1

Integrating Sensing, Computing, and

Communication in 6G Wireless Networks: Design
and Optimization
Qiao Qi, Xiaoming Chen, Ata Khalili, Caijun Zhong, Zhaoyang Zhang, and Derrick Wing Kwan Ng

Abstract—The roll-out of various emerging wireless services via coordination of the two separated systems [9]. However,
has triggered the need for the sixth-generation (6G) wireless
arXiv:2207.03634v1 [cs.IT] 8 Jul 2022

the performance gain of orthogonal spectrum sharing-based

networks to provide functions of target sensing, intelligent radar sensing and wireless communication is limited. As a
computing and information communication over the same radio
spectrum. In this paper, we provide a unified framework integrat- remedy, radar-centric communication integration was proposed
ing sensing, computing, and communication to optimize limited by conveying information over radar signals concurrently via
system resource for 6G wireless networks. In particular, two the same spectrum. For example, the amplitude and the phase
typical joint beamforming design algorithms are derived based of radar signals were controlled modulating some information
on multi-objective optimization problems (MOOP) with the goals bits [10]. Since radar signals can be changed to a limited
of the weighted overall performance maximization and the total
transmit power minimization, respectively. Extensive simulation extent, the information transmission rate of radar-centric com-
results validate the effectiveness of the proposed algorithms. munication integration is severely unsatisfactory. In contrary,
Moreover, the impacts of key system parameters are revealed sensing-aided communication was also proposed to enhance
to provide useful insights for the design of integrated sensing, the performance of wireless communication by using the
computing, and communication (ISCC). sensed information. For instance, the authors in [11] designed
Index Terms—Integrated sensing, computing, and communi- an effective beam selection scheme for millimeter-wave com-
cation, 6G wireless networks, beamforming design. munications based on the sensed environment information,
e.g., the shape, the position, and the materials of surrounding
I. I NTRODUCTION building/car/tree by using a camera. Yet, the functions of
sensing in sensing-aided communication are constrained. In
The demands of various emerging wireless services, e.g., this context, ISAC based on the integrated wireless architecture
unmanned vehicle, holographic communication, and extended is regarded as a promising approach to enable 6G wireless
reality, require wireless networks to provide integrated func- networks [12]. By optimizing the wireless transceivers, it is
tions of target sensing, intelligent computing and information possible to improve both the sensing and communication per-
communication in a timely manner [1]. In current fifth- formance simultaneously. In particular, since the base station
generation (5G) wireless networks, these functions are sep- (BS) will be equipped with a large-scale antenna array for
arated and irreciprocal for simplicity. Generally speaking, 5G 6G wireless networks, the superposition coded sensing and
wireless networks can provide enhanced mobile broadband communication signal can be effectively separated by using
communications [2], cloud and edge computing [3] but with spatial beamforming [13]. Inspired by this, the authors in [14]
limited environment sensing capability [4]. As a result, these studied a joint communication and sensing 6G cellular system
functions in 5G wireless networks cannot satisfy the require- and discussed the optimal waveform design for communication
ments of emerging wireless services. In this context, it is as well as sensing by exploiting ultra-high terahertz spectrum
desired to study the integrated sensing, computing, and com- and ultra-massive antenna array.
munication (ISCC) to improve the utilization of limited system Meanwhile, integrated computing and communication has
resources for the sixth-generation (6G) wireless networks [5]. attracted significant interests due to the demands of huge-
In fact, integrated sensing and communication (ISAC) has volume data processing [15], [16]. Originally, cloud computing
received considerable attentions in both academic and industry was commonly adopted by exploring the powerful computing
[6]-[8]. At the preliminary research stage, as a special case of capability of cloud servers. For example, the authors in [17]
ISAC, spectrum sharing between radar sensing and wireless jointly optimized the precoder of communication signals and
communication has been investigated to improve the spectral the computational resources of cloud servers to minimize the
efficiency. In particular, the co-channel interference between energy consumption at the terminals. Yet, with the increas-
radar and communication systems was carefully mitigated ing of data volume, cloud computing leads to prohibitive
communication load and latency. To address this issue, edge
Qiao Qi, Xiaoming Chen, Caijun Zhong, and Zhaoyang Zhang are with
the College of Information Science and Electronic Engineering, Zhejiang computing is applied to wireless networks [18], [19]. By
University, Hangzhou 310027, China. Ata Khalili is with the Institute for processing the data at the edge of wireless networks, e.g.,
Digital Communications, Friedrich-Alexander-University Erlangen-Nurnberg, the BS, it is likely to increase the efficiency of comput-
Erlangen 91054, Germany. Derrick Wing Kwan Ng is with the School of
Electrical Engineering and Telecommunications, the University of New South ing and communication. For instance, the transmit power of
Wales, NSW 2052, Australia. communication signals and computational resources of edge
servers in a multi-cell wireless network were jointly optimized The rest of this paper outlined as follows: Section II
to maximize the weighted sum of reductions in time and briefly introduces a 6G wireless network integrating sensing,
cost during task completion in [20]. Recently, computing is computing and communication. Section III presents two joint
implemented via cooperation among the BS and the targeted beamforming design algorithms of the weighted overall perfor-
devices, namely over-the-air computation, to further increase mance maximization and the total transmit power minimiza-
the computing efficiency and to decrease the communication tion, respectively. Section IV provides extensive simulations to
load [21]. Specifically, by taking advantage of the superpo- verify the effectiveness of the proposed algorithms and Section
sition nature of wireless multiple access channels, numerous V concludes the paper.
computing functions can be realized at the BS by coordinating Notations: We use ΩN = {1, 2, . . . , N } to denote a set
the transmit signals of the end devices. In [22], the authors for given natural number N , bold upper (lower) letters to
jointly optimized transmit and receive beamforming to combat denote matrices (column vectors), k · k to denote L2 -norm,
the interference between computing and communication sig- | · | to denote the absolute value of a scalar or the size of a
nals. As the processing capability of the end devices increases, dataset, and E{·} to denote expectation. Rm×n and Cm×n
over-the-air computation and federate learning are gradually represent the sets of m-by-n dimensional real matrix and
combined to enable intelligent computing, namely over-the-air complex matrix, respectively. (·)H , tr(·), Rank(·) indicates
federate learning (AirFL) [23]. In practice, AirFL trains local the conjugate transpose, the trace, and the rank of a matrix,
computing models at the end devices with their raw data and respectively.
aggregates a global computing model at the BS by exploiting
the over-the-air computation. In [24], the authors derived the
II. S YSTEM M ODEL
tradeoff between communication and computing, and designed
a broadband aggregation algorithm for analog AirFL to reduce
the communication latency.
)R[ZZKX
Moreover, integrated sensing and computing has also be- S Estimating targets’ reflection coefficients

come the research focus with the widespread applications (9 C Computing global model for AirFL

of sensing services. In [25], a novel distributed framework C Decoding communication signal

was developed by combing edge computing and distributed

:GXMKZ
deep learning for urban environment sensing, which can ef-
fectively improve the computing efficiency and the sensing
accuracy. The authors in [26] developed a multiple-input
9KTYUXY
multiple-output (MIMO) over-the-air computation technique
2UIGRSUJKRZXGOTOTM
for realizing multi-modal sensing with high-mobility. The
equalization scheme based on channel feedback was designed 濷 *GZGYKZ

to minimize the computation error. In addition, a real-time big

/TZKXLKXKTIKYZXKGS 9KTYOTMYZXKGS
data analytical framework for remote satellite sensing service )USS[TOIGZOUTYZXKGS )USV[ZOTMYZXKGS
was put forward to enhance the computation efficiency in the
context of massive data [27]. Fig. 1. System model for 6G wireless network integrating sensing, computing
and communication.
Commonly, previous works attempt to integrate two func-
tions in an effective way. In order to support emerging intel-
ligent applications and services in 6G wireless networks, it is We consider a 6G wireless network, c.f. Fig. 1, consisting
necessary to integrate sensing, computing and communication of a BS equipped with N antennas, and K multi-function
simultaneously. In this paper, we aim to provide feasible and sensors equipped with M antennas each, where KM ≤ N .
effective integration schemes by exploiting available resources The range of interest (ROI), including target sensing area
of 6G wireless networks. The contributions of this paper are and interference area, is partitioned into multiple cubes with
three-fold: equal size, each of which denotes a pixel point1 [28]. Without
loss of generality, there exist O clutters2 within interference
1) We put forward a general design framework of ISCC for range around the targets as interference for sensing, such a 6G
6G wireless networks. Massive MIMO techniques enable wireless network has three fundamental tasks, namely target
the networks to perform multiple-target sensing, multiple- sensing, model computing, and information communication,
dimension computing, and multiple-stream communica- which are collaboratively completed by multiple intelligent
tion simultaneously over the same wireless resources. sensors and a multi-antenna BS. Specifically, based on the
2) We design two typical joint beamforming algorithms for
1 Note that the size of the pixel represents the accuracy of sensing. In
ISCC from the perspectives of maximizing the overall
performance subject to transmit power constraints and fact, pixels are not necessarily cubic, and different side lengths can be set
individually to change the accuracy of different dimensions. According to the
minimizing the total transmit power consumption subject requirement of sensing accuracy, the size and the shape of the pixels can be
to performance requirements, respectively. determined based on the wavelength of the signal, the number of BS antennas,
3) We reveal the impacts of key system parameters on the overall size of the ROI and the wireless environment.
2 In general, the clutters can be considered as scatters over wireless channels.
the overall performance of ISCC, and provide useful Hence, the number of clutters and channel state information can be acquired
guideline for practical system design. by channel estimation [29].
acquired environmental information3, each intelligent sensor transmit beamforming vectors for corresponding sensing sig-
first performs spatial directional beamforming to control the nal, computing signal, and communication signal, respectively.
transmit directions of the superposition of its sensing, comput- Consequently, the received signal at the BS is given by
ing, and communication signals, such that the signals reach K X
I K X
O I
X X X
the BS adaptively to the channel conditions. As for target y = ri Gk,i ak,i ssens ak,m ssens
k,i + ro Fk,o k,m
sensing, the BS receives the reflected signals containing the k=1 i=1 k=1 o=1 m=1
environmental information of the target area and then designs | {z } | {z }
Sensing Interference from clutters
the corresponding sensing receivers to estimate the reflection
K X
L K X
J
coefficients of targets4 . As for model computing, by exploiting X X
+ Hk bk,l scomp
k,l + Hk ck,j scomm n , (2)
k,j + |{z}
the computing ability of intelligent sensors that train local
k=1 l=1 k=1 j=1 Noise
models with their datasets, the BS designs the corresponding | {z } | {z }
Computing Communication
computing receivers to obtain the desired computing signals
to facilitate the global model via AirFL. As for information where n is the additive white Gaussian noise (AWGN) vector,
communication, the BS designs the corresponding communi- ri ∈ C, ∀i ∈ ΩI , and ro ∈ C, ∀o ∈ ΩO , are the reflection
cation receivers to decode the communication signals of each coefficients of the i-th target and the o-th clutter, respectively.
′
sensor for data transmission. With the above framework, we Gk,i = gk,i gk,iH
∈ CN ×M represents the cascaded channel
′
can integrate sensing, computing and communication by a in which gk,i and gk,i are the channel gains from the k-th
unified transceiver. sensor to the i-th target and from the i-th target to the BS,
′
Without loss of generality, we assume that H
respectively. Fk,o = fk,o fk,o ∈ CN ×M denotes the cascaded
the k-thh sensor constructs i the sensing signals ′
channel, where fk,o and fk,o are the channel gains from the
T
ssens
k = ssens sens sens
k,1 , . . . , sk,i , . . . , sk,I ∈ RI×1 to sense I targets, k-th sensor to the o-th clutter and from the o-th clutter to the
namely I pixel points hin the target area, generatesilocal model BS, respectively. Furthermore, Hk ∈ CN ×M is the channel
T
gain from the k-th sensor to the BS. Here, we assume that
parameters scomp
k = scomp comp comp
k,1 , . . . , sk,l , . . . , sk,L ∈ RL×1
Gk,i , Fk,o , and Hk remain constant within a time slot, but
as computingh signals to perform AirFL, i and logs dataT fade over time slots independently. In the following, we will
scomm
k = scomm comm comm
k,1 , . . . , sk,j , . . . , sk,J ∈ RJ×1 as discuss the performance metrics of the three fundamental tasks
communication signals to transfer the information in each one-by-one.
time slot. Note that the functions of theses three types of
signals are different, where ssens k,i is the sensing signal for the A. Sensing for Target Observation
i-th target from the k-th sensor, which does not carry any First, we deal with the sensing signal. To improve the
data, but is just used to estimate the reflection coefficients of accuracy of target sensing, all sensors collaborate to sense
the targets, scomp
k,l is the computing signal for the pre-processed the targets. Thus, a linear unbiased estimator is deployed at
l-th local model parameter at the k-th sensor, which generates the BS to estimate the reflection coefficients of the targets. To
from AI model trained by its local dataset, and scomm k,j is obtain an accurate reflection coefficient ri , i ∈ ΩI , for target
the communication signal for the j-th measured data at the sensing, the BS carries out receive beamforming to enhance
k-th sensor, which needs to be decoded exactly at the BS. the desired signal and suppress the co-channel interference.
For ease of analysis, it is assumed that the sensing signals, Hence, the estimated reflection coefficient r̂i of the i-th target
the computing signals, and the communication signals are can be expressed as
mutually independent  and all obey Gaussian distributed  with
 H    H r̂i = viH y
sens sens comp comp
unit norm, i.e., E sk,i sk,i = E sk,l sk,l = K K I
 X X X
 H  H
= vi ri sens H
Gk,i ak,i sk,i + vi rm Gk,m ak,m ssens
k,m
E scomm
k,j scomm
k,j = 1, ∀k ∈ ΩK , i ∈ ΩI , l ∈ ΩL , j ∈ ΩJ .
k=1 k=1 m=1,m6=i
Then, the k-th sensor constructs a superposition coded K X
X O I
X
transmit signal as + viH ro Fk,o ak,m ssens
k,m
I L J k=1 o=1 m=1
X X X  
xk = ak,i ssens
k,i + bk,l scomp
k,l + ck,j scomm
k,j , (1) K
X XL
comp
J
X
i=1 l=1 j=1 + viH Hk  bk,l s +
k,l ck,j scomm
k,j

| {z } | {z } | {z } k=1 l=1 j=1
Sensing Computing Communication
+ viH n, ∀i, (3)
where ak,i ∈ CM×1 , bk,l ∈ CM×1 , ck,j ∈ CM×1 are the
where vi ∈ CN ×1 is the sensing receive beamforming vector
3 The environmental information can be roughly known by environmental adopted at the BS for the i-th target. Following relevant
detection. For example, channel conditions can be obtained by channel
estimation [30] and the location of the ROI can be obtained by beam sweeping parameter estimation works, we adopt the mean squared error
[31]. (MSE) as the performance metric of target sensing. It is seen
4 Since all the pixel points of the ROI have the same size, shape and
from Eq. (3) that the r̂i is not only a function of ssens
k,i , but also
alignment, the reflection coefficients are the only recognition features in this
work. For example, the BS can perform target classification [32] or target the other sensing signals from different sensors for different
imaging [33] by accurately obtaining their reflection coefficients. targets, the computing signals, and communication signals. In
this context, we can minimize the MSE between the estimated computing local models is given by
reflection coefficient r̂i and the actual reflection coefficient ri K
to suppress the interference from other signals and achieve a 1 X comp
scomp
l = sk,l , ∀l ∈ ΩL (5)
high-accuracy target sensing. Thus, the MSE is given by K
k=1
n o comp
MSEsens
i = E (r̂i − ri ) (r̂i − ri )H where sk,l is the l-th computing signal contributed by the
2 pre-processed local model parameter at the k-th sensor, and
K
X each local model parameter is multiplied by a weighting factor
= viH Gk,i ak,i − 1 Ri2
ξk = |D k|
|D| before transmission. To reduce the computation
k=1
K I
error in the presence of interference and channel fading, we
X X 2 need to implement receive beamforming at the BS. As a result,
2
+ Rm viH Gk,m ak,m
the received signal for AirFL at the BS can be expressed as
k=1 m=1,m6=i
K X
O X
I zH
+
X
Ro2 viH Fk,o ak,m
2 ŝcomp
l = l
y
K
k=1 o=1 m=1 K K L
  zH X zH X X
K L J = l Hk bk,l scomp
k,l + l
Hk bk,i scomp
k,i
X X 2 X 2 K K
+  viH Hk bk,l + viH Hk ck,j  k=1 k=1 i=1,i6=l
k=1 l=1 j=1 K X
I
zH X
2 + l
ri Gk,i ak,i ssens
+ σn2 viH , ∀i, (4) K k,i
k=1 i=1
where σn2is the noise power, and Rm is the root mean squared zH XK X O I
X
(RMS) value of the reflection coefficient rm , ∀m ∈ {ΩI ∪ΩO } + l
ro Fk,o ak,m ssens
k,m
K
according to the a-priori probability of occurrence for each k=1 o=1 m=1
K X J
type [32]. To present more intuitively the impact of reflection zH X zH
+ l
Hk ck,j scomm
k,j +
l
n, ∀l, (6)
coefficient estimation, we provide an application result on K K
k=1 j=1
target imaging [33]. As is shown in Fig. 2, the higher the
sensing accuracy (the smaller MSE between the estimated where zl ∈ CN ×1 is the computing receive beamforming
reflection coefficient and the actual reflection coefficient), the vector at the BS for the l-th model parameter. In general, the
more accurate the target imaging. distortion of the model aggregation is measured by the MSE
comp comp
between sl and ŝl , which can be expressed as
n o
comp
comp
H
MSEcomp
l = E ŝ comp
l − s l ŝ comp
l − s l
K
1 X H 2
= zl Hk bk,l − 1
K2
k=1
L K
1 X X 2
+ zH
l Hk bk,m
K2
m=1,m6=l k=1
K I
1 XX 2 H 2
+ Ri zl Gk,i ak,i
K2 i=1
k=1
K XO I
1 X X 2
+ Ro2 zH
l Fk,o ak,m
K2
k=1 o=1 m=1
K J
1 XX H 2 σn2 2
Fig. 2. Target Imaging results under different sensing accuracy.
+ zl Hk ck,j + kzl k , ∀l. (7)
K2 K2
k=1 j=1

Herein, we also provide a classical application case to vividly

present the function of computing, i.e. image recognition based
on AirFL[34]. It is seen from Table I that the higher computing
B. Computing for Model Aggregation
accuracy (the smaller computing MSE between the aggregated
Secondly, we handle the computing signal. In a time slot, a computing signal and the desired global model), the more
shared learning process for AirFL with a global model scomp ∈ accurate the image recognition.
RL×1 is trained cooperatively by multiple intelligent sensors
C. Communication for Information Transmission
[23], [24]. Each intelligent sensor has its own local dataset
Dk , ∀k ∈ ΩK , and the sum of local datasets is D, where Finally, we discuss the communication signal. To promote
PK the quality of the communication signal, the BS performs cor-
|D| = |Dk |. Thus, the desired global model at the BS by responding receive beamforming to mitigate the interference.
k=1
 TABLE I
H ANDWRITTEN - DIGIT IDENTIFICATION RESULTS UNDER DIFFERENT COMPUTING ACCURACY ON IID MNIST DATASET TRAINED BY CNN NETWORK

Data image
Computing accuracy: High (small MSE) 0 5 1 9 4 6 4
Computing accuracy: Medium (medium MSE) 6(×) 5 1 9 4 6 4
Computing accuracy: Low (large MSE) 0 8(×) 1 4(×) 8(×) 4(×) 4

Data image
Computing accuracy: High (small MSE) 7 1 8 5 2 9 3
Computing accuracy: Medium (medium MSE) 7 1 8 5 2 4(×) 3
Computing accuracy: Low (large MSE) 4(×) 1 8 8(×) 2 4(×) 8(×)

Hence, the received signal at the BS for the communication cation systems in terms of signal waveform, transceiver design
signal scomm
k,j , ∀k ∈ ΩK , j ∈ ΩJ , is given by and function realization. Firstly, multi-user communication
systems only consider one kind of signal for communication
comm
yk,j = uH
k,j y purpose, while in our proposed integrated system, sensing,
= uH comm
k,j Hk ck,j sk,j computing and communication signals have different types
K
X J
X and serve different purposes. Secondly, signals of the same
+ uH comm
k,j ci,n si,n type from different users interfere with each other at the BS
i=1,i6=k n=1,n6=j in multi-user communication systems, and thus transceiver
K X
X L is designed to reduce the inter-user interference. In contrary,
comp
+ uH
k,j Hi bi,l si,l for the proposed integrated system, signals of different types
i=1 l=1 interfere with each other, but signals of the same type may
XK X I help each other. In this case, the designed transceiver not
+ uH
k,j ri Gk,i ak,i ssens
k,i only needs to decrease the interference among signals of
k=1 i=1 different types, but also needs to increase the cooperative gain
XK X O XI of signals of the same type in terms of sensing and computing.
H
+ uk,j ro Fk,o ak,m ssens
k,m + uH
k,j n,(8) Finally, although different performance metrics such as SINR,
k=1 o=1 m=1 data rate and MSE, are chosen for optimization in multi-
where uk,j ∈ CN ×1 is the communication receive beamform- user communication systems, they are all used to improve the
ing vector of the j-th communication signal from the k-th performance of the communication function. While for the
sensor. Since the received SINR determines the quality of the proposed integrated system, we select the appropriate perfor-
communication signal, it is usually regarded as the perfor- mance metrics based on the functions of sensing, computing
mance metric for communication. The signal-to-interference- and communication. By solving the formulated multi-objective
plus-noise ratio (SINR) related to scomm
k,j can be expressed as optimization problem (MOOP), the desired performance of
2
three different functions can be effectively achieved.
uH
k,j Hk ck,j
Γk,j = K J
, III. J OINT D ESIGN OF S ENSING , C OMPUTING AND
P P 2 2
uH
k,j Hi ci,n + Xk,j + σn2 uk,j C OMMUNICATION
i=1,i6=k n=1,n6=j
This section aims at jointly designing transmit and receive
(9)
where beamforming vectors for ISCC in 6G wireless networks.
K X L K X I Since 6G wireless networks have different priorities among
X 2 X 2
Xk,j = H
uk,j Hi bi,l + 2 H
Rm uk,j Gi,m ai,m sensing, computing and communication for various application
i=1 l=1 i=1 m=1 scenarios, we formulate two typical categories of MOOPs
K
XX X O I to investigate the trade-offs among them. The first one is a
2
+ Ro2 uHk,j F i,o ai,m . (10) weighted overall performance maximization (WOPM) subject
i=1 o=1 m=1 to the maximum budget of transmit power. The second one
is a total transmit power minimization (TTPM) while guaran-
It is observed from (4), (7), and (9) that the performance of
teeing the quality of service (QoS) requirements on sensing,
sensing, computing, and communication are jointly determined
computing, and communication, respectively.
by the transmit beamforming vectors ak,i , bk,l , and ck,j at
the sensors, and receive beamforming vectors vi , zl , and uk,j
at the BS. Although the three performance metrics are all A. Weighted Overall Performance Maximization Design
desirable to the system, they are competitive for the system Now, we first study three individual single-objective opti-
resources. Hence, it is desired to establish a joint beamforming mization problem (SOOP) for sensing, computing, and com-
design framework to enhance the overall performance of ISCC munication in 6G wireless networks, respectively, which severs
over the limited resources in 6G wireless networks. as building blocks for the formulation of WOPM design.
Remark: It is worthy pointing out that our proposed inte- The first SOOP aims at minimizing the weighted sum-MSE
grated system is inherently different from multi-user communi- for sensing subject to the maximum transmit power budget,
 which is formulated as S-1: Weighted Sum-MSE of Sensing Γk,j and the minimum MSE (MMSE) ecomm k,j for the commu-
Minimization: nication signal scomm
k,j can be expressed as
I −1
X 1 + Γk,j = (ecomm
k,j ) , ∀k, j. (14)
min θisens MSEsens
i (11)
ak,i ,bk,l ,
ck,j ,vi i=1
I
X L
X J
X
s.t. C1: kak,i k2 + kbk,l k2 + kck,j k2 ≤ Pmax,k , Proof: Please refer to Appendix A.
i=1 l=1 j=1

where constant θisens is the weighted coefficient related to

MSEsens According to Theorem 1, we can transform the objective
i , and Pmax,k is the maximum transmit power budget
at the k-th sensor. The objective in S-1 is to minimize the function of S-3’ to
weighted summation of MSEs between estimated reflection K X
X J
comm


coefficients and actual ones for all I targets depicted in (4), min θk,j log2 ecomm
k,j , (15)
ak,i ,bk,l ,
and constraint C1 is the transmit power limitation at the k-th ck,j ,uk,j k=1 j=1
sensor. Note that each SOOP only concerns one performance which is equivalent to minimizing the weighted sum-MSE of
metric of the proposed integrated sensing, computing and communication MSEcomm with the MMSE receiver uk,j =
k,j
communication without considering others. For example, it Ξ−1 Hk ck,j , given in Appendix A. Thus, (15) can be refor-
is seen from S-1 that the communication beam ck,j and mulated as
the computing beam bk,l will naturally be zero to minimize K X
J
X

the objective, while the sensing beam ak,i will not be zero min comm
θk,j log2 MSEcomm (16)
2 k,j
K
P ak,i ,bk,l ,ck,j
due to the term viH Gk,i ak,i − 1 Ri2 in (4). Hence, no k=1 j=1
k=1 However, the transformed objective function (16) still remains
additional conditions related to communication and computing
beams are required. The second SOOP is to minimize the non-convex caused by the structure of sum of logarithmic
weighted sum-MSE for computing under the limitation of function. To this end, we introduce a weight variable ωk,j
transmit power consumption, which can be mathematically for the MSEcomm
k,j [36], and then (16) is transformed as
expressed as X J
K X
 
S-2: Weighted Sum-MSE of Computing Minimization: min comm
θk,j ωk,j MSEcomm
k,j − log2 (ωk,j ) . (17)
ak,i ,bk,l ,
L ck,j, ,ωk,j k=1 j=1
X
min θlcomp MSEcomp
l As for unconstrained optimization problems in terms of the
ak,i ,bk,l ,
ck,j ,zl l=1 weight variable ωk,j , by letting the first-derivative of the
s.t. C1, (12) objective function (17) equal to zero, we can obtain the optimal
solution as follows
where constant θlcomp
is the weighted coefficient relevant to
−1
∗
MSEcomp
l , and the objective is minimizing the weighted sum- ωk,j = ln 2 · MSEcommk,j , (18)
mation of the computation distortions for L model parameters,
which makes (16) and (17) are equivalent. After the above
which was defined in (7). The third SOOP focuses on the
conversion, S-3’ can be transformed as
weighted sum-rate maximization for communication subject
to the transmit power limitation, which is formulated as
S-3’: Weighted Sum-Rate of Communication Maximization:
S-3: Weighted Sum-Modified MSE of Communication Min-
K X
X J imization:
comm
max θk,j rk,j K X
J
ak,i ,bk,l , X  
k=1 j=1
ck,j ,uk,j min comm
θk,j ωk,j MSEcomm
k,j − log2 (ωk,j )
s.t. C1, (13) ak,i ,bk,l ,ck,j ,
uk,j ,ωk,j k=1 j=1

where rk,j = log2 (1 + Γk,j ) is the achievable rate of the j- s.t. C1. (19)
comm
th communication signal from the k-th sensor and θk,j is To jointly optimize the three performance of sensing, comput-
the weighted coefficient associated with rk,j . S-3’ is a clas- ing, and communication, we adopted the weighted sum method
sical weighted sum-rate maximization problem which is non- to formulate the corresponding MOOP [37], [38]. The WOPM
convex due to the complicated objective function involving is formulated as follows:
log functions of fractions [35]. Moreover, the transmit beam-
forming vectors {ak,i , bk,l , ck,j } and communication receive
beamforming vector uk,j are coupled in the constraints and M-1: Weighted Overall Performance Maximization:
objective function. To handle this issue as well as facilitate
the design, we transform S-3’ into its equivalent log-MSE min α1 Ψ1 + α2 Ψ2 + α3 Ψ3
ak,i ,bk,l ,ck,j ,
minimization problem via the following theorem. vi ,uk,j ,zl ,ωk,j
Theorem 1: The relationship between the received SINR s.t. C1, (20)
 K I
where Ψp , p = 1, 2, 3 is the normalized objective function5 of respectively, where Φ =
P P
Rn2 Gk,n ak,n aH H
k,n Gk,n +
S-1, S-2 and S-3, respectively. αp ≥ 0 is the priority of the k=1 n=1,n6=i
p-th objective function, which represents the preference of the K P
P O P
I K P
P L
Ro2 Fk,o ak,m aH H
k,m Fk,o + Hk bk,l bH H
k,l Hk +
system operator and is satisfied with α1 + α2 + α3 = 1. By k=1 o=1 m=1 k=1 l=1
varying the value of αp , M-1 can yield different solutions. Note PK P J
2 2
K
P
that M-1 is equivalent to S-p when αp = 1 and αq = 0, ∀p 6= q, Hk ck,m cH H
k,m Hk + σn IN , Φi = Ri Gk,i ak,i ,
k=1 m=1 k=1
which means M-1 is a general formulation of S-1, S-2 and S-3. K P
P I
Hence, we aim at solving M-1 with given priorities αp in the and Ξ = Φ + Rn2 Gk,n ak,n aH H
k,n Gk,n is defined in
k=1 n=1
following. Appendix A. Next, for the second subproblem in terms of
Since multiple variables are inter-coupled in the objective optimizing weight variables {ωk,j }, the optimal solution is

−1
function of M-1, i.e., transmit beams {ak,i , bk,l , ck,j }, receive ∗
given in (18), i.e., ωk,j = ln 2 · MSEcomm
k,j . Finally, with
beams {vi , uk,j , zl }, and weight variables {ωk,j }, M-1 is the MMSE receivers and weight variables, the last sub-
non-convex, which makes it impossible to obtain an optimal problem of optimizing transmit beams {ak,i , bk,l , ck,j } for
solution in polynomial time. In this context, we turn to find maximizing the weighted overall performance is a standard
a sub-optimal solution for exploring the trade-off relationship convex quadratic constrained quadratic programming (QCQP)
among sensing, computing, and communication in 6G wireless problem, which can be solved by an inter-point method (IPM)
networks. By examining M-1, although it is not a joint convex [[42], Chapter 11]. At first, we utilize the barrier method
function of all variables, it is a convex one in terms of transmit to transform this subproblem into an unconstrained convex
beams, receive beams, and weight variables, respectively. optimization problem by adding the logarithmic barrier func-
Based on this observation, an alternating optimization (AO) tion. Then, the Newton’s method can be applied to obtain the
method is applied to divide M-1 into three subproblems, i.e., solution. Specifically, the transformed unconstrained convex
optimizing receive beams by fixing transmit beams and weight optimization problem can be written as
variables, optimizing weight variables by fixing transmit and K
X
receive beams, and optimizing transmit beams by fixing re- min ε(α1 Ψ1 + α2 Ψ2 + α3 Ψ3 ) − log (−fkcon ), (25)
ceive beams and weight variables. In particular, enabled with ak,i ,bk,l ,ck,j
k=1
the AO method, solution procedure for M-1 will stop until I L J
P P P
the objective value converges in the iterations [39]. Let us where fkcon = kak,i k2 + kbk,l k2 + kck,j k2 − Pmax,k
first address the subproblem of optimizing receive beams i=1 l=1 j=1
{vi , uk,j , zl } while other variables remain fixed. By applying and ε > 0 is a barrier parameter that sets the accuracy of the
the Karush-Kuhn-Tucher (KKT) conditions in M-1, i.e., approximation. Then, we compute the variables’ gradients and
! Hessian for Newton’s method as follows
P3  
2ak,i
∂
∂vi αp Ψp = 0 ⇒ ∂v ∂
i
(MSEsens
i ) = 0, ∇ak,i = 2ε Tak,i ak,i − Ri2 α̃1 θisens GHk,i vi + fkcon ,
p=1  
2bk,l
P3
!

∇bk,l = K 2ε
2 Tbk,l bk,l − α̃2 θlcomp HHk zl + fkcon , (26)
∂ ∂
MSEcomp
 
αp Ψp = 0 ⇒ ∂z = 0, (21) 2c
∂zl
p=1
l l ∇ck,j = 2ε Tck,l ck,j − α̃3 θk,j
comm
ωk,j HH k,j
k uk,j + f con ,
! k
3

∂ P
αp Ψp = 0 ⇒ ∂u∂k,j MSEcomm = 0, and
∂uk,j k,j
4 2
p=1 ∇2ak,i = 2εTak,i + 2 ak,i aTk,l +
fkcon I,
(fkcon )
we can acquire the optimal receive beamforming vectors, also ∇2bk,l = 2ε b 4 T 2
K 2 Tk,l + (f con )2 bk,l bk,l + fkcon I, (27)
called MMSE receiver, as below k
4
∇2ck,j = 2εTck,l + con c cT + 2con I,
" K
! #−1 (f )2 k,j k,j fk
k
X
H H
vi = Φi ak,i Gk,i + Φ Φi , (22) with
k=1 O
!
X
K
X Tak,i = α̃1 θisens Ri2 GH H
k,i vi vi Gk,i + Ro2 FH H
k,o vi vi Fk,o
zl =Ξ−1 Hk bk,l , (23) o=1
L O
!
k=1
α̃2 X comp X
and + θl Ri2 GH H
k,i zl zl Gk,i + Ro2 FH H
k,o zl zl Fk,o
K2
−1 l=1 o=1
uk,j = Ξ Hk ck,j , (24)
K
XX J
+ α̃3 comm
θk,j ωk,j Ri2 HH H
k uk,j uk,j Hk ,
k=1 j=1
5 Since the objective value of S-1, S-2 and S-3 have different ranges, it K X
X J O
X
is desired to perform the normalization for each objective function in order
to coordinate the performance of sensing, computing, and communication
+ α̃3 comm
θk,j ωk,j Ro2 FH H
k,o uk,j uk,j Fk,o , (28)
as well as to facilitate the convergence of objective value for M-1. In this k=1 j=1 o=1
paper, we define Fp , p = 1, 2, 3, as the objective function of S-p and Ψp =
Fp − Fp∗ / Fp∗ , where Fp∗ are the corresponding performance limits of


sensing, computing and communication in M-1, which can be obtained by
respectively solving S-p by applying the same algorithm for solving M-1.
 I
2
X comp H
Algorithm 1 : WOPM-based Joint Design of ISCC for 6G
Tb
k,l = α̃1 K θisens HH H
k vi vi Hk + α̃2 θl Hk zl zH
l Hk wireless networks
i=1 comp
Input: N, M, K, O, I, L, J, σn2 , θk,i sens
, θk,l comm
, θk,j , Pmax,k , αp
K X
J
X Output: ak,i , bk,l , ck,j , vi , uk,j , zl
+ α̃3 K 2 comm
θk,j ωk,j HH H
k uk,j uk,j Hk , (29) (0)
1: Initialize Iteration index t = 1, ak,i = bk,l = ck,j =
(0) (0)
k=1 j=1 q
Pmax,k
[ 3M
, 0, . . . , 0]T , ∀k, i, l, j;
and 2: repeat
(t) (t−1) (t−1)
I
X L
α̃2 X comp H H 3: Compute vi according to (22) with ak,i , bk,l and
Tck,j = α̃1 θisens HH H
k vi vi H + θl Hk zl zl Hk (t−1)
ck,j ;
i=1
K2 (t) (t−1) (t−1)
l=1 4: Compute zl according to (23) with ak,i , bk,l and
comm
+ α̃3 θk,j ωk,j HH H
k uk,j uk,j Hk , (30) (t−1)
ck,j ;
(t) (t−1) (t−1)
where ∇x and ∇2x denote the gradient and the Hessian of the 5: Compute uk,j according to (24) with ak,i , bk,l and
(t−1)
variable for the objective function (25), respectively. Moreover, ck,j ;
α1 (t) (t) (t) (t)
α̃1 = F , α̃ = Fα2∗ , and α̃3 = F α3
. With the gradients and 6: Compute ωk,j according to (18) with vi , zl , uk,j ,
| 1∗ | 1 | 2| | 3∗ | (t−1) (t−1) (t−1)
Hessian, the transmit beams can be computed by the following ak,i , bk,l and ck,j ;
7: Initialize Iteration index m = 1, barrier parameter ε = 1,
updating rules: amplification factor ν=10;
(t+1) (t)
 −1 8: repeat
ak,i a
= ak,i − ηk,i ∇2ak,i ∇ak,i (31) 9:
a(m) b(m)
Select feasible step sizes ηk,l , ηk,l and ηk,l ;
c(m)

 −1 (m)
(t+1) (t) 10: Compute ak,i according to (31) with
bk,l = bk,l − ηk,lb
∇2bk,l ∇bk,l (32) a(m) (m−1) (t) (t) (t) (t)
ηk,l ,ak,i ,vk,i , zl , uk,j and ωk,j ;
 −1 (m)
(t+1) (t) 11: Compute bk,l according to (32) with
ck,j c
= ck,l − ηk,l ∇2ck,j ∇ck,j (33) b(m) (m−1) (t) (t) (t) (t)
ηk,l ,bk,l ,vk,i , zl , uk,j and ωk,j ;
a b c (m)
where ηk,i , ηk,i and ηk,i are the step sizes of the transmit 12: Compute ck,j according to (33) with
c(m) (m−1) (t) (t) (t) (t)
beams ak,i , bk,l and ck,j , respectively. In summary, the joint ηk,l ,ck,j ,vk,i , zl , uk,j and ωk,j ;
design of ISCC for 6G wireless networks based on the WOPM 13: if Meeting the centrality condition then
is concluded as Algorithm 1. 14: Update ε = ε ∗ ν;
15: end if
16: Update m = m + 1;
B. Total Transmit Power Minimization Design 17: until The duality gap converges
(t) (m) (t) (m) (t) (m)
Now, let us focus on another MOOP with the goal of 18: Update ak,i = ak,i , bk,l = bk,l and ck,j = ck,j ;
minimizing total transmit power consumption but ensuring the 19: Update t = t + 1;
20: until The objective value converges
QoS requirements on sensing, computing, and communication,
respectively. The design is formulated as follows:
M-2: Total Transmit Power Minimization:
 
K I L J
X X 2
X 2
X 2 receive beams by fixing transmit beams, which are given in
min  kak,i k + kbk,l k + kck,j k  (22), (23), and (24). For another subproblem for optimizing
ak,i ,bk,l ,ck,j
,vi ,uk,j ,zl k=1 i=1 l=1 j=1
transmit beams by fixing receive beams, we need to handle
sens
s.t. C2: MSEi ≤ δi , ∀i ∈ ΩI , the non-convex constraint C4, which is equivalent to
comp
C3: MSEl ≤ χl , ∀l ∈ ΩL , 1 2
uHk,j Hk ck,j ≥
C4: Γk,j ≥ γk,j , ∀k ∈ ΩK , j ∈ ΩJ , (34) γk,j
K J
where δi > 0 is the given tolerable maximum sensing X X 2 2
uH k,j Hi ci,n + Xk,j + σn2 uk,j . (35)
distortion for i-th target object, χl > 0 is the given maximum
i=1,i6=k n=1,n6=j
tolerable computation error for the l-th model parameter,
and γk,j > 0 denotes the required minimum SINR for the To address the non-convexity, we introduce an auxiliary vari-
communication signal scomm k,j . The objective function of M-2 is
able Ck,j = ck,j cH k,j and substitute it into (35), which is given
minimizing the total transmit power of all sensors, and the con- by
straints C2, C3, and C4 are the QoS requirements for sensing, 1 H H


computing, and communication, respectively. However, M-2 γk,j tr uk,j Hk Ck,j Hk uk,j ≥ (36)
is not convex because of the variables coupling in constraints K J
X X

C2, C3 and C4, i.e., transmit beams {ak,i , bk,l , ck,j } and tr uH H 2 2
k,j Hi Ci,n Hi uk,j + Xk,j + σn kuk,j k .
receive beams {vi , uk,j , zl }, which makes it hard to find the i=1,i6=k n=1,n6=j
optimal solution in the polynomial time. Thus, it is also desired
to divide M-2 into two subproblems and then solve it by Based on this, the subproblem for the transmit beamforming
the AO method. Considering the balance between the system optimization can be converted to a standard semi-definite
performance and the computational complexity, it is feasible programming (SDP) problem as follows
to apply the MMSE receivers in the subproblem for optimizing M-2’: Equivalent Subproblem of Transmit Beamforming
Optimization: Algorithm 2 : TTPM-based Joint Design of ISCC for 6G
  wireless networks
K I L J
X X X X Input: K, N, M, I, O, L, J, σn2 , P0 , δi , χl , γk,j
min  kak,i k2 + kbk,l k2 + tr (Ck,j ) Output: ak,i , bk,l , ck,j ,vi ,zl ,uk,j
ak,i ,bk,l ,Ck,j
k=1 i=1 l=1 j=1 (0) (0) (0)
1: Initialize iteration index t = 1, ak,i = bk,l = ck,j =
sens q
s.t. C5: MSEi ≤ δi , ∀i ∈ ΩI , P0
[ 3M , 0, . . . , 0]T , ∀k, i, l, j;
comp
C6: MSEl ≤ χl , ∀l ∈ ΩL , 2: repeat
(t) (t−1) (t−1)
3: Compute vi according to (22) with ak,i , bk,l and
C7: (36), (t−1)
ck,j ;
C8: Ck,j  0, ∀k ∈ ΩK , j ∈ ΩJ , (t) (t−1) (t−1)
4: Compute zl according to (23) with ak,i , bk,l and
C9: Rank (Ck,j ) = 1, ∀k ∈ ΩK , j ∈ ΩJ , (37) (t−1)
ck,j ;
sens (t) (t−1) (t−1)
where MSEi = MSEsens − 5: Compute uk,j according to (24) with ak,i , bk,l and
i (t−1)
PK PJ 2 K P
P J
ck,j ;
viH Hk ck,j + H H
tr vi Hk Ck,j Hk vi 6:
(t) (t) (t)
Obtain {ak,i , bk,l , Ck,j } by solving M-2” with fixed
k=1 j=1 k=1 j=1
comp (t) (t) (t)
and MSEl = MSEcomp
l − {vi , zl , uk,j };
(t) (t)
1
P J
K P 2 1
P J
K P
7: Obtain ck,j by EVD on Ck,j according to (39);
K2 zH
l Hk ck,j + K2 tr zH H
l Hk Ck,j Hk zl . 8: t = t + 1;
k=1 j=1 k=1 j=1 9: until convergence
Note that constraint C9 for the rank-one limitation of Ck,j is
non-convex, which makes M-2’ still non-convexity. To this
end, we adopt the semi-definite relaxation (SDR) technique,
namely discarding constraint C9. As a result, M-2’ is restated As a result, the objective value of M-1 is monotonically
as non-increasing during the iterations. Furthermore, due to the
M-2”: Transformed Subproblem of Transmit Beamforming transmit power constraint at the sensors, the objective value of
Optimization: M-1 is lower bounded. Hence, based on the monotone bounded
  convergence (MBC) theorem [40], Algorithm 1 would con-
K
X X I XL XJ
verge for a suitable number of iteration. Next, we analyze
min  kak,i k2 + kbk,l k2 + tr (Ck,j )
ak,i ,bk,l ,Ck,j the convergence property of Algorithm 2. First, the adopted
k=1 i=1 l=1 j=1
MMSE receivers {vi , zl , uk,j } is effective to ensure the QoS
s.t. C5 − C8. (38) requirements, which can help enhance the system perfor-
It is found that M-2” is a joint convex problem in terms of mance. Moreover, since M-2” is convex for transmit beams
transmit beams {ak,i , bk,l , Ck,j } due to the convex constraints {ak,i bk,l , ck,j } with given MMSE receivers, it is feasible to
C5-C8 and convex objective function, and thus it can be get the optimal solution via CVX directly, which ensures the
directly solved via CVX [41]. It is worth mentioning that for obtained objective value in the (t + 1)-th iteration is less than
the rank of the optimal solution to M-2”, we have following that in the t-th iteration. That implies the total transmit power
theorem: consumption is monotonically non-increasing in the iterations.
∗ Moreover, due to the QoS requirements in constraint C2,
Theorem 2: The rank  of the
 optimal solution C k,j to M-2”
∗ C3, and C4, the total transmit power consumption is lower
always satisfies Rank Ck,j = 1, ∀k, j.
bounded. Thus, Algorithm 2 would converge according to the
Proof: Please refer to Appendix B.
MBC theorem.
According to Theorem 2, we can perform eigenvalue de-
composition (EVD) to get the unique optimal solution c∗k,j of Complexity Analysis: Since Algorithm 1 and Algorithm 2
M-2. The EVD operation can be presented as are both iterative, the execution steps are the same for each
q iteration. Thus, we focus on analyzing the per-iteration com-
∗ max ∗ max plexity of the proposed algorithms. By noticing the operations
ck,j = λk,j (Ck,j )ξk,j , (39)
of algorithms, it is observed that the dominating computational
where λmax ∗
k,j (Ck,j ) denotes the maximum eigenvalue of Ck,j
∗
complexities arise from the steps 7-17 for Algorithm 1, i.e.,
max
and ξk,j represents the corresponding unit eigenvector. finding optimal {a(t) , b(t) , c(t) } of M-1, and the step 6 for
k,i k,l k,j
Hence, the joint design of ISCC for 6G wireless networks (t) (t) (t)
Algorithm 2, i.e., finding optimal {ak,i , bk,l , Ck,j } of M-2”.
based on TTPM is summarized as Algorithm 2.
Since M-1 and M-2” only contain second-order cone (SOC)
or linear matrix inequalities (LMI) constraints, they could
C. Convergence and Complexity Analysis of Proposed Algo- both be solved by a standard IPM [42]. As such, the worst-
rithms case runtime by IPM can be used to depict the computa-
Herein, let us discuss the property of convergence and tional complexities of algorithm [43]. Specifically, M-1 has
complexity for the proposed two algorithms. K SOC constraints of dimension 1, and the decision variable
Convergence Analysis: For Algorithm 1, since M-1 is con- n1 = O(K 2 M 2 ). For M-2”, it has KJ LMI constraints of
vex for each optimization variable while fixing the others, the dimension 1, KJ LMI constraints of dimension M , I + L
optimal solutions of every subproblem can be found. In fact, SOC constraint of dimension 1, and the decision variable
the solutions of each iteration are feasible in the next iteration. n2 = O(KM 3 ). As a result, for the solution with a given
precision ǫ > 0, the worst-case complexities of Algorithm reduces while the performance of communication gradually
1 and Algorithm 2 for per iteration are in the order of improves as α3 increases. Hence, it makes sense to select
ln(1/ǫ)ς1,2 , c.f. Table II. an appropriate set of priorities to balance the performance of
sensing, computing, and communication. Moreover, it is found
IV. S IMULATION R ESULTS that there is a big gap between the SNR at 5 dB and 0 dB
for all three performance metrics, which means increasing the
In this section, we conduct extensive simulations to verify SNR at the sensors can bring more gain for the overall system
the effectiveness of the proposed algorithms. Without loss of performance.
generality, all sensors are assumed being randomly distributed Fig. 5 shows the influence of the number of sensors K with
within the cell radius. The pass loss is modeled as PLdB = different numbers of BS antennas N on the performance of
128.1 + 37.6 log10 (d) [44], where d (km) is the distance Algorithm 1. For a given SNR at the sensors, Algorithm 1 with
between the transmitter and receiver. Moreover, it is assumed more antennas at the BS can get a better overall performance.
that all sensors have the same QoS requirements and the same This is because a higher spatial multiplexing gain provided
maximum transmit power budget, namely γk,j = γ0 , δi = δ0 , by the extra antennas are exploited to effectively enhance the
χl = χ0 , and Pmax,k = P0 . We use SNR = 10 log10 (P0 /σn2 ) performance. Furthermore, it is seen that with the increasing
to denote the transmit signal-to-noise ratio (SNR) (in dB) at of the number of sensors, both the sensing error and the
the sensors. Unless otherwise stated, the simulation parameters computation error diminishes while the communication rate
are summarized in Table III. increases. That implies the increment of the number of sensors
is conducive to acquire a more accurate estimation on the
10-3 reflection coefficient for the target object, a more accurate
1.8 100
Sensing error with SNR = 5 dB

1.6
Sensing error with SNR = 10 dB
Computation error with SNR = 5 dB
90
model aggregation for AirFL, and a higher weighted sum-rate
Computation error with SNR = 10 dB
Communication rate with SNR = 10 dB for information communication, which is quite attractive for
1.4 Communication rate with SNR = 5 dB 80
6G wireless networks with massive sensors.
Weighted Sum-rate (bit/s/Hz)

1.2 70 In Fig. 6, we show the performance of Algorithm 1 over

Weighted Sum-MSE

1 60
three baseline beamforming algorithms. They are a fixed-
MMSE (F-MMSE) algorithm with fixed MMSE receivers only
0.8 50
relevant to the channels, an AO-based match filtering beam-
0.6 40 forming (MFBF) algorithm with the match filter receivers,
0.4 30
an AO-based zero-forcing beamforming (ZFBF) algorithm
with the zero-forcing transmitters, respectively. It is seen
0.2 20
that the F-MMSE algorithm performs the worst among the
0 10 other algorithms due to the fixed receiver, which limits the
0 2 4 6 8 10 12 14 16 18 20
Iteration Number performance. The MFBF algorithm outperforms the ZFBF
algorithm in the performance of sensing, but performs worse
Fig. 3. Convergence behavior of Algorithm 1. in the aspects of computing and communication. Moreover,
it is found that Algorithm 1 performs the best for all three
First, we give the convergence performance of Algorithm performance since it obtains transmit beams and receive beams
1 with different SNRs at the sensors. As is shown in Fig. 3, both in the optimal way, which demonstrates the superiority
the sensing error and the computation error decrease while the of Algorithm 1.
communication rate increases monotonically in the iterations, Fig. 7 checks the convergence performance of Algorithm 2
and they converge to a stable equilibrium point within few under different required minimum SINRs of communication.
iterations on average. Hence, the implementation cost of It is observed that the total transmit power is progressively
Algorithm 1 is bearable for practical systems. lowering in the iterations, and then converges within no more
Then, we investigate the effect of priorities of sensing than 10 iterations on average under different required mini-
α1 , computing α2 , and communication α3 of MOOP on the mum SINRs of communication, which verifies the feasibility
performance under different transmit SNRs at the sensors. of Algorithm 2.
Here, we present two different cases. As shown in Fig. 4, for Fig. 8 investigates the impacts of the maximum tolerable
case (a), we fix the weighted coefficient of communication sensing error δ0 , the maximum tolerable computation error
α3 = 1/3. In this case, α1 + α2 = 2/3, and then we change χ0 , and the required minimum SINR of communication γ0
the value of the priority of computing α2 to obtain different on the total transmit power consumption. It can be seen that
solutions. In case (b), we fix α1 = α2 while changing the the total transmit power decreases as the tolerable maximum
priority of communication α3 to get various results. It is computation error increases, since a larger χ0 represents a
observed that in case (a) with increasing of α2 , the three more relax constraint on the computing accuracy resulting in
performance metrics have different varying trends. Specif- less power consumption. Moreover, in the whole region of
ically, the computation error slowly decreases, the sensing χ0 , the case with δ0 = 0.001 and γ0 = 0.1 dB requires the
error gradually increases, while the communication rate first smallest total transmit power consumption, while the case with
decreases and then increases. Moreover, it is seen that in case δ0 = 0.01 and γ0 = 0.1 dB requires the biggest one. That
(b) the performance of sensing and computation gradually means a larger δ0 is less restriction on sensing accuracy and
 TABLE II
T HE WORST- CASE COMPLEXITIES OF PROPOSED ALGORITHMS FOR PER ITERATION

Algorithms Complexity is√in order of ln (1/ε) ς{1,2}

Algorithm 1 p ς1 = 2K n1 K + n31


Algorithm 2 ς2 = KJ (M + 1) + 2(I + L) · n2 · KJ M 3 + n2 M 2 + n2 + 1 + I + L + n22

TABLE III
S IMULATION PARAMETERS

Parameters Values
Number of BS antennas N = 64
Sensor K = 20, M = 3
Streams O = 1, I = 1, L = 1, J = 1
Cell radius 500 m
RMS of reflection coefficient Ri = Ro = 1
Weighted coefficient θisens = θlcomp = θk,j
sens
=1
Maximum transmit SNR at the sensors SNR = 5 dB
Noise powers σn2 = −50 dBm
QoS requirements γ0 = 0.1 dB, δ0 = 0.01, χ0 = 0.01
Performance priority α1 = α2 = α3 = 1/3

(a) =1/3 (b) =

10-3 3 10-3 1 2
95 3.5 100
Sensing error with SNR = 0 dB
9 Sensing error with SNR = 5 dB
Computation error with SNR = 0 dB 90
8 3 90
Computation error with SNR = 5 dB
Weighted Sum-rate (bit/s/Hz)

Weighted Sum-rate (bit/s/Hz)

Communication rate with SNR = 5 dB 85
7 Communication rate with SNR = 0 dB
2.5 80
Weighted Sum-MSE

Weighted Sum-MSE

legend both for case (a) and case (b)

80
6
75 2 70
5
70 1.5 60
4

3 65
1 50
2 60
0.5 40
1 55

0 50 0 30
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Priority of Computation, Priority of Communication, 3
2

Fig. 4. The system performance gains versus priorities with different SNRs for Algorithm 1.

a larger γ0 signifies higher requirement on communication. higher cost on radio frequency chains. Hence, it makes sense
Hence, it makes sense to loosen performance requirements to select a proper number of BS antennas for balancing the
appropriately for reducing the total power consumption at the performance and the cost.
sensors.
Finally, we compare the performance of different algorithms
Next, we study the influences of the numbers of clutters in the sense of minimizing the total transmit power in Fig. 10.
O and BS antennas N on the performance of Algorithm 2 There are four algorithms, namely a F-MMSE algorithm, an
in Fig. 9. It is found that the total transmit power signif- AO-based MFBF algorithms, an AO-based ZFBF algorithm,
icantly increases as the number of clutters increases, since and Algorithm 2. For all algorithms, the total transmit power
clutters as interference component is harmful to the three increases as the required minimum SINR on communication
performance metrics resulting in more power consumption. γ0 increases. This is because for given sensing and computing
Moreover, Algorithm 2 with more BS antennas performs requirements, a larger γ0 means a more stricter requirement on
better thanks to exploiting more array gains. Besides, the communication leading to more transmit power consumption.
performance gain gap between N = 80 and N = 96 is less It is evident that the F-MMSE algorithm presents the most
than that between N = 64 and N = 80, which implies the deficient performance compared to the three AO algorithms.
performance enhancement provided by adding the number of In addition, although the ZFBF algorithm performs better
BS antennas is not infinite. Notice that in practical systems, the than the MFBF algorithm in the low-value region of γ0 ,
BS equipped with more antennas could effectually decrease however, it performs worse in the high-value region of γ0 .
the total transmit power consumption, but it also leads to a More importantly, the proposed Algorithm 2 achieves the best
 10-3 25
7 80
Sensing error with N = 64
0
= 1 dB
Sensing error with N = 80
0
= 0.5 dB
6 Computation error with N = 64 70
Computation error with N = 80
20 = 0.1 dB
0

Total Transmit Power (W)

Communication rate with N = 80

Weighted Sum-rate (bit/s/Hz)

5 Communication rate with N = 64 60
Weighted Sum-MSE

15

4 50

10
3 40

2 30 5

1 20
0
0 2 4 6 8 10 12 14 16 18 20
0 10 Iteration Number
3 6 9 12 15 18 21
Number of Sensors, K
Fig. 7. Convergence behavior of the proposed Algorithm 2.

Fig. 5. The system performance gains versus the number of sensors with
different numbers of BS antennas for Algorithm 1.
4.5

0
= 0.001, 0 = 0.1 dB
4
= 0.01, 0 = 0.5 dB
10-2 0
1.4 80 = 0.01, 0 = 0.1 dB
0

Total Transmit Power (W)

3.5
Sensing error
1.2 Computation error
70
3
Communication rate
Weighted Sum-rate (bit/s/Hz)

60 2.5
1
Weighted Sum-MSE

50 2
0.8
1.5
40
0.6
1
30

0.4 0.5
20 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
The Tolerable Maximum Computation Error, 0
0.2 10
Fig. 8. Total transmit power versus tolerable maximum computation error
0 0 with different maximum tolerable sensing error and required minimum SINR
F-MMSE ZFBF MFBF Algorithm 1
of communication for Algorithm 2.

Fig. 6. The comparison of performance on MOOP for different algorithms.

paper, we only make a general integration of the three iso-
lated functions of sensing, computing and communication to
performance in the whole region of γ0 , especially for high explore their relationship. To support future emerging various
requirement of communication, which verifies the advantages applications with ultra-high performance requirements, it is
of Algorithm 2 for 6G wireless networks integrating sensing, necessary to deeply integrate wireless sensing functions such
computing, and communication. as positioning, detection, and imaging with wireless transmis-
sion functions, and at the same time exploit widely distributed
V. C ONCLUSION computing power for auxiliary processing to achieve cross-
integration of sensing, computing and communication.
This paper provided a comprehensive architecture for en-
abling 6G wireless networks to integrate sensing, computing,
and communication. Two typical MOOP-based algorithms A PPENDIX A
were designed from the perspectives of maximizing the P ROOF OF T HEOREM 1
weighted overall performance and minimizing the total trans- comm
For the received communication signal yk,j at the BS, the
mit power consumption. Extensive simulations confirmed the
MSE related to the j-th communication signal from the k-th
excellent performance of the proposed algorithms. Moreover,
sensor is given by
according to the preference of applications for 6G wireless n o
networks, it was feasible to acquire desired performance of MSEcomm = E y comm − scomm
y comm − scomm
H
k,j k,j k,j k,j k,j
sensing, computing, and communication by varying the system
K X J
priorities. Besides, it was found that the overall performance X
+ uH H H
k,j Hi ci,m ci,m Hi uk,j
can be improved by increasing the suitable numbers of sensors
i=1 m=1
and BS antennas. K X L
6G is still in the initial conceptual stage, and many chal- X
+ uH H H
k,j Hi bi,l bi,l Hi uk,j
lenges faced by ISCC has not been well addressed. In this i=1 l=1
 uH H H H H
k,j Ξ uk,j − uk,j ck,j Hk Hk ck,j uk,j
3.5
= H H
uk,j Ξ uk,j
3 N = 64 1
= . (42)
N = 80
1 + Γk,j
Total Transmit Power (W)

N = 96
2.5
It is seen that the MMSE of communication signal is the inverse of
2 one plus SINR and thus helps to transform the objective function of
S-3’. The proof is completed.
1.5

1
A PPENDIX B
0.5 P ROOF OF T HEOREM 2
0
1 2 3 4 5 6 7 To start with, let us present two lemmas as follows:
Number of Clutters, O
Lemma 1: If QP = 0, based on Sylvester’s rank inequality [45],
we have Rank(Q) + Rank(P) ≤ n, where Q ∈ Ct×n and P ∈
Fig. 9. Total transmit power versus the number of clutters with different
number of BS antennas for Algorithm 2.
Cn×s , ∀t, n, s.
Lemma 2: For the two matrices of the same size Q and P, it
always holds true that Rank(Q + P) ≤ Rank(Q) + Rank(P).
8 Proof:
     
Q+P Q+P Q
7 F-MMSE Rank (Q + P) = Rank ≤ Rank = Rank
ZFBF 0 P P
Total Transmit Power (W)

MFBF
6 Algorithm 2 ≤ Rank (Q) + Rank (P) . (43)

5
Then, we derive the Lagrangian function of M-2” associated with
4 Ck,j , which is given by
K I L J
!
3 X X X X
2 2
L (Ck,j ) = kak,i k + kbk,l k + tr (Ck,j )
2 k=1 i=1 l=1 j=1
I h i
sens
X
1
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 + λi MSEi − δi
The Required Minimium SINR on Communication, 0 (dB) i=1
L h i XK X
J
X comp
Fig. 10. Performance comparison of different algorithm for total transmit + βl MSEl − χl + µk,j Tk,j
power minimization. l=1 k=1 j=1
K X
X J
− Θk,j Ck,j , (44)
K X
X I k=1 j=1
+ Rn2 uH H H
k,j Gi,n ai,n ai,n Gi,n uk,j K J
tr uH H
P P

i=1 n=1 where Tk,j = k,j Hi Ci,n Hi uk,j + Xk,j +
XK X O I
X i=1,i6=k n=1,n6=j
σn2 kuk,j k2 − γk,j
1
tr uH H


+ Ro2 uH H H
k,j Fi,o ai,m ai,m Fi,o uk,j k,j Hk Ck,j Hk uk,j . λi , βl , µk,j , and Θk,j

i=1 o=1 m=1 are Lagrange multipliers of constraint C5, C6, C7 and C8, re-
spectively. To explore the optimal solution C∗k,j under the Slater’s
+ σn2 kuk,j k − uH H H
k,j Hk ck,j − ck,j Hk uk,j + 1. (40) condition, we make use of the KKT conditions as follows:
Extracting the common factor from the above equation, we K
X J
X  
∗
J
K P L
K P
tr uH H
k,j Hi Ci,n Hi uk,j + Xk,j +
Hi ci,m cH H
Hi bi,l bH H
P P
define Ξ = i,m Hi + i,l Hi +
i=1 m=1 i=1 l=1 i=1,i6=k n=1,n6=j
K P
I K P
O P
I
1  
Rn2 Gi,n ai,n aH H
Ro2 Fi,o ai,m aH H
σn2 kuk,j k2 − tr uH ∗ H
P P
i,n Gi,n + i,m Fi,o + k,j Hk Ck,j Hk uk,j = 0, (45a)
i=1 n=1 i=1 o=1 m=1 γk,j
σn2 I, then (40) can be rewritten as
Θ∗k,j C∗k,j = 0, (45b)
MSEcomm
k,j = uH H H H
k,j Ξuk,j − uk,j Hk ck,j − ck,j Hk uk,j + 1
   H I L
H −1 H −1
= uH H
k,j − ck,j Hk Ξ Ξ uH H
k,j − ck,j Hk Ξ ∇C∗k,j L = IM +
X
λ∗i HH H
X
βl∗ HH H
k vi vi Hk + k zl zl Hk
H −H
− cH
k,j Hk Ξ Hk ck,j + 1. (41) i=1 l=1
µ∗k,j H ∗
It can be concluded that when uk,j = Ξ Hk ck,j , the −1
MSEcomm − Hk uk,j uH
k,j Hk − Θk,j = 0, (45c)
k,j γk,j
would be minimized. Consequently, the MMSE associated with the
communication signal scomm
k,j is as below λ∗i ≥ 0, βl∗ ≥ 0, µ∗k,j ≥ 0, Θ∗k,j  0. (45d)
H −H 2
ecomm
k,j = 1 − cH
k,j Hk Ξ Hk ck,j Since Xk,j + σn2 uH
k,j > 0 in (45a), we have C∗k,j 6= 0, which
Ξ H
− cH H means
k,j Hk Hk ck,j
= Rank(C∗k,j ) ≥ 1. (46)
ΞH
By exploiting the relationship between Θ∗k,j and C∗k,j in (45b), based [15] X. Ge, Y. Sun, H. Gharavi, and J. Thompson, “Joint optimization of
on Lemma 1, we obtain that computation and communication power in multi-user massive MIMO
systems,” IEEE Trans. Wireless Commun., vol. 17, no. 6, pp. 4051-4063,
∗ ∗
Rank(Θk,j ) + Rank(Ck,j ) ≤ M. (47) Jun. 2018.
[16] Q. Qi, X. Chen, C. Zhong, and Z. Zhang, “Integrated sensing, computa-
Next, substituting (46) into (47) yields tion and communication in B5G cellular Internet of Things,” IEEE Trans.
∗ Wireless Commun., vol. 20, no. 1, pp. 332-344, Jan. 2021.
Rank(Θk,j ) ≤ M − 1. (48) [17] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of radio
and computational resources for multicell mobile-edge computing,” IEEE
Then, based on Lemma 2, it is observed from (45c) that Trans. Signal Inf. Process. Netw., vol. 1, no. 2, pp. 89-103, Jun. 2015.
Rank(Υk,j ) + Rank(Θ∗k,j ) ≥ Rank(IM ), (49) [18] S. Deng, H. Zhao, W. Fang, J. Yin, S. Dustdar, and A. Y. Zomaya,
“Edge intelligence: The confluence of edge computing and artificial
 ∗
µ I L
 intelligence,” IEEE Internet of Things J., vol. 7, no. 8, pp. 7457-7469,
where Υk,j = HH k,j
u uH − λ∗i vi viH − βl∗ zl zH
P P
k γk,j k,j k,j l Hk . Aug. 2020.
i=1 l=1 [19] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge computing: Vision
Since Υk,j 6= 0 and Rank(IM ) = M , we obtain and challenges,” IEEE Internet Things J., vol. 3, no. 5, pp. 637-646, Oct.
∗ 2016.
Rank(Θk,j ) ≥ M − 1. (50) [20] T. X. Tran and D. Pompili, “Joint task offloading and resource allocation
for multi-server mobile-edge computing networks,” IEEE Trans. Veh.
Based on the relationship between (48) and (50), we can get
Technol., vol. 68, no, 1, pp. 856-868, Jan. 2019.
Rank(Θ∗k,j ) = M − 1. Substituting it into (47) and combining (46), [21] L. Chen, N. Zhao, Y. Chen, F. R. Yu, and G. Wei, “Over-the-air com-
it is concluded that putation for IoT networks: Computing multiple functions with antenna
Rank(C∗k,j ) = 1, (51) arrays,” IEEE Internet of Things J., vol. 5, no. 6, pp. 5296-5306, Dec.
2018.
which indicates the adopted SDR is tight. The proof is completed. [22] X. Chen and Q. Qi, “Convergence of energy, communication and
computation in B5G cellular Internet of Things”, Germany: Springer
Press, 2020.
[23] T. Sery, N. Shlezinger, K. Cohen, and Y. C. Eldar, “Over-
R EFERENCES the-air federated learning from heterogeneous data,” [Online]:
https://arxiv.org/pdf/2009.12787.pdf.
[1] Z. Zhang, Y. Xiao, Z. Ma, M. Xiao, Z. Ding, X. Lei, G. K. Karagiannidis, [24] G. Zhu, Y. Wang, and K. Huang, “Broadband analog aggregation for
and P. Fan, “6G wireless networks: Vision, requirements, architecture, and low-latency federated edge learning,” IEEE Trans. Wireless Commun.,
key technologies,” IEEE Veh. Technol. Mag., vol. 14, no. 3, pp. 28-41, vol. 19, no. 1, pp. 491-506, Jan. 2020.
Sep. 2019. [25] Y. Gao, L. Liu, B. Hu, T. Lei, and H. Ma, “Federated region-learning
[2] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K. for environment sensing in edge computing system,” IEEE Trans. Netw.
Soong, and J. C. Zhang, “What will 5G be?” IEEE J. Sel. Areas Commun., Sci. Eng., vol. 7, no. 4, pp. 2192-2204, Dec. 2020.
vol. 32, no. 6, pp. 1065-1082, Jun. 2014. [26] G. Zhu and K. Huang, “MIMO over-the-air computation for high-
[3] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A survey on mobility multimodal sensing,” IEEE Internet of Things J., vol. 6, no.
mobile edge computing: The communication perspective,” IEEE Commun 4, pp. 6089-6103, Aug. 2019.
Surv. Tuts., vol. 19, no. 4, pp. 2322-2358, Dec. 2017. [27] M. M. U. Rathore, A. Paul, A. Ahmad, B. -W. Chen, B. Huang, and
[4] B. Lin, S. Li, A. Nallanathan, and C. Zhao, “Deep sensing for future W. Ji, “Real-time big data analytical architecture for remote sensing
spectrum and location awareness 5G communications,” IEEE J. Sel. Areas application,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol.
Commun., vol. 33, no. 7, pp. 1331-1344, Jul. 2015. 8, no. 10, pp. 4610-4621, Oct. 2015.
[5] “Integration of sensing, communication and computing toward 6G,” [28] C. Luo, B. Deng , H. Wang, and Y. Qin, “High-resolution terahertz
World 5G Convention White Paper, Nov. 2020. coded-aperture imaging for near-field three-dimensional target,” Appl
[6] D. K. P. Tan, J. He, Y. Li, A. Bayesteh, Y. Chen, P. Zhu, and W. Tong, Opt., vol. 58, no. 12, pp. 3293-3300, Apr. 2019.
“Integrated sensing and communication in 6G: Motivation, use cases, [29] J. R. Guerci, J. S. Bergin, R. J. Guerci, M. Khanin, and M. Rangaswamy,
requirements, challenges and future directions,” in Proc. IEEE Inter. “A new MIMO clutter model for cognitive radar,” in Proc. IEEE Radar
Symp. Joint Commun. Sensing (JC&S), Feb. 2021, pp. 1-6. Conf., 2016, pp. 1-6
[30] Z. Huang, K. Wang, A. Liu, Y. Cai, R. Du, and T. X. Han, “Joint pilot op-
[7] H. Wymeersch, D. Shrestha, C. M. de Lima, V. Yajnanarayana,
timization, target detection and channel estimation for integrated sensing
B. Richerzhagen, M. F. Keskin, K. Schindhelm, A. Ramirez, and
and communication systems,” [Online]:https://arxiv.org/abs/2202.02688.
A. Wolfgang, “Integration of communication and sensing in 6G:
[31] T. Wild, V. Braun, and H. Viswanathan, “Joint design of communication
A joint industrial and academic perspective,” Jun. 2021, [Online]:
and sensing for beyond 5G and 6G systems,” IEEE Access, vol. 9, no. 1,
https://arxiv.org/pdf/2106.13023.
pp. 30845-30857, Feb. 2021.
[8] Q. Qi, X. Chen, C. Zhong, and Z. Zhang, “Integration of energy, [32] G. Alirezaei, O. Taghizadeh, and R. Mathar, “Optimum power allocation
computation and communication in 6G cellular Internet of Things,” IEEE in sensor networks for active radar Applications,” IEEE Trans. Wireless
Commun. Lett., vol. 24, no. 6, pp. 1333-1337, Jan. 2020. Commun., vol. 14, no. 5, pp. 2854-2867, May 2015.
[9] J. A. Mahal, A. Khawar, A. Abdelhadi, and T. C. Clancy, “Spectral [33] Y. Tian, Z. Zhang, C. Huang, X. Chen, and C. Zhong, “Intelligent
coexistence of MIMO radar and MIMO cellular system,” IEEE Trans. reflecting surface aided computational imaging exploiting reed-muller
Aerosp. Electron. Syst., vol. 53, no. 2, pp. 655-668, Jan. 2017. sequences,” in Proc. IEEE 94th VTC2021-Fall, 2021, pp. 1-5.
[10] A. Hassanien, M. G. Amin, Y. D. Zhang, and F. Ahmad, “Phase- [34] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. Y. Arcas,
modulation based dual-function radar-communications,” IET Radar, “Communication-efficient learning of deep networks from decentralized
Sonar & Navigation, vol. 10, no. 8, pp. 1411-1421, Oct. 2016. data,” in Proc. Int. Conf. Artif. Intell. Stat. (AISTATS), 2017, pp. 1273-
[11] W. Xu, F. Gao, S. Jin, and A. Alkhateeb, “3D scene-based beam selection 1282.
for mmWave communications,” IEEE Wireless Commun. Lett., vol. 9, no. [35] Z.-Q. Luo and S. Zhang, “Dynamic spectrum management: Complexity
11, pp. 1850-1854, Nov. 2020. and duality,” IEEE J. Sel. Topics Signal Process., vol. 2, no. 1, pp. 57-73,
[12] A. Liu, Z. Huang, M. Li, Y. Wan, W. Li, T. X. Han, C. Liu, R. Du, Feb. 2008.
D. T. K. Pin, J. Lu, Y. Shen, F. Colone, and K. Chetty, “A survey on [36] Q. Shi, M. Razaviyayn, Z-Q. Luo, and C. He, “An iteratively weighted
fundamental limits of integrated sensing and communication,” Apr. 2021, MMSE approach to distributed sum-utility maximization for a MIMO
[Online]: https://arxiv.org/pdf/2104.09954. interfering broadcast channel,” IEEE Trans. Signal Process., vol. 59, no.
[13] H. Hua, J. Xu, and T. X. Han, “Optimal transmit beamform- 9, pp. 4331-4340, Sep. 2011.
ing for integrated sensing and communication,” Apr. 2021, [Online]: [37] R. Marler and J. Arora, “Survey of multi-objective optimization methods
https://arxiv.org/pdf/2104.11871. for engineering”. Struct Multidisc Optim, vol. 26, 369-395, Mar. 2004.
[14] T. Wild, V. Braun, and H. Viswanathan, “Joint design of communication [38] A. Khalili and D. W. K. Ng, “Energy and spectral efficiency tradeoff in
and sensing for beyond 5G and 6G systems,” IEEE Access, vol. 9, pp. OFDMA networks via antenna selection strategy,” in Proc. IEEE WCNC,
30845-30857, 2021. 2020, pp. 1-6.
[39] J. C. Bezdek and R. J. Hathaway, “Convergence of alternating opti-
mization,” Neural, Parallel & Scientific Computations, vol. 11, no. 4, pp.
351-368, Dec. 2003.
[40] J. Yeh, Real analysis: Theory of measure and integration. Singapore:
World Scientific, 2006.
[41] M. Grant and S. Boyd, CVX: Matlab Software for Disciplined Convex
Programming, [Online]: http://cvxr.com/cvx, Sep. 2013.
[42] S. Boyd and L. Vandenberghe, “Convex Optimization”. Cambridge,
U.K.: Cambridge Univ. Press, 2004.
[43] A. Ben-Tal and A. Nemirovski, “Lectures on modern convex op-
timization: Analysis, algorithms, and engineering applications,” MPS-
SIAMSeries on Optimization. Philadelphia, PA, USA: SIAM, 2001.
[44] Coordinated Multi-Point Operation for LTE Physical Layer Aspects (Rel.
11), document TR 36.819, 3GPP, Feb. 2011.
[45] G. Matsaglia and G. P. H. Styan, “Equalities and inequalities for ranks
of matrices,” Linear and Multilinear Algebra, vol. 2, pp. 269-292, 1974.