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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
1

Massive MIMO Forward Link Analysis for

Cellular Networks
Geordie George, Member, IEEE, Angel Lozano, Fellow, IEEE, and Martin Haenggi, Fellow, IEEE

Abstract—This paper presents analytical expressions for the •Optimizing the number of active users as a function of
signal-to-interference ratio (SIR) and the spectral efficiency in the number of antennas and the path loss exponent.
macrocellular networks with massive MIMO conjugate beam- • Assessing the benefits of a channel-dependent power
forming, both with a uniform and a channel-dependent power
allocation. These expressions, which apply to very general net- allocation.
work geometries, are asymptotic in the strength of the shadowing. At the same time, the analysis is not without limitations,
Through Monte-Carlo simulation, we verify their accuracy for chiefly that pilot contamination is not accounted for. We
relevant network topologies and shadowing strengths. Also, since explore this aspect by contrasting our analysis with simulation-
the analysis does not include pilot contamination, we further
gauge through Monte-Carlo simulation the deviation that this
based results that include the contamination, thereby delineat-
phenomenon causes with respect to our results, and hence the ing the scope of our results.
scope of the analysis. The paper is organized as follows. The network and channel
Index Terms—Massive MIMO, stochastic geometry, shadow- models are introduced in Section II and the forward-link
ing, Poisson point process, lattice networks, PPP networks conjugate beamforming SIRs, for uniform and equalizing
power allocations, are derived in Section III. Building on some
I. I NTRODUCTION preliminary analysis presented in Section IV, Sections V–
VI subsequently characterize the SIR distributions for fixed
T HE tedious and time-consuming nature of system-level
performance evaluations in wireless networks is exacer-
bated with massive MIMO [1]–[13], as the number of active
and Poisson-distributed numbers of users. The SIRs then lead
to spectral efficiencies in Section VII. The applicability of
users per cell becomes hefty and the dimensionality of the the results to relevant network geometries is illustrated in
channels grows very large. This reinforces the interest in Section VIII and the impact of noise and pilot contamination
analytical solutions, and such is the subject of this paper. is assessed in Section IX. Finally, Section X concludes the
To embark upon the analysis of massive MIMO settings, paper.
we invoke tools from stochastic geometry that have been
II. N ETWORK M ODELING
successfully applied already in non-MIMO [14]–[20] and in
MIMO contexts [21]–[24]. We consider a macrocellular network where each base
We present expressions for the forward-link signal-to- station (BS) is equipped with Na ≫ 1 antennas while users
interference ratio (SIR) and the spectral efficiency in macrocel- feature a single antenna.
lular networks with massive MIMO conjugate beamforming,
both with a uniform and a channel-dependent power allocation, A. Large-scale Modeling
and for different alternatives concerning the number of users The BS positions form a stationary and ergodic point
per cell. The derived expressions apply to very general network process Φb ⊂ R2 of density λb , or a realization thereof,
geometries in the face of shadowing. These expressions allow: say a lattice network. As a result, the density of BSs within
• Testing and calibrating system-level simulators. any region converges to λb > 0 as this region’s area grows
• Determining how many cells need to be simulated for [25]. In turn, the user positions conform to an independent
some desired accuracy in terms of interference: the anal- point process Φu ⊂ R2 of density λu , also stationary and
ysis subsumes an infinite field of cells whereas simulators ergodic. Altogether, the models encompass virtually every
necessarily feature a finite number thereof, which, if macrocellular scenario of relevance.
too small, results in a deficit of interference and in Each user is served by the BS from which it has the
consequently optimistic performance predictions. strongest large-scale channel gain, and we denote by Kℓ the
• Gauging the impact of parameters such as the path loss number of users served by the ℓth BS.
exponent. The large-scale channel gain includes path loss with expo-
nent η > 2 and shadowing that is independent and identically
This work was supported by Project TEC2015-66228-P (MINECO/FEDER,
UE), by the European Research Council under the H2020 Framework Pro-
distributed (IID) across the links. Specifically, the large-scale
gramme/ERC grant agreement 694974, and by the U.S. NSF through award gain between the ℓth BS and the kth user served by the lth
CCF 1525904. BS is
G. George and A. Lozano are with the Department of Information and Com-
munication Technologies, Universitat Pompeu Fabra (UPF), 08018 Barcelona, Lref
Gℓ,(l,k) = η χℓ,(l,k) ℓ, l ∈ N0 , k ∈ {0, . . . , Kl − 1},
Spain. E-mail: {geordie.george, angel.lozano}@upf.edu. rℓ,(l,k)
Martin Haenggi is with the University of Notre Dame, Notre Dame, IN
46556, USA. E-mail: mhaenggi@nd.edu. (1)

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
2

with Lref the path loss intercept at a unit distance, rℓ,(l,k) the BS. This is because, as the shadowing strengthens, it comes to
link
 distance,
 and χℓ,(l,k) the shadowing coefficient satisfying dominate over the path loss and, in the limit, all BSs become
E χδ < ∞, where we have introduced δ = 2/η. When equally likely to be the serving one. Consider a region of area
expressed in dB, the shadowing coefficients have a standard A having Aλb BSs and Aλu users, all placed arbitrarily. As
deviation of σdB . In the absence of shadowing, χ = 1 and σdB → ∞, the number of users associated served
by each BS
1
σdB = 0. becomes binomially distributed, B Aλu , Aλ b
, because each
Without loss of generality, we declare the 0th BS as the user has equal probability A1λb of being served by any of the
focus of our interest and, for notational compactness, drop A λb BSs. Now, letting A → ∞ while keeping λu /λb constant,
its index from the scripting. For the large-channel gains, for the binomial distribution converges to the Poisson distribution
instance, this means that: [16].
• G0,(l,k) = G(l,k) relates the BS of interest with the kth In accordance with the foregoing reasoning, which is for-
user served by the lth BS. malized in [30], we model {Kℓ }ℓ∈N0 as IID Poisson random
• Gℓ,(0,k) = Gℓ,(k) relates the ℓth BS with the kth user variables with mean E[Kℓ ] = K̄ = λu /λb . Noting that
served by the BS of interest. the unbounded tail of the Poisson distribution needs to be
• G0,(0,k) = G(k) relates the BS of interest with its own truncated at Na , because such is the maximum number of
kth user. users that can be linearly served by a BS with Na antennas,
• K0 = K is the number of users served by the BS of our analysis is conducted with Poisson {Kℓ }ℓ∈N0 and subse-
interest. quently we verify the accuracy against simulations where the
The same scripting and compacting is applied to other quan- truncation is effected (see Examples 10 and 12). For the BS
tities. of interest specifically, we apply the Poisson PMF (probability
Let us consider an arbitrary user k served by the BS mass function)
of interest. It is shown in [25]–[27] that, as σdB → ∞,
irrespective of the actual BS positions, the propagation process K̄ k e−K̄
fK (k) = k ∈ N0 , (3)
from that user’s vantage, specified by {1/Gℓ,(k) }ℓ∈N0 and k!
seen as a point process on R+ , converges to what the typical whose corresponding CDF (cumulative distribution function)
user—e.g., at the origin—would observe if the BS locations is
conformed to a homogeneous PPP on R2 with density λb . Γ(k + 1, K̄)
Moreover, by virtue of independent shadowing in each link, FK (k) = k ∈ N0 , (4)
k!
as σdB → ∞ the propagation process from the vantage
where Γ(·, ·) is the upper incomplete gamma function.
of each user becomes independent from those of the other
users. Relying on these results, we embark on our analysis by
regarding the propagation processes C. Small-scale Modeling

1/Gℓ,(k) ℓ∈N k = 0, . . . , K − 1 (2) Let us focus on the local neighborhood of the kth user
0

as IID and Poisson, anticipating that the results be applicable served by the BS of interest, wherein the large-scale gains
under relevant network geometries and realistic values of σdB ; {Gℓ,(k) }ℓ∈N0 apply. Without loss of generality, a system-
this is validated in Section VIII. level analysis can be conducted from the perspective of this
user, which becomes the typical user in the network once we
B. Number of Users per BS uncondition from {Gℓ,(k) }ℓ∈N0 .
The modeling of {Kℓ } is a nontrivial issue. Even in a Upon data transmission from the BSs, such kth user ob-
lattice network with equal-size cells, and let alone in irregular serves
networks, disparities may arise across BSs because of the Xq
yk = Gℓ,(k) h∗ℓ,(k) xℓ + vk k = 0, . . . , K − 1, (5)
shadowing and the stochastic nature of the user locations.
ℓ
In the absence of shadowing, the users served by a BS are
those within its Voronoi cell, and their number—conditioned where hℓ,(k) ∼ NC (0, I) is the (normalized) reverse-link Na ×
on the cell area—is a random variable with mean λu times the 1 channel vector, h∗ℓ,(k) is its forward-link reciprocal, and vk ∼
cell area; for instance, it is a Poisson random variable if Φu NC (0, σ 2 ) is AWGN. The Na × 1 signal vector xℓ emitted  by
is a PPP [16]. Depending on how the cell area is distributed, the ℓth BS, intended for its Kℓ users, satisfies E kxℓ k2 = P
then, the distribution of the number of users per cell can be where P is the per-base transmit power.
computed. For instance, if Φu is PPP, such number in a lattice
network is Poisson-distributed with mean λu /λb while, for an
III. C ONJUGATE B EAMFORMING
irregular network with PPP-distributed BSs, the corresponding
distribution is computed approximately in [28], [29]. A. Transmit Signal
With shadowing, a user need not be served by the BS in
The signal transmitted by the ℓth BS is
whose Voronoi cell it is located. Remarkably though, with
strong and independent shadowing per link, a Poisson distribu- K ℓ −1
r
X Pℓ,k
tion with mean λu /λb turns out to be a rather precise model— xℓ = f sℓ,k , (6)
Na ℓ,k
regardless of the BS locations—for the number of users per k=0

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
3

where Pℓ,k is the power allocated to the data symbol sℓ,k ∼ with
NC (0, 1), which is precoded by f ℓ,k and intended for its kth   X h i
2
DEN = G(k) Pk var h∗(k) f k + G(k) Pk E h∗(k) f k
user. The power allocation satisfies
k6=k
K ℓ −1
X X KX
ℓ −1

2

Pℓ,k = P (7) + Gℓ,(k) Pℓ,k E h∗ℓ,(k) f ℓ,k + Na σ 2 .
k=0 ℓ6=0 k=0

and, with conjugate beamforming and an average power con- (12)

straint, As indicated, the analysis in this paper ignores pilot con-
tamination. It follows that, in interference-limited conditions
p ĥℓ,(ℓ,k)
f ℓ,k = Na r h k = 0, . . . , Kℓ − 1, (8) (σ 2 /P → 0), the conjugate beamforming precoders at BS ℓ
i
E kĥℓ,(ℓ,k) k2 are
f ℓ,k = hℓ,(ℓ,k) k = 0, . . . , Kℓ − 1, (13)
where ĥℓ,(ℓ,0) , . . . , ĥℓ,(ℓ,Kℓ −1) are the channel estimates gath-
from which
ered by the ℓth BS from the reverse-link pilots transmitted by 2
    2
its own users. E h∗(k) f k = E kh(k) k2 (14)
Bringing (5) and (6) together, the kth user served by the BS
of interest observes = Na2 (15)
s and
K−1
X G(k) Pk ∗
yk = h(k) f k sk    
Na var h∗(k) f k = E kh(k) k4 − Na2 (16)
k=0
X KX ℓ −1
s = Na (Na + 1) − Na2 (17)
Gℓ,(k) Pℓ,k ∗
+ hℓ,(k) f ℓ,k sℓ,k + vk , (9) = Na (18)
Na
ℓ6=0 k=0
while, because of independence between user k and other users
whose first and second terms contain, respectively, the same- k 6= k, for both ℓ = 0 and ℓ > 0 it holds that
BS and other-BS transmissions. 
2

∗
E hℓ,(k) f ℓ,k = Na . (19)
B. SIR Altogether, in interference-limited conditions,
We consider receivers reliant on channel hardening  [1], Na Pk G(k)
whereby user k served by the BS of interest regards E h∗(k) f k SIRk = P PKℓ −1 , (20)
as its precoded channel, with the expectation taken over the ℓ k=0 Pℓ,k Gℓ,(k)
small-scale fading. The fluctuations of the actual precoded which, invoking (7), further reduces to
channel around this expectation constitute self-interference,
Na Pk G(k)
such that (9) can be elaborated into SIRk = P P . (21)
s ℓ Gℓ,(k)
G(k) Pk  ∗  The foregoing ratio of channel gains can be seen to equal
yk = E h(k) f k sk
Na G(k) G
| {z } P = P(k) (22)
s
Desired signal G
ℓ ℓ,(k) G(k) + ℓ6=0 Gℓ,(k)
G(k) Pk  ∗   1
+ h(k) f k − E h∗(k) f k sk = P (23)
ℓ6=0 Gℓ,(k)
Na 1+
| {z } G(k)
Self-interference 1
s = (24)
X G(k) Pk ∗ 1 + 1/ρk
+ h(k) f k sk
Na with
k6=k
| {z } G(k)
Same-BS interference ρk = P (25)
ℓ6=0 Gℓ,(k)
s
X KX ℓ −1
Gℓ,(k) Pℓ,k ∗
+ hℓ,(k) f ℓ,k sℓ,k + vk (10) being the local-average SIR in single-user transmission [21].
Na |{z}
ℓ6=0 k=0 Hence, (21) can be rewritten as
| {z } Noise
Other-BS interference Pk /P
SIRk = Na . (26)
and the SINR is 1 + 1/ρk
  2 Two different power allocations, meaning two different
G(k) Pk E h∗(k) f k
formulations for Pk /P , are analyzed in this paper; the cor-
SINRk = (11)
DEN responding SIRs are presented next.

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
4

TABLE I: Parameter s⋆ for common values of the path loss exponent η (the
corresponding δ = 2/η is also listed). where γ(·, ·) is the lower incomplete gamma function. The
CDF of ρ satisfies
η δ s⋆ η δ s⋆
 s⋆ /θ 1
 Fρ (θ) ≃ e  0 ≤ θ < 2+ǫ
3.5 0.571 -0.672 3.9 0.513 -0.819 
3.6 0.556 -0.71 4 0.5 -0.854 
θ 1
3.7 0.540 -0.747 4.1 0.488 -0.888 Fρ (θ) = 1 − θ−δ sinc δ + Bδ 1−θ 2 ≤θ <1
(34)
3.8 0.526 -0.783 4.2 0.476 -0.922

 −δ
Fρ (θ) = 1 − θ sinc δ θ ≥ 1,
where “≃” indicates asymptotic (θ → 0) equality while
1) Uniform Power Allocation: With a uniform power allo-


cation, Pk = P/K and 2 F1 1, δ + 1; 2 δ + 2; −1/x δ
Bδ (x) = 1+2 δ (35)
Na /K x Γ(2 δ + 2) Γ2 (1 − δ)
SIRUnif
k = . (27)
1 + 1/ρk with 2 F1 the Gauss hypergeometric function. Setting
2) Equal-SIR Power Allocation: Alternatively, the SIRs can log 1 − 2δ sinc δ + Bδ (1)


be equalized by setting [31] ǫ= − 2, (36)
P s⋆
Gℓ,(k)
ℓ


1
G(k) we ensure Fρ 2+ǫ = Fρ 21 and the CDF can be taken as
Pk = P PK−1 Pℓ Gℓ,(k) (28)
constant therewithin.
k=0 G(k)
P (1 + 1/ρk ) As an alternative to the foregoing characterization, an exact
= PK−1 . (29) but integral form can be obtained for Fρ (·).
K + k=0 1/ρk
Plugged into (26), this power allocation yields an SIR, com- Lemma 2.
Z ∞  
mon to all the users served by the BS of interest, of 1 1 ei(θ+1)ω dω
Fρ (θ) = − ℑ (37)
Eq Na /K 2 π 0 1 F1 (1, 1 − δ, iθω) ω
SIR = 1 PK−1
. (30)
1+ K k=0 1/ρk where ℑ{·} denotes imaginary part and 1 F1 is the confluent
hypergeometric or Kummer function.
Introducing the harmonic mean of ρ0 , . . . , ρK−1 , namely
1 Proof. See Appendix A. 
ρK = 1
PK−1 , (31)
K k=0 1/ρk While the results in the sequel are derived using either
we can rewrite (30) as Lemma 1 or Lemma 2, there is yet another useful alterna-
tive, namely computing Fρ (·) via an approximate numerical
Na /K
SIREq = . (32) inversion of the Laplace transform L1/ρ (s) = E[e−s/ρ ],
1 + 1/ρK
presentation of which is relegated to Appendix B.
Having formulated the SIRs for given large-scale link gains, Equipped with the foregoing tools, let us proceed to char-
let us next characterize the system-level performance by acterize the spatial distribution of the SIRs in (27) and (30).
releasing the conditioning on those gains.
V. S PATIAL SIR D ISTRIBUTIONS WITH F IXED K
IV. S PATIAL D ISTRIBUTION OF ρk
To begin with, let us characterize the SIR distributions for
As the SIR formulations in (27) and (30)Pindicate, a key a fixed number of served users. Besides having their own
ingredient is the distribution of ρk = G(k) / ℓ6=0 Gℓ,(k) . To interest, the ensuing results serve as a stepping stone to their
characterize this distribution, we capitalize on results derived counterparts for K conforming to the Poisson or to any other
for the typical user in a PPP-distributed network of BSs [20], desired distribution.
[21], in accordance with the PPP convergence exposed in
Section II-A. Specifically, ρ0 , . . . , ρK−1 are regarded as IID
with CDF Fρ (·), where ρ is the local-average SIR of the typical A. Uniform Power Allocation
user in such a network [20], [21].1 In the sequel, we employ Starting from (27), we can determine the CDF of SIRUnif
k as
a slightly simplified version of the result in [21, Eq. 18] as  
presented next. θ
FSIRUnif (θ) = P ρk < (38)
k Na /K − θ
Lemma 1. [21] Define s⋆ < 0 as the solution—common (  
θ
values are listed in Table I—to Fρ Na /K−θ 0 < θ < Na /K
= (39)
1 θ ≥ Na /K
s⋆ δ γ(−δ, s⋆ ) = 0, (33)
which depends on Na and K only through their ratio, Na /K.
1 Our analysis focuses on the typical cell, rather than the typical user Now, invoking Lemmas 1–2, we can express the SIR
(or, more precisely, the typical location). However, thanks to the Poisson
convergence for strong shadowing, the propagation point process as seen from distribution for the typical user in a massive MIMO network
a user in the typical cell is statistically the same as that seen from the typical with fixed K and a uniform power allocation.
location in a shadowless PPP.

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
5

 ⋆ Na

 FSIRUnif

 k
(θ) ≃ es ( θ K −1) 0≤θ< Na /K
3+ǫ

δ  
Na /K
FSIRUnif
k
(θ) = 1 − Na
θK − 1 sinc δ + B δ
θK
Na −2 θ K 3 ≤ θ < Na2/K (40)


δ
Na Na /K
θ < NKa

FSIRUnif
k
(θ) = 1 − θK − 1 sinc δ 2 ≤

Unif 1
Proposition 1. The CDF of SIR k with a fixed
N a /K Na /K
 K satisfies
(40) with constant value within 3+ǫ , 3 . Alternatively,
the CDF can be computed exactly as
  0.8 = 2000
Z ∞  iω
 dω = 200
1 1 e 1−θ K/N a
FSIRUnif (θ) = − ℑ  
k 2 π 0  F 1, 1 − δ, iθω
1 1
 ω = 100
Na /K−θ
0.6 = 10
0 < θ < Na /K. (41)

CDF
With either of these expressions, one can readily compute
0.4
the percentage of users achieving a certain local-average
performance for given Na /K and η. Furthermore, one can
establish minima for Na /K given η and given some target
performance at a desired user percentile. 0.2

Example 1. Let η = 4. In order to ensure that no more than

3% of users experience an SIR below 0 dB, it is required that 0
Na /K ≥ 5. 0 2 4 6 8 10
SIR (dB)
B. Equal-SIR Power Allocation Fig. 1: CDFs of SIRUnif
k and SIR
Eq
for Na /K = 10 and η = 4. For SIRUnif
k ,
the dotted and solid curves correspond, respectively, to (40) and (41) in
From (32), the CDF of SIREq can be expressed as Prop. 1. The dashed curves for SIREq are obtained via Prop. 2, with the
  block arrows indicating the convergence for Na , K → ∞ as per (45).
θ
FSIREq (θ) = P ρK < (42)
N /K − θ
(  a 
θ
FρK Na /K−θ 0 < θ < NKa Example 2. FSIREq (·) and FSIRUnif
k
(·) are compared in Fig. 1,
= (43) for Na /K = 10 and η = 4. For SIRUnif
1 θ ≥ NKa k , the dotted and solid
curves correspond, respectively, to (40) and (41) in Prop. 1.
where, recall, ρK is the harmonic mean of ρ0 , . . . , ρK−1 . The dashed curves for SIREq , obtained via Prop. 2, can be
While an explicit expression such as (40) seems difficult observed to converge, as per (45), to
to obtain for FSIREq (·), an integral form similar to (41) is
forthcoming. Na
(1 − δ) = 10 (1 − 1/2) (48)
Eq
K
Proposition 2. The CDF of SIR with a fixed K is = 7 dB (49)
Z ( )
1 1 ∞ ei ω dω
FSIREq (θ) = − ℑ as Na , K → ∞.
2 π 0 K ω
1 F1 (1, 1 − δ, i θ ω/Na )
Note from the foregoing example that, for Na = 100 and
0 < θ < Na /K. (44)
K = 10 onwards, over 90% of users have an SIREq within 1 dB
Proof. See Appendix C.  of its spatial average. Besides confirming the effectiveness
of the simple SIR-equalizing power allocation for massive
What can be established explicitly is that, as Na , K → ∞ MIMO, this allows establishing Na /K rather accurately by
with ratio Na /K, directly equating NKa (1 − δ) to the desired SIR.
Na
SIREq → (1 − δ) (45)
K
VI. S PATIAL SIR D ISTRIBUTION WITH
which follows from P OISSON - DISTRIBUTED K
K−1
1 X 1  
lim = E 1/ρ (46) Let us now allow K to adopt a Poisson distribution, which,
K→∞ K ρk
k=0 as argued, captures well the variabilities caused by network
δ irregularities and shadowing. Similar derivations could be
= (47)
1−δ applied to other distributions if, for instance, one wishes to
where (47) is established in [32]. further incorporate activity factors for the users or channel
assignment mechanisms.

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
6

A. Uniform Power Allocation 1

= 100
The expectation of (40) in Prop. 1 over (3) yields the = 10
following result. 0.8 = 2000
Unif
= 200
Proposition 3. The CDF of SIRk with a Poisson-distributed
K satisfies 0.6

CDF
⌈ N3+ǫ
a /θ
⌉−1
k
1 X ⋆ Na K̄
FSIRUnif (θ) ≈ K̄ es ( θ k −1) 0.4
k
e −1 k!
k=1
⌈N a
3θ ⌉−1
1 X ⋆ K̄ k
+ K̄ es (ǫ+2) 0.2
e −1 k!
k=⌈ N3+ǫ
a /θ
⌉

⌈N a
2θ ⌉−1
" δ 
1 X Na 0
+ K̄ − 1 sinc δ
1− 0 3 6 9 12 15
e −1 θk SIR (dB)
k=⌈ N a
3θ ⌉
 # k Fig. 2: CDFs of SIRUnif Eq Unif
θk K̄ k and SIR with Na /K̄ = 10 and η = 4. For SIRk ,
+ Bδ the dotted and solid lines correspond to Prop. 3, respectively (50) and (51).
Na − 2 θ k k! The dashed curves for SIREq correspond to Prop. 4. The block arrows indicate
" # the evolution for Na , K̄ → ∞.
⌈Na /θ⌉−1  δ
1 X Na K̄ k
+ K̄ 1− − 1 sinc δ
e −1 θk k!
k=⌈ N a
2θ ⌉
Proposition 4. Since the user’s SIR is a valid quantity only

for K ≥ 1, i.e., when the BS has at least one user, FSIREq (·)
1 − Γ ⌈ Nθa ⌉, K̄
+
Na
(50) with Poisson-distributed K equals
1 − e−K̄ ⌈ θ ⌉ − 1 !
∞
X fK (k)
FSIREq (θ) = FSIREq |K=k (θ) (52)
where Bδ (·) is as per (35). Exactly, 1 − FK (0)
k=1
1 1
= −
⌈Na /θ⌉−1
X K̄ k 2 π(e − 1)
K̄
1 1 Z ∞ ( " #)
FSIRUnif (θ) = − K̄
dω
k 2 π (eK̄ − 1) k! · ℑ e iω
e 1 F1 (1,1−δ, Na )
iθω
−1 , (53)
k=1

iω
 0 ω
Z ∞   dω
e 1−θ k/Na
· ℑ   where we recalled the expression for FSIREq |K=k (·) from (44)
0  F 1, 1 − δ, iθω  ω
1 1 Na /k−θ and invoked (3)–(4).


1 − Γ ⌈ Nθa ⌉, K̄ Again, because of the convergence Na /K → Na /K̄ for
+

. (51)
2 1 − e−K̄ ⌈ Nθa ⌉ − 1 ! Na , K̄ → ∞, SIREq hardens to its fixed-K value with K = K̄.
In this case, this corresponds to Na (1−δ)/K̄ as demonstrated
in Fig. 2.
Proof. See Appendix D. 
Example 3. FSIREq (·) and FSIRUnif (·) with a Poisson-distributed
K are compared in Fig. 2, for Na /K̄ = 10 and η = 4.
For Na , K̄ → ∞ with fixed Na /K̄, it can be verified that For SIRUnif
k , the dotted and solid lines correspond to Prop.
Na /K → Na /K̄ with convergence in the mean-square sense, 3, respectively (50) and (51). The dashed curves for SIREq
and therefore in probability. As a consequence, SIRUnif , which correspond to Prop. 4.
depends on Na and K only through their ratio, progressively
behaves as if this ratio were fixed at Na /K̄ despite the Poisson VII. S PECTRAL E FFICIENCY
nature of K. This behavior is clearly in display in Fig. 2, where With each user’s SIR stable over its local neighborhood
FSIRUnif
k
(·) can be seen to approach its value for fixed K = K̄. thanks to the channel hardening, the spectral efficiency of the
typical user is directly
Ck = log2 (1 + SIRk ) (54)
B. Equal-SIR Power Allocation
and the spatial distribution thereof can be readily obtained as
 
In this case, we expect the expression in Prop. 2 over (3) FCk (ζ) = P log2 (1 + SIRk ) < ζ (55)
to obtain the following. = FSIRk (2ζ − 1) (56)

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
7

for each setting considered in the foregoing sections. Expecting the spatially averaged user spectral efficiency in
(58) over K, via the Poisson distribution in (3), we obtain
Example 4. Let η = 4, Na = 100 and K = 10. The 3%-
∞
ile user spectral efficiency, computed by numerically solving X fK (k)
C̄ Unif = C̄ Unif |K=k (60)
for ζ in FSIRUnif
k
(2ζ − 1) = 0.03 and FSIREq (2ζ − 1) = 0.03, 1 − FK (0)
k=1
respectively under uniform and equal-SIR power allocations, P∞ Na
K̄ k
are 1.6 b/s/Hz and 2.2 b/s/Hz.
Z ∞ 1 − eK̄1−1 k=1 e−z k k! dz
= log2 (e)
. (61)
0 1 F1 1, 1 − δ, z z
Example 5. Reconsidering Example 4, but with K being Pois-
son with mean K̄ = 10, the 3%-ile user spectral efficiencies, The corresponding spatial average of the sum spectral effi-
respectively under uniform and equal-SIR power allocations, ciency, recalling (57), is
are 1.51 b/s/Hz and 1.94 b/s/Hz. ∞
X
C̄ΣUnif = k C̄ Unif |K=k fK (k) (62)
From its user spectral efficiencies, the sum spectral effi-
k=1
ciency at the BS of interest can be found as Z ∞
P∞

log e k 1 − e−zNa /k K̄ k /k! dz
K−1 = K̄2 k=1

. (63)
X e 0 1 F1 1, 1 − δ, z z
CΣ = Ck , (57)
k=0 Example 7. Let η = 4 and Na = 100 with a uniform power
which is zero whenever the BS serves no users. allocation. For K = 10, (59) returns an average user spectral
Explicit expressions for the spatial averages of the user and efficiency of C̄ Unif = 2.76 b/s/Hz and an average sum spectral
  efficiency of C̄ΣUnif = 10 C̄ Unif = 27.6 b/s/Hz per BS. For K
the sum spectral efficiencies, C̄ = E Ck and C̄Σ = E[CΣ ]
with expectation over all possible propagation processes, are being Poisson with mean K̄ = 10, (61) returns C̄ Unif = 2.84
presented next. b/s/Hz while (63) yields C̄ΣUnif = 27.08 b/s/Hz per BS.

B. Average Spectral Efficiency: Equal-SIR Power Allocation

A. Average Spectral Efficiency: Uniform Power Allocation
Proposition 6. With an equal-SIR power allocation and a
Proposition 5. With a uniform power allocation and a fixed
fixed K, the spatially averaged user spectral efficiency equals
K, the spatially averaged user spectral efficiency equals Z ∞
Z ∞ 1 − e−z dz
1 − e−zNa /K dz C̄ Eq = log2 (e) K z
(64)
C̄ Unif = log2 (e)
(58) 0 1 F1 (1, 1 − δ, z/Na )
0 1 F1 1, 1 − δ, z z
and the spatially averaged sum spectral efficiency is C̄ΣEq =
and the spatially averaged sum spectral efficiency is C̄ΣUnif = K C̄ Eq .
K C̄ Unif .
Proof. See Appendix E. 
Proof. See Appendix E. 
For Na , K → ∞, recalling how the SIR hardens to (1 −
For the special case of η = 4, i.e., for δ = 1/2, Prop. 5 δ) Na /K, we have that C̄ Eq hardens to log2 1 + (1 − δ) NKa .
reduces to Expecting (64) over K, via (3), gives
Z ∞ ∞
1 − e−zNa /K dz
Z ∞
log2 e 1 − e−z X fK (k)
C̄ Unif = log2 (e) √ √ (59) Eq
C̄ = k dz
0 1 + ez πz erf z z 1 − FK (0) 0 z

k=1 1 F1 1, 1 − δ, z

√ √ Na
thanks to 1 F1 1, 1/2, z ≡ 1 + ez πz erf z, an equivalence (65)
that is further applicable in the characterizations that follow. Z ∞
log2 e 1 − e−z  F (1,1−δ,z/N
K̄

Since the denominator of the integral’s argument in (59), = K̄ e1 1 a) − 1 dz. (66)
and more generally in (58), is strictly positive for z ≥ 0, e −1 0 z
we can deduce by inspection that C̄ Unif shrinks if we increase In turn, the average sum spectral efficiency with expectation
K with a fixed Na , i.e., if we add more users with a fixed over (3) becomes
number of antennas. However, this reduction is sublinear in K Z ∞ ∞
Eq
1 − e−z X k fK (k)
and thus C̄ΣUnif grows as we add more users. Thus, K should C̄Σ = log2 (e)  k dz (67)
z
be set to the largest possible value that ensures an acceptable 0 k=1 1 F1 1, 1 − δ, z
Na
performance for the individual users—recall Example 1—and Z ∞ K̄
−z
the corresponding average performance for both individual log e 1−e K̄ e 1 1
F (1,1−δ,z/N a )
= K̄2 dz. (68)
users and cells can then be readily computed by means of e 0 z F
1 1 (1, 1 − δ, z/N a)
(58) or (59). Example 8. Let η = 4 and Na = 100 with an equal-SIR
Example 6. Example 1 established that, for η = 4, we should power allocation. For K = 10, (64) returns an average user
have Na /K ≥ 5 to ensure that no more than 3% of users fall spectral efficiency of C̄ Eq = 2.61 b/s/Hz and an average sum
below 0 dB of SIR. Plugged into Prop. 5, such Na /K returns spectral efficiency of C̄ΣEq = 10 C̄ Eq = 26.1 b/s/Hz per BS. For
C̄ Unif = 1.97 b/s/Hz while the corresponding sum spectral K being Poisson with mean K̄ = 10, (66) returns C̄ Eq = 2.69
efficiency becomes C̄ΣUnif = Na C̄ Unif /5 = 0.39 Na. b/s/Hz while (68) yields C̄ΣEq = 25.56 b/s/Hz per BS.

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
8

TABLE II: Spectral efficiency (b/s/Hz) with Na = 100 and η = 4. 1 1

K = 10 Poisson K with K̄ = 10
0.8 0.8 Analytical
C Unif 1.60 1.51
3%-ile b/s/Hz per user
C Eq 2.20 1.94 Analytical
C̄ Unif 2.76 2.84 0.6 0.6
Average b/s/Hz per user

CDF

CDF
C̄ Eq 2.61 2.69 Lattice
C̄ΣUnif
27.6 27.08 0.4 Lattice simulation
Average b/s/Hz per BS 0.4
C̄ΣEq
26.1 25.56 simulation

0.2 0.2
= {0, 10, 14} = {0, 10, 14}
0 0
For the reader’s convenience, the results in Examples 7–8 0 3 6 9 12 0 3 6 9 12
are brought together in Table II, alongside the 3%-ile user (dB) (dB)
spectral efficiencies from Examples 4–5. Not surprisingly, (a) Uniform power allocation; analysis (b) Equal-SIR power allocation; anal-
from Prop. 1. ysis from Prop. 2.
with an equal-SIR power allocation, the 3%-ile user spectral
efficiency improves significantly, but the spatial average slips Fig. 3: CDF of SIR with fixed K = 10, Na = 100 and η = 4. The analytical
results, in solid, are contrasted against hexagonal network simulations without
relative to a uniform power allocation. These results also (σdB = 0 dB) and with (σdB = 10 dB and 14 dB) shadowing.
corroborate the observation [31, Remark 4.1] that the sum
spectral efficiency with a uniform power allocation is higher 1 1
than with an equal-SIR power allocation. Fairness does come Analytical
at a price. 0.8 0.8
Analytical
0.6 0.6 Lattice
VIII. A PPLICATION TO R ELEVANT N ETWORK
CDF

CDF
simulation
G EOMETRIES Lattice
0.4 0.4
simulation
Let us proceed to verify, via Monte-Carlo, that the analytical
results in the foregoing sections closely abstract the perfor- 0.2 0.2 = {0, 10, 14}
= {0, 10, 14}
mance of network geometries of interest with practical values
0 0
for the shadowing standard deviation. The number of network 0 3 6 9 12 15 0 3 6 9 12 15
snapshots is chosen to ensure a 95% confidence interval of (dB) (dB)
±0.07% (absolute value) in the CDFs, which at the median (a) Uniform power allocation; analysis (b) Equal-SIR power allocation; anal-
corresponds to at most ±0.06 dB in SIR. from Prop. 3. ysis from Prop. 4.
Fig. 4: CDF of SIR for random K, with Na = 100, K̄ = 10 and η = 4. The
analytical results, in solid, are contrasted against hexagonal network results
A. Hexagonal Lattice Networks without (σdB = 0 dB) and with (σdB = 10 dB and 14 dB) shadowing.
Let the BS locations Φb conform to a hexagonal lattice. In
the following examples, Monte-Carlo results are generated for
the users served by the BS at the center of a lattice of 499 values of σdB is conspicuous, similar to the corresponding
hexagonal cells. The shadowing is lognormal. observation made in the context of nonmassive single-user
communication [21]. For σdB = 10 dB, with either power
Example 9. Let Na = 100, K = 10 and η = 4. To ensure allocation and regardless of whether K is fixed or a truncated
a fixed K, for every network snapshot users are dropped Poisson random quantity, the performance in a hexagonal
uniformly over the entire network until K are being served by network is within 1 dB of our analytical characterizations.
the central BS. Figs. 3a–3b demonstrate the convergence of the Furthermore, recalling that FSIRUnif (·) depends on Na and
SIR CDFs to our characterizations of FSIRUnif
k
(·) and FSIREq (·) k
K only through Na /K, the examples for a uniform power
in Props. 1–2, respectively. The dotted and dashed curves are allocation correspond verbatim to any Na and K related by a
respectively without (σdB = 0 dB) and with (σdB = 10 dB and factor of 10.
14 dB) shadowing in the hexagonal network, while the solid We further note that Example 10 confirms the negligible loss
curves correspond to our analytical characterizations. in accuracy incurred by conducting the analysis with a pure
Example 10. Still on a hexagonal network, let the users now Poisson distribution for K, rather than the actual truncated
conform to a PPP with density K̄λb while Na = 100, K̄ = 10 distribution.
and η = 4. For each snapshot, the SIRs are computed for
the users served by the central BS. (If there are over 100 B. PPP Networks
such users, a truncation takes place and 100 of them are
For an irregular deployment of BSs and the ensuing vari-
selected uniformly at random.) In Figs. 4a–4b we illustrate
ability in cell sizes, let us consider a network where the
how, as the shadowing strengthens, the SIR CDFs obtained via
BS locations are themselves PPP-distributed. Specifically, we
simulation in this hexagonal network tend to the corresponding
consider Φb = Φ∪{o}, where Φ ⊂ R2 is a homogeneous PPP
characterizations of FSIRUnif (·) and FSIREq (·) in Props. 3–4,
k and o denotes the origin [33]. Then, by Slivnyak’s theorem, the
respectively.
central BS becomes the typical BS under expectation over Φb .
The closeness of our analytical abstractions (which corre- In the Monte-Carlo examples that follow, we drop BSs (500 on
spond to σdB → ∞) to the behaviors with typical outdoor average) around the central BS and, as in the previous section,

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
9

1 1 1 Analytical 1
Analytical Analytical
Analytical
0.8 }
= 10 PPP
= 0 simulation 0.8
= 10
=0 } 0.8
PPP
simulation 0.8
PPP
simulation
PPP =0 =0
simulation
0.6 0.6 0.6 = 10 0.6 = 10

CDF

CDF
CDF
CDF

0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
0 3 6 9 12 0 3 6 9 12 0 3 6 9 12 15 18 0 3 6 9 12 15 18
(dB) (dB) (dB) (dB)
(a) Uniform power allocation; analysis (b) Equal-SIR power allocation; anal- (a) Uniform power allocation; analysis (b) Equal-SIR power allocation; anal-
from Prop. 1. ysis from Prop. 2. from Prop. 3. ysis from Prop. 4.
Fig. 5: CDF of SIR with K = 10, Na = 100, and η = 4. The analytical Fig. 6: CDF of SIR under random K, with Na = 100, K̄ = 10, and η = 4.
results, in solid, are contrasted against Poisson network simulation results The analytical results, in solid, are contrasted against PPP network simulation
without (σdB = 0 dB) and with (σdB = 10 dB) shadowing. results without (σdB = 0 dB) and with (σdB = 10 dB) shadowing.

IX. I MPACT OF N OISE AND P ILOT C ONTAMINATION

our focus is on the users associated with this central BS. The
shadowing is lognormal. A. Noise

Before proceeding, a relevant observation made in the The interference-limited regime reflects well the operating
absence of shadowing (whereby the users are confined to conditions of mature systems, justifying our SIR-based anal-
the serving BS’s Voronoi cell) is worth reproducing [33]: ysis with thermal noise neglected. Nevertheless, for the sake
the BS-to-user link distances {rℓ,(k) } as perceived by user of completeness, we herein corroborate that noise does not
k in the typical cell are distributed significantly differently significantly alter the applicability of our expressions.
from those seen from the typical location on R2 , which The signal-to-interference-plus-noise ratio (SINR) at the kth
underpin Fρ (·) in our analysis. This effect manifests in the user can be formulated as [13, Sec. 10.5.1]
following examples, although only as a very slight deviation Na PPk G2(k)
from the SIR CDF characterizations, and the convergence with SINRk =
P
(69)
shadowing is rather immediate. G(k) + ̺SNR σ 2 /P 2
ℓ Gℓ,(k) + σ /P

where ̺SNR is the ratio between the forward- and reverse-link

Example 11. With a fixed K, again effected for each snapshot signal-to-noise ratios and the reverse-link pilots are assumed
of the PPP BSs by uniformly dropping users until K of them not to be power-controlled. With a uniform power allocation,
are served by the central BS, and with Na = 100, K = 10 and the SINR specializes to
η = 4, Figs. 5a–5b show the convergence of the Monte-Carlo
CDFs to our characterizations in Props. 1–2, respectively for N /K
SINRUnif
k =   a , (70)
FSIRUnif
k
(·) and FSIREq (·). 1 + ̺SNR P G(k) σ2
1 + 1/ρk + σ2
P G(k)

In the next example, where K is allowed to be random, P

where, recall, ρk = G(k) / ℓ6=0 Gℓ,(k) . In turn, with an equal-
we validate the applicability of our results for a Poisson-
SINR power allocation,
distributed K (Props. 3–4). As advanced in Section II-B, and
expounded in more detail in [30], the simulated K acquires Na
SINREq = P   .
its Poisson nature only as the shadowing intensifies. This K−1
1+
2
̺SNR P σG(k) 1 + 1/ρk + σ2
k=0 P G(k)
causes a gap between analysis and simulation in the absence
of shadowing, but this gap closes quickly as shadowing makes (71)
its appearance (cf. Figs. 6a–6b).
Example 13. Reconsider Example 12 with λb = 1 BS/km2 .
The path loss intercept is Lref = −128 dB at 1 km, which is a
Example 12. Consider again a PPP network of BSs, with
reasonable value for such BS density at around 2 GHz. Each
the users conforming also to a PPP of density K̄λb , and with
BS emits P = 47.8 dBm on the forward link, the reverse-link
Na = 100, K̄ = 10 and η = 4. For each snapshot, the PPPs of
pilot power is 20 dB lower, and the noise figures at the user and
BSs and users are realized over the whole network region and
BS receivers are 7 and 3 dB, respectively, such that ̺SNR = 16
the SIRs are computed for the users served by the central BS
dB. The bandwidth is the standard 20 MHz of LTE, the noise
with their number again truncated at K = 100. In Figs. 6a–6b
power spectral density is N0 = −174 dBm/Hz, and the BS
we show how, as the shadowing strengthens, the SIR CDFs
antennas feature 6 dBi of vertical gain, which adds to G(k) .
obtained via Monte-Carlo quickly tend to the corresponding
The CDFs of SINRUnif Eq
k and SINR , respectively simulated via
characterizations in Props. 3–4, respectively for FSIRUnif (·) and
k (70) and (71) with shadowing (σdB = 10), are plotted in Fig. 7
FSIREq (·).
alongside their SIR counterparts replotted from Fig. 6.

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
10

1 1 1 1
Analytical Analytical

0.8
Analytical
PPP
simulation
= 10
0.8 Analytical PPP
simulation
= 10
0.8
} Lattice simulation
= 10 0.8 }
Lattice
simulation
= 10

0.6 0.6 0.6 0.6

CDF

CDF
CDF
CDF

=7
0.4 0.4 0.4 0.4
=7
=4
SINR SINR =4
0.2 0.2 0.2 0.2
} SIR } SIR
0 0 0 0
0 3 6 9 12 15 0 3 6 9 12 15 0 3 6 9 12 0 3 6 9 12
(dB) (dB) (dB) (dB)
(a) Uniform power allocation; analysis (b) Equal-SIR power allocation; anal- (a) Uniform power allocation; analysis (b) Equal-SIR power allocation; anal-
from Prop. 3. ysis from Prop. 4. from Prop. 1. ysis from Prop. 2.
Fig. 7: CDFs of SINR and SIR under random K, with Na = 100, K̄ = 10 Fig. 9: CDFs of SIR with fixed K = 10, Na = 100 and η = 4. The
and η = 4. The simulated SINR is contrasted against the SIR, both analytical hexagonal network simulations incorporate pilot contamination via (73) and
and simulated. (74), respectively for SIRUnif Eq
k and SIR , as well as shadowing (σdB = 10 dB).

1) Hexagonal Lattice Networks: In lattice networks, regular

pilot reuse patterns can be established. Precisely, we consider
pilot reuse patterns whereby each pilot is used in only one cell
within every cluster of L adjacent cells. Such reuse patterns,
illustrated in Fig. 8 for a hexagonal network with L = 4
and L = 7, are implementationally convenient. And they are
suboptimum in the face of shadowing, meaning that they are a
(a) Pilot reuse factor L = 4. (b) Pilot reuse factor L = 7. conservative choice for our assessment; with a dynamic pilot
Fig. 8: Pilot reuse patterns in a hexagonal grid network: the shaded cells reuse assignment, the impact of contamination would abate even
the same pilots. further.
Example 14. Consider again η = 4, Na = 100 and K = 10.
As in the foregoing example, an excellent match between CDFs of SIRUnif
k and SIR
Eq
with pilot contamination, respec-
the SINRs and SIRs is observed for all other cases reported tively corresponding to (73) and (74) and computed via Monte-
in the paper. Carlo for a hexagonal network with shadowing (σdB = 10 dB),
are plotted in Figs. 9a–9b. The results for pilot reuse factors
L = 4 and L = 7 are plotted alongside their counterparts
B. Pilot Contamination that forgo contamination (replotted from Fig. 3), namely the
hexagonal simulation result with L → ∞ and the analytical
Finally, in order to gauge the deviations that pilot con-
curve.
tamination may bring about, we contrast our solutions with
simulations that do incorporate contamination. Denoting by For L = 7, and even for L = 4, the contamination is seen
P the set of indices of the BSs reusing the pilots of the kth to have a minor effect. Our analysis is therefore applicable
user of interest, when the LMMSE (linear minimum mean- provided these reuse factors are feasible, and that is indeed
square error) channel estimation incorporates the ensuing the case over a broad operating range. For a fading coherence
contamination [13, Sec. 10.5.1] of Nc = 1000 symbols, which is a reasonably low value,2
the pilot overhead equals KL/1000. For L = 4 and L = 7,
PNa Pk
G2(k)
G(k) + ℓ∈P G(ℓ,k) P and with the values of K corresponding to Na = 100 or even
SIRk = P P
Gℓ,(k) + PNa Pℓ,k
G2ℓ,(k) Na = 200 antennas, such overhead is perfectly acceptable. For
ℓ ℓ∈P Gℓ,(k) + Gℓ,(l,k) P
l∈P
a condition determining the combinations of Na , K, and η
(72)
under which the contamination has a minor effect with tighter
where the reverse-link pilots are again assumed not to be pilot reuse patterns (L = 3 or even L = 1), readers are referred
power-controlled. With a uniform power allocation, and with to [13, Example 10.4].
a fixed K in all the contaminating cells, the above expression 2) PPP Networks: In nonlattice networks, the establish-
specializes to ment of pilot reuse patterns requires site-specific planning on
the part of the operator, or else some clustering algorithm [34].
Na /K
P
1+ G(ℓ,k) /G(k) As such schemes are well beyond the scope of our paper,
Unif ℓ∈P
SIRk = P Gℓ,(k) Na /K the next example abstracts them and gauges the impact of
1 + 1/ρk + P
ℓ∈P G(k) 1+ l∈P Gℓ,(l,k) /Gℓ,(k)
(73) 2 This value is obtained as N = ⌈ v ⌉ with a wavelength of λ = 0.15
c λ Bc
m, a coherence bandwidth of Bc = 370 kHz, and a velocity of v = 100
while, with an equal-SIR power allocation it becomes (74). km/h.

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
11

Na
SIREq = P P
 P Gℓ,(k)
 (74)
K−1 P Na /K
k=0 1+ ℓ∈P G(ℓ,k) /G(k) 1 + 1/ρk + ℓ∈P G(k) 1+ l∈P Gℓ,(l,k) /Gℓ,(k)

1 ACKNOWLEDGMENT
Analytical
PPP simulation; = 10 The efficient editorial handling by Prof. Kaibin Huang and
the excellent feedback provided by the reviewers are gratefully
0.8
acknowledged.

A PPENDIX A
0.6 P ROOF OF L EMMA 2
CDF

Capitalizing on the fact that the characteristic function,

0.4
ϕX (ω) = E[eiωX ], can be inverted as per the Gil-Pelaez
theorem [35] to compute
With pilot Z
contamination 1 1 ∞ dω
P[X < x] = − ℑ{e−ixω ϕX (ω)} , (76)
0.2 2 π 0 ω
we write
No contamination  
0 Fρ (θ) = P 1 − θ/ρ < 0 (77)
0 2 4 6 8 10 Z
(dB)
1 1 ∞  dω
= − ℑ ϕ1−θ/ρ (ω) (78)
2 π 0 ω
Fig. 10: CDF of SIRUnif
k for Na /K = 10 and η = 4 in a PPP network. Z ∞
The contamination is incorporated to the simulations via (73) with an average 1 1  dω
reuse factor of L = 8.5. The analytical curve is from Prop. 1.
= − ℑ eiω ϕ1/ρ (−θ ω) (79)
2 π 0 ω
Z ∞ n h io
1 1 dω
= − ℑ eiω E e−iθω/ρ (80)
2 π 0 ω
contamination in a manner that is not specific to any pilot Z ∞
1 1  dω
allocation approach. = − ℑ eiω L1/ρ (iθω) (81)
2 π 0 ω
Example 15. Consider again a PPP network with η = 4, and plug the Laplace domain characterization [36, Lemma 2]
Na = 100, K = 10 and σdB = 10. For each network h i
realization, out of 500 random pilot allocations (wherein each L1/ρ (s) = E e−s/ρ (82)
interfering BS belongs to P with probability 1/7), we select 1
the one minimizing the average contamination power per BS, = (83)
F
1 1 (−δ, 1 − δ, −s)
computed as es
= (84)
1 X X K−1
X 1 F1 (1, 1 − δ, s)
Gℓ,(l,k) , (75) into (81) to complete the proof.
|P0 |
l∈P0 ℓ∈P0 ,ℓ6=l k=0
A PPENDIX B
where P0 = {0} ∪ P is the set of all mutually contaminating CDF VIA N UMERICAL I NVERSION OF L APLACE
BSs including the central BS while |P0 | is its cardinality. The T RANSFORM
effective pilot reuse factor becomes L = 8.5 on average.
Lemma 3. [37, Prop. 1] The CDF of ρ can be computed as
In Fig. 10, the CDF of SIRUnifk with pilot contamination,
simulated via (73) with the aforedescribed procedure, is plotted B   Q+b  
eA/2 X B X (−1)q L1/ρ (θ t)
along side its contamination-free counterpart (replotted from Fρ (θ) ≈ 1 − B ℜ
2 b q=0 Dq t
Fig. 5a) and the analytical curve from Prop. 1. b=0
(85)
where t = (A + i2πq)/2 while D0 = 2 and Dq = 1 for q ≥ 1,
X. C ONCLUSION
and recall that L1/ρ (·) is in (84). In turn, the parameters A, B
While formally derived for interference-limited networks and Q control the accuracy of the approximation. Suggested
devoid of pilot contamination and asymptotic in the shadowing values for numerical Laplace transform inversions with many
strength, the expressions presented in this paper for conjugate digits of precision are A = 18.4, B = 11 and Q = 15 [38].
beamforming are broadly applicable, as confirmed by a mul- Following the recommendations of [39] for a 4-digit accuracy,
titude of examples. This broad applicability of the analysis a more relaxed—but still plentiful for our purposes—precision
makes its extension to other transmit and receive strategies, target, we obtain A = 9.21 and B = 5, with Q as large as
power allocations, user association policies, and even channel possible; as it turns out, moderate values of Q suffice to yield
conditions, highly desirable. error levels that are negligible.

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
12

Z ∞
A PPENDIX C 1 − e−zNa /K dz
= log2 (e)
(99)
P ROOF OF P ROPOSITION 2 0 1 F1 1, 1 − δ, z z
Invoking (76), where in (98) we applied [40, Lemma 1] while (99) follows
Eq from [36, Lemma 2]. This proves Prop. 5.
FSIREq (θ) = P [1 − θ/SIR < 0] (86)
Z Similarly, with an equal-SIR power allocation,
1 1 ∞ n o dω
= − ℑ ϕ1−θ/SIREq (ω) (87)  
2 π 0 ω C̄ Eq = E log2 (1 + SIREq ) (100)
Z
1 1 ∞ n iω o dω Z ∞  −z/SIREq 
E e −z
(1 − e )
= − ℑ e ϕ1/SIREq (−θω) (88) = log2 (e) dz (101)
2 π 0 ω z
Z0 ∞
where the characteristic function ϕ1/SIREq (ω) = E[ei ω/SIR ]
Eq
1 − e−z dz
= log2 (e) K z
(102)
can be computed as 0 1 F1 (1, 1 − δ, z/Na )
h iω PK−1 i
ϕ1/SIREq (ω) = E e Na (K+ k=0 1/ρk ) (89) where (101) follows again from [40, Lemma 1] while (102)
"K−1 # follows from (92) in Appendix C. This proves Prop. 6.
iωK Y iω
= e Na E e Na ρk (90)
k=0 R EFERENCES
iωK
 h i ω iK
[1] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-
=e Na Eρ e Na ρ (91) bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9,
−K no. 11, pp. 3590–3600, Nov. 2010.
= 1 F1 (1, 1 − δ, −i ω/Na) (92) [2] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
MIMO for next generation wireless systems,” IEEE Commun. Mag.,
with (91) and (92) holding, respectively, because of the IID vol. 52, no. 2, pp. 186–195, Feb. 2014.
nature of ρ1 , . . . , ρK and from (84). [3] T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo, Fundamentals
of Massive MIMO. Cambridge University Press, 2016.
[4] L. Lu, G. Y. Li, A. L. Swindlehurst, A. Ashikhmin, and R. Zhang, “An
A PPENDIX D overview of massive MIMO: Benefits and challenges,” IEEE J. Select.
P ROOF OF P ROPOSITION 3 Topics Signal Processing, vol. 8, no. 5, pp. 742–758, Oct. 2014.
[5] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta,
Since the user’s SIR is a valid quantity only for K ≥ 1, the O. Edfors, and F. Tufvesson, “Scaling up MIMO: Opportunities and
CDF of SIRUnif
k with Poisson-distributed K becomes
challenges with very large arrays,” IEEE Trans. Signal Processing,
vol. 30, no. 1, pp. 40–60, Jan. 2013.
∞ [6] E. Björnson, L. Sanguinetti, J. Hoydis, and M. Debbah, “Optimal design
X fK (k)
FSIRUnif (θ) = FSIRUnif
k |K=k
(θ) (93) of energy-efficient multi-user MIMO systems: Is massive MIMO the
k 1 − FK (0) answer?” IEEE Trans. Wireless Commun., vol. 14, no. 6, pp. 3059–3075,
k=1
Jun. 2015.
⌈Na /θ⌉−1
X fK (k) [7] E. Björnson, J. Hoydis, and L. Sanguinetti, “Massive MIMO networks:
= FSIRUnif
k |K=k
(θ) Spectral, energy, and hardware efficiency,” Foundations and Trends in
1 − FK (0) Signal Processing, vol. 11, no. 3-4, pp. 154–655, 2017.
k=1
∞ [8] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
X fK (k) of cellular networks: How many antennas do we need?” IEEE J. Select.
+ (94) Areas Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013.
1 − FK (0)
k=⌈Na /θ⌉ [9] K. T. Truong, A. Lozano, and R. W. Heath, “Optimal training in
⌈Na /θ⌉−1 continuous flat-fading massive MIMO systems,” in European Wireless
1 X K̄ k Conference, May 2014, pp. 1–6.
= K̄
FSIRUnif |K=k (θ) [10] H. Q. Ngo and E. G. Larsson, “No downlink pilots are needed in TDD
e − 1 k=1 k k! massive MIMO,” IEEE Trans. Wireless Commun., vol. 16, no. 5, pp.


2921–2935, May 2017.
1 − Γ ⌈Na /θ⌉, K̄ / ⌈Na /θ⌉ − 1 ! [11] E. Björnson, J. Hoydis, and L. Sanguinetti, “Massive MIMO has
+ (95)
1 − e−K̄ unlimited capacity,” IEEE Trans. Wireless Commun., vol. 17, no. 1, pp.
574–590, 2018.
where (94) holds because FSIRUnif
k |K=k
(θ) = 1 for θ ≥ Na /k [12] K. Upadhya, S. A. Vorobyov, and M. Vehkaperä, “Downlink perfor-
while (95) follows from the application of (3)–(4). Further mance of superimposed pilots in massive MIMO systems,” IEEE Trans.
Wireless Commun., vol. 17, no. 10, pp. 6630–6644, 2018.
applying, in (95), the explicit form for FSIRUnif
k |K=k
(·) from [13] R. W. Heath and A. Lozano, Foundations of MIMO Communication.
(40) returns (50). Cambridge University Press, 2019.
Alternatively, plugging (41) into (95) yields (51). [14] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to
coverage and rate in cellular networks,” IEEE Trans. Commun., vol. 59,
no. 11, pp. 3122–3134, Nov. 2011.
A PPENDIX E [15] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and
P ROOF OF P ROPOSITIONS 5 AND 6 M. Franceschetti, “Stochastic geometry and random graphs for
the analysis and design of wireless networks,” IEEE J. Select. Areas
With a uniform power allocation and a fixed K, Commun., vol. 27, no. 7, pp. 1029–1046, May 2009.
[16] M. Haenggi, Stochastic Geometry for Wireless Networks. Cambridge,
C̄ Unif = E[log2 (1 + SIRUnif )] (96) UK: Cambridge University Press, 2012.
   [17] J. S. Gomez, A. Vasseur, A. Vergne, P. Martins, L. Decreusefond, and
Na /K W. Chen, “A case study on regularity in cellular network deployment,”
= E log2 1 + (97) IEEE Wireless Commun. Lett., vol. 4, pp. 421–424, Aug. 2015.
1 + 1/ρ
Z ∞
[18] S. Mukherjee, “Distribution of downlink SINR in heterogeneous cellular
E[e−z/ρ ] 1 − e−zNa /K e−z networks,” IEEE J. Select. Areas Commun., vol. 30, no. 3, pp. 575–585,
= log2 (e) dz (98) Apr. 2012.
0 z

1536-1276 (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/TWC.2019.2907584, IEEE
Transactions on Wireless Communications
13

[19] H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic geometry for Geordie George (S’14 - M’18) received the M.Tech.
modeling, analysis, and design of multi-tier and cognitive cellular degree in electrical engineering from the Indian
wireless networks: A survey,” IEEE Commun. Surveys Tuts., vol. 15, Institute of Technology Madras, Chennai, India, in
no. 3, pp. 996–1019, Third Quarter 2013. 2007, and the Ph.D. degree in information and com-
[20] B. Błaszczyszyn and M. K. Karray, “Spatial distribution of the SINR in munication technologies from Universitat Pompeu
Poisson cellular networks with sector antennas,” IEEE Trans. Wireless Fabra (UPF), Barcelona, Spain, in 2017. He was a
Commun., vol. 15, no. 1, pp. 581–593, 2016. Post-Doctoral Researcher with UPF, during 2017–
[21] G. George, R. K. Mungara, A. Lozano, and M. Haenggi, “Ergodic 2018. He is currently a Research Associate with
spectral efficiency in MIMO cellular networks,” IEEE Trans. Wireless the communication systems division of the Fraun-
Commun., vol. 16, no. 5, pp. 2835–2849, May 2017. hofer Institute for Integrated Circuits IIS, Erlangen,
[22] M. D. Renzo and P. Guan, “Stochastic geometry modeling and system- Germany. His research interests are in the areas of
level analysis of uplink heterogeneous cellular networks with multi- communication theory and signal processing, with a focus on the analysis of
antenna base stations,” IEEE Trans. Commun., vol. 64, pp. 2453–2476, wireless communication systems using stochastic geometry modeling.
Jun. 2016.
[23] M. D. Renzo and W. Lu, “Stochastic geometry modeling and perfor-
mance evaluation of MIMO cellular networks using the equivalent-in-
distribution (EiD)-based approach,” IEEE Trans. Commun., vol. 63, pp.
977–996, Mar. 2015. Angel Lozano (S’90 - M’99 - SM’01 - F’14)
[24] H. S. Dhillon, M. Kountouris, and J. G. Andrews, “Downlink MIMO received a Ph.D. from Stanford University in 1998.
HetNets: Modeling, ordering results and performance analysis,” IEEE In 1999, he joined Bell Labs (Lucent Technolo-
Trans. Wireless Commun., vol. 12, pp. 5208–5222, Oct. 2013. gies, now Nokia) in Holmdel, NJ, where he was a
[25] B. Błaszczyszyn, M. K. Karray, and H. P. Keeler, “Wireless networks Member of the Wireless Communications Research
appear Poissonian due to strong shadowing,” IEEE Trans. Wireless Department until 2008. He is currently a Profes-
Commun., vol. 14, no. 8, pp. 4379–4390, Aug. 2015. sor at Universitat Pompeu Fabra (UPF), Barcelona,
[26] H. P. Keeler, N. Ross, and A. Xia, “When do wireless network signals and the co-author of the textbook “Foundations
appear Poisson?” Bernoulli, vol. 24, no. 3, pp. 1973–1994, Aug. 2018. of MIMO Communication” (Cambridge University
[27] N. Ross and D. Schuhmacher, “Wireless network signals with moder- Press, 2019). He serves as Area Editor for the IEEE
ately correlated shadowing still appear Poisson,” IEEE Trans. Inform. Transactions on Wireless Communications and as
Theory, vol. 63, no. 2, pp. 1177–1198, Feb. 2017. Editor for the IEEE Communication Technology News. He received the
[28] H. ElSawy, A. Sultan-Salem, M. S. Alouini, and M. Z. Win, “Modeling 2009 Stephen O. Rice Prize for the Best Paper published in the IEEE
and analysis of cellular networks using stochastic geometry: A tutorial,” Transactions on Communications, the 2016 Fred W. Ellersick prize to the
IEEE Communications Surveys Tutorials, vol. 19, no. 1, pp. 167–203, best paper published in the IEEE Communications Magazine, and the 2016
Firstquarter 2017. Communications Society & Information Theory Society joint paper award.
[29] H. ElSawy and E. Hossain, “Two-tier HetNets with cognitive femto- He holds an Advanced Grant from the European Research Council and was
cells: Downlink performance modeling and analysis in a multichannel a 2017 Clarivate Analytics Highly Cited Researcher.
environment,” IEEE Transactions on Mobile Computing, vol. 13, no. 3,
pp. 649–663, Mar. 2014.
[30] G. George, A. Lozano, and M. Haenggi, “Distribution of the
number of users per base station in cellular networks,” IEEE
Wireless Commun. Lett., to appear. [Online]. Available: https: Martin Haenggi (S’95 - M’99 - SM’04 - F’14)
//ieeexplore.ieee.org/document/8515110 received the Dipl.-Ing. (M.Sc.) and Dr.sc.techn.
[31] H. Yang and T. L. Marzetta, “A macro cellular wireless network with (Ph.D.) degrees in electrical engineering from the
uniformly high user throughputs,” in IEEE Veh. Techn. Conf. (VTC’14), Swiss Federal Institute of Technology in Zurich
Sep. 2014, pp. 1–5. (ETH) in 1995 and 1999, respectively. Currently
[32] M. Haenggi, “The mean interference-to-signal ratio and its key role in he is the Freimann Professor of Electrical Engi-
cellular and amorphous networks,” IEEE Wireless Commun. Lett., vol. 3, neering and a Concurrent Professor of Applied and
no. 6, pp. 597–600, Dec. 2014. Computational Mathematics and Statistics at the
[33] Y. Wang, M. Haenggi, and Z. Tan, “The meta distribution of the SIR for University of Notre Dame, Indiana, USA. In 2007-
cellular networks with power control,” IEEE Trans. Commun., vol. 66, 2008, he was a visiting professor at the University
no. 04, pp. 1745–1757, Apr. 2018. of California at San Diego, and in 2014-2015 he
[34] M. Attarifar, A. Abbasfar, and A. Lozano, “Random vs structured pilot was an Invited Professor at EPFL, Switzerland. He is a co-author of the
assignment in cell-free massive MIMO wireless networks,” IEEE Int’l monographs ”Interference in Large Wireless Networks” (NOW Publishers,
Conf. on Commun. Workshops (ICCW’18), pp. 1–6, 2018. 2009) and ”Stochastic Geometry Analysis of Cellular Networks” (Cambridge
[35] J. Gil-Pelaez, “Note on the inversion theorem,” Biometrika, vol. 38, no. University Press, 2018) and the author of the textbook ”Stochastic Geometry
3-4, pp. 481–482, 1951. [Online]. Available: http://dx.doi.org/10.1093/ for Wireless Networks” (Cambridge, 2012), and he published 15 single-
biomet/38.3-4.481 author journal articles. His scientific interests lie in networking and wireless
[36] R. K. Ganti and M. Haenggi, “SIR asymptotics in Poisson cellular communications, with an emphasis on cellular, amorphous, ad hoc (including
networks without fading and with partial fading,” in Proc. IEEE Int. D2D and M2M), cognitive, and vehicular networks. He served as an Associate
Conf. Commun., May 2016. Editor of the Elsevier Journal of Ad Hoc Networks, the IEEE Transactions on
[37] R. K. Mungara, I. Thibault, and A. Lozano, “Full-duplex MIMO in Mobile Computing (TMC), the ACM Transactions on Sensor Networks, as a
cellular networks: System-level performance,” IEEE Trans. Wireless Guest Editor for the IEEE Journal on Selected Areas in Communications, the
Commun., vol. 16, no. 5, pp. 3124–3137, May 2017. IEEE Transactions on Vehicular Technology, and the EURASIP Journal on
[38] J. Abate and W. Whitt, “Numerical inversion of Laplace transforms of Wireless Communications and Networking, as a Steering Committee member
probability distributions,” ORSA J. Compt., vol. 7, no. 1, pp. 36–43, of the TMC, and as the Chair of the Executive Editorial Committee of the
1995. IEEE Transactions on Wireless Communications (TWC). From 2017 to 2018,
[39] C. A. O’cinneide, “Euler summation for Fourier series and Laplace he was the Editor-in-Chief of the TWC. He also served as a Distinguished
transform inversion,” Commun. Statist. -Stochastic Models, vol. 13, Lecturer for the IEEE Circuits and Systems Society, as a TPC Co-chair of the
no. 2, pp. 315–337, 1997. Communication Theory Symposium of the 2012 IEEE International Confer-
[40] K. A. Hamdi, “A useful lemma for capacity analysis of fading inter- ence on Communications (ICC’12), of the 2014 International Conference on
ference channels,” IEEE Trans. Wireless Commun., vol. 58, no. 2, pp. Wireless Communications and Signal Processing (WCSP’14), and the 2016
411–416, Feb. 2010. International Symposium on Wireless Personal Multimedia Communications
(WPMC’16). For both his M.Sc. and Ph.D. theses, he was awarded the ETH
medal. He also received a CAREER award from the U.S. National Science
Foundation in 2005 and three awards from the IEEE Communications Society,
the 2010 Best Tutorial Paper award, the 2017 Stephen O. Rice Prize paper
award, and the 2017 Best Survey paper award, and he is a 2017 and 2018
Clarivate Analytics Highly Cited Researcher.

1536-1276 (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.