1

Integrated Sensing, Communication, and Compu-

tation Over-the-Air: MIMO Beamforming Design
Xiaoyang Li, Fan Liu, Ziqin Zhou, Guangxu Zhu, Shuai Wang, Kaibin Huang,
and Yi Gong

Abstract
arXiv:2201.12581v2 [cs.IT] 22 Feb 2022

To support the unprecedented growth of the Internet of Things (IoT) applications, tremendous data
need to be collected by the IoT devices and delivered to the server for further computation. By utilizing
the same signals for both radar sensing and data communication, the integrated sensing and communica-
tion (ISAC) technique has broken the barriers between data collection and delivery in the physical layer.
By exploiting the analog-wave addition in a multi-access channel, over-the-air computation (AirComp)
enables function computation via transmissions in the physical layer. The promising performance of
ISAC and AirComp motivates the current work on developing a framework called integrated sensing,
communication, and computation over-the-air (ISCCO). Two schemes are designed to support multiple-
input-multiple-output (MIMO) ISCCO simultaneously, namely the separated and shared schemes. The
separated scheme splits antenna array for radar sensing and AirComp respectively, while all the antennas
transmit a joint waveform for both radar sensing and AirComp in the shared scheme. The performance
metrics of radar sensing and AirComp are evaluated by the mean squared errors of the estimated target
response matrix and the received computation results, respectively. The design challenge of MIMO
ISCCO lies in the joint optimization of beamformers for sensing, communication, and computation at
both the IoT devices and the server, which results in a non-convex problem. To solve this problem, an
algorithmic solution based on the technique of semidefinite relaxation is proposed. The results reveal
that the beamformer at each sensor needs to account for supporting dual-functional signals in the shared
scheme, while dedicated beamformers for sensing and AirComp are needed to mitigate the mutual
interference between the two functionalities in the separated scheme. The use case of target location
estimation based on ISCCO is demonstrated in simulation to show the performance superiority.

Index Terms

Xiaoyang Li, Fan Liu, Ziqin Zhou, and Yi Gong are with the the Department of Electrical and Electronic Engineering (EEE),
Southern University of Science and Technology (SUSTech), Shenzhen, China. Guangxu Zhu is with the Shenzhen Research
Institute of Big Data, Shenzhen, China. Shuai Wang is with the Shenzhen Institute of Advanced Technology, Chinese Academy
of Sciences, Shenzhen, China. Kaibin Huang is with the Department of EEE, The University of Hong Kong, Hong Kong.
Corresponding author: Yi Gong (e-mail: gongy@sustech.edu.cn).
 2

Integrated sensing and communication (ISAC), over-the-air computation (AirComp), multiple-input

multiple-output (MIMO), beamforming.

I. I NTRODUCTION

The next-generation wireless networks (6G and beyond) have been envisioned as a vital enabler
for the emerging Internet of Things (IoT) services, such as autonomous vehicles, extended reality,
artificial intelligence, smart cities, and digital twins [1]. To support the prosperous applications,
tremendous data need to be collected from the environment by the IoT devices and delivered
to the server for further fusion and computation [2]. In conventional data processing pipelines,
the above operations are individually accomplished with limited mutual assistance and rare
integration [3]. By utilizing the same spectrum and signals for both radar sensing and data
communication, the integrated sensing and communication (ISAC) technique has broken the
barriers between data collection and delivery in the physical layer [4]. Nevertheless, the operation
of computation is still isolated as it mainly lies in the upper layers.
For reducing the overheads and improving the efficiency, it is natural to integrate the operations
of sensing, communication, and computation together. By exploiting the analog-wave addition in
a multi-access channel, over-the-air computation (AirComp) realizes fast wireless data fusion via
simultaneous transmissions in the physical layer [5]. Prompted by AirComp, the three operations
are expected to be unified within a single signal transmission, which motivates the current work
on developing a framework called integrated sensing, communication, and computation over-
the-air (ISCCO).
In the ISCCO framework, an IoT system is considered for supporting sensing, communication,
and computation simultaneously. To be specific, multiple multi-antenna IoT sensors transmit radar
signals to detect a target and data symbols to a multi-antenna server for data fusion via AirComp.
The dual-functional radar sensing and AirComp can be achieved by two schemes, namely shared
scheme and separated scheme. In the shared scheme, the whole antenna array at each sensor
are exploited for transceiving a signal both serving as a radar pulse and a data carrier. In the
separated scheme, the antennas at each sensor are divided into two groups for radar sensing and
data transmission. As the key performance indicators (KPIs) for radar sensing and AirComp are
the mean squared errors (MSE) of target estimation and function computation, respectively, there
exists a natural tradeoff between the performance of the two functionalities, which is reflected
in the beamforming design. This introduces coupling between sensing and AirComp, and hence
 3

necessitates the joint design of radar signal beamforming, data transmission beamforming, and
data aggregation beamforming. The beamforming designs together with the performance analysis
for both the shared and separated schemes are investigated in this paper.
The contributions of this work are summarized as follows.
• Transmission and Aggregation Beamforming Design in the Shared Scheme: As the
signal transmitted by each sensor in the shared scheme serves both as the radar probing
pulse and data carrier, only one beamformer needs to be designed at each sensor, which is
known as the transmission beamformer. The radar sensing target can be extracted from the
statistic information of the reflected signal via maximum likelihood estimation (MLE). As for
AirComp, a beamformer at the server is deployed for equalizing the received signals, which
is known as the data aggregation beamformer. The joint transmission and data aggregation
beamforming design is formulated as a semidefinite programming problem for minimizing
the computation error in AirComp under the constraints of radar sensing requirement and
power budget for each sensor. The solving approach based on semidefinite relaxation is
applied to obtain the desired design.
• Radar and Communication Beamforming Design in the Separated Scheme: As there
are two signals transmitted by each sensor with one for radar sensing and another for
data transmission, two corresponding beamformers are needed to be designed at each
sensor. Moreover, the existence of radar signal will exacerbate the interference at the
server, which results in a more complex performance metrics of AirComp. The coupling
between radar sensing beamformer, data transmission beamformer, and data aggregation
beamformer makes the optimization problem even challenging to be coped with. To tackle
such a problem, an orthogonal constraint for radar sensing beamforming is imposed in the
beamforming design.
• Comparison between Shared and Separated Schemes: Simulation is conducted to com-
pare the performance of the shared and separated schemes. Particularly, the beamformers
in the shared scheme try to mitigate the noise in AirComp while guaranteeing the sensing
accuracy. In the separated scheme, the beamformers are further accounted for reducing the
interference of radar signals on AirComp. The radar sensing MSE in the separated scheme
solely depends on the maximum MSE tolerance for reducing the AirComp error, while the
sensing MSE in the shared scheme is relevant to multiple parameters such as the number
of antennas.
 4

• Target Location Estimation based on ISCCO: To illustrate the performance of ISCCO

design, the use case of target location estimation is conducted. Specifically, the target
location is estimated locally by each sensor based on the information extracted from its
received radar signals and transmitted to the server via AirComp. The averaged estimated
target location received by the server is compared with the ground truth.
The remainder of the paper is organized as follows. Section II reviews the state-of-the-art
techniques of AirComp and ISAC. Section III introduces the ISCCO system model. Sections
IV and V present the shared and separated beamforming designs for achieving the best ISCCO
performance, respectively. Section VI further illustrates the performance of target location esti-
mation based on ISCCO. Simulation results are provided in Section VII, followed by concluding
remarks in Section VIII.

II. BACKGROUND OF A IR C OMP AND ISAC

A. AirComp

The idea of AirComp can be traced back to the pioneering work studying function compu-
tation in sensor networks [6], where the distributed sensing values are analogly modulated and
transmitted over a multi-access channel for reliable function computation at a server. The impor-
tance of the work lies in the counter-intuitive finding that interference caused by simultaneous
transmission can be exploited to facilitate computation. The transmission synchronization over
sensors was further investigated in [7].
Driven by the need of fast data aggregation in IoT, AirComp has been applied to supporting
function computation via data transmission from multiple sensors to the server. The functions
that can be calculated by AirComp has the general format h(·) as shown below:
K
!
X
y = h(x1 , x2 , · · · , xK ) = f gk (xk ) , (1)
k=1

where {xk } represents the distributed data samples, f (·) and gk (·) represent post-processing
at the server and pre-processing at a device, respectively. The summation in (1) is achieved
by simultaneous analog transmission to exploit the wave-addition of the multi-access channel.
Consequently, the function computation is performed “over-the-air” and the result is directly
received by the server. The class of functions having the above form is known as nomographic
functions such as averaging and geometric mean. Typical functions in this class are summarized
 5

Table I: Examples of nomographic functions in AirComp.

Name Expression
1
PK
Arithmetic Mean y= K k=1 xk
PK
Weighted Sum y = k=1 ωk xk
Q 1/K
K
Geometric Mean y= k=1 xk
βk
y= K
P
Polynomial k=1 ωk xk
qP
K 2
Euclidean Norm y= k=1 xk

in Table I. Simultaneous transmission in AirComp achieves low latency independent of the

number of devices and saves the spectrum resources.
To accelerate the computation of multiple functions, the features of multi-modal sensing as
well as the prevalence of antenna arrays at both servers and devices were exploited to enable the
multiple input multiple output (MIMO) AirComp [8]. In MIMO AirComp, the spatial degrees-
of-freedom is leveraged to spatially multiplex multi-function computation simultaneously and
reduce computation errors by noise suppression [9]. Along this vein, the MIMO AirComp was
integrated with the wireless power transfer technique to achieve self-sustainable AirComp for
low-power devices [10]. The reduced-dimension design of AirComp was investigated in [11]
for clustered IoT networks. To overcome the reliance on channel station information (CSI), a
blind MIMO AirComp technique without requiring CSI access was proposed in [12] for low-
complexity and low-latency IoT networks. Facing the practical scenario with fading channels,
the optimal power control was designed in [13] to deal with the channel distortion, while the
tradeoff between the computation effectiveness and the energy efficiency was analyzed in [14].
The hybrid beamforming for massive MIMO AirComp was studied in [15].
Due to its promising performance in fast function computation, AirComp has been deployed
in a series of IoT applications, including federated edge learning (FEEL) [16]–[24], reconfig-
urable intelligent surface (RIS) assisted communication [25], unmanned aerial vehicle (UAV)
communication [26], autonomous driving [27], and MapReduce over the edge cloud network
[29]. As for FEEL, the wave-addition of AirComp is in perfect match with the aggregation of
local training results at the server for global model updating, which has attracted great research
efforts. The existing literature have investigated the implementation of AirComp in FEEL from
different perspectives, including communication-learning tradeoff [16], devices scheduling [17],
update compression [18], beamforming design [19], hyper-parameters control [20], learning rate
 6

control [21], power control [22], digital modulation [23], and data privacy [24].
Despite the wide applications of AirComp in learning and communication systems, the in-
corporation of data sensing remains as an uncharted area in AirComp, which deserves to be
investigated in this paper.

B. ISAC

The origin of ISAC can be traced back to the early work in joint radar-communication, where
information was embedded into a group of radar pulses [30]. In practice, the S-band (2-4 GHz)
and C-band (4-8 GHz) occupied by radar applications might be shared with communication
systems [31]. Consequently, a series of treatises focuses on investigating the co-existence of
radar and communications systems. Particularly, an opportunistic spectrum sharing scheme was
proposed in [32], where the communication signals are sent when the spectrum is not occupied
by radar. Despite its easy implementation, the radar and communication functions are unable to
work simultaneously. To overcome such a drawback, a null-space projection method was carried
out to support the co-existence of MIMO radar sensing and communication [33], where the radar
signals are projected onto the null-space of the interference channel for the communication link.
Nevertheless, such projection might harm the optimality of radar signal beamforming and thus
results in performance loss for the radar sensing.
Research has been conducted on improving the performance of radar sensing and communi-
cation based on convex optimization techniques. In [34], the designs of radar beamformer and
communication covariance matrix were jointly optimized to maximize the radar sensing signal
to interference plus noise ratio (SINR) subject to specific capacity and power constraints. As
for co-existence of MIMO radar and multi-user MIMO (MU-MIMO) communications, a robust
beamforming design with imperfect CSI was proposed in [35], where the radar sensing accuracy
is maximized under the SINR requirements by communication and the power budget. Moreover,
the multi-user interference was exploited as a source of transmission power in [36], based on
which a novel beamforming design was proposed. In order to develop an optimal communication
reception strategy in the presence of the radar interference, a communication receiver was
designed in [37] to demodulate the communication data while removing the radar interference
iteratively using a successive interference cancellation (SIC) algorithm. It should be noted that
the side-information including CSI, radar probing waveforms, and communication modulation
formats need to be frequently exchanged between the radar and communication devices to
 7

support the coexistence. Though such cooperation might be achieved by deploying a control
center connecting both systems via a wireless link or a backhaul channel, the implementation
will impose extra complexity on the system [38].
To reduce the side-information exchange overheads, an advanced co-existence scheme was
proposed in [39], where a dual-functional system supporting both radar and communications
was designed. From the perspective of information theory, the performance of radar and com-
munications were unified based on the rate distortion theory [40]. The implementation in practice
was conducted by the dual-functional waveform design, which supports target detection as well as
data transmission simultaneously [41]. Along this vein, the integrated radar and communication
waveform was designed in single antenna systems [42]. As a step forward, the work of [43]
brought the integrated waveform design into the MIMO systems, where the information bits are
embedded in the sidelobe of the radar transmitting beampattern. Accounting for the multi-user
communication system, a series of transmitting beamforming designs were carried out in [44]
w.r.t. both shared and separated schemes. Aiming at reducing signal distortion, the constant
modulus waveforms was further conceived in the dual-functional beamforming design [45].
The benefit of spectrum sharing makes ISAC a popular technology that has beed applied
in a series of systems, such as millimeter-wave radar and communication networks [46], RIS
systems [47], smart home [48], edge learning systems [49], vehicular networks [50], and UAV
systems [51]. Particularly, an IEEE 802.11ad-based radar was deployed in millimeter-wave band
for supporting an automotive radar and communication network [46]. To mitigate the multi-user
interference in ISAC, the joint waveform and discrete phase shift design was carried out by
employing RIS [47]. As for smart home, the traditional sensing devices were empowered with
communication capability, while the sensing capability of WiFi signals were enhanced [48]. ISAC
was further applied to accelerate the edge learning process by designing wireless signals for the
dual purposes of dataset generation and uploading. In vehicular networks, the wireless sensing
functionality was exploited to acquire vehicles’ states and facilitate the communication [50]. In
UAV-enabled ISAC system, the maneuver and beamforming designs were jointly optimized to
communicate with multiple users and sense potential targets simultaneously [51].
Among the rich literature on ISAC, the operation of data computation is always overlooked as it
lies in the upper layers. Fortunately, AirComp enables fast function computation via transmissions
in the physical layer. Therefore, it is natural to integrate the operations of sensing, communication,
and computation via the combination of ISAC and AirComp.
 8

Table II: Notation.

Symbol Definition Symbol Definition

M Number of sensors Gim Target response matrix between sensor i and m
K Number of AirComp functions Hm Data transmission channel from sensor m to AP
Na Number of antennas at the AP Qim Direct radar channel between sensor i and m
Ns Number of antennas at each sensor Rm Radar signal channel from sensor m to AP
nc Noise of data transmission channel ηm Sensing MSE threshold of sensor m
T Number of time slots in ISCCO phase Cim Data reflection channel between sensor i and m
P Power budget of each sensor Oim Direct data channel between sensor i and m
sm [t] Radar signals of sensor m at time t Ωm [t] Interference signal for sensor m at time t
dm [t] Data symbols of sensor m at time t Ŷm Sufficient statistic matrix of sensor m
xm [t] Signals transmitted by sensor m at time t ym [t] Target reflection signal of sensor m at time t
nr Noise of radar signal channel Nc Number of antennas at sensor for data transmission
Wm Data transmission beamformer of sensor m Nr Number of antennas at sensor for radar sensing
Fm Radar signal beamformer of sensor m Ntx Number of radar transmitting antennas at sensor
A Data aggregation beamformer at AP Nrx Number of radar receiving antennas at sensor

III. S YSTEM M ODEL

The considered MIMO ISCCO system comprises one common target, one access point (AP)
equipped with Na antennas, and M = |M| radar sensors equipped with Ns antennas. Each radar
sensor can simultaneously transmit probing signals to detect the target and data symbols to the
AP for AirComp. The ISCCO phase is divided into T time slots. The operations of different
sensors are synchronized using a reference clock broadcast by the server (see e.g., [7]). The CSI
between the AP and each sensor is estimated individually at each sensor from broadcasted pilot
signals and then passed to the AP subsequently. For simplicity, channels are assumed to vary
following the block-fading model. In other words, each channel remains fixed within a phase
and varies over different phases. The maximum transmit power of each sensor is P . Two ISCCO
schemes are considered in this paper, namely the shared scheme and the separated scheme. The
notations are summarized in Table II.

A. Shared Scheme

As shown in Fig. 1 (a), all the Ns = Ntx + Nrx antennas at each sensor are shared for both
radar sensing and data transmission, where Ntx antennas are for signal transmitting and Nrx for
receiving. The signals transmitted by the sensors have dual functions for both sensing the target
 9

Ns Nr Nc
G11 H1 G11 H1

G22 Ns Na G22 Nr Nc Na
H2 H2

Target GMM Target GMM

HM HM
Ns Nr Nc
Radar sensing Radar sensing
AirComp AirComp
Sensor Access point Sensor Access point
(a) Shared design (b) Separated design

Figure 1: Integrated sensing and AirComp system.

and carrying the data symbols to the AP simultaneously. The data symbols to be transmitted
by the m-th sensor at the t-th slot can be expressed as a vector denoted by sm [t] ∈ CK×1 ,
where K represents the number of functions to be calculated via AirComp. The data symbols
are assumed to be i.i.d. among different sensors and functions with zero mean and unit variance,
i.e., Et [sm [t]sH H
m [t]] = I and Et [sm [t]si [t]] = 0, ∀i 6= m. The diagram of shared scheme is shown

in Fig. 2 and described as follows.

Transmission AirComp
beamformer channel
Data W1 H1
symbol
s1[t]
nr Data
AWGN aggregation
Ĝ 11 beamformer
Estimated Target AWGN A
parameter Ω1 G11 response nc
Interference matrix AirComp
output
WM HM

sM[t]
nr

Ĝ MM

ΩM GMM

Figure 2: Diagram of shared scheme.

By applying a transmit beamformer Wm ∈ CNtx ×K , the signal xm [t] ∈ CNtx ×1 transmitted by

 10

the m-th sensor can be expressed as

xm [t] = Wm sm [t]. (2)

Due to the limited transmit power of each sensor, the beamformer design should satisfy the
power constraint:
tr(Wm Wm H ) ≤ P, ∀m. (3)

The target reflection signal ym [t] ∈ CNrx ×1 received at the m-th sensor can be expressed as

ym [t] = Gmm Wm sm [t] + Ωm [t] + nr [t], (4)

P P
where Ωm [t] = i∈M/{m} Gim Wi si [t]+ i∈M/{m} Qim Wi si [t] is the interference signal, Gim ∈
CNrx ×Ntx and Qim ∈ CNrx ×Ntx denotes the target response matrix (TRM) and direct radar
channel (DRC) between the i-th and m-th sensors, respectively. nr [t] ∈ CNrx ×1 is an additive
white Gaussian noise (AWGN) vector with distribution NNrx (0, σr2 I).
According to [52], the sufficient statistic matrix Ŷm ∈ CNrx ×K can be attained by matched
filtering (a.k.a. pulse compression), which is
T
1X
Ŷm = ym [t]sH
m [t]
T t=1
T M T T
1 XX H 1X X 1X
= Gim Wi si [t]sm [t] + Qim Wi si [t]sH
m [t] + nr [t]sH
m [t]. (5)
T t=1 i=1 T t=1 T t=1
i∈M/{m}

According to the law of large-number, when the number of slots T is large, one can get
T
1X
si [t]sH H
m [t] ≈ Et [si [t]sm [t]] = 0, (6)
T t=1
T
1X
sm [t]sH H
m [t] ≈ Et [sm [t]sm [t]] = I. (7)
T t=1
Therefore, the sufficient statistic matrix can be expressed as

Ŷm = Gmm Wm + Nm , (8)

1
PT Nrx ×K
where Nm = T t=1 nr [t]sH
m [t] ∈ C denotes the statistic noise matrix. The distribution of
Nm is given in the lemma below.

Lemma 1 (Distribution of Nm ). Nm ∼ MN Nrx ×K (0, √σrT INrx ×Nrx , √σrT IK×K )

Proof: See Appendix A. 

 11

The corresponding probability density function (PDF) of Ŷm is

1 − T 2 tr[(Ŷ m −Gmm Wm )H (Ŷ m −Gmm Wm )]
p(Ŷm ; Gmm ) = e 2σr . (9)
(2πσr2 /T )Nrx K/2
Therefore, the MLE of Gmm can be found by minimizing the log-likelihood function

L(Ĝmm ) = tr[((Ŷm − Gmm Wm )H (Ŷm − Gmm Wm )]. (10)

The derivatives of L(Ĝmm ) w.r.t. Ĝmm is

∂L(Ĝmm ) H H
= 2Ĝmm Wm Wm − 2Ŷm Wm . (11)
∂ Ĝmm
By setting the derivatives of L(Ĝmm ) w.r.t. Ĝmm as zero, the MLE of Gmm can be obtained as
H H −1
Ĝmm = Ŷm Wm (Wm Wm ) , ∀m. (12)

Accordingly, the MSE of estimating Gmm can be computed as [4]

n o N σ2 
2 rx r H −1
MSE(Gmm ) = E kGmm − Ĝmm k = tr (Wm Wm ) . (13)
T
Given the sensing MSE threshold ηm , the sensing quality requirement of the m-th sensor is
Nrx σr2  H −1
tr (Wm Wm ) ≤ ηm , ∀m. (14)
T
Due to the long distance between the target and the AP, the target reflection signal vanishes
at the AP. Therefore, the corresponding received signal ẑ[t] ∈ CK×1 at the AP can be expressed
as
M
X
ẑ[t] = AH Hm Wm sm [t] + AH nc [t], (15)
m=1
Na ×Ntx
where Hm ∈ C is the channel between the AP and the m-th sensor, A ∈ CNa ×K is
the data aggregation beamformer at the AP. nc ∈ CNa ×1 is a AWGN vector with distribution
NNa ×1 (0, σc2 I), which is statistically independent of sm [t].
Due to the nature of analog transmission, the accuracy of AirComp is prone to the distortion
by channel fading and noise. As the goal of AirComp is to accurately compute certain functions,
the computation error becomes a natural performance metric. Following the existing literatures
(see, e.g., [9], [10]), the error is measured via the MSE between the estimated function value
and the ground truth one, i.e.,
  ! !H  
M
X M
X
Et tr  (AH Hm Wm −I)sm [t]+AH nc [t] (AH Hm Wm −I)sm [t]+AH nc [t] 
m=1 m=1

M
X
tr (AH Hm Wm − I)(AH Hm Wm − I)H + σc2 tr(AAH ).


= (16)
m=1
 12

B. Separated Scheme

As shown in Fig. 1 (b), the antennas at each sensor are divided into two groups with Ns =
Nr +Nc , where Nr antennas are for radar sensing and Nc for data transmission. The Nr antennas
are further divided into two groups with Nr = Ntx + Nrx , where Ntx antennas are for radar
signal transmitting and the remaining Nrx for receiving. The data symbols transmitted by the
m-th sensor at the t-th slot can be expressed as a vector denoted by dm [t] ∈ CK×1 , where
K represents the number of functions to be calculated via AirComp. The data symbols are
assumed to be i.i.d. among different sensors and functions with zero mean and unit variance,
i.e., Et [dm [t]dH H
m [t]] = I and Et [dm [t]di [t]] = 0, ∀i 6= m. The radar signals transmitted by

the m-th sensor at the t-th slot can be expressed as a vector denoted by sm [t] ∈ CK×1 , which
satisfies Et [sm [t]sH H
m [t]] = I and Et [sm [t]si [t]] = 0, ∀i 6= m. The data stream signals are statically

independent of the radar signals, i.e., Et [si [t]dH

m [t]] = 0, ∀i, m. The diagram of separated scheme

is shown in Fig. 3 and described as follows.

The signal transmitted by the m-th sensor can be expressed as
 
Wm dm [t]
xm [t] =  , (17)
Fm sm [t]

where Wm ∈ CNc ×K is the data transmission beamformer, Fm ∈ CNtx ×K is the radar sensing
beamformer. Due to the limited transmit power of each sensor, the beamformer design should
satisfy the power constraint:

tr(Wm Wm H ) + tr(Fm Fm H ) ≤ P, ∀m. (18)

The target reflection signal ym [t] ∈ CNrx ×1 received at the m-th sensor can be expressed as

ym [t] = Gmm Fm sm [t] + Ωm [t] + nr [t], (19)

PM P
where Ωm [t] = i=1 Cim Wi di [t]+ i∈M/{m} (Gim Fi si [t]+Qim Fi si [t]+Oim Wi di [t]). For the i-
th and m-th sensors, Gim ∈ CNrx ×Ntx is the TRM, Qim ∈ CNrx ×Ntx is the DRC, Cim ∈ CNrx ×Nc
is the data reflection channel, and Oim ∈ CNrx ×Nc is the direct data channel. nr [t] ∈ CNrx ×1 is
an AWGN vector with distribution NNrx (0, σr2 I).
According to [52], the sufficient statistic matrix Ŷm ∈ CNrx ×K can be derived as
T M
1 XX
Ŷm = (Gim Fi si [t] + Cim Wi di [t])sH
m [t]
T t=1 i=1
 13

Data
transmission
beamformer
Data W1
symbol
s1[t]
Radar sensing Data
beamformer AirComp aggregation
Radar F1 channel H1 beamformer
signal
AWGN A
d1[t]
nc
AWGN Target AirComp
nr response output
matrix
Ĝ 11 G11
Estimated
parameter Ω1 Interference
WM

sM[t]

FM HM

dM[t]
nr

Ĝ MM GMM

ΩM

Figure 3: Diagram of separated scheme.

T T
1X X 1X
+ (Qim Fi si [t] + Oim Wi di [t])sH
m [t] + nr [t]sH
m [t]. (20)
T t=1 T t=1
i∈M/{m}

According to the law of large-number, when the number of slots T is large, one can get
T
1X
si [t]sH H
m [t] ≈ Et [si [t]sm [t]] = 0, (21)
T t=1
T
1X
sm [t]sH H
m [t] ≈ Et [sm [t]sm [t]] = I, (22)
T t=1
T
1X
di [t]sH H
m [t] ≈ Et [di [t]sm [t]] = 0. (23)
T t=1
Therefore, the sufficient statistic matrix can be expressed as

Ŷm = Gmm Fm + Nm , (24)

 14

1
PT Nrx ×K
where Nm = T t=1 ns [t]sH
m [t] ∈ C . Following the similar analysis in the shared scheme,
it can be derived that Nm ∼ MN Nrx ×K (0, √σrT INrx ×Nrx , √σrT IK×K ).
Accordingly, the MSE of estimating Gmm can be computed as [4]
n o N σ2 
rx r −1
MSE(Gmm ) = E kGmm − Ĝmm k2 = tr (Fm FH
m) . (25)
T
Given the sensing MSE threshold ηm , the sensing quality requirement of the m-th sensor is
Nrx σr2  −1
tr (Fm FH
m) ≤ ηm , ∀m. (26)
T
The received signal ẑ[t] ∈ CK×1 at the AP can be expressed as
M
X
ẑ[t] = AH (Hm Wm dm [t] + Rm Fm sm [t]) + AH nc [t], (27)
m=1
Na ×Nc
where Hm ∈ C and Rm ∈ CNa ×Ntx are the channels between the AP and the m-th sensor
for data symbols and radar signals, respectively. The AWGN vector nc ∈ CNa ×1 is statistically
independent of sm [t] and dm [t]. The corresponding MSE between the estimated function value
and the ground truth one can be expressed as
 2

XM M
X
Et  AH (Hm Wm dm [t] + Rm Fm sm [t]) + AH nc [t] − dm [t]  (28)
m=1 m=1

M
X M
X
H H H
tr AH Rm Fm FH H 2 H



= tr (A Hm Wm − I)(A Hm Wm − I) + m Rm A + σc tr(AA ).
m=1 m=1

IV. D UAL - FUNCTIONAL S HARED B EAMFORMING D ESIGN

The shared scheme for radar sensing and AirComp in the ISCCO system can be formulated as
a joint optimization problem over transmission beamformer Wm at each sensor and aggregation
beamformer A at the AP. Specifically, given the MSE in (16) together with the power constraint
in (3) and the sensing quality constraint in (14), the problem can be formulated as
M
X
tr (AH Hm Wm − I)(AH Hm Wm − I)H + σc2 tr(AAH )


min
A,{Wm }
m=1

H −1

T ηm
(P1) s.t. tr (Wm Wm ) ≤ , ∀m,
Nrx σr2
tr(Wm Wm H ) ≤ P, ∀m.
Problem P1 is difficult to solve due to its non-convexity. The lack of convexity arises from the
coupling between the transmitting and aggregation beamformers. To deal with such problem, the
optimal transmission beamforming design is given in the following proposition.
 15

Proposition 1 (Optimal Transmission Beamformer at Sensor). For the shared beamforming

design, given the data aggregation beamformer A, the computation error is minimized by
adopting the following zero-forcing transmission beamformer at the sensors:

−1 H
Wm = (HH H
m AA Hm ) Hm A, ∀m. (29)

Proof: See Appendix B. 

By adopting the zero-forcing transmitting beamformer, the corresponding problem can be
formulated as
min σc2 tr(AAH )
A
T ηm
tr HH H


(P2) s.t. m AA Hm ≤ , ∀m,
Nrx σr2
−1
tr((HH H
m AA Hm ) ) ≤ P, ∀m.

−1 Na ×K
Since tr((HH H
m AA Hm ) ) is neither convex nor concave over A ∈ C , the problem (P2)
is non-convex. By introducing new variable  = AAH , the problem can be formulated as

min σc2 tr(Â)

Â
  T ηm
(P3) s.t. tr HH
m ÂH m ≤ , ∀m,
Nrx σr2
−1
tr((HH
m ÂHm ) ) ≤ P, ∀m,

rank(Â) = K,

  0.
By applying the semidefinite relaxation (SDR), the problem can be formulated as

min σc2 tr(Â)

Â
  T ηm
(P4) s.t. tr HH
m ÂH m ≤ , ∀m,
Nrx σr2
−1
tr((HH
m ÂHm ) ) ≤ P, ∀m,

  0.
The convexity of problem (P4) is established in the following lemma.

Lemma 2. Problem P4 is a convex problem.

Proof: See Appendix C. 

 16

Upon solving the problem (P4) via a convex problem solver (e.g., the cvx toolbox in MATLAB)
and attaining the globally optimal solution Â∗ , the next task is to retrieve from it a feasible
solution to the original problem denoted by Ã. Since the rank of Â∗ might be larger than K,
the Gaussian randomization algorithm proposed in [53] can be applied to extract à from Â∗ .
The main procedure is summarized in Algorithm 1.

Algorithm 1 Gaussian Randomization Algorithm for ISCCO

• Initialization: Given an SDR solution Â∗ , and the number of random samples N .
• Gaussian Random Sampling:
(1) Perform eigen decomposition [VÂ , ΣÂ ] = eig(Â∗ ).
(2) Generate N 0 random matrices Zn ∼ CN (0, I) with Zn ∈ CNa ×K , 0 ∈ CNa ×K and
I ∈ CNa ×Na .
1/2
(3) Retrieve N feasible solutions {An = VÂ ΣÂ ZH n }, such that the constraints
−1
≤ NTrxηmσ2 , ∀m


rank(An AH H H H H
n ) = K, tr((Hm An An Hm ) ) ≤ P , and tr Hm An An Hm r

can be enforced.
(4) Select the best An that leads to the minimum objective, namely A∗n =
arg minn σc2 tr(AH
n An ).

(5) Output à = A∗n as the approximated optimal data aggregation beamformer.

Remark 1 (Coupling Relationship of Radar Sensing and AirComp in the Shared Scheme). As
shown in problem (P1), one transmission beamformer needs to be designed at each sensor for
supporting both radar sensing and AirComp functionalities, which is further correlated to the data
aggregation beamformer design at the server via zero-forcing. Therefore, the shared scheme need
to guarantee the sensing MSE requirements at the price of sacrificing the AirComp accuracy.

V. D UAL - FUNCTIONAL S EPARATED B EAMFORMING D ESIGN

In contrast to the shared scheme, the separated scheme should take the joint optimization
of data transmission beamformer Wm , radar sensing beamformer Fm , and data aggregation
beamformer A into consideration. Specifically, given the MSE in (28) together with the power
constraint in (18) and the sensing quality constraint in (26), the problem can be formulated as
 17

M
X
tr (AH Hm Wm − I)(AH Hm Wm − I)H


min
A,{Wm },{Fm }
m=1
M
X
tr AH Rm Fm FH H 2 H


+ m Rm A + σc tr(AA )
m=1

−1 T ηm
tr (Fm FH


(P5) s.t. m) ≤ , ∀m,
Nrx σr2
tr(Wm Wm H ) + tr(Fm Fm H ) ≤ P, ∀m.
Following the similar approach of solving problem (P1), the zero-forcing data transmission
beamformer is adopted to minimize the AirComp MSE, i.e.,

−1 H
Wm = (HH H
m AA Hm ) Hm A, ∀m. (30)

The corresponding problem can be formulated as

M
X
tr Fm FH H H 2 H


min m Rm AA Rm + σc tr(AA )
A,{Fm }
m=1

−1 T ηm
tr (Fm FH


(P6) s.t. m) ≤ , ∀m,
Nrx σr2
−1 H
tr((HH H
m AA Hm ) ) + tr(Fm Fm ) ≤ P, ∀m.

The problem (P6) is non-convex due to the coupling variables Fm and A in the objective function.
Following a common approach in the MIMO beamforming literature (see e.g., [54]–[56]), the
radar sensing beamformer Fm is constrained to be an orthogonal matrix. Mathematically, one
√
can write Fm = αm Dm with Dm being a tall unitary matrix and thus Dm DH m = I, while αm

is a positive scaling factor. The corresponding problem can be formulated as

M
X
αm tr RH H 2 H


min m AA Rm + σc tr(AA )
A,{αm }
m=1
Ntx T ηm
(P7) s.t. ≤ , ∀m,
αm Nrx σr2
−1
tr((HH H
m AA Hm ) ) + αm ≤ P, ∀m.

It can be observed that the increasing of αm will result in larger MSE. Therefore, the minimum
∗
of MSE over αm is achieved when the minimum αm is adopted for all m, i.e.,

∗ Ntx Nrx σr2

αm = , ∀m. (31)
T ηm
 18

By introducing  = AAH and applying the SDR, the problem can be formulated as
M
X Ntx Nrx σr2  H 
min tr Rm ÂRm + σc2 tr(Â)
Â
m=1
T ηm

−1 Ntx Nrx σr2

(P8) s.t. tr((HH
m ÂHm ) ) + ≤ P, ∀m,
T ηm
  0.
It can be easily proved that problem (P8) is convex due to the linear objective and convex con-
straints. Upon attaining the globally optimal solution of problem (P8) by convex programming,
denoted by Â∗ , the remaining task is to convert it into a feasible solution of the original problem,
denoted by Ã, of rank K. To this end, the Gaussian randomization based Algorithm 1 is applied.

Remark 2 (Coupling Relationship of Radar Sensing and AirComp in the Separated Scheme). As
shown in problem (P8), the existence of radar signals results in extra AirComp error denoted by
PM Ntx Nrx σr2  H 
m=1 T ηm
tr R m ÂR m . To mitigate the interference on AirComp caused by radar signals,

the radar sensing beamformers are designed to achieve the maximum tolerance of sensing MSE,
2
i.e., NrxT σr tr (Fm FH −1


m ) = ηm .

y
Target
Sensor !i

yq(i)
yp(i) θi
y1(i)

yq(m)
θm
yp(m)
Sensor !m
y1(m) x

Figure 4: Location Estimation System based on ISCCO.

VI. TARGET L OCATION E STIMATION BASED ON ISCCO

In this section, the ISCCO scheme was applied for the use case of target location estimation.
Particularly, the location of the target is estimated by all M sensors based on their own locations
as well as the information of angle and distance extracted from the reflected radar signals. The
 19

estimated location of the target by each sensor is then transmitted to the server via AirComp,
and thus the server will obtain the averaged estimated target location. As shown in Fig. 4, the
TRM Gmm is composed of a phase delay matrix Φ(θm ) and a complex amplitude βm of the
received signal, i.e., Gmm = βm Φ(θm ). Let ϕpq (θm ) denote the element of Φ(θm ) in p-th row
and q-th column, then
ϕpq (θm ) = exp{−jω[τp (θm ) + τq (θm )]}, (32)

where ω represents the angular velocity, τp (θm ) represents the transmitting time delay between
the 1-st and p-th antennas, τq (θm ) represents the receiving time delay between the 1-st and q-th
antennas. According to [52], the phase delay between the p-th and q-th antennas at the m-th
sensor can be expressed as
2πj (m)
ϕpq (θm ) = exp{− (y + yq(m) ) sin θm }, (33)
λ p
(m) (m)
where yp and yq denote the location of the p-th and q-th antennas at the m-th sensor,
respectively. Following the derivation of Gmm , the MLE of βm and θm can be found by
minimizing the log-likelihood function:

L(βm , θm ) = tr[(Ŷm − βm Φ(θm )Wm )H (Ŷm − βm Φ(θm )Wm )]. (34)

The derivatives of L(βm , θm ) w.r.t. βm is

∂L(βm , θm ) H H H H
= 2βm tr(Wm Φ (θm )Φ(θm )Wm ) − 2tr(Wm Φ (θm )Ŷm ). (35)
∂βm
H H
According to (12), Ĝmm Wm Wm = Ŷm Wm , ∀m. Setting the derivatives as zero, one can get
H H
tr(Wm Φ (θm )Ĝmm Wm )
β̂m = , ∀m. (36)
tr(Wm ΦH (θm )Φ(θm )Wm )
H

By replacing βm with β̂m in L(βm , θm ), one can get

H tr2 (WmH H
Φ (θm )Ĝmm Wm )
L(θm ) = tr(Ŷm Ŷm ) − . (37)
tr(Wm ΦH (θm )Φ(θm )Wm )
H

As the angle θm to be estimated is only relevant with the second item in (37), one can get
tr2 (WmH H
Φ (θm )Ĝmm Wm )
θ̂m = arg max H ΦH (θ )Φ(θ )W )
. (38)
θm tr(Wm m m m

It should be noted that θ̂m cannot be expressed in closed form. Therefore, the grid search or
golden section search can be applied to find the numerical results, where the beamformer Wm
is obtained by solving problem (P1). On the other hand, the distance dm between the target and
the m-th sensor can be estimated following the free space propagation law [57]. Based on the
 20

estimated parameters (distance dˆm and angle θ̂m ) and its own location (0, ym ), the m-th sensor
can obtain its local estimation of the target location denoted by ẑm = [x̂m , ŷm ]T via

x̂m = dˆm sin θ̂m , (39)

(m)
ŷm = y1 + dˆm cos θ̂m . (40)

The target location estimated by the m-th sensor is then modulated into data symbols repre-
sented by sm = [xm , ym ]T , where
x̂m
xm = − 1, (41)
x̄
ŷm
ym = − 1, (42)
ȳ
with x̄ and ȳ denoting the statistic values of the target location at x-axis and y-axis. After
transmission beamforming Wm , the data symbols are transmitted to the AP. In the shared scheme,
the signals received at the AP can be expressed as
M
X
H
ŝ = A Hm Wm sm + AH nc , (43)
m=1

where ŝ = [x0 , y 0 ]T . The averaged estimated target location can be derived as zm = [(x0 +
1)x̄, (y 0 + 1)ȳ]T . The performance of radar sensing and AirComp will be evaluated based on the
simulation results in section VII.

VII. S IMULATION

In this section, the performance of our proposed ISCCO framework is evaluated by simulation,
where the radar sensing and AirComp channel models in shared and separated schemes are
simulated based on (4), (15), (19), and (27). The performance metric is the normalized AirComp
MSE, defined by MSE/M with the AirComp MSE given in (16) and (28) for shared and separated
schemes, respectively. The simulation parameters are set as follows unless specified otherwise.
The number of time slots is T = 1000. The number of computed functions is set to be K = 10.
There are M = 10 sensors each equipped with Ns = 12 antennas and one AP with Na =
15 antennas. In the shared scheme, Ntx = 6 antennas are for signal transmitting and Nrx =
6 antennas are for signal receiving. In the separated scheme, Nc = 4 antennas are for data
transmission and Nr = 8 antennas are for radar sensing, where Ntx = 4 antennas are for radar
signal transmitting and Nrx = 4 antennas are for radar signal receiving. All the channels are
assumed to be i.i.d. Rician fading, modeled as i.i.d. complex Gaussian random variables with
 21

non-zero mean µ = 1 and variance σ 2 = 1. In addition, the maximum transmission power is set
as P0 = 10 mW. The effective power conversion efficiency follows a uniform distribution with
ηn ∈ (0, 1). According to the settings in LTE [58], the powers of noise in radar signal channel
σr2 and data transmission channel σc2 are −79.5 dBm. Each point in the figures is obtained by
averaging over 10 simulation realizations, with independent channels in each realization.

A. Baseline Schemes

Two baselines are designed by applying antenna selection (AS) on the shared and separated
schemes, respectively. All the schemes assume the channel-inversion data precoding. Define the
sum-channel matrix Hsum = M
P
m=1 Hm . The baseline schemes with AS select the K receive

antenna observing the largest channel gains in the sum channel Hsum . For fair comparison, all
aggregation beamformers in the baseline schemes are scaled to have the same norm.

B. Function Computation Performance of ISCCO

First, the normalized AirComp MSE versus the number of antennas at the AP is evaluated
in Fig. 5 for both the shared and separated schemes. It can be observed that the normalized
AirComp MSE decreases with the increasing number of antennas at the AP. This is because
more antennas at the AP will enlarge the dimension of data aggregation beamformer, and thus
exploit the diversity gain for achieving lower AirComp MSE. It should be noted that both the
shared and separated schemes proposed in this paper can achieve lower AirComp MSE than
the baselines with AS, which verifies the necessity of beamformer optimization. Moreover, the
separated scheme has better performance than the shared one under the current system settings.
The reason is that the dual-functional signals in the shared scheme make it hard to design one
common beamformer for supporting both radar sensing and AirComp, while the interference
caused by radar signals in the separated scheme can be effectively mitigated by the dedicated
beamformer design for AirComp signals.
Fig. 6 demonstrates the normalized AirComp MSE versus the number of antennas at each
sensor in both the shared and separated schemes. One can observe that the normalized AirComp
MSE monotonically increases with the increasing number of antennas at each sensor, since
more antennas at sensors will result in larger dimension of TRM to be estimated and thus more
stringent sensing constraints. Therefore, the beamformers need to guarantee the requirements of
radar sensing at the price of scarifying the performance of AirComp. Moreover, the performance
 22

10-1
Optimized separated scheme
Separated scheme with AS
Optimized shared scheme

Normalized computation MSE

Shared scheme with AS

10-2

10-3
10 12 14 16 18 20
Number of antennas at the AP

Figure 5: Normalized AirComp MSE versus the number of antennas at the AP.

10-1
Optimized separated scheme
Separated scheme with AS
Optimized shared scheme
Normalized computation MSE

Shared scheme with AS

10-2

10-3

10-4
6 8 10 12 14 16 18
Number of antennas at each sensor

Figure 6: Normalized AirComp MSE versus the number of antennas at each sensor.

of shared scheme becomes better than that of the separated one with the increasing number
of antennas at each sensor. Such a phenomenon is caused by double effects of deploying more
antennas at each sensor. On one hand, more antennas at each sensor will result in larger dimension
of beamforming matrix for supporting the dual-functionality of signals in the shared scheme. On
the other hand, more antennas for radar sensing at each sensor will exacerbate the interference
on AirComp in the separated scheme. The similar trends also hold for the baselines with AS.
Fig. 7 illustrates the curves of the normalized AirComp MSE versus the number of sensors
 23

10-1
Optimized separated scheme
Separated scheme with AS
Optimized shared scheme

Normalized computation MSE

Share scheme with AS

10-2

10-3
5 10 15
Number of sensors

Figure 7: Normalized AirComp MSE versus the number of sensors.

for both the shared and separated schemes. It is shown that the increasing number of sensors
will result in higher normalized computation MSE, as more connected sensors make it harder
to design one common data aggregation beamformer to equalize the channels among different
sensors. Moreover, the increasing trend of normalized AirComp MSE in the separated scheme is
more drastic compared with the shared scheme, since larger number of sensors will exacerbate
the interference of radar signals on AirComp. The similar trends also hold for the baselines with
AS.
Fig. 8 further shows the curves of the normalized AirComp MSE versus the number of
functions to be computed in both the shared and separated schemes. One can observe that the
normalized AirComp MSE increases with the number of functions, which indicates that higher
computation throughput is at a cost of declining accuracy. Moreover, the separated scheme always
performs better than the shared one no matter how many functions need to be computed, which
implies that the former is more robust against the varying number of functions. The similar
trends also hold for the baselines with AS.

C. Radar Sensing Performance of ISCCO

As for radar sensing, the effects of antenna amounts at both the sensors and the AP on the
averaged sensing MSE are illustrated in Fig. 9 for both the shared and separated schemes. One can
observe that the averaged sensing MSE decreases with the increasing number of antennas at the
AP in the shared scheme, which indicates that the enlarged dimension of aggregation beamformer
 24

101
Optimized separated scheme
Separated scheme with AS
Optimized shared scheme

Normalized computation MSE

Shared scheme with AS
100

10-1

10-2

10-3
10 11 12 13 14 15
Number of functions

Figure 8: Normalized AirComp MSE versus the number of functions to be computed.

100

Shared scheme with N s = 12

-1 Shared scheme with N s = 18
10
Averaged sensing MSE

Separated scheme with N s = 12

separated scheme with N s = 18

10-2

10-3

10-4
10 12 14 16 18 20
Number of antennas at the AP

Figure 9: Effects of antenna amounts on averaged sensing MSE.

will result in higher degree of freedom for achieving lower sensing MSE. Moreover, deploying
more antennas at each sensor in the shared scheme will result in larger averaged sensing MSE as
the dimension of TRM to be estimated is enlarged. In contrast, the averaged sensing MSE does
not change with the number of antennas at neither the AP nor the sensors in the separated scheme,
since the radar sensing constraint is tighten for mitigating the interference of radar signals on
AirComp. Therefore, the sensing MSE only depends on the sensing quality requirement and is
irrelevant with other parameters.
 25

D. Target Location Estimation based on ISCCO

The use case of target location estimation based on ISCCO is demonstrated in Fig. 10. The
ground truth location of the target is set as (5, 30) m. M = 10 sensor are located at the range
[0, 20] m on the Y-axis with 2 m distance between each other. The number of antennas at each
sensor is set as Ntx = Nrx = 2 with 0.1 m space between each other. The information to be
estimated and transmitted is a vector which contains the two-dimensional location of the target.
The angles between the target and the sensors are estimated via (38), while the distance is
assumed to be perfectly estimated. The performance of the conventional radar sensing scheme
based on angle of arriving (AoA) [3] is also plotted, where the estimated target location (x0 , y0 )
is obtained by minimizing the MSE function minx0 ,y0 M x0 −xm 2
P
m=1 |θ̂m −arctan y0 −ym | via grid search,

with (xm , ym ) denoting the location of the m-th sensor. It can be observed that the estimated
target location by each sensor based on ISAC is a little deviating from the ground truth, while the
application of AirComp can alleviate such deviation by averaging the measured values of sensors
over transmission. Moreover, the target location estimated by ISCCO is more approaching the
ground truth than that by the conventional AoA.

35

30

25

20
Y-axis

15

10 Estimated target locations by sensors

Locations of sensors
5 Ground truth target location
Averaged target location by AirComp
Estimated target location by AoA
0
0 1 2 3 4 5 6
X-axis

Figure 10: Target location estimation based on ISCCO.

When the power of noise in data transmission channel is set as −59.5 dBm, the target location
estimation based on ISCCO is demonstrated in Fig. 11. It can be observed that the performance of
AirComp is deteriorated due to the strong noise. In such condition, the target location estimation
by a single sensor might have better performance, which necessitates the scheduling of sensors.
 26

35

30

25

20

Y-axis
15

10 Estimated target locations by sensors

Locations of sensors
5 Ground truth target location
Averaged target location by AirComp
Estimated target location by AoA
0
0 1 2 3 4 5 6
X-axis

Figure 11: Target location estimation based on ISCCO under strong channel noise.

VIII. C ONCLUDING R EMARKS

In this paper, an ISCCO framework has beed proposed for enabling the simultaneous radar
sensing and Aircomp to improve the spectrum efficiency in IoT systems. To this end, two designs
known as the shared and separated schemes have been investigated. In the shared scheme, all the
antennas at each sensor are exploited for transceiving dual-functional signals. In the separated
scheme, the antenna array at each sensor is divided into two sub-arrays for supporting radar
sensing and Aircomp, respectively. The non-convex problem of joint optimizing the beamformers
for radar sensing, data transmission and aggregation is solved via semidefinite relaxation together
with Gaussian randomization. This work contributes to the promising new research area of
ISCCO and many interesting follow-up research issues warrant further investigation, such as
sensor scheduling, vehicular tracking, and target surface estimation.

A PPENDIX

A. Proof of Lemma 1

By letting Sm = [sm [1], sm [2], ...sm [T ]] ∈ CK×T and Nr = [nr [1], nr [2], ..., nr [T ]] ∈ CNrx ×T ,
one can get Nm = 1
T
Nr SH
m ∈ CNrx ×K . The vectorization of Nr is a Gaussian random vector
denoted by vec(Nr ) ∼ NNrx T ×1 (0, σr2 INrx T ×Nrx T ). Correspondingly, the vectorization of Nm
can be expressed as
1 1
nm = vec(Nm ) = vec( INrx ×Nrx Nr SH
m) = (Sm ⊗ INrx ×Nrx )vec(Nr ), (44)
T T
 27

which is a linear transformation of vec(Nr ). Therefore, nm ∼ NNrx K×1 (0, Σ), where Σ =
h 2 i h 2 i 2
σr
E T 2 (Sm ⊗ INrx ×Nrx )(Sm ⊗ INrx ×Nrx )H
= E T 2 (Sm Sm ) ⊗ INrx ×Nrx = σTr INrx K×Nrx K . Un-
σr H

der such condition the PDF of nm is

1 1 H −1 n
p(nm |0, Σ) = e− 2 nm Σ m

(2π)Nrx K/2 |Σ|1/2

1 − T H
2 nm nm
= 2 e 2σr

(2π)Nr K/2 | σTr INrx K×Nrx K |1/2

1 − T 2 vec(Nm )H vec(Nm )
= e 2σr
(2π)Nrx K/2 | √σrT INrx ×Nrx |K/2 | √σrT IK×K |Nrx /2
1 − T H
2 tr[Nm Nm ]
= e 2σr
(2π)Nrx K/2 | √σrT INrx ×Nrx |K/2 | √σrT IK×K |Nrx /2
σr σr
= p(Nm |0, √ INrx ×Nrx , √ IK×K ). (45)
T T

B. Proof of Proposition 1

Given the AirComp MSE minimization objective provided in (16), it can be observed that
both M H H H


and σc2 tr(AAH ) are positive. Therefore,
P
m=1 tr (A Hm Wm − I)(A Hm Wm − I)

given any data aggregation beamformer A, the inequality

M
X
tr (AH Hm Wm − I)(AH Hm Wm − I)H + σc2 tr(AAH ) ≥ σc2 tr(AAH )


(46)
m=1

always holds. It is easy to verify that setting Wm to have the zero-forcing structure in (29)
enforces
M
X
tr (AH Hm Wm − I)(AH Hm Wm − I)H = 0,


(47)
m=1

and thus achieves the equality in (46).

C. Proof of Lemma 2
−1
Since the item (HH
m ÂHm ) is convex over  and tr(X) is linear over X, the function
−1
tr((HH
m ÂHm ) ) is convex over  according to the composition rule [59]. Since other constraints

as well as the objective function are linear functions over Â, problem (P4) is convex.

R EFERENCES

[1] W. Saad, M. Bennis, and M. Chen, “A vision of 6G wireless systems: Applications, trends, technologies, and open research
problems,” IEEE Netw., vol. 34, no. 3, pp. 134–142, 2019.
 28

[2] Y. Cui, F. Liu, X. Jing, and J. Mu, “Integrating sensing and communications for ubiquitous IoT: Applications, trends, and
challenges,” IEEE Netw., vol. 35, no. 5, pp. 158–167, 2021.
[3] F. Liu, C. Masouros, A. P. Petropulu, H. Griffiths, and L. Hanzo, “Joint radar and communication design: Applications,
state-of-the-art, and the road ahead,” IEEE Trans. Commun., vol. 68, no. 6, pp. 3834–3862, 2020.
[4] F. Liu, Y. Cui, C. Masouros, J. Xu, T. X. Han, Y. C. Eldar, and S. Buzzi, “Integrated sensing and communications: Towards
dual-functional wireless networks for 6G and beyond,” arXiv preprint arXiv:2108.07165, 2021.
[5] G. Zhu, J. Xu, K. Huang, and S. Cui, “Over-the-air computing for wireless data aggregation in massive IoT,” IEEE Wireless
Commun., vol. 28, no. 4, pp. 57–65, 2021.
[6] B. Nazer and M. Gastpar, “Computation over multiple-access channels,” IEEE Trans. Inf. Theory, vol. 53, no. 10, pp.
3498–3516, 2007.
[7] O. Abari, H. Rahul, D. Katabi, and M. Pant, “Airshare: Distributed coherent transmission made seamless,” in Proc. IEEE
Conference on Computer Communications (INFOCOM), 2015.
[8] L. Chen, N. Zhao, Y. Chen, F. R. Yu, and G. Wei, “Over-the-air computation for IoT networks: Computing multiple
functions with antenna arrays,” IEEE Internet Things J., vol. 5, no. 6, pp. 5296–5306, 2018.
[9] G. Zhu and K. Huang, “MIMO over-the-air computation for high-mobility multimodal sensing,” IEEE Internet Things J.,
vol. 6, no. 4, pp. 6089–6103, 2018.
[10] X. Li, G. Zhu, Y. Gong, and K. Huang, “Wirelessly powered data aggregation for IoT via over-the-air function computation:
Beamforming and power control,” IEEE Trans. Wireless Commun., vol. 18, no. 7, pp. 3437–3452, 2019.
[11] D. Wen, G. Zhu, and K. Huang, “Reduced-dimension design of MIMO over-the-air computing for data aggregation in
clustered IoT networks,” IEEE Trans. Wireless Commun., vol. 18, no. 11, pp. 5255–5268, 2019.
[12] J. Dong, Y. Shi, and Z. Ding, “Blind over-the-air computation and data fusion via provable wirtinger flow,” IEEE Trans.
Signal Process., vol. 68, pp. 1136–1151, 2020.
[13] X. Cao, G. Zhu, J. Xu, and K. Huang, “Optimized power control for over-the-air computation in fading channels,” IEEE
Trans. Wireless Commun., vol. 19, no. 11, pp. 7498–7513, 2020.
[14] W. Liu, X. Zang, Y. Li, and B. Vucetic, “Over-the-air computation systems: Optimization, analysis and scaling laws,” IEEE
Trans. Wireless Commun., vol. 19, no. 8, pp. 5488–5502, 2020.
[15] X. Zhai, X. Chen, J. Xu, and D. W. K. Ng, “Hybrid beamforming for massive MIMO over-the-air computation,” IEEE
Trans. Commun., vol. 69, no. 4, pp. 2737–2751, 2021.
[16] G. Zhu, Y. Wang, and K. Huang, “Broadband analog aggregation for low-latency federated edge learning,” IEEE Trans.
Wireless Commun., vol. 19, no. 1, pp. 491–506, 2019.
[17] Y. Sun, S. Zhou, Z. Niu, and D. Gündüz, “Dynamic scheduling for over-the-air federated edge learning with energy
constraints,” arXiv preprint arXiv:2106.00490, 2021.
[18] M. M. Amiri and D. Gündüz, “Federated learning over wireless fading channels,” IEEE Trans. Wireless Commun., vol. 19,
no. 5, pp. 3546–3557, 2020.
[19] K. Yang, T. Jiang, Y. Shi, and Z. Ding, “Federated learning via over-the-air computation,” IEEE Trans. Wireless Commun.,
vol. 19, no. 3, pp. 2022–2035, 2020.
[20] H. Guo, A. Liu, and V. K. Lau, “Analog gradient aggregation for federated learning over wireless networks: Customized
design and convergence analysis,” IEEE Internet Things J., vol. 8, no. 1, pp. 197–210, 2020.
[21] C. Xu, S. Liu, Z. Yang, Y. Huang, and K.-K. Wong, “Learning rate optimization for federated learning exploiting over-
the-air computation,” arXiv preprint arXiv:2102.02946, 2021.
[22] N. Zhang and M. Tao, “Gradient statistics aware power control for over-the-air federated learning,” IEEE Trans. Wireless
Commun., vol. 20, no. 8, pp. 5115–5128, 2021.
 29

[23] G. Zhu, Y. Du, D. Gündüz, and K. Huang, “One-bit over-the-air aggregation for communication-efficient federated edge
learning: Design and convergence analysis,” IEEE Trans. Wireless Commun., vol. 20, no. 3, pp. 2120–2135, 2020.
[24] D. Liu and O. Simeone, “Privacy for free: Wireless federated learning via uncoded transmission with adaptive power
control,” IEEE J. Sel. Areas Commun., vol. 39, no. 1, pp. 170–185, 2020.
[25] W. Ni, Y. Liu, Z. Yang, H. Tian, and X. Shen, “Federated learning in multi-RIS aided systems,” IEEE Internet Things J.,
2021.
[26] M. Fu, Y. Zhou, Y. Shi, T. Wang, and W. Chen, “UAV-assisted over-the-air computation,” arXiv preprint arXiv:2101.09856,
2021.
[27] S. Wang, Y. Hong, R. Wang, Q. Hao, Y.-C. Wu, and D. W. K. Ng, “Edge federated learning via unit-modulus over-the-air
computation,” arXiv preprint arXiv:2101.12051, 2021.
[28] P. Park, P. Di Marco, and C. Fischione, “Optimized over-the-air computation for wireless control systems,” IEEE Commun.
Lett., early access, 2021.
[29] F. Han, V. K. Lau, and Y. Gong, “Over-the-air computation of large scale nomographic functions in MapReduce over the
edge cloud network,” IEEE Internet Things J., early access, 2021.
[30] R. M. Mealey, “A method for calculating error probabilities in a radar communication system,” IEEE Transactions on
Space Electronics and Telemetry, vol. 9, no. 2, pp. 37–42, 1963.
[31] DARPA, “Shared spectrum access for radar and communications (SSPARC),” 2012.
[32] R. Saruthirathanaworakun, J. M. Peha, and L. M. Correia, “Opportunistic sharing between rotating radar and cellular,”
IEEE J. Sel. Areas Commun., vol. 30, no. 10, pp. 1900–1910, 2012.
[33] S. Sodagari, A. Khawar, T. C. Clancy, and R. McGwier, “A projection based approach for radar and telecommunication
systems coexistence,” in IEEE Global Communications Conference (GLOBECOM), Dec. 2012.
[34] B. Li and A. Petropulu, “MIMO radar and communication spectrum sharing with clutter mitigation,” in IEEE Radar
Conference (RadarConf), May 2016.
[35] F. Liu, C. Masouros, A. Li, and T. Ratnarajah, “Robust MIMO beamforming for cellular and radar coexistence,” IEEE
Wireless Commun. Lett., vol. 6, no. 3, pp. 374–377, 2017.
[36] F. Liu, C. Masouros, A. Li, T. Ratnarajah, and J. Zhou, “MIMO radar and cellular coexistence: A power-efficient approach
enabled by interference exploitation,” IEEE Trans. Signal Process., vol. 66, no. 14, pp. 3681–3695, 2018.
[37] L. Zheng, M. Lops, and X. Wang, “Adaptive interference removal for uncoordinated radar/communication coexistence,”
IEEE J. Sel. Topics Signal Process., vol. 12, no. 1, pp. 45–60, 2017.
[38] B. Li and A. P. Petropulu, “Joint transmit designs for coexistence of MIMO wireless communications and sparse sensing
radars in clutter,” IEEE Trans. Aerosp. Electron. Syst., vol. 53, no. 6, pp. 2846–2864, 2017.
[39] B. Paul, A. R. Chiriyath, and D. W. Bliss, “Survey of RF communications and sensing convergence research,” IEEE Access,
vol. 5, pp. 252–270, 2016.
[40] A. R. Chiriyath, B. Paul, G. M. Jacyna, and D. W. Bliss, “Inner bounds on performance of radar and communications
co-existence,” IEEE Trans. Signal Process., vol. 64, no. 2, pp. 464–474, 2015.
[41] S. D. Blunt, J. G. Metcalf, C. R. Biggs, and E. Perrins, “Performance characteristics and metrics for intra-pulse radar-
embedded communication,” IEEE J. Sel. Areas Commun., vol. 29, no. 10, pp. 2057–2066, 2011.
[42] J. Moghaddasi and K. Wu, “Multifunctional transceiver for future radar sensing and radio communicating data-fusion
platform,” IEEE Access, vol. 4, pp. 818–838, 2016.
[43] A. Hassanien, M. G. Amin, Y. D. Zhang, and F. Ahmad, “Dual-function radar-communications: Information embedding
using sidelobe control and waveform diversity,” IEEE Trans. Signal Process., vol. 64, no. 8, pp. 2168–2181, 2015.
 30

[44] F. Liu, C. Masouros, A. Li, H. Sun, and L. Hanzo, “MU-MIMO communications with MIMO radar: From co-existence
to joint transmission,” IEEE Trans. Wireless Commun., vol. 17, no. 4, pp. 2755–2770, 2018.
[45] F. Liu, L. Zhou, C. Masouros, A. Li, W. Luo, and A. Petropulu, “Toward dual-functional radar-communication systems:
Optimal waveform design,” IEEE Trans. Signal Process., vol. 66, no. 16, pp. 4264–4279, 2018.
[46] P. Kumari, J. Choi, N. González-Prelcic, and R. W. Heath, “IEEE 802.11 ad-based radar: An approach to joint vehicular
communication-radar system,” IEEE Trans. Veh. Technol., vol. 67, no. 4, pp. 3012–3027, 2017.
[47] X. Wang, Z. Fei, J. Huang, and H. Yu, “Joint waveform and discrete phase shift design for RIS-assisted integrated sensing
and communication system under Cramér-Rao bound constraint,” IEEE Trans. Veh. Technol., early access, 2021.
[48] Q. Huang, H. Chen, and Q. Zhang, “Joint design of sensing and communication systems for smart homes,” IEEE Network,
vol. 34, no. 6, pp. 191–197, 2020.
[49] T. Zhang, S. Wang, G. Li, F. Liu, G. Zhu, and R. Wang, “Accelerating edge intelligence via integrated sensing and
communication,” arXiv preprint arXiv:2107.09574, 2021.
[50] W. Yuan, F. Liu, C. Masouros, J. Yuan, D. W. K. Ng, and N. González-Prelcic, “Bayesian predictive beamforming for
vehicular networks: A low-overhead joint radar-communication approach,” IEEE Trans. Wireless Commun., vol. 20, no. 3,
pp. 1442–1456, 2020.
[51] Z. Lyu, G. Zhu, and J. Xu, “Joint maneuver and beamforming design for UAV-enabled integrated sensing and
communication,” arXiv preprint arXiv:2110.02857, 2021.
[52] I. Bekkerman and J. Tabrikian, “Target detection and localization using MIMO radars and sonars,” IEEE Trans. Signal
Process., vol. 54, no. 10, pp. 3873–3883, 2006.
[53] Z.-Q. Luo, W.-K. Ma, A. M.-C. So, Y. Ye, and S. Zhang, “Semidefinite relaxation of quadratic optimization problems,”
IEEE Signal Process. Mag., vol. 27, no. 3, pp. 20–34, May 2010.
[54] D. J. Love and R. W. Heath, “Limited feedback unitary precoding for spatial multiplexing systems,” IEEE Trans. Inf.
Theory, vol. 51, no. 8, pp. 2967–2976, 2005.
[55] J. Choi, B. Mondal, and R. W. Heath, “Interpolation based unitary precoding for spatial multiplexing MIMO-OFDM with
limited feedback,” IEEE Trans. Signal Process., vol. 54, no. 12, pp. 4730–4740, 2006.
[56] S. W. Peters and R. W. Heath, “Cooperative algorithms for MIMO interference channels,” IEEE Trans. Veh. Technol.,
vol. 60, no. 1, pp. 206–218, 2010.
[57] C. Sturm and W. Wiesbeck, “Waveform design and signal processing aspects for fusion of wireless communications and
radar sensing,” Proc. IEEE, vol. 99, no. 7, pp. 1236–1259, 2011.
[58] O. W. Schwarz and R. Minihold, “LTE system specifications and their impact on RF & base band circuits,” Rohde &
Schwarz App Note, pp. 1 – 37, 2013.
[59] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, 2004.