolute time each is
variable is sampled. Only the first and second moments are required to be invariant
with absolute time for (III.1) to be valid, such a process is called wide sense stationary
(WSS). Without WSS and ergodicity properties the second moment of a random
process is defined as
which requires knowledge of the joint probability density function, f ZZ [ zɶ(t1 ), zɶ(t2 )] of
the random process. Generally, the joint probability density function is not known.
85
Fortunately, most communication signal problems fit the WSS and ergodic criterion
such that the statistical moments can be replaced by the time average autocorrelation
function.
most well known and studied signals in information theory. The popularity of the
properties and because of the central limit theorem, which states that the sum of
identically distributed, zero mean, independent random processes tends to a zero mean
Gaussian distribution. This is a remarkable and useful property since there are many
processes which are not Gaussian. A Gaussian or normal random process follows the
1 ( x − µ X )2
f X ( x) = exp − (III.3)
2πσ X2 2σ X2
where µ X is the mean value of the process and σ X2 is the variance. Typically wireless
communication system problems deal with zero mean processes because no information
is contained in the mean and nonzero mean processes cause practical problems of
numerical headroom for digital signal processing (DSP) and voltage/current headroom
issues for baseband and RF circuits.
complex addition of two independently filtered stationary random processes, x(t) and
y(t), each assumed to be zero mean:
86
~
z (t ) = x (t ) + jy (t ) (III.4)
where the tilde is used to indicate a quadrature signal. The variables here are voltage-
like quantities; the power of each is proportional to the corresponding autocorrelation
function, and each is assumed to be stationary and identically distributed. Thus the
notation is simplified by introducing subscripts to indicate quantities at different times:
x1 = x(t ) , x 2 = x(t + τ )
y1 = y (t ) , y 2 = y (t + τ ) (III.5)
~
z1 = x1 + jy1 , ~ z 2 = x 2 + jy 2
Simplification of the notation is possible since the stationary assumption permits the
time variable to be removed from the notation because the autocorrelation and
crosscorrelation sequences only depend on the time difference, τ, and not on absolute
time. The autocorrelation function of a complex stationary random process ~ z (t ) is
defined as
~
R zz (τ ) = E[~ z * (t + τ )] = E[~
z (t ) ~ z1 ~
z 2* ] (III.6)
where E[ ] is the expected value operator and * denotes the complex conjugate. Since
random processes x(t) and y(t) are stationary, the following autocorrelation and
crosscorrelation identities apply :
R xx (τ ) = R yy (τ ) = E[ x1 x 2 ] = E[ y1 y 2 ]
R xx (0) = E[ x1 x1 ] = E[ x 2 x 2 ] = E[ y1 y1 ] = E[ y 2 y 2 ]
(III.7)
R xy (τ ) = − R yx (τ ) = E[ x1 y 2 ] = − E[ x 2 y1 ]
R xy (0) = E[ x1 y1 ] = E[ x 2 y 2 ] = 0
~
R zz (τ ) defines the statistics of the modulator’s output, which is the signal input to the
nonlinear device. It should be noted that the real part of the autocorrelation function is
an even function and the imaginary part is an odd function about τ = 0 . The average
power, Pzz, of the input signal, ~
z (t ) , is proportional to the autocorrelation function
evaluated at τ=0:
~
Pzz = AR zz (0)
= A{2 R xx (0) - jR xy (0)} (III.9)
= 2 ARxx (0)
Here A is a power scaling variable used to set the input power level.
The nonlinear device is characterized, in part, using single-tone measurements, and
can be represented as a bandpass nonlinearity with complex transfer characteristic [32]:
∑ a z (t)
N
Gɶ [ zɶ (t )] = ɶi ɶ
i
i =1 (III.10)
= aɶ1 zɶ (t ) + aɶ3 zɶ (t ) + aɶ5 zɶ (t ) + ... + aɶ N zɶ (t )
3 5 N
where a~i are complex power series coefficient and Gɶ [ zɶ(t )] is the complex power
series representation of an AM-AM and AM-PM characteristic. Only the odd terms
can be determined from single-tone complex compression characteristics, but
fortunately the odd order terms are the most important as they produce intermodulation
distortion in band and adjacent to the desired signal. It is important to realize that
(III.10) represents a general nonlinear transfer characteristic and is not a gain
expression.
88
The autocorrelation function of the signal at the output of the nonlinear device is
found by applying (III.6) to the baseband equivalent polynomial model, (III.10), with
coefficients a~ : i
~ ~
Rgg (τ ) = E G(~ [ ~
z1 )G * (~
z2 ) ]
= E a~1 ~[{
z1 + a~3 ~
z1 + a~5 ~
3
z1 a~1 ~
5
z 2 + a~3 (~
* * *
}{
z 2 ) 3 + a~5* (~
* *
z2 )5 }]
= a~ E ~
2
1z~ [ ] 1 2 [
z * + a~ a~ * E ~z (~ ]
1 3 [
z 3 )* + a~ a~ * E ~
1 ]
z (~z 5 )* +
2 1 5 1 2 (III.11)
a~ a~ E[~ z ] + a~ E[~
z ~ z ) ] + a~ a~ E[~
z (~ z ) ]+
z (~
* 3 * 2 3 3 * * 3 5 *
1 3 1 2 3 1 2 3 5 1 2
a~ a~ E[~ z ] + a~ a~ E[~
z ~ z ) ] + a~ E[~
z (~ z ) ] + ...
z (~
* 5 * * 5 3 * 2 5 5 *
1 5 1 2 3 5 1 2 5 1 2
~
Expanding Rgg (τ ) results in many algebraically intensive moment manipulations
involving x(t) and y(t). Fortunately, a previous result for the moments of complex
Gaussian random variables [88] can be used to calculate each of the terms in this
expression, namely
0 , s≠t
[
~ ~ ~ ~ *~ * ~ *
E z1z 2 ...z s z1 z 2 ...zt = ] (III.12)
E~
π
∑
zπ (1) ~
z1* E ~ [
zπ ( 2) ~ ][
z 2* ...E ~
zπ ( s ) ~
zt* ] [ ] , s=t
E zɶ1 zɶ2 zɶ3* zɶ4* = E zɶ1 zɶ3* E zɶ2 zɶ4* + E zɶ2 zɶ3* E zɶ1 zɶ4* (III.13)
[
E (~
z1~ ]
z2* ) n = n! E ~ [ ]
~
z 2* = n! Rzzn (τ )
z1~
n
(III.14)
89
expectation is zero, for example the first order term crossed with the third order term is
Assuming that the input random process is a complex Gaussian variable and application
the moment theorem (III.12) and resulting identities (III.14), (III.15) allows the
expanded form of the output autocorrelation function, (III.11), to be simplified,
yielding the compact expression:
~ ~ ~ ~
R gg (τ ) = a~1 ARzz (τ ) + 3! a~3 A3 R zz3 (τ ) + ... + N ! a~N A N R zzN (τ )
2 2 2
∑ n! a~
N
~ (III.16)
= A n R zzn (τ )
2
n
n =1
n odd
where A is a power scaling variable used to set the input power level. Evaluation of the
output power spectrum requires an estimate of the input autocorrelation function,
~
Rzz (τ ) , of the complex input process ~
z (t ) . A closed form analytical expression for the
from (III.16)
∫ R~ (τ )e dτ
~ ∞
− j 2π fτ
S gg ( f ) = gg
−∞
= ∑ n! a~ A ∫ R (τ )e
N ∞
~ − j 2π f τ
dτ
2 n n
n zz (III.17)
−∞
n =1
= ∑ n! a~ A F {R (τ )}
N
~ 2 n n
n zz
n =1
n odd
where
{ }
∞
F Rɶ zzn (τ ) = ∫ Rɶ zzn (τ )e − j 2π f τ dτ .
−∞
Thus the output power spectrum is the sum of the individual spectra of each term in the
power series, scaled by its corresponding coefficient and input power level. When
computing the output spectrum, the Fourier transform of each spectral component is
calculated first, then scaled by the input power level and the corresponding power
~
series coefficient, and added into (III.17) yielding S gg ( f ) . This derivation shows that
~
only the input autocorrelation function, Rzz (τ ) , and the complex power series
description are necessary to estimate the output power spectrum of the nonlinear
device.
1 1
w(t ) = x (t )e jωct + x(t )e − jωc t (III.18)
2 2
91
the input of a nonlinear circuit represented by the complex power series model from
(II.10)
( N −1) / 2
Gɶ ωc [ zɶ (t ) ]= ∑n =0
aɶ2 n +1 2n + 1
2n
2 n +1
x(t )
2 n +1
. (III.19)
The output autocorrelation function is found by taking the expectation of the output
signal
∑ a2 n+1 2n + 1
∑ a2 m +1 2m + 1
2 ɶ 2 ɶ*
2 n +1 2 m +1
= E 2n x1 2m x (III.20)
n +1 m +1
2
n =0 2 m =0 2
N −1 N −1
∑∑ a2 n+1aɶ2*m +1 2n + 1 2m + 1
2 2 ɶ
2 n +1 2 m +1
= 2( n + m ) E x1 x2 .
n =0 m =0 2 n + 1 m + 1
The expectation can only be evaluated if the moments of the random variable are
known. For the case of a zero mean real Gaussian process the moments are given by
[60, 90]
0 , s odd
E [ x1 x2 ...xs ] = (III.21)
∑
all distinct { [ 1 2 ] [ 3 4 ]
E xx E xx …E[x x
s −1 s ]} , s even .
pairs of subscripts
Expansion of the modulated carrier leads to many cross terms which do not evaluate to
zero as was the case with a complex baseband signal. Here are the moments evaluated
for a seventh order power series expansion
92
[ ] [ ]
n=0, m=1 : E x1 x23 = E x13 x2 = 3Rxo Rxx (τ )
[ ]
n=1, m=1 : E x13 x23 = 9 Rxo
3
Rxx (τ ) + 6 Rxx3 (τ )
[ ] [ ]
n=0, m=2 : E x1 x25 = E x15 x2 = 15Rxo2 Rxx (τ )
[ ] [ ]
n=1, m=2 : E x13 x25 = E x15 x23 = 45 Rxo
3
Rxx (τ ) + 60 Rxo Rxx3 (τ )
[ ]
n=2, m=2 : E x15 x25 = 225 Rxo4 Rxx (τ ) + 600 Rxo2 Rxx3 (τ ) + 120 Rxx5 (τ )
[ ] [ ]
n=0, m=3 : E x1 x27 = E x17 x2 = 105 Rxo
3
Rxx (τ )
[ ] [ ]
n=1, m=3 : E x13 x27 = E x17 x23 = 315 Rxo4 Rxx (τ ) + 630 Rxo Rxx3 (τ )
[ ] [ ]
n=2, m=3 : E x15 x27 = E x17 x25 = 1575 Rxo
5
Rxx (τ ) + 6300 Rxo
3
Rxx3 (τ ) + 2520 Rxo Rxx5 (τ )
[ ]
n=3, m=3 : E x17 x27 = 11025 Rxo6 Rxx (τ ) + 66150 Rxo4 Rxx3 (τ ) + 52920 Rxo2 Rxx5 (τ ) + 5040 Rxx7 (τ )
~
where Rxo = Rxx (τ = 0) .
Collecting terms of equal order and including power series coefficients for clarity
yields:
the linear term (gain expansion/compression term)
93
2
aɶ1 + 3aɶ3 Rxo + 15aɶ5 Rxo2 + 105aɶ7 Rxo
3
Rɶ xx (τ ) (III.22)
2 ~
3! a~3 + 10a~5 Rxo + 105a~7 Rxo2 Rxx3 (τ ) (III.23)
2 ~
5! a~5 + 21a~7 Rxo Rxx5 (τ ) (III.24)
2 ~
7! a~7 Rxx7 (τ ) . (III.25)
Inspection of the terms reveals a pattern to the terms yielding the following expression
for the terms of equal power
2
N −1
where the power series coefficients have been included to illustrate which terms are
associated with the series coefficients. Adding in the binomial coefficient terms along
with the power series coefficients yields an expression for the nonlinear output
autocorrelation function term
2
N −1
a2n +1 ( 2n + 1)!( 2n + 1) !
∑2 Rɶ xx2 k +1 (τ )
2 ɶ
2 k +1
Rɶ gg (τ ) = 3n− k
Rxon −k . (III.27)
n =k (n − k )!(n + 1)!n ! (2k + 1)!
94
N −1
∑R
2
Rɶ gg (τ ) = ɶ 2 k +1 (τ ) .
gg (III.28)
k =0
Thus similarly to the complex Gaussian case, the output autocorrelation function is a
sum of (N − 1) / 2 terms for an N th odd order power series; however, here each
N −1
∑S
2
Sɶgg ( f ) = ɶ 2 k +1 ( f ) (III.29)
gg
k =0
where
{ }
N −1
( 2n + 1)!( 2n + 1)! F Rɶ zz2 k +1 (τ )
∑2
2 ɶa2 n +1
Sɶ 2 k +1 ( f )= 3n− k
R n− k
(III.30)
(n − k )!(n + 1)!n ! (2k + 1)!
gg zo
n= k
and
{ }
∞
F Rɶ zz2 k +1 (τ ) = ∫ Rɶ zz2 k +1 (τ )e − j 2π f τ dτ .
−∞
Use of the moment theorem yielded a closed form expression for the output spectrum
modulated by a complex random process. For this case, the output autocorrelation
function is formulated by taking the expectation of the output of the nonlinear amplifier
∑ a2 n+1 2n + 1 n +1 *
∑ 2m + 1
*
( ) (z )
2 ɶ n 2 ɶ m +1
a m
= E 2n zɶ1 zɶ1
2 m +1
*
ɶ2 ɶ2
z
(III.31)
= 2 n + 1 m =0
2m
2 m +1
n 0
N −1 N −1
∑∑ a2n +1aɶ2*m +1 2n + 1 2m + 1 n +1 *
( ) ( zɶ )
2 2 ɶ n m +1
= 2( n + m ) E zɶ1 zɶ1 *
zɶ2m .
n +1 m +1
2
n =0 m =0 2
Thus the power series expansion and resulting cross terms in the autocorrelation
N −1 N −1
∑∑ R
2 2
Rɶ gg (τ ) = ɶ
(2 n +1)(2 m +1) (τ ) . (III.33)
n =0 m =0
Evaluation of the expectation function is more complicated for this case due to the
complex conjugate terms generated by the envelope of the carrier. The cross terms do
not evaluate to zero as they did for the complex random process case. Each term of
the expectation is evaluated by applying the complex Gaussian moment theorem (II.10)
n=0, m=0 : E ~
z1 ~
z 2* [ ] ~
= Rzz (τ )
[
n=0, m=1 : E ~z1 ~z 2 (~
z 2* )
2
] = 2R~ (τ )R zz zo
[
n=1, m=1 : E ~z12 ~z 2 ~z1* (~z 2* )
2
] = 4R~ zz
~ ~
(τ ) R zo2 + 2 R zz2 (τ ) R zz* (τ )
[
n=0, m=2 : E ~z1 ~z 22 (~z 2* )
3
] = 6R~ zz (τ ) R zo2
[
n=1, m=2 : E ~z12 ~z22 ~z1* (~
3 ~
]
z 2* ) = 12 Rzz (τ ) Rzo3 + 12 Rzz2 (τ ) Rzz* (τ ) Rzo
~ ~
[
n=2,m=2: E ~z13 ~z 22 (~z1* ) (~z 2* )
2 3
] = 36R~ zz
~ ~ ~
[~
(τ ) R zo4 + 72 R zz2 (τ ) R zz* (τ ) R zo2 + 12 R zz3 (τ ) R zz* (τ ) ] 2
[
n=0, m=3 : E ~z1 ~z 23 (~z 2* )
4
] = 24R~ (τ )R zz
3
zo
n=2, m=3:
E~[
z13 ~ ( ) (~z )
z 23 ~
z1*
2 *
2
4
] = 144R~ zz
~ ~ ~ ~ 2
(τ ) R zo5 + 432 R zz2 (τ ) R zz* (τ ) R zo3 + 144 R zz3 (τ ) R zz* (τ ) R zo [ ]
n=3, m=3:
E~[
z14 ~
z 23 ~( ) (~z )
z1*
3 * 4
2 ] = 576R~ zz
~ ~
(τ ) R zo6 + 2592 Rzz2 (τ ) R zz* (τ ) Rzo4 +
~ ~
[ 2 ~
] ~
1728R zz3 (τ ) Rzz* (τ ) R zo2 + 144 Rzz4 (τ ) R zz* (τ ) [ ] 3
97
~
where R zo = R zz (τ = 0) .
Collecting terms of equal order and adding in the power series coefficients for clarity
2 ~
a~1 + 2a~3 Rzo + 6a~5 Rzo2 + 24a~7 Rzo3 + 120a~9 Rzo4 Rzz (τ ) (III.34)
2 ~ ~
1!2! a~3 + 6a~5 Rzo + 36a~7 Rzo2 + 240a~9 Rzo3 Rzz2 (τ ) Rzz* (τ ) (III.35)
2 ~ ~
2!3! a~5 + 12a~7 Rzo + 120a~9 Rzo2 Rzz3 (τ ) Rzz* (τ ) [ ] 2
(III.36)
2 ~ ~
3!4! a~7 + 20a~9 Rzo Rzz4 (τ ) Rzz* (τ )[ ]
3
(III.37)
2 ~
[ ~ 4
4!5! a~9 Rzz5 (τ ) Rzz* (τ ) . ] (III.38)
Inspection of the terms reveals a pattern to the terms yielding the following expression
for the terms of equal power
2
N −1 k
Rɶ zzk +1 (τ ) Rɶ zz* (τ )
∑ a2n +1n !(n + 1)! n −k
2 ɶ
2 k +1
Rɶ gg (τ ) = Rzo (III.39)
n =k (n − k )! k !(k + 1)!
where the power series coefficients have been included to illustrate which terms are
associated with the series coefficients. Adding in the binomial coefficients yields, after
simplification, the expression for the nonlinear output autocorrelation function term
98
2
N −1 k
a2n +1 ( 2n + 1)! Rɶ zzk +1 (τ ) Rɶ zz* (τ )
∑
2 ɶ
2 k +1
Rɶ gg (τ ) = Rzon− k (III.40)
n =k 2 2n ( n − k )! k !(k + 1)!
N −1
∑R
2
Rɶ gg (τ ) = ɶ 2 k +1 (τ ) .
gg (III.41)
k =0
Thus similarly to the real bandpass Gaussian case the output autocorrelation
function is a sum of ( N − 1) / 2 terms for an Nth odd order power series; although, the
autocorrelation terms are products of the autocorrelation and its complex conjugate.
The output autocorrelation function is a closed form expression in terms of
autocorrelation of the input signal and the sum of the input power weighted by the
power series coefficients. Note that there are only N autocorrelation terms as
compared to N2 for the general time domain case. The output power spectrum is the
Fourier transform of the autocorrelation function
N −1
∑S
2
Sɶgg ( f ) = ɶ 2 k +1 ( f ) (III.42)
gg
k =0
where
{ }
2
N −1 k
and
{
F Rɶ zzk +1 (τ ) Rɶ zz* (τ )
k −1
}= ∞
∫−∞ Rzz
ɶ k +1 (τ ) R
ɶ*
zz
k
(τ ) e − j 2π f τ dτ .
99
Similarly to the real bandpass Gaussian case, use of the moment theorem yielded a
closed form expression for the output spectrum in terms of the input autocorrelation
evaluated from (II.25) for the general autocorrelation case. Calculation of each of the
autocorrelation terms from (III.27) and (III.40) only requires calculation of the
autocorrelation of the input signal, Rɶ zz (τ ) . Each term is a product of powers of the
-10
-20
Power Spectrum (dBm/2.4 kHz)
-30
-40
-50
-60
-70
Figure III-1: Power spectrum at 2 dBm with complex Gaussian input signal.
-20
-30
Power Spectrum (dBm/2.4 kHz)
-40
-50
-60
-70
-80
Figure III-2: Distortion spectrum at 2 dBm with complex Gaussian input signal.
102
-10
-20
Power Spectrum (dBm/2.4 kHz)
-30
-40
-50
-60
-70
-80
-90
-100
-110
-5 -4 -3 -2 -1 0 1 2 3 4 5
Frequency (MHz)
correspond to those specified for the adjacent and alternate channel ACPR
measurement in the IS-95 CDMA specification [2]. For these plots, the distortion
spectrum was compared against the main channel power to eliminate the effects of the
baseband filter response limiting the minimum distortion measurement. In addition, a
plot ACPR at 885 kHz and 1.98 MHz offsets versus output power for a real Gaussian
input signal are shown in Figure III-5 and Figure III-7 respectively. Power sweep
results from both the Gaussian moment formulation and the general autocorrelation
formulation are shown in each plot. The general autocorrelation results are solid lines
and the Gaussian moment results are data points indicated by shaped markers.
Generally the agreement is excellent between the Gaussian moment and general
autocorrelation techniques. Differences occur near the compression region where the
Gaussian moment results do not extend to the quite same dynamic range as the general
autocorrelation results.
There are differences between ACPR produced by the different models similar to
the differences observed with a CDMA input signal. The hyperbolic tangent and Cann
s=2 models both exhibit an ACPR to output power ratio of 2:1 which is expected if the
slope of the spectral regrowth versus input power is 3:1 and the desired signal versus
input power is 1:1. A plot of the ACPR slope versus output power is shown in Figure
III-8. A slope of 2:1 is expected for most class A or class AB amplifiers at lower
output power levels where the third order term of the nonlinearity should dominate the
distortion term. The triplet multi-tanh, Cann s=10, and Cann s=100 models exhibit a
2:1 slope at low output power level; however, there are noticeable notches in the
ACPR response in the 0 dBm to 6 dBm output power range. The Cann s=4 model has
a notch at lower output power levels, around -21 dBm, and the slope appears to be
asymptotically approaching 2:1 at lower output power levels. The notches observed in
104
the multi-tanh, Cann s=10, and Cann s=100 models are present; however, the notches
are not as distinct appearing more as a plateau than a notch.
Differences in ACPR between the triplet multi-tanh versus the hyperbolic tangent
models are of interest since each represents a physical circuit response. At lower
output power levels the triplet multi-tanh model has an approximate 13 dB
improvement in the ACPR than the hyperbolic tangent model at the same output power
level. Near the notch of the triplet multi-tanh response the ACPR improvement is as
much as 27 dB and 25 dB for the complex and real Gaussian signals respectively which
contrasts with the 35 dB improvement for a CDMA input signal. After the notch the
distortion rises and approaches that of the hyperbolic tangent function at high output
power levels. This result shows that the multi-tanh technique of linearity improvement
provides substantial improvement in linearity performance for the same output power
and bias current of a differential pair circuit.
There are also differences in the ACPR performance of the second offset at 1.98
MHz. The second offset is generally thought to be dominated by fifth order or higher
nonlinearities because the offset falls outside of the spectrum where the third order
distortion term dominates. This is equivalent to a two-tone intermodulation distortion
test where the input tone spacing is ± 495 kHz and the third order intermodulation
terms would appear at ± 990 kHz and the fifth order terms at ± 1.98 MHz. This is true
for the hyperbolic tangent, Cann s=2, and Cann s=100 models at higher output power
levels where the ACPR slope visually appears to be 4:1; however, at lower output
power levels the slope decreases to 2:1 indicating a third order dominant term. The
multi-tanh, Cann s=4, and Cann s=10 models show a 4:1 response down to low output
power. It should also be noted that the second offset distortion contributions are low
enough, at lower output power levels, to have little significance to system performance.
105
-10
-20
-30
Adjacent Channel Power (dBc)
-40
-50
-60
-70
-80
-90
-10
-20
-30
Adjacent Channel Power (dBc)
-40
-50
-60
-70
-80
-90
-100
Cann s=2 GM-Cann s=2
-110 Tanh GM-Tanh
Cann s=4 GM-Cann s=4
-120 Multi-Tanh GM-Multi-Tanh
Cann s=10 GM-Cann s=10
-130
Cann s=100 GM-Cann s=100
-140
-24 -22 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
Output Power (dBm)
-30
-40
-50
Adjacent Channel Power (dBc)
-60
-70
-80
-90
-100
-110
-30
-40
-50
Adjacent Channel Power (dBc)
-60
-70
-80
-90
-100
-110
12
ACPR to Output Power Slope (dBc/dBm)
10
Cann s=2
0 Tanh
Cann s=4
Multi-Tanh
-2
Cann s=10
Cann s=100
-4
-24 -22 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
Output Power (dBm)
27
26
25
24
CDMA Gain (dB)
23 Cann s=2
Tanh
22 Cann s=4
Multi-Tanh
21 Cann s=10
Cann s=100
20
19
18
17
-40 -38 -36 -34 -32 -30 -28 -26 -24 -22 -20 -18 -16 -14 -12 -10
Input Power (dBm)
-20
Adjacent Channel Power (dBc)
-40
-60
-80
-140
-24 -22 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12
Output Power (dBm)
III.7 Summary
This chapter presented a statistical approach to calculating power spectrum of a
wide sense stationary random signal passed through a bandpass nonlinearity described
by a complex power series. The moment properties of real and complex Gaussian
random variables were used to formulate the output autocorrelation function when
these signals are passed through a complex power series. The use of moment
properties results in a formulation containing ( N − 1) / 2 spectral terms compared to the
[(N − 1) / 2]2 spectral terms from the general autocorrelation formulation presented in
chapter I. Evaluation of the statistical moments are possible through the ergodic and
wide sense stationary properties of real and complex Gaussian random processes
thereby permitting substitution of time average autocorrelation functions for the
statistical moments. The simplified moment formulations for both real and complex
111
Gaussian random processes were shown to be in excellent agreement with the general
autocorrelation formulations passed through the same nonlinear limiter models.
The text of this chapter, in part, is a reprint of the material as it appears in our
published papers in IEEE Conferences [CICC01 and IMS01] and publication in IEEE
Transactions on Microwave Theory and Techniques [1999 and submitted 2003]. The
dissertation author was the primary investigator and author of these papers.
IV. Simulation and Measurements
Previous chapters presented the analysis, formulation, and spectral results for the
general autocorrelation function and Gaussian moment analysis techniques for
obtaining the output power spectrum from a signal passed through a bandpass
nonlinearity. This chapter presents results of comparing the analysis techniques against
a commercial electronic design automation EDA software package and ACPR
characterization of an integrated wireless transmitter amplifier. Characterization data
for a 900 MHz CDMA driver amplifier, and application of both the time average and
statistical autocorrelation functions to CDMA and Gaussian signals respectively.
Results of the autocorrelation function methods for determining spectral regrowth are
compared to measured data to validate the analysis techniques. Measurement
techniques for characterizing the AM-AM and AM-PM response of RFIC devices with
RF input and RF outputs as well as devices with baseband inputs and RF outputs such
is the case with superheterodyne and direct conversion transmitter RFIC devices.
112
113
triplet, Cann s=2, and Cann s=10 limiter models with CDMA and complex Gaussian
inputs in Figure IV-1 and Figure IV-2 respectively.
The envelope simulation results are in agreement with the general autocorrelation
results down to an ACPR value of approximately -65 dBc and -80 dBc for CDMA and
complex Gaussian input signals respectively. Envelope simulation results below -65
dBc are masked by the finite rejection of the baseband filter as explained previously in
Chapters I and I. The filter effects are evident by each of the ACPR traces converging
to approximately -70 dBc at lower output power levels. Unfortunately, effects of the
finite rejection of the filter can not be removed from the envelope simulation results.
This is one advantage of the general autocorrelation formulation where the distortion
terms can be clearly separated from the components correlated with input signal. The
apparent hump in the ACPR results between -10 dBm and 1 dBm output power for the
multi-tanh model with a CDMA input signal is a result of the summation of the
distortion spectrum and the finite rejection of the baseband filter.
The envelope simulation parameters were set to a stop time of 615 µsec and a
sampling time of 208.3 nsec for 1626 Hz resolution and 4.8 MHz spectral bandwidth.
Each simulation was swept in power from -50 dBm to -5 dBm in 1dB steps as
compared to the same range with 0.5 dB steps used in the general autocorrelation
results. Each envelope simulation run took approximately 90 sec on a Pentium 4
processor running at 2.4 GHz compared to 13 sec for the general autocorrelation
formulation and 6 sec for the Gaussian moment formulation with twice the number of
input power data points results using MATLAB. It should be noted that the envelope
simulation time did not include the several seconds of post processing time required to
read the data file and calculate the power spectrums from each power point. The data
files from the simulation contained the input and output signal waveforms for each
power point simulated resulting in a binary file size of 102 Mbytes per simulation. The
114
-20
Cann s=2 Env Cann s=2
Adjacent Channel Power (dBc)
-40
-50
-60
-70
-80
-22 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12
Output Power (dBm)
-10
-20
Cann s=2 Env Cann s=2
Adjacent Channel Power (dBc)
-50
-60
-70
-80
-90
-24 -22 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
Output Power (dBm)
Analyzer
836.365200 MHZ 836.365200 MHZ
Source Reference
VNA
Generator Generator
+10 dB 130°
DUT
A B
DUT Vector
Voltmeter
Figure IV-3: Measurement setups for AM-AM AM-PM characterization.
The device under test (DUT) is a 900 MHz CDMA/AMPS driver amplifier device
fabricated using a GaAs MESFET technology [62]. The device is a two stage amplifier
designed to provide 23.4 dB of power gain, in a 50 Ω system, and meet CDMA ACPR
specification requirements at an output power of 8 dBm. A schematic diagram of the
117
amplifier is shown in Figure IV-4. A vector network analyzer with a built in power
sweep function was used to measure the AM-AM and AM-PM characteristics at the
connector ports of an evaluation board. The measured AM-AM and AM-PM
characteristics for an input power range of -25 dBm to -2 dBm are shown in Figure
IV-5 and Figure IV-6 respectively. The gain compression characteristic is quite soft
showing a shallow compression characteristic leading into the knee of the curve before
entering full gain compression. There is little phase change until the knee of the gain
compression characteristic, around -10 dBm input power, then the phase transitions to
a nearly linear in dB relationship.
VDD
RF OUT
RF IN GAIN
SELECT
SWITCH
DRIVER
BY-PASS
ATTENUATOR
25
24
AM-AM Gain Compression (dB)
23
22
21
20
19
18
17
16
15
-25 -24 -23 -22 -21 -20 -19 -18 -17 -16 -15 -14 -13 -12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2
Input Power (dBm)
20
19
18
AM-PM Characteristic (Degrees)
17
16
15
14
13
12
11
10
9
8
7
6
-25 -24 -23 -22 -21 -20 -19 -18 -17 -16 -15 -14 -13 -12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2
Input Power (dBm)
Modern RFIC integrated heterodyne and direct conversion transmitter circuits [91-
96] pose an interesting challenge for measurement of AM-AM and AM-PM
characteristics because the input is a quadrature baseband signal and the output is a RF
signal. Block diagrams of a superheterodyne and direct conversion RFIC devices are
shown in A simple solution is to sweep a DC offset at the in-phase and quadrature
inputs of the device thus generating carrier feedthrough which is proportional to the
DC offset at the input. The input DC offset is swept from the fullscale input value
down to a reasonable level to achieve a linear gain and phase measurement. Gain and
phase are determined by comparing the output carrier feedthrough to a fixed reference
source using a vector voltmeter. An example measurement setup is shown in Figure
IV-8. The input signal is controlled by an arbitrary waveform generator which accepts
digital inputs from a computer and generates an analog output signal. Additional
sinusoidal signal generators are required to present local oscillator (LO) signals which
are phase locked to the same PLL reference signal to insure a stable phase comparison
measurement against the reference signal generator.
120
RF LO LC PLL
LC
I/Q
DAC
RF/BB Superheterodyne RFIC
RX LO PLL PLL
I/Q
DAC
RF/BB Direct Conversion RFIC
Figure IV-7: Block diagrams of RFIC transmitter devices.
A B
Q
Vector
Arbitrary Waveform RFIC DUT Voltmeter
Generator
Figure IV-8: RFIC transmitter measurement setup.
121
One drawback of using a swept DC input offset signal is the finite carrier
suppression of the RFIC DUT will limit the lower bound on the linear gain
measurement. Carrier suppression in the DUT is determined by mismatches between
devices and layout imbalances in the circuits internal to the DUT. As the input DC
offset is reduced the finite carrier suppression starts to contribute to the output signal
resulting in an apparent increase in gain because the output signal is not dropping in
relation to the input signal. One solution for overcoming this limitation is to use a
quadrature sinusoidal input signal instead of a DC offset. A single sideband signal is
generated which is offset to the carrier; thus, the amplitude of the sinusoid can be
measured at levels lower than the carrier feedthrough. However, this measurement
technique requires that the input signal is also carefully locked to the same PLL
reference signal as the LO and reference signal generators in order to make a stable
phase comparison measurement.
The RFIC setup shown in Figure IV-8 was used to measure the AM-AM and AM-
PM characteristics of a superheterodyne [95] and a direct conversion [91] dual band
CDMA transmitter devices. The superheterodyne RFIC requires the use of two LO
signal generators: one for the conversion from baseband to the first intermediate
frequency (IF) and a second IF LO to upconvert from IF to the RF output frequency.
The input DC offset was swept over a 40 dB range referenced to the full scale input
signal value. The RFIC supports over 90 dB of dynamic gain control range; however,
ACPR performance is typically worst at the maximum rated output power.
Accordingly, the gain control settings required to achieve rated output power values
are determined before making the AM-AM and AM-PM measurements. The measured
AM-AM and AM-PM characteristics for the superheterodyne RFIC operating in the
cell band at 8.3 dBm CDMA output power and PCS band at 9.7 dBm CDMA output
power are shown in Figure IV-9 and Figure IV-10 respectively.
122
77.0 136
76.5 134
76.0 132
Phase (degree)
Gain (dB)
75.5 130
75.0 128
74.5 126
74.0 124
73.5 122
-40 -35 -30 -25 -20 -15 -10 -5 0
Baseband Input Level (dB)
72.6 -60
72.4 -62
72.2 -64
72.0 -66
71.8 -68
Phase (degree)
Gain (dB)
71.6 -70
71.4 -72
71.2 -74
71.0 -76
70.8 -78
70.6 -80
70.4 -82
70.2 -84
-40 -35 -30 -25 -20 -15 -10 -5 0
Baseband Input Level (dB)
80.0 -52
79.5 -54
79.0 -56
78.5 -58
Phase (degree)
78.0 -60
Gain (dB)
77.5 -62
77.0 -64
76.5 -66
76.0 -68
75.5 -70
75.0 -72
-40 -35 -30 -25 -20 -15 -10 -5 0
Baseband Input Level (dB)
Figure IV-11: Cell band AM-AM AM-PM for direct conversion RFIC.
82.0 -70
81.5 -75
81.0 -80
80.0 -90
Gain (dB)
79.5 -95
79.0 -100
78.5 -105
78.0 -110
77.5 -115
77.0 -120
-40 -35 -30 -25 -20 -15 -10 -5 0
Baseband Input Level (dB)
Figure IV-12: PCS band AM-AM AM-PM for direct conversion RFIC.
125
Table IV-1: Complex power series coefficients for 900 MHz CDMA amplifier.
24 -156
23 -158
22 -160
20 -164
MODELED
19 -166
MEASURED
18 -168
17 -170
16 -172
15 -174
-24 -22 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2
Input Power (dBm)
A carrier modulated with a IS-95 CDMA reverse link signal is applied to the
amplifier circuit and the output distortion measured using a spectrum analyzer.
Specifically, an Agilent ESG series signal generator with the capability to generate a
IS-95 CDMA signal was used as the signal source and an Agilent vector signal analyzer
(VSA) was used to measure the adjacent channel power ratio (ACPR) is the standard
distortion measurement for CDMA transmitters. The VSA equipment has built in
measurement routines to measure ACPR for specified offsets to the carrier frequency.
ACPR is the ratio, in decibels, of the distortion power, in a 30 kHz bandwidth offset by
±885 kHz, and the desired channel power, in a 1.23MHz bandwidth as defined in the
IS-95 standard [2]. The measured ACPR is shown in Figure IV-14. The ACPR was
calculated using the general time average correlation function formulation presented in
Chapter I and the power series coefficients from Table IV-1. The simulated ACPR
results using the composite of the gain compression/expansion (II.26) and distortion
127
terms (II.27) agree well with the measured data shown in Figure IV-14. The ACPR
plateaus at lower output power because of the finite rejection of the CDMA baseband
filter used by the waveform generator.
-15
-20
-25
-30
ACPR (dBc)
-35
-40
-45
-50
-55
Measured ACPR
-65
-2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13
Output Power (dBm)
Figure IV-14: Measured and calculated ACPR for CDMA reverse link signal.
A complex Gaussian signal was also used to measure ACPR, using the IS-95
definition, to compare against both the complex Gaussian moment and the time average
autocorrelation formulations. The Agilent ESG signal generator is capable of
generating a complex Gaussian signal filtered with a finite impulse response (FIR) filter
with a maximally flat response. The ACPR measurements along with the predicted
ACPR from both the complex Gaussian moment and time average autocorrelation
formulations are shown in Figure IV-15. The measured and predicted ACPR are in
good agreement below an output power level of 11 dBm. Both the complex Gaussian
moment and time average autocorrelation formulations deviate from the measured data
because of the limited dynamic range of the complex power series model of the
128
nonlinear amplifier. The difference in dynamic range between using the CDMA and
complex Gaussian input signals is approximately 2 dB which is consistent with
differences found with the hyperbolic tangent limiter model when comparing Figure
II-20 and Figure III-4.
-10
-15
-20
-25
-30
ACPR (dBc)
-35
-40
-45
-50
Measured Complex Gaussian
-55
Complex Gaussian Moment Formulation
-60 Time Average Autocorrelation Formulation
-65
-2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13
Output Power (dBm)
Figure IV-15: Measured and calculated ACPR for complex Gaussian input
signal.
IV.4 Summary
This chapter presented the results of comparing the time average autocorrelation
and complex Gaussian moment formulations against a commercial transient envelope
simulation tool and measurements from an integrated CDMA amplifier. Comparisons
with a commercial simulator verify the equivalence of the time average autocorrelation
and complex Gaussian moment formulations with time domain solutions using a
Fourier transform to obtain the power spectrum. The time average autocorrelation and
complex Gaussian moment calculations provide a significant advantage in the processor
129
time and data storage required to perform a power sweep analysis of ACPR.
Although, equivalent usage of autocorrelation methods, during the design phase of a
wireless circuit, requires single tone power sweep simulations to generate the AM-AM
and AM-PM characteristics. An added benefit of this procedure is that the power
spectrum or distortion characteristics caused by different input signals can be
determined quickly once the AM-AM and AM-PM characteristics are extracted.
Comparisons of the time average autocorrelation and complex Gaussian moment
formulations with measured ACPR data from a CDMA amplifier validate the
techniques when used with real data from an actual nonlinear wireless amplifier. Thus
the techniques described can be used during the design phase with simulated AM-AM
and AM-PM characteristics or during the characterization phase with measured data.
The text of this chapter, in part, is a reprint of the material as it appears in our
published papers in IEEE Conferences [RAWCON03, CICC01, and IMS01] and
publication in IEEE Transactions on Microwave Theory and Techniques [1999 and
submitted 2003]. The dissertation author was the primary investigator and author of
these papers.
V. Conclusions
Two methods for analyzing the power spectrum of modulated carriers passed
through a nonlinear wireless circuit were presented. The first method is based on
formulating the time average autocorrelation function for a signal passed through a
complex power series behavioral model of the AM-AM and AM-PM characteristics of
a wireless amplifier. The power spectrum of the signal is obtained via the Fourier
transformation of the output autocorrelation function leading to a summation of
[( N − 1) / 2]2 for a Nth odd order power series expansion of the nonlinearity. The second
method is based on a transformation of the statistical properties of the modulated
carrier passed through a nonlinear amplifier. The second order moments of the
nonlinear terms are calculated and combined leading to a closed form expression of the
output autocorrelation in terms of the autocorrelation function of the input signal. The
statistical formulation yields ( N − 1) / 2 spectral terms for real and complex Gaussian
130
131
motivation of this work is to provide circuit designers with an efficient tool that
provides insight into the interactions between signals and nonlinear circuits. Likewise,
this work can also be used by wireless system designers to better understand the impact
of nonlinear channels on system performance. Thus the modeling and analysis
techniques presented bring circuit and system designers closer to understanding the
impact of nonlinear circuits used in wireless communication systems. Future
contributions to this work should focus on improving the modeling techniques to
incorporate behavior not accounted for in the bandpass nonlinearity assumption and to
apply the analysis techniques to communication system performance analysis. Future
work is divided into three primary areas: distortion analysis, wireless system analysis,
and behavioral modeling of nonlinear circuits.
interact terms filter by a low-pass filter to model the baseband frequency dependencies.
Such a model also requires more sophisticated measurement techniques to characterize
the frequency dependent asymmetric intermodulation response.
Appendix A: Power Series Coefficients for Limiter Models
134
135
137
138
[2] TIA/EIA, "TIA/EIA IS-95, Mobile station-base station compatibility standard for
dual-mode wideband spread-spectrum cellular systems," Telecommunications
Industry Assoc. TIA/EIA IS-95, July 1993 1993.
[10] N. Wiener, "Generalized Harmonic Analysis," Acta Mathematica, vol. 55, pp.
117-258, 1930.
145
146
[14] K. S. Kundert, The Designers Guide to SPICE and SPECTRE. Boston: Kluwer
Academic Publishers, 1995.
[16] R. J. Gilmore and M. B. Steer, "Nonlinear circuit analysis using the method of
harmonic balance-a review of the art. I. Introductory concepts," International
Journal of Microwave & Millimeter-Wave Computer-Aided Engineering, vol. 1,
pp. 22-37, 1991.
[18] V. Rizzoli, N. Neri, F. Mastri, and A. Lipparini, "A Krylov-subspace technique for
the simulation of integrated RF/microwave subsystems driven by digitally
modulated carriers," International Journal of RF & Microwave Computer-Aided
Engineering, vol. 9, pp. 490-505, 1999.
[19] V. Rizzoli, F. Mastri, C. Cecchetti, and F. Sgallari, "Fast and robust inexact
Newton approach to the harmonic-balance analysis of nonlinear microwave
circuits," IEEE Microwave and Guided Wave Letters, vol. 7, pp. 359-361, 1997.
[21] V. Borich, J. East, and G. Haddad, "An efficient Fourier transform algorithm for
multitone harmonic balance," IEEE Trans. Microwave Theory and Tech., vol. 47,
pp. 182-188, 1999.
147
[25] E. Ngoya and R. Larcheveque, "Envelop transient analysis: a new method for the
transient and steady state analysis of microwave communication circuits and
systems," Microwave Symposium Digest, 1996., IEEE MTT-S International,
1996.
[26] T. J. Aprille, Jr. and T. N. Trick, "Steady-state analysis of nonlinear circuits with
periodic inputs," Proc. of the IEEE, vol. 60, pp. 108-114, 1972.
[27] F. R. Colon and T. N. Trick, "Fast periodic steady-state analysis for large-signal
electronic circuits," IEEE Journal of Solid-State Circuits, vol. 8, pp. 260-269,
1973.
[32] M. B. Steer and P. J. Khan, "An algebraic formula for the output of a system with
large-signal, multifrequency excitation," Proc. of the IEEE, vol. 71, pp. 177-179,
1983.
148
[43] R. Baum, "The correlation function of smoothly limited Gaussian noise," IEEE
Trans. on Information Theory vol. 3, pp. 193-197, 1957.
149
[44] R. Baum, "The correlation function of Gaussian noise passed through nonlinear
devices," IEEE Trans. on Information Theory vol. 15, pp. 448-456, 1969.
[46] M. Rudko and D. Weiner, "Addendum to "Volterra Systems with Random Inputs:
A Formalized Approach"," IEEE Trans. on Communications, vol. 27, pp. 636-
638, 1979.
[47] M. Rudko and D. Weiner, "Volterra Systems with Random Inputs: A Formalized
Approach," IEEE Trans. on Communications, vol. 26, pp. 217-227, 1978.
[48] Q. Wu, M. Testa, and R. Larkin, "On design of linear RF power amplifiers for
CDMA signals," Int. J. of Microwave and Millimeter-wave Computer Aided Eng.,
vol. 8, pp. 283-292, 1998.
[49] Q. Wu, H. Xiao, and F. Li, "Linear RF power amplifier design for CDMA signals:
a spectrum analysis approach," Microwave Journal, pp. 22-40, 1998.
[54] S. O. Rice, "Envelopes of narrow-band signals," Proc. of the IEEE, vol. 70, pp.
692-699, 1982.
[64] W. R. Curtice, "GaAs MESFET modeling and nonlinear CAD," IEEE Trans.
Microwave Theory and Tech., vol. 36, pp. 220-230, 1988.
[65] T. Kacprzak and A. Materka, "Compact DC model of GaAs FETs for large-signal
computer calculation," IEEE Journal of Solid-State Circuits, vol. 18, pp. 211-
213, 1983.
[66] A. J. Cann, "Nonlinearity model with variable knee sharpness," IEEE Trans.
Aerosp. Electron. Syst., vol. AES-16, pp. 874-877, 1980.
[68] S. L. Loyka, "On the use of Cann's model for nonlinear behavioral-level
simulation," IEEE Trans. on Vehicular Tech., vol. 49, pp. 1982-1985, 2000.
151
[70] Y. Guo and J. R. Cavallaro, "Enhanced power efficiency of mobile OFDM radio
using predistortion and post-compensation," Vehicular Technology Conference,
2002. Proceedings, 2002.
[71] M. Kuipers and R. Prasad, "Pre-distorted amplifiers for OFDM in wireless indoor
multimedia communications," Vehicular Technology Conference Proceedings,
2000. VTC 2000-Spring Tokyo. 2000 IEEE 51st, 2000.
[72] H.-G. Ryu, B.-I. Jin, and I.-B. Kim, "PAPR reduction using soft clipping and ACI
rejection in OFDM system," Consumer Electronics, IEEE Transactions on, vol.
48, pp. 17-22, 2002.
[74] M. Wen-jie, R. Li-xing, and C. Kang-shen, "RF solid state power amplifier
linearization using unbalanced parallel predistortion structure," Microwave and
Millimeter Wave Technology, 2002. Proceedings. ICMMT 2002. 2002 3rd
International Conference on, 2002.
[76] P. R. Gray and R. G. Meyer, Analysis and design of analog integrated circuits,
3rd ed. ed. New York: Wiley, 1993.
[80] B. Gilbert, "The multi-tanh principle: a tutorial overview," IEEE Journal of Solid-
State Circuits, vol. 33, pp. 2-17, 1998.
[81] S. Liwei, J. C. Jensen, and L. E. Larson, "A wide-bandwidth Si/SiGe HBT direct
conversion sub-harmonic mixer/downconverter," IEEE. IEEE Journal of Solid-
State Circuits, vol. 35, pp. 1329-37, 2000.
[83] M. H. Hayes, Statistical Digital Signal Processing and Modeling. New York:
John Wiley & Sons, Inc., 1996.
[85] J. S. Lee and L. E. Miller, "Analysis of peak-to-average power ratio for IS-95 and
third generation CDMA forward link waveforms," IEEE Trans. on Vehicular
Tech., vol. 50, pp. 1004-1013, 2001.
[88] I. Reed, "On a moment theorem for complex Gaussian processes," IEEE Trans.
on Information Theory vol. 8, pp. 194-195, 1962.
[89] V. Aparin, "Analysis of CDMA signal spectral regrowth and waveform quality,"
IEEE Trans. Microwave Theory and Tech., vol. 49, pp. 2306-2314, 2001.
[95] K. Sahota, K. Gard, B. Walker, S. Szabo, W. Su, and E. Zeisel, "A base-band to
RF BiCMOS transmitter RFIC for dual-band CDMA/AMPS wireless handsets,"
Radio Frequency Integrated Circuits (RFIC) Symposium, 2000. Digest of Papers.
2000 IEEE, 2000.
[96] A. Springer, L. Maurer, and R. Weigel, "RF system concepts for highly integrated
RFICs for W-CDMA mobile radio terminals," IEEE Trans. Microwave Theory
and Tech., vol. 50, pp. 254-267, 2002.