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/LCOMM.2018.2797274, IEEE
Communications Letters
Spreading Factor Allocation for Massive Connectivity in LoRa Systems
Jin-Taek Lim and Youngnam Han
Abstract—Along with enhanced mobile broadband and ultra-
reliable low latency communication, massive connectivity has
been one of the key requirements for enabling technologies of
5G. For IoT, low power consumption and wide area coverage
for end devices (ED) are important figures of merit, for which
LoRa, SigFox and Narrow Band-IoT are dominant technologies.
In this letter, we analyze LoRa systems for increasing average
system packet success probability (PSP) under unslotted ALOHA
random access protocol. The lower bound for average system
PSP is derived when many contending EDs are transmitting by
stochastic geometry. And it is shown that the average system
PSP can be maximized by properly allocating an SF to each
traffic, which also maximizes connectivity of EDs. We formulate
an optimization problem for maximizing the average system PSP
to propose a sub-optimal SF allocation scheme to each traffic.
Analysis on PSP is validated through simulations, and compar-
ison with existing schemes reveals that our proposed scheme
achieves the highest PSP and so the maximum connectivity.
Index Terms—Low power wide area networks, internet of
things, wireless communications, unslotted aloha, LoRa, stochas-
tic geometry, interference analysis.
I. I NTRODUCTION
T HE INTERNET of Things (IoT) is widely adopted in
such areas as home automation, u-Health and environ-
ment. In IoT, it is essential that things can communicate with
their surroundings widely and collect or transmit information
with low power consumption, for which Low Power Wide
Area (LPWA) network technologies are proposed. However,
LPWA technologies achieve the objectives through the trade-
off with low data rate (typically in orders of tens of kilobits
per seconds) and higher latency (typically in orders of seconds
or minutes). Up to date, three technologies are proposed for
those use cases of low-power, delay tolerant and low rate
transmissions, Narrow Band (NB)-IoT, LoRa and SigFox.
These technologies have many different features, and espe-
cially differ in that NB-IoT adopts licensed band with existing
infrastructure, while as LoRa and SigFox adopt unlicensed
band requiring new infrastructure for service.
Lots of work have been carried on for the effect on the
coverage by analyzing the interference. The signal activity
and power levels for LoRa and SigFox are measured for the
European Industrial, Scientific, and Medical band (863-870
MHz) in the city of Aalborg, Denmark [2]. The coverage
and capacity of SigFox and LoRaWAN in a large scale
urban environments have been analyzed [3]. Moreover, the
coverage of GPRS, NB-IoT, LoRa, and SigFox is compared
in a realistic scenario by simulation [4]. Especially for LoRa,
a mathematical analysis on its coverage was conducted in [5]
via stochastic geometry [6]. However, a little work has been
done to maximize the massive connectivity which is another
important feature for LPWA IoT.
This work was supported by Institute for Information & communications
Technology Promotion (IITP) grant funded by the Korea government (MSIP)
(No.2016-0-00209).
The authors are with the Department of Electrical Engineering, Korea
Advanced Institute of Science and Technology, Daejeon 305-701, South Korea
(e-mail: jtyim@kaist.ac.kr; ynhan@kaist.ac.kr).
1089-7798 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1
6LJQDOWREHGHFRGHG
6DPHVSUHDGLQJLQWHUIHUHQFH
'LIIHUHQWVSUHDGLQJLQWHUIHUHQFH
!
!" 
s + 1
s
s
s " 1
V
)CVGYC[
V
!"1 
! '& $t #!,!" 
!+1 
0 #!"  #!  %&  W
!+2 
(a) LoRaWAN model (b) Packet collision model
Fig. 1: System model
In LoRa, the Chirp Spread Spectrum (CSS) modulation
technique [1] is proposed, which is a kind of spread spectrum
multiple access techniques accommodating multiple users in a
single channel. It is known that the CSS modulation achieves
low receiver sensitivities making long range communications
possible. Also, it is known that CSS interference comes mostly
from the same-spreading factor (SF) signals, and has a pseudo-
orthogonal characteristic with the different-SF signals [1]. In
order to analyze the massive connectivity, we must consider
low receiver sensitivities and interferences of CSS modulation.
We investigated the massive connectivity in LoRa and
proposed a sub-optimal SF allocation method by considering
the effects of same/different-SF interference under unslotted
ALOHA protocol. First, we derived the average system PSP
by the well-known stochastic geometry. Then, realizing that
massive connectivity can be achieved by maximizing average
system PSP, we formulate an optimization problem and pro-
posed an SF allocation scheme.
The rest of this letter is organized as follows. Section II
briefly introduces LoRa system. Analysis on PSP and problem
formulation are provided in Section III. Performance evalua-
tions and discussions are presented in Section IV. Conclusions
are made in Section V.
II. S YSTEM M ODEL
We consider an uplink LoRa system as shown in Fig. 1a
where N EDs are distributed in a single cell. EDs are located
from a homogeneous Poisson point process (PPP) Φ on the
Euclidean plane R2 with intensity λ = πRN̄
2 , where RC is the
C
cell radius and N̄ is the mean of Poisson random variable
N. Each ED selects an SF s and transmits a packet, where
s (∈ S = {1, 2, ..., 6} and S = |S|) is the index assigned
for corresponding SF in Tab I. And these EDs compete with
each other through unslotted ALOHA random access protocol
within a given contention time Tc .
Collision overlap time (ts,s0 ) in unslotted ALOHA: To
model the interference caused by collisions, we define the
collision overlap time ts,s0 (Fig. 1b) between the packets of
SF s and s 0 and derive its probability density function (PDF).
ls is the length of packet with SF s, s ∈ S. Because LoRa
assumes unslotted ALOHA protocol, ts,s0 does not depend on
which EDs are transmitting, but on its packet durations. Also,
in unslotted ALOHA, each ED with SF s randomly chooses
its transmission start time ts , where ts ∼ U(0, Tc0 = (Tc − ls ))
for contention window, Tc . Here, we can assume that Tc0  Tc
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/LCOMM.2018.2797274, IEEE
Communications Letters
2
TABLE I: CSS modulation properties of a 25 byte message
[5]
because Tc is sufficiently longer than ls . If we define ∆t as SF s Data Packet Receiver S N Rs Range
the time difference between the transmission start times of two Rate Duration Sensitivity dB m
colliding packets of SF s and s (i.e., ∆t = ts − ts), the
0 0
 PDF
kbps l s sec q s dBm
7 1 5.47 0.036 -123 -6 [d1, d2 ]
of ∆t can be derived as f∆t (t) = max T1c 1 −  Ttc  , 0 . From
 
8 2 3.13 0.064 -126 -9 [d2, d3 ]
f∆t (t), we can get the pdf of ts,s0 (see Fig. 1b, where ls < ls0 9 3 1.76 0.113 -129 -12 [d3, d4 ]
and ts0 < ts < ts0 + ls0 ). Assuming that ls is shorter than ls0 10 4 0.98 0.204 -132 -15 [d4, d5 ]
11 5 0.54 0.365 -134.5 -17.5 [d5, d6 ]
(Since, for ls > ls0 , it can be seen as symmetric case with ls 12 6 0.29 0.682 -137 -20 [d6, d7 ]
and ls0 interchanged, we provide analysis for this case), we
can derive the distribution of ts,s0 from f∆t (t) as follows In LoRa, the chirp symbol duration can be expressed as
ls + ls0
2 2
! T s = 2SF /BW [1], where SF is a spreading factor and BW is
1
fts, s0 (t) = Tc − ls0 − ls + δ(t) bandwidth. Ts doubles for every increase in SF. As a result, a
Tc 2Tc packet length ls , which is multiples of Ts , has a different value
(ls0 − ls )2 ls + ls0
   
1 2 1 for each SF s. Actually, an increase in packet duration for a
+ ls0 − ls − δ(t − ls ) + 2 t + 2− , chirp gives the message higher robustness to noise, resulting in
Tc 2Tc Tc T c Tc
(1) a low required signal-to-noise ratio threshold. For each SF s,
receiver sensitivity qs = −174+10 log B + N F + SN Rs where
where δ(·) is the Dirac-delta function. In (1), the first term on the constant (-174 dBm) is the thermal10 noise density, and N F
the righthand side is for the event when there is no collision is the noise figure margin at GW (6 dB). A typical example
and the second term comes from the observation that the for system parameters is provided in Tab. I when B = 125
packet with smaller SF s 0 is included in the packet of larger KHz and 25 byte message [5].
SF s packet. The last two terms are for the case when the
From qs , we can deduce that the SF to satisfy qs gets
packets overlap each other in t.
smaller as the ED gets closer to the GW. In other words,
SIR model: The reference packet to be decoded has multi-
the farther the ED is from the GW, the higher SF the ED
ple SIRs from Is0 for each s 0 (Is0 , a set of interfering EDs of
should be allocated to. It means that there are boundaries
SF s 0, s 0 ∈ S). Then, we define the SIR of packet with SF s
which is determined from the receiver sensitivity (condition I).
interfering with Is0 as
This results in the predefined SF allocation intervals [ds, ds+1 ]
Pg A0 r −α P A0 gr −α shown in Fig. 1a. However, the amount of interference is also
γs,s0 = Í ts, s 0 = , (2)
P g A r −α I s,s 0 related to these intervals, the intervals should be determined
k ∈I s 0 ls 0 k 0 k
adaptively considering condition II. To be specific, the increase
where P is the transmit power of each ED, g (or gk ) is in the number of long messages could cause higher probability
the Rayleigh fading channel between gateway (GW) and ED of collision between packets due to the increase in the longer
located at r (or interfering ED k), A0 = c/4π f which comes interval of high SF messages, causing severe same/different-
from the Friis transmission equation with the carrier frequency SF interferences. So, for EDs, it is important to assign a proper
f , c is the light velocity and α is the path loss exponent. And, [ds, ds+1 ] to meet both conditions.
r (rk ) is the distance between the GW and ED (interfering ED III. P ROBLEM F ORMULATION
k). The packets are not synchronized under unslotted ALOHA,
ts, s 0 The massive connectivity can be defined as the average
so we need to normalize interfering power value in (2) by ls0 .
The reason behind this is that if the channel coding make number of EDs which meet condition I and condition II;
P A gr −α ≥ q and γ 0 ≥ T 0 for all s 0 . Then, massive
use of an interleaver, the interfering energy between colliding 0 s s,s s,s
packets can be spread out. connectivity can be mathematically written as (3) where
SF Allocation in LoRa [5]: In order for the received IA(x) is an indicator function, fg, Is, s0 (x, i) is the joint pdf of
packet with SF s to be successfully decoded in a GW, its random variables g and Is,s0 , fR (r) = R2 and P[SF = s|r]
2r
C
received power must exceed a receiver sensitivity qs (condition is the probability that ED located at r will send a packet
Ñ
I) and its signal-to-interference ratio (SIR) contaminated by with SF s. ∫is the ’intersection’ of sets. (a) holds since
the interference with packets of SF s 0 (s 0 ∈ S), γs,s0 , must = dP = P(A). And, (b) holds because of the
∫
X
I A (x)dP X
also exceed the relative threshold Ts,s0 (condition II) for all fact that the choice of SF is determined from the predefined
s 0, simultaneously (Ts,s0 is given by the element of T in [1] interval [ds, ds+1 ]. Therefore, P[SF = s|r] is 1 in [ds, ds+1 ],
which is the threshold matrix of LoRa). and 0 otherwise.

{Average # of EDs satisfying both conditions}

∫ 2π ∫ RC ∫ ∞ ∫ ∞ Õ
= P[SF = s|r]IÑ  P A0 gr −α  (x, i) × λ × r × fg, Is, s0 (x, i) dx di dr dθ
{P A0 gr −α ≥qs }
Ñ
0 0 0 0 s 0 ∈S I s, s 0 ≥Ts, s 0
s ∈S
"( )  #
∫ RC (3)
(a)
Õ Ù Ù  2r
= N̄ P[SF = s|r]P γs,s0 ≥ Ts,s0 {P A0 gr −α
≥ qs }r × 2 dr

0 s ∈S s0 ∈S
 RC
∫ RC Õ∫ ds+1
(b)
Õ
= N̄ P[SF = s|r]PSPs (r) fR (r) dr = N̄ PSP s (r) fR (r) dr
0 s ∈S s ∈S ds

1089-7798 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
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/LCOMM.2018.2797274, IEEE
Communications Letters
3

T 0rα
Then, the average system PSP over all EDs, PSP, can where z = s,PsA0 . (a) comes from the independence among
Í ∫d
be obtained as s ∈S d s+1 PSP s (r) fR (r) dr. Since N̄ is a ts,s0 , gk and rk , and gk ∼ exp(1). (b) comes from the fact that
s
constant, we can conclude that maximizing PSP is equivalent ts,s0 for all EDs using SF s 0 are independent of each other.
to maximizing massive connectivity. By using the probability generating functional of homoge-
Average System PSP: neous PPP [6] to (7), it can be further written as :
Õ ∫ ds+1 L Is, s0 (z) =
PSP = PSP s (r) fR (r)dr. (4) ds 0 +1
∫  
s ∈S ds  1  ® xdx ª® .
ª
exp ­−2πλ ­1 − Ets, s0
© © 
ts, s 0
 1 + zP ls0 A0 x 
−α
 
And, PSP s (r) can be expressed as follows «
ds0
«  ¬ ¬
"(
Ù
)
Ù
 #
 (8)
PSP s (r) = P γs,s0 ≥ Ts,s0 {P A0 gr −α ≥ qs }r

So, we can obtain the closed form for Ets, s0 [·] applying the
s0 ∈S

"  # PDF (1) of ts,s0 , which results in (9). With (9), the integral of
Ù  (8) has the general solution which can be solved by using an
=P γs,s0 ≥ Ts,s0 P A0 gr −α ≥ qs, r P [P A0 gr −α ≥ qs |r]
0
 analytical solver, such as Mathematica. Then, by substituting
"s ∈S  # (8) into (6), and then into (5) with P[P A0 gr −α ≥ qs |r] =
exp(−qs r α /P A0 ), the lower bound of (4) can be obtained.
(a) Ù 
≥P γs,s0 ≥ Ts,s0 r P [P A0 gr −α ≥ qs |r]
0
 However, it is not possible to get the closed-form for (4), so the
" s ∈S # Riemann summation can be adopted for numerical calculation.
(b) Ö
≥ P[γs,s0 ≥ Ts,s0 |r] P[P A0 gr −α ≥ qs |r]. Optimization problem and solution: In this subsection, we
s0 ∈S propose an algorithm to determine [ds, ds+1 ] maximizing (4).
(5) We start from (5) because the difference between (4) and a
Inequality (a) is due to the fact that the conditioned probability lower bound (5) is small, which will be verified. In order
is greater than the unconditioned probability because the given to make (5) tractable, we made the following assumptions.
condition limits the minimum value of the numerator in γs,s0 The different-SF interference is ignored. Moreover, only the
for all s 0. Inequality (b) holds due to Fortuin, Kastelynm, probability event when ts,s0 = 0, which is the largest portion
Ginibre (FKG) inequality [8]. For that, we need the following due to large Tc , is considered in (1). We use the approximation,
definitions. exp(−x) ' (1 − x) for small x. With approximations, we can
For the PPP, let ω = (a1, a2, ..., an, ...) for n ∈ N, and formulate the following optimization problem :
an = 1 if E Dn is active and an = 0 otherwise. Then α+2 − d α+2 )
S
" #
Õ 2 − d2)
(ds+1 2qs (ds+1
s s
ω 0  ω if an0 ≥ an , ∀n. From (2), γs,s0 decreases as the max PSP = max −
number of interfering EDs increases, i.e., γs,s0 (ω) ≤ γs,s0 (ω 0)
d2,...,dS d2,...,dS
s=1
RC2 (α + 2)P A0 RC 2

if ω 0  ω. Thus, the event {γs,s0 ≥ Ts,s0 } for each SF s 0 l2

  
 2l 
s
is a decreasing event [7], since if ω 0 ∈ {γs,s0 ≥ Ts,s0 } and × 1 − λπ ds+1 2
− ds2 − s2
Tc Tc
ω 0  ω, then ω ∈ {γs,s0 ≥ Ts,s0 }. In other words, these events (10)
are positively correlated (i.e., The likelihood of each of these
subject to ds ≤ ds+1, ∀s ∈ S, where d1 = δ and dS+1 = RC
events being true decreases when ω increases and increases
are fixed values (δ is set to avoid the singular point in path
when ω decreases). Hence, by applying FKG inequality for
loss model). The first bracket within the summation of (10)
these events, we can get the lower bound of condition II.
is for condition I. In the second bracket (i.e., condition II
To evolve the last line of (5), we should derive the SIR
probability), we can deduce that one interfering ED reduces
Complementary Cumulative Distribution Function (CCDF) for
the probability of condition II by almost 2l s
Tc on the average.
each s 0, which can be written as 2ls
    T 0rα I 0  We conjecture that Tc comes from the uniform random choice
Ts,s0 Is,s0  (a) − s, s P A s, s
P[γs,s0 ≥ Ts,s0 |r] = P g ≥ r = EIs, s0 e 0 . of slots in unslotted ALOHA protocol. Therefore, maximizing
P A0 r −α  PSP corresponds to determining SF regions and so allocating
(6) SFs to each ED.
(a) follows from the fact that g ∼ exp(1). Then, the Laplace If α is not an integer, (10) becomes a nonlinear program-
transform of the interference Is,s0 , L Is, s0 (z) can be written as ming (NLP) problem with the bounded variation. Although the
complexity is high with multiple variables, this problem can
be solved with a global optimization solver, such as multilevel
Ö t s0

−z P s, −α 
l s 0 gk A0 rk 
L Is, s0 (z) = Ets, s0 ,gk ,rk 

e  coordinate search (MCS) in MATLAB or MathOptimizer and
k ∈Is0
Global optimization packages of Mathematica.

 
Ö
1
 And, the proposed method can be implemented as follows.
(a)
= Ets, s0 ,rk 
 
ts, s 0
 (7) ED periodically sends its location and uplink channel infor-
k ∈Is0 1 + zP ls0 A0 rk  −α 
  mation to a GW, with a long interval. Then the GW allocates
Ö   SF to each ED by the proposed SF allocation, which will be
(b) 1
= Erk   , updated with the next channel report. It might have a limitation
  
 Ets, s0 
ts, s 0
 1 + zP ls0 A0 rk−α   because EDs periodically send information to the GW, which

k ∈Is0
  

1089-7798 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
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/LCOMM.2018.2797274, IEEE
Communications Letters
4
!
ls20 + ls2 (ls0 − ls )2
   
1  1
= 1 1
0 − ls + +

Ets, s0  T c − ls ls 0 − ls −
 t 0
 1 + zP s,lss0 A0 x −α  Tc
 2Tc 1 + zP A0 x −α Tc 2Tc
   2l2 (9)
l +l 0

l 0 0
ls20 T22 (zP A0 x −α ) Tsc 2 − sTc s − Ts2
+ c
+ c
log(1 + zP A0 x −α )
zP A0 x −α (zP A0 x −α )2

requires more power and complexity, but negligible compared 45

SF12

to other schemes. And the proposed scheme is also beneficial 40 SF11

to for long-term changes in the channel. 35

SF10
30

IV. R ESULTS AND D ISCUSSIONS 25 SF9

We make simulations for performance comparison, where 20

SF8
path loss exponent is assumed to be 2.7 (suburban scenario), 15

f is assumed to be 867 MHz (for LoRa in Europe), the 10

Proposed
EIB
SF7
average number of EDs varies from 100 to 1000, and EDs are 5 EAB

distributed within a single GW coverage of radius RC = 45 0

200 400 600 800 1000 1200 1400 1600 1800 2000

km. The transmit power for each ED was 14 dBm and Tc is Fig. 2: Average number of EDs v.s. Allocation intervals
set to 60 sec. The other parameters are provided as in Tab. I. 1 1000

For the performance comparison, we adopted 3 other existing 0.9

Proposed (anal)
EIB (anal) 900
EAB (anal)
schemes such as; 0.8
Random (anal)
Proposed (sim)
800

EIB (sim)
• Equal-Interval-Based (EIB) scheme [5] which deter- 0.7
EAB (sim)
Random (sim)
700

mine the intervals [ds, ds+1 ], according to (ds+1 − ds ) = 0.6 600

0.5 500
RC /S, 0.4 400

• Equal-Area-Based (EAB) scheme which p determine the 0.3 300

intervals [ds, ds+1 ], based on ds = RC (s − 1)/S that 0.2 200

makes the areas of doughnuts equal. 0.1 100

• Random scheme which allocates SFs by uniform random 0

200 400 600 800 1000 1200 1400 1600 1800
0
2000

choice, not depending on the location of the ED.
The allocation intervals for proposed, EIB and EAB scheme
are shown in Fig. 2. From Fig. 2, it can be said that, as N
increases, it get more difficult for each ED to satisfy condition
II, so the proposed scheme makes the lowest SF region larger
to reduce collisions. In other words, when N is large, it is more
advantageous for maximizing PSP to increase the probability
for condition II than to increase the probability for condition
I.
Monte Carlo simulations and numerical analysis results are
presented in Fig. 3 for the validation of the lower bound of (5).
The gap between the numerically calculated lower bound in
(4) and simulation lies within an acceptable bound, confirming
that the overall performance of LoRa can be analyzed through
the derived lower bound. Also, this tight bound means that
each event considered in (5) is almost independent.
From Fig. 3, we can make the following discussions. First,
we can confirm that EIB scheme works worse than Random
scheme because condition II becomes difficult to be satisfied
when N increases. In case of EAB scheme when EDs were
equally allocated in each SF region, the influence of inter-
ferences can be reduced to some extent, resulting in better
performance than Random and EIB schemes. The proposed
scheme is superior to all other schemes in terms of massive
connectivity because both conditions are considered. From
simulations with N̄ = 2000, the proposed method can provide
stable 810 EDs satisfying both conditions, which is about 22%
increase in massive connectivity over EAB. These results are
based on the active EDs only, so the total number of EDs
served during the contention period will be large by the inverse
of activity factor.
Fig. 3: Average number of EDs v.s. (Average system PSP,
Massive connectivity)
V. C ONCLUSIONS
We formulated an optimization problem to maximize av-
erage system PSP when each ED attempts to transmit uplink
under unslotted ALOHA via stochastic geometry. We proposed
an adaptive SF allocation algorithm and provided performance
comparison with 4 other SF assigning schemes in Section IV.
It was observed and confirmed that our proposed allocation
was superior to other methods through simulations.
R EFERENCES
[1] C. Goursaud and J.-M. Gorce, ”Dedicated networks for IoT: PHY/MAC
state of the art and challenges,” EAI Endorsed Trans. Internet of Things,
vol. 1, no. 1, pp. 1-11, Oct. 2015.
[2] M. Lauridsen, B. Vejlgaard, I. Kovcs, H. Nguyen, and P. Mogensen,
”Interference measurements in the European 868 MHz ISM band with
focus on LoRa and SigFox,” in Proc. IEEE WCNC, San Francisco, CA,
USA, 2017.
[3] B. Vejlgaard, M. Lauridsen, H. Nguyen, I. Kovacs, P. Mogensen, and M.
Sorensen, ”Interference Impact on Coverage and Capacity for Low Power
Wide Area IoT Networks,” in Proc. IEEE WCNC, San Francisco, CA,
USA, 2017.
[4] M. Lauridsen, H. Nguyen, B. Vejlgaard, I. Kovacs, P. Mogensen and M.
Srensen, ”Coverage comparison of GPRS, NB-IoT, LoRa, and SigFox in
a 7800 km2 area,” in Proc. IEEE VTC Spring, Sydney, AUS, 2017.
[5] O. Georgiou and U. Raza, ”Low power wide area network analysis: Can
lora scale?,” IEEE Wireless Communications Letters, vol. 6, no. 2, pp.
162165, Apr. 2017.
[6] M. Haenggi, Stochastic Geometry for Wireless Networks. Cambridge
University Press, 2013.
[7] R. Vaze, K. Truong, S. Weber, and R. Heath, ”Two-way transmission
capacity of wireless ad-hoc networks,” IEEE Trans. Wireless Commun.,
vol. 10, no. 6, pp. 1966-1975, 2011.
[8] G. Grimmett, Percolation. New York, NY, USA: Springer-Verlag, 1989

1089-7798 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.