SCIENCE CHINA

Information Sciences
. RESEARCH PAPER . February 2019, Vol. 62 022304:1–022304:9
https://doi.org/10.1007/s11432-017-9420-y

An enhanced digital predistortion algorithm based

on polynomial model identification

Wanzhi MA1 , Xin QUAN1* , Bo ZHAO2 , Ying LIU1 , Wensheng PAN1 ,

Shihai SHAO1 & Youxi TANG1
1
National Key Laboratory of Science and Technology on Communication, University of Electronic Science and
Technology of China, Chengdu 631731, China;
2
Military Delegate Office in Shijiazhuang, Military Delegate Bureau of Equipment Development Department,
Shijiazhuang 050000, China

Received 24 November 2017/Revised 31 January 2018/Accepted 29 March 2018/Published online 26 December 2018

Abstract To improve the accuracy of the nonlinear distortion correction for the radio frequency (RF)
power amplifier (PA), it is necessary to precisely obtain the reverse function of the PA nonlinear model.
However, the direct inversion of the PA nonlinear model involves solving a high-order univariate polynomial,
which is difficult to apply in engineering. In this study, based on the envelope memory polynomial (EMP)
model, the high-order terms of the nonlinear model are approximated by their previously calculated values
through iterations and considered as known constants in the polynomial solution finding process, thereby
resulting in a significant reduction in computational complexity. Compared with the direct inversion method,
model of a 9th-order nonlinear, the proposed method reduces the calculation time in the coordinated rotation
digital computer (CORDIC) algorithm by at least 80%. The simulation results show that for a long-term
evolution (LTE) downlink signal, the results obtained by the proposed simplified method agree well with
those obtained by direct inversion method.
Keywords digital predistortion, power amplifier, nonlinear distortion, memory effects, radio frequency
Citation Ma W Z, Quan X, Zhao B, et al. An enhanced digital predistortion algorithm based on polynomial
model identification. Sci China Inf Sci, 2019, 62(2): 022304, https://doi.org/10.1007/s11432-017-9420-y

1 Introduction
In wireless communication systems, the power amplifier (PA) greatly affects the communication per-
formance [1]. When a broadband radio frequency (RF) signal with non-constant envelope and high
peak-to-average power ratio (PAPR), e.g., an orthogonal frequency division multiplexing (OFDM) sig-
nal, is fed to a nonlinear PA, intermodulation distortions and out-of-band leakages will be introduced
in the output, which increase the bit error rate (BER) of the communication system and interfere the
adjacent channels. To solve the above problem, the digital predistortion (DPD) method has been studied
to improve the linearity of the PA [2]. A transmitter that employs the DPD method can operate at the
saturation point and subsequently increase the efficiency of the PA.
Generally speaking, the DPD method has two structures, i.e., the indirect learning structure and the
direct learning structure [3]. The indirect learning DPD structure is used in [4], which does not require a
specific nonlinear PA model. However, when the indirect learning DPD method is adopted, noise will be
introduced in the output of the PA, which may make the adaptive algorithm to converge to an erroneous
value. Moreover, the predistorter and the post-predistorter are not interchangeable [3].
* Corresponding author (email: [email protected])

c Science China Press and Springer-Verlag GmbH Germany, part of Springer Nature 2018 info.scichina.com link.springer.com
 Ma W Z, et al. Sci China Inf Sci February 2019 Vol. 62 022304:2

Figure 1 The DPD structure based on polynomial model identification.

In the direct learning DPD structure, the PA is modeled first, and then the algebraic inverse solution
is calculated based on the model expression to obtain DPD function. The common nonlinear behavioral
models with memory include the Volterra series model, the memory polynomial (MP) model [5], the
envelope memory polynomial (EMP) model [6], the generalized memory polynomial (GMP) model [7],
the radial pruning Volterra (RPV) and simplified RPV (SRPV) models [8].
In [9], the MP is used to model the PA, and the DPD function is calculated by constructing a uni-
variate polynomial of high degree and finding its roots by QR factorization. However, the high-order QR
decomposition is very complex and is difficult to apply in engineering.
In this study, a simplified DPD function is derived by approximating the high-order terms of the EMP
model by their previously calculated values through iterations and considering them as known constants
during the polynomial solution finding process. For approximation, the simplified DPD function lead to a
slight decline in out-of-band spurious performance. Theoretical analysis and simulation results show that
the calculation error of the proposed DPD method is very close to that of the direct inversion method
in [9].
The remainder of this paper is organized as follows. Section 2 presents the system model and describes
the two core modules in detail. The algorithm derivation and simplification are discussed in Sections 3
and 4, respectively. The simulation results are presented in Section 5. And Section 6 summarizes the
paper.

2 System model

In the DPD method based on polynomial model identification, a nonlinear model f˜(·) is used to represent
a real PA f (·) as
y(n) = f˜(X, W ), (1)

where X = [x(n), x(n − 1), . . . , x(n − L)]T is a vector with L + 1 elements and x(n) denotes the input
of the PA at the nth time instant, and W is a vector composed of the PA parameters. After estimating
W , the input sequence X at the current time instant is calculated as

x(n) = g (y(n), W , |x(n)| , V ) , (2)

where |x(n)| is the modulus of x(n), V = [x(n − 1), . . . , x(n − L)]T , and y(n) is the output of the PA.
Then, the reverse function of the PA model can be calculated as f˜−1 (·) = g(·).
The DPD structure based on polynomial model identification is shown in Figure 1,
where u(n) is the input baseband signal of the DPD and x(n) is the input baseband signal of the PA,
which is also the output of the DPD. y(n) is the baseband feedback signal of the PA after the down-
conversion and filtering. The studied DPD structure contains the following two core modules: (1) PA
modeling and model identification; (2) calculation of the DPD function.
 Ma W Z, et al. Sci China Inf Sci February 2019 Vol. 62 022304:3

3 Derivation of the algorithm

3.1 PA modeling and parameter calculation

PA modeling has attracted considerable attention in recent years [10]. In the system-level behavior model
simulation, the PA can be considered as a black box in the absence of any prior knowledge of the PA.
Hence, the PA model can be designed to be relatively simple to ensure the rapid prediction of the model
performance.
Considering the nonlinearity and memory effects of the PA, the Volterra series model is always used
for PA modeling. The Volterra series model is a functional series model, and it is suitable for nonlinear
system modeling with fading memory.
The Volterra series model in the continuous time domain can be expressed as [11]
∞ Z
X +∞ ρ
Y ∞
X
y(t) = y0 (t) + hρ (τ1 , τ2 , . . . , τρ ) x(t − τi )dτi = yρ (t), (3)
ρ=1 −∞ i=1 ρ=0

where x(t) and y(t) represent the input and output signal of the PA, respectively, and
Z +∞ Z +∞ ρ
Y
yρ (t) = ··· hρ (τ1 , τ2 , · · · τρ ) x (t − τi ) dτi .
−∞ −∞ i=1

The equivalent baseband model of the Volterra series model in the discrete time domain can be expressed
as [12]
XX X
y(n) = ··· h2k+1 (l1 , l2 , . . . , l2k+1 )
k l1 l2k+1
k+1
Y 2k+1
Y
× x(n − li ) x∗ (n − li ), (4)
i=1 i=k+2

where x(n) and y(n) are the baseband input and output of the PA, respectively, and hk is the Volterra
kernel of order k. In practice, the original Volterra series model is generally intercepted with a limited
order and memory depth to represent the nonlinearity and memory effects of the target object.
Due to the high complexity of the classic Volterra series model, a number of simplified models have been
proposed, e.g., the MP, EMP, GMP, RPV and SRPV models [5–8]. To further simplify the derivation
of the reverse function of the PA model, the EMP model is adopted in this paper because of the low
complexity in the inversion process, which can be expressed as [6]

K X
X L
y(n) = wkl x(n)|x(n − l)|k−1 , K > 2, (5)
k=1 l=0

where K is the nonlinear order and L denotes the memory depth.

Letting x(n, k, l) = x(n)|x(n − l)|k−1 , Eq. (5) can be rewritten as

y(n) = W T X̃n , (6)

where

X̃n = [x(n, 1, 0), . . . , x(n, 1, L), . . . , x(n, K, L)]T , W = [w(1, 0), . . . , x(1, L), . . . , x(K, L)]T . (7)

After N sets of data are observed, the obtained data can be expressed as

Y = W T X̃, (8)
 Ma W Z, et al. Sci China Inf Sci February 2019 Vol. 62 022304:4

where Y = [y(0), y(1), . . . , y(N − 1)]T is an N -dimensional output vector and X̃ = [X̃0 , X̃1 , . . . , X̃N −1 ]T
is the input matrix with N rows and K(L + 1) columns. Then, the least squares solution of W can be
given by
ŴLS = (X̃ H X̃)−1 X̃ H Y , (9)
where X̃ H is the conjugate transpose of X̃. After ŴLS obtained, the modeling of the PA is complete,
and the parameters of the PA become known.

3.2 Calculation of the DPD function

In this paper, the exact reverse function of the PA is obtained according to the estimate of the PA model,
which is different from the adaptive algorithm in conventional direct and indirect learning structures. As
the PA model coefficients have been estimated ŴLS = [w̃0 , w̃1 , w̃2 , . . . , w̃K×L ]T , the polynomial model of
the PA can be expressed as
K L−1
X X
y(n) = w̃kl x(n)|x(n − l)|k−1 = x(n) [α(x) + β(x)] , (10)
k=1 l=0

where
K
X K L−1
X X
α(x) = w̃k0 |x(n)|k−1 , β(x) = w̃kl |x(n − l)|k−1 . (11)
k=1 k=1 l=1
From (10), x(n) can be rewritten as
  y(n)
x(n) = g y(n), ŴLS , |x(n)| , V = . (12)
α(x) + β(x)
For an ideal DPD process, the output of the PA y(n) is equal to the input of the DPD u(n) [10,13,14].
Therefore, the output of the DPD x(n) can be expressed by replacing y(n) with u(n) in (12) as
  u(n)
x(n) = g u(n), ŴLS , |x(n)| , V = , (13)
α(x) + β(x)
where g(·) denotes the DPD function.
It is observed from (13) that the value of x(n) is related to its modulus |x(n)|, its values at previous
time instants x(n − l) and the input baseband signal u(n). Hence, in [13, 14], |u(n)| is used instead of
|x(n)| and substituted into (13) to calculate x(n) first. Then, the obtained x(n) is substituted into (13)
to recalculate x(n), and this step is repeated to further improve the accuracy of the value of x(n). Herein,
differently, we attempt to calculate x(n) more precisely.
According to (13), we have
|u(n)|
|x(n)| = . (14)
|α(x) + β(x)|
Substituting (11) into (14), and regarding β(x) as a constant β (given β(x) is only related to x(n − l) for
l = 1, 2, . . . , L − 1), Eq. (14) becomes
|u(n)|
|x(n)| = P . (15)
K k−1 + β
k=1 wk0 |x(n)|

For k = 1, the coefficient w10 and the constant β in (15) can be combined, and |x(n)| can be rewritten
as
|u(n)|
|x(n)| = P , (16)
K k−1
k=1 ωk0 |x(n)|

where ω10 = w10 + β and ωk0 = wk0 (k > 1). Letting ωk0 = ak0 + jbk0 , the denominator in (16) can be
expressed as
K
X K
X
ωk0 |x(n)|k−1 = (ak0 |x(n)|k−1 + jbk0 |x(n)|k−1 )
k=1 k=1
 Ma W Z, et al. Sci China Inf Sci February 2019 Vol. 62 022304:5

K K
!
X X
k−1 k−1
= ak0 |x(n)| +j bk0 |x(n)|
k=1 k=1
v
u K !2 K
!2
u X X
= t ak0 |x(n)|k−1 + bk0 |x(n)|k−1 . (17)
k=1 k=1

By substituting (17) into (15), we could have

ϕ(|x(n)|) = C0 + C2 |x(n)|2 + · · · + C2K |x(n)|2K , (18)

2 P k−1 P k−1
where C0 = −|v(n)| and Ck = i=1 ai0 · a(k−i)0 + i=1 bi0 · b(k−i)0 . Denoting z = |x(n)|, we have

ϕ(z) = C0 + C2 z 2 + · · · + C2K−1 z 2K−1 + C2K z 2K . (19)

Note that ϕ(z) = 0 is at least a fourth-order equation for K > 2. Therefore, there is no universal
root-finding formula for the equation ϕ(z) = 0. To solve the above equation, numerical analysis, i.e.,
the iterative method by successive approximation, is typically adopted [15]. To further reduce the search
range for the root of the equation ϕ(z) = 0, the following conditions should be satisfied:
(a) 0 6 z 6 max {u(n), y(n)}.
(b) The deviation between z and |u(n)| is not greater than 10% [14].
Given the above two conditions, the novel Newton search algorithm in [16] is used to search for the
value of |x(n)|. Then, |x(n)| is substituted into (13) to calculate x(n), which is the output of the DPD
g(·).
Note that if conditions (a) and (b) are both satisfied and there are multiple real roots for the equation
ϕ(z) = 0, any one of them can be chosen as the initial value of the novel Newton search algorithm, which
will eventually converge to a stable state. If there is no real root that satisfies conditions (a) and (b), the
search result remains constant during the current iteration.

4 Simplification of the algorithm

When the novel Newton search algorithm in [16] is adopted to solve the univariate high-order equation
ϕ(z) = C0 + C2 z 2 + C3 z 3 + · · · + C2K−1 z 2K−1 + C2K z 2K = 0 (K > 2). The difference between the
obtained solution and the true value is approximately 5 × 10−15 , which is considerably smaller than that
for the traditional methods. Furthermore, the PA fitting error is typically much larger than 5 × 10−15 .
Therefore, the calculation error of the proposed DPD method can be neglected.
In Section 3, to solve the univariate high-order equation ϕ(z) = 0, the QR decomposition is used to
decompose the adjoint matrix G of the equation. For the nonlinear model with order Ka , the computa-
tional complexity and memory requirement of QR decomposition are O K 3 and O K 2 , respectively.

Even when the adjoint matrix G is a special matrix, e.g., G is a conjugate symmetric matrix, a triangu-
lar matrix, or a unitary matrix, the computational complexity and memory requirement introduced by
the QR decomposition are still O K 2 and O (K), respectively. Therefore, the derivation of the reverse

function of EMP model is simplified in this section.

Eq. (10) can be rewritten as follows:
K L−1
X X
y(n) = w̃kl x(n)|x(n − l)|k−1 = x(n) [αP (x) + αP +1 (x) + β(x)] , (20)
k=1 l=0

where
P
X
αP (x) = w̃k0 |x(n)|k−1 ,
k=1,k=odd
(21)
P
X K
X
αP +1 (x) = w̃k0 |x(n)|k−1 + w̃k0 |x(n)|k−1 .
k=2,k=even k=P +1
 Ma W Z, et al. Sci China Inf Sci February 2019 Vol. 62 022304:6

Since the differences between the frequency bands of even-order terms P k−1
P
k=2,k=even w̃k0 |x(n)| and that
of the original signal are multiples of the carrier frequency, the effects of the even-order terms are not
significant and can be replaced by P k−1
P
k=2,k=even w̃k0 |x(n − 1)| . Similarly, the nonlinear terms with
PK
orders higher than P can be approximated by k=P +1 w̃k0 |x(n − 1)|k−1 because the modulus of x(n) is
less than 1.
Let
XP XK
η(x) = w̃k0 |x(n − 1)|k−1 + w̃k0 |x(n − 1)|k−1 . (22)
k=2,k=even k=P +1

Eq. (20) can be rewritten as

y(n) = x(n) [αP (x) + η(x) + β(x)] . (23)
By replacing y(n) with u(n), x(n) can be expressed as

u(n)
x(n) = . (24)
αP (x) + η(x) + β(x)

The modulus of (24) is

u(n)
|x(n)| = . (25)
αP (x) + η(x) + β(x)
When calculating |x(n)|, all the terms that contain |x(n − q)| , q > 1 can be regarded as constant terms,
i.e., η(x) + β(x) is regarded as a constant term δ. Then, we have

|u(n)|
|x(n)| = P . (26)
P
k=1 wk0 |x(n)|k−1 + δ

Since the coefficient w10 can be combined with the constant δ for k = 1, Eq. (26) can be further simplified
as
|u(n)|
|x(n)| = P , (27)
P k−1
k=1 ωk0 |x(n)|

where ω10 = w10 + δ and ωk0 = wk0 , (k > 1). Noting that ωk0 = ak0 + jbk0 , the denominator of (27) can
be expanded as
P
X P
X
k−1
ωk0 |x(n)| = (ak0 |x(n)|k−1 + jbk0 |x(n)|k−1 )
k=1 k=1
P P
!
X X
k−1 k−1
= ak0 |x(n)| +j bk0 |x(n)|
k=1 k=1
v
u P !2 P
!2
u X X
= t ak0 |x(n)|k−1 + bk0 |x(n)|k−1 . (28)
k=1 k=1

Substituting (28) into (15), the equation of |x(n)| is rewritten as

ϕ(|x(n)|) = C0 + C2 |x(n)|2 + · · · + C2P |x(n)|2P , (29)

2 Pk−1 Pk−1 2
where C0 = −|u(n)| and Ck = i=1 ai0 · a(k−i)0 + i=1 bi0 · b(k−i)0 . Letting τ = |x(n)| , we have

ϕ(τ ) = C0 + C2 τ + · · · + C2P τ P . (30)

By comparing (19) and (30), it is observed that the complexity of solving |x(n)| is decreased since
Eq. (30) is an equation of order P rather than 2K. When solving the univariate high-order equation,
the QR decomposition method is adopted, which is the major calculation complexity for solving the
reverse function of the EMP model. Based on the Givens rotation method in [17], the numbers of
 Ma W Z, et al. Sci China Inf Sci February 2019 Vol. 62 022304:7

Table 1 Complexity analysis of QR decomposition algorithm

Number of multiplications Number of divisions CORDIC iteration times
1 3 1 5 3
2
P + 6P 2 − 2
P P 2
P + 32 P 2 − 2P

Table 2 Simulation parameters of the signal source

Source type Subcarrier number Signal bandwidth Signal power
LTE-A 5 100 MHz −10 dBm
Amplitude error

Samples

Figure 2 Estimation errors of the two methods.

multiplications and divisions and coordinated rotation digital computer (CORDIC) calculation times for
QR decomposition are shown in Table 1.
Taking the 9th-order nonlinear polynomial for example, the order of the polynomial is decreased to
P = 5 in (30) and the numbers of multiplications and CORDIC iterations are decreased by more than
82% and 83% when compared with (19), respectively.

5 Numerical and simulation results

In this section, the proposed DPD method is validated by simulations. The MP model in [4] is adopted
as the PA model, where the nonlinear order is K = 5 and the memory depth is L = 5. The detailed
parameter settings are as follows: b10 = 1.0513+0.0904j, b30 = −0.0542−0.2900j, b50 = −0.9657−0.7028j,
b11 = −0.0680 − 0.0023j, b31 = 0.2234 + 0.2317j, b51 = −0.2451 − 0.3735j, b12 = −0.0289 − 0.0054j,
b32 = −0.0621 − 0.0932j, and b52 = 0.1229 + 0.1508j.
A long-term evolution (LTE)-A downlink signal is used as the signal source whose main parameters
are shown in Table 2.
We first study the estimation error introduced by the method proposed in this paper to calculating
the DPD function x(n) = g(u(n), ŴLS , |x(n)| , V ). In the proposed method, a simplified univariate
polynomial is constructed based on the method in [9] and additional estimation error introduced by
the simplification is shown in Figure 2. As a comparison, the estimation error of the method in [9] is
also shown in Figure 2, where the output of the DPD function is precisely calculated by constructing
a univariate polynomial and the introduced estimation error is negligible when compared with the PA
modeling error. It is shown in Figure 2 that the estimation error of the proposed DPD method is fare
well with that of the precise method in [9].
In detail, the normalized mean squared errors (NMSE) of the DPD functions for the proposed DPD
method and the direct inversion method in [9] are listed in Table 3, where the PA modeling error is set
to −80 dB. As shown in Table 2, the estimation error of the proposed DPD method is very close to that
of the direct inversion method in [9] and the PA modeling error. Therefore, the additional estimation
 Ma W Z, et al. Sci China Inf Sci February 2019 Vol. 62 022304:8

Table 3 NMSE comparison of the DPD functions

DPD methods NMSE (dB)
PA modeling error −80.00
Error of the precise method in [13, 14] −72.70
Error of the precise method in [9] −80
Error of (30) in this paper with P = 7 −79.24
Error of (30) in this paper with P = 5 −78.12

Figure 3 Performance comparison of the DPD algorithms on inhibition of spectral regeneration in LTE-A system.

error of the proposed method introduced by approximating the terms in (19) is negligible.
Finally, the performance of the DPD method proposed in this paper on the inhibition of spectral
regeneration is shown in Figure 3. It is observed in Figure 3 that the proposed DPD method can
effectively suppress the nonlinear distortion introduced by the PA, and its performance is close to that
of the direct inversion algorithm in [9].
Meanwhile, with the 9th-order nonlinear equation, the conventional methods need to solve an 18th-
order equation, whereas the simplified method with P = 5 only requires solving a 5th-order equation
without decreasing calculation accuracy. As a result, the consumed time in the simulations for the pro-
posed simplified method is decreased significantly when compared with the conventional precise method.

6 Conclusion

In contrast to the DPD method in [9] based on polynomial model identification, which obtains the
exact form of the DPD function by solving the reverse function of PA polynomial model, the proposed
DPD method can greatly reduce the order of the univariate polynomial without an obvious performance
degradation. In detail, the reverse function expression based on the EMP model is derived first, and
a univariate high-order equation of the parameters in the reverse function is constructed. Second, the
closed-form of the DPD function is obtained by the above equation. Finally, the high-order univariate
polynomial is simplified by approximating the adjacent terms of the closed-form function. Because of the
effects of approximating terms on the DPD function, the increase in error of the DPD function calculated
 Ma W Z, et al. Sci China Inf Sci February 2019 Vol. 62 022304:9

by the proposed method is very small. Our research can provide a feasible design scheme for broadband
communication systems. The future work we plan is the extension of this work to cope with practical
RF impairments such as IQ imbalances and sampling jitters.

Acknowledgements This work was supported in part by National Natural Science Foundation of China (Grant Nos.
61771107, 61701075, 61771115, 61531009, 61471108), National Major Projects (Grant No. 2016ZX03001009), the Project
Funded by China Postdoctoral Science Foundation, and the Fundamental Research Funds for the Central Universities.

References
1 Guo Y C, Tang J L, Wu G, et al. Power allocation for massive MIMO: impact of power amplifier efficiency. Sci China
Inf Sci, 2016, 59: 022301
2 Guan N, Wu N, Wang H. Digital predistortion of wideband power amplifier with single undersampling ADC. IEEE
Microw Wirel Compon Lett, 2017, 27: 1016–1018
3 Ding L, Mujica F, Yang Z. Digital predistortion using direct learning with reduced bandwidth feedback. In: Proceedings
of 2013 IEEE MTT-S International Microwave Symposium Digest (MTT), Seattle, 2013. 1–3
4 Abdelaziz M, Anttila L, Kiayani A, et al. Decorrelation-based concurrent digital predistortion with a single feedback
path. IEEE Trans Microw Theory Tech, 2018, 66: 280–293
5 Braithwaite R N. Digital predistortion of an RF power amplifier using a reduced volterra series model with a memory
polynomial estimator. IEEE Trans Microw Theor Tech, 2017, 65: 3613–3623
6 Hammi O, Ghannouchi F M, Vassilakis B. A compact envelope-memory polynomial for rf transmitters modeling with
application to baseband and RF-digital predistortion. IEEE Microw Wirel Compon Lett, 2008, 18: 359–361
7 Mkadem F, Islam A, Boumaiza S. Multi-band complexity-reduced generalized-memory-polynomial power-amplifier
digital predistortion. IEEE Trans Microw Theor Tech, 2016, 64: 1763–1774
8 Kim Y H, Do Jo G, Oh J H, et al. An efficient simplified behavioral model for RF power amplifiers. In: Proceedings
of IEEE Topical Conference on Power Amplifiers for Wireless and Radio Applications, Phoenix, 2011. 65–68
9 Pan W, Liu Y, Tang Y. A predistortion algorithm based on accurately solving the reverse function of memory poly-
nomial model. IEEE Wirel Commun Lett, 2012, 1: 384–387
10 Liu Y, Pan W, Shao S, et al. A new digital predistortion for wideband power amplifiers with constrained feedback
bandwidth. IEEE Microw Wirel Compon Lett, 2013, 23: 683–685
11 Li L M, Billings S A. Estimation of generalized frequency response functions for quadratically and cubically nonlinear
systems. J Sound Vib, 2011, 330: 461–470
12 Safari N, Roste T, Fedorenko P, et al. An approximation of volterra series using delay envelopes, applied to digital
predistortion of RF power amplifiers with memory effects. IEEE Microw Wirel Compon Lett, 2008, 18: 115–117
13 Kim J, Konstantinou K. Digital predistortion of wideband signals based on power amplifier model with memory.
Electron Lett, 2001, 37: 1417–1418
14 Kim J, Konstantinou K. Method and apparatus for calculating the predistortion function from a power amplifier. US
Patent, 6498529, 2002-12-24
15 Szidarovszky F, Yakowitz S J. Principles and Procedures of Numerical Analysis. New York: Springer, 1978. 141–178
16 Lang M, Frenzel B C. Polynomial root finding. IEEE Signal Process Lett, 1994, 1: 141–143
17 Kim H S, Zhu W, Bhatia J, et al. A practical, hardware friendly MMSE detector for MIMO-OFDM-based systems.
Eurasip J Adv Sig Proc, 2008, 2008: 1–14