Cell-Free Millimeter-Wave Massive MIMO Systems With Limited Fronthaul Capacity

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2908688, IEEE Access
SUBMITTED TO IEEE ACCESS 1
Cell-Free Millimeter-Wave Massive MIMO Systems

with Limited Fronthaul Capacity
Guillem Femenias, Senior Member, IEEE, and Felip Riera-Palou, Senior Member, IEEE
Abstract—Network densification, massive multiple-input (MU-MIMO) arrangements in which the number of antenna
multiple-output (MIMO) and millimeter-wave (mmWave) bands elements at each access point (AP) is much larger than the
have recently emerged as some of the physical layer enablers for number of mobile stations (MSs) simultaneously served over
the future generations of wireless communication networks (5G
and beyond). Grounded on prior work on sub-6 GHz cell-free the same time/frequency resources. The operation of massive
massive MIMO architectures, a novel framework for cell-free MIMO schemes is based on the availability of channel state
mmWave massive MIMO systems is introduced that considers information (CSI) acquired through time division duplexing
the use of low-complexity hybrid precoders/decoders while (TDD) operation and the use of uplink (UL) pilot signals. Such
factors in the impact of using capacity-constrained fronthaul a setting allows for very high spectral and energy efficiencies
links. A suboptimal pilot allocation strategy is proposed that is
grounded on the idea of clustering by dissimilarity. Furthermore, using simple linear signal processing in the form of conjugate
based on mathematically tractable expressions for the per-user beamforming or zero-forcing (ZF)1 [3], [5].
achievable rates and the fronthaul capacity consumption, max- In UDNs, a large number of APs deployed within a given
min power allocation and fronthaul quantization optimization coverage area cooperate to jointly transmit/receive to/from a
algorithms are proposed that, combining the use of block
(relatively) reduced number of MSs thanks to the availability
coordinate descent methods with sequential linear optimization
programs, ensure a uniformly good quality of service over of high-performance low-latency fronthaul links connecting
the whole coverage area of the network. Simulation results the APs to a central coordinating node. Coordination among
show that the proposed pilot allocation strategy eludes the APs can effectively control (or even eliminate) intercellular
computational burden of the optimal small-scale CSI-based interference in an approach that was first referred to as
scheme while clearly outperforming the classical random pilot
network MIMO [6], [7], later led to the concept of coordinated
allocation approaches. Moreover, they also reveal the various
existing trade-offs among the achievable max-min per-user multipoint (CoMP) transmission [8] and, more recently, to that
rate, the fronthaul requirements and the optimal hardware of cloud radio access network (C-RAN) [9]. In a C-RAN,
complexity (i.e., number of antennas, number of RF chains). the APs, which are treated as a distributed MIMO system,
Index Terms—Cell-free, Massive MIMO, Millimeter Wave, are connected to a cloud-computing based central processing
Hybrid precoding, Constrained-capacity fronthaul unit (CPU) in charge, among many others, of the baseband
processing tasks of all APs. Conceptually similar to the C-
I. I NTRODUCTION RAN architecture, but explicitly relying on assumptions speci-
fic of the massive MIMO regime, distributed massive MIMO-
A. Motivation and previous work based UDNs have been recently termed as cell-free massive
D RIVEN by the continuously increasing demands for high

system throughput, low latency, ultra reliability, impro-
ved fairness and near-instant connectivity, fifth generation (5G)
MIMO networks [10], [11]. In these networks, a massive
number of APs connected to a CPU are distributed across
the coverage area and, as in the cellular collocated massive
wireless communication networks are being standardized [1] MIMO schemes, exploit the channel hardening and favorable
while, at the same time, insights and innovations from industry propagation properties to coherently serve a large number of
and academia are paving the road for the coming of the MSs over the same time/frequency resources. Typically using
sixth generation (6G) [2]. As stated by Marzetta et al. in simple linear signal processing schemes, they are claimed to
[3, Chapter 1], there are three basic pillars at the physical provide uniformly good quality of service (QoS) to the whole
layer that can be used to sustain the spectral and energy set of served MSs irrespective of their particular location in
efficiencies that these networks are expected to provide: (i) the coverage area.
employing massive multiple-input multiple-output (MIMO), Since the microwave radio spectrum (from 300 MHz to 6
(ii) using ultra dense network (UDN) deployments, and (iii) GHz) is highly congested, the use of massive antenna systems
exploiting new frequency bands. and network densification alone may not be sufficient to meet
Massive MIMO systems, equipped with a large number of the QoS demands in next generation wireless communications
antenna elements, are intended to be used as multiuser-MIMO networks. Thus, another promising physical layer solution that
G Femenias and F Riera-Palou are with the Mobile Communications Group,
University of the Balearic Islands, Palma 07122, Illes Balears, Spain (e-mail: 1 As stated by Björnson et al. in [4], the simple ZF precoder approaches the
{guillem.femenias,felip.riera}@uib.es). performance provided by the capacity-achieving dirty paper coding/successive
This work has been funded in part by the Agencia Estatal de Investigación interference cancellation (DPC/SIC) precoder/decoder and, thus, the use of
and Fondo Europeo de Desarrollo Regional (AEI/FEDER, UE) under project much more complex precoding/decoding schemes can only offer negligible
TERESA (subproject TEC2017-90093-C3-3-R), Ministerio de Economı́a y performance improvements when compared to the ZF approach, at the cost
Competitividad (MINECO), Spain. of increasing the amount of feedback information from the MSs to the APs.
2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also 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/ACCESS.2019.2908688, IEEE Access
is expected to play a pivotal role in 5G and beyond 5G com- hybrid precoders and assuming the availability of capacity-
munication systems is to increase the available spectrum by constrained fronthaul links connecting the APs and the CPU.
exploring new less-congested frequency bands. In particular, The main contributions of our work can be summarized as
there has been a growing interest in exploiting the so-called follows:
millimeter wave (mmWave) bands [12]–[15]. The available • The performance of both the downlink (DL) and UL of
spectrum at these frequencies is orders of magnitude higher cell-free mmWave massive MIMO systems is considered
than that available at the microwave bands and, moreover, with particular emphasis on the per-user rate, rather than the
the very small wavelengths of mmWaves, combined with the system sum-rate, by posing max-min fairness resource allo-
technological advances in low-power CMOS radio frequency cation problems that take into account the effects of imper-
(RF) miniaturization, allow for the integration of a large num- fect channel estimation, power control, non-orthogonality of
ber of antenna elements into small form factors. Large antenna pilot sequences, and fronthaul capacity constraints. Instead
arrays can then be used to effectively implement mmWave of assuming the use of rather simple uniform quantization
massive MIMO schemes (see, for instance, [16], [17] and processes when forwarding information on the capacity-
references therein) that, with appropriate beamforming, can constrained fronthauls, the proposed optimization problems
more than compensate for the orders-of-magnitude increase in assume the use of large-block lattice quantization codes able
free-space path-loss produced by the use of higher frequencies. to approximate a Gaussian quantization noise distribution.
The performance of cell-free massive MIMO using con- Optimal solutions to these problems are proposed that
ventional sub-6 GHz frequency bands and assuming infinite- combine the use of block coordinate descent methods with
capacity fronthaul links has been extensively studied in, for sequential linear programs.
instance, [11], [18]–[20]. Cell-free massive MIMO networks • A hybrid beamforming implementation is proposed where
using capacity-constrained fronthaul links have also been the RF high-dimensionality phase shifter-based preco-
considered in [21], [22] but assuming, again, the use of fully ding/decoding stage is based on large-scale second-order
digital precoders in conventional sub-6 GHz frequency bands. statistics of the propagation channel, and hence does not
Sub-6 GHz massive MIMO systems are often assumed to need the estimation of high-dimensionality instantaneous
implement a fully-digital baseband signal processing requiring CSI. The low-dimensionality baseband MU-MIMO preco-
a dedicated RF chain for each antenna element. The present ding/decoding stage can then be easily implemented by stan-
status of mmWave technology, however, characterized by dard signal processing schemes using small-scale estimated
high-power consumption levels and high production costs, CSI. As will be shown in the numerical results section,
precludes the fully-digital implementation of massive MIMO such a reduced complexity hybrid precoding scheme, when
architectures, and typically forces mmWave systems to rely on combined with appropriate user selection, performs very
hybrid digital-analog signal processing architectures. In these well in the fronthaul capacity-constrained UDN mmWave-
hybrid transceiver architectures, a large antenna array connects based scenarios under consideration.
to a limited number of RF chains via high-dimensional RF • A suboptimal pilot allocation strategy is proposed that,
precoders, typically implemented using analog phase shifters based on the idea of clustering by dissimilarity, avoids the
and/or analog switches, and low-dimensional baseband digital computational complexity of the optimal pilot allocation
precoders are then used at the output of the RF chains [23]– scheme. The performance of the proposed dissimilarity
[25]. The network of phase shifters connecting the array of cluster-based pilot assignment algorithm is compared with
antennas to the RF chains determines whether the structure that of both the pure random pilot allocation approach and
is fully or partially connected [26]. Thus, the assumptions, the balanced random pilot strategy.
methods and analytical expressions in [11], [18]–[22] cannot • For those cases in which the number of active MSs in the
by applied directly when assuming the use of mmWave network is greater than the number of available RF chains
frequency bands. Despite its evident potential, as far as we at a particular AP, a MS selection algorithm is proposed
know, besides [27], [28] there is no other research work on that aims at maximizing the minimum average sum-energy
cell-free mmWave massive MIMO systems and, furthermore, (i.e., Frobenius norm) of the equivalent channel between the
the authors of these works did not face one of the main APs and any of the active MSs, constrained by the fact that
challenges in the implementation of cooperative UDNs, that each AP can only beamform to a number of MSs less or
is, the fact that these systems require of a substantial infor- equal than the number of available RF chains.
mation exchange between the APs and the CPU via capacity-
constrained fronthaul links. Moreover, they also considered
the use of oversimplified mmWave channel models and RF C. Paper organization and notational remarks
precoding stages, without constraining the available number The remainder of this paper is organized as follows. In
of RF-chains at each AP. Section II the proposed cell-free mmWave massive MIMO
system is introduced. Different subsections are devoted to the
description of the channel model, the large-scale and small-
B. Aim and contributions scale training phases, the channel estimation process, and
Motivated by the above considerations, our main aim in the DL and UL payload transmission phases. The achievable
this paper is to address the design and performance evaluation DL and UL rates are presented in Section III and further
of realistic cell-free mmWave massive MIMO systems using developed in Appendices A and B. Section IV is dedicated to
τLc Every τL c
Large-scale Coherence Coherence Coherence • Large-scale channel estimation (Sect II.A)
Training Interval Interval Interval • RF precoder design (Sect II.B)
τc τc τc • Selection of MSs to beamform from each AP (Sect II.C)
• Pilot assignment (Sect V)
• Max-min power allocation and quantization (Sects III, IV and VI)
τc
• Every τc
Uplink pilots Uplink payload data Downlink payload data – UL pilot transmission (Sect II.D)
– Small-scale channel estimation (Sect II.E)
τp τu τd – DL payload data transmission (Sect II.F)
Fig. 1: Allocation of the samples in large-scale and short-scale – UL payload data transmission (Sect II.G)
coherence intervals. Fig. 2: Operational schedule of the proposed cell-free mm-
Wave massive MIMO system
the calculation of the capacity consumption of both the DL and

UL fronthaul links. The pilot assignment, power allocation and The propagation channels linking the APs to the MSs
quantization optimization processes are described in Sections are typically characterized by small-scale parameters that are
V and VI. Numerical results and discussions are provided in (almost) static over a coherence time-frequency interval of
Section VII and, finally, concluding remarks are summarized τc time-frequency samples (see [3, Chapter 2]), and large-
in Section VIII. scale parameters (i.e., path loss propagation losses and co-
Notation: Vectors and matrices are denoted by lower-case variance matrices) that can be safely assumed to be static
and upper-case boldface symbols. The q-dimensional identity over a time-frequency interval τLc τc . As shown in the
matrix is represented by I q . The operator det(X) represents following subsections, these channel characteristics can be
the determinant of matrix X, tr(X) denotes its trace, kXkF is leveraged to simplify both the channel estimation and the
its Frobenius norm, whereas X −1 , X T , X ∗ and X H denote precoding/combining processes. In particular, DL and UL
its inverse, transpose, conjugate and conjugate transpose (also transmissions between APs and MSs are organized in a half-
known as Hermitian), respectively. With a slight abuse of duplex TDD operation whereby each coherence interval is
notation, the operator diag(x) is used to denote a diagonal split into three phases, namely, the UL training phase, the
matrix with the entries of vector x on its main diagonal, and DL payload data transmission phase and the UL payload data
the operator diag(X) is used to denote a vector containing transmission phase, and every large-scale coherence interval
the entries in the main diagonal of matrix X. The expectation τLc the system performs an estimation of the large-scale
operator is denoted by E{·}. Finally, CN (m, R) denotes a parameters of the channel (see Fig. 1). In the UL training
circularly symmetric complex Gaussian vector distributions phase, all MSs transmit UL training orthogonal pilots allowing
with mean m and covariance R, N (0, σ 2 ) denotes a real the APs to estimate the propagation channels to every MS in
valued zero-mean Gaussian random variable with standard de- the network2 . Subsequently, these channel estimates are used
viation σ, and U[a, b] represents a random variable uniformly to detect the signals transmitted from the MSs in the UL
distributed in the range [a, b]. payload data transmission phase and to compute the precoding
filters governing the DL payload data transmission. Not shown
II. S YSTEM MODEL are guard intervals between UL and DL transmissions. In
order to ease the identification of the different tasks that
Let us consider a cell-free massive MIMO system where need to be conducted at both the CPU and the APs, Fig. 2
a CPU coordinates the communication between M APs and provides the operational schedule of the proposed cell-free
K single-antenna MSs randomly distributed in a large area. mmWave massive MIMO system and the sections of this paper
Each of the APs communicates with the CPU via error-free describing each of these tasks.
fronthaul links with DL and UL capacities CF d and CF u ,
respectively. Baseband processing of the transmitted/received A. Channel Model
signals is performed at the CPU, while the RF operations are
carried out at the APs. Each AP is equipped with an array of MmWave propagation is characterized by very high
N > K antennas and L ≤ N RF chains. A fully-connected distance-based propagation losses that lead to sparse scattering
architecture is considered where each RF chain is connected multipath propagation. Furthermore, the use of mmWave trans-
to the whole set of antenna elements using N analog phase mitters and receivers with large tightly-packet antenna arrays
shifters. Without loss of essential generality, it is assumed in results in high antenna correlation levels. These characteristics
this paper that the number of active RF chains at each of the make most of the statistical channel models used in conventio-
APs in the network is equal to LA = min{K, L}. That is, if nal sub-6 GHz MIMO research work inaccurate when dealing
K ≤ L, all APs in the cell-free network provide service to with mmWave scenarios. Thus, a modified version of the
the whole set of MSs and if K > L, instead, each AP can discrete-time narrowband clustered channel model proposed
only provide service to L out of the K MSs in the network by Akdeniz et al. in [14] and further extended by Samimi and
and, thus, an algorithm must be devised to decide which are 2 Note that channel reciprocity can be exploited in TDD systems and
the MSs to be beamformed by each of the APs. therefore only UL pilots need to be transmitted.
Rappaport in [29] will be used in this paper to capture the where the cluster c is assumed to contribute with a fraction of
peculiarities of mmWave channels. power given by
The link between the mth AP and the kth MS will be 0
N γmk,c
considered to be in one out of three possible conditions: γmk,c = PCmk 0 , (6)
outage, line-of-sight (LOS) or non-line-of-sight (NLOS) with Pmk j=1 γmk,j
probabilities: with
0 rτ −1 Zmk,j /10
γmk,j = Umk,j 10 , (7)
pout (dmk ) = max 0, 1 − e−aout dmk +bout ,

(1a)
2
Umk,j ∼ U[0, 1], Zmk,j ∼ N (0, ζ ), and the constants rτ and
pLOS (dmk ) = (1 − pout (dmk )) e−aLOS dmk , (1b) ζ 2 being treated as model parameters (see [14, Table I]).
Although the small-scale fading gains αmk,cp are assumed
pNLOS (dmk ) = 1 − pout (dmk ) − pLOS (dmk ), (1c) to be static throughout the coherence interval and then change
respectively, where dmk is the distance (in meters) between the independently (i.e., block fading), the spatial covariance ma-
AP and the MS, and, according to [14, Table I], 1/aout = 30 m, trices
n o
bout = 5.2, and 1/aLOS = 67.1 m. Those links that are in Rmk =E hmk hH mk
outage will be characterized with infinite propagation losses, C mk
while for the links that are not in outage, the propagation
X
=10− PL(dmk )/10 γmk,c
losses will be characterized using a standard linear model with c=1
(8)
shadowing as P
Xmk
× a (θmk,cp , φmk,cp ) aH (θmk,cp , φmk,cp ) ,

PL(dmk )[dB] = α + 10β log10 (dmk ) + χmk , (2)
p=1
where α and β are the least square fits of floating intercept and are assumed to vary at a much smaller pace (i.e., τLc τc ).
slope and depend on the carrier frequency and on whether the
link is in LOS or NLOS (see [14, Table I]), and χmk denotes B. Large-scale training phase: RF precoder/combiner design
the large-scale shadow fading component, which is modelled In order to exploit the UL/DL channel reciprocity using the
as a zero mean spatially correlated normal random variable TDD frame structure shown in Fig. 1, it is assumed in this
with standard deviation σχ (again, see [14, Table I] to obtain paper that the N × LA RF matrix W RF m , describing the effects
the typical values of σχ for LOS and NLOS links) whose of the active analog phase shifters at the mth AP, is common to
spatial correlation model is described in [11, (54)-(55)]. the DL (RF precoding phase) and UL (RF combining phase).
The UL channel vector hmk ∈ CN ×1 between MS k and Furthermore, denoting by Km = {κm1 , . . . , κmLA } the set
AP m will be modelled as the sum of the contributions of of LA MSs beamformed by the mth AP, it is assumed that
Cmk scattering clusters, each contributing Pmk propagation W RFm is a function of only the spatial channel covariance
paths as matrices {Rmk }k∈Km , known at the mth AP through spatial
Cmk Pmk
channel covariance estimation for hybrid analog-digital MIMO
precoding architectures (see e.g. [30]–[33]).
X X
hmk = αmk,cp a (θmk,cp , φmk,cp ) , (3)
c=1 p=1
Using eigen-decomposition, the covariance matrix of the
propagation channel linking MS k and AP m can be
where αmk,cp is the complex small-scale fading gain on the expressed as Rmk = U mk Λmk U H mk , where Λmk =
pth path of cluster c, and a (θmk,cp , φmk,cp ) represents the AP diag ([λmk,1 . . . λmk,rmk ]) contains the rmk non-null eigen-
normalized array response vector at the azimuth and elevation values of Rmk , and U mk is the N × rmk matrix of the
angles θmk,cp and φmk,cp , respectively. These angles, as stated corresponding eigenvectors. Hence, assuming the use of (con-
by Akdeniz et al. in [14, Section III.E] can be generated strained) statistical eigen beamforming [34], [35], the analog
as wrapped Gaussians around the cluster central angles with RF precoder/combiner can be designed as
standard deviation given by the root mean square (rms) angular h i
RF
spreads for the cluster. The azimuth cluster central angles are W RF
m = w mκm1 . . . wRF
mκmLA
uniformly distributed in the range [−π, π] and the elevation h i (9)
cluster central angles are set to the LOS elevation angle. = e−j∠umκm1 ,max . . . e−j∠umκmLA ,max ,
Moreover, the cluster rms angular spreads are exponentially where umk,max is the dominant eigenvector of Rmk associated
distributed with a mean equal to 1/λrms that depends on to the maximum eigenvalue λmk,max , and the function ∠x re-
the carrier frequency and on whether we are considering the turns the phase angles, in radians, for each element of the com-
azimuth or elevation directions (see [14, Table I]). The number plex vector x. Note that using the RF precoding/combining
of clusters is distributed as a random variable of the form matrix, the equivalent channel vector between MS k and AP
m, including the RF precoding/decoding matrix, is defined as
Cmk ∼ max {Poisson(σC ), 1} , (4)
T
LA ×1
g mk = W RF
m hmk ∈ C , (10)
where σC is set to the empirical mean of Cmk . The small-scale
fading gains are distributed as whose dimension is much less than the number of antennas
of the massive MIMO array used at the mth AP, thus largely
αmk,cp ∼ CN 0, γmk,c 10− PL(dmk )/10 , (5) simplifying the small-scale training phase.
C. Large-scale training phase: selection of MSs to beamform D. Small-scale training phase

from each AP Communication in any coherence interval of a TDD-based
massive MIMO system invariably starts with the MSs sending
As previously stated, in those highly probable cases in
the pilot sequences to allow the channel to be estimated at the
which the number of active MSs in the network is greater
APs. Let τp denote the UL training phase duration (measured
than the number of available RF chains at each AP (i.e.,
in samples on a time-frequency grid) per coherence interval.
K > L), the mth AP, with m ∈ {1, . . . , M }, can only
During the UL training phase, all K MSs simultaneously
beamform to a group of L out of the K MSs in the network,
transmit pilot sequences of τp samples to the APs and thus,
which are indexed by the set Km = {κm1 , . . . , κmL }. As the
the LA × τp received UL signal matrix at the mth AP is given
RF beamforming matrices at the APs are a function of only
by
the large-scale spatial channel covariance matrices and are XK
g mk0 ϕTk0 + N p m ,
p
common to both the UL and the DL, the selection of the sets Y p m = τp Pp (13)
of MSs to beamform from each AP must also be based only k0 =1
on the available large-scale CSI. Inspired by the Frobenius where Pp is the transmit power of each pilot symbol, ϕk
norm-based suboptimal user selection algorithm proposed by denotes the τp × 1 training sequence assigned to MS k, with
Shen et al. in [36], a selection algorithm is proposed that aims kϕk k2F = 1, and N p m is an LA × τp matrix of i.i.d. additive
at maximizing the sum of the average energy (i.e., average noise samples with each entry distributed as3 CN (0, σu2 (N )).
Frobenius norm) of the equivalent channels (including the Ideally, training sequences should be chosen to be mutually
corresponding beamformer) between the M APs and the K orthogonal, however, since in most practical scenarios it
MSs with the constraints that, first, the minimum average holds that K > τp , a given training sequence is assigned
energy of the equivalent channel between the M APs and any to more than one MS, thus resulting in the so-called pilot
of the active MSs must be maximized and, second, that each contamination, a widely studied phenomenon in the context of
AP can only beamform to L MSs. Note that this optimization collocated massive MIMO systems. For additional details on
problem, which tends to provide some degree of (average) the relation between channel estimation, pilot signals length
max-min fairness among MSs, can be efficiently solved by and pilot contamination error, please refer to [3], [5], [10],
using an iterative reverse-delete algorithm (similar to that used [11], [37].
in graph theory to obtain a minimum spanning tree from
a given connected, edge-weighted graph). In particular, at E. Channel estimation
the beginning of the ith iteration of the algorithm the cell-
free network is represented by a very simple edge-weighted Channel estimation is known to play a central role in the
directed graph with M source nodes and K sink nodes, where performance of massive MIMO schemes [38] and also in the
the mth source node, representing the mth AP, is connected specific context of cell-free architectures [11]. The minimum
(i)
to a group Km of sink nodes, representing the active MSs mean square error (MMSE) estimation filter for the channel
beamformed by the mth AP. The connection (edge) between between the kth active MS and the mth AP can be calculated
(i) as
the mth source node and the lth sink node in Km is weighted n o
2
by the average Frobenius norm of the equivalent channel D mk = arg min E g mk − DY p m ϕ∗k
(i)
linking the mth AP and MS l ∈ Km , that can be obtained as p D RF −1 (14)
= τp Pp Rmk Qmk ,

T 2 T where
ξml = E wRF
ml hml = wRF RF
ml Rml w ml . (11)
RF T RF ∗
RRF H
F

mk = E g mk g mk = W m Rmk W m , (15)
The average sum energy of the equivalent channels between and

the M APs and MS k at the beginning of the ith iteration is K
X
T ∗ 2
Qmk = τp Pp RRF
mk0 ϕk0 ϕk + σu2 (N )I LA . (16)
(i) k0 =1
X
Ek = ξmk , (12)
(i)
Hence, the corresponding estimated channel vector can be
m∈Mk
expressed as
−1
ĝ mk = D mk Y p m ϕ∗k = τp Pp RRF ∗
p
(i)
where Mk is the set of APs beamforming to MS k at the mk Qmk Y p m ϕk . (17)
beginning of the ith iteration. During this iteration, the reverse- 3 Note that in the UL of a fully-connected hybrid beamforming architecture
delete algorithm removes the edge (i.e., the RF chain and each reception chain is composed of N antenna elements, each connected
associated beamformer) that, first, goes out of one of those to a low-noise amplifier (LNA) characterized by a power gain GLNA and
a noise temperature TLNA . Each of the N LNAs feeds an analog passive
APs still beamforming to more than L MSs and, second, phase shifter characterized by an insertion loss LPS . The outputs of the N
has the minimum weight maximizing the minimum average phase shifters are introduced to a power combiner whose insertion losses
sum energy after removal. The algorithm begins with a fully are typically proportional to the number of inputs, that is, LPC = N LPCin .
Finally, the output of the power combiner is introduced to an RF chain
connected graph and stops when all APs beamform to exactly characterized by a power gain GRF and a noise temperature TRF . Thus, the
L MSs. Hence, note that M (K − L) iterations are needed to equivalent noise temperature of each receive chain can be obtained as Tu =
T0 (LPS LPCin −1) TRF LPS LPCin

select the sets Km for m ∈ {1, . . . , M }. N T0 + TLNA + G LNA
+ G
.
LNA
CPU AP 1 transmitted by the mth AP. Due to the distortion introduced by

CF d the quantization/unquantization processes, we have that [39],
Power allocation (Υ) x1
W BB
d1 Qd1 Fronthaul Q−1
d1 W RF
1 [40]
Q̂dm (x) , Qd −1
m (Qdm (x)) = x + q dm , (21)
sd .. .. .. ..
. . .
AP M
. where q dm is the quantization noise vector, which is assumed

CF d
xM
to be statistically distributed as q dm ∼ CN 0, σq2dm I . As
W BB
dM QdM Fronthaul Q−1
dM W RF
M shown by Zamir et al. in [39], this assumption is supported by
the fact that large-block lattice quantization codes are able to
approximate a Gaussian quantization noise distribution. Thus,
Fig. 3: System block diagram during the DL payload trans- the mathematical operations describing the CPU-AP functional
mission phase split considered in this paper can be summarized as

1/2
xm = Pm (sd ) = W RF Q̂
m dm W BB
dm Υ sd
The MMSE channel vector estimates
RF
can be shown to be (22)
distributed as ĝ mk ∼ CN 0, R̂mk , where RF BB 1/2
= W m W d m Υ sd + q dm ,
RF −1 H iT
R̂mk , τp Pp RRF RF h
mk Qmk Rmk . (18) BB T BB T
where W BBd = W d1 . . . W dM ∈ CM LA ×K , with
Furthermore, the channel vector g mk can be decomposed W BB ∈ CLA ×K denoting the

dm = wBB BB
dm1 . . . w dmK
as g mk = ĝ mk + g̃ mk , where g̃ mk is the MMSE channel baseband precoding matrix affecting the signal transmitted by
estimation error, which is statistically independent of both g mk the mth AP, and Υ = diag ([υ1 . . . υK ]) is a K ×K diagonal
and ĝ mk . matrix containing the power control coefficients in its main
diagonal, which are chosen to satisfy the following necessary
F. Downlink payload data transmission power constraint at the mth AP
T
Let us define sd = [sd1 . . . sdK ] as the K × 1 vector n K
o X
2 2
of symbols jointly (cooperatively)
transmitted from the APs E kxm kF = BB/RF
υk θmk + σq2 dm W RF
m F
to the MSs, such that E sd sH d = I K . Let us also define k=1
(23)
xm = Pm (sd ) as the N ×1 vector of signals transmitted from XK
the mth AP, where Pm (sd ) is used to denote the mathematical = BB/RF
υk θmk + σq2 dm LA N ≤ P m,
operations (linear and/or non-linear) used to obtain xm from k=1
sdn. Note that
o this vector must comply with a power constraint where we have used the definition
2
E kxm kF ≤ P m , where P m is the maximum average n 2
o
BB/RF
θmk = E W RF BB
m w dmk F
. (24)
transmit power available at AP m. Using this notation, the
signal received by MS k can be expressed as Using the proposed hybrid CAP approach, the signal recei-
M
X ved by the K MSs can be rewritten as
yd k = hTmk xm + ndk , (19) M
X
m=1 yd = H Tm W RF BB
m W d mΥ
1/2
sd
where ndk ∼ CN (0, σd2 )
is the Gaussian noise sample at MS m=1
T M
k. The vector y d = [yd 1 . . . yd K ] containing the signals X (25)
received by the K scheduled MSs in the network can then + H Tm W RF
m q dm + nd
m=1
be expressed as
M
= GT W BB
d Υ
1/2
sd + ηd ,
X
yd = H Tm xm + nd , (20) where G = [GT1 . . . GTM ]T , with Gm = W RF
T
m H m , repre-
m=1
senting the equivalent MIMO channel matrix between the K
T
where H m = [hm1 . . . hmK ] and nd = [nd1 . . . ndK ] . MSs and the M APs, including the RF precoding/decoding
As schematically represented in Fig. 3, the mathematical matrices, and
operations that symbol vector sd undergoes before being η d = GT q d + nd , (26)
transmitted, generically represented as xm = Pm (sd ), for all
includes the thermal noise as well as the quantization noise
m ∈ {1, . . . , M }, include, first, a power allocation process and
samples received from all the APs in the network. Now, using
a baseband precoding task at the CPU, second, a quantization
the classical ZF MU-MIMO baseband precoder to harness the
process of all or part of the data that must be sent from
spatial multiplexing, we have that
the CPU to the APs through the fronthaul links and, third,
∗
T ∗ −1
an unquantization process and an RF precoding task at each
W BB
d = Ĝ Ĝ Ĝ (27)
of the APs. Let us denote by Qdm (x) and Qd −1 m (x) the
quantization and unquantization mathematical operations per- or, equivalently,
formed by the compress-after-precoding (CAP)-based CPU- ∗
T ∗ −1
AP functional split on a vector of signal samples x to be W BB
d m = Ĝm Ĝ Ĝ ∀m, (28)
T
where we have assumed that G = Ĝ + G̃ and Gm = Ĝm + to jointly process the vector z u = z u T1 . . . z u TM and obtain
G̃m . Consequently, the signal received by the kth MS can be the vector of detected samples
expressed as p 1/2
y u =W BBu zu = Pu W BBu GΩ su + η u
∗
T ∗ −1 p p (34)
yd k =g Tk Ĝ Ĝ Ĝ Υ1/2 sd + ηd k
1/2
= Pu Ω su + Pu W u G̃Ω1/2 su + η u ,
BB
where η u = W BB
∗ T ∗ −1
u (q u + nu ). Again, the first term denotes

= ĝ Tk + g̃ Tk Ĝ Ĝ Ĝ Υ1/2 sd + ηd k (29)
the useful received signal, the second term contains the interfe-
√ ∗
T ∗ −1
= υk sdk + g̃ Tk Ĝ Ĝ Ĝ Υ1/2 sd + ηd k rence terms due to the use of imperfect CSI, and the third term
includes both the quantification and thermal noise samples.
where ηd k = g Tk q d + ndk . The first term denotes the useful The detected sample corresponding to the symbol transmitted
received signal, the second term contains the interference by the kth MS can then be obtained as
p h i
terms due to the use of imperfect CSI (pilot contamination), 1/2
p 1/2
yu k = Pu ωk suk + Pu W BB u G̃Ω su + ηu k , (35)
and the third term encompass both the quantification and k
thermal noise samples. where [x]k denotes the kth entry of vector x.
III. ACHIEVABLE RATES

G. Uplink payload data transmission
Analysis techniques similar to those applied, for instance, in
In the UL, the vector of received signals at the output of [3], [11], [18], [41]–[43], are used in this section to derive DL
the LA RF chains (including the RF phase shifters) of the mth and UL achievable rates. In particular, the sum of the second
AP is given by and third terms on the right hand side (RHS) of (29), for
K the DL case, and (35), for the UL case, are treated as effective
p X √
r um = Pu g mk0 ωk0 suk0 + num noise. The additive terms constituting the effective noise are, in
0 (30) both DL and UL cases, mutually uncorrelated, and uncorrela-
p k =1
= Pu Gm Ω1/2 su + num , ted with sdk and suk , respectively. Therefore, both the desired
signal and the so-called effective noise are uncorrelated. Now,
where Pu is the maximum average UL transmit power avai- recalling the fact that uncorrelated Gaussian noise represents
lable at any of the active MSs, su = [su1 . . . suK ]T denotes the worst case, from a capacity point of view, and that the
the vector of symbols transmitted by the K active MS, complex-valued fast fading random variables characterizing
Ω = diag([ω1 . . . ωK ]), with 0 ≤ ωk ≤ 1, is a matrix the propagation channels between different pairs of AP-MS
containing the power control coefficients used at the MSs, connections are independent, the DL and UL achievable rates
and num ∼ CN (0, σu2 (N )I LA ) is the vector of additive (measured in bits per second per Hertz) for MS k can be
thermal noise samples at the output of the LA RF chains of obtained as stated in the following theorems:
the mth AP. The received vector of signals at each of the
Theorem 1 (Downlink achievable rate). An achievable
APs in the network is quantized and forwarded to the CPU
rate of MS k using the analog precoders W RF m , for all
via the UL fronthaul links, where they are unquantized and
m ∈ {1, . . . , M }, and the ZF baseband precoder W BB d =
jointly processed using a set of baseband combining vectors. ∗
T ∗ −1
Using a similar approach to that employed to model the DL Ĝ Ĝ Ĝ is Rdk = log2 (1 + SINRdk ), with
transmission, the received vector of (unquantized) samples υk
from the mth AP can be expressed as4 SINRdk = PK , (36)
k0 =1 υk0 $kk0 + ση2dk
z um = Q̂um (r um ) = r um + q um , (31) where
M
X
where q um is the quantization noise vector, which is assumed ση2dk = σq2 dm tr RRF 2

mk + σd , (37)
to be statistically distributed as q um ∼ CN 0, σq2um I LA . m=1
Now, assuming the use of ZF MIMO detection, the CPU uses and
the detection matrix h n H ∗ T
oi
H −1 H $kk0 = diag E W BB
d g̃ k g̃ k W BB
d . (38)
T 0 k
W BBu = Ĝ Ĝ Ĝ = W BB d (32)
or, equivalently Proof. See Appendix A.

−1 H
H Theorem 2 (Uplink achievable rate). An achievable UL rate
T
W BB
um = Ĝ Ĝ Ĝm = W BB
dm , ∀m, (33)
for the kth MS in the Cell-Free Massive MIMO system with
limited capacity fronthaul links and using ZF MIMO detection,
4 Note that the schematic block diagram describing the UL payload data
for any M , N and K, is given by Ruk = log2 (1 + SINRuk ),
transmission phase is similar to the one shown in Fig. 3, however, the
processing flux is reversed, the power allocation is performed at the MSs, with
the quantization and unquantization tasks are performed at the APs and CPU,
Pu ωk
SINRuk = PK , (39)
respectively, and the fronthaul capacity is equal to CF u . Pu k0 =1 ωk0 δkk0 + ση2uk
where entropy of x̂ conditioned on x. Since the determinant is a log-

h n H H BB
oi concave function on the set of positive semidefinite matrices,
δkk0 = diag E G̃ wBB
uk w uk G̃ (40) it follows from Jensen’s inequality that
k0
!
BB
with wBB
uk denoting the kth row of W u , or, equivalently, 1 n
BB BB H
o
Ĉdm ≤ log2 det E W d m ΥW d m + I LA
h n
H BB H
oi σq2 dm
δkk0 = diag E W BB g̃
u k k0 g̃ 0 W u , (41) !
k K
1 X n H
o
and = log2 det υk E wBB BB
mk w mk + I LA .
M
σq2 dm
X k=1
ση2uk = σq2 um + σu2 (N ) νumk , (42) (46)
m=1 Analogously, the UL quantization process performed at the
with h n oi mth AP is given by Q̂um (r um ) = r um + q um . Thus, using
BB H
νumk = diag E W BB
u mW u m . (43) arguments similar to those used in the DL case, the required
k average rate to transfer the quantized vector Q̂um (r um ) on
the corresponding UL fronthaul link can be upper bounded as
Proof. See Appendix B. (in bps/Hz)
n o
Note that [18], as well as this paper, are based on the use Ĉum = E I Q̂um (r um ) ; r um
of ZF precoding, consequently, the corresponding signal-to-
n o n o
= E H Q̂um (r um ) − E H Q̂um (r um ) r um
interference-plus-noise ratio (SINR) equations must unavoi- ! !
dably bear some resemblance. Nevertheless, as it has been K
Pu X RF σu2 (N )
stressed in the text, the conceptual backgrounds used to obtain ≤ log2 det ωk Rmk + + 1 I LA .
σq2 um σq2 um
these analytical expressions are markedly different in the sense k=1
(47)
that our approach incorporates the influence of the mmWave-
based RF signal processing and constraints, while taking into
V. P ILOT ASSIGNMENT
account the quantization/unquantization effects associated to
the use of capacity-constrained fronthaul links. To warrant an appropriate system performance, the radio
resource management (RRM) unit must efficiently manage
both the pilot assignment and the UL and DL power control.
IV. F RONTHAUL CAPACITY CONSUMPTION
As the pilots are not power controlled, pilot assignment and
The DL quantization process performed at the mth AP can power control can be conducted independently. Since the
be expressed as length of the pilot sequences is limited to τp , there only exist

1/2

1/2
τp orthogonal pilot sequences. In a network with K ≤ τp
Q̂dm W BBd mΥ sd = W BB
d mΥ sd + q dm . (44) MSs, an optimal pilot assignment strategy simply allocates K
orthogonal pilots to the K MSs. The real pilot assignment
From standard random coding arguments [44], vector sd
problem arises when K > τp . In this case, fully orthogonal
can be safely assumed to be distributed as sd ∼ CN (0, I K)
1/2
pilot assignment is no longer possible and hence, other pilot
and thus, the quantized vector Q̂dm W BB d mΥ sd assignment strategies must be devised.

1/2
is distributed as Q̂dm W BB d mΥ sd ∼ On the one hand, designing an optimal pilot assignment

BB BB H 2
strategy aiming at maximizing the minimum rate allocated
CN 0, W d m ΥW d m + σq dm I LA . Furthermore, as to the active MSs in the network is a very difficult com-
the differential entropy of a vector x ∼ CN (ω, Θ) is given binatorial problem, computationally unmanageable in most
by H(x) = log det(πeΘ) [44], the required average rate network setups of practical interest [11]. On the other hand,
1/2
to transfer the quantized vector Q̂dm W BB dm Υ sd on using straightforward strategies such as, for instance, the pure
the corresponding DL fronthaul link can be obtained as (in random pilot assignment (RPA) scheme [45], where each
bps/Hz) MS is randomly assigned one pilot sequence out of the set
n o of τp orthogonal pilot sequences, or the balanced random
1/2 1/2
Ĉdm = E I Q̂dm W BB d mΥ sd ; W BB d mΥ sd pilot assignment (BRPA) scheme, where each MS is allocated
n
1/2
o a pilot sequence that is sequentially and cyclically selected
= E H Q̂dm W BB dm Υ sd from the ordered set of available orthogonal pilots, provides
poor performance results. In order to avoid the computational
n o
1/2 1/2
− E H Q̂dm W BB d mΥ sd W BB d mΥ sd
( !) complexity of the optimal strategies while improving the
1 BB BB H
performance of the baseline RPA or BRPA approaches, a
= E log2 det W ΥW d m + I LA , suboptimal solution is proposed in this paper that is based
σq2 dm d m
(45) on the idea of clustering by dissimilarity. This suboptimal
approach, that will be termed as the dissimilarity cluster-
where I(x̂; x) is used to denote the mutual information based pilot assignment (DCPA) strategy, is motivated by the
between vectors x̂ and x, and H(x̂|x) is the differential following key observation:
Key observation: In those scenarios where K > τp , MSs having similar large-scale propagation fingerprints are
cell-free communication is severely impaired whenever MSs allocated different pilot codes and, thus, they do not interfere to
showing very similar large-scale propagation patterns to the each other during the UL channel estimation process. In other
set of APs (that is, MSs typically located nearby) are allocated words, it aims at minimizing the residual interuser interference
the same pilot sequence. In this case, the inter-MS interference terms in both (29) and (35).
leads to very poor channel estimates at all APs and, eventually,
to low SINRs. VI. M AX - MIN POWER ALLOCATION AND OPTIMAL
The clustering algorithm proposed in this work basically QUANTIZATION
ensures that pilot sequences are only reused by MSs showing A. Downlink power control and quantization
dissimilar large-scale propagation patterns to the APs (that In line with previous research works on cell-free architectu-
is, MSs typically located sufficiently apart). Two key aspects res [10], [11], [18], [21], our aim in this subsection is to find
regarding the clustering operation are thus, on the one hand, the power control coefficients υk , for all k ∈ {1, . . . , K},
to decide which should be the large-scale propagation pattern and the quantization noise variances σq2 dm , for all m ∈
that ought to be used to represent a given MS and, on the other {1, . . . , M }, that maximize the minimum of the achievable
hand, to decide what metric should be used to measure simila- DL rates of all MSs while satisfying the average transmit
rity among the large-scale propagation patterns characterizing power and DL fronthaul capacity constraints at each AP. As
different MSs. To this end, and resting upon the premise that the logarithmic function is a monotonic increasing function
the CPU has perfect knowledge of the large-scale gains, let of its argument, maximizing the minimum achievable DL rate
T
ξ k = [ξ1k . . . ξM k ] denote the M × 1 vector containing the is equivalent to maximizing the minimum achievable SINR
average Frobenius norms of the equivalent channels linking [3], [10], [11], [18] thus, mathematically, this optimization
the kth MS to all M APs in the cell-free network. Vector ξ k problem can be formulated as
can be considered as an effective fingerprint characterizing the υk
location of MS k. Now, although no single definition of a simi- max min PK
Υ0 k∈{1,...,K} 2
larity measure exists, the so-called cosine similarity measure σ 0
qd
k0 =1 k $kk0 + σηdk
υ 0
is one of the most commonly used similarity metrics when K

X
dealing with real-valued vectors. Hence, as the fingerprint s.t. BB/RF
υk θmk ≤ P m − σq2 dm LA N, ∀ m, (51)
vectors characterizing the different MSs are non-negative real- k=1
valued, the cosine similarity measure between two fingerprint K
!
X υk
vectors ξ k and ξ k0 , defined as log2 det RBB + I LA ≤ CF d , ∀ m,
σq2 dm mk
k=1
ξ Tk ξ k0
fD (ξ k , ξ k0 ) = , (48) where we have used the definition σ q d = [σq d1 . . . σq dM ]T .
kξ k k2 kξ k0 k2
Optimization problem (51) is characterized by continuous
will be used as a proper similarity metric in our work. The objective and constraint functions of interdependent block
resulting similarity values range from 0, meaning orthogona- variables, namely, Υ and σ q d . A widely used approach for
lity (perfect dissimilarity), to 1, meaning exact match (perfect solving optimization problems of this class is the so-called
similarity). block coordinate descend (BCD) method [46], [47]. This is an
The proposed DCPA algorithm proceeds as follows. In a first iterative optimization approach that, at each iteration and in a
step, it calculates the fingerprint of an imaginary MS centroid, cyclic order, optimizes one of the blocks while the remaining
defined as variables are held fixed. As stated by the authors of [46], [47],
K
1 X convergence of the BCD method is ensured whenever each
ξC = ξk . (49)
K of the subproblems to be optimized in every iteration can be
k=1
exactly solved to its unique optimal solution. In the following
Then, it moves onward to the calculation of the cosine simila-
we show that each of the subproblems into which (51) is
rity measures among the fingerprint vectors characterizing the
decomposed can be solved to its unique optimal solution
K MSs in the network and the fingerprint of the centroid, that
and, thus, both a linear rate of convergence and optimal
is, the algorithm proceeds to the calculation of fD (ξ k , ξ C ), for
performance of the BCD approach are ensured at an affordable
all k ∈ {1, . . . , K}. The MSs are then sorted in descending or-
complexity [46], [47].
der of similarity with the centroid, that is, the algorithm obtains
The first important fact to note is that, given a power
the ordered set of subindices O = {o1 , o2 , . . . , oK }, such
that allocation matrix Υ(i−1) obtained at the (i − 1)th iteration,
fD ξ o1 , ξ C ≤ fD ξ o2 , ξ C ≤ · · · ≤ fD ξ oK , ξ C . Once
and as the achievable user rates monotonically increase with
the MSs have been sorted, the algorithm constructs τp clusters
the capacity of the fronthaul links between the APs and
of MSs, namely K1 , . . . , Kτp , with
the CPU, the optimal solution for the acceptable fronthaul
Kt =O (t : τp : K) quantization noise in the ith iteration is achieved when the
(50) fronthaul capacity constraints are satisfied with equality, that
= ot , ot+τp , ot+2τp , . . . , ∀t ∈ {1, . . . , τp },
is, when
and all MSs in cluster Kt , which are located far from each K (i−1)
!
other, are allocated the same pilot code ϕt . Note that the appli- X υk BB
det Rmk + I LA = 2CF d , ∀ m. (52)
cation of this algorithm ensures that, as far as it is possible, two σ 2 (i)
k=1 q dm
(i) TABLE I: Summary of default simulation parameters

Note that σq2 dm cannot be expressed in a closed-form algebraic
expression as it only admits a solution in the form of a Parameters Value
transcendental function Carrier frequency: f0 28 GHz
(i)
K
Bandwidth: B 20 MHz
σq2 dm = Fd Υ(i−1) , RBB

mk k=1 , CFd (53) Side of the square coverage area: D 200 m
AP antenna height: hAP 15 m
that can be numerically solved by applying mathematical
MS antenna height: hM S 1.65 m
software tools to (52). Noise figure at the MS: N F M S 9 dB
(i)
Once the optimal block of variables σ q d have been obtai- Noise figure of the LNA at the AP: N F LN A 1.6 dB
ned, the optimization problem in (51) can be rewritten in terms Gain of the LNA at the AP: GLN A 22 dB
of the power allocation matrix Υ(i) as Attenuation of the phase splitters at the AP: LP S 3 dB
(i) Attenuation of the power combiner at the AP: LP C in 3 dB
υk Noise figure of the RF chain at the AP: N FRF 7 dB
max min K M
Υ(i) 0 k∈{1,...,K} X (i)
X (i)
Available average power at the AP: P m 200 mW
υk0 γkk0 + σq2 dm tr RRF 2
mk + σd Available average power at the MS: Pu = Pp 100 mW
k0 =1 m=1
Coherence interval length: τc 200 samples
K
X (i) (i) Training phase length: τp 15 samples
s.t. BB/RF
υk θmk ≤ P m − N LA σq2 dm , ∀ m.
k=1
(54) VII. N UMERICAL RESULTS
Note that this is a convergent quasi-linear optimization pro- In this section, simulation results are obtained in order to
blem that can be solved using conventional standard convex quantitatively study the performance of the proposed cell-free
optimization methods [11], [18]. mmWave massive MIMO network with constrained-capacity
fronthaul links. In particular, we demonstrate the impact of
B. Uplink power control and quantization using different pilot allocation strategies, the effects of modi-
fying the capacity of the fronthaul links and the RF infrastruc-
In this subsection we aim at finding the power control
ture at the APs, and the repercussion of changing the density
coefficients ωk , for all k ∈ {1, . . . , K}, and quantization
of APs per area unit. For simplicity of exposition, and without
noise variances σq2 um , for all m ∈ {1, . . . , M }, that maximize
loss of essential generality, a cell-free scenario is considered
the minimum of the achievable ulink rates of all MSs while
where the M APs and K MSs are uniformly distributed at
satisfying the power control coefficient constraints at each MS
random within a square coverage area of size D × D m2 .
and the UL fronthaul capacity constraints at each AP. This
As described in subsection II-A, a modified version of the
optimization problem can be formulated as
discrete-time narrowband clustered channel model proposed
Pu ωk by Akdeniz et al. in [14] is used in the performance evaluation.
max min PK
ω0 k∈{1,...,K}
σ q u 0
Pu k0 =1 ωk0 δkk0 + ση2uk The parameters necessary to implement this channel model
can be found in [14, Table I]. Furthermore, similar to what
s.t. 0 ≤ ωk ≤ 1, ∀ k, was done by Ngo et al. in [11], a shadow fading spatial
K
!
Pu X correlation model with two components is also considered
det ωk RRF
mk + ϑm I LA ≤ 2CF u , ∀ m, (see [11, eqs. (54) and (55)]) where the decorrelation distance
σq2 um
k=1 is set to ddecorr = 50 m and the parameter δ is set to 0.5.
(55)
Default parameters used to set-up the simulation scenarios
T under evaluation in the following subsections are summarized
where σ q u = [σq u1 . . . σq uM ] , and we have used the defini-
tion ϑm = 1+σu2 (N )/σq2 um . As for the DL case, problem (55) in Table I.
admits the use of the block coordinate descend (BCD) method
where, in each
iteration,
RF the nonconvex transcendental function A. Impact of the pilot allocation process
2 K
σq um = Fu Ω, Rmk k=1 , Pu , CF u is approximated by a Our aim in this subsection is to benchmark the performance
constant calculated using the power allocation vector obtained of the proposed large-scale CSI-aware DCPA strategy against
in the previous iteration of the algorithm. That is, in the ith both the pure RPA and the BRPA schemes. Accordingly, the
iteration of the UL optimal power allocation approach, the average max-min rate per user versus the number of active
algorithm solves the optimization problem MSs is presented in Fig. 4 for each of these pilot allocation
(i) strategies and for both the DL and the UL. All results have
Pu ωk
max min , been obtained assuming the default system parameters descri-
PK (i) 2(i) (56)
Ω(i) 0 k∈{1,...,K}
Pu k0 =1 ωk0 δkk0 + σηuk bed in Table I, the use of L = 8 RF chains fully connected to
s.t. 0 ≤ ωk ≤ 1, ∀ k, uniform linear antenna arrays with N = 64 antenna elements,
K
and fronthaul links with a capacity of CF d = CF u = 64
(i)
where σq2 um = Fu Ω(i−1) , RRF

mk k=1 , Pu , CF u . Note bit/s/Hz. The first important result to note from Fig. 4 is that
that, again, this is a convergent quasi-linear optimization the pure RPA scheme is clearly outperformed by both the
problem that can be solved using conventional convex opti- BRPA and the DCPA strategies irrespective of the of active
mization methods [11], [18]. MSs in the network. In fact, the RPA scheme cannot guarantee
Downlink Uplink Downlink Uplink

18 18 18 18
DCPA DCPA CF d = 256 bps/Hz CF u = 256 bps/Hz
BRPA BRPA CF d = 64 bps/Hz CF u = 64 bps/Hz
16 16 16 CF d = 32 bps/Hz 16 CF u = 32 bps/Hz
RPA RPA
Average max-min user rate (bit/s/Hz)

CF d = 16 bps/Hz CF u = 16 bps/Hz
14 14 14 14
12 12 12 12
10 10 10 10
8 8 8 8
6 6 6 6
4 4 4 4
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
Number of users (K) Number of users (K) Number of users (K) Number of users (K)
Fig. 4: Average max-min rate per user versus the number of active Fig. 5: Average max-min rate per user versus the number of active
MSs for different pilot allocation strategies (N = 64 antennas, L = 8 MSs for different values of the fronthaul capacities (N = 64
RF chains, CF d = CF u = 64 bit/s/Hz). antennas, L = 8 RF chains, DCPA).
neither the absence of pilot reuse, even for those cases in which deduced from the analysis of the SINR expressions in (36)
K ≤ τp (in this setup, τp = 15 time/frequency samples), nor and (39). As the number of active MSs in the cell-free
the possibility of having pilots that are allocated to a high network increases, provided that it is greater than τp , the
number of MSs and/or to MSs exhibiting very similar large- term in the denominator corresponding to the residual interuser
scale propagation patterns to the APs. Therefore, the higher the interference due to pilot contamination becomes increasingly
number of active MSs, the higher the probability of having one dominant in comparison to the quantification and thermal
or more users suffering from high levels of pilot contamina- noise terms, eventually reaching the point where they can be
tion, with the consequent reduction of the achievable max-min considered virtually negligible. Under these conditions, and
user rate. If we turn our attention to results provided by the since the pre-coding filters used on both links are identical, the
BRPA and DCPA strategies, two disjoint operation regions can DL and the UL experience similar SINR values and, therefore,
be distinguished. In the first one, comprising the scenarios in tend to provide the same achievable max-min rate per user,
which K ≤ τp , both approaches allocate orthogonal pilots to except for small differences that can be attributed to, on the
the users (absence of pilot contamination) and thus naturally one and, the dissimilar amount of quantified information that
provide the same performance. In the second one, however, has to be conveyed through the corresponding fronthaul links
comprising the scenarios in which K > τp , pilots have to and, on the other hand, disparities among the thermal noise
be reused and, as a consequence, pilot contamination appears powers experienced at both the APs and the MSs.
(note the rather abrupt performance drop when going from
K ≤ τp to K > τp ). In these scenarios, based on a smart B. Modifying the capacity of the fronthaul links and the RF
exploitation of the available large-scale CSI, the proposed infrastructure at the APs
DCPA approach reduces the amount of pilot contamination The max-min achievable rate per user is plotted in Fig. 5
experienced by the worst users in the network and it clearly against the number of active MSs in the network, assuming
improves the achievable max-min user rates provided by the the use of fronthaul links with different constraining capacities
channel-unaware BRPA scheme. Note that, irrespective of the equal to 16, 32, 64 and 256 bit/s/Hz (for the network setups
scenario under evaluation, increasing the number of active under consideration, using fronthaul links with a capacity of
MSs in the system translates into a per-user performance drop, 256 bit/s/Hz is virtually equivalent to using infinite-capacity
despite the fact that the global network performance increases fronthauls). As expected, results show that increasing the
due to the exploitation of the well-known multiuser diversity fronthaul capacity is always beneficial if the main aim is to
effects. increase the achievable max-min user rate. Nevertheless, it is
Another result that is worth emphasizing, since it will re- worth stressing that, keeping all the other parameters constant,
peatedly appear in the following subsections, is that, although the marginal increment of performance produced by each new
in scenarios with high-capacity fronthaul links the achievable increment of the fronthaul capacity suffers from the law of
max-min DL user rate is higher than that provided in the diminishing returns, especially for network setups with a high
UL, as the number of active users in the network increases, number of active MSs. That is, although the performance
the performance obtained in both the DL and the UL tend increase produced by doubling the fronthaul capacity from 16
to become increasingly similar. This behavior can be easily bit/s/Hz to 32 bit/s/Hz, or even from 32 bit/s/Hz to 64 bit/s/Hz,
Downlink Uplink To understand how the RF infrastructure used at the APs

18 18 influences the performance of the proposed cell-free mmWave
N = 128 N = 128
N = 64 N = 64 massive MIMO system under constrained-capacity fronthaul
16 N = 32 16 N = 32
links, Figs. 6 and 7 show the achievable max-min user rate

N = 16 N = 16
N =8 N =8 against the number of active MSs assuming the use of uniform
14 14
linear antenna arrays with different number of elements and
fully-connected analog RF precoders with different number of
12 12
RF chains, respectively. In particular, results presented in Fig.
6 have been obtained assuming the use of an analog precoder
10 10
with L = 8 RF chains fully-connected to a linear uniform
antenna array with N = 8, 16, 32, 64 or 128 antenna elements,
8 8
whereas results presented in Fig. 7 have been obtained assu-
6 6
ming the use of L = 2, 4, 8 or 16 RF chains fully-connected to
a linear uniform antenna array with N = 64 antenna elements.
4 4
The first conclusion we may draw when looking at the results
0 10 20 30 40 50 0 10 20 30 40 50 presented in Fig. 6 is that, irrespective of the number of active
Number of users (K) Number of users (K)
MSs in the cell-free network, increasing the number of antenna
elements at the APs in scenarios with high capacity fronthaul
Fig. 6: Average max-min rate per user versus the number of active links (CF d = CF u = 64 bit/s/Hz), although moderate and
MSs for different values of the number of antennas at the APs (L = 8 subject to the law of diminishing returns, always produces an
RF chains, CF d = CF u = 64 bit/s/Hz, DCPA). increase in the achievable max-min user rate. As shown in
Fig. 7, in contrast, the impact produced by an increase in the
Downlink Uplink
number of RF chains at the APs depends on the number of
18 18 active MSs in the network. In particular, when the number
L=2 L=2
L=4 L=4
of active users is high, the interuser interference term due to
16 L=8 16 L=8 pilot contamination (imperfect CSI) dominates the factors in
L = 16 L = 16
the denominator of the SINR (i.e., makes the quantization and
14 14 thermal noises negligible) and thus, increasing the number of
RF chains is always beneficial when trying to increase the
12 12
achievable max-min user rate. When the number of active
users in the network is low, however, the quantization noise,
10 10
which is an increasing function of L, is not negligible anymore
when compared to the interuser interference term (recall that
8 8
this term is null when the number of active MSs is less than
or equal to τp ) and thus, increasing the number of RF chains
6 6
at the APs can be clearly disadvantageous.
Results presented in Figs. 5, 6 and 7 were obtained assuming
4 4
0 10 20 30 40 50 0 10 20 30 40 50 high-capacity fronthaul links with CF d = CF u = 64 bit/s/Hz.
Number of users (K) Number of users (K)
However, the amount of quantized data that has to be conveyed
from (to) the CPU to (from) the APs in the DL (UL) depends
Fig. 7: Average max-min rate per user versus the number of active on the number of antennas and RF chains at the APs (see
MSs for different values of the number of RF chains at the APs Section IV). Thus, in order to deepen in the study of the impact
(N = 64 antennas, CF d = CF u = 64 bit/s/Hz, DCPA). the RF infrastructure may have on the achievable performance
of the proposed cell-free mmWave massive MIMO system
under constrained-capacity fronthaul links, the average max-
can be justifiable, increasing the fronthaul capacity beyond 64 min user rate is plotted in Figs. 8 and 9 against the number
bit/s/Hz does not seem to be reasonable from the point of view of antenna elements and RF chains, respectively, for different
of increasing the achievable performance of the system under values of the fronthaul capacities and assuming a fixed number
the considered network setups. As observed in the previous of K = 20 active MSs in the network. In network setups using
subsection, in cell-free mmWave massive MIMO networks very high capacity fronthaul links (i.e., CF d = CF u = 256
using high-capacity fronthaul links, the achievable max-min bit/s/Hz), increasing the number of antenna elements N and/or
DL user rate is always slightly higher than that achieved in the number of RF chains L (up to L = K) is always beneficial
the UL irrespective of the number of active MSs. In scenarios as, in this case, the noise introduced by the quantization
with low-capacity fronthaul links and a large number of active process is negligible and the system can take full advantage of
MSs, however, the quantization noise experienced in the DL the increased RF resources. As the capacity of the fronthaul
is higher than its UL counterpart and thus, the achievable per- links decreases, however, the amount of noise introduced by
user rate in the UL is slightly higher that than supplied in the the quantization process increases with both N and L and,
DL. therefore, a situation arises where the potential performance
Downlink Uplink L = 10, 4, and 1 when using fronthaul links with a capacity of
10 10 64 bit/s/Hz, 32 bit/s/Hz and less than 24 bit/s/Hz, respectively.
CF d = 256 bit/s/Hz
CF u = 256 bit/s/Hz Using a number of RF chains beyond this optimal value leads
9.5 9.5
to a clear performance degradation since the potential benefits

CF d = 64 bit/s/Hz
CF u = 64 bit/s/Hz
9 9 of having extra hardware resources does not compensate for
CF u = 32 bit/s/Hz
the effects produced by the increase in quantization noise due
CF d = 32 bit/s/Hz
8.5 8.5 to fronthaul-capacity constraints.
8 8
CF u = 24 bit/s/Hz
CF d = 24 bit/s/Hz
C. Impact of the density of APs
7.5 7.5
With the aim of evaluating the impact the density of APs
7 7
per area unit may have on the performance of the proposed
CF u = 16 bit/s/Hz
cell-free mmWave massive MIMO system, Fig. 10 represents
6.5 CF d = 16 bit/s/Hz 6.5 the cumulative distribution function (CDF) of the DL and
UL achievable max-min user rate for different values of the
6 6
0 50 100 0 50 100 number of APs in the network. It has been assumed in these
Number of antennas at the AP (N) Number of antennas at the AP (N) experiments a fixed number of active MSs equal to either
K = 25 or K = 8 MSs, the use of L = 8 RF chains fully-
connected to a linear uniform antenna array with N = 64
Fig. 8: Average max-min rate per user versus the number of antennas
at the APs for different values of the fronthaul capacities (K = 20 antenna elements, and the use of DL and UL fronthaul links
users, L = 8 RF chains, DCPA). with a capacity CF d = CF u = 64 bit/s/Hz. As expected,
cell-free massive MIMO scenarios with a high density of
APs per area unit significantly outperform those with a low
Downlink Uplink density of APs per area unit in both median and 95%-likely
10 10
CF d = 256 bit/s/Hz achievable per-user rate performance. However, the achievable
CF u = 256 bit/s/Hz
9.5 9.5 max-min user rate increase due to increasing the number of
9
CF d = 64 bit/s/Hz
9
CF u = 64 bit/s/Hz APs in the network is, again, subject to the law of diminishing
CF d = 32 bit/s/Hz
returns. For instance, in scenarios with K = 25 MSs, the
8.5 8.5 CF u = 32 bit/s/Hz 95%-likely achievable user rate is equal to 2.55, 4.33, 6.11
8 8
and 6.50 bit/s/Hz for cell-frre massive MIMO networks with
M = 25, 50, 100 and 200 APs, respectively. That is, doubling
CF u = 24 bit/s/Hz
7.5 CF d = 24 bit/s/Hz 7.5 the number of APs per area unit does not result in doubling
7 7
the 95%-likely achievable user rate. Similar conclusions can
CF d = 16 bit/s/Hz be drawn when looking at either the median or the average
CF u = 16 bit/s/Hz
6.5 6.5 achievable user rates.
6 6
As was observed in results presented in previous subsections
for high-capacity fronthaul setups, when the number of active
5.5
5 10 15 20
5.5
5 10 15 20
users in the system is low, the achievable max-min rate values
Number of RF chains at the AP (L) Number of RF chains at the AP (L) in the DL are slightly higher than those achievable in the
UL. Instead, when the number of active users increases, the
achievable max-min user rates are virtually identical in both
Fig. 9: Average max-min rate per user versus the number of RF the DL and the UL. Also, note that the dispersion of the
chains at the APs for different values of the fronthaul capacities (K =
20 users, N = 64 antennas, DCPA). achievable max-min user rates around the median tends to
diminish as the density of APs increases. That is, cell-free
massive MIMO networks with a high density of APs per
improvement provided by the increase of N and/or L is area unit tend to offer max-min achievable rates that suffer
compromised by the performance reduction due to fronthaul little variations irrespective of the location of the APs (i.e,
capacity constraints. On the one hand, it can be observed in irrespective of the scenario under evaluation).
Fig. 8 that, for fixed numbers of users and RF chains, there is a
certain fronthaul capacity constraint value (near 24 bit/s/Hz in VIII. C ONCLUSION
the setup used in this experiment) under which increasing the A novel analytical framework for the performance analysis
number of antenna elements at the array is counterproductive. of cell-free mmWave massive MIMO networks has been intro-
On the other hand, results presented in Fig. 9 show that, for duced in this paper. The proposed framework considers the use
fixed numbers of users and antenna elements at the arrays, of low-complexity hybrid precoders/decoders where the RF
there is always an optimal number of RF chains to be deployed high-dimensionality phase shifter-based precoding/decoding
(or activated) at the APs that is dependent on the capacity of stage is based on large-scale second-order channel statistics,
the fronthaul links. In particular, for the network setups under while the low-dimensionality baseband multiuser MIMO pre-
consideration, the optimal number of RF chains is equal to coding/decoding stage can be easily implemented by standard
Downlink Uplink
1 1
M = 25 APs M = 25 APs
M = 50 APs M = 50 APs
Cumulative distribution function
Cumulative distribution function

M = 100 APs M = 100 APs
0.8 M = 200 APs
0.8 M = 200 APs
K = 25 MSs K = 25 MSs
0.6 0.6
K = 8 MSs K = 8 MSs
0.4 0.4
0.2 0.2
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
Max-min user rate (bit/s/Hz) Max-min user rate (bit/s/Hz)
Fig. 10: CDF of the DL and UL achievable max-min rate per user for different values of the number of APs and active MSs in the cell-free
network (N = 64 antennas, L = 8 RF chains, CF d = CF u = 64 bit/s/Hz, DCPA).
ZF signal processing schemes using small-scale estimated CSI. future work, it would be interesting to develop low-complexity
Furthermore, it also takes into account the impact of using pilot- and power-allocation techniques specifically designed
capacity-constrained fronthaul links that assume the use of to maximize the energy efficiency of cell-free mmWave mas-
large-block lattice quantization codes able to approximate a sive MIMO networks considering both the fronthaul capacity
Gaussian quantization noise distribution, which constitutes an constraints and the fronthaul power consumption. It would
upper bound to the performance attained under any practical also be interesting to explore the use of partially-connected
quantization scheme. Max-min power allocation and fronthaul RF precoding/decoding architectures, the implementation of
quantization optimization problems have been posed thanks to baseband MU-MIMO precoding/decoding other than the ZF
the development of mathematically tractable expressions for scheme, the development of new user selection algorithms, and
both the per-user achievable rates and the fronthaul capacity the investigation of the effects a non-uniform distribution of
consumption. These optimization problems have been solved MSs and/or APs may have on the performance of the proposed
by combining the use of block coordinate descent methods system.
with sequential linear optimization programs. Results have
shown that the proposed DCPA suboptimal pilot allocation A PPENDIX A
strategy, which is based on the idea of clustering by dissi- P ROOF OF T HEOREM 1
milarity, overcomes the computational burden of the optimal Following an approach similar to that proposed by Nayebi
small-scale CSI-based pilot allocation scheme while clearly et al. in [18], the signal received by the kth MS in (29) can
outperforming the pure random and balanced random schemes. be rewritten as yd k = yd k 0 + yd k 1 + yd k 2 + ndk , where the
It has also been shown that, although increasing the fronthaul useful, interuser interference, and quantization noise terms can
√ 1/2
capacity and/or the density of APs per area unit is always be expressed as yd k 0 = υk sdk , yd k 1 = g̃ Tk W BB d Υ sd ,
T
P M T
beneficial from the point of view of the achievable max-min and yd k 2 = g k q d = m=1 g km q dm , respectively. Now, con-
user rate, the marginal increment of performance produced by sidering that data symbols, quantization noise, thermal noise,
each new increment of these parameters suffers from the law of and channel-related coefficients are mutually independent, the
diminishing returns, especially for network setups with a high terms yd k 0 , yd k 1 , yd k 2 and ndk are mutually uncorrelated
number of active MSs. Moreover, simulation results indicate and thus, based on the worst-case uncorrelated additive noise
that, as the capacity of the fronthaul links decreases, the [41], the achievable DL rate for user k is lower bounded by
potential performance improvement provided by the increase Rdk = log2 (1 + SINRdk ), with
of the number of antenna elements N and/or the number of n
2
o
RF chains L is compromised by the performance reduction E |yd k 0 |
SINRdk = n o n o ,
due to the corresponding increase of the fronthaul quantization 2 2
E |yd k 1 | + E |yd k 2 | + σd2
noise. In particular, for fixed numbers of users and RF chains, n o
there is a certain fronthaul capacity constraint value (near 2
where E |yd k 0 | = υk ,
24 bit/s/Hz in the setups under consideration) under which n o n o
increasing the number of antenna elements at the array is 2 1/2 BB H ∗ T BB 1/2
E |yd k 1 | = E sH d Υ W d g̃ g̃
k k W d Υ sd
counterproductive. Similarly, for fixed numbers of users and n o
BB H ∗ T BB
antenna elements at the arrays, there is always an optimal = tr ΥE W d g̃ k g̃ k W d
number of RF chains to be deployed (or activated) at the APs K h n oi
that is dependent on the capacity of the fronthaul links. For
X H ∗ T
= υk0 diag E W BB
d g̃ g̃
k k W BB
d 0
,
k
k0 =1
and R EFERENCES
M
n
2
o X
∗
[1] M. Shafi, A. F. Molisch, P. J. Smith, T. Haustein, P. Zhu, P. D. Silva,
E qd H T

E |yd k 2 | = m g km g km q dm F. Tufvesson, A. Benjebbour, and G. Wunder, “5G: A tutorial overview
m=1 of standards, trials, challenges, deployment, and practice,” IEEE Journal
M on Selected Areas in Communications, vol. 35, no. 6, pp. 1201–1221,
X June 2017.
σq2 dm tr RRF

= mk . [2] K. David and H. Berndt, “6G vision and requirements: Is there any need
m=1 for beyond 5G?” IEEE Vehicular Technology Magazine, vol. 13, no. 3,
pp. 72–80, Sept 2018.
[3] T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo, Fundamentals
A PPENDIX B of massive MIMO. Cambridge University Press, 2016.
P ROOF OF T HEOREM 2 [4] E. Björnson, E. G. Larsson, and T. L. Marzetta, “Massive mimo: ten
myths and one critical question,” IEEE Communications Magazine,
The detected signal at the CPU corresponding to the vol. 54, no. 2, pp. 114–123, February 2016.
symbol transmitted by the kth MS in (35) can be rewrit- [5] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-
bers of base station antennas,” IEEE Transactions on Wireless Commu-
ten as yu k = yu k 0 + yu k 1 + yu k 2 + yu k 3 , where the nications, vol. 9, no. 11, pp. 3590–3600, November 2010.
useful, interuser interference, quantization noise√and thermal [6] M. K. Karakayali, G. J. Foschini, and R. A. Valenzuela, “Network
√ coordination for spectrally efficient communications in cellular systems,”
noise terms can be expressedi as yu k 0 = Pu ωk suk ,
√ h BB IEEE Wireless Communications, vol. 13, no. 4, pp. 56–61, Aug 2006.
yu k 1 = Pu W u G̃Ω1/2 su , yu k 2 = W BB

u q u k , and [7] D. Gesbert, S. Hanly, H. Huang, S. S. Shitz, O. Simeone, and W. Yu,
k
“Multi-cell MIMO cooperative networks: A new look at interference,”
yu k 3 = W BB

u nu k , respectively. As in the DL, since IEEE Journal on Selected Areas in Communications, vol. 28, no. 9, pp.
data symbols, quantization noise, thermal noise, and channel- 1380–1408, December 2010.
related coefficients are mutually independent, the terms yu k 0 , [8] R. Irmer, H. Droste, P. Marsch, M. Grieger, G. Fettweis, S. Brueck,
H. Mayer, L. Thiele, and V. Jungnickel, “Coordinated multipoint:
yu k 1 , yd k 2 and yd k 3 are mutually uncorrelated and thus, Concepts, performance, and field trial results,” IEEE Communications
based on the worst-case uncorrelated additive noise [41], Magazine, vol. 49, no. 2, pp. 102–111, February 2011.
the achievable UL rate for user k is lower bounded by [9] A. Checko, H. L. Christiansen, Y. Yan, L. Scolari, G. Kardaras, M. S.
Berger, and L. Dittmann, “Cloud RAN for mobile networks - A
Ruk = log2 (1 + SINRuk ), with technology overview,” IEEE Communications Surveys Tutorials, vol. 17,
n
2
o no. 1, pp. 405–426, Firstquarter 2015.
E |yd k 0 | [10] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta,
SINRuk = n o n o n o, “Cell-free massive MIMO: Uniformly great service for everyone,” in
2 2 2 IEEE SPAWC 2015.
E |yd k 1 | + E |yd k 2 | + E |yd k 3 |
[11] ——, “Cell-free massive MIMO versus small cells,” IEEE Transactions
n o on Wireless Communications, vol. 16, no. 3, pp. 1834–1850, 2017.
2
where E |yu k 0 | = Pu ωk , [12] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N.
Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave
mobile communications for 5G cellular: It will work!” IEEE Access,
1/2 H BB H BB
n o n o
2 1/2
E |yu k 1 | = Pu E sH
u Ω G̃ w uk w uk G̃Ω su vol. 1, pp. 335–349, May 2013.
n H o [13] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski,
BB H BB “Five disruptive technology directions for 5G,” IEEE Communications
= Pu tr ΩE G̃ wuk wuk G̃
Magazine, vol. 52, no. 2, pp. 74–80, February 2014.
K h n H oi [14] M. R. Akdeniz, Y. Liu, M. K. Samimi, S. Sun, S. Rangan, T. S.
X H BB Rappaport, and E. Erkip, “Millimeter wave channel modeling and
= Pu ωk0 diag E G̃ wBB
uk w uk G̃ 0
, cellular capacity evaluation,” IEEE Journal on Selected Areas in Com-
k
k0 =1 munications, vol. 32, no. 6, pp. 1164–1179, 2014.
BB [15] T. S. Rappaport, Y. Xing, G. R. MacCartney, A. F. Molisch, E. Mellios,
with wBB
uk denoting the kth row of W u , or, equivalently, and J. Zhang, “Overview of millimeter wave communications for fifth-
generation (5G) wireless networks - with a focus on propagation
H
n o h n oi
2 BB H
E |yu k 1 | = Pu diag E W BB u G̃ΩG̃ W u
models,” IEEE Transactions on Antennas and Propagation, vol. 65,
k no. 12, pp. 6213–6230, Dec 2017.
K h n oi [16] X. Gao, L. Dai, and A. M. Sayeed, “Low RF-complexity technologies to
BB H enable millimeter-wave MIMO with large antenna array for 5G wireless
X
H
= Pu ωk0 diag E W BB
u g̃ k0 g̃ k0 W u , communications,” IEEE Communications Magazine, vol. 56, no. 4, pp.
k
k0 =1 211–217, April 2018.
[17] S. A. Busari, K. M. S. Huq, S. Mumtaz, L. Dai, and J. Rodriguez,
and, finally, “Millimeter-wave massive MIMO communication for future wireless
n o h n oi systems: A survey,” IEEE Communications Surveys Tutorials, vol. 20,
2 BB H
E |yu k 2 | = diag E W BB H
u qu qu W u
no. 2, pp. 836–869, Second quarter 2018.
k [18] E. Nayebi, A. Ashikhmin, T. L. Marzetta, H. Yang, and B. D. Rao,
M h n oi “Precoding and power optimization in cell-free massive MIMO sys-
BB H
X
= diag E W BB H
u q um q um W u
tems,” IEEE Transactions on Wireless Communications, vol. 16, no. 7,
k pp. 4445–4459, July 2017.
m=1
[19] L. D. Nguyen, T. Q. Duong, H. Q. Ngo, and K. Tourki, “Energy effi-
M h n oi ciency in cell-free massive MIMO with zero-forcing precoding design,”
BB H
X
= σq2 um diag E W BB
um W um , IEEE Communications Letters, vol. 21, no. 8, pp. 1871–1874, Aug 2017.
k [20] H. Q. Ngo, L. Tran, T. Q. Duong, M. Matthaiou, and E. G. Larsson,
m=1
“On the total energy efficiency of cell-free massive MIMO,” IEEE
and, analogously, Transactions on Green Communications and Networking, vol. 2, no. 1,
pp. 25–39, March 2018.
n o M h n oi [21] M. Bashar, K. Cumanan, A. G. Burr, H. Q. Ngo, and M. Deb-
BB H
2
X
E |yd k 3 | = σu2 (N ) diag E W BB
u mW u m . bah, “Cell-free massive MIMO with limited backhaul,” arXiv preprint
k arXiv:1801.10190, 2018.
m=1
[22] M. N. Boroujerdi, A. Abbasfar, and M. Ghanbari, “Cell free massive [44] T. M. Cover and J. A. Thomas, Elements of Information Theory (Wiley
MIMO with limited capacity fronthaul,” Wireless Personal Communica- Series in Telecommunications and Signal Processing), 2nd ed. Wiley-
tions, October 2018. Interscience, 2006.
[23] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, [45] H. Ahmadi, A. Farhang, N. Marchetti, and A. MacKenzie, “A game
“Spatially sparse precoding in millimeter wave MIMO systems,” IEEE theoretic approach for pilot contamination avoidance in massive MIMO,”
Transactions on Wireless Communications, vol. 13, no. 3, pp. 1499– IEEE Wireless Communications Letters, vol. 5, no. 1, pp. 12–15, 2016.
1513, March 2014. [46] P. Tseng, “Convergence of a block coordinate descent method for
[24] X. Gao, L. Dai, S. Han, C. I, and R. W. Heath, “Energy-efficient hybrid nondifferentiable minimization,” Journal of Optimization Theory and
analog and digital precoding for mmWave MIMO systems with large Applications, vol. 109, no. 3, pp. 475–494, June 2001.
antenna arrays,” IEEE Journal on Selected Areas in Communications, [47] A. Beck and L. Tetruashvili, “On the convergence of block coordinate
vol. 34, no. 4, pp. 998–1009, April 2016. descent type methods,” SIAM Journal on Optimization, vol. 23, no. 4,
[25] A. F. Molisch, V. V. Ratnam, S. Han, Z. Li, S. L. H. Nguyen, L. Li, and pp. 2037–2060, 2013.
K. Haneda, “Hybrid beamforming for massive MIMO: A survey,” IEEE
Communications Magazine, vol. 55, no. 9, pp. 134–141, Sept 2017.
[26] S. Park, A. Alkhateeb, and R. W. Heath, “Dynamic subarrays for hybrid
precoding in wideband mmWave MIMO systems,” IEEE Transactions
on Wireless Communications, vol. 16, no. 5, pp. 2907–2920, May 2017.
[27] M. Alonzo and S. Buzzi, “Cell-free and user-centric massive MIMO
at millimeter wave frequencies,” in IEEE International Symposium on
Personal, Indoor, and Mobile Radio Communications (PIMRC), Oct
2017, pp. 1–5.
[28] M. Alonzo, S. Buzzi, and A. Zappone, “Energy-efficient downlink
power control in mmwave cell-free and user-centric massive
MIMO,” CoRR, vol. abs/1805.05177, 2018. [Online]. Available:
http://arxiv.org/abs/1805.05177
[29] M. K. Samimi and T. S. Rappaport, “Ultra-wideband statistical channel
model for non line of sight millimeter-wave urban channels,” in 2014
IEEE Global Communications Conference, Dec 2014, pp. 3483–3489.
[30] A. Adhikary, E. A. Safadi, M. K. Samimi, R. Wang, G. Caire, T. S.
Rappaport, and A. F. Molisch, “Joint spatial division and multiplexing
for mm-Wave channels,” IEEE Journal on Selected Areas in Communi-
cations, vol. 32, no. 6, pp. 1239–1255, June 2014.
[31] R. Méndez-Rial, N. González-Prelcic, and R. W. Heath, “Adaptive
hybrid precoding and combining in MmWave multiuser MIMO systems
based on compressed covariance estimation,” in IEEE 6th International
Workshop on Computational Advances in Multi-Sensor Adaptive Pro-
cessing (CAMSAP), Dec 2015, pp. 213–216.
[32] S. Park and R. W. Heath, “Spatial channel covariance estimation
for mmWave hybrid MIMO architecture,” in Asilomar Conference on
Signals, Systems and Computers, Nov 2016, pp. 1424–1428.
[33] S. Park and R. W. Heath Jr, “Spatial channel covariance estimation for
the hybrid MIMO architecture: A compressive sensing based approach,”
arXiv preprint arXiv:1711.04207, 2017.
[34] S. Park, J. Park, A. Yazdan, and R. W. Heath, “Exploiting spatial channel
covariance for hybrid precoding in massive MIMO systems,” IEEE
Transactions on Signal Processing, vol. 65, no. 14, pp. 3818–3832, July
2017.
[35] R. Mai, T. Le-Ngoc, and D. H. N. Nguyen, “Two-timescale hybrid
RF-baseband precoding with MMSE-VP for multi-user massive MIMO
broadcast channels,” IEEE Transactions on Wireless Communications,
vol. 17, no. 7, pp. 4462–4476, July 2018.
[36] Z. Shen, R. Chen, J. G. Andrews, R. W. Heath, and B. L. Evans, “Low
complexity user selection algorithms for multiuser MIMO systems with
block diagonalization,” IEEE Transactions on Signal Processing, vol. 54,
no. 9, pp. 3658–3663, 2006.
[37] O. Elijah, C. Y. Leow, T. A. Rahman, S. Nunoo, and S. Z. Iliya, “A
comprehensive survey of pilot contamination in massive MIMO-5G
system,” IEEE Comm. Surv. & Tutor., vol. 18, no. 2, pp. 905–923, 2016.
[38] L. Lu, G. Y. Li, A. L. Swindlehurst, A. Ashikhmin, and R. Zhang, “An
overview of massive MIMO: Benefits and challenges,” IEEE Journal of
Selected Topics in Signal Processing, vol. 8, no. 5, pp. 742–758, 2014.
[39] R. Zamir and M. Feder, “On lattice quantization noise,” IEEE Transacti-
ons on Information Theory, vol. 42, no. 4, pp. 1152–1159, Jul 1996.
[40] G. Femenias and F. Riera-Palou, “Multi-layer downlink precoding for
cloud-RAN systems using full-dimensional massive MIMO,” IEEE
Access, vol. 6, pp. 61 583–61 599, 2018.
[41] B. Hassibi and B. M. Hochwald, “How much training is needed in
multiple-antenna wireless links?” IEEE Transactions on Information
Theory, vol. 49, no. 4, pp. 951–963, April 2003.
[42] H. Yang and T. L. Marzetta, “Capacity performance of multicell large-
scale antenna systems,” in 51st Annual Allerton Conference on Commu-
nication, Control, and Computing (Allerton), Oct 2013, pp. 668–675.
[43] G. Interdonato, H. Q. Ngo, E. G. Larsson, and P. Frenger, “On the perfor-
mance of cell-free massive MIMO with short-term power constraints,”
in IEEE 21st International Workshop on Computer Aided Modelling and
Design of Communication Links and Networks (CAMAD), Oct 2016, pp.
225–230.

Cell-Free Millimeter-Wave Massive MIMO Systems With Limited Fronthaul Capacity

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Cell-Free Millimeter-Wave Massive MIMO Systems With Limited Fronthaul Capacity

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

SUBMITTED TO IEEE ACCESS 1

Cell-Free Millimeter-Wave Massive MIMO Systems

D RIVEN by the continuously increasing demands for high

SUBMITTED TO IEEE ACCESS 2

SUBMITTED TO IEEE ACCESS 3

the calculation of the capacity consumption of both the DL and

SUBMITTED TO IEEE ACCESS 4

× a (θmk,cp , φmk,cp ) aH (θmk,cp , φmk,cp ) ,

SUBMITTED TO IEEE ACCESS 5

C. Large-scale training phase: selection of MSs to beamform D. Small-scale training phase

The average sum energy of the equivalent channels between and

SUBMITTED TO IEEE ACCESS 6

CPU AP 1 transmitted by the mth AP. Due to the distortion introduced by

SUBMITTED TO IEEE ACCESS 7

III. ACHIEVABLE RATES

or, equivalently Proof. See Appendix A.

SUBMITTED TO IEEE ACCESS 8

where entropy of x̂ conditioned on x. Since the determinant is a log-

SUBMITTED TO IEEE ACCESS 9

is one of the most commonly used similarity metrics when K

SUBMITTED TO IEEE ACCESS 10

(i) TABLE I: Summary of default simulation parameters

SUBMITTED TO IEEE ACCESS 11

Downlink Uplink Downlink Uplink

Average max-min user rate (bit/s/Hz)

Average max-min user rate (bit/s/Hz)

Average max-min user rate (bit/s/Hz)

SUBMITTED TO IEEE ACCESS 12

Downlink Uplink To understand how the RF infrastructure used at the APs

Average max-min user rate (bit/s/Hz)

Average max-min user rate (bit/s/Hz)

SUBMITTED TO IEEE ACCESS 13

Average max-min user rate (bit/s/Hz)

Average max-min user rate (bit/s/Hz)

SUBMITTED TO IEEE ACCESS 14

Cumulative distribution function

SUBMITTED TO IEEE ACCESS 15

SUBMITTED TO IEEE ACCESS 16

You might also like