Sie sind auf Seite 1von 12

1

Grant-Free Massive NOMA:


Outage Probability and Throughput
Rana Abbas, Student Member, IEEE, Mahyar Shirvanimoghaddam, Member, IEEE, Yonghui Li, Senior
Member, IEEE, and Branka Vucetic, Fellow, IEEE

Abstract—In this paper, we consider a massive uncoordinated perspective, orthogonal multiple access has been shown to be
non-orthogonal multiple access (NOMA) scheme where devices strictly sub-optimal for short packet transmissions [6]. The
have strict latency requirements and no retransmission oppor- gap between the achievable sum-rate of orthogonal multiple
tunities are available. Each device chooses a pilot sequence
from a predetermined set as its signature and transmits its access and the maximum sum-rate increases as the system
arXiv:1707.07401v1 [cs.IT] 24 Jul 2017

selected pilot and data simultaneously. A collision occurs when load increases [6].
two or more devices choose the same pilot sequence. Collisions On the other hand, authors in [7] showed that the maximum
are treated as interference to the remaining received signals. sum-rate is achievable through non-orthogonal multiple access
We consider successive joint decoding (SJD) and successive (NOMA) and joint decoding. This is because NOMA allows
interference cancellation (SIC) under a Rayleigh fading and path
loss channel model. We first derive the expression for the outage multiple users to share the different resources, e.g., time, fre-
probability for the case where devices transmit at the same fixed quency, and space, either through power domain multiplexing
rate. Then, we derive the expression for the maximum achievable or code domain multiplexing [1], [8], [9]. Thus, unlike orthog-
throughput for the case where devices transmit with rateless onal multiple access, overloading is possible at the expense
codes. Thus, their code rate is adaptive to the system conditions, of increased processing complexity at the receiver [9]. As
i.e., load, received powers, and interference. Numerical results
verify the accuracy of our analytical expressions. For low data mMTC communications are uplink-oriented, this complexity
rate transmissions, results show that SIC performs close to that is at the access point (AP) and is, thus, acceptable. NOMA
of SJD in terms of outage probability for packet arrival rates up allows multiple users to share time and frequency resources
to 10 packets per slot. However, SJD can achieve almost double in the same spatial layer via power domain or code domain
the throughput of SIC and is, thus, far more superior. multiplexing. In what follows, we highlight some of the recent
works conducted on NOMA so far.
I. I NTRODUCTION
A. Related Work

M Assive Machine Type Communication (mMTC) is ex-


pected to play an important role in future 5G networks
[1]. mMTC handles the connections between the massive
For uplink coordinated NOMA, the problem of user
scheduling, resource allocation, and power control has been
widely investigated [10]–[18]. Authors in [13] proposed a
number of machine type devices (MTDs) which are char-
novel interference cancellation technique, called Triangular
acterized by their short packets, and low computational and
SIC, for asynchronous NOMA. Triangular SIC detects and
storage capabilities. Moreover, in most cases, MTDs have
exploits a priori information of all the overlapping symbols.
fixed locations and need to operate on a low power budget.
Authors also presented the bit-error rate for their proposed
Previous works have shown that coordinated access is not
technique. In [18], authors proposed a physical layer network
suitable for mMTC due to the large control overhead. The
coding scheme and cascade-computation decoding scheme for
control overhead can be as large as the payloads themselves
the case where a fixed number of users are transmitting over
in a large network which is very inefficient [2]. In contrast,
a fading channel. Their results show that cascade computation
uncoordinated access operates without or with minimal control
can outperform the iterative detection and decoding scheme.
overhead at the expense of collisions due to devices contending
Authors in [16] considered a single-cell scenario where active
to access the channel.
users follow a Poisson Point Process (PPP). All the active
One of the most common uncoordinated access schemes is
users are known at the AP and are scheduled together. Authors
the slotted ALOHA protocol. Many variants of this protocol
derived the expression for the achievable spectrum efficiency
have been proposed over the past decade [3]–[5] that aim at
for Nakagami fading. In [17], authors considered a multi-cell
improving the system throughput. However, as these protocols
scenario. Using the theory of ordered statistics and Poisson
are orthogonal in nature, i.e., transmissions take place over a
Cluster Process, they characterized the rate coverage proba-
set of non-overlapping resource units, the number of devices
bility for perfect and imperfect SIC. Their results show that
that can be supported is dependent on the number of avail-
NOMA outperforms its counterpart orthogonal multiple access
able resource units. Moreover, from an information theoretic
cluster in the cases of higher number of users per cell and
The authors R. Abbas, M. Shirvanimoghaddam, Y. Li and B. Vucetic higher target rate requirements.
are with the Center of Excellence in Telecommunications, School For uncoordinated NOMA, the problem is that the set of
of Electrical and Information Engineering, The University of Sydney, active users as well as their respective channel conditions are
NSW, Australia (email: {rana.abbas; mahyar.shirvanimoghaddam; yonghui.li;
branka.vucetic}@sydney.edu.au). not known a priori at the AP. Different approaches to solve
This paper was submitted in part to PIMRC 2017. this problem have been proposed. In [19], a message passing
2

algorithm was proposed to jointly detect user activity and their TABLE I
data. This problem was also solved in [20] using compressive N OTATION S UMMARY
sensing. In their technique, the estimated user set in each Notation Description
slot depends on prior information from previous transmissions. M Number of symbols in a time slot
This is valid when there exist temporal correlations between K Number of information bits in a packet
q Number of symbols in a pilot sequence
user transmissions. In [21], another user detection technique N Set of transmitting devices
was proposed using a set of orthogonal pilot sequences which L Set of pilot sequences
are chosen uniformly at random by the set of active users. Ls Set of singleton pilot sequences/layers/devices
As long as every pilot sequence is chosen by only one user, Nℓ Set of devices that chose the ℓth pilot sequence
Z Set of devices in collision (interfering)
users can be accurately detected at the receiver side and PT Maximum transmit power of a device
their channel state information can be accurately estimated. Pi Received power of device i
However, even when the number of pilot sequences is very P̂i Received power of the device in ℓth singleton layer
large, the probability that two or more users choose the same ρℓ SINR of layer ℓ, ℓ ∈ Ls
ρ̂ℓ SINR of the ℓth singleton layer
pilot sequence is non-zero. In this case, a collision is said to Rc Device code rate
have occurred as these devices cannot be distinguished by the Rf Device effective rate
AP. While authors in [22] suggest that devices adopt some
back-off scheme to resolve this collision, this implies that any pilot sequences scales linearly with the packet arrival rate. Our
collision will lead to the loss of all data including those devices results also show that SIC achieves a similar performance as
that did not collide. This is very wasteful of resources. SJD while reducing the decoding complexity.

B. Contributions 3) Performance Analysis for Successive Interference Can-


cellation: We then consider the case where the devices trans-
Motivated by these findings, this paper investigates the per- mit using rateless codes. In this case, the rate is determined
formance of an uncoordinated massive NOMA scheme where on the fly and varies from slot to slot based on the system
user detection and channel estimation is carried out via pilot load, received powers and interferers. The receiver stops
sequences that are transmitted simultaneously with the device’s transmissions by broadcasting a beacon when the throughput
data. In particular, we investigate the performance of massive is maximized. We derive the expression for the maximum
NOMA with collisions, which is missing from previous works throughput for the case of joint decoding and successive in-
in this area. If the AP is not able to resolve the collisions, they terference cancellation. The evaluation of the exact expression
are regarded as interference. In this case, other devices that is shown to be very complicated. Based on this, we propose
did not collide in the given slot can still be recovered. This a simplified expression and demonstrate its accuracy through
allows for a more practical transmission scheme along with simulations. Our results show that the maximum throughput
a more rounded assessment of its performance and suitability of SJD is almost double that of SIC. However, we explain
for mMTC. All in all, the work provides the following three that the maximum throughput under SIC is achievable in
major contributions. practice whereas the existence of codebooks that can achieve
1) Grant-Free Massive NOMA Scheme: We consider an the maximum throughput in the case of SJD is questionable.
uplink grant-free NOMA setting where devices jointly transmit
a randomly chosen pilot sequence along with their data. For The rest of this paper is organized as follows. Section II
this setting, there is always a non-zero probability that two describes the system model and the considered transmission
or more devices choose the same pilot sequence. The receiver scheme for uncoordinated massive NOMA. In Section III, we
is only able to estimate their aggregate power. However, it is show that the aggregate power of collisions can be modelled
unable to distinguish the devices from one and another, and as a PPP and present its distribution. In Section IV and V, we
a collision is said to have occurred. In this work, we propose derive the outage probability and maximum system throughput
to treat these codewords as interfering signals at the AP. for uncoordinated NOMA under massive access for SJD and
We derive the distribution of the number of collided devices SIC, respectively. Numerical results are presented in Section
and show that the aggregate interference power can be well- VI. Practical considerations are discussed in Section VII, and
approximated by a PPP. Finally, we present the characteristic conclusions are drawn in Section VIII.
function of the aggregate interference power which is an
essential parameter in the performance analysis of this system. In this paper, we denote by fX (x), FX (x), ψX (ω) and
2) Outage Probability of Massive NOMA: For the proposed µX (n) the probability density function (PDF), the cumulative
framework, we first consider the case where all the devices density function (CDF), the characteristic function (CF), and
transmit at the same fixed code-rate. We derive the expression the nth moment of X, respectively. The cardinality of the set
of the outage probability for the case of joint decoding and X is denoted by |X | = X. All logarithms are taken to the
successive interference cancellation. The evaluation of the base 2, unless otherwise indicated. C(x) = log(1 + x) is the
exact expression is shown to be daunting especially for the point-to-point Gaussian channel capacity with x denoting the
case of joint decoding. To overcome this problem, we propose signal-to-noise
√ ratio (SNR), Γ(·) is the Gamma function, and
a simplified expression and demonstrate its accuracy through j = −1. Finally, δ(x) is the indicator function such that
simulations. Our results show that the optimal length of the δ(x) = 1 if x > 0 and is zero otherwise.
3

II. S YSTEM M ODEL B. Transmission Scheme

A. Overview
In each slot, each device chooses a pilot sequence of length
We model the location of the devices transmitting in any q symbols independently and uniformly at random from a set
time slot as a homogeneous Poisson point process (PPP) L = {1, 2, ..., L}. We categorize the pilots sequence into three
on an annular region with minimum and maximum radii types. An idle pilot sequence is a pilot sequence which has
dmin and dmax [23]–[27], respectively. The AP is located at not been chosen by any device. A singleton pilot is a pilot
the center, and the average number of transmitting devices sequence chosen by only one device, and a collision pilot
surrounding the AP per time slot is denoted by λ. Devices sequence is a pilot sequence chosen by two or more devices.
are considered to be static, and the channel is modelled as a Each pilot sequence ℓ is made unique to a specific code book
block fading channel. That is, the devices’ channel conditions Cℓ and, thus, acts as the device’s signature. A device i that
remain constant for the duration of one packet and vary has chosen the pilot sequence ℓ encodes its data bi into a
randomly and independently from one slot to the other. For codeword xi = fℓ (bi ).
reciprocal channels, each device can make use of the pilot
signal sent periodically over the downlink channel by the AP In practice, the set of received pilot sequences is determined
to synchronize their timing to that of the AP. The impact of by the AP via the power delay profile which is constructed
asynchrony is discussed in Section VII. by performing cross-correlations between the known pilot
For each time slot, the AP initiates uplink transmissions sequences and the received signals and averaging the absolute
through beaconing. To minimize signalling overhead, we as- square values of the created channel impulse responses [28].
sume that the beacon signal carries load information based on This is based on the fact that the auto-correlation of orthogonal
which the devices adjust their code rate. The received signal pilot sequences can be approximated by a delta function, and
at the AP for a given time slot t can be expressed as: its cross-correlation with other pilot sequences yields all-zero
sequences [28]. Thus, it is essential that the chosen set of pilot
(t) sequences have good auto and cross-correlation properties,
X
y(t) = xi + w(t) , (1)
i∈N (t)
e.g., Zadoff-chu [29], Golden codes [30], m-sequences [31]. In
what follows, we assume perfect pilot sequence detection and
(t)
where xi is the codeword transmitted by device i from the power estimation, i.e., perfect device detection and channel
total set of transmitting devices N (t) with a received power of estimation.
(t)
Pi , and w(t) is a circular symmetric white Gaussian noise Some examples are illustrated in Fig. 1. In Fig. 1(a), the
with unity variance (could include inter-cell interference). We four transmitting devices choose unique pilot sequences (1, 2,
consider Rayleigh fading and a path-loss channel model. Thus, 4 and 8). Thus, there are no collisions. The received powers
(t) (t)
Pi = |hi |2 d−α i PT , where hi , di , α and PT denote the are estimated as P1 , P2 , P4 , and P8 . The remaining pilot
small scale fading gain of the ith device, the distance between sequences (3, 5, 6, 7, 9 and 10) are idle pilot sequences. In Fig.
the ith device and the AP, the path-loss exponent and the 1(b), three devices choose the first pilot sequences. Devices
transmit power, respectively. In what follows, we consider tight that choose the same pilot sequence also choose the same code
latency requirements where no retransmission opportunities book. These devices cannot be distinguished by the AP and
are available, i.e., if the transmission of the packet is not are said to have collided. We will refer to each code book as a
successful in one slot, the packet is dropped. Thus, for the layer. Thus, these devices have collided over the first layer. In
remaining part of the paper, we focus our analysis on a single what follows, we assume that the AP can distinguish between
time-slot and, thus, we drop the superscripts. collision layers and singleton layers. However, for collision
In this work, each codeword is concatenated with a pilot layers, the AP does not know the number of colliding devices
sequence of length q symbols for user detection and decoding. and is, thus, unable to decode the collision layers. Instead,
Thus, each codeword is of length M − q symbols. Assuming the AP treats them as interference whose aggregate power can
each device has a set of K information bits to transmit, the be estimated from the power delay profile. The feasibility of
code rate per device is defined as these assumptions will be discussed in Section VII.

K We denote by Nℓ the set of devices that have chosen the


Rc := bits/symbol. (2) ℓth pilot sequence of size Nℓ . For a sufficiently large number
M −q
of pilot sequences, the random variables N1 , N2 , ..., NL are
On the other hand, the effective rate per device is defined as Poisson random variables with average λ/L. This is a practical
assumption as the length of the pilot sequences need to be
K large enough for the AP to accurately calculate the received
Rf := bits/symbol. (3) powers, and the number of pilot sequences increases with its
M
length [29]. The set of devices that have collided is denoted
Rf represents the ratio of the number of information bits to by Z = {i i ∈ Nℓ , Nℓ > 1, ∀ℓ ∈ L}. Moreover, the set of

the total number of symbols transmitted. Thus, the effective
singleton layers is denoted by Ls = {ℓ Nℓ = 1, ∀ℓ ∈ L}.

rate takes into consideration the redundancy incurred by the
pilot sequence. Then, the equivalent received signal as seen by the AP can be
4

4
A. Distribution of Received Power
1 1
3
2
4 1 2
For the case where devices always transmit with maximum
2

8
1 8
transmit power PT over a channel subject to Rayleigh fading
8 7
and path-loss, the distribution of the received power of a
singleton layer is given in the following corollary. We refer
C1
C1
the readers to Appendix A for the proof. We make note that
C1
C1
C2
C3
the extension of this to Nakagami fading or any other fading
C2 C4
C4
C8
C7
C8
C2
C8
distribution is straight-forward.
q M −q q M −q q M −q

P8 Inteference P7 P8 Corollary 1. For a device uniformly distributed in an annulus


P4 P4 P8
P1 P2 P2
P3
P2
with a minimum radius of dmin and a maximum radius of dmax ,
the CDF of the received power P = |h|2 d−α (PT = 1) at the
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
origin under Rayleigh fading is
(a) (b) (c)

2
p− α Γ α2 + 1
 
Fig. 1. Different case scenarios of a network with L = 10 pilot sequences: d2
(a) N = 4, (b) N = 8 and (c) N = 2. FP (p) = 1 − 2 2 + 2 min 2 . (7)
dmax − dmin dmax − dmin
expressed as
X X B. Aggregate Interference Power
y= xn + xn′ + w, (4)
n∈Nℓ ,ℓ∈Ls n′ ∈Z We now proceed to find the distribution of the aggregate
interference power I. For that, we first find the exact distri-
where xi = fℓ (bi ) for i ∈ Nℓ . The received signal-to-
bution of the number of interferers in the following lemma,
interference-and-noise-ratio (SINR) of the singleton layers is
i.e., number of devices in collision Z. We refer the readers to
given as
Appendix B for the proof.
Pi
ρℓ = , for i ∈ Nℓ , ℓ ∈ Ls , (5) Lemma 1. Consider a total of L layers where packet arrivals
I + σ2
where I denotes the aggregate interference power caused by over each layer are Poisson distributed with an average of
these collisions given as: λ/L. Given that the number of singleton layers is Ls < L,
X the probability of having a total of n packets collide in a given
I= |hn |2 d−α
n PT . (6) time slot is expressed in (8) (top of the next page).
n∈Z
Although Z devices are only Poisson distributed in space
It is worthy of pointing out the main differences between this
and not in number, our results in Fig. 2 show that their aggre-
transmission scheme and that of Long-term Evolution (LTE).
gate power can be well-approximated by a PPP. Accordingly,
In LTE [32], the set of orthogonal sequences used for user
the distribution of the aggregate interference power I, given
detection are called preambles. Users transmit their chosen
Ls , can be well-approximated by the skewed truncated stable
preambles separately from their data on a dedicated random
distribution [33], and its CF can be expressed as
access channel, and the AP allocates resources on the uplink
α
ψI (jω) = eγI Γ(−aI )((gI −jω) −gI I )
αI
data channel for each of the detected preambles. A collision , (9)
occurs when two or more devices choose the same preamble.
These devices will transmit their packets over the same time- where the parameters αI , gI and γI determine the shape of
frequency resources, and the AP will not be able to decode the distribution. In particular, the parameters αI and γI are
them correctly. Although many techniques have been proposed related to the dispersion and the characteristic exponent of the
to resolve these collisions, e.g. [15], LTE remains a form of stable distribution, respectively, and the parameter gI is the
coordinated orthogonal transmission scheme and, thus, is not argument of the exponential function used to smooth the tail
suitable for massive access as explained in Section I. More of the stable distribution. These parameters are found through
importantly, the control signals and the data are transmitted the method of the cumulants [33] and are given as
over separate channels at different times which was shown to 2
be strictly sub-optimal for small payloads [2]. αI = , (10)
α
κI (1)(1 − αI )
III. P RELIMINARIES gI = , (11)
κI (2)
In this section, we first derive the distribution for the
−κI (1)
received power of a single user randomly located in the γI =  αI −1 . (12)
cell. Then, for Poisson packet arrivals, we show that the κI (1)(1−αI )
Γ[−αI ]αI κI (2)
colliding/interfering devices can be well-approximated by a
Poisson point process (PPP). Based on this, we find the Here, κI (n) denotes the nth cumulant of the interference
characteristic function  for the aggregate power I defined
(CF) power and is given as
as ψI (jω) := E e−jωI . These parameters will be useful in

2λZ d2−nα 2−nα
min − dmax
evaluating the outage probability and the throughput of the κI (n) = µv (n)PTn , (13)
considered uncoordinated NOMA setting. nα − 2 d2max − d2min
5

 n (L−Ls )λ
 
(L−Ls )λ
e− min{L−Ls −1,n}  λ n−c −(L−Ls −c) L
λ
L

(L − L − c) e

L L − Ls
  c
X s
−λe−λ L

Pr Z = n Ls = 1 + .

n!(1 − λe−λ )L−Ls c (n − c)!
c=1
(8)

A. Outage Probability with SJD


0.06
= 10 Consider the case where all the devices encode their data
K
= 20
using a fixed rate code with code rate Rc := M−q . In this
PDF of Aggregate Interference Power

0.05 = 30
case, we are interested in characterizing and evaluating the
outage probability of the system. That is, the probability that
0.04
a device is not successfully decoded in a given time slot. Given
Ls singleton layers with a received power vector P̂, subject
0.03
to an interference power I, a successive joint decoder first
tries to decode all Ls layers jointly. However, if the equal-rate
0.02 point (R, ..., R) ∈ RL+ is outside the Ls -dimensional capacity
s

region defined by the received SINRs ρ̂, the AP will try to


0.01 decode the strongest Ls − 1 devices by treating the weakest
layer as interference. The weakest layer is defined as the
0 singleton layer with the smallest received power. Similarly, if
-100 -90 -80 -70 -60 -50 -40 -30 Ls −1
the equal-rate point (R, ..., R) ∈ R+ is outside the Ls − 1-
Aggregate Inteference Power: I (dB)
dimensional capacity region, the AP will try to decode the
Fig. 2. Aggregate interference power for L − Ls = 200. The solid lines strongest Ls − 2 layers by treating the two weakest layers as
correspond to the actual distribution, and the dashed lines correspond to the interference. The process repeats until decoding is successful
Poisson distribution. Readers are referred to Table II for the remaining system
parameters.
or until there are no more layers to decode. Now, consider the
descending ordered set P̂(1) ≥ P̂(2) ≥ . . . ≥ P̂(Ls ) .
where v = |h|2 is a chi-squared distributed random vari- For that, the outage probability is defined as the average
able under Rayleigh-fading,
µv (n) is its nth moment, and number of singleton layers that cannot be decoded success-
λZ (Ls ) := E[Z Ls ].
fully. Given Ls singleton layers with a received power vector
Finally, the inversion theorem [33] dictates that the P, subject to an interference power I, the fraction of layers in
cumulative-density function (CDF) of I can be computed as outage can be expressed as
Z ∞ 
ψI (−jω)ejωx − ψI (jω)e−jωx

1 1 ℓ
FI (x) = − Re dω, ǫSJD (Ls , I, P̂) = 1 − max , (16)
2 2π 0 jω 0≤ℓ≤Ls Ls
(14) Pℓ !
and the probability density function (PDF) of I can be 1 P̂(c)
s.t. Rc ≤ min log 1 + PLs c=ℓ−i+1 .
computed as 1≤i≤ℓ i ′ =ℓ+1 P̂(c′ ) + I + σ
2
Z ∞ c
1
fI (x) = e−jωx ψI (−jω)dω. (15) Here, ℓ denotes the maximum number of layers that can be
2π 0 decoded out of the Ls singleton layers. In general, the outage
IV. P ERFORMANCE OF M ASSIVE NOMA WITH probability can be expressed as
S UCCESSIVE J OINT D ECODING "L
s
#
X ℓ 
In this section, we consider the case where all the singleton ǫSJD = 1 − ELs ,I,P̂ Pr φ̄Ls ,Ls , . . . , φ̄ℓ+1,Ls , φℓ,Ls ,
Ls
layers are jointly decoded at the receiver using a maximum ℓ=0
likelihood decoder. Although jointly decoding all of the sin- (17)
gleton layers is optimal in terms of the common outage event, where φℓ,Ls =
i.e., the probability that at least one layer decoded incorrectly, Pℓ !!
it is not optimal in terms of the individual outage event, i.e., 1 c=ℓ−i+1 P̂(c)
the probability that one layer is decoded incorrectly. Therefore, δ Rc ≤ min log 1 + PLs ,
1≤i≤ℓ i 2
c′ =ℓ+1 P̂(c′ ) + I + σ
we consider the successive joint decoder (SJD) where a
subset of the layers can be decoded jointly while regarding Thus, to find the outage probability, not only do we need
the remaining layers as interference. For that, we derive the to average over the different number of singleton layers as
expressions for the outage probability and the throughput. For well as the different aggregate interference powers, we also
ease of notation, we define P̂ := [P̂ℓ ]1≤ℓ≤Ls , where P̂ℓ is need to average over the numerous realizations of the received
the received power of the device in the ℓth singleton layer. power vector. Thus, computing the outage probability for
Similarly, we define ρ̂ := [ρ̂ℓ ]1≤ℓ≤Ls , where ρ̂ℓ is the SINR uncoordinated multiple access is even more daunting than
of the ℓth singleton layer. that for coordinated multiple access [34]. In what follows, we
6

make use of the massive aspect of mMTC to simplify this throughput is maximized through beaconing. The throughput
expression. is defined as the ratio of the number of successfully decoded
Consider a set of L i.i.d. random variables [Xi ]1≤i≤L , where information bits to the length of the transmitted codeword
X1 ≥ X2 ≥ ... ≥ XL . Then, from ordered statistics [35], we (M − q). This is determined by the AP and can be made
have known to the devices through beacons. Given Ls singleton
−1
 layers, a received power vector P̂, subject to an interference
lim Pr |X⌊yL⌋ − FX (1 − y)| > ν = 0, (18) power I, the maximum instantaneous throughput is given as
L→∞

for all y ∈ [0, 1] and ν > 0. Based on this, for an asymptoti- ζSJD (Ls , I, P̂) =
cally large number of singleton layers, the following equality Pℓ !
holds for all ℓ1 , ℓ2 ∈ [1, Ls ]. ℓ c=ℓ−i+1 P̂(c)
max min log 1 + PLs .
1≤ℓ≤Ls 1≤i≤ℓ i 2
Xℓ2 Z ℓ2 /Ls c′ =ℓ+1 P̂(c′ ) + I + σ
lim P̂(i) =Ls FP−1 (1 − y)dy As the throughput varies from slot to slot, the maximum aver-
Ls →∞ ℓ1 /Ls
i=ℓ1 age throughput should be averaged over all possible values of
=Ls (g(ℓ2 /Ls ) − g(ℓ1 /Ls )) , (19) Ls , I and P̂. However, from (18), the maximum instantaneous
system throughput under massive access can be expressed
where g(y) := FP−1 (1 − y)dy. For Rayleigh fading and a
R
from [34] as
path-loss channel model, g(y) is evaluated by integrating the
inverse of (7) and is expressed in (20). For the special case of ζSJD (Ls , I, P̂) =
Rayleigh fading only, we have
!
1 g(v) − g(v − uv)
max min log 1 + 2 . (24)
gRay. (y) = (1 − y) loge (1 − y) − (1 − y). (21) 0≤v≤1 0<u≤1 u g(1) − g(v) + I+σ
Ls

Using this property, the complexity of computing the outage Then, the average maximum system throughput under massive
probability of massive NOMA in (17) is significantly reduced access is derived in the following lemma. We refer the readers
as P converges to a deterministic value as Ls → ∞. The to Appendix D for the proof.
expression for that is derived in the following lemma. We refer Lemma 3. For L pilot sequences of length q and a packet
the readers to Appendix C for the proof. We make note that in arrival rate of λ packets, the average maximum system
practice, we find that this property gives sufficiently accurate throughput of massive NOMA with SJD can be expressed as
results for Ls > 10. in (25).
Lemma 2. Consider a time slot of length M symbol durations.
V. P ERFORMANCE OF M ASSIVE NOMA WITH
For a code rate Rc , L pilot sequences of length q and a packet
S UCCESSIVE I NTERFERENCE C ANCELLATION
arrival rate of λ packets, the outage probability of the massive
NOMA system with SJD can be expressed as As joint decoding is often infeasible in practice due to
Z 1 Z ∞  complexity constraints, we consider successive interference
v ∗ cancellation (SIC) in this section. In each stage of SIC, the
ǫSJD = 1−ELs e−jωISJD (v) ψI (−jω)dω , (22)
0 2π 0
strongest layer is decoded by regarding remaining layers as
∗ interference. Thus, we only need a single-user decoder. If the
where ISJD (v) is the solution to the equation below
layer is decoded successfully, the decoded layer is subtracted
! from the received signal. We assume perfect SIC. In the second
1 g(v) − g(v − uv)
Ls Rc − min log 1 + 2 = 0. stage, the second strongest layer is decoded by regarding the
0<u≤1 uv g(1) − g(v) + I+σ Ls remaining layers as interference. SIC stops when a layer is
(23) decoded unsuccessfully or when there are no more layers to
Here, I ∗ (v) denotes the largest interference power for which decode.
SJD
the largest fraction of layers that can be decoded successfully
A. Outage Probability with SIC
is v. The integral in Lemma 3 averages over the aggregate
interference power, and the expectation averages over the Given that the strongest ℓ − 1 layers have been decoded
number of singleton layers. successfully, the ℓth strongest layer is decoded successfully if
the following condition is true.
B. Maximum Throughput with SJD P̂(ℓ)
PLs ≥ 2 Rc − 1 (26)
2
We now consider the case where devices can transmit as c=ℓ+1 P̂(c) + I + σ
many coded symbols as necessary. This is feasible when Thus, given that the strongest ℓ − 1 layers have been decoded
devices use rateless channel codes to encode their information successfully, the probability that the ℓth strongest layer is
[36]. With rateless codes, the length of the codeword is decoded successfully is given as
determined on the fly and is adaptive to the load, received   PLs 
powers and interference power. Thus, the duration of the slot P̂(ℓ) − 2Rc − 1 c=ℓ+1 P̂(c) + σ
2

(M ) as well as the devices’ code rate (Rc ) also varies from one Pr I ≤ 
2 Rc − 1
slot to the other. Transmissions are stopped by the AP when the
7

2 (d2max − d2min )y − d2min



g(y) =  −α α
. (20)
(α − 2)(d2max − d2min ) Γ α2 + 1 2 ((d2max − d2min )y − d2min ) 2


 ∞  
1 ψI (−s)  −s((g(1)−g(v))Ls +σ2 )
Z
−s((g(1)−g(v−uv))Ls +σ2
ζSJD = ELs max min e −e ds . (25)
0≤v≤1 0<u≤1 u 0 s

From (18), given Ls and I, the SINR at each stage of the SIC is TABLE II
deterministic under massive access. Based on this, the outage S YSTEM PARAMETERS
probability under SIC is derived in the following proposition. Parameter Value
The definition of the outage probability is as given in Section Cell Radius 500 m
IV-A, and the proof of the proposition is similar to that of Reference Distance 50 m
Bandwidth 15 kHz
Lemma 2. Transmit Power 23 dBm
Thermal Noise Density -174 dBm/Hz
Proposition 1. For a code rate Rc , L pilot sequences of length Receiver Noise Figure 3 dB
q and a packet arrival rate of λ packets, the outage probability Path loss exponent 3.5
of massive NOMA with SIC can be expressed as
Z 1 Z ∞  suffers from collision. We also assume that q = L [29]–
v ∗ [31]. This means that the overhead associated with channel
ǫSIC (R) = 1−ELs e−jωISIC (v) ψI (−jω)dω ,
0 2π 0 estimation and user detection scales linearly with the packet
(27) arrival rate. In what follows, we evaluate the performance of
massive NOMA for different collision probabilities.
where
First, we evaluate the outage probability for massive
FP−1 (v) − 2Rc − 1 (g(1) − g(v))Ls + σ 2
 
∗ NOMA. Each time slot is of length M = 2000 [37] over
ISIC (v) := .
2 Rc − 1 which the devices transmit with a code rate of 0.1. Results for
(28) SJD and SIC are shown in Fig. 3 and Fig. 4, respectively.
∗ We first observe that the analytical results from Lemma 3
Here, ISIC (v) denotes the largest interference power for which
and Proposition 1(PROP 1) provide a good approximation of
the largest fraction of layers that can be decoded successfully
the actual performance obtained from Monte Carlo simula-
is v.
tions. When comparing the performance for different collision
probabilities, we find that the gap between the number of
B. Maximum Throughput with SIC successfully decoded devices and the number of singleton
We now consider the same setup in Section IV-B. Given Ls layers becomes more significant as β increases for both
singleton layers with a received power vector P̂, subject to an SJD and SIC. This is because collisions in NOMA act as
interference power I, the maximum instantaneous throughput interference to the remaining singleton layers which leads to
is given as more outages.
In Fig. 3(c) and Fig. 4(c), the effective sum-rate is evaluated
ζSIC (P, Ls , I) = as the product of the number of successfully decoded devices
K

FP−1 (1 − v ′ )
 and the effective rate M . In general, if the number of pilot
Ls max v log 1 + min . sequences is too small, the number of singleton layers will be
0<v≤1 ′ 0<v ≤v (g(1) − g(v ′ ))Ls + I + σ 2
low. Thus, the number of successfully decoded devices will
(29)
be low as well. On the other hand, when the number/length
The average maximum system throughput is given in the of pilot sequences is too large, the overhead will be large.
following proposition. The proof of this proposition is similar Thus, the effective rate will be low. Interestingly, it seems that
to that of Lemma 3. the optimal value of L scales linearly with the packet arrival
rate. Finally, in Fig. 3(d) and Fig. 4(d) we plot the maximum
Proposition 2. For L pilot sequences of length q and with a
payload sizes that can be transmitted and compare them to the
packet arrival rate of λ packets, the average maximum system
overhead induced by the pilot sequences. In general, we can
throughput of massive NOMA under SIC can be evaluated
say that for applications with payloads of size 100-200 bits and
from (30).
a target outage probability of ≈ 0.01, we can simultaneously
support up to 7 devices with SJD and up to 6 devices with
VI. N UMERICAL R ESULTS SIC.
In our simulations, we consider single-tone transmissions In Fig. 5, we evaluate the outage probability for a larger
with the system parameters listed in Table II [37]–[39]. code rate (Rc = 1). Here, we notice that the performance
The number of pilot sequences L is determined such that gap between SJD and SIC is more significant. In fact, from
λ
1 − e− L = β, where β denotes the probability that a device Fig. 5(b), we can see that SIC can barely decode any packet.
8

 ∞  
ψI (−s) −s((g(1)−g(v′ ))Ls +σ2 ) 
Z
−sFP−1 (1−v ′ )
ζSIC = ELs max vLs min e 1−e ds . (30)
0<v≤1 ′ 0<v ≤v 0 s

100 10 0.9 1000


SIM: = 0.10 SIM: = 0.10 = 0.10 K: = 0.10
900
LEMMA 2: = 0.10 9 E[L ]: = 0.10 = 0.05 q: = 0.10
s 0.8
SIM: = 0.05 = 0.01 K: = 0.05

Effective Sum-Rate (bits/symbol)


SIM: = 0.05 800

Number of Successful Devices

Number of Information Bits (K)


LEMMA 2: = 0.05 q: = 0.05
8 E[L s]: = 0.05
) SJD

SIM: = 0.01 0.7 K: = 0.01


SIM: = 0.01 700
LEMMA 2: = 0.01 q: = 0.01
10-1
E[L s]: = 0.01
Outage Probability (

7
600
0.6
6 500
0.5
400
5
10-2
0.4 300
4
200
3 0.3
100

10-3 2 0.2 0
2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10
Packet Arrival Rate ( ) Packet Arrival Rate ( ) Packet Arrival Rate ( ) Packet Arrival Rate ( )
(a) (b) (c) (d)

K
Fig. 3. Outage Probability for massive NOMA with SJD for different values of β. The code rate M −q
is equal to 0.1 and the slot duration is M = 2000
symbols.

100 10 0.9 1000


SIM: = 0.90 SIM: = 0.90 = 0.90 K: = 0.90
PROP 2: = 0.90 9 E[L s]: = 0.90 = 0.95
900 q: = 0.90
0.8
SIM: = 0.95 = 0.99 K: = 0.95
SIM: = 0.95

Effective Sum-Rate (bits/symbol)


800
Number of Successful Devices

Number of Information Bits (K)


PROP 2: = 0.95 8 E[L ]: = 0.95 q: = 0.95
)

SIM: = 0.99 s K: = 0.99


SIC

= 0.99
0.7 700
PROP 2: = 0.99 q: = 0.99
10-1 7
E[L s]: = 0.99
Outage Probability (

600
6 0.6
500
5 0.5
400
-2 4
10
0.4 300
3
200
0.3
2 100

10-3 1 0.2 0
2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10
Packet Arrival Rate ( ) Packet Arrival Rate ( ) Packet Arrival Rate ( ) Packet Arrival Rate ( )
(a) (b) (c) (d)

K
Fig. 4. Outage Probability for massive NOMA with SIC for different values of β. The code rate M −q
is equal to 0.1 and the slot duration is M = 2000
symbols.

100 10 7 2000
SIC = 0.10 = 0.10
9 1800
E[L s]: = 0.10 = 0.05
6 SJD
SJD = 0.01
Effective Sum-Rate (bits/symbol)

8 = 0.05 1600
Number of Successful Devices

Number of Information Bits (K)

E[L s]: = 0.05 SJD


5
Outage Probability ( )

7 = 0.01 1400
10-1
E[L s]: = 0.01 K: = 0.10
6 1200 q: = 0.10
4
K: = 0.05
5 1000 q: = 0.05
3 K: = 0.01
4 800 q: = 0.01
10-2
3 2 600
SIC
2 SIC 400
= 0.10
= 0.05
1
1 200
= 0.01

10-3 0 0 0
2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10
Packet Arrival Rate ( ) Packet Arrival Rate ( ) Packet Arrival Rate ( ) Packet Arrival Rate ( )
(a) (b) (c) (d)

K
Fig. 5. Outage Probability for massive NOMA with SJD and SIC for different values of β. The code rate M −q
is equal to 1 and the slot duration is
M = 2000 symbols.

However, although the outage probability of SJD is better, it of that of SIC, thus, SJD is in theory far more superior in
is still too high for any practical use. Moreover, as mMTC are this case. However, in the following section, we explain that
characterized by their low data rate transmissions, SIC seems the maximum throughput under SIC is achievable in practice
to be a good candidate for our uncoordinated massive NOMA whereas the achievability in the case of SJD is questionable.
scheme when low processing complexity is needed. Finally, we make note that in practice M should not be larger
Next, we evaluate the throughput for massive NOMA in than the coherence time. In this case, collision probabilities
Fig. 6 and Fig. 7. In this setup, each device has a set lower than 0.01 cannot be supported by this scheme as it will
of 1024 information bits to transmit [37]. The number and require longer pilot sequences.
length of pilot sequences is determined as before. We observe
VII. P RACTICAL C ONSIDERATIONS
that the analytical results from Lemma 4 and Proposition 2
(PROP 2) also provide a good approximation of the actual A. Detection of Collision Layers and Achievability
performance obtained from Monte Carlo simulations. We also For SIC, the detection of collision layers can be imple-
observe that the throughput gain of SJD is almost double mented through a simple algorithm. At first, the AP assumes
9

14 10 7 2500
= 0.10 q: = 0.10
9 = 0.05 M-q: = 0.10
12 6

Average Transmission Time (symbols)


)

= 0.01 M: = 0.10
SJD

2000 q: = 0.05

Number of Successful Devices


8
M-q: = 0.05

Average Code Rate (R c)


10 5
Average System Throughput (

M: = 0.05
7
q: = 0.01
1500 M-q: = 0.01
8 6 4
M: = 0.01

6 5 3
1000
SIM: = 0.10
SIM: = 0.10 4 E[L ]: = 0.10
s
4 LEMMA 3: = 0.10
2
SIM: = 0.05
SIM: = 0.05 3 E[L ]: = 0.05
s 500
LEMMA 3: = 0.05
2 SIM: = 0.01 1
SIM: = 0.01 2
E[L s]: = 0.01
LEMMA 3: = 0.01
0 1 0 0
2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10
Packet Arrival Rate ( ) Packet Arrival Rate ( ) Packet Arrival Rate ( ) Packet Arrival Rate ( )
(a) (b) (c) (d)

Fig. 6. Average system Throughput for massive NOMA with SJD for different values of β (K = 1024).

14 10 7 2500
SIM: = 0.10 SIM: = 0.10 = 0.10 q: = 0.10
PROP 2: = 0.10 9 E[L s]: = 0.10 = 0.05 M-q: = 0.10
12 6

Average Transmission Time (symbols)


)

SIM: = 0.05 = 0.01 M: = 0.10


SIC

SIM: = 0.05 2000


Number of Successful Devices

PROP 2: = 0.05 8 E[L s]: = 0.05 q: = 0.05

Average Code Rate (R c)


SIM: = 0.01 M-q: = 0.05
Average System Throughput (

10 SIM: = 0.01 5
PROP 2: = 0.01 7 M: = 0.05
E[L ]: = 0.01 q: = 0.01
s
1500 M-q: = 0.01
8 6 4
M: = 0.01

6 5 3
1000
4
4 2
3
500
2 1
2

0 1 0 0
2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10
Packet Arrival Rate ( ) Packet Arrival Rate ( ) Packet Arrival Rate ( ) Packet Arrival Rate ( )
(a) (b) (c) (d)

Fig. 7. Average system Throughput for massive NOMA with SIC for different values of β (K = 1024).

that all the layers are singleton layers and orders the layers if not null. We would like to point out that as most MTDs
according to their respective received powers in descending are stationary devices, their timing advance need not be
order. In the first iteration, the AP attempts to decode the communicated very frequently. Alternatively, time asynchrony
strongest layer. The AP can verify that the decoded informa- can provide another contention unit by which devices can be
tion is correct through a simple CRC-check. If decoding is distinguished from one another without having to increase
successful, the decoded signal is cancelled from the received the pilot sequence length and thus without introducing re-
signal. If not, the AP assumes that this is a collision layer. dundancy. Moreover, time asynchrony can help in detecting
In the second iteration, the AP attempts to decode the second the number of devices in collision layers. This allows the
strongest layer while regarding the previously undecoded layer opporutnity to resolve collisions layers and decode the packets
as interference. The AP can continue to follow the same steps of the involved devices. However, the performance analysis
until there are no more layers to decode. Detecting collision for this case scenario is not straightforward as the maximum
layers for SJD is also possible at the expense of a larger achievable rate with code-reuse is strictly lower than that with
complexity, as the AP would need decode all possible subsets unique codes and is dependent on the codes used [18].
of layers while regarding the remaining layers are interference. In general, the AP can feedback a timing advance signal
When decoding finishes, the AP would choose the largest in the acknowledgement. For that, the operator can either
subset of successfully decoded layers. choose to operate either in a fully synchronous mode or
It is straight-forward to see that the derived bounds for SIC allow the devices to introduce random asynchrony into their
are achievable when capacity achieving point to point channel transmissions to create more contention units [40].
codes are used. However, for SJD, it is uncertain whether
their exists channel codes suitable for grant-free coordinated VIII. C ONCLUSION
access with joint decoding. That is mainly because these In this paper, we considered a massive uncoordinated
codes would not only have to adapt to the different channel NOMA setting where devices choose pilot sequences from
conditions but also to the different channel loads. Finally, we a predetermined set as their signature. Then, each device
would like to point out that there exists practical low-complex encodes its data using the pilot as the signature and transmits
decoders such as belief-propagation-based decoders that can its selected pilot and data simultaneously with the rest of the
score performances close to that of maximum likelihood devices. In our proposed scheme, a collision occurs when
decoding more than one device choose the same pilot sequence. The
set of collided packets are treated as interference. For that,
B. Synchronization we show that the aggregate interference power can be well-
In conventional communication systems, synchronization approximated by a PPP. Based on this, we derive the outage
takes place beforehand. However, it has been well established probability and the maximum system throughput under suc-
now that for massive access control signals should be minimal cessive joint decoding (SJD) as well as successive interference
10

cancellation (SIC) for a Rayleigh fading and path loss channel The last equation follows from the fact that the sum of any c
model. We verified the accuracy of our derived expressions via elements of N is a Poisson distributed random variable with
simulations. Our results show that SIC performs close to that average cλ.
of SJD in terms of outage probability for packet arrival rates We want to calculate the probability that the sum of these
up to 10 packets per slot. In terms of throughput, although SJD elements is equal to some positive integer n given that Ni 6= 1
scores almost double the throughput gain of that of SIC, we for 1 ≤ i ≤ L. This can be expressed as
explained that this throughput might not be achievable with

Pr (B|Sn ) Pr (Sn ) 1 − Pr B̄|Sn Pr (Sn )
practical modulation and channel coding schemes. Pr (Sn |B) = =  .
Pr (B) 1 − Pr B̄
A PPENDIX A Moreover, we have
P ROOF OF C OROLLARY 1 n−c −(L−c)λ
c ((L − c)λ) e
 
Pr A1 ∩ A2 ∩ ...Ac Sn = Pr (A) ,

For a device uniformly distributed in an annulus of min- (n − c)!
imum radius dmin and maximum radius dmax , the CDF of
the distance d with respect to the origin follows the truncated for c ≤ min{L − 1, n}; it is zero otherwise. Given Sn , the
Pareto distribution below probability that at least one of the elements of vector N is
equal to 1 can be expressed as:
d2 − d2min
FD (d) = 2 . L
!
dmax − d2min  [
Pr B̄|Sn = Pr Ai Sn

For X = d−α , i=1
−α L
FX (x) = Pr(d < x) X X
1
= (−1)c+1 Pr (Ai1 ∩ . . . ∩ Aic |Sn )
−α
= Pr(d > x ) c=1 i1 ,...,ic :
2 1≤i1 <...<ic ≤L
x− α − d2min n−1 −(L−1)λ
=1− ((L − 1)λ) e
d2max − d2min = LPr(A) −
(n − 1)!
For Rayleigh fading, we can write   n−2 −(L−2)λ
L 2 ((L − 2)λ) e
Z ∞ Pr (A) + ...+
FP (p) = Pr(vd−α < p|v)Pr(v)dv 2 (n − 2)!
0   n−θ −(L−θ)λ
L θ ((L − θ)λ) e
(−1)θ
 
Z ∞  2
p −α 2 Pr (A)
− dmin  −v θ (n − θ)!
= 1 − v e dv.
0 d2max − d2min θ  
c ((L − c)λ)n−c e−(L−c)λ
L
(−1)c+1
X
= λe−λ ,
∞ p −α
2 c (n − c)!
− d2min −v
Z 
c=1
v
=1− e dv.
0 dmax − d2min
2 where θ = min{L−1, n}. Then, it is straight-forward to arrive
−α2 (1−Pr(B̄|Sn ))Pr(Sn )
d2min at Lemma 1 by calculating and substituting
 
p 2 1−Pr(B̄ )
=1− Γ +1 + 2 .
d2max − d2min α dmax − d2min λ
L by L − Ls and λ by L .
A PPENDIX C
P ROOF OF L EMMA 3
From (19) and (16), the outage probability converges to a
A PPENDIX B
deterministic value given below as
P ROOF OF L EMMA 2
Consider a vector N = {Ni }1≤i≤L , where Ni ∈ N are ǫSJD (Ls , I, P̂) = 1−
independent and identically distributed Poisson distributed
( !)
1 g(v) − g(u − uv)
random variables with average λ. For that, we define the max v, R ≤ min log 1 + 2 ,
0≤v≤Ls 0<u≤1 uvLs g(1) − g(v) + I+σ
Ls
following events
Ai : Event of Ni = 1. where u := ℓi and v := Lℓs . As we are considering a massive
λ
access setting, the average number of singleton layers λe− L
B: Event of Ni 6= 1, ∀i.
is taken to be asymptotically large. Moreover, as the ordered
L
X received powers converge to deterministic values under mas-
Sn : Event of Ni = n.
sive access, we can see that expression above for the outage
i=1
probability for a given number of singleton layers depends
Following from the definition of Poisson distributions, we have only on the interference power. Based on this, in step (a), we
Pr(Ai ) = Pr(A) = λe−λ , ∀i, average over Ls and I. Then, in step (b), we decompose the
L ∗
integral such that ISJD ( Lℓs ) denotes the maximum interference
Pr(B) = (1 − Pr(A))L = 1 − λe−λ , and power for which ℓ out of Ls layers can be decoded correctly.
n −Lλ
(Lλ) e It is expressed in (23). Finally, we arrive at (22) by applying
Pr (Sn ) = .
n! the inversion theorem in (14).
11

"Z ( !) #

(a) 1 g(v) − g(v − uv)
ǫSJD = 1 − ELs max v, R ≤ min log 1 + 2 fI (x)dI
0 v 0≤u≤1 uvL s g(1) − g(v) + x+σ
Ls
Z 1 
(b) ∗
= 1 − ELs vfI (ISJD (v))dv .
0

" !#
1 g(v) − g(v − uv) (c)
EI max min log 1 + I+σ2
=
0≤v≤1 0<u≤1 u g(1) − g(v) + Ls
" −z(g(v)−g(v−uv))
! #
1 ∞ e−z
Z 2
g(1)−g(v)+ I+σ (d)
EI max min 1−e L s dz =
0≤v≤1 0<u≤1 u 0 z
" Z ∞ −s((g(1)−g(v))Ls +σ2 ) #
1 e −sI

−s(g(v)−g(v−uv))Ls
 (e)
EI max min e 1−e ds =
0≤v≤1 0<u≤1 u 0 s
2
1 ∞ e−s((g(1)−g(v))Ls +σ )  −sI  
Z  (f )
max min EI e 1 − e−s(g(v)−g(v−uv))Ls ds =
0≤v≤1 0<u≤1 u 0 s
Z ∞ −s((g(1)−g(v))Ls +σ2 )
1 e  
max min ψI (−s) 1 − e−s(g(v)−g(v−uv))Ls ds.
0≤v≤1 0<u≤1 u 0 s

A PPENDIX D [5] E. Paolini, C. Stefanovic, G. Liva, and P. Popovski, “Coded random


P ROOF OF L EMMA 4 access: applying codes on graphs to design random access protocols,”
IEEE Communications Magazine, vol. 53, no. 6, pp. 144–150, June
The average system throughput can be evaluated by aver- 2015.
[6] E. MolavianJazi and J. N. Laneman, “On the second-order cost of
aging the expression in (25) over Ls and I. Step (c) follows TDMA for gaussian multiple access,” in 2014 IEEE International
from averaging over the latter and using the following identity Symposium on Information Theory. IEEE, 2014, pp. 266–270.
[ref] [41] [7] ——, “A random coding approach to gaussian multiple access channels
with finite blocklength,” in IEEE 50th Annual Allerton Conference on
Z ∞ −z Communication, Control, and Computing (Allerton), 2012, pp. 286–293.
e
1 − e−xz .

log(1 + x) = [8] L. Dai, B. Wang, Y. Yuan, S. Han, C. l. I, and Z. Wang, “Non-orthogonal
0 z multiple access for 5G: solutions, challenges, opportunities, and future
research trends,” IEEE Communications Magazine, vol. 53, no. 9, pp.
Step (d) follows from a change of variable: z = s I + σ 2 ,

74–81, September 2015.
and step (e) follows from interchanging the integration and [9] Y. Yuan, Z. Yuan, G. Yu, C. h. Hwang, P. k. Liao, A. Li, and
K. Takeda, “Non-orthogonal transmission technology in LTE evolution,”
the expectation. Finally, step (f ) follows from the definition IEEE Communications Magazine, vol. 54, no. 7, pp. 68–74, July 2016.
of CFs. [10] T. Takeda and K. Higuchi, “Enhanced user fairness using non-orthogonal
access with SIC in cellular uplink,” in IEEE Vehicular Technology
Conference (VTC Fall), 2011, pp. 1–5.
ACKNOWLEDGEMENT [11] Y. Endo, Y. Kishiyama, and K. Higuchi, “Uplink non-orthogonal access
with MMSE-SIC in the presence of inter-cell interference,” in IEEE
The authors would like to thank Dr. Wei Bao and Dr. International Symposium on Wireless Communication Systems (ISWCS),
2012, pp. 261–265.
Matthias Wildemeersch on their valuable feedback on this [12] M. Al-Imari, P. Xiao, M. A. Imran, and R. Tafazolli, “Uplink non-
work. orthogonal multiple access for 5G wireless networks,” in IEEE 11th In-
ternational Symposium on Wireless Communications Systems (ISWCS),
2014, pp. 781–785.
R EFERENCES [13] H. Haci, H. Zhu, and J. Wang, “A novel interference cancellation
technique for non-orthogonal multiple access (NOMA),” in IEEE Global
[1] C. Bockelmann, N. Pratas, H. Nikopour, K. Au, T. Svensson, C. Ste- Communications Conference (GLOBECOM), Dec 2015, pp. 1–6.
fanovic, P. Popovski, and A. Dekorsy, “Massive machine-type communi- [14] C. W. Sung and Y. Fu, “A game-theoretic analysis of uplink power
cations in 5G: physical and mac-layer solutions,” IEEE Communications control for a non-orthogonal multiple access system with two interfering
Magazine, vol. 54, no. 9, pp. 59–65, September 2016. cells,” in IEEE 83rd Vehicular Technology Conference (VTC Spring),
[2] G. Durisi, T. Koch, and P. Popovski, “Toward massive, ultrareliable, and 2016, pp. 1–5.
low-latency wireless communication with short packets,” Proceedings of [15] N. Zhang, J. Wang, G. Kang, and Y. Liu, “Uplink nonorthogonal multiple
the IEEE, vol. 104, no. 9, pp. 1711–1726, 2016. access in 5G systems,” IEEE Communications Letters, vol. 20, no. 3,
[3] G. Liva, “Graph-based analysis and optimization of contention resolution pp. 458–461, 2016.
diversity slotted ALOHA,” IEEE Transactions on Communications, [16] P. Sedtheetorn and K. Panyim, “Accurate uplink spectral efficiency for
vol. 59, no. 2, pp. 477–487, 2011. non orthogonal multiple access in nakagami fading,” in IEEE 13th Inter-
[4] C. Stefanovic, P. Popovski, and D. Vukobratovic, “Frameless ALOHA national Conference on Electrical Engineering/Electronics, Computer,
protocol for wireless networks,” IEEE Communications Letters, vol. 16, Telecommunications and Information Technology (ECTI-CON), 2016,
no. 12, pp. 2087–2090, 2012. pp. 1–4.
12

[17] H. Tabassum, E. Hossain, and M. J. Hossain, “Modeling and analysis [41] Y. Lin and W. Yu, “Downlink spectral efficiency of distributed antenna
of uplink non-orthogonal multiple access (NOMA) in large-scale cel- systems under a stochastic model,” IEEE Transactions on Wireless
lular networks using poisson cluster processes,” IEEE Transactions on Communications, vol. 13, no. 12, pp. 6891–6902, 2014.
Communications, vol. PP, no. 99, pp. 1–1, April 2017.
[18] T. Yang, L. Yang, Y. Guo, and J. Yuan, “A non-orthogonal multiple-
access scheme using reliable physical-layer network coding and cascade-
computation decoding,” IEEE Transactions on Wireless Communica-
tions, vol. PP, no. 99, pp. 1–1, 2017.
[19] C. Wei, H. Liu, Z. Zhang, J. Dang, and L. Wu, “Approximate message
passing based joint user activity and data detection for NOMA,” IEEE
Communications Letters, vol. PP, no. 99, pp. 1–1, 2016.
[20] B. Wang, L. Dai, Y. Zhang, T. Mir, and J. Li, “Dynamic compressive
sensing-based multi-user detection for uplink grant-free NOMA,” IEEE
Communications Letters, vol. 20, no. 11, pp. 2320–2323, Nov 2016.
[21] H. Nikopour and H. Baligh, “Sparse code multiple access,” in IEEE
24th Annual International Symposium on Personal, Indoor, and Mobile
Radio Communications (PIMRC). IEEE, 2013, pp. 332–336.
[22] S. Zhang, X. Xu, L. Lu, Y. Wu, G. He, and Y. Chen, “Sparse code
multiple access: An energy efficient uplink approach for 5G wireless
systems,” in IEEE Global Communications Conference. IEEE, 2014,
pp. 4782–4787.
[23] M. Z. Win, P. C. Pinto, and L. A. Shepp, “A mathematical theory of
network interference and its applications,” Proceedings of the IEEE,
vol. 97, no. 2, pp. 205–230, Feb 2009.
[24] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach
to coverage and rate in cellular networks,” IEEE Transactions on
Communications, vol. 59, no. 11, pp. 3122–3134, 2011.
[25] A. Rabbachin, T. Q. Quek, H. Shin, and M. Z. Win, “Cognitive network
interference,” IEEE Journal on Selected Areas in Communications,
vol. 29, no. 2, pp. 480–493, 2011.
[26] A. Ghosh, N. Mangalvedhe, R. Ratasuk, B. Mondal, M. Cudak, E. Visot-
sky, T. A. Thomas, J. G. Andrews, P. Xia, H. S. Jo et al., “Heterogeneous
cellular networks: From theory to practice,” IEEE Communications
Magazine, vol. 50, no. 6, 2012.
[27] T. D. Novlan, H. S. Dhillon, and J. G. Andrews, “Analytical modeling
of uplink cellular networks,” IEEE Transactions on Wireless Communi-
cations, vol. 12, no. 6, pp. 2669–2679, 2013.
[28] T. Manabe and H. Takai, “Superresolution of multipath delay profiles
measured by PN correlation method,” IEEE Transactions on Antennas
and Propagation, vol. 40, no. 5, pp. 500–509, 1992.
[29] C.-P. Li and W.-C. Huang, “A constructive representation for the fourier
dual of the zadoff–chu sequences,” IEEE Transactions on Information
Theory, vol. 53, no. 11, pp. 4221–4224, 2007.
[30] J.-C. Belfiore, G. Rekaya, and E. Viterbo, “The golden code: a 2/spl
times/2 full-rate space-time code with nonvanishing determinants,” IEEE
Transactions on Information Theory, vol. 51, no. 4, pp. 1432–1436,
2005.
[31] G. T. Buračas and G. M. Boynton, “Efficient design of event-related
fMRI experiments using m-sequences,” Neuroimage, vol. 16, no. 3, pp.
801–813, 2002.
[32] D. T. Wiriaatmadja and K. W. Choi, “Hybrid random access and
data transmission protocol for machine-to-machine communications in
cellular networks,” IEEE Transactions on Wireless Communications,
vol. 14, no. 1, pp. 33–46, 2015.
[33] M. Wildemeersch, T. Q. Quek, C. H. Slump, and A. Rabbachin,
“Cognitive small cell networks: Energy efficiency and trade-offs,” IEEE
Transactions on Communications, vol. 61, no. 9, pp. 4016–4029, 2013.
[34] D. Tuninetti and G. Caire, “The throughput of some wireless multiaccess
systems,” IEEE Transactions on Information Theory, vol. 48, no. 10, pp.
2773–2785, 2002.
[35] H. A. David, Ordered Statistics, 2nd ed. New York: Wiley, 1981.
[36] M. Shirvanimoghaddam, M. Condoluci, M. Dohler, and S. Johnson, “On
the fundamental limits of random non-orthogonal multiple access in
cellular massive IoT,” arXiv preprint arXiv:1705.10471, 2017.
[37] A. Adhikary, X. Lin, and Y.-P. E. Wang, “Performance evaluation of NB-
IoT coverage,” in IEEE 84th Vehicular Technology Conference (VTC-
Fall), 2016, pp. 1–5.
[38] Y.-P. E. Wang, X. Lin, A. Adhikary, A. Grovlen, Y. Sui, Y. Blankenship,
J. Bergman, and H. S. Razaghi, “A primer on 3GPP narrowband internet
of things,” IEEE Communications Magazine, vol. 55, no. 3, pp. 117–123,
2017.
[39] S.-M. Oh and J. Shin, “An efficient small data transmission scheme
in the 3GPP NB-IoT system,” IEEE Communications Letters, vol. 21,
no. 3, pp. 660–663, 2017.
[40] S. Verdu, “The capacity region of the symbol-asynchronous gaussian
multiple-access channel,” IEEE Transactions on Information Theory,
vol. 35, no. 4, pp. 733–751, Jul 1989.