IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 70, NO.

3, MARCH 2022 1509

Digital Predistortion of 5G Multiuser MIMO

Transmitters Using Low-Dimensional
Feature-Based Model Generation
Xiaoyu Wang , Graduate Student Member, IEEE, Yue Li , Member, IEEE,
Hang Yin , Graduate Student Member, IEEE, Chao Yu , Senior Member, IEEE,
Zhiqiang Yu , Member, IEEE, Wei Hong , Fellow, IEEE, and Anding Zhu , Senior Member, IEEE

Abstract— In this article, we present a novel digital predistor- behavior of power amplifiers (PAs) in MIMO arrays can lead
tion (DPD) system which can be updated quickly and efficiently to significant distortion that deteriorates the signal quality.
in response to the dynamic reconfiguration of multiuser multiple- Digital predistortion (DPD) has been widely recognized as
input multiple-output (MIMO) transmitters. By identifying the
shared properties of different power amplifiers (PAs) with two one of the most effective linearization techniques in modern
feature extraction stages, nonlinear behaviors of the PAs are communication systems [2], [3]. Compared with the conven-
encoded into low-dimensional features. Using the extracted fea- tional single-input single-output (SISO) transmitters (TXs),
tures as input, a novel DPD generator is employed to synthesize the MIMO TXs in 5G exhibit more complex nonlinear char-
DPD model coefficients directly. In this way, when the power acteristics [4]. Recently, many DPD techniques have been
levels or beam angles change, the DPD model can be updated
fast and accurately without capturing the output data or applying proposed to mitigate distortions generated by MIMO TXs.
linear system identification algorithms. To capture slow-varying Various DPD models have been proposed to suppress the
dynamics of the PAs, the features can also be calibrated efficiently nonlinear distortions and crosstalk induced by PAs in fully
in the background. Computational complexity and operational digital (FD) MIMO TXs [5]–[7]. In hybrid beamforming (HB)
latency can thus be reduced significantly. Simulation and exper- TXs, the linearization target may be set to the main beam
imental results demonstrate that the proposed DPD approach
achieves excellent linearization performance with low complexity, via over-the-air (OTA) test, rather than the output of each
making itself a promising linearization solution for 5G massive PA. The conventional SISO DPD models can thus still be
MIMO transmitters. adopted [8], [9]. To update the DPD model coefficients,
Index Terms— 5G, digital predistortion (DPD), feature extrac- different data acquisition strategies have been proposed. The
tion, model adaptation, multiuser, multiple-input multiple-output desired target signal for DPD can be captured using embedded
(MIMO), power amplifier (PA). antennas [10], [11], dedicated coupling structures [12] or
using OTA antennas [13], [14]. To avoid interruption of data
transmission, Wang et al. [15], [16] proposed the OTA-based
I. I NTRODUCTION
indirect identification technique. By monitoring system behav-

A T MILLIMETER-WAVE (mmWave) frequencies, thanks

to the reduced size of antenna elements, using multiple-
input multi-output (MIMO) technique with large-scale antenna
ior with a single OTA antenna, this method guarantees the
linearization performance while keeping hardware complexity
low.
arrays becomes feasible. Combining with beamforming tech- Unlike SISO systems, the configuration of 5G multi-
niques, highly focused beams can be generated to point to user (MU) MIMO systems can change rapidly, because the
specific user directions to increase communication quality beamforming weights need to track the movement of user
and save energy [1]. Nevertheless, the complex nonlinear equipment (UE) [17]. Thereby, the precoding weights and
power level of each PA in the array need to change accord-
Manuscript received July 23, 2021; revised September 19, 2021; accepted ingly [18]. It indicates that the nonlinear behavior of the PAs
October 19, 2021. Date of publication December 3, 2021; date of current
version March 4, 2022. This work was supported in part by the Science may frequently change during the operation, and DPD must
Foundation Ireland under Grant 17/NSFC/4850 and Grant 16/IA/4449 and track these variations and make corresponding adjustments in
in part by the National Natural Science Foundation of China (NSFC) under a timely manner.
Grant 61861136002. (Corresponding author: Xiaoyu Wang.)
Xiaoyu Wang, Yue Li, Hang Yin, and Anding Zhu are with the Some existing works have considered the challenges arising
School of Electrical and Electronic Engineering, University College Dublin, from beam steering. In [19], the DPD model is trained
Dublin 4, Ireland (e-mail: [email protected]; [email protected]; with data from multiple beam directions, but the perfor-
[email protected]; [email protected]).
Chao Yu, Zhiqiang Yu, and Wei Hong are with the State Key Laboratory mance is limited and its application to MU scenarios is
of Millimeter Waves, Southeast University, Nanjing 210096, China (e-mail: not considered. Other adaptation methods have proposed
[email protected]; [email protected]; [email protected]). interpolation-based solutions to quickly derive the DPD model
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TMTT.2021.3129777. coefficients under varying power levels [20] or beam direc-
Digital Object Identifier 10.1109/TMTT.2021.3129777 tions [21], [22], which can lead to high complexity in model
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
1510 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 70, NO. 3, MARCH 2022

multiuser transmitter architecture where the TX is composed

of multiple subarrays and each subarray supports one user.
This configuration is simple to understand and has a great
potential in millimeter wave applications. The same principle
can be extended to other transmitter architectures.
For simplicity, a uniform linear array (ULA) is considered
for each subarray and the ULA consists of N antenna elements
with equal spacing d and each succeeding RF chain has a
β progressive phase increase relative to the preceding one,
i.e., the phase shifts in RF chains are 0, β, 2β, . . . , (N − 1)β.
As shown in Fig. 1, in a MU HB TX, digital precoding is
applied to the signals for each UE before loading to the RF
chains, represented by
Fig. 1. Proposed feature based fast-adaptive DPD for MU MIMO
transmitters. X = SF (1)
where X = [x1 , x2 , . . . , xv , . . .], S = [s1 , s2 , . . . , sv , . . .].
extraction, considering the large number of PAs in the array. sv is the data of vth UE, and xv is the input data for
Ng et al. [23], Yin et al. [24], and Liu et al. [25] used mul- vth subarray. F is the digital precoding matrix. After digital
tiple sets of coefficients to cover the beam steering or power precoding, one DPD is assigned for each subarray. Analog
variation range. Nevertheless, the number of coefficients to be beamforming weights can be added by phase shifters in RF
identified can grow exponentially when more system parame- chains. Assuming the output of vth DPD is uv and the transfer
ters are involved. To this end, it is desirable to develop a fast- function for the nth PA in the subarray is G n , the output signals
and-efficient method to update DPD model for MU MIMO for nth PA can be represented by
systems with reasonable time and complexity constraints. yn = G n [uv e j (n−1)β ] = G n [uv ]e j (n−1)β . (2)
In this article, we present a novel model adaptation tech-
nique for MU MIMO DPD applications. Firstly, we present a In the HB TXs, the transmitted signals of all RF chains in
feature extraction technique for characterization of MU MIMO vth subarray are combined in the transmitted direction and can
arrays. Each PA in the array is represented by a core PA model, be represented by
and its behavior under different power levels is modeled by

Nv 
Nv
j (n−1)θv
separate PA features. Compared with conventional approaches, zv = yn e = G n [uv ]e j (n−1)β e j (n−1)θv (3)
the computational complexity for identifying the character- n=1 n=1
istics of the PAs is greatly reduced. Second, we propose a
novel DPD generator to synthesize the desired DPD model where Nv is the number of PAs in the vth subarray, e j (n−1)θv
coefficients directly from the PA features. It can reduce the represents the channel coefficients between UE and nth PA
DPD adaptation complexity by multiple orders of magnitude, in the subarray, and θv is the transmitted direction of the vth
because only low-dimensional matrix is involved using the subarray. For the main beam direction, the phase shifts in the
proposed approach. Finally, by integrating the array charac- transmitter compensate that in the channels, that is,
terization and DPD generator techniques, as shown in Fig. 1, e j (n−1)β e j (n−1)θv = 1. (4)
a complete dynamic DPD solution is developed for MU MIMO
systems, which can effectively update DPD models based on With the captured PA output data, the DPD model update
the operating conditions using the DPD generator and track the process can be formulated in the same way as conventional
slow-varying PA behavior by efficiently identifying the core SISO DPD scenarios. In the conventional methods, DPD
PA models in the background. model coefficients can be extracted using the indirect learning
The rest of the article is organized as follows. In Section II, architecture (ILA). If DPD model is linear in the parameters,
the challenges of MU MIMO DPD are discussed. Section III least squares (LS) can be adopted to solve the model coeffi-
introduces the new characterization method of MU MIMO TX cients
based on the feature extraction. Subsequently, the theory and  −1 H
cv = ZHv Zv Zv uv (5)
operation of the DPD generator are explained in Section IV.
Simulation results are given in Section V, followed by exper- where Zv is the DPD basis functions built by zv and cv is the
imental results in Section VI. Finally, a conclusion is drawn DPD model coefficients.
in Section VII.
B. Issues of MU MIMO DPD Adaptation
II. C HALLENGES IN MU MIMO DPD A DAPTATION
To realize beam steering while ensuring low interference
A. Signal Model of MU MIMO Systems among UEs, beamforming weights, including power and phase
In massive MIMO systems, different array configurations information, are changed according to the location of the
may be deployed. For proof of concept, in this article, UEs [26]. The operating conditions of PAs can thereby
we present the proposed DPD method based on the hybrid change rapidly due to varying power level and beam direction.
WANG et al.: DIGITAL PREDISTORTION OF 5G MULTIUSER MIMO TRANSMITTERS 1511

Fig. 3. Extraction of PA features for one PA model.

Fig. 2. Update of system configuration and DPD model when the supported
UE moves in conventional systems.
separate couplers and data acquisition receivers [15]. It is thus
desirable to characterize the entire array with a single data
As shown in Fig. 2, when the supported UE moves, channel acquisition receiver. One way to achieve it is to capture each
state information (CSI) of UEs is updated. The system needs PA one by one in a time-division manner. Another method is
to change the power and phase in the corresponding subarray to capture the combined feedback signals and then separate the
to steer the beam to the new location of the UE. Owing to components of each PA. In both cases, multiple measurements
the changed system configuration, characteristics of the PAs, are required. As the input signals will change during this
and the subarray as the whole, change as well. In this case, process, the captured signals cannot be directly combined to
new feedback information is needed to update DPD model form the target signal, i.e., far-field signal at UE. Therefore,
coefficients. it is necessary to build a forward model for each PA in the
In MU scenarios, since each user may have different power array and estimate the target signal by feeding a common input
and angle settings, the number of possible system configura- signal to the forward models.
tions can be very large, making it impossible to pre-compute Generally, we assume the PAs share the same model struc-
the DPD coefficients for all conditions in advance. This sug- ture but have independent coefficients. If the PA model is
gests that the DPD coefficients must be updated fast enough to linear in its parameters, it can be expressed in matrix format
track the varying operating condition. The new situation is very as follows:
different from the conventional systems where DPD adaptation
algorithms only need to cope with long-term variation, e.g., ynt = Uλnt (6)
temperature or device aging.
In the new application scenarios, the existing DPD adapta- where ynt is the nth PA output under tth power level consisting
tion schemes face two critical issues: latency and complexity. of N S data samples
The latency of the conventional adaptation schemes may not
be acceptable, because they need to capture enough data and ynt = [ ỹnt (N S ), ỹnt (N S − 1), . . .]T (7)
execute complex model extraction algorithms before they can
update the DPD models. The system may fail to meet the λnt is a vector including all PA model coefficients of nth PA
linearity requirement until the DPD update is finished. The operating under tth power level
issue will be worse with a large number of PAs in MIMO
arrays. Furthermore, because of the differences between PAs λnt = [λnt (1), λnt (2), . . . , λnt (PA )]T (8)
in the array, each PA needs to be characterized separately [27].
The latency from data acquisition and model extraction will
and U is the regression matrix containing all PA model basis
increase rapidly with the array size. The other serious con-
functions constructed with the PA input signal. For example,
cern is the computational complexity in model identification.
if the memory polynomials (MP) model [28] is used, the basis
To avoid interruption of data transmission, we need to estimate
functions have the form of |u(n − m)|k u(n − m), where k is
the PA model coefficients for each PA in the array [15], [16],
the nonlinearity order and m is the delay number, and U can
which can have high complexity when the array size is large.
be constructed as (9), shown at the bottom of the page.
Since the moving UEs can change the PA operating condi-
III. C HARACTERIZATION OF MU MIMO TX W ITH PA
tion, a separate model is needed for each operating condition.
F EATURE E XTRACTION
As a result, a large number of PA models are needed to cover
Considering the large number of PAs in 5G MIMO trans- all potential conditions, leading to very high-computational
mitters, it is impractical to capture the output of PAs using complexity to operate the system.

⎡ ⎤
u(N S ) u(N S − 1) ··· |u(N S )|u(N S ) ···
⎢ u(N S − 1) u(N S − 2) ··· |u(N S − 1)|u(N S − 1) ···⎥
U=⎣ ⎦. (9)
.. .. .. .. ..
. . . . .
1512 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 70, NO. 3, MARCH 2022

Algorithm 1 QR-SVD Algorithm

Input: U, Y
Output: A
1: Perform QR decomposition on U: U = QR
2: D = Q H Y
3: Perform SVD on D: D = UD D VD H
4: Determine the number of required principal components
PW by checking the diagonal elements of D and remove
those with very small magnitude
(P )
Fig. 4. Extraction of PA feature for PAs in MIMO array. 5: Form UD W using the first PW columns of UD
(P )
6: A = R −1 UD W

A. Feature Extraction of General PA Behaviors

To simplify the PA and MIMO system characterization,
we first learn a transformation matrix that extracts the essential
features from the nonlinear behaviors of PAs. In [29], it is
shown that the PA behavior under varying operating conditions
can be represented with a small number of features using a
pretraining technique, namely, (6) can be rewritten as
ynt = UAωnt (10)
where A is a transformation matrix and ωnt is the extracted
PA features for the nth PA in the array under tth power level.
As shown in Fig. 3, the PA features have a much smaller
dimension but can perfectly reconstruct the original PA model
coefficients. In other words, the useful information about PA Fig. 5. PA features generated by the core PA model and general power-
behavior can be fully encoded into the PA features. induced information.
In MIMO TXs, all the PAs are usually designed to be
identical but may differ slightly because of process, voltage,
and temperature (PVT) variations. Based on the previous the PAs behave differently, the change of their behavior under
findings, we can further extract the common properties shared different operating conditions may follow the same pattern,
by all PAs in the array. In this way, the aggregated model which can be extracted and shared by all the PAs. In this
function section, we propose to extract the shared pattern of PA
Y = U (11) behavior under different power levels.
We achieve this goal by extracting the general information
where Y = [y11 , y12 , . . . , ynt , . . .] is the output signals and about power-induced variation of PA nonlinearity into matrix
 = [λ11 , λ12 , . . . , λnt , . . .] includes model coefficients of the Bt which is shared by all the PAs. As shown in Fig. 5, the PA
PAs under different power levels, can be converted to features of nth PA under tth power level can be approximately
Y = UAW (12) decomposed as

where the PA features are aggregated in W =

[ω11 , ω12 , . . . , ωnt , . . .]. As shown in Fig. 4, A is shared ωnt ≈ Bt γn (13)
by all operating conditions and each condition has its own
features ω, so we only need to update ω when the system
where γn is the core PA model coefficients to describe the
operating condition changes.
characteristics of nth PA. It is noted that Bt solely depends
During the initial offline training, output data of each PA
on the power level, while γn is fully determined by the PA.
operating under different conditions are captured. With the
It suggests that, when PA behavior is affected by long-term
sweeping data, the important PA features can be found using
effects, like temperature or device aging, only γn needs to
singular vector decomposition (SVD). To realize this idea,
be updated. Thus, with the extracted pattern, just one set of
the QR-SVD pretraining algorithm [29] is used to extract
coefficients is enough to build the PA features for all power
the transformation matrix A. The detailed procedures are
levels of each PA. In this way, we can not only reduce
displayed in Algorithm 1.
the computational complexity, but also update PA features
more conveniently, because the information extracted from one
B. Core PA Model for Aggregation of Operating Conditions operating condition can be extrapolated to other situations.
With the PA feature extraction, the model size is signifi- The transformation matrix can be identified with the PA
cantly reduced, but we still need to extract separate PA features features extracted previously, as shown in Fig. 6. Assuming
for each operating condition of a given PA. Fortunately, while each PA has T discrete power levels, the features of all N
WANG et al.: DIGITAL PREDISTORTION OF 5G MULTIUSER MIMO TRANSMITTERS 1513

Fig. 7. Generation of DPD coefficients for one subarray with DPD generator
under lth power level.

Fig. 6. Extraction of general power-induced information B. signal of the vth subarray in the space can be represented by

Nv
zv = Uv λntn e j (n−1)(β+θv )
PAs under T different power levels can be rearranged as n=1
⎡ ⎤ 
Nv
ω11 ω21 ··· ωN 1 = Uv Aωnt n e j (n−1)(β+θv )
⎢ ω12 ω22 ··· ωN 2 ⎥
⎢ ⎥ n=1
W = ⎢ . .. .. .. ⎥ (14)
⎣ .. . . . ⎦ 
Nv

ω1T ω2T ··· ωN T = Uv A ωnt n e j (n−1)(β+θv ) . (17)

n=1

where each column is the coefficients of one PA under T From (17), the PA features of each PA in the subarray
different power levels. Perform SVD on matrix W are combined to model the spatial combination of transmitted
signals. The combined features are referred to as array fea-
W = UW W VW H . (15) tures. As shown in Fig. 7, the PA features are added together
using beam direction information as weighting factors and then
By keeping the PC most significant principal components, normalized. Therefore, the subarray to be linearized can be
that is the first PC columns of UW , we obtain modeled in a similar way as a single PA, using normalized
array features ω instead of PA features ω. From a behavioral
 T modeling point of view, the subarray can simply be viewed as
B = B1 T , B2 T , . . . , BTT = UW (:, 1 : PC ). (16)
a special PA with a different set of parameters. Thus, to derive
the relationship between the array model and the desired
If the most representative cases are covered in offline
DPD coefficients, it is sufficient to examine such relationship
training, the transformation matrices can also be generalized
on individual PAs in the array, so we derive and train the
to cover untested cases by interpolation.
DPD generator by collecting the PA features and the DPD
coefficients for the corresponding PAs.
IV. D IRECT S YNTHESIS OF DPD M ODEL C OEFFICIENTS
B. DPD Generator
In the conventional DPD, the model coefficients are usually
To derive the model structure of DPD generator, we first
extracted using linear system identification algorithms, such
consider a post-inverse model where a DPD model is cas-
as LS in (5), where a group of input–output data are gathered
caded after the PA. After model extraction, the post-inverse
and matrix operations are conducted. As mentioned earlier,
coefficients can be directly copied to the pre-inverse in DPD.
this type of model extraction is computationally complex and
Representing the PA using the extracted PA features, the output
time consuming, which is not suitable for MIMO TXs. In this
of post-inverse DPD model is
section, we present a novel DPD generator to find the value
of the model coefficients directly without using LS. The DPD Yc = f DPD (y) = f DPD (UAω) (18)
generator is a model built between PA features ω and the
corresponding DPD model coefficients c. where Y includes DPD basis functions built by PA output
signal y and f DPD represents the DPD model.
When PA characteristics (ω) vary, to ensure the cascaded
system is linear, all DPD coefficients (c) need to change
A. Array Model and DPD Formulation
accordingly. Thus, every DPD coefficient is a function of PA
Before deriving the DPD generator, let us check the array features ω. Without loss of generality, we consider i th DPD
model equation. Consider the case of HB TXs, where one DPD coefficient, c(i ), in the following derivation.
is assigned to each subarray and the DPD linearizes the signals Due to the use of SVD technique in feature extraction,
of the subarray combined in space. According to (3), if we the importance of different PA features decays exponentially,
model the PA behavior with the PA features, the transmitted and the first feature ω(1) is far more important than the rest
1514 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 70, NO. 3, MARCH 2022

TABLE I
C OMPLEXITY C OMPARISON OF DPD M ODEL A DAPTATION C ONSIDERING
C OMPLEX M ULTIPLICATION O PERATIONS

Fig. 8. Operation of DPD generator.

features. Thus, ω(1) dominates the variation of PA behavior,

so it can be viewed as the main factor in modeling c. For
the rest PA features, as their influence is much smaller than C. Offline Extraction of DPD Generator
ω(1), we only consider their linear effect. Thanks to the feature
extraction, simple polynomials are sufficient to describe the An initial offline extraction is required to determine the
relationship between PA features and DPD model coefficients. parameters of DPD generator, which does not need to operate
The model of the DPD generator can thus be expressed as during deployment. To obtain the required data for offline
training, iterative learning control (ILC) [30] measurement is
conducted on a set of randomly selected PAs in the array.

PG 
PW
Taking DPD model coefficients for nth PA under tth power
c(i ) = CGi (0) + CGi ( p, q)|ω(1)|q−1 ω( p) (19)
level as an example, the matrix format of DPD generator can
q=1 p=1
be expressed by
where CG is the coefficients of DPD generator and ω( p) is cnt = CG φnt (20)
pth entry of ω. The derived mapping can be conveniently
applied to all coefficients and the complete DPD generator where φnt = [1, ωnt T , |ωnt (1)|ωnt T , . . .]T .
is illustrated in Fig. 8. After collecting the DPD model coefficients under diverse
In the conventional DPD method, array model is calculated operating conditions, they can be represented by
with coefficients of each PA model in the subarray. Then C = CG  (21)
the target signal in space is generated by multiplying the
basis function for the subarray. DPD model is extracted using where
the target signal as the input and PA input as the output
 = [φ11 , φ12 , . . . , φnt , . . .]. (22)
with LS. Compared with the conventional model extraction
method, the complexity of DPD generator is significantly CG can then be extracted using LS
lower because it does not need large number of samples
to extract DPD model coefficients. To give a clear picture, CG = CH (H )−1 . (23)
we present a comparison of the computational complexity of It is worth noting that not all the coefficients are needed in
the different adaptation algorithms with a specific example the DPD generator model since the importance of coefficients
which uses the same setup as the experimental validation in ω varies significantly. For example, if PA features have
in Section VI-A. We assume each subarray has N = 16 PW = 6 coefficients, the first four coefficients may be
PAs. There are PA = 52 and PD = 52 basis functions in enough to represent most of characteristics and build the DPD
each original PA model and each DPD model, respectively. generator in some cases.
N S = 8192 training samples are used in the conventional Besides polynomials, other modeling methods may also be
model extraction. PA features for each PA under one power used to build DPD generator model. For example, neural net-
level have PW = 5 coefficients. Core PA model for each PA works may show good performance if we have large datasets
has PC = 5 coefficients. The nonlinear order of DPD generator to train the DPD generator.
is PG = 2. The computational complexity is measured by the
number of complex multiplications. From the results shown in
Table I, we can see that the proposed method can reduce the D. DPD Model Update for Real-Time Operation
computational complexity by several orders of magnitude. Based on the techniques presented in Sections III and IV,
Also, different from the conventional pth-order inverse we propose a novel DPD adaptation method for MU MIMO
method, the proposed technique is compatible with all types TXs to lower the latency and complexity.
of DPD models. Moreover, due to the use of unsupervised The DPD model can be updated as soon as the system
feature extraction technique, the DPD generator operates on configuration, e.g., power-level or beam direction, changes.
a reduced set of PA features, rather than the original PA The adaptation procedure is depicted in Fig. 9. For simplicity,
model coefficients, so it can achieve robust generalization only the adaptation for one subarray is shown. The adaptation
performance with few training samples. of this part does not require capturing new feedback data, and
WANG et al.: DIGITAL PREDISTORTION OF 5G MULTIUSER MIMO TRANSMITTERS 1515

Fig. 9. Adaptaion of proposed MU MIMO DPD architecture.

the new DPD coefficients can be conveniently derived based

on the features of the new system configuration. Thus, the
complexity and latency are greatly reduced.
In this case of HB TXs, the features of vth subarray ωv
can be calculated based on the new system configuration,
as discussed in Section IV-A, and sent to the DPD generator.
The new DPD model coefficients are obtained by applying
DPD generator with (20)
c v = CG φ v (24)
where φ v = [1, ωv T , |ωv (1)|ωv T , . . .]T is the basis function of
the DPD generator.
When the PA behavior is affected by slow-varying effects,
the core PA models are updated in the background with LS
estimation. Thus, the DPD generator can generate accurate Fig. 10. Relative importance of principal components for PA features and
core PA model.
DPD models using the up-to-date information stored in the
core PA models.
the dimension of PA features, we examined the principal
V. S IMULATION R ESULTS components of the transform. The relative importance of the
The performance of the proposed MU MIMO DPD method principal components is shown in Fig. 10. For clearness,
was analyzed based on a 4 × 8 ULA (four subarray and eight only the first 20 principal components are shown. By setting
RF chains in each subarray) supporting four UEs. As every the threshold of principal component importance to −70 dB,
subarray supports one UE, the DPD for each subarray is set we have PW = 7, which means the each PA feature has
to linearize the corresponding UE. For PA and DPD model, seven coefficients. The NMSE between actual PA output and
generalized memory polynomial (GMP) model [31] was used modeled PA output using PA features is listed in Table II.
with M = 6, K = 3, and L = 2 (104 coefficients). Input The worst case NMSE value is −56.8 dB, which means using
waveforms were 100-MHz long-term evolution (LTE) input seven coefficients is enough to describe the PA behavior under
signals with peak-to-average power ratio (PAPR) of 8 dB. The any operating condition.
number of possible power levels for each PA were set to 5. Matrix B was then identified using SVD to extract the
shared pattern of power-induced variations, as shown in Fig. 6.
Similar to A, the dimension of B is also determined by
A. Initial Offline Training analyzing the relative importance of principal components,
In the initial offline training stage, 32 different PA models which is shown in Fig. 10. With the same threshold, we find
working under five power levels, which is 32 × 5 = 160 PA PC = 4. The NMSE between the PA features and the modeled
operating conditions in total, were used to extract transformed PA features with core PA models is listed in Table II. The
matrices. To extract each PA model, we set a random threshold worst case of the NMSE is −44.4 dB, which suggests four
for the amplitude of signals and deleted data samples above coefficients are enough for each PA to describe the nonlinear
the threshold. As each PA model was extracted using different behavior under all power levels.
data, they could represent different nonlinear behavior. Subsequently, 80 PA operating conditions were used in
Matrix A was first extracted to encode the original PA model the extraction of DPD generator. About 56 of them were
coefficients into PA features, as shown in Fig. 4. To decide for training the DPD generator and the rest 24 were for
1516 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 70, NO. 3, MARCH 2022

TABLE II
S IMULATED P ERFORMANCE OF O FFLINE E XTRACTION

Fig. 12. Simulated linearization performance of signals received by 3rd UE

under 5th power level with 100-MHz signals. (a) Spectrum. (b) AM-AM and
AM-PM results.

TABLE III
S IMULATED L INEARIZATION P ERFORMANCE OF P ROPOSED DPD M ETHOD
ON 4 × 8 A RRAY U NDER F IVE P OWER L EVELS S UPPORTING F OUR UE S

Fig. 11. Simulated linearization performance of signals received by 1st UE

under 2nd power level with 100-MHz signals. (a) Spectrum. (b) AM-AM and
AM-PM results.

testing its accuracy. To extract the model coefficients of DPD TABLE IV

generator, the corresponding DPD model coefficients were first S IMULATED L INEARIZATION P ERFORMANCE OF P ROPOSED DPD M ETHOD
obtained by ILC-DPD [30]. The NMSE between the ideal and ON 4 × 8 A RRAY U NDER F IVE P OWER L EVELS S UPPORTING F OUR UE S
synthesized DPD coefficients is listed in Table II. The worst W ITH M UTUAL C OUPLING
case NMSE is −60.0 dB. The NMSE results show that the
extracted DPD generator can accurately reproduce the DPD
model coefficients, even for unseen test states.

B. Linearization Performance
The linearization performance of the proposed DPD method
was then simulated. Core PA model coefficients were updated
with the OTA-based indirect PA identification method [16]. Furthermore, only the coupling from the adjacent RF chains
Data blocks from 68 directions were used to update the core at each side was considered. The coupling factor was set as
PA models and calculate the PA features. It is noted that 0.5, which means that half of the signals from the adjacent RF
PA outputs were not used in DPD adaptation but they were chain was coupled. From the simulation results, we can see
estimated to verify the accuracy of the forward modeling with that, even with a high coupling factor, the mutual coupling has
OTA data acquisition. The worst case NMSE of PA output little impact on the reconstructed PA outputs. With the core
reconstruction is −43.6 dB. PA models, the worst case NMSE of the 32 reconstructed PA
After that, DPD model coefficients were obtained directly outputs is −43.7 dB. The linearization performance is listed
by DPD generator. The linearization performance of the pro- in Table IV, demonstrating that the main beam signals are
posed DPD method for all the four UEs is listed in Table III. effectively linearized with the proposed DPD method under
From the results, the proposed DPD method successfully mutual coupling effect. This is because the received signal at
linearized all UEs. The worst case NMSE and ACPR of four the main beam is still a linear combination of the PA outputs
UEs are −45.9 dB and −51.6 dBc, respectively. Figs. 11 and if only linear mutual coupling occurs. For nonlinear coupling
12 show the spectrum, AM-AM and AM-PM results of signals where the coupling affects the PA nonlinearity, the situation
received by the 1st UE under 2nd power level and 3rd UE is more complex. We will investigate that in our future work.
under fifth power level without and with the proposed DPD.
VI. E XPERIMENTAL R ESULTS
C. Effect of Mutual Coupling In this section, we validate the efficacy of the proposed
Effect of mutual coupling between antenna elements was method on two different MIMO array platforms. In both cases,
considered in the MATLAB simulation. For simplicity, only each PA element has three discrete power levels.
linear coupling between the RF chains was considered in the
simulation. In other words, the output of each RF chain with A. Linearization of Hybrid Beamforming Transmitter
coupling was generated by adding the output of the PA in its 1) Experimental Setup: The test bench for 4 × 4 URA
own RF chain with the outputs from the adjacent RF chains. was set up as shown in Fig. 13. The device used in the
WANG et al.: DIGITAL PREDISTORTION OF 5G MULTIUSER MIMO TRANSMITTERS 1517

TABLE V
E XPERIMENTAL L INEARIZATION P ERFORMANCE OF THE P ROPOSED
DPD U SING OTA D ATA A CQUISITION FOR HB TX W ITH M ODEL
E XTRACTED U NDER −6.0 D B M

Fig. 13. Test bench of the HB TX.

16 channel phase array is AWMF-0158, an integrated sili-

con chip from Anokiwave. The 2-D planar antenna array is
composed of 16 microstrip antennas. The distance between
antenna elements is 6 mm, which was designed to work at the
center frequency of 28 GHz. Due to the limited availability
of test equipment, only a single user case was tested. One
baseband input signal with bandwidth of 20 MHz and PAPR Fig. 14. Measured linearization performance of main beam with OTA data
of 6.5 dB was downloaded to the signal generator. The signal acquisition under −6.0-dBm power in HB TX. (a) Spectrum comparison and
(b) AM-AM and AM-PM results of proposed DPD.
was then sent to an in-house designed 16-channel beamform-
ing transmitter (16 PAs and 16 antennas) and upconverted
to 28 GHz. The available input power levels were −6.0, −6.5,
and −7.0 dBm. Data blocks with 16 different main beam
directions were transmitted. At the receiver side, one horn
antenna was employed for OTA test and a spectrum analyzer
(Keysight N9030A) was utilized to capture the OTA outputs.
The OTA outputs were then used to extract the core PA model
for the 16 PAs in the array according to the method in [15].
Both the outputs and the inputs were sent back to the PC
for DPD procedures. As for forward and DPD models, GMP
model was used with M = 6, K = 3 and L = 2.
In the offline training stage, each PA output was captured by
the horn antenna by turning on the single PA and turning off
the other PAs. The threshold of principal component impor-
tance was set to −55 dB for the extraction of transformation
matrices A and B. PW was set to 5 and PC was set to 5. The Fig. 15. Spectrum comparison of main beam with power level 1 (−6.0 dBm)
and power level 3 (−7.0 dBm) in HB TX.
order of DPD generator was set to 2.
2) Linearization Performance: The linearization of the pro-
TABLE VI
posed DPD method was tested on the HB TX. In the online
E XPERIMENTAL L INEARIZATION P ERFORMANCE OF C ONVENTIONAL DPD
stage, core PA model coefficients were extracted at −6.0-dBm FOR HB TX W ITH M ODEL E XTRACTED U NDER −6.0- D B M P OWER
power using OTA-based data acquisition [15]. Then they were
used to generate the DPD model coefficients when the power
is −6.5 and −7.0 dBm.
With OTA-based data acquisition, data blocks from 21 direc-
tions were used to update core PA model coefficients. The
resulting linearization performance is listed in Table V. The
spectrum, AM-AM, and AM-PM results of the main beam
namely −6.5 and −7.0 dBm. The linearization performance
without and with the proposed DPD are shown in Fig. 14.
is listed in Table VI. The spectrum of the conventional DPD is
The results of proposed DPD solution show that the proposed
shown in Fig. 15. It shows good linearization performance for
DPD method can linearize all the PA under different power
the −6.0-dBm case, but it cannot linearize main beam under
levels.
other power levels.
3) Comparison With Conventional Method: Conventional
DPD was tested for comparison. The DPD model was
extracted by capturing the main beam signals with power of B. Linearization of Multiuser MIMO Transmitter
−6.0 dBm using ILA. The extracted DPD model was then 1) Experimental Setup: The test bench for 2 × 1 ULA
applied on the main beam signals with other power levels, was set up as shown in Fig. 16. Due to the limited hardware
1518 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 70, NO. 3, MARCH 2022

Fig. 17. Experimental linearization performance of 2nd PA under 3.0-dBm

power in MU MIMO TX. (a) Spectrum comparison and (b) AM-AM and
AM-PM results of proposed DPD.
Fig. 16. Test bench of the MU MIMO TX.

resource available, the proposed DPD method was tested on

2 × 1 ULA for multiuser verification. The antenna array
is tapered slot antenna. The spacing of antenna elements
is 6 mm, slightly longer than the half-wavelength 5.5 mm
at 27 GHz. The MIMO TX consisted of two RF chains (two
PAs and two antennas) and supported two UEs. Each PA in
the array has a separate DPD for linearization. Two baseband
input signals for two UEs with bandwidth of 20 MHz were
Fig. 18. Experimental linearization performance of 2nd PA under 2.0-dBm
generated and precoded by MATLAB on PC, then downloaded power in MU MIMO TX. (a) Spectrum comparison and (b) AM-AM and
to the two signal channels provided by two single channel AM-PM results of proposed DPD.
signal generators SMW200A. PAPR of the transmitted signals
was 7 dB. The baseband signals were upconverted to IF at TABLE VII
5.5 GHz and fed into the designed RF front end. In this NMSE OF THE E XPERIMENTAL L INEARIZATION P ERFORMANCE OF THE
module, all signals were again upconverted to 27 GHz by P ROPOSED DPD M ETHOD ON MU MIMO TX
10.75 GHz local oscillator (LO) signals generated by a signal
generator (Keysight E8267D). The available power levels were
3.0, 2.5, and 2.0 dBm. The TX chain outputs were fed into
the antenna elements to form the desired radiation pattern.
To verify whether the proposed DPD solution is compatible
with different PA data acquisition approaches, in this test,
at the receiver side, one horn antenna was connected to TABLE VIII
spectrum analyzer (Keysight N9030A) for data acquisition via ACPR OF THE E XPERIMENTAL L INEARIZATION P ERFORMANCE OF THE
OTA test. The output of each PA was captured by the horn P ROPOSED DPD M ETHOD ON MU MIMO TX
antenna by only tuning on the selected PA at a time. This is
equivalent to using a coupler at PA output to capture the output
signal. As for forward and DPD models, MP model [28] was
used with M = 5 and K = 3.
In the offline training stage, each PA output was captured
by the horn antenna by turning on the selected PA and turning
off the other. The threshold of principal component importance
to be −60 dB for the extraction of transformation matrices A 3) Comparison With Conventional Method: The conven-
and B. PW was set to 3 and PC was set to 2. The order of tional adaptation method using the same setup was also tested.
DPD generator was set to 1. The DPD models were extracted based on data captured under
2) Linearization Performance: PA output was captured with 3.0-dBm power using ILA. The extracted DPD models were
the power of 3.0 dBm and then used to extract core PA model then applied on the PA under 2.5- and 2.0-dBm power. The
coefficients. The transformed PA model coefficients as well as linearization performance is listed in Table IX. The spectrum,
DPD model coefficients for other PA conditions under 2.5- and AM-AM, and AM-PM results for the second PA with the
2.0-dBm power were generated directly without PA adaptation. conventional DPD method are shown in Figs. 17 and 18.
The linearization performance is listed in Tables VII and VIII. It shows that the performance of conventional adaptation
The spectrum, AM-AM, and AM-PM results for the second degrades significantly in some scenarios, because the PAs
PA are shown in Figs. 17 and 18, respectively. The results cannot be linearized by the conventional DPD extracted at
show that the proposed DPD method can linearize all the PA a different power level, which further establishes the efficacy
with different power levels. of the proposed technique.
WANG et al.: DIGITAL PREDISTORTION OF 5G MULTIUSER MIMO TRANSMITTERS 1519

TABLE IX [8] N. Tervo, J. Aikio, T. Tuovinen, T. Rahkonen, and A. Parssinen,

E XPERIMENTAL L INEARIZATION P ERFORMANCE OF DPD M ODEL “Digital predistortion of amplitude varying phased array utilising over-
E XTRACTED U NDER 3.0 D B M P OWER ON MU MIMO TX the-air combining,” in IEEE MTT-S Int. Microw. Symp. Dig., Jun. 2017,
pp. 1165–1168.
[9] X. Liu et al., “Beam-oriented digital predistortion for 5G massive MIMO
hybrid beamforming transmitters,” IEEE Trans. Microw. Theory Techn.,
vol. 66, no. 7, pp. 3419–3432, Jul. 2018.
[10] R. Murugesu, M. Holyoak, H. Chow, and S. Shahramian, “Lineariza-
tion of mm-wave large-scale phased arrays using near-field coupling
feedback for >10Gb/s wireless communication,” in IEEE MTT-S Int.
Microw. Symp. Dig., Aug. 2020, pp. 1271–1274.
[11] S. Hesami, S. R. Aghdam, C. Fager, T. Eriksson, R. Farrell, and
J. Dooley, “Intra-array coupling estimation for MIMO transceivers uti-
VII. C ONCLUSION lizing blind over-the-air measurements,” in IEEE MTT-S Int. Microw.
Symp. Dig., Jun. 2019, pp. 468–471.
In this article, a novel DPD solution for MU MIMO TXs
[12] N. Tervo et al., “Digital predistortion of phased-array transmitter with
has been proposed. To lower the computational complexity and shared feedback and far-field calibration,” IEEE Trans. Microw. Theory
increase the adaptation speed, the concept of DPD generator Techn., vol. 69, no. 1, pp. 1000–1015, Jan. 2021.
is proposed to generate the DPD model coefficients directly [13] Q. Luo, X. Zhu, C. Yu, and W. Hong, “Single-receiver over-the-air digi-
tal predistortion for massive MIMO transmitters with antenna crosstalk,”
with transformed forward model coefficients, together with IEEE Trans. Microw. Theory Techn., vol. 68, no. 1, pp. 301–315,
the power and angle information. In this case, the acquisition Jan. 2020.
of PA output signal and the LS estimation of DPD model [14] K. Hausmair, U. Gustavsson, C. Fager, and T. Eriksson, “Modeling and
linearization of multi-antenna transmitters using over-the-air measure-
coefficients are no longer necessary, unless the PA behavior ments,” in Proc. IEEE Int. Symp. Circuits Syst. (ISCAS), May 2018,
is affected by long-term effects. Feature extraction algorithms pp. 1–4.
are employed to extract transformation matrices shared by all [15] X. Wang, Y. Li, C. Yu, W. Hong, and A. Zhu, “Digital predistortion of
PAs in the array so that only one core PA model is enough to 5G massive MIMO wireless transmitters based on indirect identification
of power amplifier behavior with OTA tests,” IEEE Trans. Microw.
describe the distinct characteristics of each PA. The core PA Theory Techn., vol. 68, no. 1, pp. 316–328, Jan. 2020.
model can thus be updated efficiently, and the calibration is [16] X. Wang, Y. Li, C. Yu, W. Hong, and A. Zhu, “OTA-based data
needed only when considering slow-varying dynamics of the acquisition and signal separation for digital predistortion of multi-user
MIMO transmitters in 5G,” in IEEE MTT-S Int. Microw. Symp. Dig.,
PAs. According to the simulation and experimental validation, Aug. 2020, pp. 845–848.
the proposed DPD method can realize the linearization of [17] A. Gaber, A. Nahler, W. Nitzold, and M. Anderseck, “USRP-based
mmWave MU MIMO TXs with low complexity, proving itself mmWave prototyping architecture with real-time RF control,” in IEEE
MTT-S Int. Microw. Symp. Dig., Jun. 2021, pp. 693–696.
as a promising solution for linearization of 5G MU MIMO
[18] Physical Layer Procedures for Control, document TS 38.213, 3GPP,
TXs. 2007.
It is worth noting that, although only power and angle [19] A. Brihuega, L. Anttila, and M. Valkama, “Neural-network-based digital
conditions are considered and demonstrated in this article, the predistortion for active antenna arrays under load modulation,” IEEE
Microw. Wireless Compon. Lett., vol. 30, no. 8, pp. 843–846, Aug. 2020.
principle of the proposed approach can be generally extended [20] Y. Guo, C. Yu, and A. Zhu, “Power adaptive digital predistortion
to other conditions, such as bandwidth and center frequency. for wideband RF power amplifiers with dynamic power transmission,”
The only difference is that the sweeping measurements are IEEE Trans. Microw. Theory Techn., vol. 63, no. 11, pp. 3595–3607,
Nov. 2015.
expected to include these varying factors and the features are [21] F. M. Barradas, T. R. Cunha, and J. C. Pedro, “Digital predistortion
generated under these dynamic conditions. of RF PAs for MIMO transmitters based on the equivalent load,” in
Proc. Integr. Nonlinear Microw. Millimetre-Wave Circuits Workshop,
Apr. 2017, pp. 1–4.
R EFERENCES [22] S. K. Dhar, A. Abdelhafiz, M. Aziz, M. Helaoui, and F. M. Ghannouchi,
“A reflection-aware unified modeling and linearization approach for
[1] S. Mumtaz, J. Rodriguez, and L. Dai, MmWave Massive MIMO:
power amplifier under mismatch and mutual coupling,” IEEE Trans.
A Paradigm for 5G. New York, NY, USA: Academic, 2018.
Microw. Theory Techn., vol. 66, no. 9, pp. 4147–4157, Sep. 2018.
[2] L. Guan and A. Zhu, “Green communications: Digital predistortion for
wideband RF power amplifiers,” IEEE Microw. Mag., vol. 15, no. 7, [23] E. Ng, Y. Beltagy, G. Scarlato, A. Ben Ayed, P. Mitran, and S. Boumaiza,
pp. 84–99, Nov./Dec. 2014. “Digital predistortion of millimeter-wave RF beamforming arrays using
[3] F. M. Ghannouchi and O. Hammi, “Behavioral modeling and predistor- low number of steering angle-dependent coefficient sets,” IEEE Trans.
Microw. Theory Techn., vol. 67, no. 11, pp. 4479–4492, Nov. 2019.
tion,” IEEE Microw. Mag., vol. 10, no. 7, pp. 52–64, Dec. 2009.
[4] C. Fager, T. Eriksson, F. Barradas, K. Hausmair, T. Cunha, and [24] H. Yin et al., “Data-clustering-assisted digital predistortion for 5G
J. C. Pedro, “Linearity and efficiency in 5G transmitters: New tech- millimeter-wave beamforming transmitters with multiple dynamic con-
niques for analyzing efficiency, linearity, and linearization in a 5G figurations,” IEEE Trans. Microw. Theory Techn., vol. 69, no. 3,
active antenna transmitter context,” IEEE Microw. Mag., vol. 20, no. 5, pp. 1805–1816, Mar. 2021.
pp. 35–49, May 2019. [25] X. Liu, W. Chen, L. Chen, F. M. Ghannouchi, and Z. Feng, “Power
[5] S. A. Bassam, M. Helaoui, and F. M. Ghannouchi, “Crossover digital scalable beam-oriented digital predistortion for compact hybrid massive
predistorter for the compensation of crosstalk and nonlinearity in MIMO MIMO transmitters,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 67,
transmitters,” IEEE Trans. Microw. Theory Techn., vol. 57, no. 5, no. 12, pp. 4994–5006, Dec. 2020.
pp. 1119–1128, May 2009. [26] M. Shafi et al., “5G: A tutorial overview of standards, trials, challenges,
[6] A. Abdelhafiz, L. Behjat, F. M. Ghannouchi, M. Helaoui, and O. Hammi, deployment, and practice,” IEEE J. Sel. Areas Commun., vol. 35, no. 6,
“A high-performance complexity reduced behavioral model and digital pp. 1201–1221, Jun. 2017.
predistorter for MIMO systems with crosstalk,” IEEE Trans. Commun., [27] C. Yu et al., “Full-angle digital predistortion of 5G millimeter-wave
vol. 64, no. 5, pp. 1996–2004, May 2016. massive MIMO transmitters,” IEEE Trans. Microw. Theory Techn.,
[7] K. Hausmair, P. N. Landin, U. Gustavsson, C. Fager, and T. Eriksson, vol. 67, no. 7, pp. 2847–2860, Jul. 2019.
“Digital predistortion for multi-antenna transmitters affected by antenna [28] L. Ding et al., “A robust digital baseband predistorter constructed
crosstalk,” IEEE Trans. Microw. Theory Techn., vol. 66, no. 3, using memory polynomials,” IEEE Trans. Commun., vol. 52, no. 1,
pp. 1524–1535, Mar. 2018. pp. 159–165, Jan. 2004.
1520 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 70, NO. 3, MARCH 2022

[29] Y. Li, X. Wang, and A. Zhu, “Complexity-reduced model adaptation Zhiqiang Yu (Member, IEEE) received the B.S.
for digital predistortion of RF power amplifiers with pretraining-based degree from the Nanjing University of Science and
feature extraction,” IEEE Trans. Microw. Theory Techn., vol. 69, no. 3, Technology, Nanjing, China, in 2002, and the Ph.D.
pp. 1780–1790, Mar. 2021. degree from Southeast University, Nanjing, in 2013.
[30] J. Chani-Cahuana, P. N. Landin, C. Fager, and T. Eriksson, “Iterative From 2002 to 2007, he was a Research Staff in
learning control for RF power amplifier linearization,” IEEE Trans. airborne radar transmitter with the Nanjing Institute
Microw. Theory Techn., vol. 64, no. 9, pp. 2778–2789, Sep. 2016. of Electronics, China Electronics Technology Group
[31] D. R. Morgan, Z. Ma, J. Kim, M. G. Zierdt, and J. Pastalan, “A gener- Corporation, Nanjing. He is currently a Lecturer
alized memory polynomial model for digital predistortion of RF power with the School of Information Science and Engi-
amplifiers,” IEEE Trans. Signal Process., vol. 54, no. 10, pp. 3852–3860, neering, Southeast University. His current research
Oct. 2006. interests include microwave and millimeter-wave
transceiver systems, beamforming networks, and phased arrays for mobile
communication.

Xiaoyu Wang (Graduate Student Member, IEEE)

received the B.E. degree in information engineering
from Southeast University, Nanjing, China, in 2015.
She is currently pursuing the Ph.D. degree with Wei Hong (Fellow, IEEE) received the B.S. degree
University College Dublin (UCD), Dublin, Ireland. from the University of Information Engineering,
She is currently with the RF and Microwave Zhengzhou, China, in 1982, and the M.S. and Ph.D.
Research Group, UCD. Her current research focuses degrees from Southeast University, Nanjing, China,
on digital predistortion for RF power amplifiers, with in 1985 and 1988, respectively, all in radio engineer-
a particular emphasis on applications to multiple- ing.
input multiple-output (MIMO) systems. Since 1988, he has been with the State Key
Laboratory of Millimeter Waves and serves for the
Director for the Laboratory since 2003, where he is
currently a Professor and the Dean of the School
of Information Science and Engineering, Southeast
University. In 1993, 1995, 1996, 1997, and 1998, he was a short-term
Visiting Scholar with the University of California at Berkeley, Berkeley,
Yue Li (Member, IEEE) received the B.E. degree in CA, USA, and at Santa Cruz, CA, respectively. He has been engaged in
information engineering from Southeast University, numerical methods for electromagnetic problems, millimeter wave theory
Nanjing, China, in 2016, and the Ph.D. degree and technology, antennas, RF technology for wireless communications, etc.
in electronic engineering from University College He has authored and coauthored over 300 technical publications with over
Dublin (UCD), Dublin, Ireland, in 2020. 9000 citations, and authored two books.
He is currently a Post-Doctoral Researcher with Dr. Hong is a fellow of CIE, the Vice President of the CIE Microwave
the RF and Microwave Research Group, UCD. His Society and Antenna Society, the Chair of the IEEE MTT-S/AP-S/EMC-S
current research interests include behavioral model- Joint Nanjing Chapter, and was an Elected IEEE MTT-S AdCom Member
ing and digital predistortion for RF power amplifiers. from 2014 to 2016. He twice awarded the National Natural Prizes, thrice
awarded the first-class Science and Technology Progress Prizes issued by
the Ministry of Education of China and Jiangsu Province Government,
etc. Besides, he also received the Foundations for China Distinguished
Young Investigators and for Innovation Group issued by NSF of China.
He served as an Associate Editor for the IEEE T RANSACTIONS ON MTT
Hang Yin (Graduate Student Member, IEEE) from 2007 to 2010, one of the Guest editors for the 5G special issue of the
received the B.E. degree in information engineering IEEE T RANSACTIONS ON AP in 2017.
and the M.E. degree in electromagnetic fields and
microwave technology from Southeast University
(SEU), Nanjing, China, in 2018 and 2021,
respectively. He is currently pursuing the Ph.D.
degree with University College Dublin (UCD),
Dublin, Ireland, from 2021.
His current research interests include the Anding Zhu (Senior Member, IEEE) received the
modeling and linearization [especially artificial Ph.D. degree in electronic engineering from Uni-
intelligence (AI) based] of radio frequency power versity College Dublin (UCD), Dublin, Ireland,
amplifiers (PAs) and transmitters. in 2004.
He is currently a Professor with the School
of Electrical and Electronic Engineering, UCD.
His research interests include high-frequency non-
linear system modeling and device characteriza-
Chao Yu (Senior Member, IEEE) received the B.E. tion techniques, high-efficiency power amplifier
degree in information engineering and the M.E. design, wireless transmitter architectures, digital sig-
degree in electromagnetic fields and microwave tech- nal processing, and nonlinear system identification
nology from Southeast University (SEU), Nanjing, algorithms.
China, in 2007 and 2010, respectively, and the Ph.D. Prof. Zhu is an Elected Member of MTT-S AdCom, the Chair of the
degree in electronic engineering from University Electronic Information Committee, and the Vice Chair of the Publications
College Dublin (UCD), Dublin, Ireland, in 2014. Committee. He is also the Chair of the MTT-S Microwave High-Power
He is currently a Professor with the State Key Lab- Techniques Committee. He served as the Secretary of MTT-S AdCom in 2018.
oratory of Millimeter Waves, School of Information He was a General Chair of the 2018 IEEE MTT-S International Microwave
Science and Engineering, SEU, and Purple Mountain Workshop Series on 5G Hardware and System Technologies (IMWS-5G) and
Laboratories, Nanjing. His current research interests a Guest Editor of the IEEE T RANSACTIONS ON M ICROWAVE T HEORY AND
include microwave and millimeter wave power amplifier modeling and lin- T ECHNIQUES on 5G Hardware and System Technologies. He is currently an
earization, and 5G massive multiple-input-multiple-output (MIMO) RF system Associate Editor of the IEEE Microwave Magazine and a Track Editor of the
design. IEEE T RANSACTIONS ON M ICROWAVE T HEORY AND T ECHNIQUES .