PROCEEDINGS OF THE IEEE 1

Indoor Millimeter-Wave Systems: Design and

Performance Evaluation
Jacek Kibiłda, Allen B. MacKenzie, Mohammad J. Abdel-Rahman, Seong Ki
Yoo, Lorenzo Galati Giordano, Simon L. Cotton, Nicola Marchetti, Walid Saad,
William G. Scanlon, Adrian Garcia-Rodriguez, David López-Pérez, Holger
arXiv:2004.08649v1 [cs.NI] 18 Apr 2020

Claussen, and Luiz A. DaSilva

Abstract

Indoor areas, such as offices and shopping malls, are a natural environment for initial millimeter-
wave (mmWave) deployments. While we already have the technology that enables us to realize indoor
mmWave deployments, there are many remaining challenges associated with system-level design and
planning for such. The objective of this article is to bring together multiple strands of research to
provide a comprehensive and integrated framework for the design and performance evaluation of indoor
mmWave systems. The paper introduces the framework with a status update on mmWave technol-
ogy, including ongoing fifth generation (5G) wireless standardization efforts, and then moves on to
experimentally-validated channel models that inform performance evaluation and deployment planning.

J. Kibiłda, N. Marchetti, and L. A. DaSilva are with CONNECT, Trinity College, The University of Dublin, Ireland, E-mails:
{kibildj,nicola.marchetti,dasilval}@tcd.ie.
L. A. DaSilva is also with the Commonwealth Cyber Initiative (CCI) and Electrical and Computer Engineering, Virginia Tech,
US, E-mail: [email protected].
A. B. MacKenzie, M. J. Abdel-Rahman, and W. Saad are with Electrical and Computer Engineering, Virginia Tech, US,
E-mails: {mackenab,mo7ammad,walids}@vt.edu.
A. B. MacKenzie is also with the Electrical and Computer Engineering Department at Tennessee Tech, US, E-mail:
[email protected].
M. J. Abdel-Rahman is also with the Electrical Engineering and Computer Science Departments at Al Hussein Technical
University, Jordan.
S. L. Cotton is with Centre for Wireless Innovation, ECIT Institute, Queen’s University Belfast, UK, E-mail: si-
[email protected]
S. K. Yoo is with School of Computing, Electronics and Mathematics, Coventry University, UK, E-mail:
[email protected]
W. G. Scanlon is with Tyndall National Institute, Ireland, E-mail: [email protected].
L. Galati Giordano, A. Garcia-Rodriguez, D. López-Pérez, and H. Claussen are with Nokia Bell Labs, Ireland, E-mails:
{lorenzo.galati giordano,holger.claussen}@nokia-bell-labs.com,[email protected],[email protected]
PROCEEDINGS OF THE IEEE 2

Together these yield insights on indoor mmWave deployment strategies and system configurations, from
feasible deployment densities to beam management strategies and necessary capacity extensions.

Index Terms

millimeter-wave communications, 5G-NR, millimeter-wave channel, network modelling, deploy-

ment planning

I. I NTRODUCTION

Our motivation in this article is to present a comprehensive framework for performance

evaluation and design practices dedicated to indoor millimeter-wave (mmWave) networks. The
framework integrates inputs from multiple areas of mmWave networking expertise, from stan-
dardization, to channel measurements and modelling, to system-level evaluations and deployment
planning. Its major contribution is insights into indoor mmWave deployment planning strategies
and system configurations that are grounded and informed by experimentally-validated channel
models.
mmWaves are a natural choice for mobile indoor deployments due to much shorter link
distances, weak penetration through walls, and large available bandwidths. In fact, the concept of
using mmWaves to provision multi-Gbps speeds over indoor areas dates back to works published
in the late 90’s [1], [2]. Later came a suite of standards that offered the possibility of ad hoc
networking, including ECMA-387 [3], IEEE 802.15.3c [4], and more recently IEEE Wireless
Gigabit 802.11ad/ay [5], [6]. Nowadays, the importance of mmWaves keeps increasing, since
they are one of the key mobile broadband networking features of the fifth generation (5G)
wireless [7], [8].
At the time of writing, the first set of 5G-New Radio (NR) specifications had already been
defined as part of the 3rd Generation Partnership Project (3GPP) Release 15, functionally frozen
in September 2018. Work on Release 16 is ongoing and is planned to be frozen in June 2020, with
Release 17 to follow. The frequency range (FR) considered for deploying mmWave technology,
as agreed in Release 15, is from 24.25 GHz to 52.6 GHz, also referred to as FR21 . FR2 currently
addresses three bands: 24.25–27.5 GHz (n257), 26.5–29.5 GHz (n258), and 37–40 GHz (n260),
all meant to support time division duplex (TDD) operation only. Usage of frequencies between
52.6 GHz and 71 GHz is currently under study in Release 17. In particular, the spectrum around

1
FR1 addresses the spectrum between 410 MHz and 7.125 GHz.
PROCEEDINGS OF THE IEEE 3

Fig. 1. Status of the millimeter-wave spectrum for 5G-NR.

60 GHz presents an attractive use case, as it: a) does not require licenses, and b) is harmonised
globally [9]. In this spectrum, the 3GPP is planning to deploy NR-based access technology –
NR–Unlicensed (NR-U) – which will incorporate extensions necessary to work in unlicensed
spectrum. NR-U will account for national restrictions of the 60 GHz band, which the 3GPP had
already identified in Release 14 [10]. In Fig. 1, we summarize the adoption of mmWave spectrum
for 5G-NR.
As part of its 5G activity, the 3GPP defined the enhanced mobile broadband (eMBB) Indoor
Hotspot scenario as a natural case study for indoor mmWave mobile networks [11]. The key
purpose of this scenario is to provision small-cell coverage of high data rates and capacity to
a user population within a confined area. Typical examples for this type of scenario include
an open office, airport hall, or shopping mall. For such a setup, one would be interested in
understanding the required density of access points, deployment locations, and effective network
settings. Understanding these system design aspects requires adequate system-level evaluation
that involves similarly adequate representation of the mmWave channel.
There are many ongoing measurement campaigns being conducted around the globe with the
aim of characterizing and modeling the mmWave propagation channel, e.g., [12]–[16]. By their
nature, all these measurements are context-specific (as any experimental work), and, depending on
the modeling technique used, pertain to different types of system-level evaluations. For example,
in an open office scenario, radio infrastructure is typically mounted to the ceiling or walls,
illuminating the main area inside [17]. In such a setup, the main factor limiting signal propagation
is shadowing by physical objects [18]. In particular, human bodies introduce extra attenuation,
referred to as body blockage, which may vary with the orientation and position of the bodies with
respect to both the device and the serving access point [19]. However, whether the communication
link is blocked or not will also have a discernible effect on the fading characteristics of the
associated wireless channel. Broadly speaking, blocked links often display a richer multipath
PROCEEDINGS OF THE IEEE 4

structure, with weaker direct component and larger delay spread, than the unobstructed ones
[14]. Moreover, it was found that within indoor locations, such as large offices and hallways,
body-related blockage has a less pronounced effect on the received signal power, most likely
due to the increased scattering in the environment [19]. The trick is to capture all of these
propagation-related effects in a model that is both accurate and amenable to the analysis at
system-level.
In what follows we describe our framework by first discussing our channel measurement
campaign and a modelling approach that we subsequently integrate into both system-level
evaluations and deployment planning. Using this framework, we provide insights on feasible
deployment densities, beam management strategies, and necessary capacity extensions. While
the framework can be applied to any indoor mmWave networking scenario, our case study
focuses on an open office environment for illustrative purposes.

II. I NDOOR MM WAVE CHANNEL

It is well known that many aspects of wireless communications, e.g., system design, network
topology and performance, are dependent upon an accurate understanding of the channel char-
acteristics. Therefore, channel characteristics at mmWave must be comprehensively studied to
allow detailed channel models to be developed. This knowledge will both inform the design of
future mmWave communications systems and help predict important performance measures such
as the achievable coverage probabilities and capacities. In wireless communications channels, the
characteristics of the received signal are often characterized in terms of path loss, shadowing and
small-scale fading. In what follows, we introduce the κ-µ fading model and provide empirical
evidence for its utility and versatility in the context of mmWave communications, in particular
for indoor hotspot eMBB applications, which is one of the five test use cases selected in [17].

A. Indoor mmWave Channel Measurements

A number of studies have recently been conducted in [5], [13], [20]–[24], which are key
to the provisioning of future 5G services at mmWave frequencies for both indoor and outdoor
environments. In most previous mmWave channel studies [22]–[24], the measurements have
considered the case where both the transmitter (TX) and receiver (RX) are stationary and mounted
on stands. In some of these studies [22], [23], obstructions in the optical line-of-sight (LOS)
path between the TX and RX are analyzed. Crucially, however, none of these studies considers
PROCEEDINGS OF THE IEEE 5

scenarios where the TX and/or RX are in close proximity to the human user body. To this end, in
this work, we have considered the case where a hypothetical user equipment (UE) (i.e., the TX in
our case) is held by the user while emulating different UE operating modes. Furthermore, the user
is moving through the environment. This is driven by the need to understand the potential impact
of UE operation mode and user mobility on the small-scale fading characteristics, knowledge of
which will be essential for determining the localized performance of future user-centric mmWave
networks.
1) Measurement System and Environments: To emulate a possible indoor hotspot eMBB use
case the mmWave channel between a UE and access point (AP) is considered. The hypothetical
UE and AP used for the mmWave channel measurements are shown in Fig. 2(a) and (b), respec-
tively. Details of the hypothetical UE and AP can be found in [25]. The channel measurements
were conducted within a hallway (17.38 m × 1.40 m) and an open office area (10.62 m × 12.23 m)
as shown in Fig. 3. Both the hallway and open office environments are located on the 1st floor
of the Institute of Electronics, Communications and Information Technology (ECIT) at Queen’s
University Belfast in the United Kingdom. Both environments featured metal studded dry walls
with a metal tiled floor covered with polypropylene-fiber, rubber backed carpet tiles, and metal
ceiling with mineral fiber tiles and recessed louvered luminaries suspended 2.70 m above floor
level. The open office area contained a number of soft partitions, cabinets, personal computers,
chairs and desks. It should be noted that both the hallway and open office environments were
unoccupied for the duration of the channel measurements and the AP was placed above a ceiling
tile with the antenna boresight oriented towards the floor, i.e., imitating a ceiling-mounted AP.
2) Usage Cases: During the measurements, the UE was operated by an adult male of 1.83 m
in height and mass 80 kg. A number of different use cases likely to be encountered in everyday
scenarios were considered. These were: (1) operating an app, where the user held the UE with
his two hands in front of his body; (2) carrying a device I, where the UE was located in the
right-front pocket of the user’s clothing; (3) carrying a device II, where the user held the UE
with his right hand beside his right leg. Herein, and for brevity, we denote the three different
UE usage cases as app, pocket and hand respectively. In this study, we considered two different
dynamic channel conditions, namely, (1) mobile LOS and (2) mobile non-LOS (NLOS), where
the user walked towards and away from the hypothetical AP in a straight line, respectively. It
is worth remarking that the NLOS channel conditions occurred when the direct optical path
between the UE and AP was obstructed by the user’s body, i.e., self-blockage.
PROCEEDINGS OF THE IEEE 6

UE AP

50 MHz I I
Signal CW I/Q Power HMC6000LP711E HMC6001LP711E Red Pitaya
Generator Splitter Q Q
Transmitter Receiver v1.4

59 mm
177 mm

(a) (b)

Fig. 2. Hypothetical (a) UE and (b) AP used for the mmWave channel measurements.

Communal
Office Office Office Office Office Office Office
Area

14 m Cylindroid
AP Column
Hallway

Open office area

9m
Advanced SOC Lab
Media Test
Cabinet

Cabinet
Network Room
Room
Laboratory

AP

Fig. 3. Indoor hallway and open office environments considered in this study along with annotated user trajectories.

To improve the validity and robustness of the parameter estimates obtained in this study, all
the measurements were repeated three times. Due to the dissimilar sizes of each environment,
the considered walking distances were different. In particular, these were 9 m and 14 m for the
open office area and hallway respectively. The average walking speed maintained by the user
throughout all of the experiments was approximately 1 m/s.
3) Path Loss: The path loss is a measure of the signal attenuation between the transmitter
and receiver as a function of the separation distance and can be expressed as [26]

P [dB] = P0 [dB] + 10α log10 (d/d0 ) , (1)

PROCEEDINGS OF THE IEEE 7

where P0 represents the path loss at the reference distance d0 , α is the path loss exponent which
indicates the rate at which the path loss increases with distance and d is the separation distance
between the transmitter and receiver. To obtain estimates for P0 and α, we first removed the
maximum effective isotropic radiated power (EIRP) and gain at the receiver from the raw signal
power received by the AP and then performed linear regression. To enable this, the elapsed
time was first converted into a distance using the test user’s velocity. The reference distance,
which should be in far field region of the antenna, was chosen to be 1 m for all environments.
The mean values of the parameter estimates for P0 and α averaged over all the trials for all
of the considered cases are given in Table I along with the body blockage, which is defined as
the difference between the path loss at the reference distance (i.e., P0 ) for the LOS and NLOS
conditions2 . The path loss exponents for the LOS scenarios were found to be smaller than that
associated with isotropic radiation in free space (α = 2). This was possibly due to the waveguide
effect which can often be present within indoor environments [28]. As anticipated, for both the
hallway and office environments, the P0 values for the NLOS were greater than those for the
LOS due to the shadowing effects caused by the test user’s body. When considering the values
of the body blockage, it was observed that the hand case had a smaller body blockage compared
to those for the app and pocket cases. This confirms the intuition that the UE to ceiling-mounted
AP channel is less susceptible to body blockage when the user is carrying the UE further away
from the body, which is the case with the hand scenario.
4) Small-Scale Fading: The κ-µ distribution has been proposed as a generalized statistical
model, which may be used to characterize the random variation of the received signal caused
by multipath fading [29]. The probability density function (PDF) of the signal envelope, H, in
a κ-µ fading channel can be expressed as
µ+1
hµ −µ (κ + 1) h2
   
2µ(κ + 1) 2 p h
fH (h) = µ−1 µ+1 exp Iµ−1 2µ κ (κ + 1) √ , (2)
κ 2 exp (µκ) Ω 2 Ω Ω

where Iv (·) represents the modified Bessel function of the first kind with order v. In terms of
its physical interpretation, κ is defined as the ratio between the total power in the dominant
signal components and the total power in the scattered signal components, µ is the number of
multipath clusters and Ω is the mean signal power given by Ω = E[H 2 ].

2
Although we define the body blockage as the difference between the path loss at the reference distance for the LOS and
NLOS conditions in this paper, the human body blockage loss is also presented in [27] using the double knife-edge diffraction.
PROCEEDINGS OF THE IEEE 8

TABLE I
AVERAGE PARAMETER E STIMATES FOR THE PATH L OSS M ODEL AND κ-µ FADING M ODEL FOR A LL OF THE C ASES

Environ- LOS NLOS Body

Use Case
ment ᾱ P¯0 [dB] κ̄ µ̄ Ω̄ ᾱ P¯0 [dB] κ̄ µ̄ Ω̄ Blockage
app 1.92 78.31 2.80 0.77 1.16 1.93 95.39 0.67 0.96 1.25 17.09 dB
Hallway pocket 1.92 82.55 2.64 0.78 1.17 1.95 95.60 0.47 1.02 1.24 13.05 dB
hand 1.93 90.42 1.89 0.88 1.18 1.94 97.49 0.89 0.99 1.22 7.06 dB
app 2.58 81.31 1.14 1.00 1.21 1.03 101.41 0.48 1.00 1.26 20.09 dB
Office pocket 1.38 92.32 1.46 0.91 1.21 1.01 102.11 0.46 1.00 1.26 9.79 dB
hand 1.52 95.74 1.24 0.93 1.21 1.38 101.83 0.50 1.04 1.24 6.09 dB

The small-scale fading was extracted by removing both the path loss and large-scale fading3
from the channel data before transforming the data to its linear amplitude. The parameter
estimates for the κ-µ fading model were obtained using a non-linear least squares routine.
Table I provides the mean parameter estimates for the κ-µ fading model averaged over all three
trials for each of the UE to AP channels.
Our first observation is that the obtained results provide evidence for the correctness of our
methodology. For all of the LOS scenarios and environments, the κ parameters were found to
be greater than unity (κ > 1), indicating the presence of a strong dominant signal component. In
contrast, for the NLOS scenarios (i.e., when the direct signal path was blocked by user’s body),
the κ parameters were smaller than unity (κ < 1), indicating the absence of a dominant signal
contribution. Additionally, for the NLOS scenarios, the κ parameters obtained for the hand case
were slightly greater than those for the app and pocket cases, suggesting that the UE to ceiling-
mounted AP link for the hand is less impacted by the human body blockage. The µ parameters
for the LOS scenarios were slightly smaller than those obtained for the NLOS scenarios. This
suggests that the signal fluctuation observed in the LOS scenarios is less impacted by multipath
clustering compared to that experienced in NLOS. As an example of the fits obtained, Fig. 4
presents the empirical PDFs of the small-scale fading alongside the κ-µ PDFs for all of the UE
usage cases while the operator walked towards the hypothetical AP in the hallway environment.
It is clear that the κ-µ fading model is in excellent agreement with the measurement data.
To ascertain the most probable model between the κ-µ and Rayleigh distributions for char-

3
The large-scale fading was extracted from the received signal power by first removing the estimated path loss using the
parameters given in Table I. The resulting dataset was then averaged using a moving window of 100 channel samples (equivalent
to a distance of 10 wavelengths).
PROCEEDINGS OF THE IEEE 9

Fig. 4. Empirical (bars) and theoretical (continuous lines) PDFs of the small-scale fading observed in the hallway environment
for the LOS scenario with (a) app, (b) pocket and (c) hand cases.

TABLE II
AICc RANK FOR ALL OF THE CONSIDERED UE- TO -AP CHANNEL MEASUREMENT DATA .

LOS NLOS
Environment Use Case
κ-µ Rayleigh κ-µ Rayleigh
app 1 2 1 2
Hallway pocket 1 2 1 2
hand 1 2 1 2
app 1 2 1 2
Office pocket 1 2 1 2
hand 1 2 1 2

acterizing the small-scale fading observed in the UE-to-AP channels, the Akaike information
criterion (AIC) was employed. More specifically, the second-order AIC (AICc ) was used in a
similar manner to the analysis employed in [30], such that

2M (M + 1)
AICc = −2 ln (l (θ|data)) + 2M + (3)
n−M −1

where ln (l (θ|data)) is the value of the maximized log-likelihood for the unknown parameter θ
of the model given the data, M is the number of estimated parameters available in the model,
and n is the sample size. It should be noted that the lower the AICc value, the more likely the
candidate model was to have generated the data. As shown in Table II, the κ-µ and Rayleigh
fading models were ranked according to their AICc . It is clear that the κ-µ distribution was
selected as the best model for all of the considered cases, suggesting that the added complexity
(i.e., additional parameters) in the κ-µ model is worthwhile.
PROCEEDINGS OF THE IEEE 10

III. S YSTEM -L EVEL M ODELLING AND P ERFORMANCE E VALUATION

Indoor mmWave networks will be characterised by much shorter distances between access
points and users as well as deployments that will be confined to closed areas and made in-
dependent from outdoor deployments due to weak out-in penetration [31]. In Section II, we
also reported that in the presence of a human user body the indoor mmWave channel becomes
highly complex with a mix of large- and small-scale fading effects. Understanding how the
basic network parameters such as access point density, antenna beamwidth, and serving distance
will impact network performance in these conditions requires new and extensive system-level
modelling approaches.
In this section we develop an analytical framework and quantify the performance of indoor
mmWave networks. Our framework bridges the statistical channel models obtained from mea-
surement campaigns, described in Section II, with stochastic network geometry, which reflects the
fact that at any given point in time the set of user locations and/or transmitter directionalities may
be considered random. Our framework yields analytical expressions describing the distribution
of the signal-to-interference and noise ratio (SINR) in a generic indoor mmWave network.
While the obtained expressions are too complex to be amenable to intuition, they allow us
to generate numerical results that we can use to cross-validate corresponding Monte Carlo
simulations. Ultimately, we use both the analytical expressions and simulations to explore the
relationship between basic network parameters, channel scenarios and network performance
indicators: coverage, area traffic capacity (ATC), and experienced data rate (EDR), which we
define in Section III-B.
Most work to date on system-level evaluations for mmWave networking are focused on large-
scale outdoor areas, e.g., [32]–[34]. Indoor area mmWave network analysis has received much
less attention, with some earlier works considering device-to-device applications [35], [36], and
more recent works addressing hotspot deployments [37]–[40]. While building on the relevant
literature, our work presented in this section goes beyond it in two important ways. First,
expanding on our work in [41], we offer analysis based on the experimentally-validated κ-µ
fading model, which allows us to consider the impact that different device usage scenarios have
on network performance. Then, the network setup we consider is aligned with network setups
recommended for system-level evaluation of mobile mmWave systems [11], [42]. For these we
provide analytical expressions describing network performance which, for special cases, can be
PROCEEDINGS OF THE IEEE 11

reduced to closed-form expressions.

A. System Model

Network Geometry: We consider a network with nTX transmitters Φ = {x0 , x1 , . . . , xnTX −1 } ⊂

W, whose locations are uniformly distributed over a finite arbitrary area W ⊂ R2 . Our analysis
takes the perspective of the reference receiver, which is an arbitrary point in W at distance ρ0
from the center of mass of W. Since we are interested in analyzing feasible networks, we consider
W to always contain the reference receiver and the transmitter serving that reference receiver
(the serving transmitter), which, in our notation, is always indexed as 0. On the horizontal
plane, the reference receiver and the serving transmitter are a distance r0 ≥ 0 apart. We will
refer to the wireless link between the reference receiver and its transmitter as the reference link.
Transmissions from remaining n0TX = nTX − 1 transmitters will be treated as interference to the
reference link. We assume that all transmitters are located at height hTX above the ground, with
the reference receiver operating at height hRX .
Directionality Gain: All transmitters and receivers utilize directional transmission and re-
ception. We assume the transmitter (receiver) antenna gain is calculated based on the cone-bulb
model [37]. In this model the main-lobe gain GTX is a function of the beamwidth ωTX , normalized
over the whole spherical surface:

1 − cos(ωTX /2) 1 + cos(ωTX /2)

GTX − gTX = 1, (4)
2 2

where gTX is the side-lobe gain. Correspondingly, we will use GRX , gRX , and ωRX to refer to the
main-lobe gain, side-lobe gain, and beamwidth of the receiver antenna. Usage of the cone-bulb
model and the main-lobe beamwidth as our parameter abstracts our analysis from the choice of
the antenna system, and makes our results applicable to scenarios with fixed directional antennas.
This could be the case, for example, in lecture halls or open offices where users remain static.
Conversion between beamwidth, antenna gain, and number of antenna elements for various types
of antennas can be found in [43]. Let us note that the cone-bulb model is also a risk-averse choice
as it leads to underestimation of coverage results [44].
Then the directionality gain is the product of transmitter and receiver antenna gains, which
depend on the beam alignment between the two [45]. Using the convention that the alignment
gain between the receiver and a transmitter is random, we can represent it as a four-dimensional
categorical random variable Gi , with i ∈ {0, . . . , nTX − 1}. This random variable maps from the
PROCEEDINGS OF THE IEEE 12

space of all possible alignments between the transmitter i and the reference receiver to the space
of alignment gains G that is formed as a product of the transmitter and receiver antenna gains,
with the corresponding probability mass function pgi = P(Gi = gi ). In practice the alignment
gain from the serving transmitter is maintained stable by beam management operations, which
corresponds to conditioning G0 = g0 , where G0 is the alignment gain on the reference link and
g0 a constant.
Blockage: In our framework, link blockage arises from self-blockage, i.e., blockage of the
LOS signal path between the receiver and the transmitter by the human user. The severity of
this blockage will depend on the usage scenarios defined in Section II: app with a device held
in front of the user, pocket with a device held in the front pocket, hand with a device carried in
the hand. For each scenario, we consider fixed blockage on the reference link4 and probabilistic
blockage on all other links. Consequently, blockage is a Bernoulli random variable, with the
success probability pLOS and pNLOS = 1 − pLOS , where the latter is referred to as the blockage
probability. Here we opt for a model where pLOS (or, indeed, pNLOS ) is simply a system parameter
that reflects the frequency of link blockages. Furthermore, we assume that each transmission link
experiences blockages independently of all other links. Numerical evaluations in [40] show that
this independence assumption results in negligible differences to the blockage probability.
Propagation Model: We use the model as proposed in Section II, but to capture the bifurcation
of its parameters (due to link blockage) we adopt the following notation for the path loss and
power fading:
(µ +1)/2
θ1,tt (µt −1)/2
 p 
lt (ri ) = γt riαt , fHi,t (h) = (µt −1)/2
h exp (−θ1,t h − θ2,t ) Iµt −1 2 θ θ
1,t 2,t h , (5)
θ2,t

where ri , with i ∈ {0, . . . , nTX − 1}, is the distance to the i-th transmitter, t ∈ {LOS, NLOS}
blockage state, γi,t path loss at the reference distance (in linear scale), αt path loss exponent,
µt (1+κt )
θ1,t = √
Ωt
, and θ2,t = µt κt . Scattering and diffraction via static objects, such as inner walls
or office furniture, are not explicitly modelled, yet their impact is included in the parameter
values given in Table I.

4
This pertains to the worst/best case analysis.
PROCEEDINGS OF THE IEEE 13

Signal-to-Interference-and-Noise Ratio: Given a transmitter at xi and blockage state t ∈

{LOS, NLOS}, the instantaneous power received at the reference receiver located in y0 ∈ W is:

Si,t = Gi Hi,t lt (ri ) , (6)

p
where ri = ||xi − y0 ||2 +(hTX − hRX )2 is the distance between the reference receiver and
transmitter i, Hi,t is the blockage-dependent power fading, and lt (ri ) is the blockage-dependent
path attenuation. Then the SINR, under blockage state t on the serving channel is given by

S0,t
SINR = , (7)
I + τ −1
Pn0TX
where S0,t is the instantaneous power received from the serving transmitter, I = i=1 Si is
the interference power, with Si = EVi [Si,Vi ] being the received signal power from transmitter i
averaged over random blockage states Vi ∈ {LOS, NLOS} on link i, and τ is the signal-noise
ratio5 .

B. Network Performance Characterization

We utilize three key performance indicators recommended by the 3GPP [11]: coverage, ATC,
and EDR. Broadly speaking, the first represents the coverage achievable in our network, the
second provides us with information on the expected user throughput, and the third represents
the throughput achievable by a subset of users with the worst channel.
Coverage: We represent the coverage as the probability that the SINR is greater than some
pre-defined coverage threshold ζ, which is equivalent to finding the complementary cumulative
distribution function (CCDF) of the SINR:

c
P(SINR > ζ) = FSINR (ζ). (8)

Area traffic capacity: ATC is a measure of the total traffic a network can serve per unit area
(in bit/s/m2 ), which can be calculated as:

Carea = λ × bw × SE, (9)

5
For simplicity, we assume the transmit power is identical across all transmitters.
PROCEEDINGS OF THE IEEE 14

where λ is the transmitter density, bw is the transmission bandwidth, and SE is the spectral
efficiency. The spectral efficiency is the average data rate per unit of spectrum and per cell,
which can be expressed6 in terms of the CCDF of the SINR as:
Z ∞ Z ∞
SE = P (log(1 + SINR) > r) dr = FcSINR (2r − 1) dr . (10)
0 0

Experienced data rate: We define the EDR as the 5th percentile of the user throughput
distribution, which can also be found numerically following the approach proposed in [46]:
Z ∞
n 1
o
qR (β = 0.05) = argmin u + FcSINR 2r/bw − 1 dr . (11)
u 1−β u

C. Analytical Results

In the following we provide our main analytical results. We start with the distribution of
the received power in Lemma 1, which we subsequently use to characterize the CCDF of the
SINR for a mmWave network distributed over an arbitrary finite area in Lemma 2. Finally, we
specialize this result to the case of a network distributed over a disk in Corollary 1.
Lemma 1 (Received Power Statistics): Given transmitter i at distance ri , directionality gain
gi , and blockage state t, the CCDF of Si,t can be expressed as:
∞ l l+µ t −1 n
θ2,t
  
X X 1 θ1,t x θ1,t
P (Si,t > x) = exp(−θ2,t ) exp − x , (12)
l=0
l! n=0
n! ϑi,t ϑi,t

where θ1,t and ϑi,t are provided in (5) and (29), respectively.
Proof: The above formula relies on a simple re-formulation of the PDF in (5), and the
subsequent calculation of the resulting cumulative distribution function (CDF), which is presented
in Appendix A.
Remark 1 (Other representation of Lemma 1): In Lemma 1 we define the received power
distribution using infinite series representation to facilitate further derivations of the SINR.
However, the distribution in question can also be represented in closed-form, as shown in [29],
utilizing the generalized Marcum-Q function.
Lemma 2 (Distribution of SINR for an Arbitrary Area): Given the distance to the serving
transmitter r0 , its directionality gain g0 and LOS blockage state t, the conditional CCDF of the

6
Noticing that the spectral efficiency is an expectation over a positive random variable log(1 + SINR).
PROCEEDINGS OF THE IEEE 15

SINR is given by

P (SINR > ζ|G0 = g0 , R0 = r0 , T0 = t) =

 ∞ l l+µ t −1 n Xn  
ζθ1,t X θ2,t
 
X 1 ζθ1,t n k
exp −θ2,t − τ
ϑ0,t τ l=0 l! n=0 n! ϑ0,t τ k=0
k
  Y
X k X X µv ki
pv pg (µv )ki θ1,v g exp(−θ2,v )Zki ,
k +k +...+k =k
k1 , k2 , . . . , kn0TX 0 g∈G
1 2 n0 1≤i≤nTX v∈{LOS,NLOS}
TX
 
!
lvki (R) θ1,v θ2,v
where Zki = ER   ki +µv 1 F1 ki + µv ; µv ; ,
 
ζθ1,t
ζθ1,t
glv (R) + θ1,v ϑ0,t
glv (R) + θ1,v
ϑ0,t

(13)

and ϑ0,t = g0 lt (r0 ), and 1 F1 (; ; ) is the confluent hypergeometric function.

Proof: The proof is provided in Appendix B.
In our numerical evaluations we will consider the case of W being a disk of fixed radius
ρ around the origin. In this case the expectation in (13) is taken with respect to the distance
between the reference receiver and a uniform point located within W with the PDF [47]:

1 2πr, 0 ≤ r ≤ ρ − ρ0

fR (r) = 2 (14)
πρ 2r arccos r2 +ρ20 −ρ2 ,
2ρ0 r
ρ − ρ0 ≤ r ≤ ρ + ρ0 .

In the case when the reference receiver is at the center of W, we get the closed-form result
stated in Corollary 1.
Remark 2 (Usability of Lemma 2): While the formula in (13) yields no immediate intuition
on mmWave network designs, it can be used to numerically evaluate indoor mmWave networks
under a variety of scenarios corresponding to:
• Arbitrarily-sized areas7 with arbitrary reference receiver locations (to test for potential
boundary effects).
• Arbitrary path loss exponents, as given in Table I.
• Other random blockage or antenna gain models.
• Deterministic deployments, in which case the expectation in (13) should be replaced with
a summation over a set of pre-defined distances.

7
Distance distributions corresponding to a variety of useful geometric shapes can be found in [48].
PROCEEDINGS OF THE IEEE 16

Corollary 1 (Specializing Lemma 2 to the case of a receiver located in the center of a disk):
Given the distance to the serving transmitter r0 , its directionality gain g0 and LOS blockage state
t, we get that Zki in Lemma 2 can be represented using the following closed-form expression:

Zki =
2γvki y 2−ki αv ζθ1,t y=√ρ2 +h2
Ψ1 (ki + µv , ki − 2/αv , µv , ki − 2/αv + 1; θ2,v , − gγv y −αv ) ,

2 ki +µv ϑ θ
ρ θ1,v (2 − ki αv ) 0,t 1,v h

(15)

where Ψ1 is the Humbert series, which can also be denoted using the Appell series notation as
F2 (a; b, −; c, c0 ; x, y).
Proof: The proof is provided in Appendix C.
Remark 3 (Computationally-Efficient Evaluation of Lemma 2): Due to the exponentially grow-
ing number of combinations that involve products of hypergeometric series, the expression in
Lemma 2 even for a relatively small number of transmitters may become cumbersome to evaluate
numerically. In order to speed up the numerical evaluations one can pre-compute the indices and
the individual terms of the multinomial expansion, and store them into two separate matrices.
During the evaluation of the expression in Lemma 2 the elements of the indices matrix can be
used to address the elements of the terms matrix, similarly to the approach proposed in [49].
Finally, calculating expectation from the definition in (10) quickly becomes computationally
challenging, in which case one may derive analytical representation to this expectation using the
approach proposed in [50].

D. Numerical Evaluations

Here we numerically evaluate system-level implications of indoor mmWave deployments,

under a variety of body blockage scenarios as described in Section II. In our evaluations, we
set the parameters for network geometry and configuration following the Indoor Hotspot eMBB
performance evaluation setup recommended by the 3GPP and International Telecommunications
Union – Radiocommunication (ITU-R)8 , in [11] and [17] respectively. For the propagation and
channel settings, we consider the models and values coming from measurements presented in
Section II. We illustrate coverage results based on numerical evaluations of the obtained analytical

8
The recommendations follow each other closely, with the ITU-R document offering more detailed assumptions, e.g., about
the ceiling or human user heights assumed.
PROCEEDINGS OF THE IEEE 17

formulas and Monte Carlo simulations 9 , while the ATC and EDR results come from Monte Carlo
simulations only due to high computational footprint of the averaging formulas in (10) and (11).
1) Evaluation Setup: The considered area has a shape of a disk of radius ρ = 12 m, which
roughly corresponds to the size of the area considered by the 3GPP. There are n0TX interfering
transmitters, located independently and uniformly within the area, and one transmitter that is
also the serving transmitter located a distance r0 apart from the reference receiver in random
direction. An example realization of our setup is illustrated in Fig. 5.
The deployment is laid out over one floor. According to our measurement setup in Section II
each transmitter is at a height of hTX = 3 m, while the reference receiver is at a height hRX =
1.5 m. Effectively the reference receiver always stays at a distance greater than the reference
distance of 1 m from the transmitter. We assume that directionalities of beams produced by n0TX
interfering transmitters are uniform and independent across space, and the transmitter/receiver
antenna side-lobe gain is gTX = gRX = -25 dBm, while the main-lobe gain is calculated, for a
given beamwidth, using (4). The receiver beamwidth is ωRX = 30° and, unless otherwise stated,
nTX = 12, ωTX = 30°, and r0 = 1 m. We assume that our network uses 200 MHz of bandwidth
(corresponding to the bandwidth of a component carrier in 5G [51]) for full buffer, downlink
transmissions, at a transmit power of 23 dBm and the noise figure of 7 dB.
Assuming channel reciprocity10 , we use the parameter values given in Table I while rounding
the µ parameters to the nearest integer value. We take this step to improve the analytical
tractability of our derivations and results (see Appendix B) meaning that they can be readily
incorporated into other communications performance analyses beyond this study. We would like
to highlight that this rounding of the µ parameter has negligible effect on the observed network
performance and is more amenable to the physical interpretation of the results, as the µ parameter
quantifies the number of clusters of scattered multipath.
In the following we will first look at the network design implications coming from our model
and consider the impact that the number of transmitters and their antenna directionality have
on the system performance. We perform our numerical analysis for an idealistic case, assuming
that our serving transmitter is at a horizontal distance of 1 m away from the reference receiver
(which, as we will show later, allows us to meet ATC and EDR targets for most of the analysed

9
Generally speaking, we will use solid lines to plot numerical values coming from analytical expressions and marks to plot
the ones coming from Monte Carlo simulations.
10
As stated in the Introduction, so far 5G NR mmWave technology has been standardized to operate in TDD mode only [52].
PROCEEDINGS OF THE IEEE 18

Fig. 5. Snapshot of the indoor mmWave hotspot layout with five transmitters. The reference receiver location is marked with a
rectangle and the transmitter locations are the small circles. The direction of transmissions for all the transmitters is illustrated
as rotation of the 2D- projection of the cone bulb antenna gain model in (4).

scenarios), with its main beam being fully aligned with the receiver beam, and an LOS channel
between the two. Subsequently, we will re-visit these assumptions and consider how network
performance changes with increased distance to the serving transmitter and when the reference
link is in NLOS state.
2) Network performance: First let us consider how the results we obtained compare to
the performance requirements imposed on the 3GPP Indoor Hotspot scenario. Following the
chairman of the 3GPP system architecture group [53], these include 1 Gbit/s of EDR, and
15 Tbps/km2 of ATC, both in the downlink direction. As we can see from Fig. 6, whether
or not we meet the performance targets is highly dependent on the usage scenario. Broadly
speaking, scenarios where the user device is held further away from the user body achieve the
ATC target with 4 transmitters in a given area and maintain values higher than target EDR
for all of the considered network densities. The reason for this is low path attenuation at the
reference distance. For that same reason Hallway pocket scenario meets target requirements. The
other scenarios require double the number of transmitters to achieve the same performance, but,
critically, do not offer target EDR.
Looking at Fig. 6(a) we see that the increasing number of transmitters leads to linear increase
in ATC, which, given the expression in (9), comes from the linear increases in the number of
available infrastructure per unit area. Due to relatively narrow beams that we use, ωTX = 30°,
this densification does not lead to increase in interference, which can be confirmed by observing
PROCEEDINGS OF THE IEEE 19

100 3000 1
Hallway app Hallway app
Hallway pocket Hallway pocket
Hallway hand 2500 Hallway hand
80
Office app Office app
Office pocket Office pocket
ATC [Tbit/s/km2]

2000 0.95
Office hand Office hand

EDR [Mbit/s]
60

Coverage
1500
40 Hallway app
1000 0.9 Hallway pocket
Hallway hand
20 Office app
500 Office pocket
Office hand
0 0 0.85
5 10 15 5 10 15 5 10 15
Number of Transmitters Number of Transmitters Number of Transmitters

(a) ATC (b) EDR (c) Coverage

Fig. 6. Impact of the number of transmitters nTX . Where nTX = 1 represents the scenario with a single serving transmitter
and no other transmitters within a given area. Other network settings are: ωTX = 30 °, r0 = 1 m, and pLOS = 0.5.

almost flat EDR curves in Fig. 6(b) (which illustrates the performance of users that would suffer
the most if any increase in interference occurred). The coverage plots in Fig. 6(c) show the
exact same story providing us with almost flat lines for each of the considered scenarios. In
conclusion, we see that for the selected network setup our network operates in the noise-limited
regime for all of the usage scenarios with the exception of the Hallway app that is characterized
by low reference distance path attenuation.
Another way to bring the data rates up would require that we increase the antenna gains
by using narrower beams, either on the transmitter (as shown in Fig. 7), or the receiver side.
Beamwidth has a much more critical impact on both the ATC and EDR. In Fig. 7(b), we can
see that for some of the scenarios the EDR drops to values below what, for example, an average
Long Term Evolution (LTE) user would experience. In order to ensure that users under all usage
scenarios enjoy target EDR would require that we use transmitter beamwidths far below our
default configuration of 30 °. Wider beams produce significant interference which greatly reduces
the performance of the worst-off users for the three scenarios with high reference distance path
attenuation. However, in the other three scenarios while the performance reduces it does so only
to a lesser degree and ATC stays above the target performance for all beamwidths considered.
When it comes to coverage (see Fig. 6(c) and Fig. 7(c)) we should remark that—as one
would expect—increasing the number of transmitters deteriorates coverage. In Fig. 6(c), coverage
degrades linearly with the number of transmitters, albeit at a negligibly small rate. This is good
news as it means that the resulting interference may be low enough to not warrant the need for
interference coordination, at least as long as directionalities of transmissions are independent
across space, an observation also made for large-scale, outdoor mmWave networks [54]. The
PROCEEDINGS OF THE IEEE 20

100 2000 1
Hallway app Hallway app
Hallway pocket Hallway pocket
80 Hallway hand Hallway hand 0.8
Office app 1500 Office app
Office pocket Office pocket
ATC [Tbit/s/km2]

Office hand Office hand

EDR [Mbit/s]

Coverage
60
0.6
1000
40 Hallway app
0.4 Hallway pocket
500 Hallway hand
20 Office app
Office pocket
0.2
Office hand
0 0
50 100 150 50 100 150 50 100 150
Beamwidth (degrees) Beamwidth (degrees) Beamwidth (degrees)

(a) ATC (b) EDR (c) Coverage

Fig. 7. Impact of the transmitter antenna main-lobe width ωTX . The antenna gain scales with ωTX according to (4). Other
network settings nTX = 12, r0 = 1 m, and pLOS = 0.5.

choice of the beamwidth has a more critical impact on the coverage performance, in Fig. 7(c). If
the coverage is to be kept at (or above) 90% mark for all of the scenarios, it is necessary that the
transmitter beamwidth stays at (or below) roughly 60 °, as wider beams may produce unmitigated
interference. Interestingly, when beams of 50 ° or more are used, we start seeing differences in
coverage between the scenarios. Similarly to changes in ATC and EDR performance this can
be explained by high discrepancy in the path attenuation at the reference distance between the
usage scenarios.
3) Network performance under design imperfections: In Fig. 8 we consider the horizontal
distance to the serving transmitter. In practical cases a particular cell will serve a particular user.
This means the serving distance may not depend on the density of transmitters or receivers. In
Fig. 8 we see that all our performance characteristics degrade with the serving distance. The
serving distance is especially critical for the coverage, as in four of our usage cases the coverage
goes below the 90% mark at roughly 5 m horizontal distance already. Interestingly this does not
affect the ATC performance which stays at relatively high values even at longer serving distances.
Yet, the performance of worst-off users becomes highly volatile to change in their distance to the
serving transmitter, and all of our scenarios fall below the 1 Gbit/s target beyond the horizontal
distance of 2 m.
We are also interested in testing how the performance changes when the reference link is
in blockage state caused by the user body. In Fig. 9 we see – following intuition – that LOS
blockage leads to performance degradation. While the coverage is affected only in a minor way,
for each of the scenarios network performance, interpreted as ATC, drops below the performance
targets, in the cases where body blockage is high (Hallway app, Hallway pocket, and Office app)
PROCEEDINGS OF THE IEEE 21

100 1200 1
Hallway app Hallway app
Hallway pocket Hallway pocket 0.9
Hallway hand 1000 Hallway hand
80
Office app Office app
Office pocket Office pocket
0.8
ATC [Tbit/s/km2]

Office hand
800 Office hand

EDR [Mbit/s]

Coverage
60 0.7
600
0.6
40 Hallway app
400 Hallway pocket
0.5
Hallway hand
20 Office app
200 0.4 Office pocket
Office hand
0 0 0.3
0 5 10 15 0 5 10 15 0 5 10 15
Serving Distance (meters) Serving Distance (meters) Serving Distance (meters)

(a) ATC (b) EDR (c) Coverage

Fig. 8. Impact of the serving distance. Other network settings nTX = 11, ωTX = 30 °, and pLOS = .5.

50 1.5
1200
LOS LOS
LOS
NLOS NLOS
40 1000 NLOS
ATC [Tbit/s/km2]

800
EDR [Mbit/s]

30

Coverage
600 1
20
400
10
200
0 0
0.5
p

nd
p

p
nd

et

nd
p
et

t
et
ap

e
ap

ap
ap

n
ck
ck

ck
ha

ck
ha

ha

d
ha

nd

et
et

ap

n
ap
ce
ay

po

ce
po

ay

po

ck
po

ck

ha
ha
ce
ay

ce
ay

ce
ffi

ay

po
ffi

po
lw

lw

ce
ce

ay
ce
ay

ay
ffi

ffi
lw

ffi
lw

lw
al

al

ce
ay

ffi
lw

O
O

O
ffi

ffi
lw

al
lw
al

al
H

O
ffi
lw

al
H
O

O
H

H
al

al

O
H
al
H

H
(a) ATC (b) EDR (c) Coverage

Fig. 9. Impact of the reference link blockage. Other network settings nTX = 11, ωTX = 30 °, r0 = 1 m, and pLOS = .5.

dropping by as much as 70%. But more critically, the 5th percentile user throughput becomes
only a small fraction of the LOS case. This is an important observation that motivates work
into blockage mitigation strategies: [55]–[57]. In our studies we also considered the impact
of blockage probability. However, for the setup we used to evaluate our system, the disparity
between LOS and NLOS channels on the interfering links was not strong enough to yield any
significant differences in performance. We can conclude that while body blockage does indeed
introduce significant attenuation to interfering links (see body blockage values in Table I), in
a well-designed network with the carefully chosen beamwidths and a reasonable number of
transmitters it should not affect network performance.
Since beam management is also a fundamental aspect to consider in mmWave systems [58],
in our numerical evaluations we also considered the impact of beam misalignments. Given our
system setup, we have observed that any misalignment between transmitter and receiver beams
substantially throttles any communication links between the two, reducing coverage to almost
PROCEEDINGS OF THE IEEE 22

zero. While this result motivates strong need for accurate beam alignment procedures, such as
[59], it is also based on a pessimistic antenna pattern model that incorporates sharp degradation
in performance in case of any misalignments. With realistic antenna patterns, misalignments
should lead to a more gradual performance degradation [60], impact of which would require
further studies.

IV. D EPLOYMENT A NALYSIS

Section III has addressed system-level modeling and performance evaluation. These tasks
are critical to understanding the basic factors that will impact the performance of mmWave
deployments across a variety of scenarios. They help us to answer questions about the impact
of factors such as base station density, antenna beamwidth, and serving distance on various
measures of performance. Nevertheless, after examining these results, actual deployment planning
for specific environments remains a challenge in mmWave systems.
In a particular environment, deployment is constrained first by the set of possible AP locations.
Then, it is necessary for deployment to consider many of the same factors considered in the
analysis in Section III. Namely, issues such as beamwidth and planned serving distance must
be considered. In addition, we saw in the last section that whether a link is LOS or NLOS has
a significant impact on system performance. For a particular deployment, the issue of LOS vs.
NLOS will depend on the location of the APs, user device orientation, as well as orientations
of all human bodies in the considered space, including that of the device user.
In this section, we discuss efficient schemes for AP deployment and beam steering in mmWave
networks. In particular, given a set of possible AP locations, we setup a stochastic optimization
problem to determine the best set of AP locations, as well as the best directions to aim the APs’
beams in order to meet coverage requirements while minimizing the deployment cost. We use
a stochastic optimization problem to model the fact that the set of user locations at any given
time is uncertain, but can be modeled as the realization of a point process. For the purposes of
this paper, we step beyond the recent work on mmWave AP deployment by assuming that each
AP generates multiple, fixed beams; dynamic beam management is left for future work.
There have been some recent works on AP deployment in mmWave networks, such as [61]–
[64]. However, these works neither consider the beam steering problem nor account for the
uncertainty in user locations. In [61], the authors proposed an automated scheme for placing
mmWave APs and gathering their line-of-sight coverage statistics, to help model small-cell
PROCEEDINGS OF THE IEEE 23

mmWave access networks. Considering the deafness and blockage problems in mmWave net-
works, in [62], [63] the authors proposed distributed schemes for association and relaying that
improve network throughput. In another AP deployment scheme, [64], the authors assumed that
APs always direct their beams in one fixed direction and considered a fixed set of UEs with
static locations. In contrast, we assume fixed beam directions, but we assume that each AP can
generate multiple beams.
Considering uncertainty in the availability of mmWave links between AP and UEs, combined
with user location uncertainty, in this section we describe a chance-constrained stochastic pro-
gramming (CCSP) [65] framework for joint AP deployment and beam steering in mmWave
networks, called DBmWave. CCSP has been recently used to model several resource allocation
problems in uncertain networks [66]–[68]. DBmWave aims at minimizing the required number of
mmWave APs to achieve a minimum network-wide coverage probability of β, which represents
the requested quality of service (QoS) level. The network-wide coverage probability constraint
formulated in this paper is in contrast to the per-user coverage probability constraint formulated
in [64]. Instead of formulating a constraint for each user to ensure that individual users are
covered with a minimum probability β, we formulate a single constraint for the entire mmWave
network that ensures that any arbitrarily selected user will be covered with this minimum
probability β. Using various reformulation techniques, we equivalently reformulate our stochastic
program as a binary linear program (BLP). Finally, we numerically analyze the performance of
DBmWave under various system settings.
Note that in addition to user location, because of significant human body shadowing in
mmWave networks, user orientation is also a significant source of uncertainty. We have addressed
this in some past work [64] and may integrate such considerations into this work in the future.

A. System Model

We consider a three-dimensional geographical area with a set N = {1, 2, . . . , N } of candidate

locations for deploying mmWave APs on the ceiling to cover the floor, as depicted in Fig. 10.
rd 1
The floor is divided into K = 2 rb
+ 2
annuli, the ith annulus consists of Mi circles, where Mi
is given by: $ % $ %
2π π
Mi =   =   . (16)
−1 rb −1 1
2 sin rd −2 rb (K−i)−rb
sin 2(i−1)
PROCEEDINGS OF THE IEEE 24

Fig. 10. Illustration of the system model considered in DBmWave.

rd represents the radius of the geographical area and rb is the radius of each circular area, as
P 2rrdb + 12
depicted in Fig. 10. We denote the set of circular areas by K, where |K|= 1 + i=2 Mi . The
kth circular area in K, denoted by Ak , is represented by a pair (ik , jk ), as illustrated in Fig. 10.
UEs are distributed in the geographical area according to the distribution fZ (z). The link
between a mmWave AP placed at location n ∈ N and the kth circular area, k ∈ K, if one of the
AP beams is steered to cover Ak , is only available with probability pnk . The maximum number
of beams that a mmWave AP can have is denoted by B.

B. Problem Statement

Given N , K, B, β, fZ (z), and pnk , n ∈ N , k ∈ K, we answer the following questions jointly

while ensuring that an arbitrarily chosen user within the geographical area of interest will be
covered with a probability ≥ β ∈ (0, 1).
1) What is the minimum number of required mmWave APs?
2) How can they be deployed optimally?
3) How can their beams be steered optimally?
PROCEEDINGS OF THE IEEE 25

C. Problem Formulation

Let ynk , n ∈ N , k ∈ K, be binary decision variables; ynk equals one if a mmWave AP is placed
at location n and one of its beams is steered to cover region k, and it equals zero otherwise. Let
Pcov be the network-wide coverage probability, i.e., the probability that an arbitrarily selected
user in the network will be covered. Then, the joint AP deployment and beam steering problem
can be formulated as:

Problem 1: Joint AP Deployment and Beam Steering

X
min 1{P ynk ≥1} (17)
{ynk } k∈K
n∈N

subject to:

Pcov ≥ β (18)
X
ynk ≤ B, ∀n ∈ N (19)
k∈K

ynk ∈ {0, 1}, ∀n ∈ N , ∀k ∈ K (20)

where 1{·} is an indicator function; 1{·} equals one if {·} is satisfied and zero otherwise, and
β ∈ (0, 1).
1) Coverage Probability Constraint: As stated earlier, the coverage probability is defined as
the probability that an arbitrarily selected user lies in a circular area that is covered by at least
one active beam. Hence, the coverage probability when the arbitrarily selected user is located
at z can be defined as:
" #
Y
P(z)
cov =E 1− (1 − δnk(z) ynk ) , (21)
n∈N

where k (z) is the index of the circular area that contains location z. δnk(z) equals one if there
is no blockage between the AP candidate location n and the circular area Ak(z) , and it equals
zero otherwise. The expectation in (21) is over blockages, which—similarly to Section III— are
assumed to be independent across links. Therefore,
Y
P(z)
cov = 1 − (1 − pnk(z) ynk ) . (22)
n∈N
PROCEEDINGS OF THE IEEE 26

Z    Z rd −2rb (K−ik )
−1 1 fRz (rz )
fZ (z) dz ≤ aik + 2 sin drz
Ak 2(ik − 1) rd −2rb (K−ik +1) 2π
   Z 2rb (ik −0.5)
−1 1 fRz (rz )
= aik + 2 sin drz ,
2(ik − 1) 2rb (ik −1.5) 2π
∀2 ≤ ik ≤ K. (24)

To compute the unconditioned coverage probability, we take the user distribution fZ (z) into
consideration as follows:
!
X Z Y
Pcov = 1 − fZ (z) dz (1 − pnk ynk ) . (23)
k∈K Ak n∈N
R
The integration Ak
fZ (z) dz over each circular area is upper-bounded by the integration over
the sector enclosed in the ik -th annulus between the two tangents of Ak , see (24). This upper
bound is used in our analysis to enhance tractability. This enables us to use the probability
  
−1 1
distribution fRz (rz ), where Rz = kzk. The term 2 sin 2(ik −1)
in (24) represents the angle
of the sector enclosed in the ik -th annulus between the two tangents of Ak . The term aik is
P R
added to ensure that k∈K Ak fZ (z) = 1. In Section IV-D, the following two user distributions
are investigated:  
r2
rz exp − z2
2σ
• Truncated Gaussian distribution, where fRgaus
z
(rz ) =  
r2
 and σ 2 represents the
σ 2 1−exp − d2
2σ
variance of the user distribution.
2rz
• Uniform distribution, where fRunif
z
(rz ) = rd2
.
2) Equivalent binary linear program: First, note that the objective function of Problem 1 is
non-linear. It can be represented in a linear form by introducing new binary decision variables,
xn , 1{P ynk ≥1} , ∀n ∈ N , and reformulating the indicator function as follows [69]:
k∈K
P
• If k∈K ynk ≥ 1 then xn = 1 can be reformulated as:

X
ynk − (M + ) xn ≤ 1 − , (25)
k∈K
P
where M is an upper bound of k∈K ynk − 1 and  > 0 is a small tolerance beyond which
we regard the constraint as having been broken. Selecting M and  to be B − 1 and 1,
P
respectively, (25) reduces to k∈K ynk ≤ B xn .
PROCEEDINGS OF THE IEEE 27

ynk ≥ 1 can be reformulated as11 :

P
• If xn = 1 then k∈K

X
ynk + m xn ≥ m + 1, (26)
k∈K
P
where m is a lower bound of k∈K ynk − 1. Selecting m to be −1, (26) reduces to
P
k∈K ynk ≥ xn .

Therefore,
X
xn = 1{P ynk ≥1} ⇐⇒ xn ≤ ynk ≤ B xn , ∀n ∈ N .
k∈K
k∈K
Q
Second, the coverage probability expression, Pcov , has the term P , n∈N (1 − pnk ynk ),
which is nonlinear in the decision variables ynk , n ∈ N , k ∈ K. Expanding P, we can see that
the nonlinear terms in P are in the form of products of binary decision variables. For example,
if N = 3, P can be expressed as:
3
X 3
Y
P=1− pik yik + p1k p2k y1k y2k + p1k p3k y1k y3k + p2k p3k y2k y3k − pik yik . (27)
i=1 i=1

To linearize a product of binary decision variables, say N

Q
i=1 yik , we introduce a new auxiliary
QN
non-negative decision variable, say yk , replace i=1 yik by yk , and add the following constraints:

yk ≤ yik , ∀i ∈ {1, 2, . . . , N }
N
X
yk ≥ yik − (N − 1)
i=1

yk ≥ 0. (28)

After reformulating the indicator function and P, as explained above, Problem 1 becomes a
BLP.

D. Numerical Analysis

1) Setup: Assuming an open indoor environment, rd and rb were selected to be 5.5 and
0.5 meters, respectively. Based on these values, we calculated the number of circular areas, as
explained in subsection IV-A, and found that K = 92. The maximum number of beams that a

11 P
Note that this condition is equivalent to k∈K ynk = 0 =⇒ xn = 0, which is already enforced by the objective function, since it aims
at minimizing the number of mmWave APs. Hence, (26) is redundant.
PROCEEDINGS OF THE IEEE 28

mmWave AP can have, B, is varied between 1 and 4. The APs are assumed to be mounted on
the ceiling, which is 10 × 10 m2 . The height of the ceiling is assumed to be 3 m, similarly to
our measurement setup in Sections II and III. The APs are deployed in a grid-based manner,
similarly to the indoor hotspot scenario presented in [17], that is, each candidate location is
with equal distance to each other, as shown in Fig. 10. Two different user distributions were
considered: (i) Gaussian distribution with mean u = 0 and variance σ = 10 and (ii) uniform
distribution. The probabilities of link availability were calculated assuming three channel effects:
path loss, small-scale fading, and blockage (the event of a user having no LOS with a certain
AP). We adopt the power-law path loss model, as presented in (1). For LOS case, we set P0 to
78.31 dB and n to 2.1. For NLOS case, we set P0 to 95.39 dB and n to 3.5. We set d0 to 1 m.
The small-scale fading is assumed to follow the κ − µ distribution given by (2). For LOS
case, we set κ to 2.80, µ to 0.77, and Ω to 1.16. For NLOS case, we set κ to 0.92, µ to 0.96,
and Ω to 1.23.
Similarly to Section III, blockages across the links between the mmWave APs and the coverage
areas are modelled as independent and identically distributed (i.i.d.) Bernoulli random variables.
pLOS denotes the probability that there is no blockage on a certain link between an AP and a
coverage area and pN LOS = 1 − pLOS denotes the probability that there is a blockage on the
link. In general, pLOS varies from one link to another; for our numerical results, pLOS is set
to 0.5 for all the links. Moreover, given the results in Fig. 6 which showed negligibly small
degradation of coverage in response to the increasing number of transmitters and similarity
in system parameters, we can assume that frequency reuse factor is  1, i.e., interference is
negligible. The Gaussian noise spectral density is set to -174 dBm/Hz.
While linearizing Problem 1 above, we assumed that, for each user, there are only three AP
candidate locations that can cover it. These three AP locations form the best (most available)
AP-user links (i.e., links with the highest pn,k values for a given k). We selected different values
of N , the number of AP candidate locations, to better characterize the behavior of the system.
We evaluated our stochastic optimization framework in terms of the required number of APs for
different coverage probabilities β. The optimization problem was solved using CPLEX.
2) Results: Fig. 11(a) shows the number of required APs as a function of the minimum
required coverage probability (β). In this figure, the number of AP candidate locations was
chosen to be N = 100, the users were assumed to be distributed according to a Gaussian
distribution. It can be seen that as β increases, more APs are needed to satisfy the coverage
PROCEEDINGS OF THE IEEE 29

50 100
25
B=4 B=4 N=100
B=3 B=3 N=25
40 B=2 80
Number of required APs

Number of required APs

B=2 20

Number of required APs

B=1 B=1
30 60
15
20 40

10
10 20

0 0 5
0.5 0.6 0.7 0.8 0.9 0.5 0.6 0.7 0.8 0.9 0.5 0.6 0.7 0.8 0.9
Minimum coverage probability ( ) Minimum coverage probability ( ) Minimum coverage probability ( )

(a) (b) (c)

Fig. 11. Number of required APs vs. minimum coverage probability for: a) Gaussian distributed users, with N = 100 beams,
b) uniformly distributed users, with N = 100 beams, and c) Gaussian distributed users and different values of N , with B = 4.

demand. Furthermore, increasing the number of beams at each AP reduces the required number
of APs. This is expected, as having more beams at an AP allows it to cover more users.
Fig. 11(b) is similar to Fig. 11(a), but assuming the users to be uniformly distributed. Both
figures show similar trends. However, the number of APs required to satisfy a certain coverage
probability is higher when the users are uniformly distributed. In the case of Gaussian distribution,
users are clustered in the geographical area (in contrast to the case of uniform distribution). This
clustering results in reducing the number of required APs.
Finally, Fig. 11(c) illustrates the effect of the number of AP candidate locations on the number
of required APs to meet a certain coverage probability. It can be seen that as N increases the
number of required APs decreases. This is due to the fact that increasing N expands the feasibility
region of the allocation problem, opening the room for better solutions (i.e., with lower objective
function value).

V. D ISCUSSION AND C ONCLUDING R EMARKS

In the preceding sections we described findings that pertain to various aspects of indoor
mmWave network design. Crucially, we showed that indoor mmWaves deployments may achieve
(or come close to achieving) performance targets of 1 Gbit/s EDR, and 15 Tbps/km2 ATC [53].
Yet, this performance can be volatile to device usage scenarios. In this section, we bring together
the results to discuss these trade-offs, and look more closely at any outstanding challenges.

A. Take Away Lessons

In the system-level analysis, we saw that our reference user achieved almost 100% coverage,
even under the presence of body blockages, in Fig. 9(c). From the deployment analysis in
PROCEEDINGS OF THE IEEE 30

Fig. 11(a) and Fig. 11(b), we saw that this does not need to be the case if multiple users are
considered. Covering each and every user above a certain reliability threshold may require deploy-
ing an excessively large number of access points, especially if we expect high reliability. These
additional transmitters will increase the network capacity, see Fig. 6(a), without deteriorating
the rate for 5th percentile users, see Fig. 6(b). However, depending on the usage scenario these
5th percentile users may operate substantially below the required performance. In Section IV,
we showed that this can be avoided by coordinating deployment locations and beam steering
according to the patterns of user locations (users that cluster in space) and user orientations
(users that face single direction). This can be observed by comparing results in Fig. 11(a) and
Fig. 11(b).
Another way to bring up the 5th percentile user performance is to increase the antenna gain
by narrowing the transmission beamwidth, as we showed in Fig. 7(b). This will come at a cost
of using more directional antennas or antenna arrays with larger number of components, and
a potential reduction in the performance of beam tracking and alignment mechanisms, which
require wider beams to operate (as reported in, e.g., [59]). This trade-off in beamwidth design can
be avoided by planning for a deployment with less stringent coverage reliability requirements,
which, as we could observe in Fig. 11(a), greatly reduces the number of necessary transmitters,
leading to reduced interference and enabling us the usage of wider beams.

B. Future Challenges

1) Exploiting Differences in Channel Characteristics Across Multiple Frequency Bands: One

way to improve link availability in mmWave networks is to make use of dual connectivity
at microwave and mmWave operating frequencies [70]. To fully exploit the extra degree of
freedom offered through dual-band operation, channel characteristics of both connections should
be adequately de-correlated (i.e., if the mmWave link goes into outage due to shadowing or a
deep fade, then it would be anticipated, that under normal circumstances, the microwave link
should act to provide a more reliable fall-back service). At present, we have a fairly good
overall understanding of the quantitative differences in propagation within the microwave and
mmWaves regions of the radio spectrum [71]. Nonetheless, important channel metrics such as
the correlation in shadowing and multipath fading across frequency bands (introduced by human
body blockages) is still largely uncharted territory, e.g., [20]. To ensure that the instantaneous
PROCEEDINGS OF THE IEEE 31

channel characteristics are suitably disparate for the envisaged multi-band operation, extensive
channel measurement campaigns, as well as follow-up system-level analyses will be necessary.
2) Coexistence between 5G-NR and WiFi: One of the key issues to address when operating
in mmWave unlicensed bands is coexistence with other wireless systems. For example, in
the 60 GHz band where the 3GPP is planning to deploy NR-U, mechanisms will have to be
developed to enable coexistence with wireless local area networks like the IEEE Wireless Gigabit
802.11ad/ay [5], [6]. As a matter of fact, the European Telecommunications Standards Institute
(ETSI) has already published a list of conformance requirements, which include limits on the
maximum emitted power or channel sensing mechanisms, necessary to ensure fair coexistence
between the systems operating in unlicensed mmWave bands [72]. Yet, more work is needed to
understand the impact of these coexistence rules on the network performance and, ultimately,
network design.
3) Network Densification Limits: Finally, in Fig. 6(a), we could see that the network capacity
of our system increases with the number of transmitters. An interesting question is to ask about
the asymptotic case, and the limits of network densification. In [73], it was observed that network
densification is limited by the channel characteristics, and antenna heights. However, this result
applies to networks operating over large areas and network settings that correspond to outdoor
deployments, e.g., variable deployment heights of tens of meters above ground level. One would
expect that for ceiling-mounted mmWave networks, with highly directional beams, we can reach
high deployment densities, without making the system interference-limited, and thus achieve high
network capacities. Yet, this would make beam alignment and tracking more challenging, thus
potentially limiting the achievable network capacity. Analysis of this network density asymptotic
regime would likely require a new modelling framework.

A PPENDIX A
P ROOF OF L EMMA 1

The power received at the reference user (as provided in (6)), conditioned on the blockage
state t and antenna gain gi , is a product of the fading random variable and a constant. This
allows us to express the PDF of Si,t using (5)
s !
(µ+1)/2  
θ1 (µ−1)/2 θ1 θ1,t θ2,t s
fSi,t (s) = (µ+1)/2 (µ−1)/2 s exp − s − θ2,t Iµt −1 2 , (29)
ϑi,t θ2 ϑi,t ϑi,t
PROCEEDINGS OF THE IEEE 32

where ϑi,t = gi lt (ri ), and Iµt −1 is the modified Bessel function of the first kind and order µt − 1.
Now, using the series representation of the modified Bessel function we get that
µt ∞ l l
θ1,t exp(−θ2,t ) X θ1,t θ2,t
 
µt −1+l θ1,t
fSi,t (s) = s exp − s . (30)
ϑµi,tt l=0
ϑl
i,t Γ(µ t + l)l! ϑi,t

Then the CDF of Si,t can be expressed as

Z ∞
P (Si,t > x) = fSi,t (s) ds
x
µt ∞ l l ∞
(a) θ1,t exp(−θ2,t ) X θ1,t θ2,t
Z  
µt −1+l θ1,t
= s exp − s ds
ϑµi,tt l=0
ϑli,t Γ(µt + l)l! x ϑi,t
∞ l Γ(l + µ , θ1,t x) (31)
X θ2,t t ϑi,t
= exp(−θ2,t )
l=0
l!
Γ(l + µt )
∞ l l+µt −1 n
θ2,t
  
(b) X X 1 θ1,t x θ1,t
= exp(−θ2,t ) exp − x ,
l=0
l! n=0 n! ϑi,t ϑi,t

where for (a) we use the expression in (30), and (b) holds only for the special case of µt being
a positive integer.

A PPENDIX B
P ROOF OF L EMMA 2

In the following, given the distance to the serving transmitter r0 , serving link gain g0 , and
the serving channel being in state t, and denoting the longer-term average power received from
the serving transmitter as ϑ0,t = g0 lt (r0 ), we derive the conditional CCDF of the SINR as
c
experienced by the reference user, i.e., Pcov = FSINR|R0 ,G0 ,T
(ζ|r0 , g0 , t). For a given threshold ζ,
PROCEEDINGS OF THE IEEE 33

this conditional CCDF can be defined as

Pcov = P S > ζ · I + τ −1



 
∞
X l l+µ
θ2,t X t −1 E[( ζθ1,t
ϑ0,t
(I + τ −1 n ζθ1,t −1
)) exp − ϑ0,t (I + τ ) ]
= exp(−θ2,t )
l=0
l! n=0
n!
 ∞ l l+µ t −1 n X n  
ζθ1,t X θ2,t
   
(a) X 1 ζθ1,t n −n+k k ζθ1,t
= exp −θ2,t − τ E[I exp − I ]
ϑ0,t τ l=0 l! n=0 n! ϑ0,t k=0
k ϑ 0,t

∞ l l+µt −1 n n  
θ2,t
 X  
ζθ1,t X 1 ζθ1,t X n k
= exp −θ2,t − τ
ϑ0,t τ l=0 l! n=0 n! ϑ0,t τ k=0
k
 !k !!
nTX nTX
X ζθ1,t X
E Gi Hi,V lV (Ri ) exp − Gi Hi,V lV (Ri ) 
i=1
ϑ 0,t i=1
 ∞ l l+µ t −1 n Xn  
ζθ1,t X θ2,t
 
(b) X 1 ζθ1,t n k
= exp −θ2,t − τ
ϑ0,t τ l=0 l! n=0 n! ϑ0,t τ k=0
k
  Y   
X k ki ζθ1,t
E (Gi Hi,V lV (Ri )) exp − Gi Hi,V lV (Ri )
k +k +...+k =k
k1 , k2 , . . . , knTX 1≤i≤n ϑ 0,t
1 2 nTX TX

 ∞ l l+µ t −1 n X n  
ζθ1,t X θ2,t
   
(c) X 1 ζθ1,t n k X k
= exp −θ2,t − τ
ϑ0,t τ l=0 l! n=0 n! ϑ0,t τ k=0
k k1 +k2 +...+knTX =k
k1 , k2 , . . . , knTX
 
µ k
!
V i
Y  (µV )ki θ1,V exp(−θ2,V ) (Gi lV (Ri )) θ1,V θ2,V
1 F1 ki + µV ; µV ; ζθ1,t ,

EV,Gi ,Ri   ki +µV
1≤i≤nTX
ζθ1,t
G l (R ) + θ ϑ
Gi lV (Ri ) + θ1,V
ϑ0,t i V i 1,V 0,t

| {z }
I1

(32)

where (a) comes from the binomial expansion, in (b) we apply the multinomial expansion and
use the fact that antenna gains, channel fading, and distances are independent across all the
interferers, (c) comes from taking the expectation with respect to the fading random variable
(one can find it by either considering series representation of the PDF of fading or considering
the integral in [74][Eq. 6.643.2]).
Now, the expectation in the expression above captures three (at least potentially) random
parameters of any interfering transmitter, which are the blockage state, directionality and distance,
PROCEEDINGS OF THE IEEE 34

and can be repesented as:

X X µv ki
I1 = pv pg (µv )ki θ1,v g exp(−θ2,v )×
v∈{LOS,NLOS} g∈G
! (33)
h lvki (R) θ1,v θ2,v i
ER  ki +µv 1 F1 ki + µv ; µv ; ζθ1,t
,
ζθ1,t
glv (R) + θ1,v ϑ0,t
glv (R) + θ1,v
ϑ0,t

where pg corresponds to the directionality gain probability model, and pv corresponds to the
blockage probability model.

A PPENDIX C
P ROOF OF C OROLLARY 1

We can find the expectation with respect to the distance of a user located in the center of a
disk as follows:
 
!
ki
(lv (r)) θ1,v θ2,v
Zki = ER   ki +µv 1 F1 ki + µv ; µv ;
 
ζθ1,t 
ζθ1,t
glv (r) + θ1,v ϑ0,t
glv (r) + θ1,v
ϑ0,t

ki !
ρ
γv (r2 + η)−αv /2
Z
2 θ1,v θ2,v
= 2 ki +µv 1 F1 ki + µv ; µv ; ζθ1,t
r dr
ρ + η)−αv /2 + θ1,v

0 ζθ1,t
gγv (r2 + η)−αv /2 + θ1,v ϑ0,t
gγv (r2
ϑ0,t
Z √ρ2 +h2 !
(a) 2γvki y −αv ki +1 θ2,v
= ki +µv ki +µv 1 F1 ki + µv ; µv ; ζθ1,t
dy
ρ2 θ1,v gγv y −αv

h ζθ1,t
gγv y −αv +1 ϑ0,t θ1,v
+1
ϑ0,t θ1,v
√
∞ j Z ρ2 +h2
2γvki X (ki + µv )j θ2,v y −αv ki +1
= ki +µv ki +µv +j dy
ρ2 θ1,v (µv )j j!

h ζθ1,t
j=0
ϑ0,t θ1,v
gγv y −αv +1

2γvki y 2−ki αv
∞ j
(ki + µv )j θ2,v ζθ1,t √ρ2 +
(b) X
= 2 F1 (ki + µv + j, ki − 2/αv ; ki − 2/αv + 1; − gγv y −αv )

k +µ (µv )j j! ϑ0,t θ1,v
ρ2 θ1,v (2 − ki αv ) j=0
i v
h

∞ X ∞ l √ 2 2
2γvki y 2−ki αv

(c) X (ki + µv )j (ki + µv + j)l (ki − 2/αv )l j ζθ1,t −αv  ρ +h
= k +µ
θ 2,v − gγv y
(µv )j (ki − 2/αv + 1)l j! l! ϑ0,t θ1,v

ρ2 θ1,vi v (2 − ki αv ) j=0 l=0 h

2γvki y 2−ki αv ζθ1,t √ρ2 +h2

(d)
= Ψ1 (ki + µv , ki − 2/αv , µv , ki − 2/αv + 1; θ2,v , − gγv y −αv ) ,

2 ki +µv ϑ0,t θ1,v
ρ θ1,v (2 − ki αv ) h

(34)
p
where for (a) we use variable transformation y = r2 + η, (b) comes from [74, 3.194.1] and
is valid for ki − 2/αv > 0, (c) uses the series representation of the hypergeometric function, (d)
PROCEEDINGS OF THE IEEE 35

uses properties of the Humbert series Ψ1 .

A PPENDIX D

TABLE III: Notation Used

Symbol Description
P Pathloss (in dB)
P0 Pathloss at the reference distance (in dB)
α Pathloss exponent
d0 Reference distance
d Separation distance between the transmitter and receiver
Ratio between the total power in the dominant signal components
κ
and the total power in the scattered signal components
µ Number of multipath clusters
Ω Mean signal power
nTX Number of access points
Φ Set of access points (point process)
W Area considered
ρ0 Distance between the reference user and the origin
ρ Radius of the disk representing the considered area
ri Distance to the i-th access point
hTX(RX) Height of an access point (the reference user)
ωTX(RX) Transmitter (receiver) beamwidth
GTX(RX) Transmitter (receiver) mainlobe gain
gTX(RX) Transmitter (receiver) sidelobe gain
Gi Alignment gain with the i-th access point
pg i Probability mass function of the event Gi = gi
pNLOS Blockage probability
Pathloss at the reference distance in linear scale given access point i and
γi,t
blockage state t
αt Pathloss exponent given blockage state t
Hi,t Blockage-dependent power fading
PROCEEDINGS OF THE IEEE 36

{κt , µt , Ωt } Parameters of the fading model given blockage state t

lt (·) Blockage-dependent pathloss
Si,t Received power from the i-th access point given blockage state t
I Interference power
τ signal-to-noise ratio
ζ signal-to-interference and noise ratio threshold
FXc (·) complementary cumulative distribution function of X
Carea area traffic capacity
nTX
λ= |W|
access point density
bw System bandwidth
SE Spectral efficiency
qX (β) β-quantile of X
fX (·) probability density function of X
N Set of candidate locations for deploying millimeter-wave access points
K Number of annuli constituting the floor of the considered area
Mi Number of circles constituting the i-th annulus
rd Radius of the geographical area of interest
rb Radius of each circular area in the geographical area of interest
K Set of circular areas in the floor of the considered area
Ak k-th circular area in K
pnk Probability that the link between an access point at location n and Ak is
available
B Maximum number of beams that an access point can have
β Requested coverage probability
Pcov Network-wide coverage probability

ACKNOWLEDGMENT

This material is based in part upon work supported by the National Science Foundation
under grant CNS-1526844, the Science Foundation Ireland under grant 14/US/I3110 and the
Department for the Economy Northern Ireland through grant USI080. The authors would also
like to thank Shubhajeet Chatterjee and Fadhil Firyaguna, for their support at various stages of
the preparation of this manuscript.
PROCEEDINGS OF THE IEEE 37

R EFERENCES

[1] S. Ohmori, Y. Yamao, and N. Nakajima, “The future generations of mobile communications based on broadband access
technologies,” IEEE Communications Magazine, vol. 38, no. 12, pp. 134–142, Dec. 2000.
[2] Y. Takimoto, “Considerations on millimeter-wave indoor LAN,” in Proceedings of the IEEE Topical Symposium on
Millimeter Waves, Jul. 1997, pp. 111–114.
[3] A. Seyedi, “On the physical layer performance of ECMA-387: A standard for 60 GHz WPANs,” in Proceedings of the
IEEE International Conference on Ultra-Wideband (ICUWB), Sep. 2009, pp. 28–32.
[4] T. Baykas et al., “IEEE 802.15. 3c: the first IEEE wireless standard for data rates over 1 Gb/s,” IEEE Communications
Magazine, vol. 49, no. 7, Jul. 2011.
[5] T. Nitsche et al., “IEEE 802.11ad: directional 60 GHz communication for multi-Gigabit-per-second Wi-Fi,” IEEE
Communications Magazine, vol. 52, no. 12, pp. 132–141, Dec. 2014.
[6] Y. Ghasempour, C. R. C. M. da Silva, C. Cordeiro, and E. W. Knightly, “IEEE 802.11ay: Next-Generation 60 GHz
Communication for 100 Gb/s Wi-Fi,” IEEE Communications Magazine, vol. 55, no. 12, pp. 186–192, Dec. 2017.
[7] X. Lin et al., “5G New Radio: Unveiling the Essentials of the Next Generation Wireless Access Technology,” IEEE
Communications Standards Magazine, vol. 3, no. 3, pp. 30–37, Sep. 2019.
[8] E. Dahlman, S. Parkvall, and J. Skold, 5G NR: The Next Generation Wireless Access Technology. Elsevier Science, 2018.
[9] S. Lagén et al., “New Radio Beam-Based Access to Unlicensed Spectrum: Design Challenges and Solutions,” IEEE
Communications Surveys Tutorials, vol. 22, no. 1, pp. 8–37, Q1 2020.
[10] Study on New Radio access technology; 60 GHz unlicensed spectrum (Release 14), 3GPP, Tech. Rep. 38.805, Mar. 2017.
[11] Study on scenarios and requirements for next generation access technologies, 3GPP, Tech. Rep. 38.913 V15.0.0, Jul. 2018.
[12] Channel modeling and characterization, MiWEBA, Deliverable 5.1, Jun. 2014.
[13] 6–100 GHz channel modelling for 5G: Measurement and modelling plans in mmMAGIC, mmMagic, White Paper 2.1, Feb.
2016.
[14] G. R. Maccartney, T. S. Rappaport, S. Sun, and S. Deng, “Indoor office wideband millimeter-wave propagation
measurements and channel models at 28 and 73 GHz for ultra-dense 5G wireless networks,” IEEE Access, vol. 3, pp.
2388–2424, 2015.
[15] S. Hur et al., “Proposal on millimeter-wave channel modeling for 5G cellular system,” IEEE Journal of Selected Topics
in Signal Processing, vol. 10, no. 3, pp. 454–469, Apr. 2016.
[16] S. K. Yoo, S. L. Cotton, R. W. Heath, and Y. J. Chun, “Measurements of the 60 GHz UE to eNB channel for small cell
deployments,” IEEE Wireless Communications Letters, vol. 6, no. 2, pp. 178–181, Apr. 2017.
[17] Guidelines for Evaluation of Radio Interface Technologies for IMT-2020, ITU-R WP 5D, Tech. Rep. M.2412, Dec. 2017.
[18] Y. Niu, Y. Li, D. Jin, L. Su, and A. V. Vasilakos, “A survey of millimeter wave communications (mmWave) for 5G:
opportunities and challenges,” Wireless Networks, vol. 21, no. 8, pp. 2657–2676, Nov. 2015.
[19] S. K. Yoo, S. L. Cotton, Y. J. Chun, W. G. Scanlon, and G. A. Conway, “Channel Characteristics of Dynamic Off-
Body Communications at 60 GHz under Line-of-Sight (LOS) and Non-LOS Conditions,” IEEE Antennas and Wireless
Propagation Letters, vol. 16, pp. 1553–1556, Feb. 2017.
[20] Study on channel model for frequencies from 0.5 to 100 GHz, 3GPP, Tech. Rep. 38.901 V15.0.0, Jun. 2018.
[21] A. Maltsev, A. Pudeyev, A. Lomayev, and I. Bolotin, “Channel modeling in the next generation mmWave Wi-Fi: IEEE
802.11ay standard,” in Proceedings of the European Wireless Conference, May 2016, pp. 1–8.
[22] METIS channel models, METIS, Deliverable 1.4 v3, Jul. 2015.
PROCEEDINGS OF THE IEEE 38

[23] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, “Wideband millimeter-wave propagation measurements
and channel models for future wireless communication system design,” IEEE Transactions on Communications, vol. 63,
no. 9, pp. 3029–3056, Sept. 2015.
[24] E. M. Vitucci et al., “Analyzing radio scattering caused by various building elements using millimeter-wave scale model
measurements and ray tracing,” IEEE Transactions on Antennas and Propagation, vol. 67, no. 1, pp. 665–669, Jan. 2019.
[25] S. K. Yoo, S. L. Cotton, Y. J. Chun, and W. G. Scanlon, “Fading characterization of UE to ceiling-mounted access point
communications at 60 GHz,” in Proceedings of the European Conference on Antennas and Propagation (EuCAP), Apr.
2018.
[26] T. S. Rappaport, Wireless communications: principle and practice. New Jersey: Prentice Hall PTR, 2002.
[27] V. Raghavan et al., “Statistical blockage modeling and robustness of beamforming in millimeter-wave systems,” IEEE
Transactions on Microwave Theory and Techniques, vol. 67, no. 7, pp. 3010–3024, Jul. 2019.
[28] ——, “Millimeter wave channel measurements and implications for PHY layer design,” IEEE Transactions on Antennas
and Propagation, vol. 65, no. 12, pp. 6521–6533, Dec. 2017.
[29] M. D. Yacoub, “The κ-µ distribution and the η-µ distribution,” IEEE Antennas Propagation Magazine, vol. 49, no. 1, pp.
68–81, Feb. 2007.
[30] S. K. Yoo and S. L. Cotton, “Composite fading in non-line-of-sight off-body communications channels,” in Proceedings
of EuCAP, Apr. 2017, pp. 286–290.
[31] S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter-wave cellular wireless networks: Potentials and challenges,”
Proceedings of the IEEE, vol. 102, no. 3, pp. 366–385, Mar. 2014.
[32] M. R. Akdeniz et al., “Millimeter wave channel modeling and cellular capacity evaluation,” IEEE Journal on Selected
Areas in Communications, vol. 32, no. 6, pp. 1164–1179, Jun. 2014.
[33] T. Bai and R. W. Heath, “Coverage and rate analysis for millimeter-wave cellular networks,” IEEE Transactions on Wireless
Communications, vol. 14, no. 2, pp. 1100–1114, Feb. 2015.
[34] E. Turgut and M. C. Gursoy, “Coverage in heterogeneous downlink millimeter wave cellular networks,” IEEE Transactions
on Communications, vol. 65, no. 10, pp. 4463–4477, Oct. 2017.
[35] K. Venugopal, M. C. Valenti, and R. W. Heath, “Interference in finite-sized highly dense millimeter wave networks,” in
Proceedings of the Information Theory and Applications Workshop (ITA), Feb. 2015, pp. 175–180.
[36] ——, “Device-to-Device Millimeter Wave Communications: Interference, Coverage, Rate, and Finite Topologies,” IEEE
Transactions on Wireless Communications, vol. 15, no. 9, pp. 6175–6188, Sep. 2016.
[37] F. Firyaguna, J. Kibiłda, C. Galiotto, and N. Marchetti, “Coverage and Spectral Efficiency of Indoor mmWave Networks
with Ceiling-Mounted Access Points,” in Proceedings of the IEEE Global Communications Conference (GLOBECOM),
Dec. 2017.
[38] S. Niknam, B. Natarajan, and R. Barazideh, “Interference analysis for finite-area 5G mmWave networks considering
blockage effect,” IEEE Access, vol. 6, pp. 23 470–23 479, 2018.
[39] S. M. Azimi-Abarghouyi, B. Makki, M. Nasiri-Kenari, and T. Svensson, “Stochastic Geometry Modeling and Analysis of
Finite Millimeter Wave Wireless Networks,” IEEE Transactions on Vehicular Technology, vol. 68, no. 2, pp. 1378–1393,
Feb. 2019.
[40] F. Firyaguna, J. Kibiłda, C. Galiotto, and N. Marchetti, “Performance Analysis of Indoor mmWave Networks with Ceiling-
Mounted Access Points,” IEEE Transaction on Mobile Computing, pp. 1–1, 2020.
[41] J. Kibiłda et al., “Performance Evaluation of Millimeter-Wave Networks in the Context of Generalized Fading,” in
Proceedings of the IEEE GLOBECOM, Dec. 2018.
PROCEEDINGS OF THE IEEE 39

[42] IMT Vision – Framework and Overall Objectives of the Future Development of IMT for 2020 and Beyond, ITU–R, Tech.
Rep. M.2083–0, Sep. 2015.
[43] C. A. Balanis, Antenna theory: analysis and design. John Wiley & Sons, 2016.
[44] X. Yu, J. Zhang, M. Haenggi, and K. B. Letaief, “Coverage analysis for millimeter wave networks: The impact of directional
antenna arrays,” IEEE Journal on Selected Areas in Communications, vol. 35, no. 7, pp. 1498–1512, Jul. 2017.
[45] J. G. Andrews et al., “Modeling and analyzing millimeter wave cellular systems,” IEEE Transactions on Communications,
vol. 65, no. 1, pp. 403–430, Jan. 2017.
[46] J. Kibiłda and G. De Veciana, “Dynamic Network Densification: Overcoming spatio-temporal variability in wireless traffic,”
in Proceedings of the IEEE GLOBECOM, Dec. 2018.
[47] Z. Khalid and S. Durrani, “Distance distributions in regular polygons,” IEEE Transactions on Vehicular Technology, vol. 62,
no. 5, pp. 2363–2368, Jun. 2013.
[48] J. Guo, S. Durrani, and X. Zhou, “Outage probability in arbitrarily-shaped finite wireless networks,” IEEE Transactions
on Communications, vol. 62, no. 2, pp. 699–712, Feb. 2014.
[49] D. Torrieri and M. C. Valenti, “The outage probability of a finite ad hoc network in Nakagami fading,” IEEE Transactions
on Communications, vol. 60, no. 11, pp. 3509–3518, Nov. 2012.
[50] Y. J. Chun et al., “A Comprehensive Analysis of 5G Heterogeneous Cellular Systems Operating Over κ–µ Shadowed
Fading Channels,” IEEE Transactions on Wireless Communications, vol. 16, no. 11, pp. 6995–7010, Nov. 2017.
[51] S. Parkvall, E. Dahlman, A. Furuskar, and M. Frenne, “NR: The New 5G Radio Access Technology,” IEEE Communications
Standards Magazine, vol. 1, no. 4, pp. 24–30, Dec. 2017.
[52] Base Station (BS) radio transmission and reception, 3GPP, Tech. Spec. 38.104 V15.2.0, Jul. 2018.
[53] T. Norp, “5G Requirements and Key Performance Indicators,” Journal of ICT Standardization, vol. 6, no. 1, pp. 15–30,
May 2018.
[54] A. K. Gupta, J. G. Andrews, and R. W. Heath, “On the Feasibility of Sharing Spectrum Licenses in mmWave Cellular
Systems,” IEEE Transactions on Communications, vol. 64, no. 9, pp. 3981–3995, Sep. 2016.
[55] C. Tatino, I. Malanchini, N. Pappas, and D. Yuan, “Maximum Throughput Scheduling for Multi-Connectivity in Millimeter-
Wave Networks,” in Proceedings of the International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and
Wireless Networks (WiOpt), May 2018.
[56] V. Petrov et al., “Achieving End-to-End Reliability of Mission-Critical Traffic in Softwarized 5G Networks,” IEEE Journal
on Selected Areas in Communications, vol. 36, no. 3, pp. 485–501, Mar. 2018.
[57] F. Firyaguna, J. Kibiłda, and N. Marchetti, “Application of Flexible Numerology to Blockage Mitigation in 5G-mmWave
Networks,” in Proceedings of the IEEE GLOBECOM, Dec. 2019.
[58] Study on New Radio Access Technology Physical Layer Aspects (Release 14), 3GPP, Tech. Rep. 38.802, Sep. 2017.
[59] A. Alkhateeb, Y.-H. Nam, M. S. Rahman, J. Zhang, and R. W. Heath, “Initial beam association in millimeter wave cellular
systems: Analysis and design insights,” IEEE Transactions on Wireless Communications, vol. 16, no. 5, pp. 2807–2821,
May 2017.
[60] M. Rebato, L. Resteghini, C. Mazzucco, and M. Zorzi, “Study of Realistic Antenna Patterns in 5G mmWave Cellular
Scenarios,” in Proceedings of the IEEE International Conference on Communications (ICC), May 2018.
[61] S. S. Szyszkowicz, A. Lou, and H. Yanikomeroglu, “Automated Placement of Individual Millimeter-Wave Wall-Mounted
Base Stations for Line-of-Sight Coverage of Outdoor Urban Areas,” IEEE Wireless Communications Letters, vol. 5, no. 3,
pp. 316–319, Jun. 2016.
[62] H. Shokri-Ghadikolaei, Y. Xu, L. Gkatzikis, and C. Fischione, “User association and the alignment-throughput tradeoff
PROCEEDINGS OF THE IEEE 40

in millimeter wave networks,” in Proceedings of the IEEE International Forum on Research and Technologies for Society
and Industry (RTSI), Sep. 2015, pp. 100–105.
[63] X. Yuzhe, H. Shokri-Ghadikolaei, and C. Fischione, “Distributed association and relaying with fairness in millimeter wave
networks,” IEEE Transactions on Wireless Communications, vol. 15, no. 12, pp. 7955–7970, Dec. 2016.
[64] M. N. Soorki, M. J. Abdel-Rahman, A. MacKenzie, and W. Saad, “Joint access point deployment and assignment in
mmWave networks with stochastic user orientation,” in Proceedings of WiOpt, May 2017.
[65] P. Kall, S. W. Wallace, and P. Kall, Stochastic programming. Springer, 1994.
[66] M. J. Abdel-Rahman and M. Krunz, “Stochastic guard-band-aware channel assignment with bonding and aggregation for
DSA networks,” IEEE Transactions on Wireless Communications, vol. 14, no. 7, pp. 3888–3898, Jul. 2015.
[67] S. Chatterjee, M. J. Abdel-Rahman, and A. B. MacKenzie, “Virtualization framework for cellular networks with downlink
rate coverage probability constraints,” in Proceedings of the IEEE GLOBECOM, Dec. 2018.
[68] M. M. Gomez et al., “Market-Driven Stochastic Resource Allocation Framework for Wireless Network Virtualization,”
IEEE Systems Journal, vol. 14, no. 1, pp. 489–499, Mar. 2020.
[69] J. Linderoth, “Lecture notes on integer programming,” Jan. 2005. [Online]. Available: http://homepages.cae.wisc.edu/
∼linderot/classes/ie418/lecture2.pdf

[70] NR; Multi-connectivity; Overall description; Stage-2 (Release 16), 3GPP, Tech. Rep. 37.340, Jan. 2020.
[71] M. Shafi et al., “Microwave vs. Millimeter-Wave Propagation Channels: Key Differences and Impact on 5G Cellular
Systems,” IEEE Communications Magazine, vol. 56, no. 12, pp. 14–20, Dec. 2018.
[72] Multiple-Gigabit/s radio equipment operating in the 60 GHz band; Harmonised Standard covering essential requirements
of article 3.2 Directive 2014/53/EU, ETSI Std. EN 302 567 v2.1.1, Jul. 2017.
[73] D. Lopez-Perez and M. Ding, “A brief history on the theoretical analysis of dense small cell wireless networks,” arXiv
preprint arXiv:1812.02269, 2018.
[74] I. S. Gradshteyn and I. M. Ryzhik, Table of integrals, series and products, 7th ed. Academic Press, 2007.