Professional Documents
Culture Documents
DOI 10.1007/s00034-015-0151-0
Gustavo H. C. Oliveira
gustavo@eletrica.ufpr.br
Ricardo Schumacher
schumacher.ric@gmail.com
Eduardo G. Lima
elima@eletrica.ufpr.br
1 Introduction
Modern wireless systems must deal with the communication of a huge amount of
information, at very high data rates, through a band-limited air interface and in a
power-efficient manner [23]. Spectral efficiency is achieved by modulating a radio
frequency (RF) carrier signal with a complex-valued envelope information having a
high ratio between the peak and average amplitude levels. As a consequence, nonconstant amplitude envelopes demand for linearity in the transmitter chain to avoid
interferences between adjacent channels. In this scenario, the RF power amplifier (PA)
present at the transmitter chain plays a major role. Indeed, the traditional design of
RF PAs, based on solid-state transistors operating in linear classes (A, B or AB), is
subject to a trade-off between linearity and efficiency [4].
To achieve a highly linear and highly efficient amplification, a cost-effective solution
is to include a digital baseband predistortion (DPD) to distort the baseband envelope
signal before the RF amplification [11,12,20]. The DPD is designed to have an inverse
transfer characteristic regarding the RF PA. Therefore, the RF PA is able to operate in
nonlinear regimes to improve the power efficiency, and at the same time the overall
transmitter linearity is maintained within an acceptable value.
In DPD applications, the implementation of the RF PA inverse characteristic is the
final goal, which may be accomplished using the direct or indirect learning architectures. In the indirect architecture, firstly a post-distorter (PoD) topology is a priori
chosen and subsequently the PoD is copied as a DPD. In fact, both the PoD and the
DPD have the same PA inverse characteristic, but differ from each other because of
their position in the cascade connection with the PA. The main motivation for the indirect learning is to simplify the identification of the inverse system, once the same set of
input and output data can be used for the extraction of the PoD and PA model parameters if the roles of the input and output data are exchanged. In the direct architecture,
the parameter identification of a DPD model is directly addressed without the use of
a PoD. However, the DPD output signal is not available. In fact, only the input and
output signals of the cascade connection of the DPD followed by the PA are available.
As a consequence, the DPD parameter identification must minimize the error between
desired and estimated PA output signals. Hence, the physical PA (or a PA model) must
be included in the error expression. In this scenario, unless the inclusion of the physical
PA in a real-time optimization loop could be feasible, a model for the PA is mandatory. Alternatively, the DPD can have a topology that is the inverse of a PA model, in
which the DPD parameters are derived from the PA ones using analytical expressions
relating them. Even though the nonlinear and dynamic behaviors observed at the forward and inverse PA characteristics can be different from each other [7,10,24], it is a
common practice in the literature [7,22] to apply the same topology for modeling both
the forward and inverse PA characteristics. In fact, it can be shown that forward and
inverse models are approximately the same under certain conditions [7]. Therefore,
the successful design of a DPD scheme is strongly conditioned by the availability of
a high-accurate and low-complexity model for representing the inverse (in the case
of both the direct and indirect learning architectures) and also forward (only in the
case of the direct learning architecture) RF PA characteristics [1,15,19,21]. On the
one hand, the model must be able to accurately represent the PA nonlinear behaviors
izations is used to describe such models. The set of poles which parameterizes the
model functions is chosen by using a constrained nonlinear optimization approach.
The paper is organized as follows. Some background information related to lowpass equivalent PA behavioral modeling is given in Sect. 2. In Sect. 3, the class of
low-pass equivalent OBF Volterra models is described. In particular, the use of models
based on TakenakaMalmquist function sets and state-space realizations is motivated
in this section. Section 4 shows how TMV models can be extracted even without
prior information about the basis function dynamics. In Sect. 5, the accuracy of the
TMV model is fairly compared with the accuracies of the CV and the OBF Volterra
models previously reported in the literature for the low-pass equivalent PA behavioral
modeling. Finally, Sect. 6 addresses the conclusions of this work.
(1)
where R(t) and (t) are the real-valued amplitude and angle components, respectively,
of the complex-valued envelope signal x(t),
(2)
where S(t) and (t) + (t) are the real-valued amplitude and angle components,
respectively, of the complex-valued envelope signal y (t).
In discrete-time PA behavioral models, the computational complexity varies according to the sampling interval. Shorter sampling intervals demand larger computational
cost, while longer sampling intervals require reduced complexity. Indeed, in a PA
behavioral model that relates the real-valued RF signals, x(t) and y(t), the sampling
interval is constrained to be very short, on the order of the reciprocal of the carrier
frequency, to obey the Nyquist sampling criterion. In this way, the number of previous samples required to an accurate prediction of the long-term memory effects is
extremely high (on the order of a thousand past samples). For linearization purposes,
where computational cost is a bottleneck, a low-pass equivalent description, working just with the complex-valued envelopes signals, x(t)
P0
M0
M0
p=1 1 =0 2 =1
M0
p = p1
h p (1 , . . . , p )
p
i=1
x(n i )
(3)
where x(n) and y(n) are the input and output signals at the sample time n, respectively,
P0 is the polynomial order truncation, and M0 is the memory length. In (3), the
assumption that the Volterra kernels, h p (1 , . . . , p ), are symmetric functions of their
arguments was exploited. The kernel symmetry assumption does not imply any loss
of generality.
Assuming a polynomial approximation at the RF system level, (3) relates the RF
output signal y of (2) to the RF input signal x of (1). To estimate the output signal
at a single time instant, it is required the knowledge about the input signal at that
particular time instant and also over a period of time in past. Indeed, due to the presence of memory effects, each input applied at an arbitrary time instant is immediately
sensed by the output and continues to be sensed by the output at future time instants.
In particular, the effects of past input samples over the instantaneous output sample
become negligible only after the time interval between them is longer than the memory length of the slowest memory effect, which is usually called long-term memory
effect, attributed to the bias circuit and on the order of the reciprocal of the envelope
bandwidth.
Therefore, an accurate PA behavioral model at the RF level using (3) must formulate
the instantaneous sample of the output based on the instantaneous and previous (up to
the memory length M0 ) samples of the input. Hence, the number of previous samples
required to provide an accurate prediction of the long-term memory effects is extremely
high (on the order of a thousand past samples).
Following the procedure described in [2], the low-pass equivalent representation
of (3) that only generates physical in-band contributions, describing the relationship
between the complex-valued envelopes x(n)
M
M
P
p=1 1 =0 2 =1
M
M
M
h 2 p1 (1 , . . . , 2 p1 )
p
i=1
x(n
i )
2
p1
M
2 p1 =2 p2
x (n i )
(4)
i= p+1
low-pass equivalent models have lower computational complexities than models at the
RF system level, CV models still lead to a huge number of parameters, since the number
of parameters increases very fast with P and M, and M is usually large to include all
the PA memory effects. In particular, [9] reports an analytical expression relating the
number of parameters in (4) as a function of P and M.
It is well known that Volterra models with fewer parameters can be obtained if the
Volterra kernels are expanded into an infinite series of OBFs [18,27]. Concerning lowpass equivalent Volterra modeling, each kernel is approximated by a series truncated
in N2 p1 + 1 terms:
h 2 p1 (1 , . . . , 2 p1 )
N2 p1
N2 p1
k1 =0
ck1 ,...,k2 p1
2
p1
k2 p1 =0
2 p1,ki (i ).
(5)
i=1
The CV model structure then becomes an OBF Volterra model structure described as
yOBF (n) =
2 p1 N2 p1
P N
N2 p1
p=1 k1 =0 k2 =k1
ck1 , ,k2 p1
p
i=1
N2 p1
N2 p1
N2 p1
l2 p1,ki (n)
2
p1
l2p1,ki (n),
k2 p1 =k2 p2
(6)
i= p+1
where ck1 ,...,k2 p1 are the kernel expansion coefficients and l2 p1,k (n) is the output of
= x(n
+ 1), as
M
2 p1,ki (i )x(n
i ) = 2 p1,ki (q)x(n).
(7)
l2 p1,ki (n) =
i =0
From now on, we define p = 2 p 1 and the subscript associated with k in (5), (6),
and (7) is omitted to simplify the notation.
For each nonlinearity order ( p ) in (6), the set of orthonormal functions { p ,k (q)}
is completely parameterized by the set of poles {a p ,k }. In general, the poles {a p ,k }
can assume real or complex values. Considering that low-pass equivalent models
relate complex-valued envelope signals, complex poles are not constrained to appear
in conjugate pairs. Concerning OBF Volterra model approximations, the expansion
coefficients ck1 ,...,k p may converge faster to zero if the model basis functions present a
variety of poles and if these poles are selected in an optimal way. Therefore, when the
OBF Volterra models are compared to the CV models under the same modeling accuracy, low-pass equivalent models with a reduced number of parameters are obtained
by the OBF Volterra models.
3.1 Expansion Using TakenakaMalmquist Functions
The TakenakaMalmquist functions can be derived by imposing the orthonormality
property to the sequence of rational functions [6]
Model
CV
a p ,k = 0
LV
a p ,k = p
KV
a p ,k = p + j p
a p ,k = p ,k + j p ,k
TMV
p ,k (q) =
qd
; d = 0 or 1
q a p ,k
(8)
where a p ,k can be either real or complex such that |a p ,k | < 1. This procedure is
essentially a GramSchmidt orthonormalization construction and leads to the so-called
TakenakaMalmquist functions [6,16]
p ,k (q) = q
1 |a p ,k |2 k1
1 a p ,r q
q a p ,k
r =0
q a p ,r
; k = 0, . . . , N p
(9)
It can be observed that when the functions in (9) are applied in (6), the choice of
d determines if the resulting TMV model will be causal or strictly causal. The PA
instantaneous output y (n) is strongly dependent on the PA instantaneous input x(n),
and therefore, d = 1 represents the most appropriate choice for the basis functions.
The generalization of the low-pass equivalent Volterra models based on the
TakenakaMalmquist functions, for d = 1, is presented in Table 1. As shown in Table
1, the simplest choice for the dynamics of { p ,k (q)} corresponds to {a p ,k } = 0,
which implies
(10)
p ,k (q) = q k k = 0, , N p
k). In addition, if N p = M for p = 1, 3, . . . , P0 (or, equivand l p ,k (n) = x(n
alently, for p = 1, . . . , P), then the model structure of (6) is equivalent to the CV
model structure in (4). Such CV models are widely applied in PA models even if
prior information about the system dynamics cannot be introduced in the model basis
functions.
In contrast, a priori knowledge about the system dynamics can be incorporated
into a LaguerreVolterra (LV) model. In LV models, the so-called Laguerre functions
given by
q 1 2p
1 q k
p
; k = 0, . . . , N p
(11)
p ,k (q) =
q p
q p
are used. In this case, each set of orthonormal functions { p ,k (q)} in (9) is parameterized by one dynamic which assumes a real value p . In [28], the authors proposed
simplified LV models for the low-pass equivalent PA behavioral modeling, assuming
that all the model functions are parameterized by a single real dynamic { p } = .
In the same context, models with one complex pole parameterizing each set of
TakenakaMalmquist functions { p ,k (q)} were proposed in [8]. The authors referred
to these models as KautzVolterra (KV) models. We will also use this terminology,
although we acknowledge that (two-parameter) Kautz functions are differently defined
in [6,16,26].
Table 1 also shows that the most general choice for the dynamics of { p ,k (q)} leads
to the TMV models, where the poles for each nonlinearity order of the OBF Volterra
model can assume different (real or complex) values. In these models, it is possible to
incorporate a wide variety of system dynamics. For example, six different dynamics
can be incorporated into a third-order (P0 = 3) TMV model with N1 = N3 = 2,
whereas the LV and KV models with the same orders would have at most two different
dynamics introduced.
(12)
(13)
and ()T denotes the transpose. To calculate each state vector, firstly let us assume that
TakenakaMalmquist functions in (9) are denoted by p ,k (q) for d = 0; then one
can define the p th state equation as
l p (n + 1) = A p l p (n) + B p x(n),
(14)
where the pair of matrices A p C(N p +1N p +1) and B p C(N p +11) are completely defined by the set of poles {a p ,k } and constructed in such a way that the states
of (14) are given by
T
l p (n) = p ,0 (q) p ,N p (q) x(n)
T
= l p ,0 (n) l p ,N p (n) .
(15)
This approach allows the shifted function output signals l p ,k (n) = q d p ,k (q)x(n)
(16)
,0 z
Yp ,0 (z) =
p ,1 (z) 1 a
U
p
z ap
,0
,1 z
Yp ,1 (z) =
p ,2 (z)
U
,1
1 ap
,Np
z ap
z Yp
,Np
(z)
,Np
(17)
Observe that, since strictly causal Volterra models are more appropriate for the RF PA
behavioral modeling, we can use (17) rather than (14) to calculate the state vectors
l1 (n), l3 (n), . . . , l P0 (n) throughout a set of input samples x(n),
n = 0, . . . , K . As
suggested by [26], to get rid of the initial condition effects, it is possible to take
l p (0) = 0 and consider each state vector l p (n) only for n n 0 (instead of n 0).
In (17), A p and B p for each nonlinearity order p can be obtained from a cascade
connection of N p + 1 first-order all-pass filters as presented in Fig. 1. Each filter of
the cascade connection has a state-space realization defined in the Z -domain as
z L p ,k (z) = A p ,k L p ,k (z) + B p ,k U p ,k (z),
Y p ,k (z) = C p ,k L p ,k (z) + D p ,k U p ,k (z),
(18)
(19)
where
X (z) = U p ,0 (z) and L p ,k (z) are the Z -transforms of x(n)
and l p ,k (n), respectively. In this approach, the matrices A p ,k , B p ,k , C p ,k and D p ,k are defined as
B
A
C p ,k D p ,k
p ,k
p ,k
a p ,k
=
1 |a p ,k |2
1 |a p ,k |2
a p ,k
(20)
so that the transfer function from the input U p ,k (z) to the state L p ,k (z) is given by
L p ,k (z)
= (z I A p ,k )1 B p ,k =
U p ,k (z)
1 |a p ,k |2
z a p ,k
(21)
(22)
Hence, matrices A p and B p are obtained from the overall cascade construction
shown in Fig. 1, according to:
A p ,0
0
,1 C p ,0
,1
A
B
p
p
B p ,2 D p ,1 C p ,0
B p ,2 C p ,1
A p =
B p ,3 D p ,2 D p ,1 C p ,0
B p ,3 D p ,2 C p ,1
.
..
..
.
B p ,N p D p ,N p 1 D p ,1 C p ,0 B p ,N p D p ,N p 1 D p ,2 C p ,1
B p ,0
B p ,1 D p ,0
B p ,2 D p ,1 D p ,0
B p =
B p ,3 D p ,2 D p ,1 D p ,0
..
B p ,N p D p ,N p 1 D p ,0
0
..
..
.
.
A p ,N p
(23)
4 Model Extraction
Volterra models based on the OBF expansion are linear in the expansion coefficients.
Hence, standard linear algorithms such as the least squares can be used to identify
the TMV model parameters throughout a set of data samples [x(n),
where the vector (n) contains all products l p ,k (n)l p ,k (n) l p ,k (n) from (6) and
is a parameter vector containing all the expansion coefficients ck1 , ,k p so that
yOBF (n) = (n)T . It can be observed that the product terms in (n) for n = 0, . . . , K
can be easily evaluated using the state-space realizations proposed in Sect. 3.
The estimate which minimizes (24) with respect to , i.e.,
= arg min J ( )
(25)
(26)
(27)
T
y = y (0) y (1) y (K ) .
(28)
Unfortunately, in many cases prior information about the system dynamics is not
available. In such cases, it is not easy to select the poles of (9) since the low-pass
equivalent OBF Volterra models of (6) are nonlinear in their basis function dynamics.
To overcome this issue, we propose a redefinition of the objective function J as
follows1
K
2
,
J (a) =
(29)
y (n) (n, a)T (a)
n=0
where the parameter vector a contains all the basis function poles from the model. It
should be observed that the objective function now depends only on the vector a since
in (24) can be replaced by its estimate:
1
(a)
= X(a)H X(a)
X(a)H y .
(30)
As aforementioned, this is a nonlinear constrained optimization problem on the variable a, which can be described by
a = arg min J (a),
a
(31)
5 Experimental Validation
In this section, the proposed TMV model is applied to the behavioral modeling of a
device under test (DUT). The DUT is a GaN HEMT class AB RF PA. This RF PA is
excited by a carrier signal at 900 MHz modulated by a 3GPP WCDMA signal which
has a 3.84 MHz bandwidth. The input and output data are obtained using a Rohde and
Schwartz FSQ vector signal analyzer (VSA). The VSA sampling frequency is 30.72
MHz. The output and input measurements are first normalized. The set of 5420 input
output measurements is then divided into two subsets: one for modeling identification
(having 3320 samples) and one for modeling validation (having 2100 samples). In this
section, the results are reported in terms of the validation data. The peak-to-average
1 In [5], a different approach based on a gradient descent procedure is proposed to search for the basis
function poles. Such algorithm is fed with an approximation of the gradient descent cost function.
1st
3r d
5th
0.133
0.442
0.382
0.133
0.442
0.382
KV model
0.105 + j0.109
0.443 + j0.009
0.376 + j0.012
0.105 + j0.109
0.443 + j0.009
0.376 + j0.012
TMV model
0.166 j0.181
0.499 + j0.161
0.678 j0.013
+0.699 + j0.147
0.581 j0.138
0.473 j0.037
LV model
power ratios (PAPRs) of the input signals used for the modeling identification and
validation are equal to 5.0 and 4.7 dB, respectively.
The modeling accuracy of the proposed TMV, CV, LV, and KV models are carefully
compared under the same computational complexity. For a fair comparison among the
different models, the polynomial order truncation and the number of basis functions
for each nonlinearity order were set to the same values for all the studied models.
In particular, all the models are fifth-order Volterra models (P0 = 5) with two basis
functions for each nonlinearity order (N1 = N3 = N5 = 1). There is no theoretical
justification for these choices. These specific values were empirically chosen because
they provide an excellent trade-off between modeling error and number of parameters. Table 2 shows the pole sets for the first, third, and fifth nonlinearity orders used
by each model. These sets are obtained using the constrained nonlinear optimization approach described in Sect. 4, with initial pole sets taken at the origin, that is,
{a p0 ,k } = 0, p0 , k, since it is considered that there is no prior information about the
system dynamics. Here, we have used the interior point algorithm [3] included in the
MATLAB toolbox.
The modeling accuracy is evaluated based on error signals containing the difference
between the desired and estimated outputs. Specifically, two metrics are computed:
the normalized mean-square error (NMSE) and the adjacent channel error power ratio
(ACEPR), according to their definitions reported in [9].
The performances of the models in terms of NMSE and ACEPR are reported in
Tables 3 and 4, respectively. The ACEPR computations have employed bandwidths of
3.84 MHz for the main and adjacent channels, as well as a 5-MHz separation between
the main and adjacent channels. It is clear that great improvements in NMSE and
ACEPRs (for the lower and upper adjacent channels) occur when the OBF Volterra
models are used. While the accuracies of the LV and KV models are very close to
each other (in fact, their pole sets shown in Table 2 are also very close), a clear and
significant reduction in modeling error is provided by the TMV model. For instance,
reductions of up to 1.6 dB in NMSE and ACEPR values are observed if the TMV model
is used instead of the LV or KV models. In fact, by considering two different complex
poles for each nonlinearity order, considerable improvements in NMSE and ACEPR
metrics are achieved by the TMV model in comparison with previous approaches
Model
NMSE (dB)
CV
27.7
LV
36.2
KV
36.2
TMV
37.8
Model
CV
36.2
35.4
LV
45.2
43.7
KV
45.2
43.8
TMV
46.8
45.1
30
measured out
40
error for CV
error for LV
50
error for KV
60
70
80
90
100
110
885
890
895
900
905
910
915
Frequency (MHz)
Fig. 2 PSDs of the measured output signal and the error signals obtained from the different models
having a unique pole (either real or complex) for each nonlinearity. An illustration of
the superior modeling accuracy of the TMV model is provided by Fig. 2, which shows
the power spectral densities (PSDs) of the error signals for the different models. The
PSD of the measured (and normalized) output signal is also included in Fig. 2 by two
reasons: to emphasize the bandwidth and the center frequency of the main channel,
and to illustrate how the error power levels are small in comparison with the output
power levels. Observe that the PSD of the TMV error signal is clearly lower than the
PSDs of the other error signals, especially at the main and adjacent channels.
Up to this point, attention was paid to confirm, in this case study, the better performance of the TMV model with respect to the CV, LV, and KV models. From now
on, focus is turned on investigating how close the TMV estimations are from the measurements. Therefore, in Figs. 3, 4, and 5, only measured and TMV-estimated data
are included. Figure 3 shows the instantaneous (normalized) amplitude of the output
measured
TMV model
0.8
Normalized |V
out
| (V)
0.6
0.4
0.2
0.2
0.4
0.6
0.8
1.2
1.4
measured
TMV model
0.15
0.1
0.05
0
0.05
0.1
0.15
0.2
0.4
0.6
0.8
1.2
1.4
Normalized |V | (V)
in
Real part
0.5
1.5
2.5
Imaginary part
time (us)
1
0
measured
TMV model
1
0.5
1.5
2.5
time (us)
Fig. 5 Top) real and bottom) imaginary components of the output envelope as a function of time: measured
and estimated by TMV
Table 5 NMSE performances of the different models when applied to the modeling of the PA inverse
characteristic
Model
NMSE (dB)
for P0 = 3
NMSE (dB)
for P0 = 5
NMSE (dB)
for P0 = 7
NMSE (dB)
for P0 = 9
CV
25.1
26.3
26.8
26.9
LV
27.2
31.0
32.5
33.7
KV
27.9
31.5
32.9
34.7
TMV
29.4
32.4
35.1
36.1
Table 6 ACEPR performances of the different models when applied to the modeling of the PA inverse
characteristic
Model
Worst ACEPR
(dB) for P0 = 3
Worst ACEPR
(dB) for P0 = 5
Worst ACEPR
(dB) for P0 = 7
Worst ACEPR
(dB) for P0 = 9
CV
32.5
34.6
34.7
35.2
LV
32.7
38.2
39.4
41.4
KV
33.4
38.6
40.3
41.9
TMV
36.3
40.8
43.6
44.1
30
measured out
40
error for CV
error for LV
50
error for KV
error for TMV
60
70
80
90
100
885
890
895
900
905
910
915
Frequency (MHz)
Fig. 6 PSDs of the measured output signal and the error signals obtained from the different models, when
applied to the modeling of the PA inverse characteristic
in Sect. 4 is employed to obtain the pole sets. The initial poles for the LV model are all
set to 0.15, and the initial poles for the KV and TMV models are all set to 0.15+ j0.15.
The modeling accuracy of the TMV, CV, LV, and KV models are compared under the
same computational complexity. Tables 5 and 6 report the NMSE and ACEPR results,
respectively. In Table 6, only the worst ACEPR result (between the ACEPR values
for the lower and upper adjacent channels) is reported. Again, the TMV model shows
a higher accuracy than the LV, KV, and CV under the same number of parameters.
In particular, for P0 = 7, the TMV improves the NMSE and ACEPR results by 2.2
and 3.3 dB, respectively, in comparison with the KV model. Furthermore, the superior
modeling accuracy of the TMV model is observed by the PSDs of the error signals
shown in Fig. 6.
6 Conclusion
This paper presents how the RF PA low-pass equivalent behavioral modeling can be
performed using Volterra models based on the TakenakaMalmquist function expansion (TMV models). This new approach generalizes classical and OBF Volterra models
previously used in this context since TakenakaMalmquist function sets enable each
pole to have its unique value (either real-valued, complex-valued or even zero). The
use of state-space realizations for the TMV model representation is highlighted. A
new approach for selecting the model basis function poles using a constrained nonlinear optimization is also proposed. Based on measurements taken on a GaN HEMT
class AB RF PA, it is possible to observe that TMV models with multiple complex
dynamics can provide better approximations than CV, LV, and KV models under the
same computational complexity, quantified by improvements of 1.6 dB in NMSE and
of 1.6 and 1.3 dB in upper and lower ACEPRs, respectively. Moreover, when applied
to the modeling of the PA inverse transfer characteristic, the proposed TMV proved
References
1. M. Bahoura, FPGA implementation of high-speed neural network for power amplifier behavioral
modeling. Analog Integr. Circuits Signal Process. 79(3), 507527 (2014)
2. S. Benedetto, E. Biglieri, R. Daffara, Modeling and performance evaluation of nonlinear satellite links
a Volterra series approach. IEEE Trans. Aerosp. Electron. Syst. 15(4), 494507 (1979)
3. R.H. Byrd, M.E. Hribar, J. Nocedal, An interior point algorithm for large-scale nonlinear programming.
SIAM J. Optim. 9(4), 877900 (1999)
4. S. Cripps, RF Power Amplifiers for Wireless Communications, 2nd edn. (Artech House, Norwood, MA,
2006)
5. R. Hacioglu, G.A. Williamson, Reduced complexity Volterra models for nonlinear system identification. EURASIP J. Appl. Signal Process. 4, 257265 (2001)
6. P.S.C. Heuberger, P.M.J. Van den Hof, B. Wahlberg, Modeling and Identification with Rational Orthogonal Basis Functions (Springer, London, 2005)
7. M. Isaksson, D. Rnnow, A parameter-reduced volterra model for dynamic RF power amplifier modeling based on orthonormal basis functions. Int. J. RF Microw. Comput.-Aided Eng. 17(6), 542551
(2007)
8. M. Isaksson, D. Rnnow, A KautzVolterra behavioral model for RF power amplifiers, in IEEE MTT-S
International Microwave Symposium Digest, (San Francisco, CA, 2006), pp. 485488
9. M. Isaksson, D. Wisell, D. Rnnow, A comparative analysis of behavioral models for RF power
amplifiers. IEEE Trans. Microw. Theory Tech. 54(1), 348359 (2006)
10. H. Jiang, P.A. Wilford, Digital Predistortion for Power Amplifiers Using Separable Functions. IEEE
Trans. Signal Process. 58(8), 41214130 (2010)
11. P.B. Kenington, High Linearity RF Amplifier Design (Artech House, Norwood, 2000)
12. B.M. Lee, R.J.P. de Figueiredo, Adaptive predistorters for linearization of high-power amplifiers in
OFDM wireless communications. Circuits Syst. Signal Process. 25(1), 5980 (2006)
13. V. Mathews, G. Sicuranza, Polynomial Signal Processing (Wiley, New York, 2000)
14. S.D. Mitchell, G.H.C. Oliveira, Analysing a power transformers internal response to system transients
using a hybrid modelling methodology. Int. J. Electr. Power Energy Syst. 69, 6775 (2015)
15. F. Mkadem, M.C. Fares, S. Boumaiza, J. Wood, Complexity-reduced Volterra series model for power
amplifier digital predistortion. Analog Integr. Circuits Signal Process. 79(2), 331343 (2014)
16. B. Ninness, F. Gustafsson, A unifying construction of orthonormal bases for system identification.
IEEE Trans. Autom. Control 42(4), 515521 (1997)
17. G.H.C. Oliveira, W.C. Amaral, K. Latawiec, An introduction to models based on Laguerre, Kautz and
other related orthonormal functionsPart II: non-linear models. Int. J. Model. Identif. Control 16,
114 (2012)
18. G.H.C. Oliveira, W.C. Amaral, K. Latawiec, CRHPC using Volterra models and orthonormal basis
functions: an application to CSTR plants, in Proc. IEEE Control Applications Conf. (Istanbul, 2003),
pp. 718723
19. B. zgl, J. Langer, J. Noguera, K. Vissers, Software-programmable digital pre-distortion on new
generation FPGAs. Analog Integr. Circuits Signal Process. 78(3), 573587 (2014)
20. W. Pan, Y. Liu, S. Shao, Y. Tang, A method to reduce sampling rate of the ADC in feedback channel
for wideband digital predistortion. Circuits Syst. Signal Process. 33(8), 26552665 (2014)
21. J.C. Pedro, S.A. Maas, A comparative overview of microwave and wireless power-amplifier behavioral
modeling approaches. IEEE Trans. Microw. Theory Tech. 53(4), 11501163 (2005)
22. M. Rawat, K. Rawat, F.M. Ghannouchi, S. Bhattacharjee, H. Leung, Generalized rational functions for
reduced-complexity behavioral modeling and digital predistortion of broadband wireless transmitters.
IEEE Trans. Instrum. Meas. 63(2), 485498 (2014)
23. D. Raychaudhuri, N.B. Mandayam, Frontiers of wireless and mobile communications. Proc. IEEE
100(4), 824840 (2012)
24. M. Schetzen, The Volterra and Wiener Theories of Nonlinear Systems (John Wiley, New York, 1980)