This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
1

Joint Transceiver Beamforming Design for

Hybrid Full-Duplex and Half-Duplex Ad-Hoc Networks
Jiancao Hou, Member, IEEE, Zhaohui Yang, Member, IEEE,
and Mohammad Shikh-Bahaei, Senior Member, IEEE

Abstract—In this paper, we propose a joint transceiver beamforming design method for hybrid full-duplex (FD) and half-duplex (HD)
ad-hoc networks to cancel co-channel interference, thereby to improve system spectral efficiency. To characterize network
performances, we derive a general expression of transmission capacity upper bound (TC-UB) plus its two compact versions by using a
stochastic geometry model. Due to the proposed beamforming design and hybrid-duplex consideration, the exact TC and conventional
methods to obtain TC-UBs are not applicable. This motivates us to exploit the UB of the largest eigenvalue of desired signals, Alzer’s
inequality for incomplete gamma function, and dominating interference region to formulate one general TC-UB and two of its compact
versions. The numerical results show that the proposed beamforming method outperforms the existing beamforming strategies in
terms of exact TC, especially when the number of transmit antennas is larger than the number of receiver antennas per node pair. In
addition, the derived general TC-UB can provide relatively close TC performance as the exact ones, and its two compact versions can
at least give order-wise TC performance. Moreover, we find the break-even points, where FD outperforms HD with different system
configurations.

Index Terms—In-band full-duplex, MIMO beamforming, transmission capacity, ad-hoc network, outage probability.

1 I NTRODUCTION
Due to rapid expansion of applications and services, next tial domain SI cancellation, which is based on beamforming
generations of wireless communications, e.g., 5G and be- techniques, can nullify the SI at the transceiver front end
yond, are expected to rely on even higher spectral efficiency [12], [13], [14]. In [15], the merit of spatial domain SI can-
and/or lower transmission latency [2], [3]. Among various cellation has been highlighted. Apart from the SI effects,
methods to improve system spectral efficiency, in-band full- some applications like FD enabled BS, which allow si-
duplex (IBFD) as one of promising techniques has recently multaneous uplink (UL) and downlink (DL) transmissions,
re-emerged and gained more attentions [4], [5], [6]. Con- also introduce inter-user interference, as DL reception is
ventionally, most common communication terminals, e.g., affected by UL transmission. In [16], with the presented
base stations (BSs), relaying nodes, and mobile users, which SI and inter-user interference, the authors jointly designed
function as both transmitter and receiver, operate in half- the transceiver beamforming vectors for BS to maximize
duplex (HD) mode or out-of-band full-duplex (FD) mode, the system achievable sum-rate. Furthermore, the authors
where their transmission and reception take place either in [17] combined FD with massive multiple-input multiple-
at different time slots, or over different frequency bands output (MIMO) for simultaneous UL and DL cellular com-
aiming to avoid co-channel interference. IBFD, on the other munications and optimized the transceiver antennas ratio at
hand, enables the communication terminals to transmit and BS to maximize the achievable capacities.
receive simultaneously over the same frequency band. Po- With large-scale networks consideration, one key ques-
tentially, it can double the spectral efficiency. Beyond that, tion is how system with aggregate interferences perform? In
IBFD can also enable the terminals to detect collisions or the literature, many research works have been conducted
receive feedback whilst transmitting their own data, which to quantify performances of large-scale networks using
can largely reduce transmission latency. stochastic geometry model, which is an efficient tool for
One of the key challenges of implementing IBFD is analysing ad-hoc and heterogeneous cellular networks. In
the presence of self-interference (SI) at terminals, where SI this case, transmission capacity (TC), which characterizes
denotes the interference that is caused by terminals’ self- maximum nodes’ density with successful transmission, has
transmitted signals. To handle this SI effect, researchers have been investigated intensively [18], [19], [20], [21], [22]. In
contributed to develop many advanced techniques, e.g., comparison to transport capacity or system throughput, the
propagation-domain antenna separation [7], [8], analog- advantage of TC lies largely in its amenability to precise
domain [9], [10], and/or digital-domain SI cancellation [9], analysis, which allows the impact of physical layer fading
[10], [11]. In general, the most obvious approach is to assess effects on link layer scheduling policies to be more pre-
and subtract SI from the received signal. Alternatively, spa- cisely characterized [18]. Considering MIMO HD ad-hoc
networks, the TC performances have been characterized
efficiently by exploiting linear beamforming designs (e.g.
• J. Hou, Z. Yang and M. Shikh-Bahaei are with the Department of Engi-
neering, King’s College London, London, United Kingdom, WC2R 2LS. maximum ratio combine (MRC), eigen-beamforming (EB),
E-mail: {jiancao.hou, yang.zhaohui, m.sbahaei}@kcl.ac.uk. The material and partial zero-forcing (P-ZF)) plus the statistical property
in this paper has been in part presented at IEEE Global Commun. Conf. of the accumulated inter-node interference via 1-D Poisson
2018 [1].
point process (PPP) modelling [20], [21], [22].

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
2

By extending network analysis to FD scenarios, the antennas. By exploiting spatial degrees of freedom (DoF),
authors in [23] and [24] first studied the performances of we first propose a joint transceiver beamforming design
ad-hoc and multi-tier heterogeneous networks, respectively, method per node pair to mitigate SI, either at transmitter
considering a mixture of the nodes operating in either FD side or receiver side, and the nearest inter-node interfer-
or HD mode, and with imperfect SI cancellation. Inter- ers/interferer pairs effects, thus the network TC perfor-
estingly, both papers come with the same conclusion that mance can be improved in comparison to existing beam-
operating all nodes in pure FD or pure HD maximizes forming methods.
the area spectral efficiency compared with a mixture of • To shed insight on network performances, we derive
the two modes. Moreover, the authors in [25] analysed the a general TC-UB plus its two compact versions by using
transport capacity upper and lower bounds for an IBFD ad- a stochastic geometry model. Due to the unknown distri-
hoc network, and concluded that nearly doubled transport bution of the largest eigenvalue of desired signal power
capacity in comparison to HD mode can be observed only per link and hybrid-duplex communication modelling, the
with relatively small paired link distance. Furthermore, conventional way to derive the TC-UB is not applicable.
the authors in [26] investigated a hybrid FD/HD cellular This motivates us to exploit the upper bound of the largest
network and showed that the average spectral efficiency of eigenvalue of desired signals, Alzer’s inequality of incom-
FD mode could nearly double that of HD mode at the cost plete gamma function, and dominating interference region
of reduced network coverage. All above mentioned works to obtain the TC-UBs.
assumed single-antenna per node, whereas multi-antenna • Simulation results show the benefit of the proposed
nodes based FD large-scale networks have been studied in beamforming design in comparison to existing beamform-
[27], [28], [29]. In [27], the authors derived the achievable ing methods. In addition, we examine the tightness of the
sum-rate upper and lower bounds of ad-hoc network by derived TC-UBs, the SI channel estimation error impact,
assuming the SI has been perfectly eliminated. In addition, and the non-linear hardware distortion effect on TC perfor-
the authors in [28] analysed the spectral efficiency of MIMO mance. Moreover, we also compare FD and HD operations
FD cellular networks according to the ZF strategy to pre- to elucidate transceiver antenna configurations and transmit
vent intra-cell interference. Considering both SI and inter- power regions that FD can outperform HD.
node interference cancellation at receiver side, the authors Note that, due to unknown expression for the maximum
in [29] investigated the throughput of FD MIMO small- singular value of desired signal and the mixed HD and FD
cell networks and drew the condition when cancelling SI inter-node interference cancellation, the performance gap
was more beneficial than cancelling the nearest inter-node between the derived TC-UBs and the exactly simulated TC
interference. can hardly be avoid. Such gap is relatively small when
To the best of our knowledge, the joint transceiver the probability of FD node pairs is small, and the number
beamforming design in generalized hybrid FD and HD ad- of receive antenna is larger than the number of transmit
hoc networks has not been well studied, especially when antenna; Or when the probability of FD node pairs is large,
TC is considered as the key performance indicator. In ad- and the number of receive antenna is smaller than the
dition, the above reviewed FD, HD and hybrid FD/HD number of transmit antenna. In other cases, the derived TC-
schemes did not take the antenna conserved scenario into UBs can at least be used to identify order-wise performance.
account. These motivate us to consider the following inter- The rest of the paper is organized as follows. Section
esting questions: Apart from the existing beamforming ap- II presents the system model of hybrid FD/HD ad-hoc
proaches, is there any more efficient beamforming method network with prepossessing/post-processing beamforming
to improve network performances? Subsequently, how can structure. Section III provides the proposed joint transceiver
we derive the mathematical formula to facilitate the analysis beamforming design to mitigate SI and the nearest inter-
of system performances? In this paper, we will answer these node interferers/interferer pairs. The general TC-UB expres-
questions by proposing a joint transceiver beamforming sion plus its two compact versions are derived in Section
design, deriving its corresponding TC-UBs, and verifying IV. Numerical results are provided in Section V. Section
the results via simulations. VI concludes the paper. Throughout this paper, the basic
This work is an extension to the ones in [1], whereas, notations have been summarized in Tab. I.
in [1], only pure FD nodes existed in the considered net-
work, and the derivation of TC-UB was only focusing on
exploiting dominating interference region. By contrast, in 2 S YSTEM M ODEL
this paper, we generalize the system model by involving hy- Consider a FD enabled ad-hoc network, where the deployed
brid FD and HD nodes in the network. Such consideration transceiver pairs follow an independently marked PPP, e.g.,
complicates the derivation of TC-UBs, because the average Φ = {(ak , bk )}, on a 2-D plane R2 × R2 , and the ground
number of cancelled inter-node interference needs to take process ΦA = {ak } is a homogeneous PPP with density
a mix of FD and HD nodes into account. Apart from the λ. The location of each marked (or paired) node can be
derived compact TC-UBs similar as in [1], we also derive written as bk = ak + L(cos ϕk , sin ϕk ), where L is a constant
a tighter TC-UB based on exploiting Alzer’s inequality for distance between ak and bk , and ϕk is the angle following
incomplete gamma function. Moreover, we examine the independent and uniform distribution over [0, 2π]. Here, we
non-linear hardware distortion and channel estimation error define a set ΦB consisting all marked nodes {bk }. ΦB can
effects on our considered TC performance. Specifically, our be treated as a displaced version of ΦA , which is also a
contributions in this paper are summarized as below. PPP with density λ. We assume each node pair can work in
• Consider each transceiver pair equipped with multiple either FD or HD mode with probability p1 or p2 (= 1 − p1 ),

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
3
TABLE 1: Summarizes the basic notations in the paper
where sbk and sak represent data symbols at node bk and
Symbol Usage node ak , respectively; wbk ∈ C Nt ×1 and wak ∈ C Nt ×1 with
RN , C N The set of real and complex N -tuples, respectively. unit norm denote the pre-processing vectors for node bk
dxe, bxc The smallest integer greater than or equal to x and and node ak , respectively; et,bk ∈ C Nt ×1 and et,ak ∈ C Nt ×1
the largest integer less than or equal to x, respec-
tively.
are transmit-side non-linear hardware distortion vectors for
bxe The nearest integer of x. node bk and node ak , respectively; E{|sbk |2 } = E{|sak |2 } =
rank(·) The rank of a matrix. P as the transmit power.
|·| The absolute value of a scalar.
k·k Euclidean norm of a vector or Frobenius norm of a Accordingly, the received signal at the typical node a0
matrix. with size of Nr × 1 is given by
E{·} The expectation of random variable(s).
−α/2
X
(·)−1 Inverse of a matrix. ya0 = L−α/2 Ha0 ,b0 xb0 + rbk Ha0 ,bk xbk
diag(a) A diagonal matrix whose diagonal is vector a. bk ∈ΦHD /b0
{·} A set. X 
−α/2

(ak , bk ) The kth number of transceiver node pair. + rbk Ha0 ,bk xbk + ra−α/2
k
H a ,a
0 k
x ak
(·)H The vector (or matrix) conjugate transpose. (ak ,bk )
∈ΦFD /(a0 ,b0 )

+ δHa0 xa0 + va0 + er,a0 , (4)

respectively. In FD mode, ak and bk are transmitting to each
where xbk ∈ C Nt ×1 and xak ∈ C Nt ×1 including xb0 and
other concurrently, while in HD mode, the transmission is
xa0 are the transmitted signals at node bk and node ak ,
only from one node to the other at a time, e.g., from bk to
respectively; Ha0 ,bk ∈ C Nr ×Nt and Ha0 ,ak ∈ C Nr ×Nt de-
ak . As a result, the set of FD mode pairs, e.g., ΦFD , can be note Rayleigh fading channels from node bk to node a0
treated as a thinning marked PPP with density p1 λ, and the
and from node ak to node a0 , respectively, in addition,
set of HD mode pairs, e.g., ΦHD , can be treated as a thinning
each entry follows independent identical complex Gaussian
marked PPP with density p2 λ. From the marking theorem
distribution with zero mean and unit variance to model
[31], these two node pair sets are independent. We select a
a richly scattered fading channel with independent fading
node pair, e.g., {(a0 , b0 )}, as the typical node pair, which
coefficients between different antennas similar to [20], [21];
will be used to analyse network performances.1 −α/2 −α/2
L−α/2 , rbk and rak represent path-loss attenuation
Each transceiver node is equipped with Nt transmit
from node b0 to node a0 , from node bk to node a0 , and from
antennas and Nr receive antennas. We assume all nodes
node ak to node a0 , respectively, where α > 2 is the path-
share the same time-frequency resource, and each frame
loss exponent; δ ∈ {0, 1} is the indicator function that ‘1’
transmitted per node is smaller or equal to the coherence
refers the typical link in FD mode and ‘0’ refers the typical
time of all nodes. Here, we divide the transmission frame
link in HD mode; va0 ∈ C Nr ×1 denotes the additive white
into two parts, whereof the first part is dedicated to pilot
Gaussian noise (AWGN) at node a0 with zero mean and
signals for channel estimation, and the second part is used
variance σ02 ; er,a0 ∈ C Nr ×1 denotes receive-side non-linear
for payload data. In this case, each node firstly broadcasts
hardware distortion vector for typical node a0 .
pilot signals, and at the same time, receive pilot signals
for channel estimation. The accurate channel acquisition By applying the post-processing vector za0 ∈ C Nr ×1 , the
depends on the number of orthogonalized pilot signals received signal at the typical node a0 can be expressed as
within the frame per node. If there are no enough such −α/2
X −α/2
pilot signals, the estimated channel will be contaminated zH
a0 ya0 = L ha0 ,b0 sb0 + rbk ha0 ,bk sbk
bk ∈ΦHD /b0
and inaccurate. In this paper, we assume each node can X 
−α/2

accurately estimate the fading channels of its intended pro- + rbk ha0 ,bk sbk + ra−α/2
k
ha0 ,ak sak
cessed links, but not SI channels due to the severe non-linear (ak ,bk )
∈ΦFD /(a0 ,b0 )
hardware distortion [32]. To satisfy the former accurate
+ δha0 sa0 + va0 + er,a0 + ẽt,a0 , (5)
channel estimation, the network has to sacrifice spectrum
efficiency by adding enough number of pilot signals. For where ha0 ,b0 , zH H
a0 Ha0 ,b0 wb0 , ha0 ,bk , za0 Ha0 ,bk wbk ,
imperfect SI, if a node ak works in FD mode, the actual SI ha0 ,ak , za0 Ha0 ,ak wak , ha0 , za0 Ha0 wa0 , va0 , zH
H H
a0 va0 ,
channel (i.e. Hak ∈ C Nr ×Nt ) can be modelled as [12] er,a0 , zH a0 er,a0 , and
Hak = Ĥak + ∆ak , ∀k, (1)  X −α/2
ẽt,a0 , za0 L−α/2 Ha0 ,b0 et,b0 +
H
rbk Ha0 ,bk et,bk
where Ĥak denotes the estimated version of Hak , and
bk ∈ΦHD /b0
∆ak is the estimation error following complex Gaussian X 
−α/2

2
distribution with zero mean and variance σSI . Furthermore, + rbk Ha0 ,bk et,bk + ra−α/2
k
Ha0 ,ak et,ak
(ak ,bk )
due to the non-linear hardware distortion, the transmitted ∈ΦFD /(a0 ,b0 )
signal can be formulated as 
+ δHa0 et,a0 .
xbk = wbk sbk + et,bk , bk ∈ (ΦFD ∪ ΦHD ), (2)
xak = wak sak + et,ak , ak ∈ ΦFD , (3) Then, the performance metric, signal-to-interference plus
1. From Slivnyak’s theorem [31] and the stationarity of Φ, the typical
noise ratio (SINR) at node a0 , can be expressed as
node pair can be selected randomly, e.g., (a0 , b0 ), and the properties of L−α |ha0 ,b0 |2
node a0 ∈ ΦA (or b0 ∈ ΦB ) represent the properties of other nodes in SINRa0 = ,
|va0 |2 +|er,a0 |2 +|ẽt,a0 |2
the same set. IkHD + IkFD + δ|ha0 |2 + P

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
4

(6) i.e., |zH 2

a0 Ĥa0 wa0 | = 0. To obtain the optimal beamforming
−α solution of (9), one heuristic thinking is to allow all nodes
where IkHD , 2
P
bk ∈ΦHD /b0 rbk |ha0 ,bk | is the aggregate chase each others’ transceiver beams and jointly make the
interference received from P transmitter nodes worked in
−α final decision, where the global channel state information
HD mode; IkFD = 2
(ak ,bk )∈ΦFD /(a0 ,b0 ) (rbk |ha0 ,bk | +
−α 2
and/or the formulated beamforming vectors need to be
rak |ha0 ,ak | ) is the aggregate interference received from shared at the cost of large signalling overhead. In addition,
transmitter and receiver nodes worked in FD mode. the proof of algorithm convergence could be a challenge
Furthermore, according to [32], [33], the transceiver non- especially with finite number of DoFs. In contrast, assume
linear hardware distortion vectors at node ak , e.g., et,ak and that each node pair can only access its related links’ channel
er,ak , follows complex Gaussian distribution, such as knowledge. In this case, we propose a distributed beam-
forming design, where each transmit node first formulates
  n o
et,ak ∼ CN 0, κt P diag E wak waHk , (7)
its pre-processing vector individually, and then broadcasts
it for post-processing vector’s formulation at each receive
  n o
er,ak ∼ CN 0, κr diag E ỹak ỹaHk , (8)
node.
where ỹak denotes undistorted receive signal at the node The design criteria for the typical node pair in FD mode
ak , e.g., ỹa0 = ya0 − er,a0 ; κt ∈ R+ and κr ∈ R+ represent follows two cases: Nt > Nr and Nt ≤ Nr . In the case
the distortion coefficient for the transmit chain and receive where Nt > Nr , the SI effect, i.e., zH a0 Ĥa0 wa0 , can be
chain, respectively. nullified by wa0 itself irrespective of za0 . Then, the node
Remark 1: Theoretical derivation of network performance a0 can spend all its Nr − 1 receive DoFs to cancel maximum
of the proposed model is quite difficult if the transceiver Nr − 1 nearest inter-node interferers and/or interferer pairs.
non-linear hardware distortions are taken into account. This Let’s first decompose Ĥa0 into spatial modes via singular
is because the variables coupling among node distance and value decomposition (SVD) as Ĥa0 = Ua0 Σa0 VaH0 , where
hardware non-linear distortion, e.g., see (5), will prevent Ua0 ∈ C Nr ×Nr and Va0 ∈ C Nt ×Nt are unitary matrices, and
us from identifying the distributions of different variables. Σa0 ∈ RNr ×Nt is a rectangular diagonal matrix. Then, the
Thus, for the rest of paper, we derive our theoretical results (rank(Ĥa0 ) + 1) to Nt column(s) of Va0 construct a matrix
without considering the non-linear hardware distortion ef- Ṽa0 (or a vector ṽa0 ) with size of Nt × (rank(Ĥa0 ) + 1 : Nt )
fect, but their impacts will be examined via numerical spanning null space of Ĥa0 . Then, to improve transmission
simulations in Sec. V. quality of the desired channel link from a0 to b0 , i.e.,
Hb0 ,a0 , we define H̃b0 ,a0 , Hb0 ,a0 Ṽa0 and perform the SVD
3 T HE P ROPOSED T RANSCEIVER B EAMFORMING of H̃b0 ,a0 as Ũb0 ,a0 Σ̃b0 ,a0 ṼbH0 ,a0 . Then, the pre-processing
D ESIGN vector at typical node a0 can be formulated as
(1)
In this section, we design the pre-processing and post- wa0 = Ṽa0 ṽb0 ,a0 , (10)
processing beamforming vectors to boost the desired signal (1)
power, cancel SI and also l (≤ Nr − 1) nearest inter- where ṽb0 ,a0
is the first column of Ṽb0 ,a0 . Similar way to
node interferers/interferer pairs. Here, we denote lHD as formulate wb0 and pre-processing vectors for other node
the number of interferers in HD mode to be cancelled and pairs in FD mode. Correspondingly, the post-processing
lFD (= l − lHD ) as the number of inter-node interferer pairs vector zb0 at typical node b0 can be formulated as
in FD mode to be cancelled. It is worth noting that, to (1)
−1 H
zb0 = (sH
b0 ũb0 ,a0 ) sb0 , (11)
reduce signalling burdens of higher layer protocol design
(1)
and simplify the network analysis similar to [29], if one where is the first column of Ũb0 ,a0 , and sb0 ∈
ũb0 ,a0
of a FD inter-node interferer pair nodes is counted as one C Nr ×1 is used to span null space of Hb0 ,bk wbk ,
of l nearest inter-node interferers, we will cancel the FD bk,k=1,...,lHD ∈ΦHD /b0 , Hb0 ,bk wbk , bk,k=1,...,lFD ∈ΦFD /b0 , and
inter-node interferer pair (i.e. require two DoFs) before Hb0 ,ak wak , ak,k=1,...,lFD ∈ΦFD /a0 . Here, we must guarantee
moving to the next nearest inter-node interferer/interferer lHD + 2lFD ≤ Nr − 1. Similar way to formulate za0 and
pair cancellation. Thus, by fixing the number of transceiver post-processing vectors for other node pairs in FD mode.
antennas per node, our objective is to In the case where Nt ≤ Nr , Ĥa0 has full column rank.
|zH 2 The SI plus Nr − 2 nearest inter-node interferers should be
max a0 Ha0 ,b0 wb0 | (9)
za0 ,wbk ,wak cancelled with Nr − 1 receiver DoFs.2 Thus, to improve
0
s.t. |zH 2
a0 Ha0 ,bk wbk | = 0, ∀bk ∈ ΦHD , transmission quality of the desired link, we formulate the
0 pre-processing vector wa0 at typical node a0 as
|zH 2
a0 Ha0 ,bk wbk | = 0, ∀bk ∈ ΦFD ,
(1)
|zH 2
a0 Ha0 ,ak wak | = 0, ∀ak ∈ ΦFD /a0 , wa0 = vb0 ,a0 , (12)
|zH 2
a0 Ĥa0 wa0 | = 0, if δ = 1, where
(1)
vb0 ,a0
denotes the first right singular vectors of
0 channel matrix Hb0 ,a0 . Similar way to formulate wb0 and
where ΦHD ⊂ ΦHD denotes the node set including all lHD
0 pre-processing vectors for other nodes in FD mode. Corre-
nearest inter-node interferers in HD mode; ΦFD ⊂ ΦFD
spondingly, the post-processing vector zb0 at typical node b0
denotes the node set including all lFD nearest inter-node in-
terferer pairs in FD mode. In addition, if the typical receiver
2. In this paper, we assume SI always has priority to be cancelled
a0 works in FD mode (i.e. δ = 1), the SI interference effect first. More detailed analysis regarding to when cancelling the SI first is
from its estimated version needs to be cancelled as well, more beneficial can be found in [29].

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
5

can be formulated as with SINR threshold β . In the paper, since the typical node
(1)−1 H pair can work in either FD mode or HD mode with probabil-
zb0 = (sH
b0 ub0 ,a0 ) sb0 , (13)
ity p1 or p2 , respectively, the overall TC performance at the
(1)
where ub0 ,a0
denotes the first left singular vectors of typical node pair is the sum of the products of each mode’s
channel matrix Hb0 ,a0 , and sb0 with size of Nr × 1 occurrence probability and its respective TC, which is
in this case is used to span null space of Hb0 ,bk wbk ,
c() = p1 c(, δ = 1) + p2 c(, δ = 0), (17)
bk,k=1,...,lHD ∈ΦHD /b0 , Hb0 ,bk wbk , bk,k=1,...,lFD ∈ΦFD /b0 , and
Hb0 ,ak wak , ak,k=1,...,lFD ∈ΦFD /a0 . Here, we must guarantee where c(, δ) denotes the conditional TC when the typical
lHD + 2lFD ≤ Nr − 2. Similar way to formulate za0 and node pair works in either FD mode (i.e. δ = 1) or HD
post-processing vectors for other nodes in FD mode. mode (i.e. δ = 0). It is worth noting that, to complete a
For the node pairs worked in HD mode, the pre- bidirectional transmission, each node pair should exchange
processing and post-processing beamforming vectors can messages between its two nodes. In this case, if a node
be formulated exactly following the case where Nt ≤ Nr pair works in HD mode, two time slots are needed to
in FD mode discussed above, except that the receiver in this exchange messages. In other words, for a fair comparison
case will have Nr − 1 DoFs to cancel maximum Nr − 1 and fixing the bidirectional communication time duration,
nearest inter-node interferers and/or interferer pairs. Fol- the calculation of OP for HD node pair should double its
lowing above beamforming design, the received signal at target rate or halve its transmission rate, i.e., by setting SINR
typical node a0 after SI and l nearest inter-node interferers threshold to be 2(2−δ)R − 1 similar to the work in [35].
cancellation can be expressed as
−α/2 √ −α/2
X
zHa0 ya0 = L γa0 ,b0 sb0 + rbk ha0 ,bk sbk 4.1 Derivation of General TC Upper Bound
bk ∈ΦHD /b0
 k>lHD
 To derive general TC expression, i.e., c(, δ = 1 or δ = 0),
−α/2 we firstly need to exploit the probability of successful trans-
X
+ rbk ha0 ,bk sbk + ra−α/2
k
ha 0 ,ak
sak
(bk ,ak )
mission for the typical node a0 . According to the proposed
∈ΦFD /(b0 ,a0 ) beamforming design, such probability can be given by
k>lFD
0
Z SINRa0 ≥ β]
s(λ, δ), Pr[
+ δha0 sa0 + va0 , (14) ^
∞
= Fγca (xβLα )fIall (x)dx, (18)
where γa0 ,b0 is the largest eigenvalue of for H̃a0 ,b0 H̃H
a0 ,b0 0
0 ,b0

the case where Nt > Nr or the largest eigenvalue of

0 where β = 2(2−δ)R − 1 is the target SINR threshold for the
Ha0 ,b0 HHa0 ,b0 for the case where Nt ≤ Nr , and ha0 , typical node working either in FD mode or in HD mode de-
zHa0 ∆a0 wa0 is the residual SI effect due to the SI channel es- pending on δ ; Fγca ,b (·) denotes the complementary cumu-
timation error. Based on the proof in [21] and similar to [23], 0 0
lative distribution function (CCDF) of γa0 ,b0 ; fIall (·) denotes
[34], the residual SI power after implementing the proposed
0 the probability density function (PDF) of the sum of all co-
beamforming design can be determined as |ha0 |2 = σSI 2
, and
2 2 channel interference plus noise normalized by the transmit
the noise power can be determined as |va0 | = σ0 . Then, the 0 0 σ2
corresponding SINR at typical node a0 can be expressed as power P , i.e., Iall , IkHD + IkFD + δσSI 2
+ P0 .
Following the claim in [20], the explicit form of CCDF
L−α γa0 ,b0 of γa0 ,b0 is difficult to find due to the unknown sum-of-
SINR
^a =
0 σ02
, (15)
0
IkHD + IkFD + δσSI
2 +0
exponentials-and-polynomials form of Fγca ,b (·). However,
P 0 0
0 0 if the largest eigenvalue γa0 ,b0 comes from a Wishart matrix,
where IkHD rb−α |ha0 ,bk |2 , and IkFD
P
, bk ∈ΦHD /b0 ,
k>lHD
k its bounds can be defined, and the corresponding CCDF
P −α 2
+ ra−α |ha0 ,ak |2 ). can always be expressed as a sum of terms of the form
(bk ,ak )∈ΦFD /(b0 ,a0 ) (rbk |ha0 ,bk |
k>lFD
k
xi e−jx . Thus, the Laplace transform of the accumulated
co-channel interference can be exploited to find a general
4 H YBRID -D UPLEX T RANSMISSION C APACITY expression of the successive transmission probability upper
In general, there is a probability that high spatial through- bound, thereby, to obtain the general TC upper bound.
put is obtained accompanied with unacceptably high out- It is worth noting that such TC upper bound in general
age, which results in a large number of wasted transmis- can help to quickly identify the performance limitation of
sions. This motivates the statistic TC framework, following the considered network with different parameters setup.
Slivnyak’s theorem and the stationarity of Φ, to be exploited Moreover, with the proposed beamforming method, the
as the maximum allowable spatial density of successful derived TC upper bound can also be used to compare to
transmission multiplied by their data rates given an outage other TC bounds with different beamforming strategies. In
probability (OP) constraint at the typical receiver [19]. For a this case, potential system designers will be able to fast
given OP constraint  and its corresponding expression, i.e., identify whether the proposed beamforming method can
q(λ), as the function of user density λ, the TC formula is lead to better performances in comparison to others.
given by [19] In detail, according to the proposed transceiver beam-
forming design in Sec. III, for the case where the node
c() = q −1 ()(1 − )R,  ∈ [0, 1], (16) works in FD/HD mode and Nt ≤ Nr , γa0 ,b0 is obtained
−1
where λ = q () is the inverse of q(λ) that satisfies the directly from Ha0 ,b0 HH H
a0 ,b0 , where Ha0 ,b0 Ha0 ,b0 is a Wishart
target OP , and R = log2 (1+β) is a target transmission rate matrix due to independent and identical complex Gaussian

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
6

distribution on each element of Ha0 ,b0 . In this case, we have and

[36] Z 2π
dϕ
kHa0 ,b0 k2 F (t0 , r) , α . (26)
≤ γa0 ,b0 ≤ kHa0 ,b0 k2 , (19) 1+ 0 t0 (r2 + L2 + 2r cos ϕ)− 2
Nt
In addition, we also defined
where kHa0 ,b0 k2 follows Γ(Nt Nr , 1) distribution, and Z ∞  
its corresponding CCDF has the series expansion as 1
PNt Nr −1 xk −x ΩFD (t0 , lFD ) , 2π − F (t0
, r) rdr,(27)
c
FkH 2 (x) =
1 + t0 r−α
k! e . For the case where the E{rlFD }
a0 ,b0 k k=0
node works in FD mode and Nt > Nr , γa0 ,b0 is obtained with
from H̃a0 ,b0 H̃H a0 ,b0 . Here, we have Γ(E{lFD } + 12 )
− 21
E{rlFD } ≤ (p1 λπ) . (28)
H̃a0 ,b0 H̃H H H
a0 ,b0 = Ha0 ,b0 Ṽb0 Ṽb0 Ha0 ,b0
Γ(E{lFD })
(a) Proof. See Appendix A.
= Ha0 ,b0 Tb0 Λb0 TH H
b0 Ha0 ,b0 , (20)
From above equations (23) to (28), we can easily identify
where Ṽb0 is formulated from the estimated SI channel Ĥb0 that the derived SP-UB is strictly monotonically decreasing
similar to formulated Ṽa0 in Sec. III. In addition, Step (a) in function w.r.t. λ. Thus, by setting s(λ, δ) = 1 −  and
(20) is obtained by applying the eigenvalue decomposition implementing the bi-section searching algorithm [38], the
of Ṽb0 ṼbH0 , where Λb0 = diag([1, . . . , 1, 0, . . . , 0]), and Tb0 user density λ that satisfies the target OP  can be obtained,
| {z } | {z }
Nt −Nr Nr and the overall TC-UB at the typical node pair can be
is a Nt × Nt unitary matrix whose columns are the eigen- formulated according to (16) and (17) at the beginning of
vectors of Ṽb0 ṼbH0 . Then, by applying the unitary invariance this section.
property on Ha0 ,b0 , we can claim that H̃a0 ,b0 H̃H a0 ,b0 is also
a Wishart matrix regardless of the distribution of the esti-
4.2 The TC Upper Bound: Compact Versions
mated SI channel Ĥb0 . In this case, we have
The general TC-UB derived in above subsection can only be
kH̃a0 ,b0 k2 obtained by implementing the bi-section numerical method.
≤ γa0 ,b0 ≤ kH̃a0 ,b0 k2 , (21)
Nr Enlightened by the work in [19], in this subsection, we
where kH̃a0 ,b0 k2 follows Γ[Nr (Nt − Nr ), 1] distribution, exploit the dominating interference region to help to find a
and its corresponding CCDF has the series expansion as more compact and manageable expression for TC-UB with
FkcH̃
PNr (Nt −Nr )−1 xk −x the pay of relatively loosing the bounds. As defined in
2 (x) = k=0 k! e . Consequently, the
a0 ,b0 k [19], the dominating interferer is an interferer which itself
following Theorem gives the general upper bound of suc-
is enough to trigger system outage. In our case, due to the
cess probability defined in (18).
proposed nearest inter-node interference cancellation, only
Theorem 1: Consider the proposed transceiver beamform- one interferer (plus the residual SI if the typical node is in FD
ing design and the provided upper bounds of γa0 ,b0 as in mode) may not be enough to trigger system outage. There-
(19) and (21).3 By defining t , βLα , n , [Γ(N̂ + 1)]−1/N̂ , fore, we generalize the definition of dominating interference
N̂ , Nt Nr for HD case and FD case where Nt ≤ Nr , and region by exploiting d (≥ 1) interferers and/or interferer
N̂ , Nr (Nt − Nr ) for FD case where Nt > Nr , the success pairs after the l nearest interference cancellation. Here, we
probability upper bound, i.e., SP-UB, can be formulated as have
" #
N̂
" ! # (a) L−α γa0 ,b0
X
k−1 N̂ 0
s(λ, δ) ≤ Pr 2 ≥ β
s(λ, δ) ≤ (−1) LIall (t ) , (22) HD00 + I FD00 + δσ 2 + σ0
k " Ik k SI P #
k=1 0 t ,knt (b) L−α γa0 ,b0
where ≤ Pr 2 ≥ β , (29)
00 000
IkHD + IkFD + δσSI 2 + σ0
P

LIall (t0 )' exp − λ(p2 ΩHD (t0 , lHD ) + p1 ΩFD (t0 , lFD )) 00
where, in Step (a), IkHD , rb−α |ha0 ,bk |2
P
bk ∈ΦHD /b0 k
σ02
 
lHD +dHD ≥k>lHD
0 2
− t δσSI + . (23) with dHD , bp2 de as the average dominating interferers in
P 00
HD mode, and IkFD , (bk ,ak )∈ΦFD /(b0 ,a0 ) (rb−α |ha0 ,bk |2 +
P
k
In (23), we defined lFD +dFD ≥k>lFD
Z ∞ ra−α
k
|ha0 ,ak |2 ) with dFD = d − dHD as the average
ΩHD (t0 , lHD ) , (2π − F (t0 , r)) rdr (24) dominating interferer pairs in FD mode. In addition,
E{rlHD } 00 000
in
P Step (b) of (29), we have IkFD ≥ IkFD ,
with r −α
(|ha0 ,bk | 2
+ |ha0 ,ak | 2
) by assuming
(bk ,ak )∈ΦFD /(b0 ,a0 ) bk
− 21 Γ(E{lHD } + 12 ) lFD +dFD ≥k>lFD

E{rlHD } ≤ (p2 λπ) (25) rbk = rak , ∀k ∈ ΦFD /(b0 , a0 ), where such operation will
Γ(E{lHD }) lose certain tightness of the derived TC-UB, on the other
hand, it can provide a upper bound of s(λ, δ) and simplify
the performance analysis as [23].
Following the explanation in [19], we call dominating
3. Here, we don’t consider success probability lower bound because
such bound is usually not sufficiently tight especially when the number interferers because they must be within some distances of
of transceiver antennas are small [37]. the origin. Define rbl+1 ≤ . . . ≤ rbk ≤ . . . ≤ rbl+d as

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
7

the distances between d dominating interferers/interferer (OP-LB). Unlike the work in [19] and references in it, where
pairs and the typical receiver after the l nearest inter- the user density λ can be easily extracted from the formu-
00 000
node interference cancellation. We set IkHD + IkFD ≥ lated OP equation or via its asymptotic analytical form,
−α PlHD +dHD l +d in our case, since λ is included in γ(·, ·) in (31), it is not
rbl+d [ k=lHD +1 |ha0 ,bk |2 + k=lFD +1 (|ha0 ,bk |2 + |ha0 ,ak |2 )]
P FD FD

to further upper bound the success probability similar to straightforward to get the expression for λ. This motivates
the work in [21]. Therefore, the radius of the dominating us to propose two independent methods to find λ. For the
interference region is given by first method, Taylor expansion onto (31) gives
γ(E{l} + d, λπ E Rd2 )
( ) 
L−α γa0 ,b0 l
q (λ, δ, d) ≥
Rd , rbl+d : 2 < β
00 000
IkHD + IkFD + δσSI 2 + σ0 Γ(E {l} + d)
 P (π E Rd2 )E{l}+d
(a)  L−α γa0 ,b0  =
≥ rbl+d : < β Γ(E{l} + d)
σ 2
r−α I + δσ 2 + 0 ∞
(−π E Rd2 )m λm+E{l}+d
  
bl+d k SI P
X
· . (32)
! α1 m! m + E{l} + d
Ik m=0
= σ02
, (30)
2 −
β −1 L−α γa0 ,b0 − δσSI Then, Newton-Raphson method as in [40] can be used to
P
iteratively find λ. Specifically, let’s start from an initial guess
where λ0 for a root of the function q l (λ, δ, d) = , which is
lHDX
+dHD lFDX
+dFD
1 1
Ik , |ha0 ,bk |2 + (|ha0 ,bk |2 + |ha0 ,ak |2 ), λ0 = 2 [Γ(E{l} + d)(E{l} + d)] E{l}+d , (33)
π E {Rd }
k=lHD +1 k=lFD +1
where λ0 is formulated by solving q l (λ0 , δ, d) =  with a
and Step (a) gives a receive power lower bound of dominat-
single term of (32). To refine this approximation, we have
ing interferers. It is worth noting that ‘<’ within the bracket
of (30) indicates the dominating interferers/interferer pairs q l (λ0 , δ, d) − 
geometrically located inside a disk centered at the typical
λ1 = λ0 − , (34)
q l0 (λ0 , δ, d)
receiver a0 . 0
where q l (λ0 , δ, d) is the first order derivative of q l (λ0 , δ, d)
Then, the SP-UB can be formulated by exploiting its in terms of λ0 . Consequently, this should give
equivalent event that not all d dominating interferers 
−E{l}−d
λn+1 = λn + λn eλn πE{Rd } λn π E Rd2
2
and/or interferer pairs after the proposed interference can-
cellation is located within the dominating interference re- · Γ(E{l} + d, λn π E Rd2 + ( − 1)Γ(E{l} + d)),


gion. Thus, we have
(35)
su (λ, δ, d) , Pr rbl+d
 
 > Rd where Γ(a, b) is the upper incomplete gamma function.
γ(l + d, λπRd2 )

(a)
= 1−E Then, following the standard TC formula, we have
Γ(l + d)
cu (, δ, d) ≤ λn+1 (1 − )R,
(  )
(b) γ(l + d, λπ E Rd2 ) (36)
≤ 1−E u
Γ(l + d) where c (, δ, d) is the TC-UB on conditions of δ and d. It
is worth noting that, in general, a few number of iterations
γ(E{l} + d, λπ E Rd2 )

(c)
≤ 1− . (31) will lead to a sufficiently accurate value of λ.
Γ(E{l} + d) For the second method, the expression of λ can be
In (31), Step (a) is formulated by exploiting the PDF of rbl+d obtained with the help of following Lemma.
as in [39], where E{·} is w.r.t. all random variables involved Lemma 1: Based on Theorem 1 in [41], and given
in the form, i.e., l, dHD , dFD , γa0 ,b0 , |ha0 ,bk |2 , and |ha0 ,ak |2 . Z ∞
p

1
 Z x
p
In addition, γ(l + d, λπRd2 )/Γ(l + d) is the regularized lower e−t dt = Γ 1 + − e−t dt, (37)
x p 0
incomplete gamma function, and the function in terms of
Rd2 is convex for l ≥ λπRd2 − d + 1. Thus, applying Jensen’s the regularized
 incomplete Gamma function
inequality w.r.t. Rd2 yields Step (b), where the outer E{·} γ(E{l} + 1, λπ E Rd2 )/Γ(E{l} + 1) in (31) can be lower
is w.r.t. l, and the inner E{·} is w.r.t. the random variables bounded by
included in Rd . Then, Step (c) is formed by applying an- γ(E{l} + d, λπ E Rd2 ) h

2
iE{l}+d
other Jensen’s  inequality and exploiting the convexity of ≥ 1 − e−κλπE{Rd } , (38)
Γ(E{l} + d)
γ(l + d, λπ E Rd2 )/Γ(l + d) interms

of l. In this case, the −1/(E{l}+d)
function is convex for l ≥ λπ E Rd2 − d. It is worth noting where κ = [Γ (E{l} + d + 1)] and the equality
that, the compact TC-UBs derived in this subsection are only holds when E{l} + d = 1.
valid when l ≥ λπRd2 − d + 1 and l ≥ λπ E Rd2 − d. This is

Proof. The proof is ignored since (38) can be easily derived
because we need to preserve the direction of inequalities in
with some mathematical manipulations.
(31). For other conditions, we have to resort to the general
TC-UB derived in Sec. IV-A to analyze the network. Based on Lemma 1, we can further transfer the OP-LB,
To obtain compact TC-UBs, we first need to find the i.e., 1 − su (λ, δ, d), to
expression of λ by solving q l (λ, δ, d) , 1 − su (λ, δ, d) = , h 2
iE{l}+d
where q l (λ, δ, d) denotes the conditional OP lower bound q l (λ, δ, d) ≥ 1 − e−κλπE{Rd } . (39)

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
8

Nr − ω b Nr2−ω c
 
Then, the expression of user density λ can be easily found C8 = p1 , (51)
by solving q l (λ, δ, d) = . So that, the TC-UB in this case can 2
be given by and ζ = 1 if Nr2−ω is not an integer, otherwise ζ = 0. In
( − 1)R  1
 addition, for all above expressions, ω ∈ {1, 2}, ω = 1 if
cu (, δ, d) ≤ log 1 −  E{l}+d . (40) the typical node works in HD mode or FD mode where
κπ E{Rd2 }
Nt > Nr , otherwise ω = 2.
It is worth noting that (36) is tighter than (40), since Newton-
Raphson method iteratively finds the exact value of λ with- Proof. See Appendix B.
out any equation bounding process like (38). Then, given the number of dominating interferers d,
E{Rd2 } can be expressed as
4.3 Exact Expression of Random Variables   α2 
To obtain the final expressions of general SP-UB from (??)  lHDX
 +dHD lFDX+dFD 

2 2
and compact versions of TC-UB from (36) and (40), respec- E{Rd } = E  |ha0 ,bk | + ψa0 ,(bk ,ak ) 
tively, we should find the exact expressions for E{lHD },

 k=lHD +1 
k=lFD +1 
E{lFD }, E{l}, and E{Rd2 }. Lemma 2 provides the analytical ( 2 )
−α
σ02

2 γa0 ,b0 2
expressions for E{lHD }, E{lFD }, and E{l}. · βαE − βδσ SI − β
Lemma 2: Given probabilities of FD/HD selection, i.e., p1 Lα P
and p2 , and transceiver antenna configurations, the average ( l +d lFDX+dFD
! α2 )
(a) HDXHD
number of cancelled hybrid-duplex interferers/interferer ≥ E |ha0 ,bk |2 + ψa0 ,(bk ,ak )
pairs can be expressed as k=lHD +1 k=lFD +1
| {z }
E{l} = C1 + C2 + C3 , (41) ,ηa0

where − α2
E{γa0 ,b0 } σ02

2
r −ω
2
C1 = (Nr − ω)pN , (42) ·β α − βδσSI − β , (52)
2 Lα P
b Nr2−ω c !
where ψa0 ,(bk ,ak ) , |ha0 ,bk |2 + |ha0 ,ak |2 , and Step (a)
Nr − k − ω k Nr −2k−ω
X 
C2 = (Nr − k − ω) p1 p2 is obtained by applying Jensen’s inequality on the sec-
k=1
k ond E{·} of the equation. In addition, the first E{·} in
!
Nr − k − ω k Nr −2k−ω+1 Step (a) is w.r.t. dHD , dFD , |ha0 ,bk |2 and ψa0 ,k . By giving

+ p1 p2 , (43)
dHD and dFD , we have lk=l
P HD +dHD
k−1 HD +1
|ha0 ,bk |2 ∼ Γ(dHD , 1),
PlFD +dFD
Nr − ω k=lFD +1 ψa0 ,(bk ,ak ) ∼ Γ(2dFD , 1), ηa0 ∼ Γ(d + dFD , 1),
    N −ω
d r e 2/α
C3 = ζ Nr − − ω p1 2 , (44) and ηa0 follows a generalized Gamma distribution with
2 2/α
parameters [1, (d + dFD )α/2, α/2]. Thus, E{ηa0 |d, dFD } =
and ζ ∈ {0, 1}, ζ = 0 if Nr2−ω is an integer, otherwise ζ = 1. Γ(d + dFD + 2/α)/Γ(d + dFD ). Moreover, since dFD (or dHD )
Similarly, the average number of cancelled HD interferers is follows Binomial distribution with parameters d and p1 (or
E{lHD } = C4 + C5 , (45) p2 ), we have
d
!
where n 2o X d k Γ(d + k + α2 )
b Nr2−ω c−%
E ηa0 =
α
p1 (1 − p1 )d−k , (53)
X
!
Nr − k − ω k Nr −2k−ω k=0
k Γ(d + k)
C4 = (Nr − 2k − ω) p1 p2 ,
where ab denotes b-combinations from a given set of a


k=0
k
(46) elements.
b Nr2−ω c !
Following the discussion in Sec. IV-A, the eigenvalue
X Nr − k − ω k Nr −2k−ω+1
C5 = (Nr −2k −ω +1) p1 p2 , γa0 ,b0 can be upper bounded as (19) and (21). Therefore,
k−1
k=1 the corresponding expected values are upper bounded by
(47)
and % ∈ {0, 1}, % = 1 if (Nr − ω)/2 is an integer, otherwise, E{kHa0 ,b0 k2 } = Nt Nr and E{kH̃a0 ,b0 k2 } = Nr (Nt −
% = 0. Moreover, we also have Nr ), respectively.4 Subsequently, by inserting the calculated
2/α 0
E{ηa0 }, E{γa0 ,b0 }, E{|ha0 |2 }, and E{|va0 |2 } into (52), we
E{lFD } = C6 + ζC7 + (1 − ζ)C8 , (48) obtained the lower bound of E{Rd2 }. Then, by inserting the
where derived E{lHD } into (25) and E{lFD } into (28), we obtain
b Nr2−ω c−1 ! the general SP-UB in (22). By inserting the derived E{l} and
Nr − k − ω k Nr −2k−ω

X E{Rd2 } into (36) or (40), we obtained the compact versions
C6 = k p1 p2
k of TC-UB. It is worth noting that, to find the optimal d, we
k=1
can minimize cu (, δ, d) in a numerical way.
!
Nr − k − ω − 1 k+1 Nr −2k−ω−1

+ p1 p2 , (49)
k
5 N UMERICAL R ESULTS AND D ISCUSSION
In this section, we first compare the proposed beamforming
!
Nr − ω Nr − b Nr2−ω c − ω
 
b Nr2−ω c
C7 = p1 design method with the existing beamforming strategies,
2 b Nr2−ω c 
Nr −ω
Nr −2b 2 c−ω b Nr2−ω c+1 4. For both antenna configurations, we select γa0 ,b0 upper bound to
· p2 + p1 , (50) keep the same direction of inequality in (52).

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
9

0.7 10 0
Proposed method
SVD
SVD + partial ZF
0.6 Partial ZF
Transmission Capacity (linear scale)

Transmission capacity (log scale)

0.5

10-1

0.4

0.3
Simulation, (N t >Nr =2)
UB-CV-1, (N t >Nr =2)
10-2
0.2 UB-CV-2, (N t >Nr =2)
UB-general, (N t >Nr =2)
Simulation, (N r >Nt =2)
0.1 UB-CV-1, (N r >Nt =2)
UB-CV-2, (N r >Nt =2)
UB-general, (N r >Nt =2)
0 10-3
11 12 13 14 15 16 17 18 5 6 7 8 9 10 11
Total number of antennas per node N Total number of antennas per node N

Fig. 1: Transmission capacity versus the total number of antennas per Fig. 2: Transmission capacity versus the total number of antennas per
node in comparison among different beamforming methods, where node, where L = 1, P = 20, σSI 2 = σ 2 = 0.001,  = 0.1, R = 5 and
2 0
L = 1, P = 20,  = 0.1, σi,SI = 0.01, β = 1, p1 = 1, and p1 = 1 − p2 = 0.8.
FD FD
Nt > Nr = 5. 100

i.e., SVD based method as in [20]; SVD plus partial ZF based

method as in [21], and purely partial ZF based method
as in [22], in terms of the exact TC performances. For a
Transmission capacity (log scale)

fair comparison, we assume all these existing beamforming

10-1
strategies will spend one of their receiver DoFs to cancel
the estimated SI. After that, we examine the accuracy of
our derived TC-UBs, evaluate the impact of CSI error and
hardware distortions, and also test the TC performances for Simulation, (N t >Nr =2)
different HD/FD operation ratios. Following, we identify -2
UB-CV-1, (N t >Nr =2)
10
the break-even point between pure FD mode and pure HD UB-CV-2, (N t >Nr =2)
UB-general, (N t >Nr =2)
mode. Throughout the experiments, we assume that the Simulation, (N r >Nt =2)
simulated ad-hoc network lies on a 2-D disk and contains UB-CV-1, (N r >Nt =2)
a number of transceiver node pairs, which are marked UB-CV-2, (N r >Nt =2)
UB-general, (N r >Nt =2)
Poisson random variables with mean equal to 2000. The disk
10-3
area is adjusted according to the density, and we assume the 5 6 7 8 9 10 11

free space path-loss exponent α is equal to 2.25. The number Total number of antennas per node N

of iterations for the simulated results is 105 .

Fig. 3: Transmission capacity versus the total number of antennas per
node, where L = 1, P = 20, σSI 2 = σ 2 = 0.001,  = 0.1, R = 5 and
0
5.1 Diffident Beamforming Methods Comparison p1 = 1 − p2 = 0.2.
In this experiment, we compare the proposed beamforming
and leave SI to be cancelled at receiver side. In addition,
design method with existing beamforming strategies in
the proposed method outperforms another two strategies
terms of TC performances getting from the exactly simu-
throughout whole presented N configurations. It is worth
lated results and configuring transceiver antenna per node
noting that, for the case that Nt ≤ Nr , the proposed
as Nt > Nr . As shown in Fig. 4, the proposed beamforming
beamforming method will be converged to the SVD plus
method outperforms SVD plus partial ZF based strategy
partial ZF based strategy.
when the total number of antennas per node is no less than
12. This is because our proposed method takes the advan-
tage of the specified transceiver antenna configuration, i.e., 5.2 TC Upper Bounds Verification
Nt > Nr . In this case, the estimated version of SI can be In this experiment, we examine the tightness of derived
cancelled at transmitter side, and it will leave an additional TC-UB in Sec. IV-A (i.e. ‘UB-general’) and two compact
DoF at receiver side for inter-node interferer cancellation. versions of TC-UB in Sec. IV-B (i.e. ‘UB-CV-1’ and ‘UB-
However, when N = 11, the proposed method shows a bit CV-2’) in comparison to the exactly simulated results with
worse performance by comparing with SVD plus partial ZF different transceiver antenna configurations and different
based strategy. This is because, in our case, most of DoFs FD/HD operation ratios (i.e. p1 and p2 ). As shown in Fig. (2)
at transmitter side have been used to cancel the estimated and Fig. (3), the derived ‘UB-general’ gives relative small
version of SI and leave few DoFs at transmitter side to performance gap when the probability of FD node pairs is
boost the desired signal. In contrast, SVD plus partial ZF small, and the number of receive antenna is larger than the
strategy is able to use the DoFs to boost the desired signal, number of transmit antenna, or, when the probability of FD

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
10
0
10-1 10

Transmission capacity (log scale)

Transmission capacity (log scale)

10-1
10-2

Increasing N r
10-2
Simulation, (N t =2,N r =6)
10-3 UB-general, (N t =2,N r =6)
Simulation, (N t =4,N r =4)
2 2 UB-general, (N t =4,N r =4)
UB-general, ( =0.1, t= r =0) Simulation, ( =0.001, = r =0.001)
SI SI t
Simulation, (N t =6,N r =2)
2 2
Simulation, ( SI
=0.1, t= r =0) UB-general, ( SI
=0, t= r
=0)
UB-general, (N t =6,N r =2)
2 2
UB-general, ( =0.001, t= r =0) Simulation, ( =0, = =0)
SI SI t r 10-3
2 2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Simulation, ( SI
=0.001, t= r =0) Simulation, ( SI
=0, t= r
=0.001)
Probability of FD node pairs in the network
10-4
5 6 7 8 9 10 11
Total number of antennas per node N
Fig. 5: Transmission capacity versus probability of FD node pairs in the
network with various transceiver nodes’ antenna configurations, where
L = 1, P = 20, σSI2 = 0.0001, σ 2 = 0.001, and R = 5.
Fig. 4: Transmission capacity versus the total number of antennas per 0
node with various SI channel estimation error and hardware distortion
configurations, where Nt > Nr = 2, L = 1, P = 20, σ02 = 0.001, and
R = 5. as they follow the same distribution. 2) Due to limited
number of inter-node interference cancellation, the impact
node pairs is large, and the number of receive antenna is of inter-node interference CSI error may not as strong as
smaller than the number of transmit antenna. In addition, SI CSI error. As shown in the figure, by increasing either
for ‘UB-general’ with other cases and two compact TC-UBs, the variance of SI channel estimation error or the value of
same order-wise performance trend as the simulated results hardware distortion coefficients, the TC performance will
can be observed. As we mentioned above, the key factors decrease. In addition, increasing the value of hardware
which affect the tightness of derived ‘UB-general’ includes: distortion coefficients will also enlarge the TC gap between
1) the upper bound of the largest eigenvalue; 2) inequality the simulated result and ’UB-general’ ones especially when
implemented in (63); 3) the expected number of cancelled the total number of antennas per node N is small. Such
nearest FD interferer pairs involved in (28), respectively. As gap is even enlarged when the variance of SI channel
Fig. (2) shown, when the network involves more FD node estimation error is increased. However, as the number of
pairs (e.g. p1 = 0.8), the tightness of derived ‘UB-general’ antennas per node increasing, the enlarged performance
is quite sensitive to larger number of FD interferer pairs’ gap is reduced. This is mainly because the beamforming
cancellation, where the third impact factor is dominant the gain in this case dominates the TC performance. It is worth
2
performance. In this case, precise estimation of the distances noting that, following additional test, when σSI = 0.1,
between the nearest FD interferer pairs and the typical κt = κr = 0.001, the TC performance is almost zero. For
receiver is quite important. As Fig. (3) shown, when the ease of distinguishing the results, we did not plot the curve
network involves more HD node pairs (e.g. p1 = 0.2), the in the figure. Similar performance is also applied for the case
expected number of cancelled FD interferer pairs is not a where κt = κr = 0.1.
big impact factor. Therefore, the second impact factor (i.e., In Fig. (5), we also examine the TC performances for
inequality implemented in (63)) will dominate the perfor- a full range of FD/HD operation ratios with different
mance. In addition, the impact of the largest eigenvalue is transceiver antenna configurations. As shown in the fig-
normally vanished when the difference between the number ure, ‘UB-general’ gives relative small gap to the simulated
of transmit antennas and the number of receive antennas are results for certain cases. Specifically, for the case where
quite large. It is worth noting that, if the channel matrices is Nt = 2, Nr = 6, ‘UB-general’ is tight around small values
rank deficient, reflecting to Fig. (2) and Fig. (3), we will have of p1 . This is because the bound of E{rlFD } in (28) gradually
high number of antennas but can only handle less number of enlarges the performance gap as p1 increases. In contrast, for
co-channel interference in comparison to the full rank case. the case where Nt = 6, Nr = 2, few number of FD interferer
To evaluate the impact of CSI error and hardware dis- pair needs to be cancelled. Thus, E{rlFD } effect is vanished,
tortions, Fig. (4) compares both simulated and ‘UB-general‘ and ‘UB-general’ becomes tight around large values of p1 .
TC performances with different SI channel estimation errors, The performance of the case where Nt = 4, Nr = 4 is in the
2
i.e., σSI , and hardware distortion coefficients for both trans- middle of the above two. Moreover, performance gaps at
mit and receive chains, i.e., κt and κr . The reasons we only small values of p1 for all antenna configurations are mainly
picked SI channel estimation error and hardware distortion because of the inequality used in (63), and such effect is
to examine the system performance are in two fold: 1) Our weaken when p1 increases. Such observations align with
beamforming design mainly relies on the estimated version the discussion for Fig. (2) and Fig. (3). In addition, when
of CSI, where both SI and inter-node interference channel Nt = 6, Nr = 2, the simulated result shows that hybrid
estimation errors will have similar impact on SINR as long FD/HD deployment provides better TC performance than

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
11

0.14 0.25
2
Simulation, Pure FD ( SI
=0.01) Simulation, Pure FD ( 2
=0.04)
SI
2 2
Simulation, Pure FD ( SI
=0.015) Simulation, Pure FD ( =0.03)
0.12 SI
Simulation, Pure HD Simulation, Pure HD
2
UB-gerneral, Pure FD ( =0.01) 0.2 UB-gerneral, Pure FD ( 2
=0.04)
SI SI
Transmission capacity (linear scale)

Transmission capacity (linear scale)

2
break-even point UB-gerneral, Pure FD ( SI
=0.015) UB-gerneral, Pure FD ( 2
=0.07)
0.1 SI
UB-gerneral, Pure HD UB-gerneral, Pure HD

0.15
0.08

break-even point

0.06
0.1

0.04

0.05
break-even point
0.02

break-even point
0 0
5 10 15 20 25 5 10 15 20 25
Transmit power P (dB) Transmit power P (dB)

Fig. 6: Transmission capacity versus the signal-to-noise ratio in compar- Fig. 7: Transmission capacity versus the signal-to-noise ratio in compar-
ison between pure FD case and pure HD case, where L = 1, σ02 = 0.001, ison between pure FD case and pure HD case, where L = 1, σ02 = 0.001,
 = 0.1, R = 5, Nt = 5, and Nr = 3.  = 0.1, R = 5, Nt = 3, and Nr = 5.

the cases that operating all nodes in pure FD and pure increased number of receive antennas configured in Fig. (7),
HD modes. Such observation confirms that hybrid FD/HD the performance gap between ‘UB-general’ and simulated
deployment can also lead to the optimal performance if result is enlarged. This is why break-even points for both
the advanced beamforming method, like our proposed one, 2
‘UB-general’ and simulated result occur with different σSI
is implemented. Apart from those, with the fixed total values.
number of transceiver antennas per node, as shown in
Fig. (5), the TC performance can be improved accompanied
6 C ONCLUSION
with the increased number of receive antennas, i.e., Nr .
This is because, with marginal SI effect, more inter-node In this paper, we have proposed a joint transceiver beam-
interferers/interferer pairs cancellation will lead to better forming design for hybrid FD/HD ad-hoc network to mit-
TC performance. igate SI and the nearest inter-node interferers/interferer
pairs. To characterize the proposed system performance,
we have derived a general TC-UB plus its two compact
5.3 Full-Duplex vs. HD-Duplex
versions. The results show that, the proposed beamforming
In this experiment, we use both simulated result and de- method outperforms the existing beamforming strategies,
rived ‘UB-general’ to find break-even points, where pure especially when the number of transmitter antennas is larger
FD (i.e. p1 = 1) and pure HD (i.e. p0 = 0) can have the than the number of receive antennas per FD node. In addi-
same TC performance. In addition, we also identify whether tion, the tightness of derived general TC-UB plus its two
the derived ‘UB-general’ can follow the same performance compact versions has been examined, where the derived
trend as the simulated result w.r.t. different levels of SI general TC-UB can offer relative tight TC performances for
cancellation and transceiver antenna configurations. Fig. (6) certain FD/HD operation ratios set-up and transceiver an-
gives the comparison for the case where Nt > Nr . As shown tenna configurations. For other cases, the order-wise perfor-
2
in the figure, break-even points happen when σSI = 0.015, mance can be observed. Moreover, we have also identified
and pure FD outperforms pure HD when transmit power the conditions when pure FD outperforms pure HD.
P is less than 6 dB for simulated result and 11 dB for ‘UB-
general’. Although the break-even point is shifted right a
bit, the TC performance trends between ‘UB-general’ and ACKNOWLEDGMENTS
simulated result are still the same, where the occurrence This work was supported by the Engineering and Physical
of the break-even point is due to the halved achievable Science Research Council (EPSRC) through the SENSE grant
rate (or doubled target rate) for HD transmission in a EP/P003486/1.
bidirectional message exchange. In this case, when transmit
power P is small, only a few transceiver pairs in pure HD
R EFERENCES
mode can achieve the target rate. As a result, pure FD is
better than pure HD. In contrast, when transmit power [1] J. Hou and M. Shikh-Bahaei, “Transmission capacity of full-duplex
MIMO ad-hoc network with limited self-interference cancella-
P is large, more transceiver pairs in pure HD mode can tion,” in Proc. IEEE Global Commun. Conf., Dec. 2018, pp. 1–7.
satisfy the target rate with less inter-node interference. Thus, [2] S. Lien, S. Shieh, Y. Huang, B. Su, Y. Hsu, and H. Wei, “5G
pure HD outperforms pure FD. Similar performance trend new radio: Waveform, frame structure, multiple access, and initial
can be observed in Fig. (7), where pure FD works better access,” IEEE Commun. Mag., vol. 55, no. 6, pp. 64–71, Jun. 2017.
[3] V. Towhidlou and M. Shikh-Bahaei, “Improved cognitive network-
than pure HD with small transmit power and/or in high ing through full duplex cooperative arq and harq,” IEEE Wireless
target rate environments. On the other hand, due to the Commun. Lett., vol. 7, pp. 218–221, Apr. 2018.

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
12

[4] A. Sabharwal, P. Schniter, D. Guo, D. W. Bliss, S. Rangarajan, [27] H. Ju, D. Kim, H. V. Poor, and D. Hong, “Bi-directional beam-
and R. Wichman, “In-band full-duplex wireless: Challenges and forming and its capacity scaling in pairwise two-way communica-
opportunities,” IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. 1637– tions,” IEEE Trans. Wireless Commun., vol. 11, no. 1, Jan. 2012.
1652, Sept. 2014. [28] R. K. Mungara, I. Thibault, and A. Lozano, “Full-duplex MIMO in
[5] D. Kim, H. Lee, and D. Hong, “A survey of in-band full-duplex cellular networks: System-level performance,” IEEE Trans. Wireless
transmission: From the perspective of phy and mac layers,” IEEE Commun., vol. 16, no. 5, pp. 3124–3137, May 2017.
Commun. Surv. Tutor., vol. 17, no. 4, pp. 2017–2046, 2015. [29] I. Atzeni and M. Kountouris, “Full-duplex MIMO small-cell net-
[6] M. Naslcheraghi, S. A. Ghorashi, and M. Shikh-Bahaei, “FD works with interference cancellation,” IEEE Trans. Wireless Com-
device-to-device communication for wireless video distribution,” mun., vol. 16, no. 12, pp. 8362–8376, Dec. 2017.
IET Commun., vol. 11, no. 7, pp. 1074–1081, May 2017. [30] N. Jindal, J. G. Andrews, and S. Weber, “Multi-antenna commu-
[7] J. I. Choi, M. Jain, K. Srinivasan, P. Levis, and S. Katti, “Achieving nication in ad hoc networks: Achieving MIMO gains with SIMO
single channel, full duplex wireless communications,” in Proc. transmission,” IEEE Trans. Commun., vol. 59, no. 2, pp. 529–540,
ACM MobiCom, 2010, pp. 1–12. Feb. 2011.
[31] M. Haenggi, Stochastic Geometry for wireless Networks. Cambridge,
[8] E. Everett, A. Sahai, and A. Sabharwal, “Passive self-interference
U.K.: Cambridge University Press, 2012.
suppression for full-duplex infrastructure nodes,” IEEE Trans.
[32] O. Taghizadeh, A. C. Cirik, and R. Mathar, “Hardware im-
Wireless Commun., vol. 13, no. 2, pp. 680–694, Feb. 2014.
pairments aware transceiver design for full-duplex amplify-and-
[9] M. Duarte and A. Sabharwal, “Full-duplex wireless communica- forward MIMO relaying,” IEEE Trans. Wireless Commun., vol. 17,
tions using off-the-shelf radios: feasibility and first results,” in no. 3, pp. 1644–1659, Mar. 2018.
Proc. 44th Asilomar Conf. Signals Syst. Comput., Nov. 2010, pp. 1558– [33] A. C. Cirik, S. Biswas, O. Taghizadeh, and T. Ratnarajah, “Robust
1562. transceiver design in full-duplex MIMO cognitive radios,” IEEE
[10] M. Jain and et al., “Practical, real-time, full duplex wireless,” in Trans. Veh. Technol., vol. 67, no. 2, pp. 1313–1330, Feb. 2018.
Proc. ACM MobiCom, Sept. 2011, pp. 301–313. [34] Y. Deng, L. Wang, M. Elkashlan, M. D. Renzo, and J. Yuan, “Mod-
[11] M. Duarte, C. Dick, and A. Sabharwal, “Experiment-driven char- eling and analysis of wireless power transfer in heterogeneous
acterization of full-duplex wireless systems,” IEEE Trans. Wireless cellular networks,” IEEE Trans. Commun., vol. 64, no. 12, pp. 5290–
Commun., vol. 11, no. 12, pp. 4296–4307, Dec. 2012. 5303, Dec. 2016.
[12] S. Huberman and T. Le-Ngoc, “MIMO full-duplex precoding: a [35] J. N. Laneman, D. N. C. Tse, and G. W. Wornell, “Cooperative
joint beamforming and self-interference cancellation structure,” diversity in wireless networks: Efficient protocols and outage
IEEE Trans. Wireless Commun., vol. 14, no. 4, pp. 2205–2217, Apr. behavior,” IEEE Trans. Inf. Theory, vol. 50, no. 12, pp. 3062–3080,
2015. Dec. 2004.
[13] D. Senaratne and C. Tellambura, “Beamforming for space division [36] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge, United
duplex,” in Proc. IEEE Int. Conf. Commun., Jun. 2011, pp. 1–5. Kingdom: Cambridge University Press, 1985.
[14] C. Psomas, M. Mohammadi, I. Krikidis, and H. A. Suraweera, [37] I. Atzeni and M. Kountouris, “Performance analysis of partial in-
“Impact of directionality on interference mitigation in full-duplex terference cancellation in multi-antenna UDNs,” in Proc. Asilomar
cellular networks,” IEEE Trans. Wireless Commun., vol. 16, no. 1, Conf. Signals, Syst., Comput. (ASILOMAR), Nov. 2016, pp. 699–703.
pp. 487–502, Jan. 2017. [38] R. Burden, A. Burden, and J. Faires, Numerical Analysis, 10th ed.
Boston, MA: Brooks Cole, Jan. 2015.
[15] T. Riihonen, S. Werner, and R. Wichman, “Mitigation of loopback
[39] M. Haenggi, “On distances in uniformly random networks,” IEEE
self-interference in full-duplex MIMO relay,” IEEE Trans. Signal
Trans. Inf. Theory, vol. 51, no. 10, pp. 3584–3586, Oct. 2005.
Process., vol. 59, no. 12, pp. 5983–5993, Dec. 2011.
[40] E. Suli and D. Mayers, An Introduction to Numerical Analysis, 1st ed.
[16] M. Mohammadi, H. A. Suraweera, Y. Cao, I. Krikidis, and C. Tel- Cambridge, UK: Cambridge University Press, Sept. 2003.
lambura, “Full-duplex radio for uplink/downlink wireless access [41] H. Alzer, “On some inequalities for the incomplete gamma func-
with spspatial random nodes,” IEEE Trans. Commun., vol. 63, tion,” Math. Comput., vol. 66, no. 4, pp. 771–778, Apr. 1997.
no. 12, pp. 5250–5266, Dec. 2015.
[17] R. Sultan, K. G. Seddik, Z. Han, and B. Aazhang, “Asymptotic
behavior analysis and performance optimization in full duplex
massive MIMO,” in Proc. IEEE Global Commun. Conf., Dec. 2017,
pp. 1–6.
[18] S. Weber, J. G. Andrews, and N. Jindal, “The effect of fading,
channel inversion, and threshold scheduling on ad hoc networks,”
IEEE Trans. Inf. Theory, vol. 53, no. 11, pp. 4127–4149, Nov. 2007.
[19] ——, “An overview of the transmission capacity of wireless net-
works,” IEEE Trans. Commun., vol. 58, no. 12, pp. 3593–3604, Dec.
2010. Jiancao Hou (S’09-M’15) received the B.Sc.
[20] A. M. Hunter, J. G. Andrews, and S. Weber, “Transmission capacity degree in information science (honor pro-
of ad hoc networks with spatial diversity,” IEEE Trans. Wireless gram) from China Agricultural University, Beijing,
Commun., vol. 7, no. 12, pp. 5058–5071, Dec. 2008. China, in 2008, the M.Sc. degree (with Distinc-
[21] R. Vaze and R. W. Heath, “Transmission capacity of ad-hoc net- tion) in radio frequency communication systems
works with multiple antennas using transmit stream adaptation from the University of Southampton, Southamp-
and interference cancellation,” IEEE Trans. Inf. Theory, vol. 58, ton, U.K., in 2009, and the Ph.D. degree in elec-
no. 2, pp. 780–792, Feb. 2012. trical engineering from the University of Surrey,
[22] K. Huang, J. G. Andrews, D. Guo, R. W. Heath, Jr., and R. A. Berry, Guildford, U.K., in 2014. From 2014 to 2016,
“Spatial interference cancellation for multiantenna mobile ad hoc he was a Research Fellow with the Institute for
networks,” IEEE Trans. Inf. Theory, vol. 58, no. 3, pp. 1660–1676, Communication Systems, University of Surrey.
Mar. 2012. Since 2017, he has been a Research Associate with the Department of
Informatics, King’s College London, London, U.K., where he has been
[23] Z. Tong and M. Haenggi, “Throughput analysis for full-duplex
strongly involved in the EPSRC SENSE Project. His current research
wireless networks with imperfect self-interference cancellation,”
interests include radio resource allocation and interference manage-
IEEE Trans. Commun., vol. 63, no. 11, pp. 4490–4500, Nov. 2015.
ment for full-duplex dense networks, stochastic geometry theory, MIMO
[24] J. Lee and T. Q. S. Quek, “Hybrid full-/half-duplex system analysis beamforming techniques, and unmanned aerial vehicle aided wireless
in heterogeneous wireless networks,” IEEE Trans. Wireless Com- communications. He also serves as a TPC Member for IEEE flagship
mun., vol. 14, no. 5, pp. 2883–2895, May 2015. conferences, such as ICC and GLOBECOM.
[25] D. Qin and Z. Ding, “Transport capacity analysis of wireless in-
band full duplex ad hoc networks,” IEEE Trans. Commun., vol. 65,
no. 3, pp. 1303–1318, Mar. 2017.
[26] S. Goyal, C. Galiotto, N. Marchetti, and S. S. Panwar, “Throughput
and coverage for a mixed full and half duplex small cell network,”
in Proc. IEEE Int. Conf. Commun., May. 2016, pp. 1–7.

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.3001071, IEEE
Transactions on Mobile Computing
13

Zhaohui Yang received the B.S. degree in in-

formation science and engineering from Chien-
Shiung Wu Honors College, Southeast Univer-
sity, Nanjing, China, in 2014, and the Ph.D. de-
gree in communication and information system
from the National Mobile Communications Re-
search Laboratory, Southeast University, Nan-
jing, China, in 2018. He is currently a Post-
doctoral Research Associate with the Center
for Telecommunications Research, Department
of Engineering, King’s College London, London,
U.K. His research interests include rate splitting, full-duplex, federated
learning, mobile edge computing, and UAV. He was an exemplary re-
viewer for IEEE Transactions on Communications in 2019.

Mohammad Shikh-Bahaei (S’96-M’00-SM’08)

received the B.Sc. degree from the University of
Tehran, Tehran, Iran, in 1992, the M.Sc. degree
from the Sharif University of Technology, Tehran,
in 1994, and the Ph.D. degree from the King’s
College London, U.K., in 2000. He has worked
for two start-up companies, and for National
Semiconductor Corporation, Santa Clara, CA,
USA (now part of Texas Instruments Inc.), on
the design of third-generation mobile handsets,
for which he has received three U.S. patents as
inventor and co-inventor, respectively. In 2002, he joined the King’s
College London as a Lecturer, where he is currently a full Professor. He
has since authored or co-authored numerous journal and conference
articles. He has been involved in research in the area of wireless com-
munications and signal processing for 25 years both in academic and
industrial organizations. His research interests include secure commu-
nication, resource allocation in full-duplex and cognitive dense networks,
visual data communications over the IoT, applications in healthcare, and
communication protocols for autonomous vehicle/drone networks. He
is a fellow of the IET and the Founder and the Chair of the Wireless
Advanced (formerly SPWC) Annual International Conference from 2003
to 2018.

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 18:36:29 UTC from IEEE Xplore. Restrictions apply.