You are on page 1of 15

This article has been accepted for publication in IEEE Transactions on Vehicular Technology.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

Joint Mode Selection and Resource Allocation for D2D and Femtocell
Users in Dense Heterogeneous Networks with Full Frequency Reuse
Laleh Eslami, Ghasem Mirjalily, Senior Member, IEEE, and Timothy N. Davidson, Fellow, IEEE

Abstract- We consider the problem of joint mode selection the users, including reuse (underlay) and dedicated
and resource allocation for D2D and femtocell users in a three- (overlay) modes [5]. In the reuse mode, devices use the
tier dense heterogeneous network in which all the users can same resources as other users. This method offers improved
reuse the spectrum. The goal is to maximize their total spectral efficiency at the cost of increased levels of
weighted sum rate, subject to minimum rate requirements
and maximum tolerable interference on the cellular system.
interference. In the dedicated mode, users utilize dedicated
Since the spectrum can be fully reused, the co-tier and cross- resources. This method offers reduced interference at the
tier interferences among the users result in a mixed integer cost of spectrum underutilization. Also, devices with the
non-linear, non-convex program that is difficult to solve capability to transmit in D2D mode can establish their
directly. Using insights into the structure of the interference, communication by using either one of the forms of D2D or
we derive a close, conservative approximation of the joint by using the (macro) cellular system [6,7].
problem that has a convex relaxation that is tight. That
enables good solutions to the joint problem to be obtained In order to enhance the capacity of the system we
using a customized iterative algorithm employing the consider, the macrocell users (MUs) and the small cell
Lagrange dual decomposition method. Since the proposed users are allowed to access the same spectrum resources to
iterative scheme is semi-distributed, the signaling overhead is transmit their data. Therefore, mutual interference between
mitigated. To further reduce the computational complexity, a the macrocell and the small cells (cross-tier interference),
low-complexity (primal) decomposition-based method is also
introduced, in which we select the transmission mode
and among the small cells (co-tier interference) is a
heuristically based on the traffic level in the network, and then significant challenge [8]. The existence of D2D users
sequentially perform admission control, power control, and (DUs) in reuse mode makes this complicated interference
sub-channel allocation for the femtocell users and D2D users. environment even more challenging. Therefore, effective
Our simulation results indicate that the proposed iterative and robust resource allocation and interference
and heuristic algorithms, on average, achieve around 94% management schemes must be adopted to prevent the
and 82% of the sum rate of the optimal Branch & Bound degradation of the network performance. Moreover, in a
method, respectively, and do so at much lower computational D2D-enabled HetNet, further gains can be obtained if users
costs. are allowed to transmit in different transmission modes [9].
Index Terms: Heterogeneous networks, femtocell, D2D Hence, there has been growing interest in investigating
communications, mode selection, resource allocation. joint resource allocation and mode selection in D2D-
enabled HetNets.
I. INTRODUCTION
A. Background B. Related Work
To meet the increasing data demand of emerging wireless Although the body of work on joint mode selection and
communication services and to provide better quality of resource allocation for DUs is significant, much of the work
service (QoS), new technologies are needed to improve does not consider HetNets. For instance, the authors in [10]
traditional cellular networks. Heterogeneous networks consider mode selection and channel allocation only for
(HetNets) and Device-to-Device (D2D) communications DUs and power assignment for both the D2D and cellular
are emerging as promising solutions [1,2]. HetNets were users to maximize the overall throughput of the network.
proposed to increase the network capacity and coverage at Their proposed approach is based on reinforcement
a reduced expense. In a HetNet architecture, a macro cell learning and a heuristic scheme in a classic macrocell
coexists with different small cells such as femtocells, network. The work in [11] investigates mode selection, and
picocells, and microcells [3]. In the D2D technique, two channel and power allocation for DUs in a macrocell
devices that share a high-quality communication channel network (and not a HetNet), and proposes a capacity-
form a D2D pair and communicate directly with each other oriented approach based on the hierarchical game. Among
without communicating through any base station. Due to the studies that consider a HetNet as the system model [12-
the short transmission range, the D2D devices can employ 21], few of them address the resource allocation or mode
low transmission powers, which reduces the interference selection for both the DUs and small cell users [12-16].
imposed on other users. As a result, the performance of Intending to maximize the total data rate, the work in [12]
cellular networks concerning data rate, delay, and power focuses on power allocation and mode selection for both
consumption can be enhanced [4]. Focusing on the inband DUs and femtocell users (FUs), considering a full
scenario, several transmission modes can be considered for frequency reuse technique. The authors in that work
propose a power control with interference-limited area and
This work was supported in part by the Natural Science and Engineering
Research Council of Canada (NSERC) under Grant RGPIN-2020-06832. a mode selection method for MUs, DUs, and FUs based on
Laleh Eslami and Ghasem Mirjalily are with the Department of Electrical the distance thresholds achieved considering target signal
Engineering, Yazd University, Yazd, Iran. (Email: to interference values. However, they did not investigate
laleh.eslami@stu.yazd.ac.ir; mirjalily@yad.ac.ir). Timothy N. Davidson channel allocation, nor did they discuss the possibility of
is with the Department of Electrical & Computer Engineering, McMaster
University, Hamilton, Canada (Email: davidson@mcmaster.ca). transmitting in the dedicated mode for the users. In
addition, their proposed scheme is fully centralized, which

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

Table I. Comparison of related work

Ref. Perspective Small cell Mode Cellular Reuse Dedicated Power Channel Full Frequency
(SE or EE) selection mode mode mode control allocation reuse
[10] SE × DU ✓ ✓ ✓ AU DU ✓
[11] SE × DU × ✓ ✓ DU DU ✓
[12] SE multiple DU/SU ✓ ✓ × DU/SU × ✓
[13] EE one AU ✓ × ✓ AU AU ×
[14] EE one DU/SU ✓ ✓ × DU/SU × ✓
[15] SE multiple DU × ✓ ✓ DU/SU DU/SU ×
[16] SE multiple DU/SU ✓ ✓ × DU/SU DU/SU ×
[17] SE/EE multiple DU × ✓ ✓ AU AU ×
[18] SE one DU ✓ ✓ ✓ DU DU ✓
[19] EE multiple DU × ✓ ✓ AU DU ✓
[20] EE multiple AU ✓ × ✓ AU AU ×
[21] EE dense DU ✓ ✓ × DU DU ×
This work SE dense DU/SU ✓ ✓ ✓ DU/SU DU/SU ✓
SE: spectral efficiency EE: energy efficiency DU: D2D users AU: all users SU: small-cell users ✓: included ×: not included

is impractical in dense HetNets because of the high executed for the users in all three tiers, channel allocation,
complexity and overhead imposed on the network. The and mode selection are only considered for DUs. In addition,
authors in [13] consider a HetNet consisting of a macro base the D2D transmission mode (in reuse and dedicated
station (MBS), one pico-base-station (PBS), and a number of manners) is the only transmission mode studied in that work
user devices to be associated with the MBS or PBS, or to and the cellular mode is not investigated. In [20], mode
construct a direct D2D link to minimize the power selection, resource assignment, and bandwidth allocation are
consumption. They overlook the opportunity for spectrum considered for the users in a HetNet using the Markov
reuse, and, instead, use orthogonal spectrum bands for the decision process and a deep reinforcement learning
MUs, DUs, and picocell users. This strategy reduces the algorithm. However, the reuse mode is ignored in that work.
spectral efficiency in a dense HetNet. In [14], a joint mode The authors in [21] seek to minimize the total power
selection, time allocation, and power control scheme is consumed by DUs in a 3-tier HetNet through the mode
proposed for users in a HetNet to minimize energy selection and spectrum allocation for the users using the
consumption. However, they investigate their problem only proximal algorithm. Nevertheless, they partition the
for two user devices. spectrum into two different sub-bands for the use of
Some works use fractional frequency reuse (FFR) macrocell users and small-cell users, which is inefficient in
architectures to better control the cross-tier and co-tier dense HetNets. For convenience, a comparison of the related
interference in the HetNet [15,16]. The authors in [15] work discussed above is summarized in Table I.
propose a new FFR architecture in a D2D-enabled HetNet
and present algorithms that perform mode selection, C. Motivation and Contribution
admission control, power control, and channel allocation for As discussed in the previous subsection and shown in
MUs, DUs, and FUs intending to maximize network sum Table I, although significant achievements have been made
rate. The authors in [16] use an FFR architecture and propose in previous works, there are some limitations that need to be
a joint channel allocation, mode selection, and power control overcome. In particular, considering the joint resource
scheme using the Lagrange dual decomposition method only allocation and mode selection for both the small cell users
for users located within the coverage area of femtocells. and DUs in a dense deployment has been missed in most of
Although FFR reduces the complexity of interference the previous studies. Furthermore, as the impacts of mode
management, this comes at the cost of reduced spectral selection, power control, and channel allocation on the
efficiency. system performance are coupled, ignoring one of them in the
As mentioned above, some works consider joint mode analysis may degrade the system performance. Also,
selection and resource allocation only for DUs in HetNets deployment of the full frequency reuse (spectrum sharing
[17- 21]. In [17], joint optimization of the energy efficiency with a frequency reuse factor of 1) can further improve the
and the spectral efficiency, including selection between the spectral efficiency while ensuring the minimum required
D2D-reuse mode and the D2D-dedicated mode, is rates of users, if proper mode selection and resource
investigated for the D2D communication in a HetNet. The allocation schemes are applied. This motivates a joint mode
authors solve the problem by decomposing it into several selection, power control, and channel allocation for both the
sub-problems. The work in [18] incorporates mode selection, DUs and FUs by considering all possible modes in dense
resource allocation, and power control for DUs in a HetNet. heterogeneous networks in which the users are allowed to
The dedicated mode is selected based on the distance reuse the spectrum.
between each D2D pair and the availability of orthogonal In this paper, we focus on the capacity-oriented version of
resources. An interference criterion is also applied to select that problem. In particular, we formulate the problem of
between the reuse and cellular modes for the DUs. However, maximizing the uplink rate for the DUs and FUs under their
the number of active users and femtocells is very limited in minimum rate requirements, and the maximum tolerable
that work. In [19], the authors use fuzzy C mean clustering, interference on the cellular system. To solve the resulting
sequential max search, Dinkelbach link, and Genetic non-linear non-convex mixed integer programming problem,
algorithms to solve the problem of joint mode selection and we present two sub-optimal algorithms, in which we perform
resource allocation. In that work, although power control is

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

mode selection and resource (power and channel) allocation


for DUs and FUs. FBS

The major contributions of this paper are summarized as FBS

follows: FU

• To the best of our knowledge, this is the first work MU


FBS

considering joint mode selection (with all possible modes),


MBS
power control, and channel allocation for both the DUs FBS

and FUs in a dense scenario. FBS

• Unlike most previous works that assign orthogonal PDU

frequency bands for femtocells and D2D users or for D2D Link Femtocell link Macrocell link
Femtocell coverage
area

neighboring femtocells, we consider the chance of a Fig. 1 Illustration of the three-tier network model
shared frequency band among the femtocells and D2D
users (full frequency reuse) to improve the spectral A. System Model
efficiency.
• Unlike the previous works, both the downlink and uplink We consider a single-cell HetNet that consists of macrocell
channel resources are considered for the channel users (MUs), K Femtocells (FCs), the kth of which serves Fk
allocation; For a D2D direct transmission, reusing is femtocell users (FUs), and 𝑁𝑑 pairs of users with the
scheduled on the uplink, while free channel resources from capability of both the D2D and cellular communications,
both the uplink and downlink are selected for the dedicated which are referred as potential D2D users (PDUs). All the
manner. users coexist within the same uplink spectral resource (full
• The formulated problem for joint mode selection and frequency reuse within the cell). Fig. 1 illustrates a typical
resource allocation is a non-linear non-convex mixed model of the network. It is assumed that FUs have been
integer problem that is difficult to solve. In this regard, we previously associated with the corresponding FCs. The term
use insight into the structure of the interference to (f, k) is used to represent FU f in FC k. We consider a dense
construct an approximate problem with a convex tight deployment of the FCs as it is one of the key features in next-
relaxation. Globally optimal solutions to the relaxed generation cellular networks. A D2D pair is referred by
approximate problem are efficiently obtained by subscripts d and d' for the transmitter and receiver,
developing a customized iterative algorithm based on the respectively. It is assumed that the total uplink frequency
Lagrange dual decomposition method. As the bandwidth contains N sub-channels (SCHs), where each
approximation is conservative, those solutions are SCH has the same bandwidth. In our system model, the MUs
guaranteed to be feasible for the original problem, and operate on separate SCHs (i.e., they operate orthogonally
since the approximation is mild, they are typically good with respect to the other MUs). Hence, without loss of
solutions to the original problem. generality, we will model MU n as taking SCH n, and hence
• The proposed iterative approach is a semi-distributed subscript n refers to either MU n or SCH n. An SCH n may
(hybrid) approach, and thus the computational burden and be shared between an MU, a PDU, and an FU in each FC.
signaling overhead to the MBS is reduced. We define 𝑁fr,ul and 𝑁fr,dl as the number of unused (free)
• To further reduce the computational complexity, a low- uplink and downlink SCHs, respectively, and use
complexity decomposition-based method is also proposed, superscripts M, D, and F to accordingly represent parameters
in which we select the transmission modes heuristically, related to the MUs, PDUs, and FUs.
and then sequentially perform admission control, power Depending on their QoS and rate requirements, and those
control, and resource allocation for the D2D and femtocell of other users, we will optimize the choice of the mode in
users. The proposed heuristic mode selection algorithm is which PDUs and the FUs transmit their data. For each PDU,
designed based on the traffic situation of the network, we will choose either the D2D-reuse mode, the D2D-
which is categorized into high-level, medium-level, and dedicated mode, or the cellular-dedicated mode. (These
low-level classes. modes are described below.) Since the distances between the
• It is proved from numerical simulation results that the FUs and their serving base stations is comparatively short,
proposed iterative scheme achieves a near-optimal for the FUs we will choose either the cellular-dedicated
performance. Also, it is demonstrated that the proposed mode, or cellular-reuse mode. In the D2D-reuse mode, the
heuristic mode selection and resource allocation scheme transmitter and receiver of a PDU pair create a direct
not only gains a proper performance, but also fits dense transmission link on an uplink SCH that is shared with an
HetNets because of much lower computational cost. MU. In the D2D-dedicated transmission mode, the PDU pair
The rest of the paper is organized as follows: In Section II, create a direct transmission link on a dedicated unused SCH
first, the system model is described; and then, the problem of (either downlink or uplink). When a user is in the cellular-
sum rate maximization for the D2D and femtocell users is dedicated transmission mode, it transmits data through the
formulated. The algorithms for solving the problem are MBS or femtocell base station (FBS) using one of the unused
presented in Sections III and IV. Evaluation results are uplink SCHs and one of the unused downlink SCHs. In the
provided in Section V, and finally, the paper is concluded in cellular-reuse mode, those two SCHs are shared with an MU.
Section VI. Since our goal will be to optimize the mode selections, SCH
II. SYSTEM MODEL AND PROBLEM allocations and uplink transmission powers, in the cellular
FORMULATION
modes we will assume that the MBS or FBS employs
sufficient power on the assigned downlink SCH to enable
reliable communication at the chosen rate. Therefore, when

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

Table II. Summary of notations transmission power of FU f in FC k on SCH n in mode m=1


Notation Description 𝐹
(cellular-reuse mode), 𝑔𝑓,𝑘,𝑑,𝑛 is the interfering channel gain
𝒲 Set of SCHs assigned to uplink transmission 𝐹
from co-channel FUs to the PDU d, and 𝑥𝑓,𝑘,𝑛,𝑚 ∈ {0,1} is
𝒲fr,ul (𝒲fr,dl) Set of free SCHs for the uplink (downlink)
ℳ𝒰 Set of MUs the SCH-mode allocation parameter for FUs, where
𝐹
𝒦 Set of FCs 𝑥𝑓,𝑘,𝑛,𝑚 = 1 means that FU f in FC k takes SCH n in mode
ℱ𝒰𝑘 Set of FUs in FC k 𝐹
m, and 𝑥𝑓,𝑘,𝑛,𝑚 = 0, otherwise. Hence, the total interference
ℱ𝒰 Set of all the FUs
𝒟𝒰 plus noise on DU d in mode 1 is:
Set of PDUs 𝐷 𝑀
𝒜𝑓,𝑘 Set of admissible SCHs for FU f in FC k 𝐼𝑑,𝑛,1 = 𝐵𝑛0 + 𝑃𝑛𝑀 𝑔𝑛,𝑑 ′

𝒜𝑑 Set of admissible SCHs for PDU d 𝐾 𝐹𝑘 (1)


𝑁 Number of SCHs in the bandwidth 𝒲 𝐹 𝐹 𝐹
+ ∑ ∑ 𝑥𝑓,𝑘,𝑛,1 𝑃𝑓,𝑘,𝑛,1 𝑔𝑓,𝑘,𝑑,𝑛
𝑁fr,ul (𝑁fr,dl ) Number of unused uplink (downlink) SCHs
𝑘=1 𝑓=1
𝑁𝑑 Number of PDU pairs 𝐷 𝐷
K Number of FCs and 𝐼𝑑,𝑛,2 = 𝐼𝑑,𝑛,3 = 𝐵𝑛0 , where the term 𝑛0 represents the
Fk Number of FUs served by femtocell k variance of thermal noise, which is modeled as being a zero-
𝑃𝑛𝑀 Transmission power of MU n on SCH n mean Gaussian random variable that is independent from
𝐷
𝑃𝑑,𝑛,𝑚 𝐹
(𝑃𝑓,𝑘,𝑛,𝑚 ) Transmission power of PDU d (FU f in FC k) on sample to sample.
SCH n in mode m Similarly, the maximum achievable uplink rate of FU f in
𝑚𝑖𝑛
𝑅𝑓,𝑘 (𝑅𝑑𝑚𝑖𝑛 ) Minimum rate requirement for PDU d (FU f in 𝐹
FC k on SCH n while being in mode m is given by 𝑅𝑓,𝑘,𝑛,𝑚 =
FC k) (bps)
𝐹 𝐹
Γ𝑑 (Γ𝑓,𝑘 ) SNR gap in DU d (FU (k, f)) 𝑃𝑓,𝑘,𝑛,𝑚 ℎ𝑓,𝑘,𝑛 𝐹
𝐵𝑙𝑜𝑔2 (1 + 𝐹 ), where ℎ𝑓,𝑘,𝑛 is the uplink channel
Γ𝑓,𝑘 𝐼𝑘,𝑛,𝑚
γ𝑀 Maximum tolerable interference on the MBS 𝐹
𝑡ℎ
for each SCH
gain of FU f in the FC k on SCH n, and 𝐼𝑘,𝑛,𝑚 is the
γ𝐹𝑡ℎ Maximum tolerable interference on the interference plus noise existing on SCH n in FC k while FU
femtocells for each SCH f is in mode m. Also, Γ𝑓,𝑘 is the SNR gap for FU (𝑘, 𝑓). When
γ𝐹𝐷
𝑡ℎ Maximum allowed interference caused by
FUs on each PDU
an FU transmits in mode m=1 (cellular-reuse mode), the total
γ𝐷𝐹
𝑡ℎ Maximum allowed interference caused by interference plus noise on the corresponding FC is:
𝑁𝑑
DUs on FCs for each SCH
𝐹 𝑀 𝐷 𝐷 𝐷
𝐷
𝑥𝑑,𝑛,𝑚 𝐹
(𝑥𝑓,𝑘,𝑛,𝑚 ) SCH-mode allocation parameter for PDU d 𝐼𝑘,𝑛,1 = 𝐵𝑛0 + 𝑃𝑛𝑀 𝑔𝑛,𝑘 + ∑ 𝑥𝑑,𝑛,1 𝑃𝑑,𝑛,1 𝑔𝑑,𝑘,𝑛
(FU f of FC k) on SCH n in mode m (2)
𝑑=1
𝑿𝑫 (𝑿𝑭 ) SCH-mode allocation matrix for PDUs (FUs) 𝐾 𝐹𝑘
𝑷𝑫 (𝑷𝑭 ) Power allocation matrix for PDUs (FUs)
𝐹 𝐹 𝐹
+ ∑ ∑ 𝑥𝑓,𝑘,𝑛,1 𝑃𝑓,𝑘,𝑛,1 𝑔𝑓,𝑘,𝑑,𝑛
formulating the downlink component of the cellular modes, 𝑘=1 𝑓=1
𝑀 𝐷
it is sufficient to focus on the allocation of the downlink where 𝑔𝑛,𝑘 and 𝑔𝑑,𝑘,𝑛 are the interfering channel gains from
SCHs. MU n and PDU d transmitting on SCH n to the FBS k,
𝐷
To begin the definition of the notations of this paper (which respectively. Also, 𝑥𝑑,𝑛,𝑚 ∈ {0,1} is the SCH-mode
𝐷 𝐷
are summarized in Table II for convenience), we let 𝑅𝑑,𝑛,𝑚 allocation parameter for PDUs, where 𝑥𝑑,𝑛,𝑚 = 1 means that
𝐷
denote the maximum achievable uplink rate of PDU d on PDU d takes SCH n in mode m, and 𝑥𝑑,𝑛,𝑚 = 0, otherwise.
SCH n in mode m, where m can be 1, 2 and 3, representing When an FU transmits in mode m=2 (cellular-dedicated
𝐹
the D2D-reuse mode, the D2D-dedicated mode, and the mode), 𝐼𝑘,𝑛,𝑚 = 𝐵𝑛0 .
𝐷
cellular-dedicated mode, respectively. Let ℎ𝑑,𝑛,𝑚 denote the
channel gain between the transmitter of PDU d and its B. Problem Formulation
receiver on SCH n in mode m. (The receiver is the MBS
when d is in mode 3, whereas it is another PDU, d', when d In this subsection, we formulate the problem of joint mode
is in mode 1 or 2). For receivers that treat interference as selection and resource allocation for the PDUs and FUs with
noise, the pragmatic rate limit for reliable communication the aim of maximizing their total uplink rate subject to the
from PDU d on SCH n in mode m is 𝑅𝑑,𝑛,𝑚 𝐷
= 𝐵𝑛 𝑙𝑜𝑔2 (1 + minimum rate requirements, the maximum tolerable
𝐷 𝐷 𝐷 interference on the cellular system, and the maximum
𝑃𝑑,𝑛,𝑚 ℎ𝑑,𝑛,𝑚 ⁄(Γ𝑑 𝐼𝑑,𝑛,𝑚 )), where 𝐵𝑛 is the bandwidth of
transmission powers of the users. As is evident from the
SCH n, which is assumed to be the same for all the SCHs above expressions, and is common in HetNet designs, our
𝐷
(𝐵𝑛 = 𝐵), 𝑃𝑑,𝑛,𝑚 is the transmission power of PDU d on SCH approach seeks to perform mode selection and resource
𝐷
n in mode m, 𝐼𝑑,𝑛,𝑚 is the interference plus noise on SCH n allocation in a “channel-adapted” manner, and hence we will
when PDU d is in mode m, and Γ𝑑 is the SNR gap, which assume that the node that performs the design (typically the
accounts for the practical modulation and coding schemes MBS) will have been provided with channel state
used by user d. There is no interference link when a PDU is information (CSI) for the direct and interfering channels
transmitting on a dedicated SCH (m=2,3). The interference (through control channels), and that it will communicate the
𝑀
caused by MU n when the PDU is in mode 1 is 𝑃𝑛𝑀 𝑔𝑛,𝑑′ , appropriate components of the solution to the other nodes.
𝑀
where 𝑃𝑛 indicates the transmission power of MU n The following inequalities state that each SCH can be
𝑀 assigned to at most one PDU and at most one FU in each FC,
transmitting on SCH n, and 𝑔𝑛,𝑑′ denotes the interfering
channel gain from MU n to the PDU receiver d'. The respectively:
𝑁𝑑 3
interference caused by the co-channel FUs is
𝐷
𝐹𝑘 𝐹 𝐹 𝐹 𝐹 ∑ ∑ 𝑥𝑑,𝑛,𝑚 ≤ 1, ∀𝑛, (C1)
∑𝐾𝑘=1 ∑𝑓=1 𝑥𝑓,𝑘,𝑛,1 𝑃𝑓,𝑘,𝑛,1 𝑔𝑓,𝑘,𝑑,𝑛 , where 𝑃𝑓,𝑘,𝑛,1 is the
𝑑=1 𝑚=1

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

𝑁𝑑
𝐹𝑘 2 𝐷 𝐷 𝐷
∑ 𝑥𝑑,𝑛,1 𝑃𝑑,𝑛,1 𝑔𝑑,𝑀𝐵𝑆,𝑛
𝐹
∑∑ 𝑥𝑓,𝑘,𝑛,𝑚 ≤ 1, ∀𝑛, ∀𝑘, (C2) 𝑑=1
𝐾 𝐹𝑘
𝑓=1 𝑚=1 (C11)
𝐹 𝐹 𝐹
+ ∑ ∑ 𝑥𝑓,𝑘,𝑛,1 𝑃𝑓,𝑘,𝑛,1 𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛 ≤ γ𝑀
𝑡ℎ , ∀𝑛
The above constraints are applied to control the
𝑘=1 𝑓=1
implementation complexity that results from increasing the
number of users sharing the same SCH simultaneously. In With these constraints established, the total weighted sum-
order to maintain fairness among the users, we assume that rate maximization problem can be formulated as:
each PDU or FU can take at most one SCH, which is 𝑁𝑑 3
reflected in the following constraints: 𝑁 ∑ ∑ 𝑤𝑑𝐷 𝑥𝑑,𝑛,𝑚
𝐷 𝐷
𝑅𝑑,𝑛,𝑚 (P1)
𝑁 3
𝑑=1 𝑚=1
𝐷
∑ ∑ 𝑥𝑑,𝑛,𝑚 ≤ 1, ∀𝑑, (C3) max ∑ 𝐾 𝐹𝑘 2
𝑿𝑫 ,𝑿𝑭 ,𝑷𝑫 ,𝑷𝑭
𝑛=1 𝑚=1 𝑛=1 𝐹 𝐹 𝐹
𝑁 2 + ∑ ∑ ∑ 𝑤𝑓,𝑘 𝑥𝑓,𝑘,𝑛,𝑚 𝑅𝑓,𝑘,𝑛,𝑚
𝐹
∑ ∑ 𝑥𝑓,𝑘,𝑛,𝑚 ≤ 1, ∀𝑘, ∀𝑓 ∈ ℱ𝒰𝑘 ,
( 𝑘=1 𝑓=1 𝑚=1 )
(C4) s.t.
𝑛=1 𝑚=1
Also, since a cellular transmission requires both the uplink (C1)-(C11)
𝐷
and downlink SCHs, the following constraint guarantees that 𝑥𝑑,𝑛,𝑚 ∈ {0,1} , ∀𝑑, ∀𝑛, ∀𝑚 ∈ {1,2,3}. (C12)
the number of PDUs and FUs in the cellular-dedicated mode 𝐹
𝑥𝑓,𝑘,𝑛,𝑚 ∈ {0,1}, ∀𝑘, ∀𝑓 ∈ ℱ𝒰𝑘 , ∀𝑛, ∀𝑚 ∈ {1,2}. (C13)
does not exceed the minimum number of free uplink and
downlink SCHs: where 𝑿𝑫 and 𝑿𝑭 are matrix variables whose elements are
𝐷 𝐹
𝑁 𝑁𝑑 𝐾 𝐹𝑘 𝑥𝑑,𝑛,𝑚 and 𝑥𝑓,𝑘,𝑛,𝑚 , respectively. Also, 𝑷𝑫 and 𝑷𝑭 are matrix
𝐷 𝐹 𝐷 𝐹
∑ (∑ 𝑥𝑑,𝑛,3 + ∑ ∑ 𝑥𝑓,𝑘,𝑛,2 ) ≤ min{𝑁fr,ul , 𝑁fr,dl } variables whose elements are 𝑃𝑑,𝑛,𝑚 and 𝑃𝑓,𝑘,𝑛,𝑚 ,
(C5)
𝑛=1 𝑑=1 𝑘=1 𝑓=1 𝐷 𝐹
respectively, and 𝑤𝑑 and 𝑤𝑓,𝑘 are the weights accordingly
As the number of free SCHs for dedicated use is limited, it assigned to the PDU d and the FU f in the FC k.
should be guaranteed that there are enough free SCHs to be Problem (P1) is a non-convex mixed-integer non-linear
taken by PDUs and FUs in the cellular-dedicated and by optimization problem, and hence it is inherently difficult to
PDUs in the D2D-dedicated modes: tackle directly. (Branch-and-Bound techniques can be
𝑁 𝑁𝑑 𝐾 𝐹𝑘 attempted for small instances of the problem, but since the
𝐷
∑ (∑(2𝑥𝑑,𝑛,3 𝐷
+ 𝑥𝑑,𝑛,2 𝐹
) + ∑ ∑ 2𝑥𝑓,𝑘,𝑛,2 ) problem is both non-convex and mixed-integer, in general
(C6) settings, they will impose an overwhelming computational
𝑛=1 𝑑=1 𝑘=1 𝑓=1
burden.) In order to generate good solutions in a tractable
≤ 𝑁fr,ul + 𝑁fr,dl
way, in the following section, we will construct a close,
In constraint (C6), the use of both uplink and downlink SCHs conservative approximation of the problem (P1) that can be
in the cellular modes has been considered. The constraints optimally solved using convex optimization techniques. In
related to the maximum transmission powers of PDUs and particular, we will develop a customized Lagrange dual
FUs are formulated as: decomposition method that uses the iterative sub-gradient
𝑁 3
𝐷 𝐷
algorithm; cf. [22]. To further reduce the computational cost,
𝐷
0 ≤ ∑ ∑ 𝑥𝑑,𝑛,𝑚 𝑃𝑑,𝑛,𝑚 ≤ 𝑃𝑚𝑎𝑥 , ∀𝑑, (C7) in Section IV, a heuristic solution is proposed for problem
𝑛=1 𝑚=1 (P1) based on the traffic situation of the network. The
𝑁 2
proposed heuristic solution decomposes the (primal)
𝐹 𝐹 𝐹 problem into four sub-problems: mode selection, admission
0 ≤ ∑ ∑ 𝑥𝑓,𝑘,𝑛,𝑚 𝑃𝑓,𝑘,𝑛,𝑚 ≤ 𝑃𝑚𝑎𝑥 , ∀𝑘, ∀𝑓 ∈ ℱ𝒰𝑘 , (C8) control, power control, and channel assignment. This
𝑛=1 𝑚=1
approach reduces the computational cost because it is
𝐷 𝐹
where 𝑃𝑚𝑎𝑥 and 𝑃𝑚𝑎𝑥 are the maximum transmission power composed of low-complexity algorithms for solving the
of PDUs and FUs, respectively. The minimum rate mentioned sub-problems.
requirement for the PDUs and FUs is imposed by constraints
(C9) and (C10): III. ITERATIVE SOLUTION
𝑁 3
𝐷
∑ ∑ 𝑥𝑑,𝑛,𝑚 𝐷
𝑅𝑑,𝑛,𝑚 ≥ 𝑅𝑑𝑚𝑖𝑛 , ∀𝑑, (C9) In this section, a computationally-efficient semi-
𝑛=1 𝑚=1 distributed solution strategy is developed for finding good
𝑁 2 solutions to the problem (P1). To begin with, it is observed
𝐹
∑ ∑ 𝑥𝑓,𝑘,𝑛,𝑚 𝐹
𝑅𝑓,𝑘,𝑛,𝑚 𝑚𝑖𝑛
≥ 𝑅𝑓,𝑘 , ∀𝑘, ∀𝑓 ∈ ℱ𝒰𝑘 , (C10) that the rates in mode 1 are not concave functions of the
𝑛=1 𝑚=1
powers, and the mode selection variables are binary-valued.
To tackle the non-concavity, we use bounds on the
In a heterogeneous cellular network, a serious challenge is
interference to construct a conservative approximation of the
to manage the interference perceived by the cellular network
problem (P1). Then, the binary variables in that approximate
from other users transmitting in the reuse mode, as it can
problem (problem (P2) below) are relaxed to lie in the
degrade the intended service. Therefore, the following
interval [0,1].
constraint is formulated to control the maximum tolerable
To construct the conservative approximation of problem
interference level from the PDUs and FUs transmitting in the
(P1), we first add constraints on the cross-tier interference
reuse mode to the MBS on each SCH:
from the PDUs to the FCs and the FUs to the PDUs. Also,

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

𝐾 𝐹𝑙
since we consider a dense deployment of the FCs, it is
essential to control the co-channel interference in the FC ∑ ∑ 𝑃̃𝑓,𝑙,𝑛,1
𝐹 𝐹
𝑔𝑓,𝑙,𝑘,𝑛 ≤ γ𝐹𝑡ℎ , ∀𝑘, ∀𝑛, ̃ 12)
(C
layer. In particular, we add the constraints: 𝑙=1 𝑓=1
𝐾 𝐹𝑘 𝑙≠𝑘
𝐾 𝐹𝑘
𝐹 𝐹 𝐹
∑ ∑ 𝑥𝑓,𝑘,𝑛,1 𝑃𝑓,𝑘,𝑛,1 𝑔𝑓,𝑘,𝑑,𝑛 ≤ γ𝐹𝐷
𝑡ℎ , ∀𝑑, ∀𝑛 (3)
∑ ∑ 𝑃̃𝑓,𝑘,𝑛,1
𝐹 𝐹
𝑔𝑓,𝑘,𝑑,𝑛 ≤ γ𝐹𝐷 ̃ 13)
(C
𝑡ℎ , ∀𝑑, ∀𝑛,
𝑘=1 𝑓=1
𝑘=1 𝑓=1
𝑁𝑑 𝑁𝑑
𝐷
∑ 𝑥𝑑,𝑛,1 𝐷
𝑃𝑑,𝑛,1 𝐷
𝑔𝑑,𝑘,𝑛 ≤ γ𝐷𝐹 (4) ̃ 14)
𝑡ℎ , ∀𝑘, ∀𝑛 ∑ 𝑃̃𝑑,𝑛,1
𝐷 𝐷
𝑔𝑑,𝑘,𝑛 ≤ γ𝐷𝐹
𝑡ℎ , ∀𝑘, ∀𝑛, (C
𝑑=1 𝑑=1
𝐾 𝐹𝑙 (C12) and (C13)
𝐹 𝐹
∑ ∑ 𝑥𝑓,𝑙,𝑛,1 𝑃𝐹𝑓,𝑙,𝑛,1 𝑔𝑓,𝑙,𝑘,𝑛 ≤ γ𝐹𝑡ℎ , ∀𝑘, ∀𝑛 (5)
As 𝑷̃ 𝑫 and 𝑷
̃ 𝑭 lie in convex sets and the Hessian matrix of
𝑙=1 𝑓=1
𝑙≠𝑘 the objective function with respect to them is a negative
where γ𝐹𝐷
𝑡ℎ is the maximum allowed interference caused by semi-definite matrix, the objective function is concave [22].
FUs on each PDU, and γ𝐷𝐹 𝑡ℎ is the maximum allowed
In addition, all the constraints, with the exception of (C12)
interference caused by the PDUs on each FC. The maximum and (C13), are convex as well. To obtain the convex
total interference level from other co-channel FCs on an FC relaxation of (P2), we simply relax (C12) and (C13) so that
is denoted by γ𝐹𝑡ℎ . the mode decision variables lie in the interval [0,1] rather
With these constraints in place, we can bound the than in the set {0,1}. The resulting problem is convex, and
interference plus noise terms that appear in equations (1) and hence a globally optimal solution (to the relaxed version of
(2). In particular, we use the following bounds on the the problem (P2)) can be efficiently obtained. Given the
interference terms that arise in mode 1: 𝐼𝑑,𝑛,1 𝐷
≤ 𝐼̃𝑑,𝑛,1
𝐷
= nature of the constraints in (P2), whenever the problem is
𝐵𝑛0 + 𝑃𝑛𝑀 𝑔𝑛,𝑑𝑀
+ γ𝐹𝐷 , and 𝐼 𝐹
≤ ̃
𝐼 𝐹
= 𝐵𝑛 + feasible, there is always a point in the relative interior of the
𝑡ℎ 𝑘,𝑛,1 𝑘,𝑛,1 0
𝑀 feasible set, and hence Slater’s constraint qualification
𝑃𝑛𝑀 𝑔𝑛,𝑘 + γ𝐷𝐹 𝐹
𝑡ℎ + γ𝑡ℎ . For notational convenience, for
reveals that the problem has zero duality gap [22]. Therefore,
m=2,3, we define 𝐼̃𝑑,𝑛,𝑚 𝐷 𝐷
= 𝐼𝑑,𝑛,𝑚 and 𝐼̃𝑘,𝑛,2
𝐹 𝐹
= 𝐼𝑘,𝑛,2 . Now, if we can use the Lagrange dual decomposition method to solve
we define the actual transmitted powers 𝑃̃𝑑,𝑛,𝑚 𝐷
= (P2) [23]. The Lagrange function of (P2) can be written as:
𝐷 𝐷 ̃ 𝐹 𝐹 𝐹
𝑥𝑑,𝑛,𝑚 𝑃𝑑,𝑛,𝑚 and 𝑃𝑓,𝑘,𝑛,𝑚 = 𝑥𝑓,𝑘,𝑛,𝑚 𝑃𝑓,𝑘,𝑛,𝑚 , we obtain the ̃𝑫, 𝑷
𝐿(𝑿𝑫 , 𝑿𝑭 , 𝑷 ̃ 𝑭 , 𝚲)
following lower bounds on the rates: 𝑅𝑑𝐷,𝑛,𝑚 ≥ 𝑅̃𝑑𝐷,𝑛,𝑚 = 𝑁 𝑫 ̃𝑫 𝐹𝐷
= ∑𝑁 𝑑
𝑛=1 ∑𝑑=1(𝐿𝑑,𝑛 (𝑿 , 𝑷 , 𝚲) + 𝜆13 (𝑑, 𝑛)γ𝑡ℎ ) +
𝐵 log 2 (1 + 𝑃̃𝑑𝐷,𝑛,𝑚 ℎ𝑑𝐷,𝑛,𝑚 ⁄Γ𝑑 𝑥𝑑𝐷,𝑛,𝑚 𝐼̃𝑑𝐷,𝑛,𝑚 ), and 𝑅𝐹𝑓,𝑘,𝑛,𝑚 ≥
𝐾 𝐹 𝑭 ̃𝑭
̃𝐹𝑓,𝑘,𝑛,𝑚 = 𝐵log (1 + 𝑃
𝑅 ̃𝐹𝑓,𝑘,𝑛,𝑚 ℎ𝐹𝑓,𝑘,𝑛 ⁄Γ𝑓,𝑘 𝑥𝐹 𝐹
𝐼̃𝑘,𝑛,𝑚 ). These ∑𝑁 𝑘
𝑛=1 ∑𝑘=1 (∑𝑓=1 𝐿𝑘,𝑓,𝑛 (𝑿 , 𝑷 , 𝚲) −
2 𝑓 , 𝑘 , 𝑛, 𝑚
lower bounds on the rates are concave functions of 𝑃̃𝑑,𝑛,𝑚
𝐷
𝜆12 (𝑘, 𝑛) ∑𝐾𝑙=1 ∑𝑓=1
𝑙 𝐹 𝐹
𝑃̃𝑓,𝑙,𝑛,1 𝐹
𝑔𝑓,𝑙,𝑘,𝑛 ) + Ψ. (6)
and 𝑃̃𝑓,𝑘,𝑛,𝑚 , respectively. Applying these bounds, we obtain
𝐹 𝑙≠𝑘
where 𝚲 represents the set of Lagrange multiplier variables,
the following conservative approximation of the problem
{𝜆1 , ⋯ , 𝜆14 }, accordingly related to the constraints of the
(P1): 𝐾
problem (P2). Also, Ψ = ∑𝑁 𝑛=1(𝜆1 (𝑛) + ∑𝑘=1 𝜆2 (𝑛, 𝑘) +
𝑀 𝑁𝑑 𝐷
𝑁𝑑 3 𝜆11 (𝑛) × γ𝑡ℎ ) + ∑𝑑=1(𝜆3 (𝑑) + 𝜆7 (𝑑)𝑃𝑚𝑎𝑥 −
𝐹𝑘
𝑁 ∑ ∑ 𝑤𝑑𝐷 𝑥𝑑𝐷,𝑛,𝑚 𝑅̃𝑑𝐷,𝑛,𝑚 𝜆9 (𝑑)𝑅𝑑𝑚𝑖𝑛 ) + ∑𝐾 𝐹
𝑘=1 ∑𝑓=1(𝜆4 (𝑘, 𝑓) + 𝜆8 (𝑘, 𝑓)𝑃𝑚𝑎𝑥 −
𝑚𝑖𝑛
max ∑ 𝐾
𝑑=1 𝑚=1
𝐹𝑘 2 (P2) 𝜆10 (𝑘, 𝑓)𝑅𝑓,𝑘 ) + ∑𝐾 𝑁 𝐹
𝑘=1 ∑𝑛=1(𝜆12 (𝑘, 𝑛)γ𝑡ℎ +
̃ ,𝑷
𝑫 𝑭 𝑫
𝑿 ,𝑿 ,𝑷 ̃ 𝑭
𝐷𝐹
𝑛=1
+∑∑ ∑ 𝑤𝑓𝐹,𝑘 𝑥𝑓𝐹,𝑘 ,𝑛,𝑚 𝑅̃𝑓𝐹,𝑘 ,𝑛,𝑚
𝜆14 (𝑘, 𝑛)γ𝑡ℎ ) + 𝜆5 𝑁min + 𝜆6 (𝑁fr,ul + 𝑁fr,dl ), 𝑁min =
( 𝑘=1 𝑓=1 𝑚=1 ) min{𝑁fr,ul , 𝑁fr,dl }, and:
s.t. ̃ 𝑫 , 𝚲)
𝐿𝑑,𝑛 (𝑿𝑫 , 𝑷
(C1)-(C6) and 𝐷 𝐷 𝐷
𝑁 3
= 𝜆5 𝑥𝑑,𝑛,3 − 𝜆6 (𝑥𝑑,𝑛,2 + 2𝑥𝑑,𝑛,3 )
− 𝜆11 (𝑛)𝑃̃𝑑,𝑛,1
𝐷 𝐷
𝑔𝑑,𝑀𝐵𝑆,𝑛
0 ≤ ∑ ∑ 𝑃̃𝑑,𝑛,𝑚
𝐷 𝐷
≤ 𝑃𝑚𝑎𝑥 , ∀𝑑, ̃ 7)
(C 3
𝑛=1 𝑚=1
+ ∑ (𝑤𝑑𝐷 𝑥𝑑,𝑛,𝑚
𝐷
𝑅̃𝑑,𝑛,𝑚
𝐷
𝑁 2 (7)
𝑚=1
0 ≤ ∑ ∑ 𝑃̃𝑓,𝑘,𝑛,𝑚
𝐹 𝐹
≤ 𝑃𝑚𝑎𝑥 , ∀𝑘, ∀𝑓 ∈ ℱ𝒰𝑘 , ̃ 8)
(C − (𝜆1 (𝑛) + 𝐷
𝜆3 (𝑑))𝑥𝑑,𝑛,𝑚
𝑛=1 𝑚=1 − 𝜆7 (𝑑)𝑃̃𝑑,𝑛,𝑚
𝐷
𝑁 3
+ 𝜆9 (𝑑)𝑥𝑑,𝑛,𝑚 𝑅̃𝑑,𝑛,𝑚
𝐷 𝐷
)
𝐷
∑ ∑ 𝑥𝑑,𝑛,𝑚 𝑅̃𝑑,𝑛,𝑚
𝐷
≥ 𝑅𝑑𝑚𝑖𝑛 , ∀𝑑, ̃ 9)
(C 𝐾
𝑛=1 𝑚=1
𝑁 2
− 𝑃̃𝑑,𝑛,1
𝐷 𝐷
∑ 𝜆14 (𝑘, 𝑛)𝑔𝑑,𝑘,𝑛
𝑘=1
𝐹
∑ ∑ 𝑥𝑓,𝑘,𝑛,𝑚 𝑅̃𝑓,𝑘,𝑛,𝑚
𝐹 𝑚𝑖𝑛
≥ 𝑅𝑓,𝑘 , ∀𝑘, ∀𝑓 ∈ ℱ𝒰𝑘 , ̃ 10)
(C
𝑛=1 𝑚=1
𝑁𝑑 𝐾 𝐹𝑘

∑ 𝑃̃𝑑,𝑛,1
𝐷 𝐷
𝑔𝑑,𝑀𝐵𝑆,𝑛 + ∑ ∑ 𝑃̃𝑓,𝑘,𝑛,1
𝐹 𝐹
𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛 ≤ γ𝑀
𝑡ℎ , ∀𝑛
̃ 11)
(C
𝑑=1 𝑘=1 𝑓=1

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

+
(𝑤𝑑𝐷+𝜆9 (𝑑))𝐵 𝐷
Γ𝑑 𝐼̃𝑑,𝑛,𝑚
[ − 𝐷 ] , 𝑚≠1
𝜆7 (𝑑) ln 2 ℎ𝑑,𝑛,𝑚
𝑃̂𝑑.𝑛.𝑚
𝐷
= + (10)
(𝑤𝑑𝐷 +𝜆9 (𝑑))𝐵 𝐷
Γ𝑑 𝐼̃𝑑,𝑛,𝑚
[ 𝐷 − ] , 𝑚=1
{ (𝜆7 (𝑑)+𝜆11 (𝑛)𝑔𝑑,𝑀𝐵𝑆,𝑛 −∑𝐾 𝐷
𝑘=1 𝜆14 (𝑘,𝑛)𝑔𝑑,𝑘,𝑛 ) ln 2
𝐷
ℎ𝑑,𝑛,𝑚

+
𝐹 𝐹
(𝑤𝑓,𝑘 +𝜆10 (𝑘,𝑓))𝐵 Γ𝑓,𝑘 𝐼̃𝑘,𝑛,𝑚
[ − 𝐹 ] , 𝑚=1
𝐹 𝐹 𝑁 ℎ𝑓,𝑘,𝑛
(𝜆8 (𝑘,𝑓)+𝜆11 (𝑛)𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛 +∑𝐾 𝑙 𝐹 𝑑 𝐹
𝑙=1 ∑𝑓=1 𝜆12 (𝑘,𝑛)𝑔𝑓,𝑙,𝑘,𝑛 −∑𝑑=1 𝜆13 (𝑑,𝑛)𝑔𝑓,𝑘,𝑑,𝑛 ) ln 2
𝑃̂𝑓.𝑘.𝑛.𝑚
𝐹
= 𝑙≠𝑘 (11)
𝐹 +
(𝑤𝑓,𝑘 +𝜆10 (𝑘,𝑓))𝐵 𝐹
Γ𝑓,𝑘 𝐼̃𝑘,𝑛,𝑚
[ − 𝐹 ] , 𝑚=2
{ 𝜆8 (𝑘,𝑓)ln 2 ℎ𝑓,𝑘,𝑛

̃ 𝑭 , 𝚲)
𝐿𝑘,𝑓,𝑛 (𝑿𝑭 , 𝑷 𝐷
Here, 𝐽𝑑,𝑛,𝑚 = 𝐵(𝑤𝑑𝐷 + 𝜆9 (𝑑)) log 2 (1 +
𝐹
= 𝜆5 𝑥𝑓,𝑘,𝑛,2 𝐹
− 2𝜆6 𝑥𝑓,𝑘,𝑛,2 𝑃̂𝑑,𝑛,𝑚
𝐷 𝐷
ℎ𝑑,𝑛,𝑚 ⁄Γ𝑑 𝐼̃𝑑,𝑛,𝑚
𝐷
) − 𝜆1 (𝑛) − 𝜆3 (𝑛) − 𝜆7 (𝑑)𝑃̂𝑑,𝑛,𝑚
𝐷
.
− 𝜆11 (𝑛)𝑃̃𝑓,𝑘,𝑛,1
𝐹 𝐹
𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛 𝐹 𝐹
Also, by defining 𝐽𝑓,𝑘,𝑛,𝑚 = 𝐵 (𝑤𝑓,𝑘 + 𝜆10 (𝑘, 𝑓)) log 2 (1 +
𝑁𝑑
𝑃̂𝑓,𝑘,𝑛,𝑚
𝐹 𝐹
ℎ𝑓,𝑘,𝑛 ⁄Γ𝑓,𝑘 𝐼̃𝑘,𝑛,𝑚
𝐹
) − 𝜆2 (𝑛, 𝑘) − 𝜆4 (𝑘, 𝑓) −
− 𝑃̃𝑓,𝑘,𝑛,1
𝐹 𝐹
∑ 𝜆13 (𝑑, 𝑛)𝑔𝑓,𝑘,𝑑,𝑛
(8) 𝜆8 (𝑘, 𝑓)𝑃̂𝑓,𝑘,𝑛,𝑚 , we get:
𝐹
𝑑=1
2 𝐹
𝜆11 (𝑛)𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛
𝐹
+ ∑ (𝑤𝑓,𝑘 𝐹
𝑥𝑓,𝑘,𝑛,𝑚 𝑅̃𝑓,𝑘,𝑛,𝑚
𝐹
𝑁𝑑
𝑚=1
𝐹
𝐽𝑓,𝑘,𝑛,𝑚 − 𝑃̂𝑓,𝑘,𝑛,𝑚
𝐹
𝐹
− (𝜆2 (𝑛, 𝑘) + 𝐹
𝜆4 (𝑘, 𝑓))𝑥𝑓,𝑘,𝑛,𝑚 + ∑ 𝜆13 (𝑑, 𝑛)𝑔𝑓,𝑘,𝑑,𝑛
( 𝑑=1 )
− 𝜆8 (𝑘, 𝑓)𝑃̃𝑓,𝑘,𝑛,𝑚
𝐹
𝐹
𝐺𝑘,𝑓,𝑛,𝑚 = 𝐾 𝐹𝑙
𝐹
+ 𝜆10 (𝑘, 𝑓)𝑥𝑓,𝑘,𝑛,𝑚 𝑅̃𝑓,𝑘,𝑛,𝑚
𝐹
) − ∑ ∑ 𝜆12 (𝑘, 𝑛)𝑃̂𝑓,𝑘,𝑛,𝑚
𝐹 𝐹
𝑔𝑓,𝑙,𝑘,𝑛 , 𝑚=1
𝑙=1 𝑓=1
𝑙≠𝑘
Then, the dual optimization problem for (P2) can be
𝐹
expressed as [23]: { 𝐽𝑓,𝑘,𝑛,𝑚 − 𝜆5 − 2𝜆6 , 𝑚=2
(14)
min max ̃𝑫, 𝑷
𝐿(𝑿𝑫 , 𝑿𝑭 , 𝑷 ̃ 𝑭 , 𝚲) (P3)
𝚲 𝑿𝑫 ,𝑷
̃ 𝑫 ,𝑿𝑭 ,𝑷
̃𝑭 This observation generates a natural approach for
rounding the relaxed SCH-mode allocation parameters
Problem (P3) consists of two sub-problems: maximizing 𝐷 𝐹 𝐷
𝑥𝑑,𝑛,𝑚 and 𝑥𝑓,𝑘,𝑛,𝑚 to zero or one. In particular, as 𝑥𝑑,𝑛,𝑚 and
the Lagrange function to achieve the optimal values for 𝐹 𝐷
̃ 𝑫 , and 𝑷
𝑿𝑫 , 𝑿𝑭 , 𝑷 ̃ 𝑭 , and then minimizing it to find the 𝑥𝑓,𝑘,𝑛,𝑚 are in the interval [0,1], when 𝐺𝑑,𝑛,𝑚 < 0 (or
𝐹 𝐷
optimal Lagrange multipliers. According to the first order 𝐺𝑘,𝑓,𝑛,𝑚 < 0), the maximum value is achieved via 𝑥𝑑,𝑛,𝑚 =
𝐹 𝐷 𝐹
KKT necessary conditions for optimality [22]: 0 (or 𝑥𝑓,𝑘,𝑛,𝑚 = 0). When 𝐺𝑑,𝑛,𝑚 > 0 (or 𝐺𝑘,𝑓,𝑛,𝑚 > 0), the
𝐷
𝜕𝐿(𝑿𝑫 ,𝑿𝑭 ,𝑷
̃ 𝑫 ,𝑷
̃ 𝑭 ,𝚲) 𝜕𝐿(𝑿𝑫 ,𝑿𝑭 ,𝑷
̃ 𝑫 ,𝑷
̃ 𝑭 ,𝚲) (9) maximum value can be achieved via 𝑥𝑑,𝑛,𝑚 = 1 (or
𝐷 = 0 and =0 𝐹
𝑥𝑓,𝑘,𝑛,𝑚 = 1). Hence:
𝜕𝑃̃𝑑,𝑛,𝑚 𝜕𝑃̃𝐹 𝑓,𝑘,𝑛,𝑚
𝐹 1, (𝑑̂, 𝑚 𝐷
̂) = argmax 𝐺𝑑,𝑛,𝑚
Therefore, 𝑃̃𝑑,𝑛,𝑚
𝐷 ̃𝑓,𝑘,𝑛,𝑚 can be obtained directly from
and 𝑃 𝑥̂𝑑𝐷̂ ,𝑛,𝑚̂ = { 𝑑,𝑚 ∀𝑛 (15)
the equations in (9). When 𝑥𝑑,𝑛,𝑚 𝐷 𝐹
and 𝑥𝑓,𝑘,𝑛,𝑚 are positive, 0, otherwise
𝐷
then optimal powers can be found from 𝑃𝑑,𝑛,𝑚 = 1, (𝑓̂, 𝑚 𝐹
̂) = argmax 𝐺𝑓,𝑘,𝑛,𝑚
𝐹
𝑃̃𝑑,𝑛,𝑚 ⁄𝑥𝑑,𝑛,𝑚 and 𝑃𝑓,𝑘,𝑛,𝑚 = 𝑃̃𝑓,𝑘,𝑛,𝑚 ⁄𝑥𝑓,𝑘,𝑛,𝑚 , and when
𝐷 𝐷 𝐹 𝐹 𝐹 𝑥̂𝑘,𝑓 ̂ ={
̂ ,𝑛,𝑚 𝑓,𝑚 ∀𝑛, ∀𝑘 (16)
𝐷 𝐹
0, otherwise
𝑥𝑑,𝑛,𝑚 and 𝑥𝑓,𝑘,𝑛,𝑚 are zero, optimal powers are inherently
zero. Thereby, the optimum non-zero values for the In other words, SCH n is assigned to the PDU 𝑑̂ in
transmission power are achieved by equations (10) and (11) 𝐷
transmission mode 𝑚 ̂, if it maximizes 𝐺𝑑,𝑛,𝑚 . Also, in each
in which [𝑥]+ = max {0, 𝑥}. Now, having the values of the FC k, SCH n is assigned to the FU 𝑓̂ in transmission mode
matrices 𝑷 ̂ 𝐷 and 𝑷 ̂ 𝐹 and 𝚲, the Lagrange function can be 𝑚 𝐹
̂, if it maximizes 𝐺𝑓,𝑘,𝑛,𝑚 .
written as:
𝐷 𝐹 𝐷 𝐷 𝐹 𝐹 Now that the interim solutions for 𝑷𝑫 , 𝑷𝑭 , 𝑿𝑫 , and 𝑿𝑭
𝐿(𝑥𝑑,𝑛,𝑚 , 𝑥𝑓,𝑘,𝑛,𝑚 ) = 𝐺𝑑,𝑛,𝑚 𝑥𝑑,𝑛,𝑚 + 𝐺𝑘,𝑓,𝑛,𝑚 𝑥𝑓,𝑘,𝑛,𝑚 +𝐶 (12)
have been found for the given values of the Lagrange
where C is a constant value, and multipliers, those multipliers 𝜆1 , ⋯ 𝜆14 in the minimization
𝐷
problem are updated using the sub-gradient method [24]:
𝜆11 (𝑛)𝑔𝑑,𝑀𝐵𝑆,𝑛
𝐾 𝜆𝑡+1 𝑡 𝑡 𝑡 +
𝑖 (𝑛) = [𝜆𝑖 − 𝜀𝑖 𝑓𝑖 ] , 𝑖 = 1,2, ⋯ ,14. (17)
𝐷
𝐽𝑑,𝑛,𝑚 − 𝑃̂𝑑,𝑛,𝑚
𝐷
( 𝐷
),𝑚 = 1
𝐷 + ∑ 𝜆14 (𝑘, 𝑛)𝑔𝑑,𝑘,𝑛 where t denotes the iteration index, and is the subgradient 𝑓𝑖𝑡
𝐺𝑑,𝑛,𝑚 =
𝑘=1
𝐷
of the Lagrange function at
𝐽𝑑,𝑛,𝑚 − 𝜆6 , 𝑚=2 𝑡 𝑡 𝑡 𝐹 𝑡
𝐷
𝐷
(𝑥𝐷𝑑,𝑛,𝑚 ) , (𝑥𝐹𝑓,𝑘,𝑛,𝑚 ) , (𝑃𝑑,𝑛,𝑚 ) , (𝑃𝑓,𝑘,𝑛,𝑚 ) , and 𝜆𝑡𝑖 . The
{ 𝐽𝑑,𝑛,𝑚 − 𝜆5 − 2𝜆6 , 𝑚=3
(13) positive step sizes 𝜀𝑖𝑡 are chosen such that [23]:

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306


decomposition technique may take many iterations to
∑ 𝜀𝑖𝑡 = ∞ , lim 𝜀𝑖𝑡 = 0 , ∀𝑖 ∈ {1,2, ⋯ ,14}. (18) converge. In this section, a more efficient, heuristic solution
𝑡→∞
𝑡=1 strategy is proposed. We will show that it often provides
According to the diminishing step size rule, if the sequence solutions with similar performance to the iterative algorithm
𝜀𝑖𝑡 converges to zero and is non-summable, updates in the
developed in the previous section, and that it does so at a
subgradient method converge to the minimum [25].
much lower cost. In the proposed heuristic method, the
The algorithm for the iterative solution of problem (P1) is
summarized in Algorithm 1. Given that this algorithm is transmission modes of the FUs and PDUs are selected based
developed with the goal of real-time implementation, the on a heuristic mode selection criterion, and then, admission
iterations continue until the convergence condition is met or control, power assignment, and SCH allocation are
the maximum number of iterations 𝑇𝑚𝑎𝑥 is reached. performed for them. The proposed heuristic scheme for
To discuss the optimality of the Lagrange dual selecting the transmission mode is based on the network
decomposition method, consider the dual problem (P3). The traffic situation, which is classified into high, medium, and
Lagrange function is a concave function with respect to low levels according to the total number of FUs and PDUs
̃ 𝑫 and 𝑷
𝑷 ̃ 𝑭 . Therefore, its stationary point gives the optimal and the number of free SCHs in the network. These situations
values for transmission powers. Furthermore, the optimal are discussed in the following sections.
binary values for 𝑿𝑫 and 𝑿𝑭 are achieved by the proposed
natural approach for rounding the relaxed 𝑿𝑫 and 𝑿𝑭 . On the A. High-level traffic
other hand, the subgradient method generally gives a lower
When there is no free SCH in the network, the traffic level
bound to the problem (P3). In our proposed algorithm,
however, it will converge to the minimum because the step is defined as high. Indeed, if the existing MUs in the
size rule holds. Therefore, the Lagrange dual decomposition macrocell occupy all the SCHs, i.e., 𝑁fr,ul = 0, the network
method gives the optimal solutions to the approximate is in a high-level traffic situation. In this case, PDUs and FUs
problem. can only transmit in the reuse mode i.e., m=1. Algorithm 2
As seen in this section, the convergent Lagrange dual illustrates the procedure for efficiently generating good
method for solving the relaxed version of the problem (P2) solutions to the problem (P1) in the high-level traffic
finds optimal solutions, in which all the mode decision situation. To this end, it decomposes the problem into the
variables are either 0 or 1. Hence, it is implicitly shown that admission control, SCH allocation, and power control
the relaxation of the problem (P2) is tight. Indeed, we have subproblems. Let us assume that PDU d and FU f in FC k and
constructed efficient, globally optimal solutions for the non- FCs 𝑘′ = 1: 𝑘1 reuse SCH n. Also, we assume that the
convex approximated problem (P2). As the approximation is
transmission power values of FU f in FCs 𝑘′ = 1: 𝑘1 are
conservative, those solutions are guaranteed to be feasible
for the original problem (P1), and since the approximation is previously determined. In this case, considering that they are
mild, they are typically good solutions to the problem (P1). in mode m=1, the problem of maximizing their achievable
rates is written as follows:
𝐷 𝐷
𝑃𝑑,𝑛,1 ℎ𝑑,𝑛,1
Algorithm 1. Iterative solution of problem (P1) 𝑤𝑑𝐷 𝐵log 2 (1 + )
Γ𝑑 𝐼̃𝑑,𝑛,1
𝐷
1: t=0,
2: Initialize 𝑇𝑚𝑎𝑥 , Lagrangian multipliers and step sizes 𝜀i0 = 𝜏𝑖
max 𝐹 𝐹
𝑘1
𝐷
𝑃𝑑,𝑛,1 𝐹
,𝑃𝑓,𝑘,𝑛,1
𝐹
𝑃𝑓,𝑘,𝑛,1 ℎ𝑓,𝑘,𝑛 𝐹
3: Repeat +𝑤𝑓,𝑘 𝐵log 2 (1 + ) + ∑ 𝑤𝑓,𝑘 ′ 𝑅𝑓,𝑘 ′ ,𝑛
̃ 𝐹
Γ𝑓,𝑘 𝐼𝑘,𝑛,1
4: Calculate 𝑃̂𝑑,𝑛,𝑚
𝐷
∀𝑑 ∈ 𝒰, ∀𝑛 ∈ 𝒲, 𝑚 ∈ {1,2,3} from (10); ( ′
𝑘 =1 )
5: Calculate 𝑃̂𝑓,𝑘,𝑛,𝑚
𝐹
∀𝑘 ∈ 𝒦, ∀𝑓 ∈ ℱ𝒰𝑘 , ∀𝑛 ∈ 𝒲 , 𝑚 ∈ {1,2} (P4)
from (11); s.t.
6: for 𝑛 = 1: 𝑁 𝐷 𝐷 ̂ 1)
0 ≤ 𝑃𝑑,𝑛,1 ≤ 𝑃𝑚𝑎𝑥 , (C
7: MBS updates (𝑑̂ , 𝑚 ̂ ) from (15), and assigns the values
𝐷
𝑥̂𝑑,𝑛,𝑚 ̂ 𝐷.
for matrix 𝑿
𝐹 𝐹
0 ≤ 𝑃𝑓,𝑘,𝑛,1 ≤ 𝑃𝑚𝑎𝑥 , ̂ 2)
(C
8: for 𝑘 = 1: 𝐾 𝐷
𝑅𝑑,𝑛,1 ≥ 𝑅𝑑𝑚𝑖𝑛 , ̂ 3)
(C
9: FBS updates (𝑓̂ , 𝑚̂ ) from (16), and assigns the values 𝐹 𝑚𝑖𝑛 ̂ 4)
𝐹 ̂𝐹.
𝑅𝑓,𝑘,𝑛.1 ≥ 𝑅𝑓,𝑘 , (C
𝑥̂𝑓,𝑘,𝑛,𝑚 for matrix 𝑿
𝐷 𝐷 𝐹 𝐹
10: end for 𝑃𝑑,𝑛,1 𝑔𝑑,𝑀𝐵𝑆,𝑛 + 𝑃𝑓,𝑘,𝑛,1 𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛
11: end for 1𝑘 𝐹 𝐹 𝑀 ̂ 5)
𝜏 + ∑𝑘′=1 𝑃𝑓,𝑘 ′ ,𝑛,1 𝑔𝑓,𝑘 ′ ,𝑀𝐵𝑆,𝑛 ≤ γ𝑡ℎ , (C
12: Set 𝜀i𝑡+1 = 𝑖 ,
√𝑡+1
13: MBS updates multipliers 𝜆1 , 𝜆5 , 𝜆6 , and 𝜆11,
𝐷
𝑃𝑑,𝑛,1 𝐷
𝑔𝑑,𝑘,𝑛 ≤ γ𝐷𝐹
𝑡ℎ ,
̂ 6)
(C
14: FBS updates 𝜆2 , 𝜆12 , and 𝜆14, 𝐹 𝐹 𝑘1 𝐹 𝐹 ̂ 7)
𝑃𝑓,𝑘,𝑛,1 𝑔𝑓,𝑘,𝑑,𝑛 + ∑𝑘′=1 𝑃𝑓,𝑘 ′ ,𝑛,1 𝑔𝑓,𝑘 ′ ,𝑑,𝑛 ≤ γ𝐹𝐷
𝑡ℎ , (C
15: PDUs update 𝜆3 , 𝜆7 , 𝜆9 , and 𝜆13,
16: FUs update 𝜆4 , 𝜆8 , and 𝜆10 , ∑𝑘𝑙=1
1 𝐹
𝑃𝑓,𝑙,𝑛,1 𝐹
𝑔𝑓,𝑙,𝑘,𝑛 ≤ γ𝐹𝑡ℎ . ̂ 8)
(C
17: Set t=t+1. 𝑙≠𝑘
18: until Convergence or t = 𝑇𝑚𝑎𝑥 .
𝐹
where 𝑅𝑓,𝑘 ′ ,𝑛 = 𝐵log 2 (1 + 𝑃𝑓,𝑘′,𝑛,1 𝐹
ℎ𝑓,𝑘′,𝑛 ⁄Γ𝑓,𝑘′ 𝐼̃𝑘′,𝑛,1
𝐹
) is the
IV. HEURISTIC SOLUTION (pragmatic) achievable rate from co-channel FUs. This term
is independent of the optimization parameters, and thus it can
At convergence, Algorithm 1 generates jointly near- be removed from the objective function. We will discuss
optimal solutions for all the variables in the problem (P1). solving the mentioned subproblems in the following
However, all the parameters are updated in each iteration, subsections.
and hence the computational cost per iteration can be
significant. Furthermore, the customized Lagrange dual

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

𝑘
Algorithm 2. Heuristic solution of problem (P1) for 𝐷 γ𝑀 𝐹 𝐹 1 𝐹 𝐹
𝑡ℎ −𝑃𝑚𝑎𝑥 𝑔𝑓,𝑘,𝑀𝐵𝑆 −∑𝑘=1 𝑃𝑓,𝑘′,𝑛,1 𝑔𝑓,𝑘′,𝑀𝐵𝑆
high-level traffic situation
where 𝑃ub1 = 𝐷 .
𝑔𝑑,𝑀𝐵𝑆,𝑛
1: Define the set of PDU-FU pairs form 𝐷𝒰 and ℱ𝒰 as 𝒰. 𝐹
Similarly, an upper bound for 𝑃𝑓,𝑘,𝑛,1 can be expressed as:
2: Define 𝑿𝑫 , 𝑿𝑭 , 𝑷𝑫 , 𝑷𝑭 = ∅. 𝐹 𝐹
3: for each PDU-FU pair (𝑑, (𝑓, 𝑘)) in 𝒰 𝑃𝑓,𝑘,𝑛,1 ≤ 𝑃ub1 (21)
Define 𝒜𝑑,𝑘,𝑓 = ∅ 𝑘
4: γ𝑀 𝐷 𝐷 1 𝐹 𝐹
𝑡ℎ −𝑃𝑚𝑎𝑥 𝑔𝑑,𝑀𝐵𝑆,𝑛 −∑𝑘=1 𝑃𝑓,𝑘′,𝑛,1 𝑔𝑓,𝑘′ ,𝑀𝐵𝑆,𝑛
𝐹
5: for each SCH 𝑛 ∈ 𝒲 where 𝑃ub1 = 𝐹 . Two
𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛
6: if inequalities (22) and (23) are satisfied
𝐷 𝐹
7: Set 𝒜𝑑,𝑓,𝑘 = 𝒜𝑑,𝑓,𝑘 ∪ {𝑛} other upper bounds for 𝑃𝑑,𝑛,1 and 𝑃𝑓,𝑘,𝑛,1
are obtained from
8: end if ̂ ̂ ̂
constraints (C1) and (C6), and (C2) and (Ĉ 7), respectively.
9: end for
10: end for
̂ ̂
Finally, constraints (C1)- (C7) can be simplified as:
𝑁𝑑 𝑁𝑑 ℱ𝒰
11: Define 𝒜𝑓,𝑘 = ⋃𝑑=1 𝒜𝑑,𝑓,𝑘 , 𝒜𝑘 = ⋃𝑑=1 ⋃𝑓=1𝑘 𝒜𝑑,𝑓,𝑘 and
ℱ𝒰
Γ𝑑 𝐼̃𝑑,𝑛,1
𝐷
𝑅𝑑𝑚𝑖𝑛 ⁄𝑤𝑑𝐷𝐵𝑛 ) 𝐷
γ𝐷𝐹
𝑡ℎ 𝐷 (22)
𝒜𝑑 = ⋃𝐾𝑘=1 ⋃𝑓=1𝑘 𝒜𝑑,𝑓,𝑘 𝐷 (2 ≤ min (𝑃𝑚𝑎𝑥 , 𝐷 , 𝑃ub1 )
ℎ𝑑,𝑛,1 𝑔𝑑,𝑘,𝑛
12: if 𝒜𝑑 = ∅, or 𝒜𝑘,𝑓 = ∅
13: PDU d or FU f in FC k stays silent and 𝐷𝒰 = 𝒰\{𝑑} or
ℱ𝒰 = ℱ𝒰\{(𝑓, 𝑘)}, Γ𝑓,𝑘 𝐼̃𝑘,𝑛,1
𝐹
𝑚𝑖𝑛 𝐹
𝐹 (2𝑅𝑓,𝑘 ⁄𝑤𝑓,𝑘 𝐵𝑛
)
14: Update 𝒰. ℎ𝑓,𝑘,𝑛
15: end if 𝑘𝐹 𝐹 (23)
16: while 𝒰 ≠ ∅ γ𝐹𝑑
𝑡ℎ − ∑𝑘′=1 𝑃𝑓,𝑘′,𝑛,1 𝑔𝑓,𝑘 ′ ,𝑑,𝑛
1
𝐹
𝐹 ,
17: Sort PDU-FU pairs in 𝒰 based on the number of
≤ min (𝑃𝑚𝑎𝑥 𝐹 , 𝑃ub1 )
𝑔𝑓,𝑘,𝑑,𝑛
elements in their admissible set 𝒜𝑑,𝑘,𝑓 ≠ ∅
in the ascending order The inequalities (22) and (23) can be used to find 𝒜𝑑,𝑘,𝑓 , that
18: for each PDU-FU pair (𝑑 ∗ , (𝑓 ∗ , 𝑘 ∗ )) in the sorted 𝒰
ℎ 𝐷∗ 𝐹
ℎ𝑓∗ ,𝑘∗ ,𝑛
is the set of admissible SCHs for each pair of PDU d and FU
19: Find 𝑛∗ = argmax ( 𝑤𝑑𝐷∗ ̃𝐷𝑑 ,𝑛,1 + 𝑤𝑓𝐹∗,𝑘 ∗ ), f in FC k. In Algorithm 2, Steps 3 to 15 are related to the
𝑛∈𝒜𝑑∗ ,𝑘∗ ,𝑓∗ 𝐼 ∗
𝑑 ,𝑛,1 𝐼̃𝑘𝐹∗,𝑛,1

20: 𝑥𝑓𝐹∗,𝑘 ∗,𝑛∗,1


=1, admission control. Each SCH will be added to the set of
21: if an SCH 𝑛 ≠ 𝑛∗ has not been already assigned to 𝑑 ∗ admissible SCHs if these inequalities hold for the PDU-FU
22: Set 𝑥𝑑𝐷∗,𝑛∗,1 = 1 pair (𝑑, (𝑓, 𝑘)). A PDU d or an FU (𝑓, 𝑘) will be silent if the
23: Find the feasible 𝑃𝑑𝐷∗,𝑛∗,1 and 𝑃𝑓𝐹∗,𝑘 ∗,𝑛∗,1 from possible number of SCHs in its admissible set 𝒜𝑑 or 𝒜𝑘,𝑓 is zero.
solutions A.1-A.7, It should be noted that the co-tier interference constraint
24: else
̂ 8) will be checked after the admission and power control
(C
25: Set 𝑃𝑓𝐹∗,𝑘 ∗,𝑛∗,1 = min(𝑃𝑚𝑎𝑥𝐹
, γ𝑀 𝐹
𝑡ℎ ⁄𝑔𝑓 ∗ ,𝑘 ∗ ,𝑀𝐵𝑆,𝑛∗ )
26: end procedures are finished.
27: Remove (𝑑 ∗ , (𝑓, 𝑘 ∗ )), 𝑓 ∈ ℱ𝒰𝑘 ∗ from 𝒰, 2) SCH Allocation and Power control: For SCH allocation,
28: Remove 𝑛∗ from the admissible set of PDU-FU pairs with
𝑑 ≠ 𝑑∗,
PDU-FU pairs are sorted based on the number of elements in
29: Remove (𝑑, (𝑓 ∗ , 𝑘 ∗ )), 𝑑 ∈ 𝐷𝒰 from 𝒰, their admissible set, in ascending order. In this way, the user
30: Sort the members of 𝒰 based on 𝒜𝑑,𝑓,𝑘 ≠ ∅ in the ascending pairs with fewer options of SCH choices have higher
order priorities in SCH allocation, and are assigned first. The index
31: end for
32: end while
𝑤𝑑𝐷 ℎ𝑑,𝑛,1
𝐷
⁄𝐼̃𝑑,𝑛,1
𝐷 𝐹
+ 𝑤𝑓,𝑘 𝐹
ℎ𝑓,𝑘,𝑛 ⁄ 𝐼̃𝑘,𝑛,1
𝐹
is used to determine
33: Sort FCs in 𝒦 based on the number of SCHs assigned to them in which SCH in the admissible set is the best choice to be taken
descending order. by that PDU-FU pair. In Algorithm 2, steps 16 to 32 are
34: for each FC 𝑘 in the sorted 𝒦 related to the SCH allocation and power control; steps 16 to
35: for each SCH n= 1:N
36: if constraint (C ̂ 8) of (P4) is not satisfied 22 determine the best SCH for each pair. Step 21 assures that
37: Find 𝑓 ∗ | 𝑥𝑓𝐹∗,𝑘,𝑛,1 = 1 and 𝑑 ∗ | 𝑥𝑑𝐷∗,𝑛,1 = 1, each PDU is assigned at most one SCH (Constraint (C3)),
38: Set 𝑥𝑓𝐹∗,𝑘,𝑛,1 = 0, 𝑃𝑓𝐹∗,𝑘,𝑛,1 = 0, step 27 guarantees that the SCH is shared with at most one
39: Set 𝑃𝑑𝐷∗,𝑛,1 = min(𝑃𝑚𝑎𝑥 𝐷
, γ𝑀 𝐷
𝑡ℎ ⁄𝑔𝑑 ∗ ,𝑀𝐵𝑆,𝑛 ).
FU in each FC, step 28 applies constraints (C1) and (C2),
40: end if and step 29 imposes constraint (C4).
41: end for In steps 23 to 25, the transmission power of the users is
42: end for allocated. For the power control, we consider problem (P4)
43: Assign any remaining SCH to one of the remaining FUs with the
higher expected rate again. As the Hessian matrix of the objective function in the
𝐹
(𝑃𝑓,𝑘,𝑛,1 𝐹
= 𝑃ub2 𝐷
, 𝑃𝑑,𝑛,1 𝐷
= 𝑃ub2 ). problem (P4) is a negative definite matrix with respect to
𝐷 𝐹
𝑃𝑑,𝑛,1 and 𝑃𝑓,𝑘,𝑛,1 , the objective function is concave [22]. In
addition, all the constraints are convex as well. Therefore,
problem (P4) is a convex optimization problem. Given the
1) Admission control: First, we investigate the admission nature of the constraints in (P4), whenever the problem is
conditions regarding the constraints of the problem (P4). feasible, there is always a point in the relative interior of the
̂ 5) in the problem (P4) can be expressed as:
Constraint (C feasible set, and hence Slater’s constraint qualification
𝑘 𝐹
𝐷
γ𝑀 𝐹 𝐹 1 𝐹
𝑡ℎ − 𝑃𝑓,𝑘,𝑛,1 𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛 − ∑𝑘′=1 𝑃𝑓,𝑘′,𝑛,1 𝑔𝑓,𝑘 ′ ,𝑀𝐵𝑆,𝑛 reveals that the problem has zero duality gap [22]. Thereby,
𝑃𝑑,𝑛,1 ≤ the KKT conditions are both necessary and sufficient for
𝐷
𝑔𝑑,𝑀𝐵𝑆,𝑛 (19)
global optimality. Hence, if a feasible KKT point is located,
𝐹 𝐹 it is necessarily a globally optimal solution. The possible
and since 𝑃𝑓,𝑘,𝑛,1 ≤ 𝑃𝑚𝑎𝑥 , an upper bound can be found for
𝐷
𝑃𝑑,𝑛,1 as: KKT points are given in Appendix A.
𝐷 𝐷 Steps 33 to 42 ensure that constraint (C ̂ 8) related to the co-
𝑃𝑑,𝑛,1 ≤ 𝑃ub1 (20)
tier interference among the FCs holds for the achieved
resource allocation. To this end, the FCs are arranged in

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

10

descending order based on the number of SCHs assigned to Algorithm 3. Heuristic solution of problem (P1) for low-level
them. Then, the total interference caused by the co-channel traffic situation
𝐹
users in other FCs is calculated for each SCH in each FC in 1: Define 𝑿𝑫 , 𝑿𝑭 , 𝑷𝑫 , 𝑷𝑭 = ∅, 𝒲fr,ul =𝒲fr,ul .
the list. If it is higher than the pre-assigned maximum level, 2: for each FU (f,k)
𝐹 𝐹
3: Find 𝑛∗ = arg max ( ℎ𝑓,𝑘,𝑛 ), 𝑛 ∈ 𝒲fr,ul
the SCH is taken from the related FU in that FC. 𝐹 𝐹 𝐷 𝐷
4: if 𝑃max ℎ𝑓,𝑘,𝑛∗ ≥ max 𝑃max ℎ𝑑,𝑛∗,𝑚 , ∀𝑑 ∈ 𝐷𝒰 ,
Consequently, the PDU previously matched with that FU for 𝐹 𝑚𝑖𝑛
∀𝑚 ∈ {2,3} and 𝑅𝑓,𝑘,𝑛 ∗ ,2 ≥ 𝑅𝑓,𝑘
the SCH sharing will take the SCH shared only with one of 𝐹 𝐹 𝐹
5: Set 𝑃𝑓,𝑘,𝑛∗,2 = 𝑃max, 𝑥𝑓,𝑘,𝑛∗,2 = 1,
the MUs, and thus can transmit with the power level of 6: Update 𝒲fr,ul = 𝒲fr,ul \𝑛∗,
𝐷
𝑃ub2 = min(𝑃𝑚𝑎𝑥 𝐷
, γ𝑀 𝐷
𝑡ℎ ⁄𝑔𝑑,𝑀𝐵𝑆,𝑛 ). Indeed, the proposed 7: else
𝐷
algorithm considers higher priorities in the co-tier 8: Find 𝑑 ∗ = arg max {ℎ𝑑,𝑛 ∗ ,𝑚 } , 𝑚 ∈ {2,3}

interference observation for the FCs with fewer SCH 9: if 𝑀𝑆𝐼2,3 (𝑑 , 𝑛 ) ≥ 2 and 𝑅𝑑𝐷∗,𝑛∗,3 ≥ 𝑅𝑑𝑚𝑖𝑛
𝑃𝐷𝑈 ∗ ∗

options. This idea can improve the channel access rate and 10: Set 𝑃𝑑𝐷∗,𝑛∗,3 = 𝑃max 𝐷
, 𝑥𝑑𝐷∗,𝑛∗,3 = 1.
consequently increase the achievable rate. Finally, there may 11: Update 𝒲fr,ul = 𝒲fr,ul \𝑛∗, 𝐷𝒰 = 𝐷𝒰\𝑑 ∗ .
𝐹 𝑚𝑖𝑛
12: else if 𝑅𝑓,𝑘,𝑛 ∗ ,2 ≥ 𝑅𝑓,𝑘
be exist one or multiple SCHs that have not been assigned to 𝐹 𝐹 𝐹
13: Set 𝑃𝑓,𝑘,𝑛∗,2 = 𝑃max , 𝑥𝑓,𝑘,𝑛 ∗ ,2 = 1.
any FUs or PDUs. In step 43, each remaining SCH is
14: Update 𝒲fr,ul = 𝒲fr,ul \𝑛∗ ,
allocated to one of the remaining PDUs or FUs with the 15: else
maximum expected rate. The transmission power, in this 16: Update 𝒲fr,ul 𝐹
= 𝒲fr,ul𝐹
\𝑛∗ and go to step 3
case, is only limited by the user’s maximum power and the 17: end if
𝐷
maximum interference on the MBS, i.e., 𝑃ub2 = 18: end if
𝐷
, γ𝑀 𝐷 𝐹 𝐹 𝑀 𝐹 19: end if
min(𝑃𝑚𝑎𝑥 𝑡ℎ ⁄𝑔𝑑,𝑀𝐵𝑆,𝑛 ) and 𝑃ub2 = min(𝑃𝑚𝑎𝑥 , γ𝑡ℎ ⁄𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛 ). 20: end for
21: for each DU d ∈ {𝐷𝒰}
𝐷
B. Low-level traffic 22: Find 𝑛∗ = arg max ( ℎ𝑑,𝑛,2 ), 𝑛 ∈ 𝒲fr,ul
𝐷 𝑚𝑖𝑛
23: if 𝑅𝑑,𝑛∗,2 ≥ 𝑅𝑑
When the number of free SCHs is equal to or larger than 24: Update 𝒲fr,ul = 𝒲fr,ul \𝑛∗.
the total number of PDUs and FUs, i.e., 𝑁𝑑 + ∑𝐾 𝑘=1 𝐹𝑘 ≤ 25: else
𝐷 𝐷 𝑚𝑖𝑛
𝑁fr,ul , the network is in a low-level traffic situation. In fact, 26: Find 𝑛∗ = arg max ( ℎ𝑑,𝑛,2 ), |𝑅𝑑,𝑛 ∗ ,2 ≥ 𝑅𝑑 , 𝑛 ∈ 𝒲fr,dl
there are enough free SCHs in the macrocell to be occupied 27: Update 𝒲fr,dl = 𝒲fr,dl \𝑛∗.
28: end if
by all the PDUs and FUs. In this case, we assume that SCH 29: Set 𝑃𝑑,𝑛𝐷 𝐷 𝐷
∗ ,2 = 𝑃max , 𝑥𝑑,𝑛 ∗ ,2 = 1,
reuse is not used, and thus the FUs can transmit in the 30: end for
cellular-dedicated mode, i.e., m=2, while the PDUs may
transmit in one of the cellular-dedicated or D2D-dedicated On the other hand, a PDU in D2D-dedicated mode can use
modes, i.e., m=2,3. As there is no interfering transmission in either a free downlink or uplink SCH. In steps 21-30, the
this situation, the users can transmit with the maximum remaining PDUs take one of the remaining free SCHs in the
transmission power. In this case, each user takes a free SCH uplink sub-band on which they achieve the highest D2D
that has the highest gain, and operates in a dedicated manner. channel gain, if their minimum rate requirements are
Algorithm 3 illustrates the heuristic solution of problem (P1) guaranteed, otherwise free SCHs in the downlink sub-band
in the low-level traffic scenario. Steps 2 to 20 search for the are assigned to them. It should be noted that the constraint
free SCHs in the uplink sub-band on which FUs and PDUs related to the minimum required rate is always checked
have the highest rate. Since the FUs need both the uplink and before an SCH is assigned to a user.
downlink SCHs for the cellular communication, a higher
priority is considered for them. The algorithm searches for
C. Medium level traffic
the uplink SCH on which each FU has the maximum rate on
it first. Now for each FU and the selected SCH, the When there are a number of free SCHs in the network, but
achievable rate is calculated. If it is higher than the that number is smaller than the total number of PDUs and
achievable rates of all the PDUs (in both the D2D and FUs, i.e., 0 < 𝑁fr,ul < 𝑁𝑑 + ∑𝐾 𝑘=1 𝐹𝑘 , the traffic level is
cellular modes), the selected SCH is assigned to that FU. defined as medium. In this case, the PDUs can transmit in all
Otherwise, if a PDU can achieve the highest gain in the of their possible transmission modes, i.e., m ∈ {1,2,3}. Also,
cellular dedicated mode (m=3) on that SCH, the following FUs can transmit in all of their possible transmission modes,
index is calculated for it: i.e., m ∈ {1,2}. Algorithm 4 demonstrates the heuristic
𝐷
𝑃𝐷𝑈
ℎ𝑑,𝑛,3 (24) solution for problem (P1) when the traffic level is medium.
𝑀𝑆𝐼2,3 (𝑑, 𝑛) = 𝐷
ℎ𝑑,𝑛,2 According to the proposed scheme, the difference between
𝐷
where ℎ𝑑,𝑛,3 is the channel gain from PDU d to the MBS the number of users and the number of free SCHs is
(cellular mode) and ℎ𝑑,𝑛,2 𝐷
is the channel gain from the considered as the number of users to transmit in D2D-reuse
(m=2 for PDUs) or cellular-reuse mode (m=1 for FUs), i.e.,
transmitter of PDU d to its receiver d' (D2D mode) on SCH
n. This index is the mode selection criterion for the PDUs. 𝑁𝑑 + ∑𝐾 𝑘=1 𝐹𝑘 − 𝑁fr,ul users are selected to take SCHs shared

Since the cellular transmission mode consumes twice the with MUs using the following indices:
spectral resources of the D2D transmission, only a PDU with 𝑤𝑑𝐷 ℎ𝑑,𝑛,1
𝐷
(25)
MSI1𝑃𝐷𝑈 (𝑑) = min ( 𝑀 ) , 𝑛∈𝒲
𝑃𝐷𝑈
𝑀𝑆𝐼2,3 ≥ 2 is allowed to be selected for the cellular- 𝑛 𝑔𝑛,𝑑′
𝐹 𝐹
dedicated mode on a given SCH. 𝐹𝑈
𝑤𝑓,𝑘 ℎ𝑓,𝑘,𝑛 (26)
MSI1,2 (𝑓, 𝑘) = min ( 𝑀 ), 𝑛 ∈ 𝒲
𝑛 𝑔𝑛,𝑘

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

11

𝐷
where ℎ𝑑,𝑛,1 is the gain of transmission from the PDU Table IV. Simulation parameters
transmitter d to its receiver d' (D2D-reuse mode) on SCH n, Para. value Para. value Para. value
𝑀
and 𝑔𝑛,𝑑′ is the gain of the interfering transmission from
every single MU in the network to d', the receiver of the PDU 𝑅 500m 𝑃𝑛𝑀 ~𝑈(0,24)dBm γ𝑀
𝑡ℎ -50dBm
𝐹 𝐹
d. The proposed index in (25) considers the ratio that has the 𝑅𝑑2𝑑 10m 𝑃max 18dBm 𝛾𝑡ℎ -50dBm
lowest value as the mode selection criterion for that PDU 𝑅𝑓 25m 𝐷
𝑃max 12dBm 𝐹𝐷
𝛾𝑡ℎ -60dBm
𝑀
pair. Also, 𝑔𝑛,𝑘 is the gain of the interfering transmission 𝐵𝑛 100𝐾𝐻𝑧 𝑅∙𝑚𝑖𝑛 350 𝐾𝑏𝑝𝑠 𝐷𝐹
𝛾𝑡ℎ - 60dBm
from every single MU in the network to the FBS k. The
𝑁 120 𝑁𝑑 20 𝑛0 -174 dBm
proposed index in (26) calculates the ratio of useful link gain
of each FU f in the FC k to the interfering link gain of every K 20 𝐹𝑘 3, ∀𝑘 |ℳ𝒰| ~𝑈(0, 𝑁 )
single MU to the FBS k, and then considers the ratio with the
lowest value as the mode selection criterion for that FU. A. Complexity Analysis
After calculating the above indices for all the PDUs and FUs,
We begin our experiments with an analysis of the
they are arranged in descending order based on their index
values and the first 𝑁𝑑 + ∑𝐾 computational complexity. Table III shows the asymptotic
𝑘=1 𝐹𝑘 − 𝑁fr,ul users (PDUs or
FUs) in the sorted list are selected for the reuse mode (D2D- worst-case complexity of the proposed iterative scheme in
reuse mode for PDUs and cellular-reuse mode for FUs). For Algorithm 1, the proposed heuristic scheme, and the Branch
reuse mode users, admission control, SCH allocation, and & Bound method. In Table III, ∆ is the iteration number of
power assignment are carried out like the procedure in the subgradient method. It can be seen that the complexity of
Algorithm 2. The remaining PDUs can transmit in one of the the heuristic algorithm increases linearly with the number of
D2D-dedicated or cellular-dedicated modes, while the uplink SCHs, and log-linearly with the number of PDUs, FCs,
remaining FUs will transmit in the cellular-dedicated mode. and FUs in each FC. Also, the complexity of the iterative
Algorithm 3 selects the proper transmission modes and algorithm grows quadratically with the number of uplink SCHs,
assign the free SCHs to the remaining users. The overall FCs, FUs in each FC, and PDUs, and linearly with the number
heuristic solution for problem (P1) when the traffic level is of iterations. Although the complexity growth of the iterative
medium is given in Algorithm 4. algorithm is faster than the heuristic algorithm, both are
within the polynomial time. In contrast, the worst-case
Algorithm 4. Heuristic solution of problem (P1) for medium-level computational complexity of the Branch & Bound searching
traffic situation
algorithm for finding the optimum solution for problem (P1)
1: Calculate the index in (25) for each PDU 𝑑,
2: Calculate the index in (26) for each FU (𝑓, 𝑘), is exponential. It is observed that finding the optimum
3 Arrange PDUs and FUs in descending order based on their index solution has significantly higher complexity than the
values,
proposed algorithms. This complexity difference will be
4: Select the first 𝑁𝑑 + ∑𝐾𝑘=1 𝐹𝑘 − 𝑁fr,ul users in the sorted list and
run Algorithm (2) for them; dramatically significant in dense heterogeneous networks
5: Run Algorithm (3) for the remaining users. with a large number of FCs and PDUs.

V. EVALUATION B. Simulation Results

In this section, we assess the performance of the proposed We will denote the radius of the macrocell by R. In the
joint mode selection resource allocation schemes for the considered macrocell, the FBSs are distributed uniformly
random in the network, and the FUs are distributed
three-tier HetNet in comparison with (i) the optimum
uniformly-random in the coverage area of each FC, which
solution, obtained (at substantial computational expense)
has radius 𝑅𝑓 . Also, the MUs are distributed uniformly-
using the Branch and Bound (B&B) algorithm, and (ii)
random within the coverage of the macrocell. We consider
Random SCH/Mode allocation with the proposed Power 𝑁𝑑 PDU transmitters located in the macrocell with uniform
Control (RSMPC) approach. The RSMPC selects reuse- distribution, where their receivers are distributed uniformly-
mode PDUs and FUs randomly in medium and high traffic random within 𝑅𝑑2𝑑 meters of them. The available uplink
situations. Also, the SCHs to be shared among all three tiers bandwidth is 12 MHz, divided into 120 SCHs with equal
are chosen in a random manner. For the reuse mode users, frequency bandwidths of 100 kHz. Also, the SNR gap is set
the proposed power control is carried out; otherwise, the to 0 dB, and all the rate weights are set to 1. For the
transmission power is set to the maximum level and SCHs calculation of channel gains on each link, path loss,
are assigned to the users randomly. The performance metric shadowing, and fast fading are considered. The shadowing
is the total sum rate (TSR) of PDUs and FUs with all the rate has a log-normal distribution with a standard deviation of 4
weights set to 1. dB and 8 dB for inside the femtocell and outside the
femtocell, respectively, and the fast fading is modeled by
Table III: Computational complexity comparison independently and identically distributed Rayleigh random
Iterative scheme 𝒪((9𝑁 2 𝑁𝑑2 + 4𝑁 2 𝐾 2 𝐹𝑘2 )∆) variables with unit variance. The path-loss model (in
Heuristic scheme 𝑂(𝑁𝑑 𝐾𝐹𝑘 (2𝑁 + log 2 𝑁𝑑 𝐾𝐹𝑘 )) decibels) for the macrocell and femtocell links is PL=128.1 +
B&B (optimum) scheme 𝑂(𝑁𝑑 + 𝐾𝐹𝑘 )𝑁
37.6 log10 (𝑟), and for the D2D links is PL=148.1 +
40 log10 (𝑟), where r is the distance between the transmitter
and the receiver in km [26]. The main parameters used in the
simulations are shown in Table IV unless stated otherwise.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

12

100 Fig. 2 shows the TSR of PDUs and FUs when their
minimum rate requirement (𝑅.𝑚𝑖𝑛 ) increases from 200 to 900
kbps in a high-traffic situation for case I (𝑃max 𝐷 =

80 𝐹 𝐷
12dBm , 𝑃max = 18 dBm) and case II (𝑃max = 6dBm , 𝑃max 𝐹 =
TSR (Mbps)

9 dBm). As seen in Fig. 2, the TSR decreases because a larger


𝑅.𝑚𝑖𝑛 shrinks the feasible region in the optimization problem.
60 In particular, for higher values of 𝑅.𝑚𝑖𝑛 , the admission
constraints become tighter, and thus fewer reuse-mode users
Iterative-I Iterative-II are admitted to take the SCHs. Therefore, the decreasing
40 Heuristic-I Heuristic-II trend of the curves becomes faster. Also, the performance of
RSMPC-I RSMPC-II the proposed schemes is considerably higher than the
200 300 400 500 600 700 800 900
RSMPC approach, which confirms the importance of
Minimum rate requirement (kbps) suitable admission control and SCH allocation. In addition,
Fig. 2. Performance versus the minimum rate requirement it can be seen that a higher maximum transmission power
results in a significant improvement in the TSR.
Iterative-I Iterative-II
120
Heuristic-I Heuristic-II
Fig. 3 plots the TSR of PDUs and FUs with respect to their
RSMPC-I RSMPC-II maximum transmission power in a high-traffic situation for
100 𝐹𝐷 𝐷𝐹 𝐹𝐷 𝐷𝐹
two cases of 𝛾𝑡ℎ = 𝛾𝑡ℎ = −60dBm (Case I) and 𝛾𝑡ℎ = 𝛾𝑡ℎ =
TSR (Mbps)

80 −100dBm (Case II). It is observed that the TSR improves


considerably as the maximum transmission power increases,
60 except for the RSMPC approach in case II; because in case
40
II, the admission constraints become tighter with decreasing
𝐹𝐷 𝐷𝐹
𝛾𝑡ℎ and 𝛾𝑡ℎ , and the TSR will degrade if an appropriate
20 admission control, mode selection, and SCH allocation
approach is not applied.
0
4 8 12 16 20 24
Fig. 4 compares the performance of the proposed
Maximum transmission power of PDUs and FUs (dBm) algorithms with the RSMPC approach as the number of
FBSs increases in a high-traffic situation. It is seen that the
Fig. 3. Performance versus the maximum transmission power
increasing trend of the TSR becomes slow as the number of
FBSs increases because the reuse gain achieved from the
110
Iterative user diversity is limited by the interference constraints (C̃ 11)-
Heuristic ̃ 14).
(C
RSMPC
90 Fig. 5 shows the TSR when the interference limit on the
MBS (𝛾𝑡ℎ 𝑀
) increases in the high traffic situation. We can see
TSR (Mbps)

70 that the performance of the heuristic scheme is close to that


of the iterative scheme, even though it has a significantly
50 lower computational cost. In addition, the performance of all
schemes improves with an increase in 𝛾𝑡ℎ 𝑀
. This is because
𝑀
30 as 𝛾𝑡ℎ increases, the PDUs and FUs are more likely to be
admitted to take an SCH, and the problem constraints
3 6 9 12 15 18 21 24
become looser. However, the increasing trend becomes slow
𝑀
Number of FBSs
for higher values of 𝛾𝑡ℎ , because the sum rate will be limited
Fig. 4. Performance versus the number of femtocells
by other constraints, such as the maximum transmission
powers, the co-tier interference among the FCs, and the
100
cross-tier interference between the PDUs and FUs.
90
170
Iterative
TSR (Mbps)

80 Heuristic High traffic


RSMPC 140
70
TSR (Mbps)

Low traffic

60 110

Medium traffic
50 80 Iterative
Heuristic
-90 -80 -70 -60 -50 -40 -30 -20 RSMPC
Interference limit on the MBS (dBm)
50
Fig. 5. Performance versus the interference limit on the MBS 40 50 60 70 80 90 100 110 120
Number of MUs
The performance of the proposed algorithms is evaluated
with an average of 3000 independent realizations. Fig. 6. Performance versus the number of MUs

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

13

Fig. 6 illustrates the impact of the network traffic situation 10.0


on the TSR of PDUs and FUs. In this case, the number of B&B Heuristic
Iterative RSMPC
MUs is increased from 40 to 120, the number of uplink SCHs 9.5
is set to 120 and the total number of PDUs and FUs is fixed
9.0

TSR (Mbps)
at 50. Therefore, the number of free SCHs decreases from 80
to 0. When the number of MUs is less than or equal to 70 8.5
(𝑁fr,ul ≥ 50), the performance decreases slowly. In this case,
the network is in the low-traffic situation, and an increment 8.0
in the number of MUs reduces the number of free SCHs. 7.5
When the number of free SCHs decreases, the number of
SCH options is limited for each PDU or FU and thus the TSR 7.0
decreases slightly. When the number of MUs is between 80 30 35 40 45 50 55 60 65 70
and 110 (10 ≤ 𝑁fr,ul ≤ 40), the network traffic is medium, Coverage radius of the femtocell (m)

and thus, the reduction in the number of free SCHs has a (a)
more significant negative effect on the sum rate. When the
number of MUs is 120, there is no free SCH in the network 10.0
(high-traffic situation), and the acquired TSR results from 9.5
9.0
the reuse gain over the pure cellular case. Generally, the TSR
8.5
degrades when the network traffic level moves from low to

TSR (Mbps)
8.0
medium and high because when the number of free SCHs is
7.5
reduced while the number of PDUs and FUs is fixed, the
7.0
probability of taking an SCH by each user and the sum rate B&B
6.5
decreases. Iterative
6.0 Heuristic
In order to evaluate our proposed schemes compared with RSMPC
5.5
the optimum B&B algorithm in reasonable amount of time,
5.0
we consider a smaller network including an MBS with a
30 35 40 45 50 55 60 65 70
radius of 150m, two FCs each of which serves one FU, one
(b)
pair of PDU and 3, 1, and 0 free SCHs accordingly for low,
6.0
medium, and high traffic situations. Figure 7 depicts the
performance of the network with respect to the coverage 5.5
radius of FCs, in the corresponding low (a), medium (b), and 5.0
B&B
TSR (Mbps)

high (c) traffic situations, respectively. It can be seen that the 4.5 Iterative
Heuristic
performance reduces monotonically as the femtocell radius 4.0 RSMPC
increases. This behavior is expected since the channel gains 3.5
of FUs decline as the radius of FCs increases. On the other 3.0
hand, a bigger radius of FCs results in a more complicated
2.5
interference environment. The combination of these two
2.0
factors results in the observed trend in the figures. It is also
observed that the performance of the proposed iterative 30 35 40 45 50 55 60 65 70
scheme is close to the optimal B&B scheme (around 94% of (c)
the optimum performance in the worst case). The higher Fig. 7. Performance versus the Radius of FCs in a (a): low, (b):
difference between the iterative and heuristic approaches in medium, and (c): high traffic situation.
the low-traffic situation stems from the fact that we ignore
the reuse mode in the heuristic scheme for low-traffic levels. iterative algorithm based on the Lagrange dual
In contrast, the iterative scheme benefits from that, and thus, decomposition method. To further reduce the computational
more options are available for allocating SCHs to the users complexity, a low-complexity decomposition-based method
in the iterative scheme. Also, the decreasing trend of the was also derived, in which we performed, in sequence, mode
curves becomes faster in the medium and high-traffic cases selection, admission control, power control, and SCH
when the radius increases. The reason is that the reuse gain allocation for the FUs and PDUs. The proposed heuristic
decreases because the interference level becomes more mode selection algorithm is based on the traffic situation of
intense. the network, which is classified into high, medium, and low-
level traffic situations. The simulation results validate the
VI. CONCLUSION
efficiency of the proposed approaches. Particularly, the
In this paper, we considered a communication scenario with performance of the iterative scheme is around 94% of that of
a dense existence of PDUs and FUs in a single macrocell, in the optimum B&B approach. It is also observed that the
which the aim was maximizing their weighted sum rate while heuristic scheme performs close to the near-optimal iterative
satisfying the minimum rate requirement and constraining solution, and yet has a much lower computational cost. In
the imposed interference on the MBS. In order to improve future work, we plan to consider a dynamic power allocation
the system capacity, a full frequency reuse framework was for the MUs and formulate the problem from a capacity-
utilized. The maximization problem was modeled as a mixed fairness perspective. In order to achieve enhanced
integer non-linear programming problem, and good interference mitigation, we intend to incorporate non-
solutions were efficiently obtained using a customized

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

14

orthogonal multiple access (NOMA) into our D2D-enabled [6] P. Bithas, K. Maliatsos, and F. Foukalas, “An SINR-aware joint mode
selection, scheduling, and resource allocation scheme for D2D
heterogeneous network.
Communications,” IEEE Trans. Veh. Technol., vol. 68, no. 5, pp. 4949
– 4963, 2019.
Appendix A. [7] C. Chen, C. Sung, and H. Chen, “Capacity Maximization Based on
Optimal Mode Selection in Multi-Mode and Multi-Pair D2D
Possible KKT points of the problem (P4) Communications,” IEEE Trans. Veh. Tech., vol. 68, no. 7, pp. 6524-
6534, 2019.
The complementary slackness conditions of the KKT
[8] Y. Jiang, et al., “Joint power and bandwidth allocation for energy
optimality conditions will lead to different possible cases, efficient heterogeneous cellular networks,” IEEE Trans. Commun.,
which should be investigated. For the problem (P4), after vol. 67, no. 9, pp. 6168 – 6178, 2019.
removing cases that result in a direct contradiction or [9] L. Eslami, and G. Mirjalily, “Throughput enhancement of D2D enabled
infeasible points, seven cases remain to be investigated. For cellular systems using a situation-aware mode selection approach,”
𝐷 Proc. IEEE International Conf. Signal & Image Processing,
each PDU-FU pair, the transmission powers 𝑃𝑑,𝑛,1 and
𝐹 Shenzhen, China, 2018.
𝑃𝑓,𝑘,𝑛,1 are calculated via the following equations in each [10] Y. Du, W. Zhang, S. Wang, J. Xia and H. A. Mohammad, “Joint
case. Here, 𝜇1 − 𝜇7 are the Lagrange multipliers according Resource Allocation and Mode Selection for Device-to-Device
to the constraints in (C ̂ 1)- (Ĉ 7) of the problem (P4), Communication Underlying Cellular Networks,” IEEE Access, vol. 9,
pp. 29020-29031, 2021.
respectively. It is obvious that the accepted solution should
[11] Sun, Y., Miao, M., Wang, Z. et al., “Resource Allocation Based on
be feasible and must not violate the Lagrange multiplier Hierarchical Game for D2D Underlaying Communication Cellular
condition of the related case. Networks,” Wireless Pers. Commun., vol. 117, pp. 281–291, 2021.
𝐷 𝐷 𝐹 𝐹 [12] X. Liu, H. Xiao and A. T. Chronopoulos, “Joint mode selection and
𝑃𝑑,𝑛,1 = 𝑃𝑚𝑎𝑥 , 𝑃𝑓,𝑘,𝑛,1 = 𝑃𝑚𝑎𝑥 , 𝜇1 , 𝜇2 > 0 A.1
𝑘
power control for interference management in D2D-enabled
γ𝑀 𝐹 𝐹 1 𝐹 𝐹
𝑡ℎ −𝑃𝑚𝑎𝑥 𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛 −∑𝑘′=1 𝑃𝑓,𝑘′,𝑛,1 𝑔𝑓,𝑘′,𝑀𝐵𝑆,𝑛
𝐷 heterogeneous cellular networks,” IEEE Trans. Veh. Tech., vol. 69,
𝑃𝑑,𝑛,1 = 𝐷
{ 𝑔𝑑,𝑀𝐵𝑆,𝑛 no. 9, pp. 9707-9719, 2020.
𝐹 𝐹
𝑃𝑓,𝑘,𝑛,1 = 𝑃𝑚𝑎𝑥 , 𝜇2 , 𝜇5 > 0 A. 2 [13] G. Apostolos, K. Konstantinos, N. Aikaterini, F. Foukalas and T.
𝑀 𝐷 𝐷 𝑘1 𝐹 𝐹 Khattab, “Energy efficient spectrum allocation and mode selection for
γ𝑡ℎ − 𝑃𝑚𝑎𝑥 𝑔𝑑,𝑀𝐵𝑆,𝑛 − ∑𝑘′=1 𝑃𝑓,𝑘′,𝑛,1 𝑔𝑓,𝑘 ′ ,𝑀𝐵𝑆,𝑛
𝐹
𝑃𝑓,𝑘,𝑛,1 = mission-critical D2D communications,” IEEE Conference on
{ 𝐹
𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛 Computer Communications Workshops, pp. 435-440, 2016.
𝐷
𝑃𝑑,𝑛,1 = 𝐷
𝑃𝑚𝑎𝑥 𝜇1 , 𝜇5 > 0 A. 3 [14] P. Chu, X. Wang, D. Wang, and L. Yu, “A D2D mode selection scheme
𝑘1 𝐹 𝐹 with energy consumption minimization underlaying two-tier
γ𝑀
𝑡ℎ −
𝐷
𝑃𝑑,𝑛,1 𝐷
𝑔𝑑,𝑀𝐵𝑆,𝑛 − ∑𝑘′=1 𝑃𝑓,𝑘′,𝑛,1 𝑔𝑓,𝑘 ′ ,𝑀𝐵𝑆,𝑛
𝐹
𝑃𝑓,𝑘,𝑛,1 = heterogeneous cellular networks,” IEEE 28th Annual International
𝐹
𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛 Symposium on Personal, Indoor, and Mobile Radio Communications,
̃𝐷 pp. 1-5, 2017.
𝐷
Γ𝐼𝑑,𝑛,1 𝑚𝑖𝑛
𝑃𝑑,𝑛,1 = 𝐷 (2𝑅𝑑 ⁄𝐵𝑛 − 1), 𝜇3 , 𝜇5 > 0 A. 4 [15] L. Eslami, G. Mirjalily and T. N. Davidson, “Spectrum-efficient QoS-
{ ℎ𝑑,𝑛,1 aware resource assignment for FFR-based D2D-enabled
𝐹 𝐹 𝑘1 𝐹 𝐹
γ𝑀
𝑡ℎ − 𝑃𝑓,𝑘,𝑛,1 𝑔𝑓,𝑘 ′ ,𝑀𝐵𝑆,𝑛 − ∑𝑘 ′ =1 𝑃𝑓,𝑘 ′ ,𝑛,1 𝑔𝑓,𝑘 ′ ,𝑀𝐵𝑆,𝑛 heterogeneous networks,” IEEE Access, vol. 8, pp. 218186-218198,
𝐷
𝑃𝑑,𝑛,1 = 𝐷 2020.
𝑔𝑑,𝑀𝐵𝑆,𝑛
̃𝐹 [16] M. Wang, H. Gao, X. Su, and T. Lv, “Joint channel allocation, mode
𝐹
Γ𝐼𝑘,𝑛,1 𝑅 𝑚𝑖𝑛⁄𝐵 selection and power control in D2D-enabled femtocells,” Proc. IEEE
𝑃𝑓,𝑘,𝑛,1 = 𝐹 (2 𝑑 𝑛 − 1), 𝜇4 , 𝜇5 > 0 A. 5
{ ℎ𝑓,𝑘,𝑛 Mil. Commun. Conf., pp. 454–459, Baltimore, USA, 2016.
𝑘1
γ𝑀 𝐷 𝐷 𝐹 𝐹
𝑡ℎ − 𝑃𝑑,𝑛,1 𝑔𝑑,𝑀𝐵𝑆,𝑛 − ∑𝑘′=1 𝑃𝑓,𝑘′,𝑛,1 𝑔𝑓,𝑘 ′ ,𝑀𝐵𝑆,𝑛
[17] H. Gao, M. Wang, and T. Lv, “Energy efficiency and spectrum
𝐹
𝑃𝑓,𝑘,𝑛,1 = 𝐹 efficiency tradeoff in the D2D-enabled hetnet,” IEEE Trans. Veh.
𝑔𝑓,𝑘,𝑀𝐵𝑆,𝑛 Tech., vol. 66, no. 11, pp. 10583–10587, 2017.
𝐷
γ𝐷𝐹
𝑡ℎ [18] Y. Huang, A. A. Nasir, S. Durrani, and X. Zhou, “Mode selection,
𝑃𝑑,𝑛,1 = 𝐷 , 𝜇5 , 𝜇6 > 0 A. 6 resource allocation, and power control for D2D-enabled two-tier
{ 𝑔𝑑,𝑘,𝑛
cellular network,” IEEE Trans. Commun., vol. 64, no. 8, pp. 3534–
𝐹 𝐹 𝐹 𝑘 𝐹 3547, 2016.
γ𝑀 1
𝑡ℎ − 𝑃𝑓,𝑘,𝑛,1 𝑔𝑓,𝑘,𝑀𝐵𝑆 − ∑𝑘′=1 𝑃𝑓,𝑘′,𝑛,1 𝑔𝑓,𝑘 ′ ,𝑀𝐵𝑆,𝑛
𝐷 [19] A. A. Algedir and H. H. Refai, “Energy Efficiency Optimization and
𝑃𝑑,𝑛,1 = 𝐷
𝑔𝑑,𝑀𝐵𝑆,𝑛 Dynamic Mode Selection Algorithms for D2D Communication Under
γ𝐹𝐷 −
𝑘1 𝐹 𝐹
∑𝑘′=1 𝑃𝑓,𝑘′,𝑛,1 𝑔𝑓,𝑘 ′ ,𝑑,𝑛
HetNet in Downlink Reuse,” IEEE Access, vol. 8, pp. 95251-95265,
𝐹 𝑡ℎ
𝑃𝑓,𝑘,𝑛,1 = 𝐹 , 𝜇5 , 𝜇7 > 0 A. 7 2020.
{ 𝑔𝑓,𝑘,𝑑,𝑛
[20] T. Zhang, K. Zhu, and J. Wang, “Energy-Efficient Mode Selection and
Resource Allocation for D2D-Enabled Heterogeneous Networks: A
REFERENCES Deep Reinforcement Learning Approach,” IEEE Trans. Wirel.
Commun., vol. 20, no. 2, pp. 1175-1187, 2021.
[1] G. Liu, W. Feng, Z. Han, and W. Jiang, “Performance analysis and
[21] A. Galanopoulos, F. Foukalas, and T. Khattab, “Energy efficient
optimization of cooperative full-duplex D2D communication
spectrum allocation and mode selection for D2D communications in
underlaying cellular networks,” IEEE Trans. Commun., vol. 18, no.
heterogeneous networks,” IEEE Trans. Signal Inf. Process. Netw., vol.
11, pp. 5113 – 5127, 2019.
6, pp. 382-393, 2020.
[2] G. Wu, Q. Li, R. Q. Hu, and Y. Qian, Overview of heterogeneous
[22] S. Boyd and L. Vandenberghe, “Convex optimization,” New York,
networks, “Heterogeneous Cellular Networks,” 1st ed., H. Sun and R.
NY: Cambridge University Press, 2004.
Q. Hu, Eds. United Kingdom: John Wiley & Sons, ch. 1, 2013.
[23] S. Low and D. Lapsley, “Optimization flow control. I. basic algorithm
[3] Y. Zou, T. Wu, M. Sun, J. Zhu, M. Qian, and C. Liu, “Secrecy outage
and convergence,” IEEE/ACM Trans. Netw., vol. 7, no. 6, pp. 861–
analysis of non-orthogonal spectrum sharing for heterogeneous
874, 1999.
cellular networks,” IEEE Trans. Commun., vol. 67, no. 9, pp. 6626 –
6640, 2019. [24] D. Bertsekas, “Convex optimization theory,” Belmont, Massachusetts:
Athena Scientific, 2009.
[4] L. Song, D. Niyato, Z. Han, E. Hossain, “Basics of D2D
communications, wireless device-to-device communications and [25] S Boyd, L Xiao, A Mutapcic, “Lecture notes of EE392o,” Stanford
networks,” New York, NY: Cambridge University Press, Ch. 1, 2015. University, Autumn Quarter 2004,
https://web.stanford.edu/class/ee392o/subgrad_method.pdf.
[5] J. Liu, N. Kato, J. Ma, and N. Kadowaki, “Device-to-device
communication in LTE-advanced networks: A survey,” IEEE [26] 3GPP, Further advancements for E-UTRA physical layer aspects
Commun. Surveys Tuts., vol. 17, no. 4, pp. 1923–1940, 2015. (3GPPTR 36.814), Tech. Rep., 2010.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Vehicular Technology. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TVT.2023.3283306

15

LALEH ESLAMI received the B.Sc.


degree in electrical engineering
(Electronics) from Isfahan University of
Technology, Isfahan, Iran, in 2006, and the
M.Sc. and Ph.D. degrees in electrical
engineering (Telecommunications) from
Malek Ashtar University of Technology,
Tehran, Iran, and Yazd University, Yazd, Iran, in 2010 and
2022, respectively. Her current research interests include
wireless communications and heterogeneous cellular
networks with special emphasis on resource allocation and
interference management.

GHASEM MIRJALILY (Senior Member,


IEEE) received his Ph.D. degree in
telecommunication engineering, in 2000.
Since then, he has been with Yazd
University, Yazd, Iran, where he is a
professor. He was a visiting researcher at
McMaster University, Canada, in 1998 and
a visiting research scientist at Shenzhen Research Institute of
Big Data (SRIBD), Chinese University of Hong Kong,
Shenzhen, in summers 2017 and 2018. His current research
interests include Data Communication Networks, Wireless
Networks, Network Virtualization and Service Chaining.

TIMOTHY N. DAVIDSON (Fellow,


IEEE) received the B.Eng. (Hons. I) degree
in electronic engineering from The
University of Western Australia (UWA),
Perth, WA, Australia, in 1991, and the
D.Phil. degree in engineering science from
the University of Oxford, Oxford, U.K., in
1995. He is a Professor with the Department of Electrical and
Computer Engineering, McMaster University, Hamilton,
ON, Canada. Previously, he has served as an Acting Director
of the School of Computational Engineering and Science for
two years, and as an Associate Director for three years. His
research interests lie in the general areas of communications,
signal processing, and control. He was the recipient of the
1991 J. A. Wood Memorial Prize from UWA, the 1991
Rhodes Scholarship for Western Australia, and a 2011 Best
Paper Award from the IEEE Signal Processing Society. He
was an Associate Editor for the IEEE Transactions on Signal
Processing, the IEEE Transactions on Wireless
Communications, and the IEEE Transactions on Circuits and
Systems. He was also a Guest Co-Editor of issues of the
IEEE Journal on Selected Areas in Communications, the
IEEE Journal of Selected Topics in Signal Processing, and
the EURASIP Journal on Advances in Signal Processing. He
was a General Co-Chair for the 2014 IEEE International
Workshop on Signal Processing Advances in Wireless
Communications, a Technical Program Co-Chair for the
2014 IEEE Global Conference on Signal and Information
Processing, and the Technical Chair for the 2015 Asilomar
Conference on Signals, Systems and Computers. He has also
served as the Chair of the IEEE Signal Processing Society’s
Technical Committee on Signal Processing for
Communications and Networking. He is a Registered
Professional Engineer in the Province of Ontario.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 11,2023 at 00:05:49 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.

You might also like