IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO.

7, JULY 2019 3437

Wirelessly Powered Data Aggregation for IoT via

Over-the-Air Function Computation: Beamforming
and Power Control
Xiaoyang Li , Student Member, IEEE, Guangxu Zhu , Student Member, IEEE, Yi Gong , Senior Member,
IEEE, and Kaibin Huang , Senior Member, IEEE

Abstract— As a revolution in networking, the Internet of Things Index Terms— Over-the-air computation, wireless power trans-
(IoT) aims at automating the operations of our societies by fer (WPT), multi-modal sensing, multiple-input multiple-output
connecting and leveraging an enormous number of distributed (MIMO), beamforming.
devices (e.g., sensors and actuators). One design challenge is
I. I NTRODUCTION
efficient wireless data aggregation (WDA) over the dense IoT
devices. This can enable a series of the IoT applications ranging
from latency-sensitive high-mobility sensing to data-intensive
distributed machine learning. Over-the-air (function) computa-
I N THE near future, tens of billions of Internet-of-
things (IoT) devices (e.g., sensors and actuators) are
expected to be deployed to automate the operations of our
tion (AirComp) has emerged to be a promising solution that societies and make the ambient environment smart. Among
merges computing and communication by exploiting analog-wave
addition in the air. Another IoT design challenge is battery others, there exist two design challenges for IoT. The first is
recharging for dense sensors which can be tackled by wireless fast wireless data aggregation (WDA), namely fast collection
power transfer (WPT). The coexisting of AirComp and WPT in and processing of data distributed at dense IoT devices by
the IoT system calls for their integration to enhance the perfor- wireless transmission. WDA is an enabling operation for a
mance and efficiency of WDA. This motivates the current work series of IoT applications such as fusion of sensing values in
on developing the wirelessly powered AirComp (WP-AirComp)
framework by jointly optimizing wireless power control, energy environmental monitoring [1], aggregation of mobile updates
and (data) aggregation beamforming to minimize the AirComp in federated machine learning [2], and distributed consensus
error. To derive a practical solution, we recast the non-convex in fleet driving [3]. Fast WDA is needed to regulate latency in
joint optimization problem into the equivalent outer and inner cases with ultra-dense devices and/or high mobility (e.g., for
sub-problems for (inner) wireless power control and energy sensors carried by drones or vehicles). A promising solution
beamforming, and (outer) the efficient aggregation beamforming,
respectively. The former is solved in closed form while the latter is over-the-air (function) computation (AirComp), which real-
is efficiently solved using the semidefinite relaxation technique. izes fast WDA by simultaneous transmissions and exploiting
The results reveal that the optimal energy beams point to analog-wave addition in a multi-access channel [22]. The
the dominant Eigen-directions of the WPT channels, and the other challenge for IoT is powering dense energy-constrained
optimal power allocation tends to equalize the close-loop (down- sensors for WDA and other operations. One attractive solution
link WPT and up-link AirComp) effective channels of different
sensors. The simulation demonstrates that the controlling WPT is wireless power transfer (WPT) using microwaves, whose
provides additional design dimensions for substantially reducing feasibility has been proved in practical sensor networks [4].
the AirComp error. To facilitate efficient implementation for IoT, it is natural to
tackle the two said challenges simultaneously by pursuing the
Manuscript received April 2, 2019; revised April 21, 2019; accepted
April 21, 2019. Date of publication May 7, 2019; date of current ver- fusion of two corresponding technologies: AirComp and WPT.
sion July 10, 2019. This work was supported in part by the Hong Kong The resultant design challenge lies in the joint optimization
Research Grants Council under Grant 17209917 and Grant 17259416, of their key operations at servers (or fusion centers), namely
in part by the Shenzhen Science and Technology Program under Grant
JCYJ20170817110410346, and in part by the Peng Cheng Laboratory under energy beamforming and power control for WPT, and (data)
Grant PCL2018KP002. This paper was presented in part at the IEEE aggregation beamforming for AirComp. This motivates the
Globecom Workshops, Abu Dhabi, 2018. The associate editor coordinating current work on developing a framework called wirelessly
the review of this paper and approving it for publication was D. Niyato.
(Corresponding author: Kaibin Huang.) powered over-the-air computation (WP-AirComp).
X. Li is with the Department of Electrical and Electronic Engineering In WP-AirComp framework, we consider an IoT system
(EEE), Southern University of Science and Technology, Shenzhen 518055, supporting down-link WPT and up-link AirComp. To be spe-
China, and also with the Department of EEE, The University of Hong Kong,
Hong Kong (e-mail: [email protected]). cific, a multi-antenna server transmits energy to power multiple
G. Zhu and K. Huang are with the Department of Electrical and sensors (or other types of devices) so that they can transmit
Electronic Engineering, The University of Hong Kong, Hong Kong sensing data for WDA at the server based on AirComp. The
(e-mail: [email protected]; [email protected]).
Y. Gong is with the Shenzhen Engineering Laboratory of Intelligent Infor- server controls the amount of energy harvested by sensors
mation Processing for IoT, Southern University of Science and Technology, via energy beamforming and power allocation to different
Shenzhen 518055, China (e-mail: [email protected]). beams. Since the transmit power of sensors depends on the
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. harvested energy, the WPT operations affect the analog-signal
Digital Object Identifier 10.1109/TWC.2019.2914046 superposition in the array observations at the server.
1536-1276 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
3438 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 7, JULY 2019

This introduces coupling between WPT and AirComp, and TABLE I

hence necessitates the joint design of wireless power control, E XAMPLES OF N OMOGRAPHIC F UNCTIONS T HAT ARE A IR C OMPUTABLE
energy and (data) aggregation beamforming. Via the joint
design, WPT contributes additional dimensions for reducing
the computation error in AirComp. The contributions of this
work are summarized as follows.
• Decomposition Based Design Approach: The said joint
design is formulated as an optimization problem for
minimizing the computation error in AirComp under
a power constraint for WPT. Without compromising
optimality, a decomposition approach is derived that
decomposes the non-convex problem into two tractable data samples, denoted as {xk }, which is characterized by the
nested sub-problems. The inner sub-problem concerns following form:
WPT optimization, involving the joint design of power K 
control and energy beamforming, given the aggregation 
beamformer. The outer sub-problem is the optimization y = h(x1 , x2 , · · · , xK ) = f gk (xk ) , (1)
k=1
of aggregation beamformer. The approach is applied
to develop an efficient convex-optimization algorithm where f (·) and gk (·) represent post-processing at the server
for solving the non-convex joint design problem of the and pre-processing at a device, respectively. The class of
simplified multiple-input single-output (MISO) case with functions having the above form is known as nomographic
single-antenna sensors, corresponding to scalar-function functions such as averaging and geometric mean. Typical
AirComp. The results are then extended to the multiple- functions in this class are summarized in Table I. The conven-
input-multiple-output (MIMO) case with multi-antenna tional approach for WDA decouples data collection, namely
sensors, corresponding to vector-function AirComp. distributed transmission of {xk }, and function computation,
• Wireless Power Control and Energy Beamforming: namely computing h({xk }) in (1). The approach is incapable
The said inner sub-problem is solved in closed form. The of supporting fast WDA as the application of any traditional
solution has simple structures that facilitate the imple- orthogonal multi-access scheme for data collection results in
mentation of WP-AirComp. First, the optimal policy for the linear scaling of latency with the number of devices.
wireless power control attempts to equalize the multiple In contrast, AirComp merges data collection (or else radio
cascaded WPT-AirComp channels to facilitate the outer resources) and computation. Specifically, the summation in (1)
sub-problem of designing aggregation beamformer. Sec- is implemented by simultaneous analog transmission to exploit
ond, each optimal energy beam points to the direction of the wave-addition of the multi-access channel. Consequently,
the corresponding WPT channel vector in the MISO case the function computation is performed “over-the-air” and the
or the dominant eigenvector of the channel matrix in the result is directly received by the server, thus giving the name
MIMO case. of the technology. Simultaneous transmission in AirComp
• Data Aggregation Beamforming: The said outer achieves low latency independent of the number of devices,
sub-problem is shown to be solvable by the powerful and thereby enables fast WDA.
technique of semi-definite relaxation (SDR). It is demon- The idea of AirComp can be traced back to the pioneering
strated by simulation that the joint design with WPT leads work studying function computation in sensor networks [23].
to significant gain in computation accuracy compared In [23], structured codes (e.g., lattice codes) are designed
with decoupled designs. for reliable function computation at a server based on dis-
Organization: The remainder of the paper is organized as tributed sensing values analog modulated and transmitted
follows. Section II reviews the state-of-the-art techniques of over a multi-access channel. The importance of the work
AirComp and WPT. Section III introduces the WP-AirComp lies in the counter-intuitive finding that interference caused
system model. Section IV presents the problem formulation by simultaneous transmission can be exploited to facilitate
for the joint design of power control, energy and aggregation computation. Subsequently, it was proved that the simple
beamforming. The solution for the MISO case is presented analog transmission without coding is optimal in terms of
in Section V. The extension to the MIMO case is given minimizing function distortion in the case of independent
in Section VI. Section VII further analyzes the difference Gaussian data sources [24]. Nevertheless, coding is still useful
between WP-AirComp framework and traditional designs. if the sources follow more complex distributions, such as
Simulation results are provided in Section VIII, followed by bivariate Gaussian [25], correlated Gaussian [26], and sum
concluding remarks in Section IX. of independent Gaussian [23]. The promising performance
(with optimality in certain cases) of simple analog Air-
Comp has led to an active area focusing on its robustness
II. S TATE - OF - THE -A RT OF A IR C OMP AND WPT
and implementation [22], [27]–[29]. In particular, techniques
A. WDA via AirComp for optimal linear decentralized estimation and distributed
Data aggregation in IoT can be posed as the mathematical power control are proposed in [27]. Another vein of research
problem of computing at a server a function h(·) of distributed focuses on transforming AirComp from theory into practice

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
LI et al.: WIRELESSLY POWERED DATA AGGREGATION FOR IoT VIA OVER-THE-AIR FUNCTION COMPUTATION 3439

by prototyping [22] and addressing practical issues such as power transfer (SWIPT), which remains active. Recent studies
transmission synchronization over sensors [29]. focus on applying SWIPT to a series of communication
It is also worth mentioning that inspired by the success systems, including MIMO communication [9], [10], orthogo-
of AirComp in computation-centric sensor networks, the rel- nal frequency-division multiple access (OFDMA) [11], [12],
evant principles have been applied to design new schemes two-way transmission [13], relaying [14]–[16], and cogni-
for rate-centric communication networks. The compute-and- tive networking [17]. Practical SWIPT designs account-
forward relaying schemes as proposed in [30] decodes and ing for imperfect channel state information were devel-
forwards linear functions of the transmitted messages. The oped in [18]. More recently, WPT was also considered for
integer-forcing linear receiver designed in [31] spatial mul- unmanned aerial vehicle (UAV) assisted communication [19],
tiplexes lattice codewords. Furthermore, the well-known phys- where an UAV serves as a mobile relay and its transmit
ical layer network coding leveraging analog wave addition power and trajectory are jointly optimized to maximize the
generalizes the celebrated network coding schemes invented throughput.
for wired networks to wireless networks [32]. Another important application of WPT is in sensor
IoT will feature multi-modal sensing and the prevalence networks, with energy constrained sensors. Recharging the
of antenna arrays (especially highly compact millimeter-wave batteries of hundreds to thousands of sensors is a chal-
arrays) at both servers and devices [33]. A multi-modal sensor lenging problem that can be solved efficiently by WPT [4].
can simultaneous monitor multiple parameters of the envi- For multi-user WPT systems, energy beamforming at dif-
ronment (e.g., pressure, light, humidity, and pollution), or of ferent power beacons needs to be coordinated to enable
a user state (e.g., motion, location, and speech). The need efficient energy multicasting. To this end, a collaborative
of WDA over multi-modal sensors and other data-intensive energy-beamforming scheme is proposed in [20] for efficiently
IoT applications (e.g., federated machine learning) calls for powering a sensor network. For large-scale wirelessly powered
the acceleration of computation rates in AirComp. While sensor networks, a novel framework of backscatter sensing
prior works mostly target single-antenna sensor network and was recently proposed in [21], where low-cost passive sensors
support scalar-function AirComp, recent research has started upload their sensing data to a drone mounted reader by
to explore the possibility of using antenna arrays to enable the concurrently reflecting the beamed power signal from a power
vector-function AirComp. A technique called MIMO AirComp beacon in a designed probabilistic manner. Then statistical
leverages spatial degrees-of-freedom to spatially multiplex inference algorithms can be devised for sensing value recovery
multi-function computation (i.e., vector-function computation) without the knowledge of channel state information.
and reduce computation errors by noise suppression [28]. Though energy beamforming in wirelessly powered commu-
The key design challenge lies in the optimization of the nication and sensor networks have been widely investigated,
receive beamformer, called aggregation beamformer, with the most of prior work focuses on rate maximization via opti-
objective of minimizing the error of computed functions, mizing the WPT efficiency. The design of wirelessly powered
w.r.t., the desired result in (1). On the other hand, traditional in-network computation, such as WP-AirComp in this work,
multi-user MIMO beamforming aims at a different objective remains as uncharted area.
that is to minimize the sum distortion of individual data
symbols. The difference in objective results in a fundamental
III. S YSTEM M ODEL
shift in the beamforming design principle. To be specific,
multi-user beamforming attempts to decouple simultaneously In this section, we consider the mobile IoT system
transmitted symbols (or equivalently sum-rate maximization), illustrated in Fig. 1, where WP-AirComp is deployed for
yielding the classic zero-forcing or minimizing mean-squared- fast WDA. As shown in Fig. 1 (c), the mobile server is
error (MMSE) design [34]. In contrast, aggregation beam- multi-functional serving as a power beacon, an access point
forming tries to balance the effects of individual MIMO (AP) and a data fusion center. The server is provisioned
channels so as to accurately compute the function in (1). The with an array of NAP antennas. We consider both the cases
principle is reflected in the beamformer design in [28] that is of single-antenna and multi-antenna sensors, corresponding
obtained from the weighted sum of multi-user MIMO channels to the equivalent cases of MISO and MIMO channels. Let
projected onto a Grassmann manifold. NSN denote the number of antennas at each of total K
sensors. Time is divided into symbol durations, each of which
lasts t0 seconds and is called a (time) slot. WP-AirComp is
B. Energy Beamforming for Wirelessly Powered implemented based on the harvested-then-transmit protocol
Communication that alternates WPT and AirComp phases with fixed slots. The
WPT was originally developed for point-to-point power energy a sensor harvests in a WPT phase is applied to power
delivery using radio waves [5]. Recently, the technology has transmission in the following AirComp phase. The operations
been further developed to power communication networks [6]. of different sensors are synchronized using a reference clock
One research focus is energy beamforming, referring to the broadcast by the server (see e.g., [29]). The channel state
use of an antenna array at a transmitter to beam energy in information (CSI) is estimated individually at each sensor from
the direction of a targeted receiver [7], [8]. The integration broadcasted pilot signals and then passed to the AP subse-
of WPT with traditional wireless communication has created quently. For simplicity, channels are assumed to vary following
a new area, called simultaneous wireless information and the block-fading model. In other words, each channel remains

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
3440 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 7, JULY 2019

Fig. 1. A high-mobility IoT system aggregating distributed sensing data using WP-AirComp.

TABLE II k, a receive beam is steered to harvest the transferred energy

N OTATION with the beamforming vector denoted by vk ∈ CNSN ×1 and
vkH vk = 1. With the assumption of K ≤ NAP , dedicated
sharp energy beams towards individual sensors can be formed
such that harvesting the side lobes of unintended beams gives
a sensor negligible energy.
Remark 1 (Scalability of AirComp): Theoretically, our pro-
posed AirComp can be also applied in large scale networks,
while the computation accuracy may compromise to some
extent due to the challenge in equalizing many sensors’
channels by a common aggregation beamformer. However,
this may not be an issue in the next generation ultra-dense
cellular/IoT networks, where the populations of IoT devices
and access points are comparable [36].
Remark 2 (Sharp Energy Beam): As discussed in
Remark 1, in practice each base station serves a finite
number of devices (say tens) even though the total number
of devices can be billions. Therefore, with a massive number
of antennas at each base station, it is reasonable to make the
assumption of K ≤ NAP . For the possible case of K  NAP ,
energy beamforming does not yield much power gain due to
the insufficient spatial degree of freedom. In this scenario,
isotropically broadcasting the power together with user
scheduling is more critical for ensuring the wireless power
fixed within a (WPT/AirComp) phase and varies over different efficiency. This, however, is outside the scope of the current
phases. The notation is summarized in Table II. paper.
A. WPT Phase Moreover, the energy-conversion process at a sensor, say
sensor k, is characterized by a fixed efficiency factor, denoted
Consider the WPT phase as shown in the top callout of as αk . According to the classic energy harvesting model [7],
Fig. 2. The server serves as a power beacon that allocates the amount of energy harvested by sensor k in a WPT phase
power for different sensors and delivers the power by energy of T slots is given by
beamforming. Let Pk denote the power allocated to sensor k.
2
The server transmission power is limited by a fixed value P0 , Qk = αk uH
k Gk vk  Pk T, (3)
giving the following total power constraint: 2
where uH k Gk vk  represents the effective channel power

K
gain after energy beamforming.
Pk ≤ P0 . (2)
k=1

Let Gk ∈ CNAP ×NSN represent the MIMO channel for the B. AirComp Phase
link from the server to sensor k, and uk ∈ CNAP ×1 with Consider the AirComp phase as shown in the bottom callout
uHk uk = 1 denote the transmit beamforming vector. At sensor of Fig. 2. Each sensor records the values of L heterogeneous

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
LI et al.: WIRELESSLY POWERED DATA AGGREGATION FOR IoT VIA OVER-THE-AIR FUNCTION COMPUTATION 3441

Fig. 2. Block diagram of WP-AirComp operations.

 
parameters of an external time-varying process such as the 
K
y = h (d1 , d2 , · · · , dK ) = f gk (dk ) , (4)
ambient environment, or a human being. The measurement
k=1
generates a vector symbol in each slot, represented by dk =
[dk1 , dk2 , ..., dkL ]T ∈ RL×1 for sensor k. Powered by WPT,
each sensor transmits vector symbols using its array to the
server. On the other hand, the server doubly serves as an AP where f (·) and {gk (·)} represent the post-processing
and a fusion center in the current phase. For WDA, the server and pre-processing functions (see Table I). Note that the
aims at computing a vector function of the distributed vector pre-processing functions are in general identical across
symbols {dk }K k=1 . Let y = [y1 , y2 , · · · , yL ] ∈ R
L×1
denote different sensors, i.e., gk (·) ≡ g (·), ∀k. For example,
the desired computation output, called the target-function we have gk (x) = x, ∀k for computing arithmetic
vector. Then y is a nomographic function, denoted as h , of K mean, and gk (x) = log(x), ∀k for geometric mean.
simultaneous observations of parameter , namely {dk }K k=1 . Let the vector symbol pre-processed and transmitted
Following the definition in (1), we have by sensor k using linear analog modulation be denoted as

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
3442 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 7, JULY 2019

sk = [g1 (dk1 ), g2 (dk2 ), · · · , gL (dkL )] and the desired vector

T
The above lemma implies that a vanishing MSE(ŝ, s) can
K
s = k=1 sk . To facilitate the power-control design and also lead to a vanishing MSE(ĥ, h), making the former a
reduce transmission power, the symbols are normalized to reasonable surrogate of the latter as the AirComp performance
have zero mean and unit variance, i.e., E{sk sH k } = I, where metric. One can check that this lemma holds for all the typical
the normalization factor for each data type is uniform for all AirComputable functions in Table I.
sensors and can be inverted at the server. The WDA process Substituting (5) into (7), the MSE can be explicitly written
and performance metric are described as follows. as a function of the aggregation beamforming and precoding
Given synchronized symbol boundaries, all sensors transmit as follows:
their vector symbols simultaneously. The distortion of array 
K
observations at the server with respect to the target-function MSE(A, {Bk }) = tr((AH Hk Bk − Bk )
vectors due to channel noise is suppressed using aggregation k=1
beamforming. To this end, the WDA process attempts to ×(AH Hk Bk −Bk )H )+σn2 AH A, (8)
achieve coherent combining of K vector symbols at the
server in each slot. Let A ∈ CNAP ×L denote the aggregation where A and {Bk } are jointly optimized in the sequel under
beamforming matrix and Bk ∈ CNSN ×L the data precoding the criteria of MMSE. Since the transmission energy of each
matrix at sensor k. Then the vector symbol received by the sensor cannot exceed the harvested energy, the transmission
server after AirComp is given by: power constraint for sensor k can be obtained using (3) as

K αk θk T H
k )≤
tr(Bk BH uk Gk vk 2 Pk , (9)
ŝ = A H
Hk Bk sk + A n, H
(5) T0
k=1
where T0 denotes the number of vector symbols transmitted
NAP ×NSN
where Hk ∈ C represents the MIMO chan- within each AirComp phase and the factor θk ∈ (0, 1)
nel for the link from sensor k to the server, and n is represents the fraction of harvested energy allocated for trans-
the additive white Gaussian noise (AWGN) vector with mission. For ease of notation, we denote γk = αkTθ0k T as the
independent and identically distributed (i.i.d.) CN (0, σn2 ) effective power conversion efficiency accounting for all the
elements.
  The corresponding   T is ĥ =
estimated function power discounting factors in the subsequent analysis.
K K
f1 k=1 g1 (dk1 ) , . . . , fL k=1 gL (dkL ) ∈ RL×1 .
Due to the nature of analog transmission, the accuracy IV. J OINT D ESIGN OF WPT AND A IR C OMP :
of AirComp is prone to the distortion by channel fading P ROBLEM F ORMULATION
and noise. The goal of AirComp is to accurately compute The joint design of the WPT and AirComp phases in the
certain functions. Hence, the computation error becomes the WP-AirComp system can be formulated as a joint optimization
natural performance metric. Following the existing literatures problem over wireless power control, energy and aggregation
(see, e.g., [27], [34]), we measure the error using the mean- beamforming. Specifically, given the computation error in (8)
squared-error (MSE) between the estimated function value and the two power constraints in (2) and (9), the problem can
and the ground truth one. be formulated as:
In the current work, the objective is to minimize the MSE

K
between
Kthe estimation Kfunction ĥ and the desired function (P1) min tr((AH Hk Bk −I)(AH Hk Bk −I)H )
h = [ k=1 dk1 , . . . , k=1 dkL ]T ∈ RL×1 , defined as A,{Bk },
 {uk },{vk },{Pk } k=1

MSE(ĥ, h) = E (ĥ − h)(ĥ − h)H . (6) +σn2 tr(AH A)

2
k ) ≤ γk uk Gk vk  Pk , ∀k,
s.t. tr(Bk BH H
However, such a metric limits the applicability of the design
to general nomographic function computation,
 since it requires
 K
K Pk ≤ P0 ,
specifying the component functions f g (d
k=1  k ) . Note
k=1
that for calculating nomographic functions with linear
k uk = 1, ∀k,
uH
post-processing (e.g., arithmetic mean, weighted sum, and
polynomial), minimizing MSE(ĥ, h) is equivalent to minimiz- vk vk = 1, ∀k.
H

ing Problem P1 is difficult to solve due to its non-convexity.


MSE(ŝ, s) = E (ŝ − s)(ŝ − s)H . (7) The lack of convexity arises from the coupling between the
transmit and receive beamformers for both the WPT and
For continuous nomographic functions with nonlinear post- AirComp phases. To simplify the problem and shed light on
processings, the asymptotic equivalence between MSE(ĥ, h) the optimal solution structure, we first consider the simplified
and MSE(ŝ, s) is shown in the following Lemma 1. case of single-antenna sensors, resulting in MISO channels,
Lemma 1 (Asymptotically Equivalent Metrics): According and thus called the MISO case. Note that in the MISO case,
to Cauchy continuity [35], for any continuous function f (·), only scalar-function AirComp is feasible. In the optimization
we have problem for this case, the receive energy beamformers at
E[(f (ŝ)−f(s))(f (ŝ)−f(s)) ] → 0, if E[(ŝ−s)(ŝ−s) ] → 0. sensors varnish and the data precoders reduce to scalars
H H
denoted as {bk }, making the solution tractable. Let the k-th

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
LI et al.: WIRELESSLY POWERED DATA AGGREGATION FOR IoT VIA OVER-THE-AIR FUNCTION COMPUTATION 3443

vector channels for AirComp and WPT be denoted as hk ∈ The name of η is given based on the fact that reducing it
CNAP ×1 and gk ∈ CNAP ×1 respectively, and the aggregation suppresses the effect of channel noise by increasing symbol
beamformer vector as a ∈ CNAP ×1 . Then the original problem energy (or equivalently {b∗n } as observed from (10)). Since
in P1 is simplified for the MISO case as: minimizing η is equivalent to maximizing η1 , one can incor-
porate the power constraint into the objective function and

K
(P2) min aH hk bk − 12 + σn2 aH a have the following equivalent max-min problem:
a,{bk },{uk },{Pk } 2 2
k=1 max min γk uH
k gk  hk f  Pk
H
2 2 f ,{uk },{Pk } k
s.t. bk  ≤ γk uH
k gk  Pk , ∀k,
K 
K

Pk ≤ P0 , s.t. Pk ≤ P0 , f H f = 1,
k=1
k=1
uH = 1, ∀k. k uk = 1, ∀k.
uH
k uk
Thought this problem has a simpler form, it remains non-
Problem P2 is solved in the next section. The insights are
convex. To tackle the challenge, we recast the problem into an
leveraged to develop a practical solution in Section VI. equivalent outer-inner sub-problems formulation, having the
Last, we remark that an alternative and also natural formu-
following form:
lation based on maximization of the receive signal-to-noise
⎧ 2 2
⎫
ratio (SNR) leads to strategies fundamentally different from ⎪
⎪ max min γk uHk gk  hk f  Pk ⎪
H
⎪
computation-error minimization. More details are given in ⎪
⎪ {uk },{Pk } k ⎪
⎪
⎪
⎨ ⎪
⎬
Section VII. K
(P3) max s.t. Pk ≤ P0 ,
⎪
f H f =1 ⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
V. J OINT D ESIGN OF WPT AND A IR C OMP : MISO C ASE ⎪
⎩
k=1 ⎪
⎭
uk uk = 1, ∀k.
H
In this section, we consider the MISO case corresponding   
to single-antenna sensors and solve the design problem in inner sub-problem
  
P2. To this end, a decomposition based solution approach is outer sub-problem
developed. The approach and the solution of problem P2 are
discussed in the following sub-sections. As shown in Fig. 3, the inner sub-problem solves the
optimal power allocation and energy beamforming, while the
outer sub-problem aims at further optimizing the aggregation
A. Decomposition Approach beamformer. The two sub-problems are tractable and solved
Problem P2 remains non-convex and difficult to solve separately in the following sub-sections.
directly. We overcome the difficulty by decomposing it into
two solvable sub-problems. To begin with, channel-inversion B. Optimal Wireless Power Control and Energy Beamforming
precoding conditioned aggregation beamforming is shown to A close observation of the inner sub-problem in P3 indicates
be optimal as follows. that the optimal energy beamforming can be firstly solved in
Lemma 2 (Optimal Sensor Precoder): For the MISO case, closed form, based on which the optimal power allocation can
given a data aggregation beamformer a, the computation error be then derived as a function of the normalized aggregation
is minimized by the following channel-inversion precoders at beamformer f . The first key result of this sub-section is
sensors: presented as follows.
1
b∗k = √ H , ∀k, (10) Proposition 1 (Optimal Energy Transmit Beamformer):
ηf hk Consider the MISO case. For each sensor, the optimal power
where f is defined by normalizing the aggregation beam- beam should point to the direction of the corresponding WPT
former, and η ∈ R+ is chosen to satisfy the sensor trans- channel, i.e.,
gk
mission power constraints in problem P2, i.e., to guarantee
√ u∗k = . (11)
that a = ηf . gk 
 Proof: In the inner sub-problem of P3, uH 2
k gk  ≤
Proof: See Appendix A.
2 2
By substituting the result in Lemma 2, problem P2 can be uk  gk  with the equality holds if and only if uk has
transformed into the equivalent problem of minimizing the the same direction as gk . 
variable η, called the denoising factor: By substituting (11), the inner sub-problem is transformed
min η into the following problem:
η,f ,{uk },{Pk }
1 (P4) max min γk gk 2 hH 2
k f  Pk
2 {Pk } k
s.t. 2
≤ γk uH
k gk  Pk , ∀k,
ηhH
k f  
K

K s.t. Pk ≤ P0 .
Pk ≤ P0 , f H f = 1, k=1
k=1 To solve problem P4, a necessary condition for the optimal
k uk = 1, ∀k.
uH solution is derived as follows.

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
3444 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 7, JULY 2019

Fig. 3. Outer-inner problem in WP-AirComp design and the required signaling procedure.

Lemma 3 (Optimal Wireless Power Control): For the MISO it is important to select sensors with both reliable WPT and
case, given the aggregation beamformer f , the optimal power AirComp channels.
control should have the following form:
C C. Optimal Aggregation Beamforming
Pk∗ = 2 2
, ∀k, (12)
γk hH
k f  gk  Given the solution of the inner sub-problem in the preceding
sub-section, the outer sub-problem in P3 for aggregation
where C is some constant.
beamformer optimization can be simplified by substituting
Proof: See Appendix B. 
the optimal energy beamformer in (11) and optimal power
Based on Lemma 3, problem P4 reduces to the following:
allocation in (14):
max C P0
C max K 1

K f
C k=1 γk f H hk 2 gk 2
s.t. ≤ P0 ,
γk hH 2
k  gk 
f 2 s.t. f H f = 1.
k=1

which is trivial and the solution is Note that f H hk 2 = tr(hk hH

k ff ), an equivalent problem
H

P0 is formulated as the following :

C ∗ = K . (13)
k=1 γk hH
1
2 2

K
1
k f  gk  min
f γk tr(hk hH ff H )g 2 P
k k 0
Then combining (12) and (13) gives the following second key k=1
result of this sub-section. s.t. f f = 1.
H

Proposition 2 (Optimal Power Allocation): For the MISO

Though having a simple structure, this problem is still chal-
case, given the aggregation beamformer f , the optimal power
lenging due to the non-convex norm constraint on f . To tackle
allocation is given by
the constraint, the celebrated SDR technique is applied where
P0 the non-convex constraint is relaxed by its convex hull.
Pk∗ = K . (14)
γk hH 2 2 1 Lemma 4 (Convex Hull Relaxation [37]): Let Ω1 = {X :
k f  gk  k=1 γk hH 2
k f  gk 
2
X = WWH , WH W = Id } and Ω2 = {X : tr(X) = d, 0

Remark 3 (Optimal WPT Strategies): It can be observed X

I}, wherein W is of the size m by d and X has the
form Proposition 1 that the optimal power beam points in dimension of m by m. The second condition 0
X
I
the direction of its corresponding WPT channel. With the means that both X and I − X are positive semi-definite. Then,
beams thus steered, the optimal power allocation attempts to Ω2 is the convex hull of Ω1 , and Ω1 is the set of extreme
equalize the effective close-loop channels of different sensors points of Ω2 .
that cascade the WPT channels specified by the gains gk 2 Thereby relaxing the problem gives:
2
and AirComp channels specified by the gains hH k f .

K
1
Remark 4 (Sensor Scheduling): The current design focuses (P5) min
2
on a set of sensors selected for WDA. The issue of scheduling F̂
k=1 k k F̂)gk 
γ tr(hk hH P0
is not addressed but important. To be specific, deep fading
s.t. tr(F̂) = 1, 0
F̂
I,
in a cascading WPT-AirComp channel will lead to highly
noisy data collected from the corresponding sensor and thereby where F̂ = ff H . Then, the convexity of problem P5 is
amplify the computation error. This is reflected in high established in the following lemma.
transmit power for inverting a poor channel (see (14)) and Lemma 5 (Convexity of Problem P5): Problem P5 is a
thereby reduces the average SNR of aggregated data. Thus, convex problem.

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
LI et al.: WIRELESSLY POWERED DATA AGGREGATION FOR IoT VIA OVER-THE-AIR FUNCTION COMPUTATION 3445

Algorithm 1 Gaussian Randomization Algorithm for MISO The proof of Lemma 6 is similar to that of Lemma 2 shown
WP-AirComp in Appendix A, thus omitted for brevity. Let F denote the
• Initialization: Given an SDR solution F̂∗ , and the num- normalized aggregation beamformer with tr(FFH ) = 1 and
√
ber of random samples M . thus A = ηF with η being the norm of A.Then problem
• Gaussian Random Sampling: P1 can be reduced to the following form by substituting (15):
(1) Perform eigen decomposition [VF̂ , ΣF̂ ] = eig(F̂∗ ).
(2) Generate M random vectors zm ∼ CN (0, I) with min η
η,F,{uk },
0 ∈ CNAP ×1 and I ∈ CNAP ×NAP . {vk },{Pk }
(3) Retrieve M feasible solutions {fm } from {zm } via 1  H −1
 2
VF̂ Σ
1/2 H
zm
s.t. tr (F Hk HH k F) ≤ γk uH
k Gk vk  Pk , ∀k,
fm = F̂
, m = 1, ..., M η
1/2 H
VF̂ Σ zm 
F̂ K
(4) Select the best fm thatleads to the minimum objective, Pk ≤ P0 , tr(FFH ) = 1,
∗
= arg minm k=1 γ tr(h hH fm1f H )g 2 P0 .
K
namely fm k=1
k k k m k
∗
(5) Output f̃ = fm k uk = 1 ∀k, vk vk = 1 ∀k.
uH H
as the approximated optimal normal-
ized aggregation beamformer.
Note that the first set of power constraints in this problem
can be rewritten as:
Proof: See Appendix C.  1 γk uH Gk vk 2 Pk
≤ min  Hk , (16)
Upon solving the problem P5 via a convex problem solver η k tr (F Hk HH F)−1
k
(e.g., the cvx toolbox in MATLAB) and attaining the globally
optimal solution F̂∗ , the next task is to retrieve from it a thus the minimum η in problem P2 is achieved when the above
feasible solution to the original problem denoted by f̃ . Since constraint is active (i.e., the equality holds). Since minimizing
the rank of F̂∗ might be larger than one, the Gaussian random- η is equivalent to maximizing η1 , one can move the power
ization algorithm proposed in [38] can be applied to extract constraint to the objective function and have the following
f̃ from F̂∗ . The main procedure of applying the algorithm to equivalent max-min problem:
the current design is summarized in Algorithm 1. γk uH G v 2 Pk
Remark 5 (Optimality of SDR Solution): It is worth point- max min  k k kH 
F,{uk },{vk },{Pk } k tr (FH Hk Hk F)−1
ing out that the SDR technique can lead to an optimal solution.
If a rank one F̂∗ is solved, the global optimal solution can 
K
s.t. Pk ≤ P0 , tr(FFH ) = 1,
be immediately achieved by extracting the dominant eigen-
vector of F̂∗ as the normalized data aggregation beamformer,
k=1

k uk = 1 ∀k, vk vk = 1 ∀k.
uH H
i.e., f ∗ = [VF̂ ]:,1 . As shown in the simulation later, F̂∗ has a
high probability to be rank one. Similar to the MISO counterpart, this problem can be recast
Remark 6 (Complexity Analysis): According to
as the following outer-inner sub-problems:
[38], the complexity order of Algorithm 1 is
1/2
O(max{NAP , K}4 NAP log(1/)), where , NAP and ⎧ ⎫
⎪
⎪ γk uH Gk vk 2 Pk ⎪
K denotes the solution accuracy, number of AP antennas, ⎪
⎪ max min  k ⎪
⎪
⎪
⎪ {uk },{vk },{Pk } k tr (FH Hk HH F)−1 ⎪
⎪
⎪
and number of sensors, respectively. ⎪
⎪
k ⎪
⎪
⎪
⎨ K ⎪
⎬
(P6) max s.t. Pk ≤ P0 ,
VI. J OINT D ESIGN FOR WPT AND tr(FFH )=1⎪
⎪ ⎪
⎪
⎪
⎪ k=1 ⎪
⎪
A IR C OMP : MIMO C ASE ⎪
⎪ ⎪
⎪
⎪
⎪ uH
uk = 1 ∀k, ⎪
⎪
⎪
⎩
k
⎪
⎭
In the preceding section, we consider the MISO case with vk vk = 1 ∀k,
H
single-antenna sensors. In this section, the results are extended   
to the general MIMO case with multi-antenna sensors. In par- inner sub-problem
  
ticular, the original vector-function WP-AirComp problem in outer sub-problem
problem P1 is solved. The solution builds on the outer-inner
framework developed in the previous section. where the inner sub-problem contains MIMO power control
To further develop the framework for the MIMO case, and energy beamforming optimization, and the MIMO aggre-
the non-convex problem P1 is first simplified by show- gation beamformer design gives the outer sub-problem.
ing the optimality of the zero-forcing (channel inversion)
precoder conditioned on the aggregation beamformer as
A. Optimal Wireless Power Control and Energy Beamforming
follows.
Lemma 6 (Optimal MIMO Precoder): Given an aggrega- We firstly tackle the inner sub-problem in problem P6. The
tion beamformer A, the computation error is minimized by optimal energy transmit and receive beamformers are derived
the following zero-forcing precoders: as shown below.
Proposition 3 (Optimal MIMO Energy Beamformers): For
B∗k = (AH Hk )H (AH Hk HH
k A)
−1
, ∀k. (15) each sensor, the optimal energy transmit/receive beamformers

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
3446 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 7, JULY 2019

should point to the left/right dominant singular vector of the to a sensor k is inversely proportional to its effective close-loop
WPT channel matrix, namely, channel gain that combines the dominant singular value of
umax (Gk ) vmax (Gk ) WPT channel matrix Gk and the beamed AirComp channel
u∗k = and vk∗ = , (17) after aggregation beamforming, i.e., tr (FH H 1HH F)−1 (see
umax (Gk ) vmax (Gk ) ( k k )
Proposition 4).
where umax (Gk ) and vmax (Gk ) denote the left and right
dominant singular vectors of matrix Gk , respectively. B. Optimal Aggregation Beamformer
Proof: Starting from the well-known Rayleigh-quotient
inequality Given the solution of the inner sub-problem presented
above, the outer sub-problem in P6 for aggregation beam-
2 2 2 2
k Gk vk  ≤ uk  σmax (Gk )vk  ,
uH (18) former optimization can be obtained by substituting the
optimal WPT strategies in (17) and (21) into problem P6.
where the equality holds if and only if uk and vk have
It follows that:
the same direction with the left and right dominant singular  
vector umax (Gk ) and vmax (Gk ). Therefore, one can readily 
K
tr (FH Hk HH k F)
−1
min 2
note that the objective in problem P6 can be maximized with F γk σmax (Gk )P0
k=1
respect to uk and vk by setting them as shown in (17), which
completes the proof.  s.t. tr(FFH ) = 1.
By substituting (17) in Lemma 3, the inner sub-problem is This problem is difficult to solve due to the non-convex
−1
transformed into the following problem: objective involving tr((FH Hk HH k F) ). To overcome the
γ σ 2 (G )P difficulty, we adopt the following inequality relaxation of the
(P7) max min  k max Hk k  objective function:
{Pk } k tr (FH Hk Hk F)−1
−1 L

K
tr((FH Hk HHk F) )≤ , (22)
s.t. Pk ≤ P0 , λmin (HH
k FF Hk )
H

k=1 where the equality holds given a well-conditioned channel,

where σmax (Gk ) denotes the dominant singular value of the i.e., the singular values of Hk are identical. Using (22),
WPT channel matrix Gk . To solve problem P7, a necessary a relaxed version of such problem is posed as follows:
condition for the optimal solution is provided as follows. 
K
L
Lemma 7 (Optimal Wireless Power Control): For the min 2
MIMO case, the optimal power control should have the F
k=1
γk σmax (Gk )λmin (HH
k FF Hk )P0
H

following form: s.t. tr(FFH ) = 1.

 −1

∗ Ctr (FH Hk HH k F) According to [38], by introducing a new variable F̂ =
Pk = 2
, ∀k, (19)
γk σmax (Gk ) FFH , an equivalent formulation of such problem is obtained
where C is some constant. as follows:
Proof: See Appendix D.  K
L
Based on Lemma 7, problem P7 reduces to the following: min
2 (G
F k min (Hk F̂Hk )P0
)λ
γ σ H
k=1 k max
max C s.t. tr(F̂) = 1, rank(F̂) = L, F̂  0.
C
 

K
Ctr (FH Hk HH k F)
−1
It can be observed that the only non-convex constraint in this
s.t. ≤ P0 ,
2
γk σmax (Gk ) problem is rank(F̂) = L, thus we may remove it to obtain the
k=1
following relaxed version:
which can be easily solved with the optimal solution:

K
L
∗ P0 (P8) min
C =  . (20) γ σ 2 (Gk )λmin (HH
k F̂Hk )P0
F
K tr((FH Hk HH k
F)−1 ) k=1 k max
k=1 2
γk σmax (Gk )
s.t. tr(F̂) = 1, F̂  0.
Thus the optimal power control policy is given as follows.
Proposition 4 (Optimal MIMO Power Allocation): For the The convexity of problem P8 is shown in the following lemma.
MIMO case, the optimal power allocation is given by Lemma 8 (Convexity of Problem P8): Problem P8 is a con-
 −1
 vex problem.
∗ P0 tr (FH Hk HH k F) Proof: See Appendix E. 
Pk = K tr((FH Hk HH . (21)
2 k F)
−1
) Upon attaining the globally optimal solution of problem
γk σmax (Gk ) k=1 2
γk σ (Gk ) max P8, denoted by F̂∗ , the remaining task is to convert it into
Remark 7 (Optimal WPT for MIMO Case): Similar to its a feasible solution of the original problem, denoted by F̃,
MISO counterpart, (21) indicates that the optimal transmit of rank L. To this end, the Gaussian randomization algorithm
beam points to the left dominant eigen-direction of the WPT for the MISO case in Algorithm 1 is generalized to the MIMO
channel matrix, and the optimal receive beam points to the case for searching the close-to-optimal approximate solution,
right dominant eigen-direction. Moreover, the allocated power as summarized in Algorithm 2.

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
LI et al.: WIRELESSLY POWERED DATA AGGREGATION FOR IoT VIA OVER-THE-AIR FUNCTION COMPUTATION 3447

Algorithm 2 Gaussian Randomization Algorithm for MIMO strategies are fundamentally different. The strategy from SNR
WP-AirComp maximization tends to allocate more power to a sensor with
• Initialization: Given an SDR solution F̂∗ , and the num- better channel condition so as to maximize the sum effective
ber of random samples M . channel gain. In contrast, the strategy from computation error
• Gaussian Random Sampling: minimization in problem P3 attempts to equalize effective
(1) Perform eigen decomposition [VF̂ , ΣF̂ ] = eig(F̂∗ ). channel gain across different sensors (see Proposition 2).
(2) Generate M random matrices Zm ∼ CN (0, I) with This suggests the fundamental difference between WDA and
Zm ∈ CNAP ×L , 0 ∈ CNAP ×L and I ∈ CNAP ×NAP . conventional data communication.
(3) Retrieve M feasible solutions {Fm } from {Zm }
by extracting the L left dominant singular vectors from B. How Does WPT Help AirComp?
1/2
VF̂ ΣF̂ ZH m , denoted by Vm , and normalizing it by
√ As mentioned, wireless power control provides an additional
its norm L, namely Fm = √1L Vm , such that the design dimension for reducing the AirComp error. In the
constraint tr(Fm FHm ) = 1 can be enforced. sequel, we provide insights to understand the performance
(4) Select the best Fm that leads to gain from relevant design presented in the preceding sections.
∗
the minimum
K objective, namely F m = To help exposition, a benchmark scheme is considered that
arg minm k=1 γ σ2 (G )λmin (H LT
H F FH H )P T . equally allocates wireless power, i.e., Pk = PK0 . Consider the
k max k m k0 0
m k

(5) Output F̃ = F∗m as the approximated optimal aggregation beamformer design problem in problem P3. Note
normalized aggregation beamformer. that under equal power allocation, problem P3 is converted to
a NP-hard problem as shown below:
K
(P̃) min max
VII. F URTHER D ISCUSSION f k γk gk 2 hH
k f
2 P0
In this section, we provide further discussion to gain more s.t. f H f = 1.
insights into the properties of the WP-AirComp design in the Comparing the objective functions of problem P̃ and the one
preceding sections. with wireless power control gives:

K
1 K
A. Computation Error Minimization versus SNR ≤ max . (23)
Maximization k=1
γk gk 2 hH
k f  2P
0 k γk g k  2 hH f 2 P
k 0

An intuitive and alternative design criterion for In other words, wireless power control reduces the computa-
WP-AirComp could be one that maximizes the (total) tion error.
receive SNR at the server. This criterion, however, leads
to a completely different strategy from the counterpart that C. AirComp for Digital Communication
minimizes the computation error, as shown in the sequel.
Communication in existing wireless sensor network is
To clarify this point, we consider the simple MISO case for
mainly based on digital modulation technique such as quadra-
example. Let ρ denote the receive SNR and it can be defined
ture amplitude modulation (QAM), which can be understood
based on the channel model in (5) as follows:
as a quantized version of the analog amplitude modulation

K
bk 2 aH hk hH a required for AirComp. Therefore, it is also possible to imple-
ρ= k
ment the AirComp with the more commonly used QAM
σn2 aH a
k=1 modulation. This affects the AirComp by introducing an addi-
1 
K
2 2
tional quantization noise to the computed value. Fortunately,
= γk uH
k gk  hk f  Pk ,
H
the noise effect is expected to vanish as the number of sensors
σn2
k=1
grows due to the implicit noise averaging mechanism in
√
where the second equality is attained by substituting a = ηf , AirComp.
and the power constraint bk 2 = γk uH 2
k gk  Pk . Thus,
the SNR maximization problem can be casted as: VIII. S IMULATION

K In this section, the performance of our proposed
2 2
(Max SNR) max γk uH
k gk  hk f  Pk
H WP-AirComp framework is evaluated by simulation on
f ,{uk },{Pk } MATLAB, where the WPT and AirComp channel models
k=1

K are simulated based on (3) and (5), respectively. The perfor-
s.t. Pk ≤ P0 , mance metric is the normalized computation error, defined by
k=1 MSE/K with MSE given in (8). The simulation parameters
f f = 1,
H are set as follows unless specified otherwise. The number of
sensors is K = 5. In the MISO case, the number of antennas
k uk = 1, ∀k.
uH
at the server is set as NAP = 20. The number of computed
Although the above problem and problem P3 differ slightly functions is set to be L = 1 as only 1 antenna is available
only in the objective function, the resultant resource allocation at sensor side. In the MIMO case, the antenna setting at the

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
3448 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 7, JULY 2019

TABLE III
P ROBABILITY OF SDR S OLUTION B EING O PTIMAL

server is given by NAP = 30, while at sensor side we assume

NSN = 5 antennas for energy receive beamforming or data
precoding, which is equal to the number of computed function
L = 5. All the WPT and Aircomp channels are assumed to
be i.i.d. Rician fading, modeled as i.i.d. complex Gaussian
random variables with non-zero mean μ = 1 and variance
σ 2 = 1. In addition, the maximum transmission power is set
as P0 = 1. The effective power conversion efficiency follows
a uniform distribution with ηn ∈ (0, 1), and the noise variance
is assumed to be 1.

A. Baseline Schemes
We consider three baseline schemes designed based on
two classic approaches: antenna selection (AS) and eigen-
mode beamforming (EB). All three schemes assume the
channel-inversion data precoding in (10) or (15) and also
the optimal energy beamforming in (11) or (17) depending
on whether MISO or MIMO case is considered. The dif-
ference between the three schemes lie in the aggregation
beamformer and the wireless power  allocation policy. Define
the sum-channel matrix Hsum = K k=1 Hk . For the first two
baseline schemes, the optimal power allocation is used by
solving the inner sub-problems of problem P3 and P6 under
the condition that the aggregation beamformer is set to be AS
or EB to enhance the receive SNR. The AS scheme selects
the L receive antenna observing the largest channel gains in
the sum channel Hsum , while the EB scheme selects the L
largest eigenmode of Hsum for AirComp and thus consists of
the L dominant left eigenvectors of Hsum . The third baseline
scheme assumes equal power allocation Pk = PK0 and solves
Fig. 4. The effects of server transmission power on the computation error
the resultant outer sub-problems of problem P5 and P8 to of AirComp.
attain the aggregation beamformer. For fair comparison, all
aggregation beamformers in the baseline schemes are scaled
to have the same norm. observed that the normalized computation error decreases with
the increasing maximum transmission power, which coincides
B. Optimality of the SDR Solution with the intuition that larger transmitting power will result
in higher SNR and thus less estimation error. Moreover,
Consider the MISO case in Section IV, the probability
of SDR solution for problem P5 to be globally optimal, our proposed solution outperforms all other baseline schemes
throughout the whole considered range of transmission power.
i.e., rank(F̂∗ ) = 1, is evaluated under various antenna settings
at the server as summarized in Table III. It can be observed Particularly, there are significant gaps between the base-
that the probability of SDR solution to be optimal is more line schemes of AS, EB and our proposed WP-AirComp
scheme, which demonstrate the merit of solving the outer
than 85% for different settings. Moreover, the probability is
observed to increase with the growth of the antenna numbers at sub-problem for aggregation beamformer optimization. Inter-
the server, which implies that equipping the server with more estingly, the gap between the baseline scheme of EPA and
the WP-AirComp scheme is relatively narrow, suggesting that
antennas will boost the AirComp accuracy to some extent.
optimizing the power allocation is not so critical in terms of
AirComp accuracy.
C. Computation Performance of WP-AirComp Fig. 5 further compares the proposed solution with all three
First, the performance of the normalized computation error baseline schemes under varying number of antennas at the
under varying maximum transmission power are evaluated server. It is observed that, for all four schemes, the normalized
in Fig. 4 for both the MISO and MIMO cases. It can be computation error monotonically decreases with the growth

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
LI et al.: WIRELESSLY POWERED DATA AGGREGATION FOR IoT VIA OVER-THE-AIR FUNCTION COMPUTATION 3449

Fig. 5. The effects of server antenna numbers on the computation error of Fig. 6. The effects of sensor numbers on the computation error of AirComp.
AirComp.

of the antenna numbers at the server due to the increasing

diversity gain. Similar observations to those in Fig. 4 also
apply here, showing that our proposed WP-AirComp scheme
performs consistently well in various settings of the system
parameters.

D. Effects of Different System Parameters on WP-AirComp

Performance
We further quantify the effects of different system parame-
ters on the proposed WP-AirComp framework. Fig. 6 shows
the curves of the normalized computation error versus the
number of sensors under different antenna settings at the
server. One can observe that the computation error increases Fig. 7. The effects of function numbers on the computation error of AirComp.
with the number of sensors but decreases with the number
of antennas at the server. This aligns with our intuition
that more connected sensors makes it harder to design one functions, which indicates that higher computation throughput
common data aggregation beamformer to equalize all different is at a cost of declining accuracy.
sensors’ channels, while having more antennas at the server Fig. 8 further illustrates the effects of imperfect CSI on the
can significant boost the computation accuracy by exploiting computation error of AirComp. The imperfectness of CSI can
the spatial diversity gain. be captured by a random perturbation adding to the of the
Fig. 7 shows the curve of the normalized computation real channel coefficients. We model the random perturbation
error versus the number of computed functions under different as an additive complex Gaussian noise Ne ∼ CN (0, σe2 )
power budget at the server. It can be observed that the with zero mean and variance of σe2 . Three different levels of
computation error increases with the number of computed corruption are considered in this simulation, i.e., σe2 = 0.1,

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
3450 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 7, JULY 2019

and thus achieves the equality in (24), combining with a = ηf

completes the proof.

B. Proof of Lemma 3
Since 0 ≤ Pk ≤ P0 , there exists Pk such that,
∃ C, s.t., γk gk 2 hH 2
k f  Pk = C, ∀k.

If γk gk 2 hH 2
k f  Pk are not equal at the optimal point.
Assume that γmin gmin2 f H hmin 2 Pmin
∗
is the minimum
2 H 2 ∗
and γmax gmax  f hmax  Pmax is the maximum. Let
 ∗  ∗
Pmin = Pmin + δ and Pmax = Pmax − δ, where δ is small
 
enough to ensure that Pmin < Pmax . Without violating the
power constraint,
Fig. 8. The effects of imperfect CSI on the computation error of AirComp.
γmin gmin2 f H hmin 2 Pmin

> γmin gmin2 f H hmin 2 Pmin
∗
.
By such contradiction, equal γk gk 2 f H hk 2 Pk∗ is proved
σe2 = 0.25 and σe2 = 0.5 reflecting the mild, moderate and to be the optimal strategy.
severe CSI estimation error, respectively. It can be observed
that while the computation error of AirComp increases with
C. Proof of Lemma 5
the imperfectness level, our design is robust against imperfect
CSI, especially at the mild and moderate cases. It is worth According to [39], tr(hk hH k F̂) can be regarded as a linear
pointing out that the robustness of the AirComp can be further function of F̂. Since the channel gain cannot be negative,
enhanced by using robust optimization techniques in [40], i.e., hk  0, combining with the constraint 0
F̂
I,
which, however, is out of the scope of this paper. tr(hk hH
k F̂) is always positive. Based on the composition
1
rule of scalar functions, γ tr(h hH F̂)g 2
is convex. Since
k k k k  P0
IX. C ONCLUDING R EMARKS the summation keeps the convexity, the objective function of
In this paper, we have proposed the WP-AirComp frame- problem P5 is convex. Combining with the convex constraints
work for the joint design of wireless power allocation, energy leads to the result.
and aggregation beamforming for effective WDA in IoT
systems. The framework design is tractable via an intelligent D. Proof of Lemma 7
decomposition of the original non-convex problem into an Since 0 ≤ Pk ≤ P0 , there exists Pk such that,
outer-inner form separating the complex design into sub- γ σ 2 (G )P
problems. The decomposition approach not only yields useful ∃ C, s.t.,  k max Hk k  = C, ∀k.
insights into the optimal solution structure, but also approaches tr (FH Hk Hk F)−1
2
globally optimum with a high probability. The additional γk σmax (Gk )Pk
If tr((F Hk HH
are not equal at the optimal point.
design dimension created by wireless power control is shown
H
k F)
−1
)
2 ∗
γ σmax (Gmin )Pmin
to be able to boost the AirComp accuracy. The work points to Assume that tr min is the minimum and
( (FHH HH F)−1
min min )
the promising new research area of WP-AirComp where many 2
γmax σmax (Gmax )Pmax ∗
 ∗
tr((FH Hmax HH −1 ) is the maximum. Let Pmin = Pmin + δ
interesting research issues warrant further investigation, such max F)
 ∗
as sensor scheduling, sensor clustering, and multiple servers and Pmax = Pmax − δ, where δ is small enough to ensure that
 
cooperation. Pmin < Pmax . Without violating the power constraint,
2 
γmin σmax (Gmin )Pmin γ σ 2 (G )Pmin ∗
A PPENDIX   > min max min .
tr (F Hmin Hmin F)
H H −1 tr (F Hmin Hmin F)−1
H H

A. Proof of Lemma 2 γ σ2 (G )P ∗
By such contradiction, equal tr (Fk Hmax k k
is proved to
Given the computation-error minimization objective pro- ( Hk HH
k F)
−1
)
vided in (8), it is easy be the optimal strategy.
K to note that both the first and the second
terms within, i.e., k=1 aH hk bk −12 and aH a are positive.
As a result, for any given data precoder a, we have the E. Proof of Lemma 8
following inequality: According to [39], λmin (HH k F̂Hk ) can be regarded as a

K concave function of F̂. Since the channel gain cannot be
aH hk bk − 12 + σn2 aH a ≥ σn2 aH a. (24) negative, i.e., Hk  0, combining with the constraint 0

k=1 F̂
I, λmin (HHk F̂Hk ) is always positive. Based on the com-
It is easy to verify that setting {bk } to have the zero-forcing position rule of scalar functions, the equivalent MSE function
structure in (10) enforces for each sensor γ σ2 (G )λ L (HH F̂H )P is convex. Since
k max k min k k 0
the summation keeps the convexity, the objective function of

K
aH hk bk − 12 = 0, problem P8 is convex. Combining with the convex constraints
k=1 leads to the result.

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
LI et al.: WIRELESSLY POWERED DATA AGGREGATION FOR IoT VIA OVER-THE-AIR FUNCTION COMPUTATION 3451

R EFERENCES [24] M. Gastpar, “Uncoded transmission is exactly optimal for a simple

Gaussian ‘Sensor’ network,” IEEE Trans. Inf. Theory, vol. 54, no. 11,
[1] A. Giridhar and P. R. Kumar, “Toward a theory of in-network computa- pp. 5247–5251, Nov. 2008.
tion in wireless sensor networks,” IEEE Commun. Mag., vol. 44, no. 4, [25] A. B. Wagner, S. Tavildar, and P. Viswanath, “Rate region of the
pp. 98–107, Apr. 2006. quadratic Gaussian two-encoder source-coding problem,” IEEE Trans.
[2] H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and Inf. Theory, vol. 54, no. 5, pp. 1938–1961, May 2008.
B. A. Y. Arcas. (2016). “Communication-efficient learning of deep net- [26] R. Soundararajan and S. Vishwanath, “Communicating linear functions
works from decentralized data.” [Online]. Available: https://arxiv.org/ of correlated Gaussian sources over a MAC,” IEEE Trans. Inf. Theory,
abs/1602.05629 vol. 58, no. 3, pp. 1853–1860, Mar. 2012.
[3] M. di Bernardo, A. Salvi, and S. Santini, “Distributed consensus strategy [27] J.-J. Xiao, S. Cui, Z.-Q. Luo, and A. J. Goldsmith, “Linear coherent
for platooning of vehicles in the presence of time-varying heterogeneous decentralized estimation,” IEEE Trans. Signal Process., vol. 56, no. 2,
communication delays,” IEEE Trans. Intell. Transp. Syst., vol. 16, no. 1, pp. 757–770, Feb. 2008.
pp. 102–112, Feb. 2015. [28] G. Zhu and K. Huang, “MIMO over-the-air computation for
[4] K. W. Choi, L. Ginting, P. A. Rosyady, A. A. Aziz, and D. I. Kim, high-mobility multi-modal sensing,” IEEE IoT J., to be published.
“Wireless-powered sensor networks: How to realize,” IEEE Trans. [29] O. Abari, H. Rahul, D. Katabi, and M. Pant, “AirShare: Distributed
Wireless Commun., vol. 16, no. 1, pp. 221–234, Jan. 2017. coherent transmission made seamless,” in Proc. IEEE INFOCOM,
[5] Y. Zeng, B. Clerckx, and R. Zhang, “Communications and signals design Hong Kong, Apr./May 2015, pp. 1742–1750.
for wireless power transmission,” IEEE Trans. Commun., vol. 65, no. 5, [30] B. Nazer and M. Gastpar, “Compute-and-forward: Harnessing
pp. 2264–2290, May 2017. interference through structured codes,” IEEE Trans. Inf. Theory, vol. 57,
[6] K. Huang and X. Zhou, “Cutting the last wires for mobile communi- no. 10, pp. 6463–6486, Oct. 2011.
cations by microwave power transfer,” IEEE Commun. Mag., vol. 53, [31] J. Zhan, B. Nazer, U. Erez, and M. Gastpar, “Integer-forcing linear
no. 6, pp. 86–93, Jun. 2015. receivers,” IEEE Trans. Inf. Theory, vol. 60, no. 12, pp. 7661–7685,
[7] S. Bi, C. K. Ho, and R. Zhang, “Wireless powered communication: Dec. 2014.
Opportunities and challenges,” IEEE Commun. Mag., vol. 53, no. 4, [32] B. Nazer and M. Gastpar, “Reliable physical layer network coding,”
pp. 117–125, Apr. 2015. Proc. IEEE, vol. 99, no. 3, pp. 438–460, Mar. 2011.
[8] M. Xia and S. Aissa, “On the efficiency of far-field wireless power [33] J. Gubbi, R. Buyya, S. Marusic, and M. Palaniswami, “Internet of
transfer,” IEEE Trans. Signal Process., vol. 63, no. 11, pp. 2835–2847, Things (IoT): A vision, architectural elements, and future directions,”
Jun. 2015. Future Generat. Comput. Syst., vol. 29, no. 7, pp. 1645–1660, Sep. 2013.
[9] J. Park and B. Clerckx, “Joint wireless information and energy transfer [34] A. Klein, G. K. Kaleh, and P. W. Baier, “Zero forcing and minimum
in a two-user MIMO interference channel,” IEEE Trans. Wireless mean-square-error equalization for multiuser detection in code-division
Commun., vol. 12, no. 8, pp. 4210–4221, Aug. 2013. multiple-access channels,” IEEE Trans. Veh. Technol., vol. 45, no. 2,
[10] G. Yang, C. K. Ho, R. Zhang, and Y. L. Guan, “Throughput optimization pp. 276–287, May 1996.
for massive MIMO systems powered by wireless energy transfer,” IEEE [35] H. L. Royden and P. Fitzpatrick, Real Analysis. New York, NY, USA:
J. Sel. Areas Commun., vol. 33, no. 8, pp. 1640–1650, Aug. 2015. Macmillan, 1998.
[11] D. W. K. Ng, E. S. Lo, and R. Schober, “Wireless information and power [36] M. Kamel, W. Hamouda, and A. Youssef, “Ultra-dense networks:
transfer: Energy efficiency optimization in OFDMA systems,” IEEE A survey,” IEEE Commun. Surveys Tuts., vol. 18, no. 4, pp. 2522–2545,
Trans. Wireless Commun., vol. 12, no. 12, pp. 6352–6370, Dec. 2013. 4th Quart., 2016.
[12] M. Zhang and Y. Liu, “Energy harvesting for physical-layer security in [37] M. L. Overton and R. S. Womersley, “On the sum of the largest
OFDMA networks,” IEEE Trans. Inf. Forensics Security, vol. 11, no. 1, eigenvalues of a symmetric matrix,” SIAM J. Matrix Anal. App., vol. 13,
pp. 154–162, Jan. 2016. no. 1, pp. 41–45, Jan. 1992.
[13] P. Popovski, A. M. Fouladgar, and O. Simeone, “Interactive joint transfer [38] Z.-Q. Luo, W.-K. Ma, A. M.-C. So, Y. Ye, and S. Zhang, “Semidefinite
of energy and information,” IEEE Trans. Commun., vol. 61, no. 5, relaxation of quadratic optimization problems,” IEEE Signal Process.
pp. 2086–2097, May 2013. Mag., vol. 27, no. 3, pp. 20–34, May 2010.
[14] Z. Ding, I. Krikidis, B. Sharif, and H. V. Poor, “Wireless information [39] S. P. Boyd and L. Vandenberghe, Convex Optimization. Cambridge,
and power transfer in cooperative networks with spatially random U.K.: Cambridge Univ. Press, 2004.
relays,” IEEE Trans. Wireless Commun., vol. 13, no. 8, pp. 4440–4453, [40] M. Hong and Z.-Q. Luo, “Signal processing and optimal resource
Aug. 2014. allocation for the interference channel,” Academic Press Library Signal
[15] C. Zhong, H. A. Suraweera, G. Zheng, I. Krikidis, and Z. Zhang, Process., vol. 2, pp. 409–469, Jan. 2014.
“Wireless information and power transfer with full duplex relaying,”
IEEE Trans. Commun., vol. 62, no. 10, pp. 3447–3461, Oct. 2014.
[16] Z. Zheng, L. Song, D. Niyato, and Z. Han, “Resource allocation in
wireless powered relay networks: A bargaining game approach,” IEEE
Trans. Veh. Technol., vol. 66, no. 7, pp. 6310–6323, Jul. 2017. Xiaoyang Li (S’13) received the B.S. degree in elec-
[17] D. W. K. Ng, E. S. Lo, and R. Schober, “Multiobjective resource trical and electronic engineering from the Southern
allocation for secure communication in cognitive radio networks with University of Science and Technology (SUSTech)
wireless information and power transfer,” IEEE Trans. Veh. Technol., in 2016. He is currently pursuing the Ph.D. degree
vol. 65, no. 5, pp. 3166–3184, May 2016. in electrical and electronic engineering with The
[18] C.-F. Liu, M. Maso, S. Lakshminarayana, C.-H. Lee, and T. Q. S. Quek, University of Hong Kong (HKU). His research
“Simultaneous wireless information and power transfer under different interests include wireless power transfer, mobile
CSI acquisition schemes,” IEEE Trans. Wireless Commun., vol. 14, no. 4, crowd sensing, over-the-air computation, and convex
pp. 1911–1926, Apr. 2015. optimization.
[19] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput maximization for
UAV-enabled mobile relaying systems,” IEEE Trans. Commun., vol. 64,
no. 12, pp. 4983–4996, Dec. 2016.
[20] J. Xu, Z. Zhong, and B. Ai, “Wireless powered sensor networks: Col-
laborative energy beamforming considering sensing and circuit power Guangxu Zhu (S’14) received the B.S. and M.S.
consumption,” IEEE Wireless Commun. Lett., vol. 5, no. 4, pp. 344–347, degrees in information and communication engineer-
Aug. 2016. ing from Zhejiang University in 2012 and 2015,
[21] G. Zhu, S.-W. Ko, and K. Huang, “Inference from randomized trans- respectively. He is currently pursuing the Ph.D.
missions by many backscatter sensors,” IEEE Trans. Wireless Commun., degree with the Department of electrical and elec-
vol. 17, no. 5, pp. 3111–3127, May 2018. tronic engineering, The University of Hong Kong.
[22] O. Abari, H. Rahul, and D. Katabi. (2016). “Over-the-air function His research interests include edge intelligence,
computation in sensor networks.” [Online]. Available: https://arxiv.org/ distributed machine learning, and 5G technologies,
abs/1612.02307 such as massive MIMO, mmWave communication,
[23] B. Nazer and M. Gastpar, “Computation over multiple-access chan- and wirelessly powered communication. He was a
nels,” IEEE Trans. Inf. Theory, vol. 53, no. 10, pp. 3498–3516, recipient of the Hong Kong Postgraduate Fellow-
Oct. 2007. ship (HKPF) and the Best Paper Award from WCSP 2013.

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.
3452 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 7, JULY 2019

Yi Gong (S’99–M’03–SM’07) received the Kaibin Huang (S’05–M’08–SM’13) received the

Ph.D. degree in electrical engineering from The Ph.D. degree in electrical engineering from The
Hong Kong University of Science and Technology, University of Texas at Austin. Since 2014, he has
Hong Kong, in 2002. He was with the Hong Kong been an Assistant Professor with the Department of
Applied Science and Technology Research Institute, electrical and electronic engineering, The Univer-
Hong Kong, and Nanyang Technological University, sity of Hong Kong. His research interests include
Singapore. He is currently a Professor with the MIMO communications, wireless power transfer,
Southern University of Science and Technology, and machine learning. He frequently serves on the
Shenzhen, China. His research interests include technical program committees of the major IEEE
cellular networks, mobile computing, and signal conferences in wireless communications. He was a
processing for wireless communications and related recipient of the Best Paper Award from the IEEE
applications. From 2006 to 2018, he served on the Editorial Board of the GLOBECOM in 2006 and the IEEE Communications Society Asia–Pacific
IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS and the IEEE Outstanding Paper Award in 2015. Most recently, he served as the Lead Chair
T RANSACTIONS ON V EHICULAR T ECHNOLOGY. for the Wireless Communications Symposium of the IEEE GLOBECOM
2017, the Communication Theory Symposium of the IEEE GLOBECOM
2014, the TPC Co-Chair for the IEEE PIMRC 2017, and the IEEE CTW 2013.
He was an Editor of the IEEE J OURNAL ON S ELECTED A REAS IN C OM -
MUNICATIONS Series on Green Communications and Networking and the
Special Issue on Communications Powered by Energy Harvesting, the IEEE
W IRELESS C OMMUNICATIONS L ETTERS , and the IEEE/KICS J OURNAL OF
C OMMUNICATION AND N ETWORKS . He is currently an Editor of the IEEE
T RANSACTIONS ON G REEN C OMMUNICATIONS AND N ETWORKING and the
IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS.

Authorized licensed use limited to: Sungkyunkwan University. Downloaded on August 10,2022 at 15:50:26 UTC from IEEE Xplore. Restrictions apply.