Improving Cell-Free Massive MIMO Networks Performance A User Scheduling Approach

7360 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO.
11, NOVEMBER 2021
Improving Cell-Free Massive MIMO Networks

Performance: A User Scheduling Approach
Juwendo Denis , Member, IEEE, and Mohamad Assaad , Senior Member, IEEE
Abstract— Cell-Free (CF) massive multiple-input multiple- should be able to accommodate the traffic growth, enable new
output (MIMO) system is a distributed antenna system, wherein use cases, provide seamless connectivity and ensure that users
a large number of access points wish to simultaneously com- demands together with specific performance requirements are
municate with a relatively small number of users. Similar to
co-located massive MIMO system, pilot contamination and multi- satisfied [3]. Massive multiple-input multiple-output (MIMO)
user interference are two major impediments to CF massive emerges as an appealing technology for 5G physical-layer [4],
MIMO network performance improvement. One can mitigate [5] that enables to increase the network spectral efficiency [6].
the detrimental effect of pilot contamination and multi-user Massive MIMO, which is often referred to as co-located
interference through judicious resource allocation. In this work, massive MIMO, consists of deploying a combination of large-
we formulate and investigate the problem of frequency assign-
ment for a CF massive MIMO. The formulated optimization scale antenna arrays at the access-points (APs). Its counterpart
problem is proven to be generally NP-hard. Local optima can known as distributed massive MIMO consists of disseminating
be found by approaches such as the Lagrangian method that the serving antennas over a relatively large geographical
however, has considerably slow rates of convergence. To circum- area. An instance of distributed massive MIMO is cell-free
vent this issue, we propose an alternative solution being of two- (CF) massive MIMO system which has attracted tremendous
folds. Firstly, we reformulate the problem as a grouping strategy
which enables to attenuate the effect of intra-group multi-user attention recently from the research community [7]–[10]. Cell-
interference. In the second fold, frequencies are assigned in a non- free massive MIMO systems are composed of very large
overlapping fashion to each scheduled group to palliate the effect number of APs that simultaneously and coherently serve, on
of inter-group interference and pilot contamination. To further the same time/frequency resource, a relatively small number
improve the performance of the proposed approach, power of users [8]. This distributed configuration enables to achieve
coefficients are allocated to the users via a sequential convex
approximation (SCA)-based framework. The effectiveness of the higher coverage probability, at the expense of higher back-
proposed algorithms is then verified through extensive numerical haul overhead, than its counterpart co-located massive MIMO
simulations which demonstrate a non-negligible improvement in system [11].
the performance of the studied scenario. In a CF massive MIMO network, all APs are communi-
Index Terms— Massive multiple-input multiple-output cating with the users using same time/frequency resource.The
(MIMO), users scheduling, semidefinite programming, pilot network will incur multi-user interference that limit the per-
contamination, multi-user interference. formance of the system. It was shown in [8] that multi-user
interference can be eliminated under the assumption that the
number of transmit antennas goes to infinity. However, this
I. I NTRODUCTION condition is very untenable for practical scenarios. In addition,
the escalating number of mobile devices that are foreseen
I N THE past decades, wireless communication systems
have experienced incessant growth in terms of both mobile
connected devices and mobile data traffic. This growth is fore-
to request services from the APs will lead to unprecedented
amount of interference that needs to be prudently managed
seen to escalate during the years ahead [2]. In addition to that, in order to avoid severe performance deterioration. Therefore,
there is an increasing demand from users for higher data rate. novel and efficient approaches should be investigated given
This implies that 5G and beyond related wireless networks that traditional multi-user interference management methods
deem to be infeasible or inefficient for 5G and beyond wireless
Manuscript received February 5, 2020; revised September 15, 2020 and networks in general [12], and CF massive MIMO systems in
March 5, 2021; accepted May 14, 2021. Date of publication May 31, 2021;
date of current version November 11, 2021. This article was presented at particular.
the IEEE International Workshop on Signal Processing Advances in Wireless There exists a large body of literature focusing on the
Communications (SPAWC) 2018 [1]. The associate editor coordinating the mitigation of multi-user interference for 5G-based networks
review of this article and approving it for publication was S. K. Mohammed.
(Corresponding author: Juwendo Denis.) [13]–[18] (and references therein). The issue of interference
Juwendo Denis is with the School of Engineering and Applied Sci- management for 5G heterogeneous networks was addressed
ences (SEAS), Harvard University, Cambridge, MA 02138 USA (e-mail: in [17]. In [15], [16], [18], the authors leverage the spatial
[email protected]).
Mohamad Assaad is with the Laboratoire des Signaux et Systems (L2S, signature of the channel vectors, and invoke the princi-
CNRS), CentraleSupelec, University Paris-Saclay, 91190 Gif-sur-Yvette, ple of user scheduling, to attenuate the effect of multi-
France (e-mail: [email protected]). user interference and to reduce overhead signaling pertaining
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TWC.2021.3083139. to channel state information (CSI) acquisition. A common
Digital Object Identifier 10.1109/TWC.2021.3083139 theme in existing schemes [14]–[16], [18] is that the APs
1536-1276 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: VIT University. Downloaded on December 15,2021 at 10:12:40 UTC from IEEE Xplore. Restrictions apply.
DENIS AND ASSAAD: IMPROVING CF MASSIVE MIMO NETWORKS PERFORMANCE 7361
are equipped with antennas in the form of uniform linear Different from [1] where we solely proposed an algorithm
array (ULA) and the covariance matrix of the user’s channel to improve favorable propagation for CF massive MIMO, the
follows a one-ring local scattering model. Although, these key contributions of this paper are as follows:
findings cannot directly apply to CF massive MIMO networks,
1) We investigate the benefits of dynamic scheduling of
the standpoint of user grouping can be exploited to design
algorithms to efficiently combat multi-user interference in cell- users for a TDD CF massive MIMO network by opti-
mizing the total downlink achievable rate subject to each
free systems.
user’s QoS requirement. The problem is formulated as
As in co-located massive MIMO, cell-free massive MIMO
networks can achieve good performance at the expense of a binary optimization problem which we prove to be an
NP-hard problem. In addition, a local solution using the
perfect CSI acquisition at the APs. For time-division duplex
(TDD) cell-free massive MIMO networks, the APs can rely Lagrangian method is provided. This Lagrangian-based
on the assumption of uplink/downlink channel reciprocity to solution, which is obtained by means of gradient-descent
approach, has however a very low rates of converge in
estimate the downlink CSI [8] from known pilot signaling of
the users. Considering the huge number of users that will be general.
2) For the purpose of providing faster convergence-based
active in 5G networks, non-orthogonal training sequences must
and more efficient approaches, we establish an alterna-
be utilized during the uplink training stage. Consequently,
the network will incur pilot contamination, a major limiting tive design optimization problem from the perspective
of user grouping. Firstly, we design a framework that
factor of the performance of massive MIMO network [6].
It has been argued in [19] that the capacity increases without partition users into group according to the level of
bound under the condition that the covariance matrices of their mutual interference. Such paradigm can palliate
intra-group multi-user interference. Secondly, groups
the users are asymptotically linearly independent and the pre-
coder is the centralized minimum mean-square error (MMSE). are scheduled on different radio resources to miti-
gate the both the effect of inter-group interference
Although this result is theoretically achievable, it is unattain-
as well as the impact of interference due to pilot
able for practical implementations. In fact, the MMSE pre-
coder requires the computation of the inverse of a matrix that contamination.
3) We invoke the semidefinite relaxation (SDR) [23]
scales with the number of APs and antennas. On top of that,
the acquisition of the covariance matrix is also impeded by method to design a polynomial time solvable random-
pilot contamination during the training phase [20]. Therefore, ized procedure in the same vein of Gaussian random-
ization [23] to efficiently solve the formulated user
there is still a need to provide efficient and practical (in terms
of computation) schemes to combat pilot contamination. One grouping optimization problem.
4) In addition, we design a polynomial time solvable algo-
possible way to do so is to resort to resource allocation such
rithm based on sequential convex approximation (SCA)
as pilot assignment and power control for cell-free massive
MIMO [8], [10], [21]. In this work, we advocate that the to find an efficient local solution to the problem of fre-
quency assignment for the purpose of group scheduling.
impact of pilot contamination can be circumvented through
radio resource allocation. We establish the convergence to a stationary point of
Existing literature about performance evaluation for CF the proposed SCA-based algorithm. Moreover, we con-
sider the power control optimization problem which is
massive MIMO networks has been done considering single
frequency system. However, for dense networks such as CF efficiently solved by using an SCA-based approach.
massive MIMO network single frequency system implies The rest of this work is structured as follows. In Section II,
aggressive frequency reuse that leads to co-channel multi- the CF massive MIMO system model is presented. The
user interference that deteriorate the quality of service (QoS) resource management problem formulation is given in
of the users [22]. This might be avoided by leveraging Section III together with the Lagragian-based solution.
the fragmentation of the frequency band along with proper The user grouping approach is introduced in Section IV.
management of the resulting sub-frequency with the goal of In Section V, the description of our proposed SDR-based algo-
achieving performance enhancement. rithm is provided followed by the associated sub-frequency
Motivated by the above discussion, we investigate the band assignment as well as the problem power control allo-
impact of sub-frequency bands allocation on the issues of pilot cation. In section VI, simulation results demonstrating the
contamination and multi-user interference which are two major effectiveness of our proposed methods are provided. And we
limiting factors for the performance gains of a CF massive conclude the paper in Section VII.
MIMO system. Despite the great deal of existing works on Notation: In this paper, x ∼ CN (μ, σ 2 ) represents a
CF massive MIMO [7]–[10], [21](and references therein), circularly symmetric complex random variable drawn from
no prior work has attempted to address these issues from a Gaussian distribution with mean μ and variance σ 2 . The
the perspective of radio resource assignment. In this paper, natural logarithmic function is denoted as log(·). Given a
we consider a dynamic radio resource allocation approach and nonempty set A, |A| stands for the cardinal of A. [x]ba is
evaluate its impact on the performance of TDD CF massive the Euclidean projection of x onto the interval[a, b] while
MIMO networks. (x)+ max(0, x).
7362 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 11, NOVEMBER 2021
φk 2 = τ . The channel estimate can be obtained in a decen-

tralized fashion at each AP m. Define Gm [g m1 , · · · , g mk ],
the N × K received pilot matrix signal at the mth AP is
expressed as
√
Y m,p = τ ρp Gm Φ† + N m,p (2)
where ρp is the transmit power during the training phase,

and N m,p ∈ CN ×K is the complex additive Gaussian noise
matrix at the mth AP. The elements of N m,p are i.i.d. random
variables that follow a standard normal distribution. The mth
AP can perform MMSE [24] estimation of g mk to obtain the
corresponding channel estimate, g mk which is given by [10]
√
Fig. 1. Illustration of a Cell-Free massive MIMO network. τ ρp βmk
mk =
g 2
K
ρp τ j=1 βmj φH φ
k j + 1
⎛ ⎞
II. S YSTEM M ODEL A ND P RELIMINARIES K
√ √
In this work, we consider a cell-free massive MIMO system × ⎝ τ ρp g mk + τ ρp g mj φk φj + wmk ⎠ (3)
H
j=1,j=k
that consists of M access points (APs) that serve simul-
taneously on a fully-cooperative fashion K mobiles users where wmk ∼ CN (0, 1) is the kth column of the matrix
(UEs), equipped each with a single omni-directional antenna N m,p Φ. The quality of the channel estimation is captured
as depicted in Figure 1. The APs are assumed to have each N through the mean-squared error (MSE) E{g mk − g mk 2 }.
omni-directional antennas. Furthermore, we suppose that the A good estimation quality is usually represented by a small
number of APs is much higher than the number of UEs i.e. MSE. Also, the channel estimate in (3) is corrupted by pilot
K M . All users are arbitrarily distributed over a relatively signals sent by other users, in case K > τ , leading to pilot
large geographical area. The APs and the UEs are assumed contamination which degrades user’s achievable rate [6]. It is
to be perfectly synchronized in both time and frequency. worth mentioning that, we only assume a given transmit power
Moreover, we suppose that the APs are managed by a central ρp during each training phase as this paper does not focus
processing unit (CPU) to which they are connected through on the impact of adjusting ρp on the quality of channel
a error-free back-haul link. The CPU can handle part the estimates. Similar to [25], this paper assumes imperfect chan-
physical layer information process such as data coding and nel reciprocity. Denote g mk g mk − g mk as the channel
decoding. estimation error. By considering random realizations of the
The channel of a user is considered to be frequency-flat slow MMSE channel estimate and the channel estimation error
fading on each orthogonal frequency division multiplexing mk and
during an arbitrary coherence block, it holds true that g
(OFDM) subcarrier. In the sequel, the subcarrier index will be g mk are uncorrelated and are respectively distributed as [7]
omitted for the sake of simplicity. Let g mk ∈ CN ×1 , denote ⎛ ⎞
the complex channel vector between the kth user and the mth 2
AP. It can be modeled as follows ⎜ τ ρp βmk ⎟
mk ∼ CN ⎝0,
g IN ⎠ ,
K H 2
g mk = βmk hmk (1) ρp τ j=1 βmj φk φj + 1
⎛ ⎛ ⎞ ⎞
where hmk ∼ CN (0, IN ) with IN being the N × N identity 2
⎜ ⎜ τ ρp βmk ⎟ ⎟
matrix, m = 1, · · · , M, k = 1, · · · , K, denote the small- g mk ∼ CN ⎝0,⎝βmk − K ⎠ IN ⎠
H 2
scale fading coefficients which are independent and identically ρp τ j=1 βmj φk φj + 1
distributed (i.i.d) while βmk , m = 1, · · · , M, k = 1, · · · , K, (4)
the large scale fading coefficients that include path-loss and
shadowing.
The system is operating according to a TDD protocol. Each A. Downlink Payload
coherence interval Tc is divided between uplink training and The downlink signal intended to each UE is precoded at the
downlink data transmission. By exploiting uplink/downlink APs with conjugate beamforming. Accordingly, the transmit
channel reciprocity [6], the downlink CSI can be estimated signal of the mth AP to all users is
through uplink training. During the uplink training phase, each
√ √
K
user is assigned a training sequence that spans τ < Tc channel
y m,d = ρd ∗mk dk
ηmk g (5)
uses. The pilot sequences used in the channel estimation
k=1
phase can be represented by a matrix Φ ∈ Cτ ×K . The kth
column of the matrix denoted by φk ∈ Cτ ×1 accounts for where dk with E{|dk |2 } = 1, denotes the data symbol intended
the pilot sequence used by the kth UE. Each element of φk for user k, ρd accounts for the downlink transmit power and
is of unit-magnitude so that it has a constant power level ηmk , the power coefficient between from the mth AP to user k
chosen to satisfy E{ym,d 2 } ≤ ρd at each AP. The received III. R ESOURCE A LLOCATION P ROBLEM F ORMULATION
signal at the kth user is given by Before we proceed to formulate the radio resource allocation
⎛ ⎞ problem, let us define the binary variable ϑk ∈ {0, 1}, and
√ M K
√ ϑ = {{ϑk }Kk=1 }=1 . The value of ϑk takes one if the APs can
L
rk = ρd ⎝ ηmj g ∗mj dj ⎠ + nk ,
mk g (6)
m=1 j=1
communicate with the k UE on the th sub-frequency band,
otherwise it takes value 0. Moreover, the power coefficient
where nk ∼ CN (0, 1) denotes the additive white Gaussian allocation from one AP to any use on each sub-frequency band
noise at user k. Similarly to [7], [10], we assume that is assumed to be fixed. Our goal is to dynamically assign the
UEs have only statistical knowledge of the channel esti- available radio resources for the purpose of maximizing the
mate. Accordingly, the net downlink rate of the kth user is total throughput of the network subject to QoS constraint. The
given by [10] optimization problem is mathematically formulated as follow

K
L
Rk = B 1 −
τ
× log2 max f (ϑ)
ϑk R (9a)
k
Tc ϑ∈{0,1}(K×L)
⎛ 2 ⎞ k=1 =1
M √ L
⎜ ρ d ηmk N ν mk ⎟ k ,
m=1 s.t. Rk ≤ ϑk R ∀ , ∀k (9b)
⎝1 + M K ⎠ (7)
N ρd m=1 ηmk βmk νmk + ρd N j=k Ikj + 1 =1
where Rk is the minimum rate requirement for user k and R

k
where B is the spectral bandwidth, 1 − τ
Tc is the fraction in units of bits/s/Hz, the net downlink throughput, on the lth
sub-frequency band.
of the channel uses spent for downlink data transmission
and
R = B 1 − τ × log
k 2
2 L Tc
τ ρp βmk ⎛ 2 ⎞
νmk , M
H 2 ν
ρp τ K β φ φ + 1 ⎜ ρ d m=1 N η mk mk ⎟
j=1 mj k j ⎜ ⎟
M 2 ⎜1 + M ⎟
√ 2 ⎝ N ρd m=1 ηmk βmk νmk + ρd N j=k Ikj ϑj + 1⎠
K
νmj ∗
Ikj N ηmj βmk φ j φk
m=1
βmj
(10)

M
+ ηmj βmk νmj (8) where B
L is the width of a sub-frequency band, ηmk denotes
m=1 the power coefficient on the lth frequency sub-band and
M 2
2
ν ∗
βmk φ
mj
B. Dynamic Radio Resource Assignment Ikj N
ηmk j φk
m=1
βmj
For 5G and beyond cells structured networks, radio
resources are intended to be allocated in terms of resource-
M

+ ηmj βmk νmj
block (RB) on a per-cell basis [26]. For instance, given the
m=1
signal bearer is modulated using orthogonal frequency division 2
τ ρp βmk
multiple access (OFDMA), each base station assigns each
νmk K 2 (11)

k φj ϑj + 1
τ ρp βmk + ρp τ j=k βmj φH
RB to only one user at a given time. The same RBs can be
reused simultaneously within different cells which will create

inter-cell interference. Contrarily, the users in cell-free systems Remark 1: From the expression of νmk in (11), one may
do not experience cells boundaries [27]. In a TDD cell-free notice that for a given sub-frequency band l and two users
network, all APs coherently communicate with all users on k, k1 , if ϑk = 1 while ϑk1 = 0, it can be observed that the
the same radio resource at the same time. The network may channel estimate of user k on sub-frequency band l will not
incur substantial amount of multi-user interference, on top of be affected by pilot contamination from user k1 regardless
unavoidable pilot contamination, which will severely affect that both users are assigned the same pilot vector. This leads
the performance of the system. In this work, to palliate the to believe that dynamic sub-frequency band allocation can
damaging effect of both interference and pilot contamination, potentially palliate the effect of pilot contamination.
we consider dynamic assignment of bandwidth for the users. Problem (9) is a mixed-integer programming which is more
Specifically, the available bandwidth is divided equally into L likely difficult to be solved in the global optimum [28].
orthogonal parts, which we refer to as sub-frequency bands Therefore, it of interest to us to investigate its complexity
that are allocated to the users in order to diminish the effect status.
of interference. In other words, a user may be scheduled only Proposition 1: The optimization problem formulated in (9)
on few sub-frequency bands in contrast to existing literature is NP-hard.
[7]–[10] where a user is scheduled over the entire Proof: The proof of Proposition 1 is relegated to
bandwidth. Appendix A
Proposition 1 advocates that it is impossible to find a where γ ≥ 0, υ ≥ 0 are the Lagrange multipliers. Define
polynomial time solvable algorithm capable of seeking the
∂L (ϑ, γ, υ) K
∂
R ∂
R
global optimum for problem (9) unless P = NP. Nevertheless, =−
j
ϑj γj − γk ϑk k + R
k
a local solution can be found by mean of classical approaches ∂ϑk j=k
∂ϑ k ∂ϑ k
such that Lagrangian approach.
∂R K
∂R
− ϑ
−R k
− ϑ j
k k j
∂ϑk j=k
∂ϑ
k
A. Lagrangian-Based Solution
Before proceeding in describing the Lagrangian-based fre- +ς 1 − 2ϑk

(17)
quency band allocation solution, we rewrite the binary con- The Karush-Kuhn-Tucker (KKT) conditions [31] are given by
straint ϑk ∈ {0, 1}as the difference of two concave-functions
∂L (ϑ , γ , υ )
0 ≤ ϑk ≤ 1 = 0, ∀ , ∀k
∂ϑk
L

K
2 (12) L
ϑk − ϑk ≤0 υk 1 −

ϑk
= 0, ∀k
k=1 =1
=1

and define the following set:
L
γk
Rk − k
ϑk R = 0, ∀ , ∀k
(9b)
D ϑ: =1
0≤ϑ≤1
Considering that it is extremely complicated to find a closed
Problem (9) can be equivalently recast as form solution from (17), a stationary point ϑ can be found
by invoking the gradient descent method [32] where at the nth
min − f (ϑ)
ϑ∈D iteration.
L

K
2
s.t. ϑk − ϑk ≤ 0 ∀k, ∀ (13) ϑk [n] = ϑk [n − 1]
k=1 =1 1
∂L (ϑ, γ[n − 1], υ[n − 1])
We have the following proposition −δϑ |ϑ=ϑ[n−1] ∀ , ∀k (18)
Proposition 2: For large value of ς 1, the optimization ∂ϑk 0
problem (13) is equivalent to where δϑ > 0 is a sufficiently small positive number. The
L Lagrange multipliers can be updated by the sub-gradient

K
2
min L (ϑ, ς) −f (ϑ) + ς ϑk − ϑk approach [32, Chapter 6]. The Lagrangian-based method is
ϑ
k=1 =1 summarize in Algorithm 1 where δυ > 0, δγ is a sufficiently
s.t. ϑ ∈ D (14) small positive number. Let κ1 denotes the time complexity
The proof of Proposition 2 can be derived in a similar fashion required to compute the gradient of ϑk , with a solution
as the proof of [29, Proposition 1] or [30, Proposition 2]. accuracy of 1 . The worst case computational complexity for
It can be inferred from Proposition 2 that strong duality holds. the gradient descent is O(KLκ1 log(−11 )). Since the objective
Consequently, our goal is to solve the optimization problem function is differentiable, the complexity of the sub-gradient
(14) which we rewrite as approach to calculate the Lagrange multipliers is computing
L in a similar fashion as the computational complexity of the

K
L
K
2
max k
ϑk R −ς ϑk − ϑk gradient method [33]. Then, the overall computational com-
0≤ϑ≤1 plexity of the Lagrangian-based solution to find a station-
k=1 =1 k=1 =1
L ary point to problem (15) is given by O(Kκ2 log(−1 2 ) +
s.t. Rk ≤ , ∀k
ϑk Rk
Kκ3 log(−1 −1
3 )) · O(KLκ1 log(1 )) where O(κ2 log(2 ))
−1
−1
=1 and O(κ3 log(3 )) are the computational complexity of the

L Lagrangian multiplier γ and υ, respectively. 2 , 3 are the
1≤ ϑk , ∀k (15) solution accuracy for the sub-gradient method to compute
=1 γ and υ, respectively. The convergences condition can be
The Lagrangian associated with the optimization problem (15) chosen to be (L(ϑ[n], γ[n], υ[n])−L(ϑ[n−1], γ[n−1], υ[n−
is formulated as 1]))/L(ϑ[n − 1], γ[n − 1], υ[n − 1]) < 1 . It is worth noticing
that the convergence of Algorithm 1 can be extremely slow

K L
L (ϑ, γ, υ) = γk Rk − k
ϑk R especially when higher accuracy is required since the local
k=1 =1
solution is computed by means of gradient descent and sub-
gradient approaches.
K L
+ υk 1− ϑk Although the Lagrangian-based scheme can find a local
k=1 =1 solution to problem (9), this method has rather low-rate
L
K
of convergence. This motivates us to provide an alternative
− − ς ϑ − ϑ 2
ϑlk R (16) approach with higher rates of convergence and that can enable
k k k
k=1 =1 the attenuation of the effect of multi-user interference and
Algorithm 1 Lagrangian-Based Approach to Solve and we define the mutual interference weight as follow
Problem (15) Ikj + Ijk
1: Input A feasible ϑ[0] to problem (15) and a solution ωkj = ωjk (21)
accuracy 1 > 0. 2
2: Set n ← 0;
3: repeat A. Graphical Modeling
4: Initialize γ[0], υ[0] and set m ← 0; We start by constructing a spatial correlation graph that
5: repeat captures the level of interference for a set of users which
6: m ← m + 1; are active simultaneously. More specifically, we design an
7: Update γ, υ as undirected interference graph G(V, E). The set of vertices
+ V, with |V| = K, of the considered graph accounts for

L
υk [m] ← υk [m − 1] + δυ 1 −
ϑk [n] , ∀k the users in the coverage area while each edge ekj ∈ E is
=1 associated with a weight ωkj defined in (21) set to capture the
γk [m] ← (γk [m − 1] + δγ interference level. Using the constructed graph G(V, E), our
+ goal is to formulate an optimization problem to partition users
L
Rk −
ϑk [n]R
, ∀k (19) into groups. The considered optimization framework aims at
k
=1
minimizing the mutual interference weight between the users
that are assigned to the same group.
8: until the subgradient approach converges Define the following variable
9: n←n+1 ;
10: Update ϑ[n] by ϑk [n] ← 1 if user k is allocated to the c-th group
1 xk,c = (22)
∂L(ϑ,γ[m],υ[m])
ϑk [n − 1] − δϑ

∂ϑk
|ϑ=ϑ[n−1] , ∀ , ∀k. 0 otherwise
0
11: until the gradient descent method converges.
12: Output the stationary solutions ϑ[n].
The user’s scheduling problem is formulated as the following
combinatorial optimization problem

C
pilot contamination. This is done from the standpoint of user max ωkj (1 − xk,c )xj,c
xk,c ∈{0,1}, ∀c,∀k
c=1 k∈V j∈V,j=k
grouping.

C
s.t. 1 ≤ xk,c ≤ α, ∀k ∈ V
IV. S CHEDULING -BASED R ESOURCE A LLOCATION c=1

Our proposed scheduling approach aims at clustering users xk,c ≤ τ, ∀c = 1, . . . , C, (23)
to palliate the effect of multi-user interference. The motivation k∈V
lies in the fact that by grouping together users that cause
small amount of interference to each other, the intra-group where C denote the total number of groups and α, the maxi-
interference can be made as small as possible. Regarding inter- mum number of groups a user is permitted to be assigned to
group interference and pilot contamination interference, their concurrently. We also refrain each group to have a maximum
effect can be alleviate by scheduling each group on different number of users, i.e τ , to avoid intra-group pilot contamina-
sub-frequency bands. tion. Problem (23) is a combinatorial optimization problem
Our scheduling procedure is executed at the CPU side and therefore very difficult to solve. In the next section,
and is involved large scale fading of the users’ channel and we propose an algorithm to seek an efficient solution to
some other parameters. The group coefficients are sent to problem (23).
the access points for downlink data payload. As discussed
subsequently, the users will be partitioned into groups from the V. S EMIDEFINITE R ELAXATION -BASED A PPROACH
perspective of graph theory. To be more specific, a weighted
graph will be constructed with the weight being the mutual In this section, we turn our attention in developing a
interference, which needs to be defined, between two users. polynomial-time solvable algorithm to find an efficient sub-
Towards this end, we consider the interference between two optimal solution to problem (23). However, let us first
users j and k (we assume that both users are assigned same rewrite the optimization problem (23) in a more compact
pilot sequence) in the absence of all the other users. It is way. Towards this end, we define the following change of
written as variables
M 2 xc (x1,c · · · , xK,c ) , yc 2xc − 1K ,
√ηmk τ ρp βmj βmk ⎛ ⎞
Ikj = N 0 w21 · · · wK1
m=1
τ ρp βmj + ρp τ βmk + 1 ⎜ w12 0 · · · wK2 ⎟
⎜ ⎟

M 2
ηmj τ ρp βmk βmj W ⎜ . .. . . .. ⎟ (24)
+ (20) ⎝ .. . . . ⎠
m=1
τ ρp βmj + ρp τ βmk + 1 w1K w2K · · · 0
where 1K is a column vector which entries are 1. Using (24), Algorithm 2 A Randomized Algorithm to Solve Problem (25)
problem (23) can be equivalently reformulated as 1: input an optimal solution Zc , ∀c to problem (29).

2: Generate L vector samples ξ c , , = 1, · · · , L uniformly

1
C
min y Wyc distributed on the unit sphere;
4 c=1 c
3: Generate ζ c , = (ζ1,c

, · · · , ζK+1,c

, ) according to

C
s.t. (2 − C)1K ≤ yc ≤ α1K
1 if vc,k ξc , ≥ 0
ζk,c , = (30)
c=1 −1 otherwise
Tr (diag (yc )) ≤ τ , ∀c
yc ∈ {−1, 1}K , ∀c (25) 4: Verify that ζ c , meet the constraint of (27).
C
5: Compute , = arg min,=1,··· ,L 41 c=1 ζ c , Wζ c ,;
where α = 2α − C and τ = 2τ − K. Furthermore, we want to
express the variables in problem (25) to have only quadratic 6: output the solution
zc = ζ ,
c , ∀c.
terms which have nicer property. To do so, let us define the
following change of variables

W 0K×1 combinatorial problem (27). However, dropping the rank one
zc yc , 1 , W
01×K 0 constraint, i.e. rank (Z) = 1, the optimization problem (28)
can be relaxed as
0K×K π 0K×K p
W1 , Qk
1
(26) C
π 0 p 0
min Tr WZc
4 c=1
where π 12 1K×1 , and p is a K × 1 vector with entries zeros
except the kth entry which is 12 . Therefore, by combining (26)
C
and (25), the combinatorial optimization problem (23) can be s.t. 2 − C ≤ Tr (Qk Zc ) ≤ α, ∀k
compactly written as c=1
Tr (W1 Zc ) ≤ τ , ∀c
1
C
min z Wzc Zc 0, diag (Zc ) = 1K+1 , ∀c (29)
4 c=1 c
Problem (29) is a standard convex optimization problem.

C
Therefore, we can invoke interior-point based solvers such
s.t. 2 − C ≤ z
c Qk zc ≤ α, ∀k as CVX [35] to efficiently findingits global optimal solution

with a worst case complexity of O (K + 1)3.5 log(−1
c=1
4 ) [36]
z
c W1 zc ≤ τ , ∀c where 4 is the solution accuracy. If the optimal solution
zc ∈ {−1, 1}K+1, zc,K+1 = 1, ∀c (27) obtained in solving problem (29) is rank one, it is therefore
also the optimal solution to the optimization problem (28).
To find an efficient sub-optimal solution to problem (27),
However, it is not generally the case since problem (29) is
we resort to the semidefinite relaxation (SDR) method [23]
a relaxed instance of problem (28). Consequently, solving
which has been widely known as a powerful tools capable of
(29) cannot not always guarantee to yield an optimal solution
handling combinatorial optimization problems [34]. The appli-
Zc satisfying the rank one constraint. However, the opti-
cation of SDR method
comes
after
a key observation
namely,
mal solution Zc 0, ∀c to problem (29) can serve as a
z
c Qk zc = Tr zc Qk zc = Tr Qk zc zc . Following [23],
basis for finding an approximate solution to the optimization
SDR approach consists of making the change of variables
problem (28).
Zc = zc z
c . This change of variable is equivalent to substitut-
One suggestion for obtaining an approximate solution is
ing Zc = zc z
c by using a positive semidefinite (PSD) matrix Z
to resort to powerful randomization approaches [37] such as
together with an additional rank one constraint. Consequently,
the Gaussian randomization [23]. We propose a randomized
the combinatorial problem (27) can be equivalently recast as
procedure in the same vein as the randomized technique
1
C
in [37] that takes as input the optimal solution to problem
min Tr WZc (29) and outputs a feasible solution to problem (28). Given
4 c=1
Zc 0, ∀c, an optimal solution to problem (29), it can be

C
decomposed using Cholesky factorization as Zc = V c Vc , ∀c
s.t. 2 − C ≤ Tr (Qk Zc ) ≤ α, ∀k
with Vc = (vc,1 , · · · , vc,K+1 ). The proposed randomized
c=1
scheme is summarized below.
Tr (W1 Zc ) ≤ τ , ∀c
Zc 0, diag (Zc ) = 1K+1 , ∀c
A. Disjoint Allocation of Sub-Frequency Band
rank (Zc ) = 1, ∀c (28)
The proposed grouping strategy enables to partition together
The optimization problem (29) is a non-convex problem users that cause less interference to each other. In other
because the convexity of the constraint rank (Z) = 1. There- words, the proposed user grouping policy only tackles intra-
fore, it is as difficult to finding the optimal solution to group interference. We now turn our attention to investigate
the optimization problem (28) than solving its counterpart solutions to palliate inter-group interference. We advocate that
inter-group interference can be eliminated by scheduling the The optimization problem (31) can be recast as
groups on different sub-frequency bands. The problem of
1
L C
group scheduling is formulated as max θc Rk,c (33a)
θc ∈{0,1}, ∀, ∀c L c=1
=1 k∈Γ(c)
1
L C
1
L C
max θc Rk,c
θc ∈{0,1}, ∀, ∀c L s.t. Rk ≤ θc Rk,c , ∀k (33b)
c=1
=1 k∈Γ(c) L c=1
=1
1 θc + θc̃ ≤ 1, ∀c = c̃, ∀

L C (33c)
s.t. Rk ≤ θc Rk,c , ∀k
L Let θ {{θc }L
=1 }c=1 and define the following set
C
c=1
=1

θc θc̃ = 0, ∀c = c̃, ∀ (31) (33b), (33c)
X θ:
0≤θ≤1
where θc = 1 if group c is scheduled for downlink trans-
mission on the lth sub-frequency band, otherwise θc = 0. It is straightforward to demonstrate that the set X is a convex
Moreover, Rk,c
is the rate of the kth user within group c set. Using Proposition 2, problem (33) can be reformulated as
on the lth sub-frequency band. It can be written as (32)
1
L C
,c
(see bottom of the page), where ηmk denotes the power max ψ (θ) θc Rk,c
θ∈X L
coefficient from the mth AP to user k within group c and on c=1
=1 k∈Γ(c)
the lth sub-frequency band, xk,c is solution to problem (23),
L
C

and the prelog factor 1 − |Γ(c)| accounts for the number of −ς1 θc − (θc )2 (34)
Tc
channel uses destined for downlink transmission on a given =1 c=1
sub-frequency band. The power coefficients are assumed to The optimization problem (34) is a non-convex optimization
be fixed during the sub-frequency band allocation. More- problem because the objective function is not a concave func-
over, a maximum of |Γ(c)| ≤ τ users is scheduled within tion. In the subsequent subsection, we propose an algorithm
a given group which is exclusively assigned a given sub- to efficiently solve problem (34).
frequency band. Therefore, during the uplink training stage
on that particular sub-frequency band, the length of pilot B. Sequential Convex Approximation-Based Solution
sequences assigned to the users can simply be |Γ(c)| which
explains why the fraction of channels uses dedicated to The proposed approach is motivated by the sequential
downlink transmission on that particular sub-frequency band parametric convex approximation (SCA) method [38]. Let θ be
is |Γ(c)|
Tc . Accordingly, the interference term Ikj,c is given a feasible point to problem (34). Our goal is to conservatively
M ,c approximate the objective function of problem (34) at the
by Ikj,c = m=1 ηmj βmk νmj,c + 1. The pilot contami- point θ by invoking first order convex condition [31]. More
nation term that appears initially in the interference Ikj in
specifically,
(32) vanishes thanks to our grouping strategy that allocates

orthogonal pilot vector to users that belong to the same (θc )2 ≥ (θc )2 + 2θc θc − θc (35)
group.
Remark 2: In (32), νmk,c is given by νmk,c = By using (35), the optimization problem (34) can be conser-
2
τ ρp βmk
. The denominator vatively approximated by
2
τ ρp βmk +ρp τ K
j=k,j ∈Γ(c)
/ βmj |φH
k
φj | +1
1
L C
excludes users in same group because our user partition
max ψ θ; θ θc Rk,c
strategy restricted the maximum number of users in each θ∈X L c=1 =1 k∈Γ(c)
group to be τ in problem (23). Consequently, the users that
C
L
belong to the same group are assigned pair-wisely orthogonal
training vector. In addition, the group scheduling problem −ς1 θc − (θc )2 − 2θc θc − θc (36)
=1 c=1
(31) prevents two different groups to be scheduled on the
same sub-frequency band. By combining our proposed user The optimization problem (36) is a convex problem and
partition strategy and proposed group scheduling approach, therefore can be solved via interior-point based solvers such as
the pilot contamination in the denominator of νmk,c will CVX [35]. It is worth recalling that we obtained problem (36)
vanish. Therefore, the combination of both approaches will by a restrictive approximation of the non convex optimization
mitigate the effect of pilot contamination. problem (34) around the feasible point θ. Intuitively, the
⎛ 2 ⎞
M
⎜ ρd m=1 N
,c
ηmk νmk,c ⎟
|Γ(c)| ⎜ ⎟

Rk,c xk,c B 1 −

log2 ⎜1 + M ⎟ (32)
Tc ⎝ N ρd m=1 ηmk βmk νmk,c + ρd N j∈Γ(c),j=k Ikj,c + 1 ⎠
,c K

performance of this approximation can be further enhanced Secondly, the function
ψ θ;θ is a tight lower bound to
by simply using the optimal solution obtained from previous ψ (θ), that is ψ θ; θ = ψ θ . Thirdly, it holds true that
iteration and by sequentially solving the standard convex ∂ψ (θ;θ)
|θ→θ = ∂ψ(θ)
∂θ |θ→θ . Hence, our proposed Algorithm 3
problem (36). To be more specific, denote θ[n] the solution ∂θ
is special instance of the successive-upper-bound minimiza-
of the problem (36) at the nth iteration and set θ[n] = θ[n]. tion (SUM) algorithm [39]. Based on [39, Theorem 1], any
We solve the following convex optimization problem at the limit point to problem (37) generated by the SCA-based
(n + 1)-iteration Algorithm 3 is a stationary point to problem (34).

+ 1] = arg max ψ θ; θ[n]
θ[n (37)
θ∈X
The proposed SCA-based algorithm to obtain solution to D. Power Control

problem (34) is summarized in Algorithm 3. Once the solution of the sub-frequency allocation problem is
found, the performance of our proposed scheme can be further
Algorithm 3 SCA-Based Algorithm to Solve Problem (34) improved by reducing the intra-group interference through
power control. The power control optimization problem is
1: Input A solution accuracy 5 > 0 and a feasible point θ[0]
formulated as
for problem (34).
1
L C
2: Set n = 0;
3: Repeat max θc Rk,c
,c
{ηmk } L c=1
=1 k∈Γ(c)
4: n=n+1 ;
5:
Compute θ[n] by solving problem (37) using CVX [35]; 1
L
C
6:
Update θ[n] = θ[n] . s.t. Rk ≤ θc Rk,c

, ∀k
L
=1 c=1
7: Until a predefined stopping criterion.
8:
Output the approximated solutions θ[n].
K L C
,c
N xk,c θc ηmk νmk,c ≤ 1, ∀m
k=1 =1 c=1
There are CL variables in the optimization problem ,c
ηmk ≥ 0, ∀m, ∀k, ∀ , ∀c (38)
(36) and K + C(C − 1)L linear and convex constraints.
Therefore, the worst-case computational
complexity of the where the second constraint is the total power budget con-
3
SCA-based Algorithm 3 is O (CL) (K + C(C − 1)L) . straint at each AP. Define the following change of variables:

The
stopping criterion
can be chosen
to be (ψ θ; θ[n] −
t,c ,c
ηmk
ψ θ; θ[n − 1] )/(ψ θ; θ[n − 1] ) > 5 . To further address mk
2
the efficiency of the proposed SCA-based Algorithm 3, it is ρd N 2
M
t,c νmk,c
m=1 mk
important to investigate its convergence analysis. This is done
γk,c 2
through next subsection. N ρd j∈Γ(c) M ,c
m=1 tmj βmk νmj,c + 1

|Γ(c)|
C. Convergence Analysis Δk,c = xk,c B 1 − (39)
Tc
Theorem
1: !
The sequence of objective function
+∞
By using (39), the optimization problem (38) can be equiva-
ψ θ[n]; θ[n] generated by the proposed SCA
n=1 lently reformulated as:
Algorithm 3 is a convergent sequence.

Proof: At (n + 1)-th iteration, the solution θ[n] is used 1
L C

as an iterate value (ref. step 6) to problem (37). Therefore max θc Δk,c log2 1 + γk,c
{t,c ,c
mk },{γk }
L
c=1
=1 k∈Γ(c)
+ 1]; θ[n]
ψ θ[n = ψ θ[n + 1]; θ[n]
(40a)
M
2

≥ ψ θ[n];
θ[n] s.t. N ρd t,c βmk νmj,c + 1 ≤
mj
= ψ θ[n] j∈Γ(c) m=1

2
ρd N 2
M ,c

(ā) m=1 tmk νmk,c

≥ ψ θ[n]; θ[n − 1] , ∀k, ∀ , ∀c (40b)

γk,c
where (ā) follows from (35). The objective function is non-
1
L C
θ is bounded above.
decreasing at each iteration.Moreover,! Rk ≤
θc Δk,c log2 1 + γk,c , ∀k (40c)
+∞ L

Therefore, the sequence ψ θ[n]; θ[n] converges. c=1
=1
n=1
K
L C 2
Corollary 1: Any limit point of sequence { y[n]}+∞
n=1 for N xk,c θc t,c νmk,c ≤ 1, ∀m (40d)
mk
problem (37) generated by the proposed Algorithm 3 is a k=1 =1 c=1
stationary point to the optimization problem (34). t,c
mk ≥ 0, ∀m, ∀k, ∀ , ∀c (40e)
Proof:
Firstly, it is worth noticing that the func-
tion ψ θ; θ is a continuous function of both θ and θ. γk,c ≥ 0, ∀k, ∀ , ∀c (40f)
One can notice that the objective function in (40a) is a TABLE I

concave function and that all constraints from (40c) to (40f) S IMULATION PARAMETERS
are convex constraints. However, problem (40) is a non-convex
problem due to constraint (40b). The right hand side of (40b)
possesses some nice properties which we can leverage to solve
problem (40). Indeed, it is a convex function. This can be
proven by deriving the Hessian. Therefore, problem (40) can
also be solved by invoking the principle of SCA. This is done
by conservatively approximate the right hand side of (40b) via during the group scheduling stage.
the first order Taylor expansion. Accordingly, the optimization
1 l
L C
problem (40) can be approximated as max min c Rk,c
k lc ∈{0,1}, ∀l, ∀c L c=1
l=1 k∈Γ(c)
1
L
C
max θc Δk,c
log2 1 + γk,c s.t. c + c̃ ≤ 1, ∀c = c̃ ∀ (43)
{t,c ,c
mk },{γk }
L
=1 c=1 k∈Γ(c)
Using similar arguments as the ones adopted to solve problem
M
2 (33), i.e the result of Proposition 2 and the first order convex
s.t. N ρd t,c
mj βmk νmj,c + 1 condition [31], we can apply the SCA algorithm principle to
j∈Γ(c) m=1 the following conservative approximation of problem (43)
⎛ 2
M ,c
⎜ m=1 tmk νmk,c max
≤ ρd N 2 ⎝ + ζ̂≥0,0≤lc ≤1, ∀l, ∀c

γk,c

L
C

M ζ̂ − ς2 c − (c )2 − 2c c − c
,c
2 m=1 tmk νmk,c t,c
mk − t,c
mk =1 c=1

γk,c 1 L C
2 ⎞ s.t. ζ̂ ≤ lc Rk,c , ∀k
M ,c L
l=1 c=1 k∈Γ(c)
m=1 tmk νmk,c ⎠,
−
γk,c − γk,c

γk,c c + c̃ ≤ 1, ∀c = c̃ ∀ (44)
∀k, ∀ , ∀c(40c) − (40f) (41)
VI. N UMERICAL R ESULTS
At each iteration of the SCA approach, the closed-form In this section, we provide numerical results in order to eval-
expression for the solutions of problem (41) which can be uate the performance of the proposed algorithms. All results
found by solving the KKT conditions is given by are obtained using Monte Carlo simulations by averaging,
unless stated otherwise, over 500 realizations for the position
⎡ ⎤+ of the APs and the users. All APs and users are uniformly
⎢ Δk,c θc (1 + ωk ) ⎥ distributed over a square region of width 1 Km. The large-
⎢ ⎥
=⎢ 2 − 1⎥

γk,c scale fading coefficient is modeled as [8]
⎣ M
t,c
ν ⎦
μlk,c L ln(2)ρd N 2 m=1 mk mk,c
βmk [dB]
γk,c ⎧ rmk
⎪
⎨−35.7 − 35 log10 1m + zmk rmk ≥ 50m
ρd N μlk,c t,c
t,c
mk =
mk
(42) = −61.2 − 20 log10 r1m mk
10m ≤ rmk < 50m

γk,c ρd j∈Γ(c) μlj,c βmj + λm xk,c θc ⎪
⎩
−81.2 otherwise
(45)
where {μlk,c }, {ωk } and {λm } which can be solved using the
subgradient approach [31], are Lagrange coefficients associ- where rmk is the distance from the mth AP to the kth
ated respectively with first, second and third constraint of the user. zmk is the log-normal shadowing with standard deviation
optimization problem (41). The convergence of the SCA-based of 8 dB. All other parameters are summarized in table I.
algorithm for the power control problem can be derived in a For cases without power allocation, the APs transmit with
similar fashion as in [38]. uniform power allocation. That is (i) ηmk = N K1 ν for
k mk
1
canonical CF, (ii) ηmk = LN K
νmk
the power on each sub-
k
frequency band and (ii) = LN C 1K ν
,c
ηmk for each sub-
E. Max-Min Fairness k mk,c
frequency band and for each group. For our proposed scheme,
Most existing works on cell-free massive MIMO have we implement the randomization Algorithm 2 and SCA-based
considered the problem of max-min fairness [7]–[9], [21]. Algorithm 3 which we term as Proposed SDR + SCA. When-
To be able to provide a fair comparison between our proposed ever power control is performed alongside the grouping and
approach and existing literature in Section VI, we now con- frequency allocation, we term it as Proposed SDR + SCA +
sider the max-min optimization problem which is implemented Power. We first compare the cumulative distribution of the
Fig. 2. Cumulative distribution of the per-user downlink achievable rate for

K = 20, M = 50, L = 10, ρd = 1, α = 4 and C = 5.
Fig. 4. (a)CDF of per-user achievable downlink rate for M = 100, K = 30,

L = 10, ρd = 1, C = 5 and α = 4 . (b) Comparison of Max-min downlink
rate for K = 30, L = 10, ρd = 1, C = 5 and α = 4.
approach with power allocation and without power allocation.

This indicates that proper power control can further improve
the performance of the system.
We compare the performance of the Lagrangian-based solu-
tion Algorithm 1 with the proposed SDR + SCA in terms
of sum of downlink achievable rate versus downlink transmit
power in Figure 3 (a). From Figure 3 (a), it can be observed
that the proposed scheduling-based approach outperforms the
Lagrangian-based solution. In fact, there is a performance gain
varying from 9.1% to 11.23% between our the alternative
proposed SDR+SCA and the Lagrangian-based scheme with-
Fig. 3. (a) Average Sum of Downlink Achievable Rate versus Downlink out power allocation and the gain increases from 24.08% to
Transmit Power for K = 20, M = 50, L = 10, ρd = 1, α = 4 and C = 5. 27.31% when power control is taking into account.
(b) Evolution of Sum of Downlink Achievable Rate K = 20, M = 50,
L = 10, α = 4 and C = 5.
Now, we numerically evaluate the convergence of the SCA
approach of the power control problem. The performance is
given in terms of evolution of the sum downlink rate versus
per-user downlink achievable rate for the Proposed SDR + the number of iterations in Figure 3 (b). It can be concluded
SCA + Power and the Lagrangian-based approach in Figure 2. by observing 3 (b) that the proposed SCA-based approach to
It can be seen an improvement of 15% between the Proposed solve the power control problem converges.
SDR+SCA and Lagrangian-based solution. This improve- In Section III, we stated that the Lagrangian-based
ment increases to 42% if power control is performed. Algorithm 1 is generally very slow in terms of convergence
Moreover, there is a gain of 21% between the proposed speed. We verify this assertion by mean of simulations. It is
worth pointing out that simulations were running on a Win- changes of variables:
dows desktop with 4 Intel i7 cores and 8GB of RAM. Given 2
2 3
the same error of tolerance, the simulations revealed that there M N τ ρp βmk N ρd M τ ρp βmk
ak , bk + 1,
is a computation time gain of 85% of the proposed SDR+SCA ρp βmk + 1 ρp βmk + 1
2
as compared to the Lagrangian-based approach. Algorithm 1 τ ρp βmj bk Δkj
converges on average 7 times slower than the Proposed Δkj ρd N M βmk , b̃k , Δkj
ρp βmj + 1 ak ak
SDR+SCA.
We proceed to compare our approach with existing Consequently, the optimization problem (9) is reduced to

scheme [8], that is the canonical form of representing CF K
1
massive MIMO where all users are scheduled simultaneously max ϑk log2 1 + K
ϑ∈{0,1}K b̃k + j=k Δkj ϑj
over the available bandwidth. We implement the random pilot k=1

allocation scheme as well as [8, Algorithm 1] which we term 1
as : Canonical CF+ Pilot allocation and implement the power s.t. Rk ≤ ϑk log2 1 + K , ∀k (A.1)
b̃k + j=k Δkj ϑj
allocation given in [8, Algorithm 2]. Firstly, we compare the
CDF of the per-user downlink rate in Figure 4 (a). It can be which is equivalent to
seen from Figure 4 (a) that our proposed approach outperforms
the existing scheme [8]. Secondly, We provide comparison
K
ϑk
max log2 1+ K
in terms of maximum minimum downlink rate versus the ϑ∈{0,1}K b̃k + Δkj ϑj
j=k
number of APs. It can be observed from Figure 4(b) that our
k=1

ϑk
proposed scheme outperforms existing approaches. There is s.t. Rk ≤ log2 1 + K , ∀k (A.2)
a performance gain varying from 9.1% to 27% between our b̃k + j=k Δkj ϑj
proposed scheduling-based solution SDR + SCA + Power
To prove the NP-hardness, our goal is to introduce a polyno-
and the considered baseline [8]. This comparison leads us to
mial time reduction in order to transform the vertex packing
believe that partitioning users into groups followed by non-
problem into the optimization problem (A.2 ). To proceed with
overlapping group scheduling can palliate the effect of pilot
the proof, we write the following definition
contamination.
Definition 1: Consider an undirected graph G = {V, E},
where V = {1, · · · , |V|} is the set of vertices in G while
VII. C ONCLUSION E = {(j, k) : j ∈ V k ∈ V} denotes the set of edges in G.
An independent set of G is defined as a subset, say G ⊂ G,
In this work, we advocated that efficient radio resource
where no two nodes in G are connected. That is, ∀ vj , vk ∈ G
allocation can lead to diminution of the effect of pilot contam-
implies that (vj , vk ) ∈
/ E. The vertex packing problem refers
ination and can mitigate multi-user interference for a Cell-Free
as the problem of finding an independent set with a predefined
massive MIMO network. We investigated the problem of
size.
sub-frequency bands allocation. Firstly, we locally solved
Let us construct an undirected with |V| = K vertices. For any
the NP-hard problem formulation by mean of Lagrangian-
couple (vj , vk ) ∈ G, let
based method which has substantially low convergence speed.
We also proposed a more appealing alternative method that 1, if vj and vk are connected.
consists of addressing the problem from the perspective of Δkj =
0, otherwise.
users partitioning. We proposed a semidefinite relaxation based
algorithm to efficiently solve the problem. In addition to that, 1

and let b̃k = 2|V| , Rk = R ≤ log2 1 + 2|V| , ∀k. The
we designed a successive convex approximation algorithm transformation of the maximum independent set problem into
to find a stationary point to the problem of sub-frequency problem (A.2 ) is given through the following Lemma.
bands allocation as well as the power control problem. The Lemma 1: Solving the special instance optimization prob-
simulation results demonstrated that the alternative proposed lem (A.2 ) is equivalent to finding a maximum independent set
scheme outperforms the Lagrangian-based solution. Moreover, G of size |G| of the undirected graph G. Moreover, the opti-
the simulation results showed that the proposed scheduling- mal solution s to problem (A.2 ) satisfies log2 1 + 2|V|
based method enables to attain a performance gain compared |G| ≤ s .
to existing schemes. Proof of Lemma 1 is given in Appendix B. Given that the
vertex packing problem is an NP-hard problem, it follows
A PPENDIX A that the optimization (A.2 ) is NP-hard. This concludes our
P ROOF OF P ROPOSITION 1 proof.
To prove that problem (9) is NP-hard, we consider investi-
gating the NP-hardness of a special case. More specifically, we A PPENDIX B
assume that L = 1 and suppose that the number of active users P ROOF OF L EMMA 1
is constrained by the number of orthogonal pilot sequences, i.e We proceed with the proof of Lemma 1 by firstly demon-
K ≤ τ . Before proceeding with the proof, define the following strating the necessary condition. Suppose that there exists an
independent set G of size |G|. Define ϑk , k = 1, · · · , K as where (a) follows from the fact reducing the interference
follow: leads to higher rate. We now consider a feasible solution
D for problem (A.2 ), where D is constructing as follow
1, if k ∈ G .
ϑk = (A.3)
D = (D /S) ∪ S, S includes the same index elements as

0, otherwise.
S, with only one element being 1 and all other elements 0.
Since no two nodes in G are connected, that is Δkj = Without loss of generality, suppose that the only element with
0, ∀
(k, j) ∈ G with k = j. It follows that Rk ≤ 1 is k0 . The corresponding objective function attained with the
feasible set D is
ϑk
log2 1 + b̃ + K Δ ϑ
k j=k

= log2 1 + 2|V| . Hence, solu-
kj j
⎛ ⎞
tion provided in (A.3 ) is feasible for the optimization problem ⎜ ⎟
(A.2 ) and the corresponding objective function is
sD = log2 ⎜
⎝1 +
ϑk
⎟=
Δkj ϑj ⎠

ϑk
k∈D b̃k +

j∈D,j=k

k∈(G/S)∪{k0 }
⎛ ⎞
s = log2 1 +
k∈G b̃k + j∈G,j=k Δkj ϑj ⎜ 1 ⎟
log2 ⎜ ⎟
= log2 1 + 2|V| = |G| log2 1 + 2|V| (A.4) ⎝1 + b̃k + Δkj · 1⎠
k∈G

j∈(G/S)∪{k0 },j=k
with s ≤ s . To prove the sufficient condition, we suppose
1
= log2 1+ 1 +
that problem (A.2 ) possesses an optimal solution. Denote the
optimal objective function value s and the optimal set as D .
2|V| k∈G/S
Define the following set 1
log2 1+
G= {vk |ϑk = 1, 1 ≤ k ≤ K} ⊆ V (A.5)
1

2|V|
+ j∈(G/S)∪{k0 },j=k Δkj · 1

(b) 1
|V|
It can be inferred from (A.5 ) that set G encompasses all > log2 2 + log2 1 + 1
nodes or elements of D that have value 1. Our goal is k∈G/S

2|V|
+ j∈G,j=k Δkj · 1

to demonstrate that the set G is an independent set. Before 1
starting with the proof, we have the following claim > |S| + log2 1 + 1
Claim 1: The graph characterized by combining all nodes
k∈G/S 2|V|
in G is an incomplete graph. > s (A.7)
Claim 1 is proven in Appendix C. Given from Claim 1
that the graph formed by the nodes in G is not be complete The inequality (b) comes from the monotonicity of the
logarithm function, that is log2 2|V| < log2 1 + 2|V| .

graph, we need to show that no two nodes in the optimal
set are connected. This is done by contradiction. Suppose not, Equation (A.7 ) leads to sD > s . This result contra-
namely there exists a subset S ⊂ G with at least two nodes that dicts with the fact that s is the global optimum for prob-
are connected. Under this consideration, the optimal objective lem (A.2 ). Therefore, it does not exist subsets in G with
function value s is given by a least two adjacent nodes which implies that G is an
independent set.
ϑk

s = log2 1 + A PPENDIX C
k∈D b̃k + j∈D ,j=k Δkj ϑj
P ROOF OF C LAIM 1
1
= log2 1 + Suppose that the graph formed by all nodes in G is

k∈G
b̃k + j∈G,j=k Δkj · 1 complete. Then, the optimal objective value function s is
given by
1
= log2 1 + 1
+ j∈S,j=k 1 · 1 s =
ϑk
k∈S

2|V|

log2 1 +
k∈D b̃k + j∈D ,j=k Δkj ϑj
1
+ log2 1 + 1
|V| + j∈G,j=k Δkj · 1 = log2 1 +
1
(a)
k∈G/S

2

k∈G

b̃k + j∈G,j=k Δkj · 1

1
≤ |S| log2 1 + 1 + log2 1 2|V|
2|V|
+1
k∈G/S

= |G| log2 1 +
1 + (|G| − 1)2|V|
(A.8)
1 Constructing a feasible set D constituting only of one active

+ 1
2|V|

node, say k1 ∈ G, i.e D = (D /G) ∪ k1 . The associated
1 objective function s̃ is given by s̃ = log2 1 + 11 =
< |S| + log2 1 + 1

2|V|
k∈G/S 2|V|
+ j∈G,j=k Δkj · 1 log2 1 + 2|V| . Let us

verify the sign of s̃ − s

=

|V|
|V| 2
(A.6) log2 1 + 2 − |G| log2 1 + 1+(|G|−1)2|V| . Towards that

2|V| 2|V|
end, define the functions f˜(x) = x log2 1 + 1+(x−1)2 |V| value is given by 2 log2 1 + 1+2 |V| . Consequently, f˜1 (x) is

and f˜1 (x) = log2 1 + 2|V| − f˜(x) for 2 ≤ x ≤ |V|. The an increasing function of x and its minimum is attained at
derivative of f˜(x) with respect to x is given by
|V| 2|V|
log2 1 + 2 − 2 log2 1 +
∂ f˜(x) 2|V| 1 + 2|V|
= log2 1 +
∂x 1 + (x − 1)2|V| 1 + 3 · 2 + 22+|V| + 23|V|
|V|
= log2 > 0 (A.11)
−x22|V| 1 + 22+|V| + 22+2|V|
2
1 (1+(x−1)2|V| )
+ 2|V|
which means that s̃ − s > 0. This leads to a contradiction
log(2) 1 +
1+(x−1)2|V| since s is the global maximum. Therefore, the graph formed
2|V| by the nodes in G cannot be complete.
= log2 1 +
1 + (x − 1)2|V|
1 x22|V| R EFERENCES
−
log(2) 1 + x2|V| 1 + (x − 1)2|V|
[1] S. E. Hajri, J. Denis, and M. Assaad, “Enhancing favorable propagation
in cell-free massive MIMO through spatial user grouping,” in Proc. IEEE
The second derivative of f˜(x) with respect to x can be found 19th Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC),
as follow Jun. 2018, pp. 1–5.
[2] Cisco White Paper. (Feb. 2019). Cisco Visual Networking Index:
−22|V| Global Mobile Data Traffic Forecast Update, 2017–2022. [Online].
∂ 2 f˜(x) 1 (1+(x−1)2|V| )2 Available: https://www.cisco.com/c/en/us/solutions/collateral/service-
=
∂x2 log(2) 1 + 2|V| provider/visual-networking-index-vni/white-paper-c11-738429.pdf
1+(x−1)2|V| [3] Ericsson White Paper. (2014). 5G Radio Access. [Online]. Available:
1 22|V| (2|V| − 1 + x2 22|V| ) https://pdfs.semanticscholar.org/9a06/bcadf0f4e3770260e0193746d5365
+ b6c9114.pdf
log(2) 1 + x2|V| 2 1 + (x − 1)2|V| 2 [4] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta, and
F. Tufvesson, “Scaling up MIMO: Opportunities and challenges with
1 22|V| very large arrays,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60,
=− Jan. 2013.
log(2) 1 + x2|V| 1 + (x − 1)2|V| [5] F. Boccardi, R. W. Heath, Jr., A. Lozano, T. L. Marzetta, and P. Popovski,
1 22|V| (2|V| − 1 + x2 22|V| ) “Five disruptive technology directions for 5G,” IEEE Commun. Mag.,
+ vol. 52, no. 2, pp. 74–80, Feb. 2014.
log(2) 1 + x2|V| 2 1 + (x − 1)2|V| 2 [6] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-
bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9,
1 22|V| (21+|V| − 2 − x2|V| ) no. 11, pp. 3590–3600, Nov. 2010.
=
log(2) 1 + x2|V| 2 1 + (x − 1)2|V| 2 [7] E. Nayebi, A. Ashikhmin, T. L. Marzetta, and H. Yang, “Cell-free
massive MIMO systems,” in Proc. 49th Asilomar Conf. Signals, Syst.
Comput., Nov. 2015, pp. 695–699.
By evaluating the sign of the second derivative, it is straightfor- [8] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta,
2
ward demonstrating that it is always negative for x ≥ 2 − 2|V| . “Cell-free massive MIMO versus small cells,” IEEE Trans. Wireless
∂ f˜(x) Commun., vol. 16, no. 3, pp. 1834–1850, Mar. 2017.
Therefore, the maximum value of ∂x is attained at x = 2 [9] E. Nayebi, A. Ashikhmin, T. L. Marzetta, H. Yang, and B. D. Rao,
and is equal to “Precoding and power optimization in cell-free massive MIMO sys-
tems,” IEEE Trans. Wireless Commun., vol. 16, no. 7, pp. 4445–4459,
∂ f˜(x) ∂ f˜(x) 2|V| Jul. 2017.
max = |x=2 log2 1 + [10] H. Q. Ngo, L. N. Tran, T. Q. Duong, M. Matthaiou, and E. G. Larsson,
2≤x≤|V| ∂x ∂x 1 + 2|V| “On the total energy efficiency of cell-free massive MIMO,” IEEE Trans.
1 2(1+2|V|) Green Commun. Netw., vol. 2, no. 1, pp. 25–39, Nov. 2017.
− (1+|V|)
(A.9) [11] S. Zhou, M. Zhao, X. Xu, J. Wang, and Y. Yao, “Distributed wireless
log(2) ((1 + 2 )(1 + 2|V| ) communication system: A new architecture for future public wireless
y (1+2y)
access,” IEEE Commun. Mag., vol. 41, no. 3, pp. 108–113, Mar. 2003.
2 2
Define f˜2 (y) = ln(1 + 1+2 y ) − (1+2(1+y) )(1+2y ) , y ≥ 2. The [12] E. Hossain, M. Rasti, H. Tabassum, and A. Abdelnasser, “Evolu-
tion toward 5G multi-tier cellular wireless networks: An interference
˜
derivative of f2 (y) is computed as management perspective,” IEEE Wireless Commun., vol. 21, no. 3,
pp. 118–127, Jun. 2014.
∂ f˜2 (y) 2y log(2) 22y+1 log(2) (3 · 2y + 2) [13] R. F. Ustok, P. A. Dmochowski, P. J. Smith, and M. Shafi, “Interference
= − 2 2 cancellation with jointly optimized transceivers in multiuser multicellu-
∂y (1 + 21+y ) (1 + 2y ) (1 + 2y ) (1 + 21+y )
lar networks,” IEEE Trans. Veh. Technol., vol. 67, no. 8, pp. 7219–7229,
log(2)2y 1 + 2y − 2y+1 − 22y+2 Aug. 2018.
= 2 2 ≤0 [14] J. Ma, S. Zhang, H. Li, N. Zhao, and V. C. M. Leung, “Interference-
(1 + 2y ) (1 + 21+y ) alignment and soft-space-reuse based cooperative transmission for multi-
cell massive MIMO networks,” IEEE Trans. Wireless Commun., vol. 17,
Hence, f˜2 (y) is a decreasing function of y and its maximum no. 3, pp. 1907–1922, Mar. 2018.
[15] H. Xie, F. Gao, S. Zhang, and S. Jin, “A unified transmission strategy for
value is attained at y = 2, i.e TDD/FDD massive MIMO systems with spatial basis expansion model,”
IEEE Trans. Veh. Technol., vol. 66, no. 4, pp. 3170–3184, Apr. 2017.
˜ 4 1 32
max f2 (y) = log2 1 + − = −0.1779 < 0 [16] C. Sun, X. Gao, S. Jin, M. Matthaiou, Z. Ding, and C. Xiao, “Beam divi-
2≤y 5 log(2) 45 sion multiple access transmission for massive MIMO communications,”
(A.10) IEEE Trans. Commun., vol. 63, no. 6, pp. 2170–2184, Jun. 2015.
[17] L. Sanguinetti, A. L. Moustakas, and M. Debbah, “Interference manage-
ment in 5G reverse TDD HetNets with wireless backhaul: A large system
Therefore, by combining (A.9 ) and (A.10 ), it can be con- analysis,” IEEE J. Sel. Areas Commun., vol. 33, no. 6, pp. 1187–1200,
cluded that f˜(x) is a decreasing function of x. Its maximum Jun. 2015.
[18] A. Adhikary, J. Nam, J.-Y. Ahn, and G. Caire, “Joint spatial division and [36] C. Helmberg, F. Rendl, R. J. Vanderbei, and H. Wolkowicz, “An interior-
multiplexing—The large-scale array regime,” IEEE Trans. Inf. Theory, point method for semidefinite programming,” SIAM J. Optim., vol. 6,
vol. 59, no. 10, pp. 6441–6463, Oct. 2013. no. 2, pp. 342–361, 1996.
[19] E. Bjornson, J. Hoydis, and L. Sanguinetti, “Massive MIMO has [37] M. X. Goemans and D. P. Williamson, “Improved approximation algo-
unlimited capacity,” IEEE Trans. Wireless Commun., vol. 17, no. 1, rithms for maximum cut and satisfiability problems using semidefinite
pp. 574–590, Jan. 2018. programming,” J. ACM, vol. 42, no. 6, pp. 1115–1145, Nov. 1995.
[20] D. Neumann, M. Joham, and W. Utschick, “Covariance matrix estima- [38] A. Beck, A. Ben-Tal, and L. Tetruashvili, “A sequential parametric
tion in massive MIMO,” IEEE Signal Process. Lett., vol. 25, no. 6, convex approximation method with applications to nonconvex truss
pp. 863–867, Jun. 2018. topology design problems,” J. Global Optim., vol. 47, no. 1, pp. 29–51,
[21] T. C. Mai, H. Q. Ngo, M. Egan, and T. Q. Duong, “Pilot power control May 2010.
for cell-free massive MIMO,” IEEE Trans. Veh. Technol., vol. 67, no. 11, [39] M. Razaviyayn, M. Hong, and Z.-Q. Luo, “A unified convergence
pp. 11264–11268, Nov. 2018. analysis of block successive minimization methods for nonsmooth
[22] G. T. version Release 12. (2014). LTE; Scenarios and Require- optimization,” SIAM J. Optim., vol. 23, no. 2, pp. 1126–1153, Jan. 2013.
ments for Small Cell Enhancements for E-UTRA and E-UTRAN.
[Online]. Available: https://www.etsi.org/deliver/etsi_tr/136900_136999/
136932/12.01.00_60/tr_136932v120100p.pdf
[23] Z.-Q. Luo, W.-K. Ma, A. So, Y. Ye, and S. Zhang, “Semidefinite
relaxation of quadratic optimization problems,” IEEE Signal Process.
Mag., vol. 27, no. 3, pp. 20–34, May 2010.
[24] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Juwendo Denis (Member, IEEE) received the B.E. degree in electrical
Theory. Upper Saddle River, NJ, USA: Prentice-Hall, 1993. engineering and the M.Sc. degree from the Institute of Communications
[25] E. Bjornson, J. Hoydis, M. Kountouris, and M. Debbah, “Massive Engineering, National Tsing Hua University, Hsinchu, Taiwan, in 2011 and
MIMO systems with non-ideal hardware: Energy efficiency, estima- 2013, respectively, and the Ph.D. degree in communications engineering
tion, and capacity limits,” IEEE Trans. Inf. Theory, vol. 60, no. 11, from the CEDRIC Research Laboratory, Conservatoire National des Arts
pp. 7112–7139, Nov. 2014. et Métiers, Paris, France, in 2016. He was a Postdoctoral Researcher with
[26] E. Hossain, L. B. Le, and D. Niyato, Radio Resource Management in CentraleSupelec, Gif-sur-Yvette, France and a Former Post-Doctoral Research
Multi-Tier Cellular Wireless Networks. Hoboken, NJ, USA: Wiley. Fellow with the Electrical and Computer Engineering Department, University
[27] G. Interdonato, H. Q. Ngo, P. Frenger, and E. G. Larsson, “Downlink of Illinois at Chicago. He is currently a Post-Doctoral Fellow with the
training in cell-free massive MIMO: A blessing in disguise,” IEEE Trans. School of Engineering and Applied Sciences (SEAS), Harvard University. His
Wireless Commun., vol. 18, no. 11, pp. 5153–5169, Nov. 2019. current research interests include cross-layer optimization, resource allocation,
[28] M. R. Garey and D. S. Johnson, Computers and Intractability; interference management, channel estimation and users scheduling for 5G, and
A Guide to the Theory of NP-Completeness. New York, NY, USA: beyond wireless communications systems
W. H. Freeman, 1990.
[29] A. Khalili, S. Zarandi, and M. Rasti, “Joint resource allocation and
offloading decision in mobile edge computing,” IEEE Commun. Lett.,
vol. 23, no. 4, pp. 684–687, Apr. 2019.
[30] E. Che, H. D. Tuan, and H. H. Nguyen, “Joint optimization of
cooperative beamforming and relay assignment in multi-user wireless Mohamad Assaad (Senior Member, IEEE) received the M.Sc. and Ph.D.
relay networks,” IEEE Trans. Wireless Commun., vol. 13, no. 10, degrees in telecommunications from Telecom ParisTech, France, in 2002 and
pp. 5481–5495, Oct. 2014. 2006, respectively. Since 2006, he has been with the Telecommunications
[31] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.: Department, CentraleSupélec, where he is currently a Professor. He is also
Cambridge Univ. Press, 2004. a Researcher with the Laboratoire des Signaux et Systèmes (L2S, CNRS).
[32] D. P. Bertsekas, Nonlinear Programming. Belmont, MA, USA: He has co-authored one book and more than 120 publications in journals
Athena Scientific, 1999. and conference proceedings. He has given in the past successful tutorials
[33] S. Boyd, L. Xiao, and A. Mutapcic, “Subgradient methods,” Stan- on 5G systems at various conferences including IEEE ISWCS’15 and IEEE
ford Univ., CA, USA, Tech. Rep., 2004. [Online]. Available: WCNC’16 conferences. His research interests include 5G and beyond systems,
https://web.stanford.edu/class/ee392o/subgrad_method.pdf fundamental networking aspects of wireless systems, MIMO systems, resource
[34] L. Vandenberghe and S. Boyd, “Semidefinite programming,” SIAM Rev., optimization, and machine learning in wireless networks. He also serves
vol. 38, no. 1, pp. 49–95, Mar. 1996. regularly as a TPC member or the TPC co-chair for top-tier international
[35] M. Grant and S. Boyd. (Mar. 2014). CVX: MATLAB Software for conferences. He is currently an Editor for the IEEE W IRELESS C OMMU -
Disciplined Convex Programming, Version 2.1. [Online]. Available: NICATIONS L ETTERS and the Journal of Communications and Information
http://cvxr.com/cvx Networks.

Improving Cell-Free Massive MIMO Networks Performance A User Scheduling Approach

Uploaded by

Copyright:

Available Formats

Improving Cell-Free Massive MIMO Networks Performance A User Scheduling Approach

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

Improving Cell-Free Massive MIMO Networks Performance A User Scheduling Approach

Uploaded by

Copyright:

Available Formats

7360 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO.

11, NOVEMBER 2021