IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 69, NO.

5, MAY 2021 2753

Efficient Interpolation Method for Wireless

Communications and Signal Processing
Applications
Shu-Chen Lin , Student Member, IEEE, Kevin Chuang , Senior Member, IEEE,
Chun-Wei Chang , Member, IEEE, and Jau-Horng Chen , Senior Member, IEEE

Abstract— Interpolation involves the process of estimating

values at positions in between the adjacent samples, which has
very fundamental and applied implications in signal process-
ing and communication systems. In this work, a new method
of cubic spline interpolation (CSI) is proposed and derived
mathematically from the input samples respecting the time-
domain boundary conditions, and the corresponding piecewise
polynomials are optimized in the frequency domain to achieve
good spectral properties. A comparative analysis of several
prior arts, including arbitrary frequency response polynomial
(AFRP) and basis spline (B-spline) is also presented, giving a Fig. 1. Block diagram of a generic transmitter.
clear overview of the advantages associated with the proposed
technique in terms of implementation complexity, latency, and in-
band and out-of-band spectral performance. The measurement
results with three different radio frequency (RF) power amplifier [2], motion compensation [3], and rendering [4], [5], the 2-D
(PA) prototypes using an input stimulus comprised of different
instantaneous bandwidth (iBW) carrier aggregated 5G new radio interpolation method is applied to obtain interpolated sam-
(NR) waveform are demonstrated to highlight the improved per- ples or a continuous function from original discrete samples,
formance in the context of digital predistortion (DPD) correction. whereas in digital wireless communication system, the 1-D
Finally, the architecture of the proposed method is illustrated, interpolation is essential for DSP, including sample rate con-
which is suitable for real-time high-speed digital implementation. version (SRC) [6], signal synchronization [7], and interpolated
Index Terms— Cubic spline interpolation, filters, interpolation, data access in lookup table (LUT)-based digital predistortion
microwave amplifier linearization techniques, power amplifiers (DPD) [8]–[13]. In this work, we will demonstrate two exam-
(PAs). ples of how the proposed interpolation method can be used
I. I NTRODUCTION to improve the performance of modern wireless transceiver
systems: signal synchronization and DPD as highlighted in

T HE concept of interpolation is widely adopted in many

digital signal processing (DSP) applications. In image
processing, for operations such as scaling [1], edge detection
the system block diagram shown in Fig. 1.
The purpose of interpolation is to estimate interpolated
samples or a continuous function in between a set of existing
Manuscript received October 21, 2020; revised December 31, 2020 and samples. There are several intuitive implementations, such as
February 1, 2021; accepted February 2, 2021. Date of publication March 26, nearest neighbor and linear interpolation, which require fewer
2021; date of current version May 5, 2021. This work was supported computational resources at a cost of considerable approx-
in part by the Ministry of Science and Technology, Taiwan, under Grant
MOST 109-2221-E-002-166-MY3 and Grant MOST 106-2221-E-002-222- imation error. Accordingly, they are mostly applicable to
MY3. (Corresponding author: Shu-Chen Lin.) high oversampling ratio (OSR) signals. In this work, OSR is
Shu-Chen Lin was with NanoSemi Inc., Waltham, MA 02451 USA. defined as the ratio of data sampling rate to the instantaneous
He is now with the Department of Engineering Science and Ocean
Engineering, National Taiwan University, Taipei 10617, Taiwan (e-mail: bandwidth (iBW). The limitation of OSR is undesirable in
[email protected]). many use cases. As a result of that, various methods with
Kevin Chuang is with NanoSemi, which is now part of MaxLinear, Waltham, higher accuracy have been proposed [14]–[17]. Among the
MA 02451 USA (e-mail: [email protected]).
Chun-Wei Chang was with the Department of Engineering Science and proposed approaches, basis spline (B-spline) interpolation is
Ocean Engineering, National Taiwan University, Taipei 10617, Taiwan. He one such method, which can be implemented efficiently in
is now with NXP Semiconductors N. V., Chandler, AZ 85224 USA (e-mail: digital hardware systems for image processing applications
[email protected]).
Jau-Horng Chen is with the Department of Engineering Science and Ocean [16]–[19]. In this approach, a noncausal infinite impulse
Engineering, National Taiwan University, Taipei 10617, Taiwan (e-mail: response (IIR) is needed during the calculation. Since the
[email protected]). images are mostly represented in finite samples, noncausal
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TMTT.2021.3061563. IIR operation can be achieved easily in such applications. In
Digital Object Identifier 10.1109/TMTT.2021.3061563 modern wireless communication systems, the signal is usually
0018-9480 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Ningbo University. Downloaded on May 17,2021 at 03:25:09 UTC from IEEE Xplore. Restrictions apply.
2754 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 69, NO. 5, MAY 2021

represented by infinite samples, which makes backward cal-

culation in noncausal IIR filter not suitable. The most popular
way for achieving signals time synchronization in a digital
front-end system is by applying a low-pass filter (LPF) after
the expanding operation [20]. Another attractive alternative is
polynomial-based interpolating filters, and one of the meth-
ods is cubic spline interpolation (CSI). Compared to linear
interpolation, the third-order piecewise polynomials are used
in CSI to achieve a more accurate approximation [21]. A real-
time implementation of CSI with a fixed moving window
has been proposed in [22]. Another approach of polynomial- Fig. 2. Illustration of interpolating fractional sample.
based interpolating filter is the arbitrary frequency response
polynomial (AFRP) [23]–[26], where piecewise polynomials
are used to represent finite impulse response (FIR) that can A. Interpolation LPF
be optimized for desired frequency response through linear The upsampling of a signal x[n] by a ratio L can be denoted
programming techniques [27]. Since the linearity require- as x L [n] as given by
ment becomes more and more stringent in modern wireless  n
communication systems, DPD is needed in both handset x , n = 0, ±L . . .
x L [n] = L (3)
and base station devices [28]–[30]. Typically, time alignment 0, otherwise.
with integer-sample accuracy is not sufficient for low OSR
applications, so fractional time alignment is essential for DPD To remove unwanted aliasing components, it is essential to
applications [31], [32]. Accordingly, an efficient interpolation apply an LPF after the upsampling procedure. The digital LPF
technique can reduce the overall digital implementation cost. can be designed and implemented efficiently with an FIR filter,
A new concept of CSI is first introduced through simula- as shown in [14], [27], and [35].
tion in [33]. This article aims to explain the theory and its
applications in detail with regard to the characteristics of a B. Interpolating Polynomials
proposed polynomial-based method based on measurement
Another approach to interpolating filter is to use poly-
results. It also demonstrates how the proposed polynomial-
nomial functions to reconstruct a continuous function x̂(t)
based method can be implemented efficiently by the Farrow
from the discrete sequence x[n]. Because of the continuous
structure [34].
waveform reconstruction, interpolating polynomials can be
adopted for different interpolation precision without additional
II. I NTERPOLATION M ETHODS AND T HEORY design overhead, unlike the FIR filter design.
To restore a continuous waveform from the discrete
Consider a continuous time-varying signal x̂(t) that can be sequence x[n], piecewise polynomials qn are utilized for each
sampled into a discrete-time signal x[n] as sampling interval nTs ≤ t ≤ (n +1)Ts , as shown in Fig. 2. The
  generalized form of qn can be expressed as
n
x[n] = x̂ (1) ∞

fs qn (τ ) = c[n + m]ψ N (τ − mTs ) (4)
m=−∞
where n denotes the integer sample index from −∞ to ∞, fs
is the sampling rate, and the corresponding sampling period where c[n] is the coefficient function, ψ N (t) indicates the
is Ts = 1/ f s . The discrete sequence can be transformed to Nth-order synthesis polynomial function, and τ is the time
the frequency domain X[exp( j ω)] by the Fourier transform variable of each polynomial qn within a sampling period
operation F(x[n]), where ω is the angular frequency. Then, defined as τ = (t − nTs )/Ts . According to Fig. 2 and (4),
the Fourier transform of a time-shifted sequence x[n − n 0 ] the interpolated samples can be obtained by inserting the
can be expressed as desired fractional delay into each polynomial as
  x[n + α] = qn (α) (5)
F(x[n − n 0 ]) = e− j ωn 0 X e j w . (2)
where the fractional delay α is bounded by 0 ≤ α < 1.
As shown in (2), time delay is equivalent to phase shift in the 1) Cardinal B-Spline Interpolation: Cardinal B-spline inter-
frequency domain. Accordingly, the most straightforward way polation uses B-spline as the synthesis function ψ N in (4). The
to align two digital sequences with unknown relative delay is Nth-order B-spline function is defined in [36] as
to apply fast Fourier transform (FFT) and inverse fast Fourier
N +1

transform (IFFT) to compensate for the phase offset in the (−1)n L+1
frequency domain. However, it needs numerous computational B N (t) = Cn (t − nTs ) L u(t − nTs ) (6)
n=0
L!
resources and is rarely adopted for real-time applications.
Therefore, it is more practical to time-align signals using where u(t) is the step function and CnL+1 is a combination
interpolating filters [14]. defined as (L + 1)!/(L + 1 − n)!n!. To satisfy the condition of

Authorized licensed use limited to: Ningbo University. Downloaded on May 17,2021 at 03:25:09 UTC from IEEE Xplore. Restrictions apply.
LIN et al.: EFFICIENT INTERPOLATION METHOD FOR WIRELESS COMMUNICATIONS AND SIGNAL PROCESSING APPLICATIONS 2755

passing through all knot points, the coefficient function c[n]

can be obtained by digital convolution-inverse operation as
shown in [37]

 −1
c[n] = B N (nTs ) ∗ x[n] (7)

where the Z transform B N (z) of B-spline function can be

expressed as

N +1

B N (z) = B N (nTs )z −n . (8)
n=0

Fig. 3. Impulse response of linear interpolation, conventional four-point CSI,

The cardinal B-spline interpolation can approximate an ideal and proposed eight-point CSI.
sinc function with an order that approaches infinity [37].
However, the convolution-inverse operation in (7) cannot be
implemented by a causal IIR, which results in higher compu- where the second derivatives qn in (14a) and (14b) can be
tational complexity. solved in matrix form directly as given by
2) Cubic Spline Interpolation: Cubic spline interpolation ⎡ ⎤⎡  ⎤
solves the piecewise polynomials in time domain directly. 4 1 0 ··· 0 q2,C S I (0)
Rewriting the qn function in (4) as a classical third-order ⎢1 4 1 0 0⎥⎢ q3,C  ⎥
S I (0) ⎥
⎢ ⎥⎢
polynomial ⎢ .. . . . . . . ⎥⎢ .. ⎥
⎢ . . . . 0⎥⎢ ⎥
⎢ ⎥⎢  . ⎥
⎣0 0 1 4 1⎦⎣q ⎦
M−2,C S I (0)
qn,C S I (τ ) = c3 [n]τ 3 + c2 [n]τ 2 + c1 [n]τ + c0 [n] (9) 
0 · · · 0 1 4 q M−1,C S I (0)
⎡ ⎤
x[3] − 2x[2] + x[1]
where variables c3 [n], c2 [n], c1 [n], and c0 [n] in each monomial ⎢ x[4] − 2x[3] + x[2] ⎥
⎢ ⎥
of (9) are the derived coefficients of each order in (4). ⎢ .
.. ⎥
= 6⎢ ⎥. (15)
As shown in [21], CSI defines the following conditions to ⎢ ⎥
⎣x[M − 1] − 2x[M − 2] + x[M − 3]⎦
solve these four variables:
x[M] − 2x[M − 1] + x[M − 2]

qn,C S I (0) = x[n] , qn,C S I (1) = x[n + 1] (10) To solve (15), all samples need to be collected first, which is
 
qn,C S I (1) = qn+1,C S I (0) (11) not desirable for real-time implementation. Instead, a mov-

qn,C S I (1) = 
qn+1,C S I (0) (12) ing fixed-size window is employed in [22], which makes
the computational resources of a real-time implementation
reasonable. Although real-time CSI acquires fewer resources
where (10) is the approximation curve that passes through all for implementation, it still suffers from the nonnegligible
samples, and (11) and (12) are the continuous first derivative approximation error. The impulse responses and frequency
and second derivative, respectively. For a set of M samples, responses of a linear interpolation and four-point (M = 4)
there exists M − 1 polynomials, which is equivalent to 4M − 4 CSI are shown in Figs. 3 and 4, respectively. The inadequate
variables according to (9). Therefore, it needs two additional frequency filtering response shown in Fig. 4 would introduce
boundary conditions at the first and the last polynomials. unwanted aliasing error when the OSR is lower than a certain
Natural spline [21] is one such candidate, which defines zero value [14]. If the passband rejection requirement is 40 dBc,
edge second derivatives and can be expressed as the applicable operational bandwidth (BW) ratio of linear
interpolation and CSI is 0.18 f s and 0.36 f s , respectively,
 
S I (0) = 0 , q M−1,C S I (1) = 0.
q1,C (13) as shown in Fig. 4. Furthermore, de Carvalho and Hanson [22]
shows that the OSR limitation of CSI cannot be improved by
changing window size.
From the results of (9)–(13), each coefficient of the third-order
polynomial can be solved as 3) Arbitrary Frequency Response Polynomial: Vesma and
Saramaki [23], Ridha et al. [24], and Vesma [25] proposed
  a new approach that uses piecewise polynomial functions
qn+1,C S I (0) − qn,C S I (0) to represent impulse response. By setting the coefficient
c3 (n) = (14a)
6 function c[n] to x[n], (4) can be considered as an linear

qn,C S I (0) time-invariant (LTI) system with the corresponding impulse
c2 (n) = (14b)
2 response ψ N , consisted of piecewise Nth-order polyno-
c1 (n) = x[n + 1] − x[n] − c3 (n) − c2 (n) (14c) mials. In [25], the impulse response with finite sequence
c0 (n) = x[n] (14d) length, which continues M×Ts in the time domain, where

Authorized licensed use limited to: Ningbo University. Downloaded on May 17,2021 at 03:25:09 UTC from IEEE Xplore. Restrictions apply.
2756 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 69, NO. 5, MAY 2021

Fig. 5. Target frequency response with tolerance error δ(ω).

Fig. 4. Frequency response of linear interpolation and conventional four-point

CSI.

TABLE I
L IST OF Ak , Bk , Ck AND Dk

Fig. 6. Frequency response comparison of conventional CSI, cubic B-Spline

interpolation, AFRP, proposed CSI, and ideal sinc function.

proposed method can be defined as

M
−1

2

qn (0) = ak x[n + k]. (17)

k=−( M2 −1)
M corresponds to M-point interpolation, is defined as
Set the coefficients ak real and even as
2 −1
M
N
ψ N (t) = ck (n) f k (t − nTs ) ak = a−k (18)
n=− M k=0
⎧ 2
k where k is the coefficient index such that 1 ≤ k ≤ M/2 − 1.
⎪
⎨ 2t Substituting the second derivative (17) and (18) back to (14a)–
− 1 , 0 ≤ t < Ts
f k (t) = Ts (16) (14d), the third-order polynomial can be expressed as
⎪
⎩0, otherwise. M

2

Because of the symmetry property, the desired frequency qn (τ ) = x[n + k]Fk (τ ) (19)
response can be obtained by optimizing 4L coefficients in (16). k=−( M
2 −1)
The biggest advantage of AFRP is the design flexibility of a
where Fk (τ ) is a piecewise polynomial defined by
given filter specification. However, it needs larger complexity
compared to other methods. Fk (τ ) = Ak τ 3 + Bk τ 2 + Ck τ + Dk (20)

where Ak , Bk , Ck , and Dk are the coefficients for the third-

III. P ROPOSED I NTERPOLATION M ETHOD order, second-order, linear, and constant terms, respectively.
The derivation of a novel CSI architecture is demonstrated Table I summarizes how coefficients are obtained in order.
in this section, aiming to improve the frequency response From the results of (19) and (20) and Table I, the impulse
while maintaining the low-complexity and low-latency imple- response of the proposed method can be expressed as
mentation. Inherited from the conventional CSI, the proposed  
n (mod L)
method adopts the same continuous conditions from (14a)– h[n] = F Ln +1 1 − (21)
L
(14d). However, instead of solving the matrix defined in (15),
the proposed method uses a linear combination of neighboring where L indicates the upsampling ratio and number of points
samples with tunable coefficients ak . For an M-sample inter- {1, (L − 1)/L, (L − 2)/L, . . . , 1/L} inserted into each poly-
polation with an even integer M, the second derivative of the nomial function.

Authorized licensed use limited to: Ningbo University. Downloaded on May 17,2021 at 03:25:09 UTC from IEEE Xplore. Restrictions apply.
LIN et al.: EFFICIENT INTERPOLATION METHOD FOR WIRELESS COMMUNICATIONS AND SIGNAL PROCESSING APPLICATIONS 2757

Fig. 7. Implementation block diagram of AFRP in [23]–[26].

Fig. 8. Proposed CSI architecture.

After obtaining the impulse response of (21), the corre-

sponding frequency response can be derived as
2 −1
ML
 
H e j ω = h[0] + 2 h[m] cos mω. (22) Fig. 9. Block diagram of L-order B-spline interpolation.
m=1
Because of the symmetry property in (18), only the positive
response where m ≥ 0 is concerned. From (17)–(22), it is
shown that the corresponding frequency response H [exp( j ω)]
can be optimized by governing the coefficients between a0 and
a M/2−1 . Therefore, the spectral performance of the proposed
CSI method can be improved to achieve the desired charac-
teristics in the target frequency range. As shown in Fig. 5,
for a band-limited signal with a cutoff frequency of ω p ,
the frequency response G[exp( j ω)] of the ideal interpolator
has a unity gain in the passband −ω p ≤ ω ≤ ω p and infinite
decay in the stopbands ωs1 and ωs2 . The goal of optimization is
to approximate the frequency response H [exp( j ω)] of the pro- Fig. 10. Block diagram of the measurement setup using three different PA
designs consisted of (a) class-AB pHEMT, (b) class-AB LDMOS, and (c)
posed CSI in (22) to the ideal response G[exp( j ω)] respecting Doherty LDMOS power devices.
the target frequency range {ω∈} until the difference is within
the desired tolerance error δ. It can be expressed as IV. C OMPUTATIONAL C OMPLEXITY
   
W (ω)| Ge j ω − H e j ω | ≤ δ (23)
In this work, we propose to use an eight-point (M = 8) CSI
where W (ω) is the weighting function and  is the set of to provide sufficient flexibility for the frequency-domain opti-
target frequency ranges. The optimized coefficients in (17) can mization while maintaining low latency and low complexity.
be obtained by solving (23) with linear programming. The corresponding coefficients {a0 , . . . , a3 } are optimized to

Authorized licensed use limited to: Ningbo University. Downloaded on May 17,2021 at 03:25:09 UTC from IEEE Xplore. Restrictions apply.
2758 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 69, NO. 5, MAY 2021

TABLE II
P ERFORMANCE C OMPARISON OF S TATE - OF - THE -A RT I NTERPOLATION M ETHODS

be −4.9697, 3.1651, −0.7966, and 0.1185, respectively, and

its impulse response is shown in Fig. 3. Moreover, the cor-
responding frequency response can be calculated according
to (22). Fig. 6 shows the comparison between the proposed
CSI, conventional CSI, AFRP in [23], cardinal cubic B-
spline interpolation, and an ideal sinc interpolator that has a
rectangular frequency response. It can be observed that the
proposed CSI achieves a better in-band gain flatness and out-
of-band suppression compared to conventional CSI.
In general, the DSP implementation of interpolation known
as the Farrow structure [34] is widely adopted as an efficient
architecture. For the implementation of AFRP, the modified
Farrow structure is proposed in [23], which can be realized
efficiently, as shown in Fig. 7. Horner’s method highlighted in
Fig. 11. Measured output spectrum comparison of class-AB LDMOS PA
the bottom of Fig. 7 is used to solve the interpolated sample with different interpolation methods.
qn (α) with a given fractional index α. Inherently, the proposed
CSI can be implemented with a similar Farrow structure, CSI provides the lowest computational complexity over other
as shown in Fig. 8. Compared to the same window size M, techniques with operational BW greater than 0.4 f s , as summa-
the proposed method only requires M/2 coefficients, which is rized in Table II. The implementation resources are estimated
just one fourth of the AFRP, suggesting that fewer multipliers with 8× upsampling ratio (L = 8) and 16-bit data precision
and logic gates are needed. Moreover, the proposed method conditions. All computations shown in Table II are assumed to
only needs one multiplier for every additional M. be done sequentially, which means that the number of blocks
The procedure of cardinal B-spline interpolation can be is equal to the number of operations. As shown in [37, Fig. 2],
separated into prefiltering and postfiltering blocks as the postfilter B83 (z) can be simplified to
overall block diagram shown in Fig. 9. The prefiltering block  
1 0
calculates the coefficient function c[n] in (7) by convolution- B83 (z) = B ∗ B 0
∗ B 0
∗ B 0
(z) ∗ B13 (z). (25)
inverse operation. The postfiltering block involves c[n] and 83 8 8 8 8

B-spline function expanding, which can be expressed as the Each B80 (z) can be considered as a moving average filter,
following equations: and B13 (z) can be implemented by a three-tap FIR filter.
   The comparison of approximation error for each interpolation
c[n], k = Ln t
c L [k] = , B L (t) = B
N N
(24) method is also presented in Table II. It is worthwhile to note
0, otherwise L that an ideal interpolator exhibits 0 dB attenuation in-band and
where the convolution of two functions in (24) indicates the ∞ dB suppression at 0.5 f s edge. Unlike other prior methods
interpolated samples as defined in (7). It can either be com- as shown in Table II, the proposed CSI achieves sufficient
pleted by an FIR filter directly as shown in [18] or simplified accuracy with operational BW up to 0.58 f s while meeting a
to an L-times repeatedly moving average filter before a low- stopband rejection of 40 dBc.
cost B-spline filter as shown in [37]. On the other hand,
the transfer function of the prefiltering block is derived in V. M EASUREMENT R ESULTS
[37], which has been shown that the poles are reciprocal pairs To evaluate the performance of the proposed interpolation
{α, α − 1} for every order N. In other words, the prefiltering method, a wireless transmitter line-up at 2140-MHz radio
block cannot be implemented by a casual stable IIR, where frequency (RF) frequency is prepared. The system block
the necessary condition is every pole being inside the unit diagram is shown in Fig. 10 where the interpolation is
circle. It needs an additional noncausal IIR that requires utilized in the signal synchronization block. The baseband
backward calculation as shown in [38]. Otherwise, it can signal is loaded to a vector signal generator (VSG) in the
be approximated by an FIR filter, which needs a relatively TX path and downconverted to the baseband by a spectrum
larger amount of multiplications [26]. As a result, the proposed analyzer (SA) in the RX path with a sampling frequency

Authorized licensed use limited to: Ningbo University. Downloaded on May 17,2021 at 03:25:09 UTC from IEEE Xplore. Restrictions apply.
LIN et al.: EFFICIENT INTERPOLATION METHOD FOR WIRELESS COMMUNICATIONS AND SIGNAL PROCESSING APPLICATIONS 2759

TABLE III
P ERFORMANCE C OMPARISON U NDER D IFFERENT C ONDITIONS

Fig. 12. Measured in-band PSD error comparison of Doherty LDMOS PA

with different interpolation methods.

of 150 MHz. The measurement is performed on three different

power amplifiers (PAs). Two of them are class-AB PAs that are
designed with an NXP AFT20S15N LDMOS transistor and an
Avago ATF-511P8 enhancement-mode pHEMT, respectively.
The third prototype is a commercial Doherty PA designed with
the NXP A2I20H060NR1 LDMOS transistor. Two different
carrier configurations featuring signal aggregation of multiple
5G new radio (NR) 1024-QAM 20-MHz component carriers
are used in the measurement. The crest factor reduction (CFR)
is applied on the input signal to achieve a 9-dB peak-to-
average power ratio (PAPR). In order to assess the impact of
interpolation accuracy on the performance of linearized wire- Fig. 13. AM-AM and AM-PM characteristics of (a) class-AB pHEMT and
(b) class-AB LDMOS PA with DPD (red line) and without DPD (blue line)
less transmitter chain, an iterative-learning-control (ILC) DPD using 60-MHz carrier-aggregated 5G NR signal.
correction technique [39] is used in conjunction with different
interpolation approaches. Note that all the DSP functions and
techniques discussed in this article are modeled and evaluated and error vector magnitude (EVM), respectively. To further
in MATLAB. The measurement results are shown in Figs. 11 analyze in-band performance, the measured power spectral
and 12 and Table III. The AM-AM and AM-PM characteristics density (PSD) error is shown in Fig. 12. In this example,
of 60- and 80-MHz measurements with and without DPD are the measurement result demonstrates a low OSR (i.e., 80/150
shown in Figs. 13 and 14, respectively. Note that Doherty PA = 0.53 f s ). It can also be seen from Table III that the proposed
can achieve 37.5% drain efficiency at 9-dB back-off. From method achieves in-band PSD error as low as −46 dB at the
Table III, it can be seen that linear interpolation achieves edge of passband, whereas the conventional CSI and linear
the worst out-of-band and in-band performance, which can interpolation can only achieve −25 and −22 dB, respectively.
be demonstrated by adjacent channel leakage ratio (ACLR) Consequently, the subcarriers at the edge of passband could

Authorized licensed use limited to: Ningbo University. Downloaded on May 17,2021 at 03:25:09 UTC from IEEE Xplore. Restrictions apply.
2760 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 69, NO. 5, MAY 2021

R EFERENCES
[1] J. Shi and S. E. Reichenbach, “Image interpolation by two-dimensional
parametric cubic convolution,” IEEE Trans. Image Process., vol. 15,
no. 7, pp. 1857–1870, Jul. 2006.
[2] L. Romani, M. Rossini, and D. Schenone, “Edge detection meth-
ods based on RBF interpolation,” J. Comput. Appl. Math., vol. 349,
pp. 532–547, Mar. 2019.
[3] H. Lakshman, H. Schwarz, and T. Wiegand, “Generalized interpolation-
based fractional sample motion compensation,” IEEE Trans. Circuits
Syst. Video Technol., vol. 23, no. 3, pp. 455–466, Mar. 2013.
[4] A. Guetat, A. Ancel, S. Marchesin, and J.-M. Dischler, “Pre-integrated
volume rendering with non-linear gradient interpolation,” IEEE Trans.
Vis. Comput. Graphics, vol. 16, no. 6, pp. 1487–1494, Nov. 2010.
[5] S. Lee, G. Jounghyun Kim, and S. Choi, “Real-time depth-of-field
rendering using anisotropically filtered mipmap interpolation,” IEEE
Trans. Vis. Comput. Graphics, vol. 15, no. 3, pp. 453–464, May 2009.
[6] X. Huang, Y. J. Guo, and J. A. Zhang, “Sample rate conversion using
B-Spline interpolation for OFDM based software defined radios,” IEEE
Trans. Commun., vol. 60, no. 8, pp. 2113–2122, Aug. 2012.
[7] T.-B. Deng, “Discretization-free design of variable fractional-delay FIR
digital filters,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 48, no. 6,
pp. 637–644, Jun. 2001.
[8] S. Boumaiza, J. Li, M. Jaidane-Saidane, and F. M. Ghannouchi, “Adap-
tive digital/RF predistortion using a nonuniform LUT indexing function
with built-in dependence on the amplifier nonlinearity,” IEEE Trans.
Microw. Theory Techn., vol. 52, no. 12, pp. 2670–2677, Dec. 2004.
[9] K.-F. Liang, J.-H. Chen, and Y.-J.-E. Chen, “A quadratic-interpolated
LUT-based digital predistortion technique for cellular power amplifiers,”
IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 61, no. 3, pp. 133–137,
Mar. 2014.
[10] A. Molina, K. Rajamani, and K. Azadet, “Concurrent dual-band digital
predistortion using 2-D lookup tables with bilinear interpolation and
extrapolation: Direct least squares coefficient adaptation,” IEEE Trans.
Microw. Theory Techn., vol. 65, no. 4, pp. 1381–1393, Apr. 2017.
[11] A. Molina, K. Rajamani, and K. Azadet, “Digital predistortion using
lookup tables with linear interpolation and extrapolation: Direct least
Fig. 14. AM-AM and AM-PM characteristics of (a) class-AB LDMOS and squares coefficient adaptation,” IEEE Trans. Microw. Theory Techn.,
(b) Doherty LDMOS PA with DPD (red line) and without DPD (blue line) vol. 65, no. 3, pp. 980–987, Mar. 2017.
using 80-MHz carrier-aggregated 5G NR signal. [12] Q. A. Pham, D. Lopez-Bueno, T. Wang, G. Montoro, and P. L. Gilabert,
“Partial least squares identification of multi look-up table digital pre-
distorters for concurrent dual-band envelope tracking power amplifiers,”
be affected with the conventional CSI and linear interpolation IEEE Trans. Microw. Theory Techn., vol. 66, no. 12, pp. 5143–5150,
methods. Furthermore, according to the aforementioned results Dec. 2018.
shown in Figs. 4 and 6, the frequency response near Nyquist [13] J. Ren, “Digital predistorter for short-wave power amplifier with improv-
ing index accuracy of lookup table based on FPGA,” IEEE Access, vol. 7,
rate suffers greatly from attenuation and aliasing. Therefore, pp. 182881–182885, Dec. 2019.
any operation near the Nyquist frequency is inaccurate and [14] R. W. Schafer and L. R. Rabiner, “A digital signal processing approach
thus can be ignored. Even though the PSD error near the to interpolation,” Proc. IEEE, vol. 61, no. 6, pp. 692–702, Jun. 1973.
[15] H. Hou and H. Andrews, “Cubic splines for image interpolation and
Nyquist rate is progressively worse, the level of aliasing digital filtering,” IEEE Trans. Acoust., Speech, Signal Process., vol. 26,
error is still inadequate to damage the spectral improvement, no. 6, pp. 508–517, Dec. 1978.
as shown in Fig. 11. [16] M. Unser, A. Aldroubi, and M. Eden, “B-spline signal processing. I.
Theory,” IEEE Trans. Signal Process., vol. 41, no. 2, pp. 821–833,
Feb. 1993.
[17] M. Unser, A. Aldroubi, and M. Eden, “B-spline signal processing.
VI. C ONCLUSION II. Efficiency design and applications,” IEEE Trans. Signal Process.,
vol. 41, no. 2, pp. 834–848, Feb. 1993.
In this work, a new method of interpolation is derived [18] L. Kai-Yu, W. Wen-Dong, Z. Kai-Wen, L. Wen-Bo, and X. Gui-Li,
analytically and proposed for higher performance and ease of “The application of B-spline based interpolation in real-time image
enlarging processing,” in Proc. 2nd Int. Conf. Syst. Informat. (ICSAI),
real-time implementation. Compared to the conventional CSI, Shanghai, China, Nov. 2014, pp. 823–827.
the proposed method improves the frequency-domain proper- [19] S. Abbas, M. Irshad, and M. Z. Hussain, “Adaptive image interpolation
ties significantly while maintaining low complexity. A design technique based on cubic trigonometric B-spline representation,” IET
Image Process., vol. 12, no. 5, pp. 769–777, May 2018.
example of eight-point CSI is demonstrated and compared to [20] H. Keung Kwan and A. Jiang, “FIR, allpass, and IIR variable fractional
conventional polynomial-based interpolation techniques. The delay digital filter design,” IEEE Trans. Circuits Syst. I, Reg. Papers,
results indicate that the proposed CSI has adequate approxima- vol. 56, no. 9, pp. 2064–2074, Sep. 2009.
tion accuracy for signal synchronization application in modern [21] S. A. Dyer and J. S. Dyer, “Cubic-spline interpolation: Part 1,” IEEE
Instrum. Meas. Mag., vol. 4, no. 1, pp. 44–46, Mar. 2001.
wireless communication systems while having the lowest [22] J. M. de Carvalho and J. V. Hanson, “Real-time interpolation with cubic
computational complexity. The measurement of a low OSR splines and polyphase networks,” Can. Electr. Eng. J., vol. 11, no. 2,
application shows that the proposed synchronization method pp. 64–72, Apr. 1986.
[23] J. Vesma and T. Saramaki, “Interpolation filters with arbitrary frequency
can achieve the best in-band flatness and EVM performance response for all-digital receivers,” in Proc. IEEE Int. Symp. Circuits
compared to linear interpolation and conventional CSI. Syst., Atlanta, GA, USA, May 1996, pp. 568–571.

Authorized licensed use limited to: Ningbo University. Downloaded on May 17,2021 at 03:25:09 UTC from IEEE Xplore. Restrictions apply.
LIN et al.: EFFICIENT INTERPOLATION METHOD FOR WIRELESS COMMUNICATIONS AND SIGNAL PROCESSING APPLICATIONS 2761

[24] H. Ridha, J. Vesma, T. Saramaki, and M. Renfors, “Derivative approx- Kevin Chuang (Senior Member, IEEE) was born
imations for sampled signals based on polynomial interpolation,” in in Taipei, Taiwan, in 1985. He received the B.S.
Proc. 13th Int. Conf. Digit. Signal Process., Santorini, Greece, 1997, degree (Hons.) in electrical engineering from the
pp. 939–942. University of California at Santa Barbara (UCSB),
[25] J. Vesma, “A frequency-domain approach to polynomial-based interpo- Santa Barbara, CA, USA, in 2007, and the M.S. and
lation and the farrow structure,” IEEE Trans. Circuits Syst. II, Analog Ph.D. degrees in electrical and computer engineering
Digit. Signal Process., vol. 47, no. 3, pp. 206–209, Mar. 2000. from the Georgia Institute of Technology, Atlanta,
[26] J. Vesma and T. Saramaki, “Polynomial-based interpolation Filters— GA, USA, in 2009 and 2011, respectively.
Part I: Filter synthesis,” Circuits, Syst. Signal Process., vol. 26, no. 2, From 2011 to 2014, he was a Member of Technical
pp. 115–146, Apr. 2007. Staff with the MIT Lincoln Laboratory, Lexington,
[27] T. Baran, D. Wei, and A. V. Oppenheim, “Linear programming algo- MA, USA, where he developed RF and mixed-signal
rithms for sparse filter design,” IEEE Trans. Signal Process., vol. 58, ICs for high-linearity receivers and transmitters in multiple system-on-a-chip
no. 3, pp. 1605–1617, Mar. 2010. (SoC) design cycles. In 2014, he co-founded NanoSemi, Inc. Waltham, MA,
[28] Y. Ma, Y. Yamao, Y. Akaiwa, and C. Yu, “FPGA implementation USA, which was subsequently acquired by MaxLinear, Waltham, MA, USA,
of adaptive digital predistorter with fast convergence rate and low in 2020, where he was a Domain Leader in RF Systems and pioneered
complexity for multi-channel transmitters,” IEEE Trans. Microw. Theory the development of wideband linearization systems. From May 2020 to
Techn., vol. 61, no. 11, pp. 3961–3973, Nov. 2013. October 2020, he was a Lead RF Systems Engineer with Wafer, LLC, Beverly,
[29] Q. A. Pham, G. Montoro, D. Lopez-Bueno, and P. L. Gilabert, “Dynamic MA, USA, developing distributed electronically steerable array systems
selection and estimation of the digital predistorter parameters for power for satellite communications. Since 2020, he has been a Senior Principal
amplifier linearization,” IEEE Trans. Microw. Theory Techn., vol. 67, Systems Engineer with MaxLinear, working on the radio development for
no. 10, pp. 3996–4004, Oct. 2019. wireless infrastructure. His technical contributions are at the intersection of
[30] C. D. Presti, D. F. Kimball, and P. M. Asbeck, “Closed-loop digital wireless communications, RF electronics, and signal processing algorithms.
predistortion system with fast real-time adaptation applied to a handset His research interests are focused on the characterization and linearization of
WCDMA PA module,” IEEE Trans. Microw. Theory Techn., vol. 60, nonlinear communication systems, including RF power amplifiers.
no. 3, pp. 604–618, Mar. 2012. Dr. Chuang is a member of Tau Beta Pi, Eta Kappa Nu, and the National
[31] J. Kral, T. Gotthans, and M. Harvanek, “Analytical method of fractional Society of Collegiate Scholars. He is also the General Chairman of the
sample period synchronisation for digital predistortion systems,” in Proc. 2022 IEEE Radio and Wireless Week. He was a recipient of the 2007 Joseph
27th Int. Conf. Radioelektronika (RADIOELEKTRONIKA), Apr. 2017, J. Sayovitz Scholarship for outstanding achievement in electrical engineering
pp. 1–5. at UCSB. He serves on the Executive Committee of the IEEE Radio and
[32] S. Traverso and J.-Y. Bernier, “Low complexity time synchronization Wireless Week and the IEEE MTT-S Boston Section.
based on digital predistortion coefficients,” IEEE Microw. Wireless
Compon. Lett., vol. 29, no. 3, pp. 240–242, Mar. 2019.
[33] S.-C. Lin, K. Chuang, and J.-H. Chen, “Efficient implementation of
cubic spline interpolator,” in Proc. IEEE Radio Wireless Symp. (RWS),
San Antonio, TX, USA, Jan. 2020, pp. 287–290.
[34] C. W. Farrow, “A continuously variable digital delay element,” in Proc.
IEEE . Int. Symp. Circuits Syst., vol. 3, Espoo, Finland, Jun. 1988,
pp. 2641–2645.
Chun-Wei Chang (Member, IEEE) received the
[35] T.-B. Deng, S. Chivapreecha, and K. Dejhan, “Bi-minimax design of
M.S. and Ph.D. degrees in engineering science and
even-order variable fractional-delay FIR digital filters,” IEEE Trans.
ocean engineering from National Taiwan University,
Circuits Syst. I, Reg. Papers, vol. 59, no. 8, pp. 1766–1774, Aug. 2012.
Taipei, Taiwan, in 2011 and 2018, respectively.
[36] L. L. Schumaker, Spline Functions: Basic Theory. New York, NY, USA:
From 2011 to 2014, he was a Design Engineer
Wiley, 1981.
with Richwave Technology, Taipei, where he was
[37] M. Unser, A. Aldroubi, and M. Eden, “Fast B-spline transforms for
involved in designing RF components and CMOS
continuous image representation and interpolation,” IEEE Trans. Pattern
FM receivers. From 2015 to 2017, he was an
Anal. Mach. Intell., vol. 13, no. 3, pp. 277–285, Mar. 1991.
Intern Design Engineer with NXP Semiconductors
[38] J. T. Olkkonen and H. Olkkonen, “Fractional delay filter based on
N. V., Chandler, AZ, USA, where he was involved
the B-spline transform,” IEEE Signal Process. Lett., vol. 14, no. 2,
in designing high-power RF power amplifiers for
pp. 97–100, Feb. 2007.
4G/5G wireless infrastructure. In 2018, he joined NXP Semiconductors N.
[39] J. Chani-Cahuana, P. N. Landin, C. Fager, and T. Eriksson, “Iterative
V., where he is currently an RF/Wireless Development Engineer. He holds
learning control for RF power amplifier linearization,” IEEE Trans.
two issued U.S. and foreign patents. His current research interests include
Microw. Theory Techn., vol. 64, no. 9, pp. 2778–2789, Sep. 2016.
analog/RF IC design, advanced receiver/transmitter architectures, and power
amplifier (PA) linearization techniques for cellular networks.

Jau-Horng Chen (Senior Member, IEEE) received

the B.S. degree in electrical engineering from
Shu-Chen Lin (Student Member, IEEE) received National Taiwan University, Taipei, Taiwan, in 2001,
the B.S. degree in engineering science and and the M.S. and Ph.D. degrees in electrical and
ocean engineering from National Taiwan University, computer engineering from the Georgia Institute of
Taipei, Taiwan, in 2015, where he is currently pur- Technology, Atlanta, GA, USA, in 2002 and 2006,
suing the Ph.D. degree. respectively.
From 2015 to 2018, he was a Research Assistant From 2006 to 2008, he was a Design Engineer with
with the Department of Engineering Science and Freescale Semiconductor, Tempe, AZ, USA, where
Ocean Engineering, where he was involved in RF he was involved in designing dc–dc converters and
power amplifiers linearization for 4G/5G wireless digital predistortion linearizers for cell phone power
infrastructure. From 2019 to 2020, he was an Intern amplifiers (PAs). In 2008, he returned to National Taiwan University, where
Engineer with NanoSemi, Inc. Waltham, MA, USA. he is currently a Full Professor and the Associate Chair of the Department of
His current research interests include the use of signal processing strategies for Engineering Science and Ocean Engineering. He holds five U.S. patents. His
linearity improvement of RF power amplifiers and high-efficiency transmitters current research interests include analog/RF IC design, high-efficiency PAs,
in wireless communication systems. and field-programmable gate array (FPGA) prototyping.

Authorized licensed use limited to: Ningbo University. Downloaded on May 17,2021 at 03:25:09 UTC from IEEE Xplore. Restrictions apply.