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/ACCESS.2019.2940947, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

Performance Analysis and Optimization

of a N -class Bipolar Network
PLÍNIO S. DESTER1 , PAULO CARDIERI2 , AND PEDRO H. J. NARDELLI3,4 , AND
JOSÉ M. C. BRITO5
1
School of Electrical and Computer Engineering, State University of Campinas, Brazil (e-mail: [email protected])
2
School of Electrical and Computer Engineering, State University of Campinas, Brazil (e-mail: [email protected])
3
School of Energy Systems, LUT University, Finland (e-mail: [email protected])
4
Centre for Wireless Communications, University of Oulu, Finland
5
National Institute of Communications of Brazil (e-mail: [email protected])
Corresponding author: Plínio S. Dester (e-mail: [email protected]).
This work is supported in part by the Foundation for Research Support of the State of São Paulo under Grant 2017/21347-0, the Brazilian
National Council for Scientific and Technological Development under Grant 311485/2015-4, the Brazilian Fund for Technological
Development of Telecommunications under Grant 01.14.0231.00, the Academy of Finland via ee-IoT under Grant ICT2023/n.319009 and
FIREMAN consortium under Grant CHIST-ERA/n.326270.

ABSTRACT This paper presents necessary and sufficient conditions for the stability of an interference-
limited high-mobility N -class bipolar network, where the positions of the sources follow a “mother”
Poisson point process (PPP) and each source transmits to the closest terminal within the pool of potential
receivers generated by a corresponding “son” PPP. Each transmitter has a buffer of infinite capacity, where
packets arrive according to a Bernoulli distribution. We derive simple closed formulas (a finite combination
of elementary functions) for the stationary packet success probability and mean packet delay. We also
establish enlightening relations regarding system parameters and performance measures for all classes,
including average transmitter-receiver separation distances, densities of transmitters, traffic intensities,
target transmission rates, and average packet delays. We formulate and solve analytically two optimization
problems that minimize the mean packet delays and maximize the channel throughput per unit of area. These
results indicate that the proposed analysis is capable of assessing the trade-off involved in radio-resource
management when different classes of users are considered.

INDEX TERMS Design optimization, Queueing analysis, Stochastic processes, Wireless networks.

I. INTRODUCTION sharing radio resources.

FFICIENT use of radio resources has always been an Each of these subnetworks will be dedicated to serve
E important aspect in the deployment of large-scale wire-
less communications systems [1]. This situation is now more
a particular type of application and/or scenario, with its
own requirements, such as coverage, transmission rates and
exacerbated due to the growth of the number of applications maximum acceptable latency [5]. It seems to be a consensus
that require higher data rates like video streaming services in the academic and industrial communities that the goals
[2]. On top of it, the number of terminals is exponentially imposed on 5G systems will only be achieved through the
increasing; in a not very far future, there will be more use of heterogeneous networks [6]. Although extensively
machines communicating to each other than humans [3]. studied by the scientific community, there is still a lack of
In a new world of the Internet of Things (IoT), wireless understanding of the theoretical limits of interference-limited
communications systems will need to serve both human- wireless networks where the relative positions of terminals
and machine-type communications while being capable of and their activities are unknown.
delivering data rates of up to tens of Gb/s, latency in the order We are interested here in studying the performance of
of milliseconds, and reduced energy consumption to 10% of a network composed by N classes of users that share the
the current values. All these are set as targets for the fifth- same radio resources, namely radio spectrum and transmit
generation cellular system (5G System) [4]. power. Packets arrive at the transmitters randomly, which are
In one scenario envisioned for the 5G systems, a number modeled by queues where packets waiting for transmission
of subnetworks will co-exist in the same geographic area, are stored. Each of these N classes has its own characteristics

VOLUME X, 2019 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

(e.g., terminal density, transmit power and traffic intensity)

and quality of service requirements (communication link
quality and maximum tolerable mean delay). IoT scenarios
covered by 5G systems can be viewed in this way since
thousands of wireless terminals connected to sensors access
the wireless network to transmit their messages [7]. Each
sensor is associated with a specific application that imposes
different requirements: automatic control systems may need
millisecond-delay and 99.999% reliability [8], while teleme-
try may accept minutes of delay and weaker reliability con-
straints [9]. The wireless connections may involve an access
point (cellular mode) or direct device-to-device links (D2D
mode) [10]. User Class 1
In the scenario investigated here, links of all user classes
User Class 2
share the same channel, causing mutual interference, which
in turn makes the queues of the terminals coupled since the User Class 3
transmission success probability of a transmitter (i.e., the
service rate of the associated queue) depends on the state of FIGURE 1. Example of a bipolar random network, with N = 3 user classes.
The filled geometric shapes represent the transmitters, and the potential
the queues of other transmitters (whether their queues are receivers are represented by the corresponding empty geometric shapes.
either empty or non-empty). The analysis of networks with Each transmitter communicates with the closest potential receiver, as it is
shown by the lines. Unconnected empty geometric shapes represent inactive
coupled queues is still a fertile research field, especially when receivers.
the capture model is adopted [11]. This model assumes that a
packet is successfully received if the corresponding signal to
interference plus noise ratio (SINR) at the receiver is above of stable networks, we show that the portion of the radio
a certain threshold (in contrast to the collision model [11] resource allocated to each class of users is well-defined by
where a packet transmission is successful only if there are no a simple expression relating its average delay, intensity of
concurrent transmissions). The concept of stochastic domi- traffic, density of terminals, and the minimum acceptable link
nance was employed in several works found in the literature signal-to-interference ratio . Our results evince the interplay
(see, for instance, [12], [13]) to determine the conditions for among stability, traffic intensity, density of users and outage
queue stability under coupled queues scenarios. Stochastic probability in a scenario where terminals share radio re-
dominance is based on the comparison of the original system sources, through a simple formulation that allows for insights
network with a simpler and sub-optimal system that, for into the existing trade-offs among key network parameters
example, transmits dummy packets for a given set of users. for N classes of users (Proposition 2 and Corollary 1).
If it is proven that the sub-optimal network is stable, then the The scenario considered in this paper was studied by other
original (dominated) network is also stable [14]. authors, as discussed in the following paragraphs. Liu et al.
Stamatiou and Haenggi [15] investigated the scenario [18] derived analytic expressions for the mean delay and the
described above, by combining the use of the stochastic throughput of the channel, but their solutions are not in closed
dominance technique with stochastic geometry and queueing form1 [19]. Also, they considered static PPPs and, in this
theory results. They studied the stability and the delay of case, it is not possible to find the conditions for network
random networks, where terminals are located according to stability, since for almost2 all PPPs there is a subset of
a Poisson point processes (PPP), and obtained necessary and unstable users, i.e., users with a transmission distance greater
sufficient conditions for stability in a network with one or than a critical value, such that their queues are unstable.
two classes of users. Afterwards, many other works explored Zhong et al. [20] also studied the problem of stability
this combination between queueing theory and stochastic in Poisson networks under random access, but for a static
geometry to study random networks (see, for instance, [16]– network. The authors introduce the concept of -stability,
[18] and references therein). according to which a network is said -stable when the
The present work extends the results presented by Sta- portion of unstable queues in the network is less than .
matiou and Haenggi [15], generalizing the formulation that The necessary condition and the sufficiency condition for -
describes the behavior of users in a random network with N stability are then determined. However, the authors do not
classes of users. We consider that each transmitter (generated analyze the queueing delay.
by a “mother” PPP) chooses the closest terminal within the Another work that explores this combination of stochastic
pool of potential receivers generated by its “son” PPP to geometry and queueing theory was presented by Gharbier et
transmit its packet, as illustrated in Fig. 1. We establish the
1 In this present paper, a closed form expression is defined as a finite com-
necessary and sufficient conditions relating user densities,
bination of elementary functions, which are limited to sum, multiplication,
transmit power levels and traffic intensities that ensure sta- exponentiation and their inverses in the field of complex numbers C.
bility for the terminal queues of all classes. For the case 2 Almost in the sense of having probability measure one.

2 VOLUME X, 2019

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

al. [21], where the authors compare the performance (in terms A. MAIN CONTRIBUTIONS
of transmission success probability, buffer queue length, ac- The main contributions of our work when compared to the re-
cess delay time, and scalability) of scheduled protocol with sults presented in the aforementioned papers, particularly the
random access protocol for the uplink in the context of IoT. one presented by Stamatiou and Haenggi [15], are twofold.
The conclusion is that the best performing protocol depends Firstly, we have extended the analysis of stability and delay
on the operational scenario regarding device densities and in random-access wireless network to the case of a network
traffic rates. with an arbitrary number N of user classes. Secondly, we
Gharbieh et al. [22] investigated the delay and stability have expanded this analysis to show that the channel sharing
in cellular wireless networks where base stations serve IoT mechanism in the investigated scenario can be seen as a
devices, which are modeled as queues with geometric arrival. process of partitioning a fixed and well-defined quantity into
The performance of power-ramping and backoff transmis- portions, each portion allotted to each user class, the size of
sion strategies was studied, by modeling the time evolution which varying in accordance with the user class parameters.
of queues using discrete time Markov chains. Due to the More specifically, the contributions of the paper are:
interaction among queues, caused by mutual interference,
the proposed Markov chains were solved using iterative • We propose a tractable scenario to study the perfor-
algorithms. mance and stability of a bipolar network with an arbi-
In another related work, Zhong et al. [23] studied the delay trary number N of classes of users sharing the same
in wireless networks, considering however the downlink of a channel; a simple and elegant expression relating mean
heterogeneous cellular network, with K tiers of base stations. delays, arrival rates, user densities, mean link distances
Downlink traffic arrives at the base stations in bursts, to be and bit rates of all N classes is derived for the case of
delivered to users, which are spatially distributed according stable network. This expression clearly shows that each
to a Poisson point process. Each user has a delay require- class of user takes a well-defined portion of the available
ment in terms of maximum acceptable mean packet delay. finite resource in the RF channel (Proposition 2).
Different scheduling mechanisms are investigated, including • a closed form [19] solution to the fixed-point system
random access, FIFO and robin-round scheduling. of equations that determine the stationary transmission
Yang and Quek also investigated the downlink perfor- success probabilities for N user classes is found;
mance of a wireless network composed of user terminals • an intuitive equation is presented relating link quality,
served by small access points (SAP) [24]. The locations of packet arrival rate, density of users and stationary mean
SAPs and user terminals are modeled by means of Poisson delay (Proposition 1);
point processes, and SAPs are equipped with a set of queues, • we prove the necessary and sufficient conditions that
each one dedicated to a user terminal. The SAPs share the determine whether a given network is stable (Theorems
same downlink channel, causing mutual interference among 1 and 2);
transmissions, which in turn results in a interacting queues • we establish a simple necessary condition for stability
scenario. The main focus of the paper is on the effects that does not depend on the transmit powers (Corol-
of spatial geometry of interfering terminals and the traffic lary 1);
intensity on the SAP coverage. Numerical results show that • the optimum transmit powers per user class that achieve
higher traffic intensity requires lower SINR threshold if the the optimum stationary mean delays for each user class
coverage is to remain unchanged. (Proposition 3) are derived;
Chen et al. [25] studied delay and throughput in a cognitive • the optimum packet arrival rates per user class that
radio network, in which secondary users share the channel achieve the maximum channel throughput per unit of
with a single primary user. Secondary users are allowed to area (Proposition 4) are derived;
access the channel with a probability that depends on the • we conclude that depending on the channel and user
length of the queue of the primary user. Using results from classes, the best strategy to maximize channel through-
queueing theory and stochastic geometry, the authors devel- put is to share the channel, instead of using one single
oped an analytic framework to investigate the relationship class per channel;
between the delay of the primary users and the throughput The paper is organized as follows: Section II describes
of the secondary network. Optimization problems are then the model used throughout the paper and provides some
proposed and solved to maximize the secondary network important results from the literature to be used in the fol-
throughput under delay constraints for the primary users. lowing sections; Section III presents the main results of the
Following a different approach from the ones aforemen- paper, i.e., necessary and sufficient conditions for stability
tioned, Kountouris et al. investigated the delay in wireless when we have N interacting traffic classes, it also shows
networks using tools from stochastic network calculus [26]. a simple expression for the stationary mean delay and the
For a static Poisson network, they derived bounds on the packet success probability; Section IV applies the obtained
delay violation probability and on the effective capacity results in two general scenarios: one scenario optimizes the
distribution, using network calculus. transmission power of different traffic classes with different
delay requirements sharing the same channel and the other
VOLUME X, 2019 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

TABLE 1. Notations and symbols used in the paper its queue length {Qi n (t)} of infinite capacity. The packet
arrival probability at each queue is denoted by an and the
Symbol Definition/explanation medium access probability by pn . Within each slot, the first
Z, Z+ set of integers and non-negative integers
R, R+ set of real numbers and non-negative reals event to take place for each TX with a non-empty queue is the
α ∈ (2, ∞) path loss exponent medium access decision with probability pn . If it is granted
δ ∈ (0, 1) , 2/α access and the signal to interference ratio (SIR)4 is greater
N ∈ Z+ \{0} number of traffic classes
than a threshold θn > 0, a packet is successfully transmitted
N , {1, 2, . . . , N }
n∈N refers to the n-th traffic class and leaves the queue. Then, we have the arrival of the next
pn ∈ (0, 1) medium access probability packet with probability an . The last event to take place is the
an ∈ (0, 1) packet arrival rate per time slot displacement of the transmitters and destinations.
a ∈ (0, 1)N = (a1 , a2 , . . . , aN )
ps,n ∈ (0, 1) packet success probability The queue lengths of the i-th TX, traffic class n are
θn ∈ R+ SIR threshold for successful communication Markov Chains represented by
R n ∈ R+ average transmission distance
Dn ∈ (1, ∞) average packet transmission delay Qi n (t + 1) = (Qi n (t) − Bi n (t))+ + Ai n (t), t ∈ N,
D ∈ (1, ∞)N = (D1 , D2 , . . . , DN )
P n ∈ R+ transmission power where (·)+ , max{·, 0}, {Ai n (t)} are iid Bernoulli random
P ∈ RN + = (P1 , P2 , . . . , PN ) variables of parameter an and represents the arrival process,
Φn ⊂ R2 Poisson point process for the transmitters
λ n ∈ R+ density of Φn Bi n (t) = ei n (t) 1SIRi n >θn
2 δ
φ n ∈ R+ , 4 Rn θn πδ/ sin(πδ)
|| · || euclidean norm represents the departure process, where {ei n (t)} are iid
1{·} indicator function Bernoulli random variables of parameter pn , the constant θn
represents the SIR threshold for successful communication,
and the SIR of user i and traffic class n is given by
optimizes the throughput per unit of area; Section V con-
cludes the paper. Pn hi n,i n ||Xi n − Yi n ||−α
SIRi n = X ,
The notations used in the paper are summarized in Table 1. Pk hj k,i n ej k 1Qj k >0 ||Xj k − Yi n ||−α
(j,k)6=(i,n)
II. PRELIMINARIES (1)
A. SYSTEM MODEL where the dependence on t has being omitted, {hj k,i n (t)}t
We consider a network composed by N classes of users are iid exponential distributed random variables of parameter
that share the same radio resources. Time is slotted and for one and represent the Rayleigh fading coefficient from the j-
each time slot t ∈ N and each traffic class n ∈ N , th TX of class k to the i-th RX of class n. Coefficient α > 2
{1, 2, . . . , N }, we have a homogeneous Poisson point pro- is the path loss exponent.
cess (PPP) denoted by Φn (t) ⊂ R2 of density λn , which
represents the position of the sources. These PPP are inde- B. DEFINITIONS AND RESULTS FROM THE
pendent from each other and from the past. Each transmitter LITERATURE
(TX) of traffic class n transmits with power Pn . The position This subsection is devoted to present the usual definitions
of the transmitters are given by {Xi n (t)}i , i ∈ N, i.e., found in the literature and some known results that are useful
Φn (t) = {Xi n (t)}i . More precisely, for each time slot the to prove the main results of the paper. Let us start by defining
position Xi n (t) of the i-th TX is reallocated following the the concept of stability in a queueing system. Let us use the
high-mobility random walk model [27]. The i-th TX of traffic definition proposed by Szpankowski [14], presented next.
class n communicates with a receiver (RX) located at Yi n (t).
Thus, the distance between the i-th transmitter of class n and Definition 1. A queue Q(t) is stable if for x ∈ Z+
its destination is given by Ri n (t) = ||Xi n (t) − Yi n (t)||. lim P(Q(t) < x) = F (x) and lim F (x) = 1,
t→∞ x→∞
In this work, we assume that each transmitter is associated
to a “son” PPP that models the locations of its potential where F : R+ −→ [0, 1] is the limiting distribution function.
receivers. The receiver associated to the i-th transmitter of
The system network is stable when the queue of the typical
class n is chosen as the closest point in the respective son
user from n-th class is stable for all n ∈ N . The stability
PPP to the point Xi n (t). As a consequence, {Ri n (t)}t are iid
region R is given by the set of arrival rates a ∈ [0, 1)N that
Rayleigh random variables3 [28, Eq. (2.35)]. Rayleigh dis-
makes the system network stable.
tributed TX-RX separation distance has been used in several
Throughout the paper, stationary success transmission
other works investigating similar scenarios (e.g., [29]–[31]).
probability ps,n refers to the limiting probability of a suc-
We denote the mean transmission distance E[Ri n (t)] by Rn .
cessful transmission from a typical user of class n, i.e.,
The occupation of the buffer at each TX is represented by
3 The
ps,n , lim P(SIRi,n (t) > θn ).
iid random variables for the TX-RX separation distance are of grave t→∞
importance for the theoretical model. Otherwise, there is at least one unstable
queue and, consequently, the queueing network is unstable. 4 We assume thermal noise is negligible; refer to [32] for further details.

4 VOLUME X, 2019

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

This probability does not depend on the user i, by symmetry. traffic class n ∈ N are given by
The stationary mean delay Dn to transmit packets of class n P δ !−1
is defined as the limiting (as t tends to infinity) expected time φn j Pj aj λj
ps,n = 1 + δ P , (4)
a packet spends in the buffer and the server. Pn 1 − j φj aj λj
The following results from the literature are used in many 1 − an
proves throughout the paper. In a wireless network, let us Dn = , (5)
pn ps,n − an
assume that (i) the separation distance between a given pair
TX - RX is equal to r, (ii) the positions of the interferers where the sums are taken over the set of user classes N ,
2
(users who will transmit packets in a given time slot) follow δ , 2/α and φn , 4 Rn θnδ πδ/ sin(πδ).
a PPP of density λeff , and (iii) every transmitter has the Proof. The buffer (plus server) is a discrete time Geo/Geo/1
same transmit power. Then, the probability of a successful queue [15] and the equation for the delay Dn is found in
transmission between TX and RX is given by [33, Sec. III.A] the literature [35, Chapter 4.6]. As t → ∞, the effective
(n)
PPP density of active sources λeff for each traffic class
α
P(SIR > θ) = E[e−θr I ] n ∈ N converges (by hypothesis) to λn pn ρn , where ρn =

an /(pn ps,n ) is the load of the queue (or the probability of
= exp −π Γ(1 + δ)Γ(1 − δ) θδ r2 λeff , (2)
having a non-empty queue), which is the ratio between the
(n)
P arrival rate and the service rate of packets. Thus, λeff =
where δ , 2/α and I , X∈Φ ||X||−α is the interference λn an /ps,n .
received by RX normalized by the transmit power and Φ is a Then, to calculate the transmission success probability
PPP of density λeff , which is the effective density of active ps,n , we use (3), which assumes that the link distance
sources. This result can be generalized for the case of N between TX and RX is constant. Thus, by deconditioning
classes of interferes, each class n ∈ N with an effective the transmission success probability on Ri n , we take into
(n)
density λeff and transmit power Pn . The proof with two user account that Ri n is Rayleigh distributed, that is
classes was presented by Yin et al. [34, Proposition 3] and
it is easily extended to N user classes as follows. Let the ps,n = lim P(SIRi n (t) > θn )
t→∞
Z ∞
analyzed transmitter be from class k ∈ N , then
= lim P(SIRi n (t) > θn | Ri n (t) = r) fRn (r) dr
0 t→∞
Z ∞ ( !)
" !# πr π r2 X P δ (k)
k
θ rα X = 2 exp − 2 1 + φn λ
P δ eff
dr
P(SIRk > θ) = E exp − Pn In 0 2Rn 4Rn k∈N n
Pk !−1
n∈N
Y    φn X δ ak λk
α Pn = 1+ δ Pk . (7)
= E exp −θ r In Pn ps,k
Pk k∈N
n∈N
Y   This expression can be rearranged as
πδ Pnδ (n)
= exp −πr2 θδ λ   X
sin(πδ) Pkδ eff Pnδ 1 − ps,n ak λk
n∈N
! = Pkδ . (8)
X P δ (n) φn ps,n ps,k
2 δ πδ n k∈N
= exp −πr θ λ ,
sin(πδ)
n∈N k
P δ eff Note that the right-hand side does not depend on n. Then, for
(3) all j ∈ N , we can write
   
Pjδ 1 − ps,j Pnδ 1 − ps,n
where we used Euler’s reflexion formula Γ(1 + δ)Γ(1 − = . (9)
φj ps,j φn ps,n
δ) = (πδ)/ sin(πδ) and In is the interference from the n-
th class normalized by the transmit power. It is assumed that For each j, we can solve the above equation for ps,j and plug
{In }n∈N is iid. it into the sum of (7). Then, we can solve it for ps,n , which
ends the proof.

III. MULTIPLE-CLASS NETWORK The following theorem shows the conditions for which the
As described in Section II, we consider a network with network is stable, i.e., it presents the region formed by all
N classes of users. The following proposition presents the arrival rates a that make the system stable.
stationary success probability and mean delay when transmit- Theorem 1. A necessary and S sufficient condition for the
ting a packet in a stable network. The results that guarantee network stability is that a ∈ ν∈V Cν , where V is the space
stability are presented later in the paper, in Theorem 1. of all bijective functions from N to N (or
P0permutations) and
Cν is defined in (6) with the convention k=1 · = 0.
Proposition 1. If the network is stable, then the stationary
success probability and mean delay for a typical user of Proof. See Appendix B.
VOLUME X, 2019 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

( Pn−1 )
φν(n) aν(n) 1 − k=1 φν(k) aν(k) λν(k)
Cν , a ∈ [0, 1)N 0≤ δ < Pn−1 PN ∀n ∈ N (6)
Pν(n) pν(n) − aν(n) δ
k=1 Pν(k) aν(k) λν(k) +
δ
k=n Pν(k) pν(k) λν(k)

Multiplying both sides of the above equation by Pn an λn and

summing over all n ∈ N result in
X pn an
φn λ n <1
pn − an
n∈N

0.2 after some manipulations. Thus, a ∈ S0 .

Now, let us show that S0 ⊂ R for some P ∈ RN + . In
a3

0.1 particular, let us choose Pn = φn an /(pn − an ), n ∈ N .

Then, the inequalities that describe the region R can be
0 rewritten as one unique inequality
0 0 P
1 − k φk ak λk
0.1 0.1 1< P 2
k φk ak λk /(pk − ak )
0.2 0.2
a2 a1 that does not depend on n anymore. It is easy to show that
0.3
this inequality is the same as the one that defines the region
FIGURE 2. Stability region R according to Theorem 2 for p1 = 1/3, S0 . Thus, if a ∈ S0 , then a ∈ R for that choice of P (or any
p2 = 2/3, p3 = 1, φ1 λ1 = 1, φ2 λ2 = 2, φ3 λ3 = 3, φ1 /P1 = 1/3, scalar multiple). This ends the proof.
φ2 /P2 = 1/2, φ3 /P3 = 1.

Theorem 1 requires verifying N×N ! inequalities, whereas

the following theorem is equivalent and it has N inequalities.
Thus, Theorem 2 presents a simpler form of verifying the
conditions for stability.
Theorem 2. The system network is stable iff a ∈ R, where 0.2
a3


0.1
R , a ∈ [0, 1)N an < pn ,
P 
φn an 1 − k φk ak λk 0
< P ∀n ∈ N . 0 0
Pnδ pn − an δ
k Pk ak λk
0.1 0.1
S
Proof. See Appendix C. We show that R = ν∈V Cν .
0.2 0.2
a2 a1
0.3
Figure 2 shows an example of stability region for N = 3
FIGURE 3. Maximum stability region S0 according to Corollary 1 for
classes, where we only used 3 non-linear inequalities instead p1 = 1/3, p2 = 2/3, p3 = 1, φ1 λ1 = 1, φ2 λ2 = 2, φ3 λ3 = 3.
of 18. The following corollary establishes a simple result on
stability, which is used in Section IV to propose and solve Figure 3 shows the region of arrival rates, according to
optimization problems regarding delay and throughput. Corollary 1, for which it is possible to find transmit powers
that make the network stable. On the other hand, out of this
Corollary 1. There exists a vector of transmit powers
region, the system is always unstable. It is worth mentioning
P ∈ RN+ such that the network is stable iff a ∈ S0 , where
that φn is related to the quality of the link between receiver
( ) and transmitter for class n ∈ N ; the larger is the value of φn ,
X φn λ n
N
S0 = a ∈ [0, 1) 0≤ 1 <1 . the poorer is the quality of the link. Also, the stability region
n∈N an
− p1n R showed in Fig. 2 is contained in S0 . This is expected, since
we used the same parameters for both sets and Corollary 1
Proof. First, let us show that R ⊂ S0 for all P ∈ RN
+ . If considers the best case scenario, where we can choose the
a ∈ R, then for all n ∈ N transmit powers P for each a.
P From now on, we assume that whenever there is a packet in
φn an 1 − k φk ak λk
< P δ . the buffer, the corresponding TX attempts to transmit, i.e., the
Pnδ pn − an k Pk ak λk medium access probability pn = 1 for all n ∈ N . We discuss
6 VOLUME X, 2019

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

in Appendix A the validity of this assumption and the high- available to all users of the channel. Larger α results in
mobility assumption. The motivation is that, when the access higher sin(2π/α)
2π/α , meaning that a larger amount of resource
probability of all classes is equal to one, we maximize the is available to users. This can be explained recalling that
stability region R. This is easy to see with the inequalities larger path loss exponent leads to stronger isolation among
of Theorem 2, where the right-hand side does not depend on links sharing the channel and, consequently, more users can
pn and the left-hand side decreases monotonically with pn . be accommodated in the network. Therefore, the larger the
Thus, the stability region is maximized when pn = 1 for path loss exponent α, the larger (smaller) the terms λn , Rn ,
all n ∈ N . The same occurs in Corollary 1. This result is θn , an (Dn ) can be. The identity (11) also tells us that the
surprising and might be explained from the fact that we have n-th class of user takes a well-defined portion of the amount
independence between adjacent time slots and, therefore, for of resource available in the network, which is given by the n-
each time slot there is a new scenario (a new effective PPP). th term in the summation. This means that the values of λn ,
Then, it makes sense to always try re-transmission. This Rn , θn , an and Dn for a given class n can be adjusted, while
2 2/α
approach also minimizes the mean delay according to (5), keeping the quantity λn Rn θn DD n an
n −1 1−an
unchanged. For
since the success probability ps,n in (4) does not depend on instance, we can make a direct exchange between decreasing
the access probability in a stable network. the delay Dn and decreasing the arrival rate of packets an
Using Proposition 1, Proposition 2 is proposed, which (by controlling the ratio of transmit power levels), such that
presents an equation that relates all performance parameters the term DD n an
n −1 1−an
remains constant; or else, increase the
independently of the transmission powers. Also, the condi- arrival rate of packets and decrease the density of users, such
tions for stability are extensively simplified, see Corollary 1. an
that the term λn 1−a n
remains constant. Therefore, Propo-
Proposition 2. If the network is stable and pn = 1 for sition 2 reveals, through a simple expression, the interplay
all n ∈ N , then the following identities hold (at stationary among traffic intensity, mean delay, density of users, link
state): distance, and outage probability, when the network is stable.
X Dn an
φn λn = 1, (10)
Dn − 1 1 − an IV. INTERPRETATION AND APPLICATION
n∈N

and In this section, we solve two optimization problems using

    the proposed formulation, applied to scenarios of different
φj Dj 1 φk Dk 1
Pjδ Dj −1 1−aj −1 = Pkδ Dk −1 1−ak −1 ∀ j, k ∈ N . classes of terminals sharing a radio channel.
Proof. We start with the terms of the sum,
Dn an (i) an A. DELAY OPTIMIZATION
φn λn = φn λ n
Dn − 1 1 − an 1 − ps,n Let us consider the scenario with N classes sharing a chan-
 
δ λn an φn ps,n nel. Each class may represent a particular user application,
= Pn with each application having a different delay requirement in
ps,n Pnδ 1 − ps,n
the network. Let us suppose we are interested in adjusting the
(ii) Pnδ λpns,n
an
= P , transmit power of each traffic class, such that the weighted
δ λj aj
j Pj ps,j average delay among all classes is minimized. This problem
is addressed as follows. For fixed arrival rates of vector a
where (i) comes from (5) with pn = 1 and (ii) comes from that satisfies Corollary 1, i.e., for a ∈ S0 , let us minimize
(8). Summing over N ends the proof of the first identity. For the delays D by changing the ratio between the transmit
the second relation of Proposition 2, we use (5) once again to powers P . Each traffic class requires a different response
find time, then we weight the optimization problem with the
 
φn Dn 1 φn ps,n vector (c1 , c2 , . . . , cN ) ∈ RN
+ . The larger the coefficient of
−1 = δ .
Pnδ Dn − 1 1 − an Pn 1 − ps,n a class, the smaller the mean delay to deliver packets for that
Comparing this expression with (9) ends the proof. class. Then, we have

Proposition 2 is an elegant form to see that a channel is a X X cn (1−an )

limited resource regarding traffic intensity and delay. Let us min cn Dn = min P δ
!−1 ,
P ∈RN P ∈RN φn j Pj aj λj
rewrite the identity (10) in terms of physical parameters, + n∈N + n∈N 1+ δ 1−P φ a λ
Pn j j j j
−an

N
X (12)
2 Dn an sin(2π/α)
4 λn Rn θn2/α = . (11) where Dn is given by Proposition 1. Note that as thermal
n=1
Dn − 1 1 − an 2π/α noise is not considered in our model, we have a degree of
freedom for the optimum solution P ∗ , which agrees with the
an
Note that 1−a n
and sin(2π/α)
2π/α are monotonic increasing func- formulation in (12).
tions and DD n
n −1
is a monotonic decreasing function. The right
hand-side of (11) can be seen as the amount of resource Proposition 3. The minimum of the optimization problem
VOLUME X, 2019 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

102

Normalized transmit power per unit of area

∗
D1∗ − 1 λ1 P1δ
∗
D2∗ − 1 λ2 P2δ
50%
D∗ − 1 δ∗
101 P 3 ∗ λ3 P3
Mean waiting time [slots]

(c D − 1)

40%

100
30%

10−1
20%

10−2 10%
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Arrival Rate a Arrival Rate a
(a) Optimum delays (b) Optimum transmit powers

FIGURE 4. These figures represent the optimization of a 3-class network with the following parameters: a1 = a2 = a3 = a, φ1 λ1 = 0.1, φ2 λ2 = 0.3,
10 5 1
φ3 λ3 = 0.6 and c1 = 16 , c2 = 16 , c3 = 16 .

(12) is attained by As an example, let us consider a 3-class network, where

 √  Class 1 has a more restrictive delay requirement than Class
∗δ β an An cn An
Pn = P +P √ , n ∈ N, 2, which is more restrictive than Class 3. We consider that all
λn an 1 − k Ak k ck Ak classes have the same arrival rate of packets, i.e., a1 = a2 =
(13)
an a3 = a. Figure 4(a) shows the expected waiting time of a
where β is any positive real constant, An , φn λn and
1−an packet before a successful transmission, which is Dn∗ − 1,
the sums are over N .
(n = 1, 2, 3) since a transmission takes exactly one time
Proof. Since weP have one degree of freedom for the solution slot. As expected, the optimization resulted in monotonic
P
P ∗ , let us set j Pjδ aj λj = 1 − j φj aj λj to exten- increasing functions and D1∗ < D2∗ < D3∗ for all a.
sively simplify the algebraic manipulations. Then, we use Figure 4(b) shows the normalized5 transmit powers per
the Karush-Kuhn-Tucker conditions [36, Section 3.3.1] in the unit of area λn Pn∗ δ as a function of a. In this case, we
Lagrangian function do not have a clear hierarchy among the transmit powers,
X cn (1 − an ) P For n ∈ N , if the
as it depends on the traffic intensity.
L(P , µ) =  −1 network is close to saturation, i.e., k A Pk tends to 1, then
φ
n∈N 1 + nδ
Pn
− an the normalized λn Pn∗ δ approaches An / k Ak and, at first
   order, it does not depend on the coefficients c1 , c2 , . . . , cN .
X X On
+ µ Pjδ aj λj − 1 − φj aj λj  , P the other hand, if the network is at low ∗δ
traffic, i.e.,
√ k A k tends
P √ to 0, then the normalized a n λn Pn approaches
j∈N j∈N
cn An / k ck Ak .
where µ ∈ R is the Lagrange multiplier. The objective
function is strictly convex (the Hessian is a diagonal matrix B. THROUGHPUT OPTIMIZATION
with positive eigenvalues) and the feasible region is a hyper- Now, let us maximize the total throughput of the channel
plane, therefore the solution is the global optimum. Now, we per unit of area with the constraint that the system is stable.
return to the original problem that does not have the artificial Since each TX performs re-transmissions until the packet
constraint. Thus, we multiply the solution by an arbitrary is correctly received by RX, then all packets are success-
constant β > 0 to obtain the general solution. fully transmitted (eventually) in a stable system. Thus, the
P throughput per TX is given by the packet arrival rate an . The
It is interesting to note that we must have k Ak < 1,
density of users per unit of area is given by λn , then the
by Corollary 1. Therefore, Pn∗ δ is always a positive
throughput of the n-th user class per unit of area is simply
quantity. Also, if cn = An for all n ∈ N ,
λn an and the throughput per unit of area of the entire system
then the optimum delays arePall equal and given by
−1
D1 = D2 = · · · = DN = (1 − k Ak ) . Thus, we can al- 5 Whenever we refer to normalized f P ∗ δ , it means that we choose β in
n n
ways choose transmit powers, such that we have the same Proposition 3 such that k fk Pk∗ δ = 1 and fn is any function that depends
P
mean delay for all classes! on n, for example fn = λn or fn = an λn .

8 VOLUME X, 2019

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

0.4 a∗1 λ1 a∗1

a∗2 λ2 a∗2

Optimum throughput per area

a∗3 1.5 λ a∗
P3 3∗
λa
Optimum arrival rates

0.3 one-class

1
0.2

0.5
0.1

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
 
(a) Optimum arrival rates (b) Optimum throughput. The dashed curve corresponds to the scenario of
using only the best performing class

102 Normalized transmit power per unit of area 100%

D1∗ −1 λ1 P1δ
∗

D2∗ −1 ∗
λ2 P2δ
D3∗ −1 80% ∗
λ3 P3δ
Mean waiting time [slots]

101
60%

40%
100

20%

10−1 0%
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
 
(c) Optimum delays (d) Optimum transmit powers

FIGURE 5. These figures represent the optimization of the throughput and mean delay of a 3-class network with the parameters given in Table 2.

is the sum of the throughput for all classes n ∈ N . Using As the parameter  increases, the system becomes less sensi-
Corollary
P 1, we can formulate the optimization problem as ble to perturbations6 . Now, the optimization problem is posed
max n λn an . However, the strict inequality in the region as X
a∈S0
S0 of Corollary 1 results in an optimization problem that is max λn an . (14)
a∈S
not well-posed. In this case, if the optimum solution lies in n∈N

the boundary of the feasible region, then the solution does not The following proposition presents the solution to the opti-
exist. To circumvent this problem, we propose a new region mization, i.e., the optimum arrival rates a∗ that maximize the
S ⊂ S0 by adding an arbitrarily small parameter  ∈ (0, 1) throughput per unit of area and maintain the system stable.
in the inequality, i.e., the new region is given by
Proposition 4. If
X q 
maxn φn
λk φk φk −1 < 1 − , (15)
k∈N
( )
X an
N 6 When the system parameters suffer a sufficiently small change, the
S = a ∈ [0, 1) φn λ n ≤1− .
1 − an system remains stable
n∈N
VOLUME X, 2019 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

then the solution of (14) is attained by bad link quality, such that it is better (regarding throughput
P √ efficiency) to reallocate this user class to another channel.
∗ k λkPφn φk
an = 1 − , n ∈ N, (16)
1 −  + k λk φ k Through Proposition 3 we find the best configuration of
where the sums are over N . If the inequality (15) is not transmit powers per unit area that minimizes the sum of
satisfied, then the m-th class is excluded, where m = the mean delays. Figure 5(c) shows the result of this opti-
arg maxn φn , and the inequality is checked again. mization, which is a direct application of (13). It is worth
noting that as we increase  the system is farther from
Proof. It is a direct application of the Karush-Kuhn-Tucker instability, which corresponds to having a smaller delay to
conditions [36, Section 3.3.1] in the Lagrangian function transmit packets, as we can see in Fig. 5(c), and a smaller
" #
X X throughput, as shown in Fig. 5(b). Figure 5(d) shows the
ak
L(a, µ) = λn an + µ φk λk − (1 − ) , optimum distribution of power per unit of area required by
1 − ak
n∈N k∈N each user class. Notice that the first traffic class, which has
where µ ∈ R is the Lagrange multiplier associated with the best link quality uses the smallest power per unit of area.
the constraint of stability. Equation (15) guarantees that the However this behavior is more intricate; it also depends on
solution a∗ ∈ [0, 1)N . The objective function is convex the density of users of the corresponding class. Notice, for
(affine function) and the region S is strictly convex (the example, the inversion between user class 2 and 3 as we
Hessian of the function that defines the region is a diagonal increase  in Fig. 5(d).
matrix with negative eigenvalues), therefore the presented
solution is the global optimum and it is unique.
In this system, we still have freedom to choose the transmit V. CONCLUSIONS
powers P , as long the network remains stable. The best way
In this paper, it is proposed a modified model to study the
of choosing P is by minimizing the delays, which we have
stability and delay of slotted Aloha in high-mobility Poisson
already done in Subsection IV-A, where the arrival rates a ∈
bipolar networks. The model presented in the paper considers
S ⊂ S0 are fixed. When the optimization is performed in this
a scenario where transmitters with buffer communicate with
sequence (maximization of throughput, then minimization of
the closest receiver belonging to a son Poisson point process.
delay), we have the optimum throughput (per unit of area)
We derived necessary and sufficient conditions for stability
and the optimum delays for each class. Later in this section,
in a network with N user classes; we also provided simple
we illustrate this procedure with a numerical example.
closed-form expressions for the packet success probability
In order to solve the optimization problem (14) we did not
and mean delay. As shown by the results in the paper, the
have to handle with the transmit powers P directly, which
advantage of using this model as a base to model other
would make the solution and the problem formulation more
network effects is its analytic tractability. As an example, we
cumbersome. This shows the usefulness of Corollary 1.
were able to derive simple conditions to verify the stability
Let us illustrate the optimization problem with a system for
of a interference-limited network with undetermined transmit
which the parameters are shown in Table 2. Figure 5(a) shows
powers (see Corollary 1). We also solved (analytically and
the optimum arrival rates a∗ that maximizes the throughput
in closed form) two optimization problems regarding the
per unit of area. It is quite interesting that the optimum
minimization of the delays in a network (see Proposition 3)
TABLE 2. Network parameters for Fig. 5
and maximization of the total throughput per unit of area (see
Proposition 4). An interesting insight from the optimization
Parameters Values problem is that the best solution to maximize the throughput
(λ1 , λ2 , λ3 ) = (1, 2, 3) of a channel is not necessarily using solely the traffic class
(φ1 , φ2 , φ3 ) = (0.3, 0.5, 0.4) with the best link quality, i.e., a mix with other traffic classes
(c1 , c2 , c3 ) = ( 13 , 13 , 13 )
may result in a better use of the channel. All in all, this paper
provides a simple way to evaluate the existing trade-offs
solution is not necessarily using solely the class with the involved in the design of wireless networks when different
best link quality (i.e., the class with the smallest φ, which is classes of nodes co-exist. In the future, we plan to further
Class 1 in this example). In Fig. 5(b) it is shown the optimum extent this analysis to more practical machine-type commu-
throughput for each class and the total throughput of the nications (5G) scenarios, including long range technologies
system. For comparison, we plotted a dashed curve represent- and transmission of critical messages.
ing the total throughput if we use only the best performing
user class, regarding throughput. The dashed curve is below Another interesting and direct result from Corollary 1 is to
the optimum total throughput for all . Therefore, the best provide an upper bound for the total throughput per unit of
solution is always a combination of all traffic classes, as long area in a stable system, which is given by 1/(minn φn ). This
as (15) is satisfied. On the other hand, if this equation is not is not a tight bound, however it is interesting on its own, due
satisfied, it means that there is at least one user class with a to its simplicity and the fact that it does not depend on the
10 VOLUME X, 2019

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

density of users. The proof follows is important to emphasize that we did not plot the mean delay
X 1 X D, because in a static PPP there might exist a set of unstable
ak λk ≤ φk ak λk users, whose queues and delays tend to infinity. This would
minn φn
k∈N k∈N raise the average delay to infinity too. Then, we chose to plot
1 X ak
< φk λk the mean load ρ, which is equal to 1 for unstable users and
minn φn 1 − ak does not tend to infinity as the mean delay D.
k∈N
−1 To establish Proposition 2, we suppose that the access
< (minn φn ) , (17)
probability p is equal to one for all users. In the context of
where the last inequality comes from Corollary 1. high-mobility this approach makes sense, since the typical
. user sees a different interference scenario for each time slot.
Thus, it makes sense to attempt a re-transmission every time
APPENDIX A HIGH-MOBILITY ASSUMPTION slot until the packet is successfully transmitted. This also
In this appendix, we address the high-mobility assumption, minimizes the mean delay D, which is in accordance with
which may not be realistic in real wireless networks, since the (5), as ps,n does not depend on pn in a stable network.
mobility of transmitters does not change drastically between There is another scenario, which does not require high-
adjacent time slots. Therefore, the independence assumption mobility to achieve spatial independence between adjacent
would not hold. Nevertheless, in a stable wireless network time slots. This scenario is a network that uses the frequency-
which has a small packet arrival rate a per user or a small hopping scheme over a set of channels [37]. For each time
access probability p, the correlation might be sufficiently slot there is a different PPP pattern, since the transmitting
small such that the independence (high-mobility) assumption nodes select with equal probability one channel to transmit.
is reasonable. In [29] the authors show that if the access Thus, the spatial correlation between time slots decreases
probability p is sufficiently small, then the independence with the number of channels available for selection.
assumption provides a good approximation.
When the packet arrival rate a or the access probability APPENDIX B PROOF OF THEOREM 1
p are small, the typical user sees a significantly different Proof. Using the concept of stochastic dominance [38, Sec-
PPP of transmitters, which justifies the independence (high- tion 2.1.2], it is possible to derive necessary and sufficient
mobility) assumption. We verified this claim through simula- conditions for stability. In the dominant network, all the
tions and the result is shown in Fig. 6, where it is used one traffic classes in the set D ⊂ N transmit dummy packets.
2
traffic class with λcR = π/4, α = 3 and θ = 1. The mean If the dominant network is stable, then the original network
is stable. On the other hand, if the queues of the traffic
1 classes in D are not empty in the original network, then
this system behaves exactly as the dominant network (both
a = 0.30
systems are indistinguishable [14, Section 3.2]). Therefore, if
0.8 = 0.20
the dominant network is unstable, then the original network
= 0.10
will be unstable as well. In order to have necessary and
= 0.05 sufficient conditions, we must perform this verification for
ρ = P(Q > 0)

0.6 all D ⊂ N .
Let us start with D = N , i.e., all users transmit dummy
packets. For each step of the verification, we remove the
0.4
stable traffic class from the set D. This procedure repeats
until the set D becomes empty. In order to attain stability
of the dominant network we must have an arrival rate smaller
0.2
than the service rate [39]. Thus, a sufficient condition for the
first traffic class stability is, for any queue i of this class (by
symmetry),
0
0 0.2 0.4 0.6 0.8 1
!−1
p N
g i,1 > θ1 ) = p1 φ1 X δ
a1 < p1 P(SIR 1+ δ P pk λk ,
FIGURE 6. Queue load ρ as a function of the access probability p. Simulation P1 k=1 k
results with a static network are presented in marks and the theoretical results
with the high-mobility assumption are presented in curves.
g represents the signal-interference ratio in the
where SIR
load of the queues ρ, which is equivalent to the percentage dominant network and the second equality comes from the
(n)
of queues with packets to transmit, are plotted as a function same procedure to obtain (7) with λeff = λn (all users
of the access probability p for several values of arrival rate are active, since every TX transmits dummy packets). This
a. As expected, for small values of a or p, the theoretical guarantees stability for the first traffic class. Let us remove
model presents good estimations of the average queue load. It it from the set D. Then, we calculate the stationary success
VOLUME X, 2019 11

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

(1)
probability of the first traffic class pes,1 for this dominant APPENDIX C PROOF OF THEOREM 2
network. At steady state, we have Proof. The proof consists of showing that the set R is equal
!!−1 to
S the set defined in Theorem 1. First, let us prove that
XN
(1) φ1 δ a1 δ ν∈V Cν ⊂ R. For that we suppose a ∈ Cν and we show
pes,1 = 1 + δ P1 p1 λ1 + Pk pk λk ,
P1 (1)
p1 pes,1 k=2 a ∈ R for all ν ∈ V by induction. For simplicity of
exposition let us take Cν with ν : n 7−→ n, n ∈ N . We
(1) assume the inequality
which can be solved for pes,1 ,
PN
(1) 1 − φ1 λ 1 a 1 φN −j aN −j 1 − k=1 φk ak λk
pes,1 = PN . < P N
. (19)
1+ φ1 δ PNδ −j pN −j − aN −j δ
k=1 Pk ak λk
k=2 Pk pk λk
P1δ
is true for all j ∈ {0, . . . , m − 1} and we prove that it is also
The next step is to verify the conditions of stability for the
true for j = m. First, we have to prove the base case m = 1.
second traffic class, when the first traffic class is at steady
Since a ∈ Cν , then for j = 0 (n = N )
state. After that, we remove the second traffic class from the
PN −1
set D and calculate the stationary success probability of the φN aN 1 − k=1 φk ak λk
two stable traffic classes in the dominant network. We repeat < PN −1
PNδ pN − aN δ δ
k=1 Pk ak λk + PN pN λN
these steps until we remove all traffic classes, i.e, D = {}. PN
We show this by induction; we suppose stability of the traffic 1 − k=1 φk ak λk + φN aN λN
= PN .
δ δ
classes 1, 2, . . . , j − 1. Let D = {j, j + 1, . . . N }; the j-th k=1 Pk ak λk + PN (pN − aN )λN
traffic class is stable, given that all the traffic classes in N \ D Then, using simple manipulations, we can show that the
are stable, when above inequality is equivalent to
g i,j > θj )
aj < pj P(SIR PN
φN aN 1 − k=1 φk ak λk
  −1 < PN .
φj
j−1
X ak
N
X PNδ pN − aN δ
k=1 Pk ak λk
= pj 1 + δ  P δ λk (j) + Pkδ pk λk  ,
Pj k=1 k pes,k k=j Thus, the base case m = 1 is true. Now, for j = m and
a ∈ Cν , we know that
(18)
φN −m aN −m
(j)
where pes,k is the k-th traffic class success probability (1 ≤ PNδ −m pN −m − aN −m
k < j) at steady state in the dominant network at the PN −m−1
1 − k=1 φk ak λk
j-th step. To calculate this probability, we must solve the < PN −m−1 PN
following system of equations. For k ∈ {1, 2, . . . , j − 1} k=1 Pk ak λk + k=N −m Pkδ pk λk
δ
PN PN
  −1 1 − k=1 φk ak λk + k=N −m φk ak λk
j−1
X XN = PN .
(j) φ k a` δ
PN δ
pes,k = 1 + δ  P δ λ` (j) + P`δ p` λ`  . k=1 Pk ak λk + k=N −m Pk (pk − ak )λk
Pk `=1 ` pes,` `=j
Through simple manipulations we can show that the above
Using an analogous approach as the one presented in the inequality is equivalent to
proof of Proposition 1, we have that for k ∈ {1, 2, . . . , j −1}, φN −m aN −m
Pj−1 δ PN !−1 PNδ −m pN −m − aN −m
δ
(j) φk `=1 P` a` λ` + `=j P` p` λ` PN PN
pes,k = 1 + δ Pj−1 . 1 − k=1 φk ak λk + k=N −m+1 φk ak λk
Pk 1 − `=1 φ` a` λ` < PN PN .
δ δ
k=1 Pk ak λk + k=N −m+1 Pk (pk − ak )λk
Comparing the last two equations, it is easy to see that
Now, we only need to verify that
j−1
X N
X
a` PN PN
P`δ λ` + P`δ p` λ` 1 − k=1 φk ak λk + k=N −m+1 φk ak λk
(j) PN PN
`=1 pes,` `=j δ δ
Pj−1 δ PN k=1 Pk ak λk + k=N −m+1 Pk (pk − ak )λk
δ PN
`=1 P` a` λ` + `=j P` p` λ` 1− φk ak λk
= Pj−1 . < PN k=1 .
1 − `=1 φ` a` λ` δ
k=1 Pk ak λk
Finally, we can use this result to rewrite (18) as, for all j ∈ N , Again, simple manipulations lead to the equivalent inequality
Pj−1 PN PN
φj aj 1 − k=1 φk ak λk k=N −m+1 φk ak λk 1 − k=1 φk ak λk
0≤ δ < Pj−1 PN . PN < PN ,
Pj pj − aj δ
k=1 Pk ak λk +
δ
k=j Pk pk λk
δ δ
k=N −m+1 Pk (pk − ak )λk k=1 Pk ak λk

This concludes the proof, since the extension for the other which is true from the base case. This can be seen by
partitions of N is analogous. multiplying (19) by PNδ −j (pN −j − aN −j ) at both sides of
12 VOLUME X, 2019

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

the inequality and summing over j ∈ {0, . . . , m − 1}. Thus, We can also prove this by contradiction and then, for simplic-
Cν ⊂ R for the mapping ν : n 7−→ n. The extension for ity of exposition, suppose that a2 is the one that satisfies this
another instances
S of ν ∈ V is analogous. This concludes the restriction. We repeat this procedure until we reach all traffic
proof that ν∈V Cν ⊂ R. classes. Let us show the j-th step for completeness, j ∈ N .
However,
S we still need to prove the converse, that is Suppose that for all n ∈ {j, j + 1, . . . , N },
R ⊂ ν∈V Cν . Note that the set of arrival rates that makes Pj−1
φn an 1 − k=1 φk ak λk
the system stable in Theorem 1 requires that at least one an ≥ P PN . (25)
Pnδ pn − an j−1 δ δ
(n ∈ N ) satisfies k=1 Pk ak λk + k=j Pk pk λk

φn an 1 Multiplying (25) by Pnδ (pn − an )λn > 0 at both sides,

< PN . (20)
Pnδ pn − an δ
k=1 Pk pk λk
summing over n ∈ {j, j + 1, . . . , N } and manipulating we
have
Let us show that R requires the same restriction by contra-   
N
X N
X XN
diction. Suppose that there exist a ∈ R such that
 φn an λn   Pkδ ak λk + Pkδ (pk − ak )λk 
φn an 1 n=j k=1 k=j
≥ PN ∀n ∈ N . (21)
Pnδ pn − an δ
k=1 Pk pk λk j−1
! N 
X X
Multiplying (21) by Pnδ (pn − an )λn > 0 at both sides and ≥ 1− φk ak λk  Pnδ (pn − an )λn  . (26)
summing over n ∈ N we have k=1 n=j

XN PN Once again, multiplying (23) by Pnδ (pn − an )λn > 0 at both

Pnδ (pn − an )λn
φn an λn ≥ n=1 PN , sides, summing over n ∈ {j, j +1, . . . , N } and manipulating
δ
n=1 k=1 Pk pk λk we have
  !
which is equivalent to XN N
X
! N !  φn an λn  δ
Pk ak λk (27)
XN X N
X
δ
φn an λn Pk pk λk ≥ Pnδ (pn − an )λn . n=j k=1

N
! N 
n=1 k=1 n=1 X X
(22) < 1− φk a k λ k  Pnδ (pn − an )λn  .
k=1 n=j
Since a ∈ R, then
P Then, through some manipulations on (26) and (27), we have
φn an 1 − k φk ak λk
Pnδ pn − an
< P δ
∀n ∈ N . (23) ! N 
k Pk ak λk XN X
0≤ 1− φk ak λk  Pnδ (pn − an )λn 
Again, multiplying (23) by Pnδ (pn −an )λn > 0 at both sides,
k=1 n=j
summing over n ∈ N and performing some manipulations   !
N
X N
X
we have
N
! N ! − φn an λn  Pkδ ak λk
X X n=j k=1
δ
φn an λn Pk ak λk (24)
n=1 k=1 < 0.
N
! N
!
X X As expected, we have a contradiction. Then, we must have at
< 1− φk ak λk Pnδ (pn − an )λn .
least one an , n ∈ {j, j + 1, . . . , N } that satisfies
k=1 n=1
Pj−1
Then, through some manipulations on (22) and (24), we have φn an 1 − k=1 φk ak λk
< Pj−1 PN . (28)
! N ! Pnδ pn − an δ δ
X N XN X k=1 Pk ak λk + k=j Pk pk λk
δ δ
0≤ Pn an λn − 1 − φk ak λk Pn pn λn We assume that this is satisfied by the j-th class and in this
n=1 k=1 n=1
case, R ⊂ Cν for ν : n 7−→ n. Without the assumption of the
< 0, ordering in which (28) is satisfied, we conclude that (28) must
which clearly is a contradiction, since R is a non-empty set. hold
S for at least one permutation
S of N . This region Sis exactly
Thus, there exists at least one an , n ∈ N that satisfies (20). ν∈V Cν . Therefore, R ⊂ ν∈V C ν . Finally, R = ν∈V Cν .
For simplicity of exposition, let us suppose that the arrival
rate an that satisfies this restriction is from the first traffic
REFERENCES
S (n = 1). The next step is to show that as in the set
class
[1] H. Lee, S. Vahid, and K. Moessner, “A survey of radio resource manage-
ν∈V Cν , the set R also requires that we have at least one ment for spectrum aggregation in LTE-advanced,” IEEE Communications
an , aside from a1 , that satisfies Surveys & Tutorials, vol. 16, no. 2, pp. 745–760, 2014.
[2] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. Soong,
φn an 1 − φ1 λ1 a1 and J. C. Zhang, “What will 5G be?” IEEE Journal on selected areas in
< PN .
Pnδ pn − an P1 a1 λ1 + k=2 Pkδ pk λk
δ communications, vol. 32, no. 6, pp. 1065–1082, 2014.

VOLUME X, 2019 13

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 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/ACCESS.2019.2940947, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

[3] T. Taleb and A. Kunz, “Machine type communications in 3GPP networks: Transactions on Wireless Communications, vol. 17, no. 9, pp. 5885–5899,
potential, challenges, and solutions,” IEEE Communications Magazine, Sep. 2018.
vol. 50, no. 3, 2012. [26] M. Kountouris, N. Pappas, and A. Avranas, “Qos provisioning in large
[4] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski, wireless networks,” in 2018 16th International Symposium on Modeling
“Five disruptive technology directions for 5G,” IEEE Communications and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt),
Magazine, vol. 52, no. 2, pp. 74–80, February 2014. May 2018, pp. 1–6.
[5] A. Ghosh et al., “Heterogeneous cellular networks: From theory to prac- [27] F. Baccelli, B. Błaszczyszyn et al., “Stochastic geometry and wireless net-
tice,” IEEE Communications Magazine, vol. 50, no. 6, pp. 54–64, June works: Volume II applications,” Foundations and Trends® in Networking,
2012. vol. 4, no. 1–2, pp. 1–312, 2010.
[6] D. Liu, L. Wang, Y. Chen, M. Elkashlan, K. Wong, R. Schober, and [28] J. F. C. Kingman, Poisson processes. Clarendon Press, 1992, vol. 3.
L. Hanzo, “User association in 5g networks: A survey and an outlook,” [29] M. Haenggi and R. Smarandache, “Diversity polynomials for the analysis
IEEE Communications Surveys Tutorials, vol. 18, no. 2, pp. 1018–1044, of temporal correlations in wireless networks,” IEEE Transactions on
Secondquarter 2016. Wireless Communications, vol. 12, no. 11, pp. 5940–5951, 2013.
[7] T. Qiu, N. Chen, K. Li, M. Atiquzzaman, and W. Zhao, “How can heteroge- [30] S. S. Kalamkar and M. Haenggi, “The spatial outage capacity of wireless
neous internet of things build our future: A survey,” IEEE Communications networks,” IEEE Transactions on Wireless Communications, vol. 17, no. 6,
Surveys & Tutorials, 2018. pp. 3709–3722, June 2018.
[8] P. Popovski, J. J. Nielsen, C. Stefanovic, E. de Carvalho, E. Strom, [31] X. Lin, J. G. Andrews, and A. Ghosh, “Spectrum sharing for device-
K. F. Trillingsgaard, A.-S. Bana, D. M. Kim, R. Kotaba, J. Park et al., to-device communication in cellular networks,” IEEE Transactions on
“Wireless access for ultra-reliable low-latency communication: Principles Wireless Communications, vol. 13, no. 12, pp. 6727–6740, Dec 2014.
and building blocks,” IEEE Network, vol. 32, no. 2, pp. 16–23, 2018. [32] M. Haenggi, Stochastic geometry for wireless networks. Cambridge
[9] L. Liu, E. G. Larsson, W. Yu, P. Popovski, C. Stefanovic, and E. de Car- University Press, 2012.
valho, “Sparse signal processing for grant-free massive connectivity: A [33] M. Haenggi et al., “Stochastic geometry and random graphs for the
future paradigm for random access protocols in the internet of things,” analysis and design of wireless networks,” IEEE Journal on Selected Areas
IEEE Signal Processing Magazine, vol. 35, no. 5, pp. 88–99, 2018. in Communications, vol. 27, no. 7, 2009.
[10] F. Jameel, Z. Hamid, F. Jabeen, S. Zeadally, and M. A. Javed, “A survey of [34] C. Yin, C. Chen, T. Liu, and S. Cui, “Generalized results of transmission
device-to-device communications: Research issues and challenges,” IEEE capacities for overlaid wireless networks,” in 2009 IEEE International
Communications Surveys & Tutorials, 2018. Symposium on Information Theory, June 2009, pp. 1774–1778.
[11] P. Cardieri, “Modeling interference in wireless ad hoc networks,” IEEE [35] S. K. Bose, An introduction to queueing systems. Springer Science &
Communications Surveys & Tutorials, vol. 12, no. 4, pp. 551–572, 2010. Business Media, 2013.
[12] R. R. Rao and A. Ephremides, “On the stability of interacting queues [36] D. P. Bertsekas, Nonlinear programming. Athena scientific Belmont,
in a multiple-access system,” IEEE Transactions on Information Theory, 1999.
vol. 34, no. 5, pp. 918–930, Sep 1988. [37] D. Tse and P. Viswanath, Fundamentals of wireless communication. Cam-
[13] V. Naware, G. Mergen, and L. Tong, “Stability and delay of finite-user slot- bridge university press, 2005.
ted aloha with multipacket reception,” IEEE Transactions on Information [38] S. Kompella, A. Ephremides et al., “Stable throughput regions in wireless
Theory, vol. 51, no. 7, pp. 2636–2656, July 2005. networks,” Foundations and Trends in Networking, vol. 7, no. 4, pp. 235–
[14] W. Szpankowski, “Stability conditions for some distributed systems: 338, 2014.
Buffered random access systems,” Advances in Applied Probability, [39] R. M. Loynes, “The stability of a queue with non-independent inter-
vol. 26, no. 2, pp. 498–515, 1994. arrival and service times,” in Mathematical Proceedings of the Cambridge
[15] K. Stamatiou and M. Haenggi, “Random-access poisson networks: stabil- Philosophical Society, vol. 58, no. 03. Cambridge Univ Press, 1962, pp.
ity and delay,” IEEE Communications Letters, vol. 14, no. 11, pp. 1035– 497–520.
1037, 2010.
[16] A. Munari, F. Rossetto, P. Mähönen, and M. Petrova, “On the stability of a
full-duplex aloha network,” IEEE Communications Letters, vol. 20, no. 12,
pp. 2398–2401, 2016.
[17] Y. Zhou and W. Zhuang, “Performance analysis of cooperative communi-
cation in decentralized wireless networks with unsaturated traffic,” IEEE
Transactions on Wireless Communications, vol. 15, no. 5, pp. 3518–3530,
2016.
[18] L. Liu, Y. Zhong, H. H. Yang, M. Sheng, T. Q. Quek, and J. Li, “Mean
packet throughput analysis of downlink cellular networks with spatio-
temporal traffic,” in GLOBECOM 2017-2017 IEEE Global Communica-
tions Conference. IEEE, 2017, pp. 1–6.
[19] J. M. Borwein, R. E. Crandall et al., “Closed forms: what they are and why
we care,” Notices of the AMS, vol. 60, no. 1, pp. 50–65, 2013.
[20] Y. Zhong, M. Haenggi, T. Q. S. Quek, and W. Zhang, “On the stability
of static poisson networks under random access,” IEEE Transactions on
Communications, vol. 64, no. 7, pp. 2985–2998, July 2016.
[21] M. Gharbieh, H. ElSawy, H. Yang, A. Bader, and M. Alouini, “Spatiotem-
poral model for uplink iot traffic: Scheduling and random access paradox,”
IEEE Transactions on Wireless Communications, vol. 17, no. 12, pp.
8357–8372, Dec 2018.
[22] M. Gharbieh, H. ElSawy, A. Bader, and M. Alouini, “Spatiotemporal
stochastic modeling of iot enabled cellular networks: Scalability and
stability analysis,” IEEE Transactions on Communications, vol. 65, no. 8,
pp. 3585–3600, Aug 2017.
[23] Y. Zhong, T. Q. S. Quek, and X. Ge, “Heterogeneous cellular networks
with spatio-temporal traffic: Delay analysis and scheduling,” IEEE Journal
on Selected Areas in Communications, vol. 35, no. 6, pp. 1373–1386, June
2017.
[24] H. H. Yang and T. Q. S. Quek, “Spatiotemporal analysis for sinr coverage
in small cell networks,” IEEE Transactions on Communications, pp. 1–1,
2019.
[25] Z. Chen, N. Pappas, M. Kountouris, and V. Angelakis, “Throughput
with delay constraints in a shared access network with priorities,” IEEE

14 VOLUME X, 2019

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.