See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/327136847

Leakage Based Users Selection for Hybrid Beamforming in MillimeterWave

MIMO

Conference Paper · September 2018

DOI: 10.1109/PIMRC.2018.8580820

CITATIONS READS
3 114

4 authors:

Mohamed Shehata Matthieu Crussière

Mitsubishi Electric Corporation Institut National des Sciences Appliquées de Rennes
11 PUBLICATIONS 33 CITATIONS 204 PUBLICATIONS 954 CITATIONS

SEE PROFILE SEE PROFILE

Maryline Helard Patrice Pajusco

Institut National des Sciences Appliquées de Rennes, IETR Institut Mines-Télécom
268 PUBLICATIONS 1,301 CITATIONS 123 PUBLICATIONS 641 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Power Estimation Techniques on FPGAs for Digital Signal Processing Applications View project

High rate WPAN View project

All content following this page was uploaded by Mohamed Shehata on 05 February 2019.

The user has requested enhancement of the downloaded file.

 2018 IEEE 29th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC)
Leakage Based Users Selection for
Hybrid Beamforming in MillimeterWave MIMO
M. Shehata1, M. Crussière1, M. Hélard1 and P. Pajusco2
1
Univ Rennes, INSA Rennes, CNRS, IETR - UMR 6164, F-Rennes, France
2
Institut Mines-Telecom, Telecom Bretagne, CNRS UMR 6285 Lab-STICC, Brest, France
Abstract—Millimeter-Wave (mmWave) channels are
characterized as being sparse and Line of Sight (LoS)
dominated. Applying hybrid beamforming in mmWave
systems guarantees Multi-User (MU) transmission with low
hardware requirements. Zero Forcing (ZF) precoding can
be utilized at the digital layer of the hybrid beamformer to
mitigate the Inter User Interference (IUI). However, due to
the sparse nature of the mmWave channels, a ZF precoder
may provide limited performance due to the highly correlated
MU channel matrix. The aim of this paper is hence to address
the interference mitigation problem in the analog domain
before handling it at the ZF digital level. Signal to Leakage
and Noise Ration (SLNR) based users selection algorithm is
introduced for maximizing the Spectral Efficiency (SE) of the
system with low complexity requirements. We finally highlight
by simulations the potential gains achieved by the introduced
interference aware users selection approach.
Index Terms—Multi-User MIMO, mmWave, Users
Selection, Condition Number, Signal to Leakage and Noise
Ratio (SLNR), Hybrid Beamforming
I. INTRODUCTION
The millimeter-Wave (mmWave) technology has emerged
as one of the main building blocks for the fifth generation (5G)
wireless communication systems [1]. A clear rationale for this is
the ability of mmWave systems to extend the utilized spectrum
grid, therefore allowing for much larger available bandwidths
that can be utilized for extending the network’s coverage and
boosting the data rate. But beyond that, due to the small
wavelength used, the mmWave systems offer the other potential
advantage of allowing for massive Multiple Input Multiple
Output (MIMO) antenna arrays with reasonable form factors.
Hence, such systems are expected to unleash the gains of
massive MIMO arrays and pencil beamforming approaches at
both the Base Station (BS) and User Equipment (UE) sides [2].
On the other side, implementing mmWave systems means
dealing with a lot of practical limitations at the hardware
level, specifically in the analog part and in the mixed analog
to digital devices. Besides, the channel characteristics are
completely different compared to these at lower range of
frequencies. Concerning the hardware limitations, applying
the digital beamforming/combining for massive MIMO arrays
at mmWaves requires one Radio Frequency (RF) chain and
one Digital to Analog Converter (DAC) / Analog to Digital
Converter (ADC) per antenna. This is limited by the existing
hardware components at such frequencies, which are quite ex-
pensive, complex and power hungry [3]. Moreover, the wireless
978-1-5386-6009-6/18/$31.00 ©2018 IEEE 1144
mmWave propagation channel imposes more challenges for
having robust communication links between the transmitter
and the receiver. The two main limitations imposed by the
mmWave channel are the high path-loss due to the high carrier
frequency and the blockage due to the small wavelength [4].
Henceforth, in order to tackle both the high path-loss limitation,
together with the hardware limitation, hybrid analog-digital
beamforming recently emerged as a potential solution [5].
Hybrid beamforming enables employing massive antenna
arrays, with much less RF chains and mixed analog-digital
devices. Through such an approach, it is possible to tackle the
high channel path-loss by taking the advantage of the massive
antenna array, while respecting some hardware limitations. In
addition, hybrid beamforming schemes still allow for serving
Multi-User (MU) scenarios through Spatial Division Multiple
Access (SDMA) since multiple RF chains are considered.
Hybrid beamforming focus mainly on serving multiple
streams for Single User or MU scenarios in the spatial domain.
Many work in the literature recently focused on MU hybrid
beamforming [6]–[8] by decoupling the beamforming design
problem into two problems in the analog and digital domains
that can be addressed separately or jointly. In [8], the authors
decouple the two problems and solve the analog problem using
the Line of Sight (LoS) beamsteering to maximize the received
signal by each UE. Then, in the digital part the MU interference
is tackled through applying the Zero Forcing (ZF) digital beam-
forming. This framework is crucial to consider at mmWave fre-
quencies, due to the high LoS dominated channels and clustered
UEs deployments planned for mmWave small cells leading to
high spatial correlation of the MU channel. These high spatially
correlated channel matrices turns out to be ill-conditioned for in-
version, thus limiting the performance of the ZF beamforming.
It is then understood that users selection algorithms are
crucial for mmWave hybrid beamforming techniques that rely
on ZF at the digital layer. A lot of works have been dedicated to
users selection for fully digital ZF precoding [9]–[11]. In most
of the contributions, the selection was mainly done based on the
orthogonality of the channels of the selected UEs using linear
algebraic approaches for the ’orthogonality’ terminology. Other
works translated the orthogonality metric into other simpler, yet
similar metrics to relax the complexity constraints imposed in
the orthogonality based algorithms such as: the angle between
subspaces [12], [13], also based on chordal distance in [14] and
based on matrix determinant for representing the orthogonality
in [15]. Complementary to ’orthogonality’ based approaches,
the authors in [16] introduced a statistical based terminology
which is the ’spatial correlation’ expressed by the condition
number of the MU channel matrix. In this work, it was shown
that imposing the condition number as the constraint in the
Semi-Orthogonal User Selection (SUS) algorithm [9] could
improve the performance of the ZF beamformer.
In this paper, we introduce a novel UEs’ selection algorithm
for hybrid beamforming systems relying on the condition
number as a user selection metric. The condition number is
chosen as our metric for evaluating the correlation between the
UEs. The main idea of the proposed algorithm is to decrease
the correlation of the MU channel matrix by choosing the
UEs which have the lowest condition number of their mutual
MU channel matrix. The introduced algorithm is derived
analytically and compared with the baseline SUS technique
[9]. We show that our approach is particularly adequate when
applied to sparse channels such as mmWave channels, since
the condition number can in that case be easily approximated.
Accordingly, as proposed in this paper, it is then possible to
design a suboptimal yet efficient user selection algorithm based
on maximizing the Signal to Leakage and Noise ratio (SLNR)
of the selected UEs. We show with complexity analysis and
simulation results the potentials of this sub-optimum approach
in mmWave LoS dominated environments.
The contribution of this paper can be summarized as follows:
• Interference aware analog layer: The main objective in this
paper is to involve the analog layer in hybrid beamforming,
in tackling the interference, unlike [8]. Due to the analog
constraints forced on the analog domain, (specifically
constant amplitude phase-shifters), the analog beamforming
and/or power allocation solutions are not capable of taking
into account the interference mitigation in the analog part.
Henceforth, we choose the interference aware UEs selection
approaches to address that problem in the analog domain,
since it isn’t affected by the analog hardware constraints,
yet it enhances the performance of the ZF precoding in
the digital layer significantly as will be shown later in the
simulation results.
• Analytically deriving novel condition number based UEs
selection.
• Introducing a low complexity sub-optimum UEs selection
algorithm based on the SLNR metric: Maximizing the
SLNR has been utilized before for precoding [17] and power
allocation [18] [19], serving as a low complexity approach
for maximizing the SE of the system. However, up to the
authors knowledge, it wasn’t utilized for UEs selection
before. Thus, in this paper, we apply this metric for UEs
selection, since it acts as an attractive candidate for mmWave
systems UEs selection problem.
II. SYSTEM AND CHANNEL MODEL
Throughout the paper a MU Multiple Input Single Output
(MISO) is considered in the downlink Orthogonal Frequency
Division Multiplexing (OFDM) narrow band scenario. The BS
is equipped with M transmit antennas according to a Uniform
Linear Array (ULA) architecture. The BS has a set of available
UEs to serve denoted as K, however it can serve a maximum
number of UEs |C|, where C is the set of the selected UEs to be
1145
served. The cardinality of the served set of UEs is limited to the
number L of RF chains available in the BS, such that |C|≤L.
In the sequel, we assume |C| = L, to achieve the maximum
multiplexing and coverage gains (degrees of freedom). At the
transmitter side the hybrid beamforming strategy is applied,
while at the receiver side each UE k ∈ K has only a single
receive antenna and thus it can’t apply combining schemes
and can be served only by a single stream.
The streams received by the set of selected UEs C,
r=[R1,R2,...,R|C|]T ∈C|C|×1 are expressed as:
r=Hx+n (1)
where n = [N1,N2,...,N|C|]T ∈ C|C|×1 is a vector denoting
the independent and identically distributed (i.i.d.) Additive
White Gaussian Noise (AWGN), where n ∼ N (0, σ2), and
σ2 is the noise variance. H∈C|C|×M represents the channel
matrix between the |C| single receive antenna selected UEs
and the M transmit antennas BS.
The M precoded symbols transmitted from the BS are
denoted by the vector x = [X1,X2,...,XM ]T ∈ CM×1 and
represented as:
1
x= √ Fs (2)
ζ
H
such that ζ =E[(T r(F F)] is the power normalization factor,
F is the precoding matrix which will be described in more
details in the next section. s = [S1,S2,...,S|C|]T ∈ C|C|×1,
represents the vector of the transmitted symbols.
The channel model used throughout our work corresponds
to the sparse geometric channel model introduced in [8], [20]. It
is characterized by a limited number of scatterers contributing
by a single propagation path each, to model the mmWave
channel. For each UErk, the channel vector hk is denoted as:
Pk
MX
hk = αk,paHt (φk,p ) (3)
Pk p=1
such that the complex amplitude of propagation path p which
is one of the received Pk paths by UE k is denoted as αk,p
given that E[|αk,p|2]= α̂. The Angle of Departure (AoD) for
each path p for UE k is represented as φk,p and is uniformly
∆φ]
distributed in the interval [− ∆φ
2 , 2 ], where ∆φ is the angular
spread. The transmit array steering vector is represented as
at(φk,p). As we consider an ULA array at the BS, the steering
vector at(φk,p) is defined as,
1
at(φk,p)= √ [1,ejβ(φk,p ),...,ej(M−1)β(φk,p )]T (4)
M
where β(φk,p) is defined as
2π
β(φk,p)= dsin(φk,p) (5)
λ
such that λ represents the wavelength of the signal and d is
the inter-element antenna spacing. Finally, the MU channel
matrix H∈C|C|×M can be written as
H=[hT1 ,hT2 ,...,hT|C|]T (6)
III. BEAMFORMING TECHNIQUE
In this section we introduce the hybrid beamforming
strategy considered throughout the paper. The aim of this
strategy concatenates two objectives which are: maximizing
the Signal to Noise Ratio (SNR) in the analog domain for
each UE k separately and minimizing the MU interference in
the digital domain. The RF and Base Band (BB) architectures 1) Problem Formulation for Digital ZF: Given a digital
are decoupled as in [8]. Then the problem of the sum SE ZF precoding system is considered, the sum SE maximization
maximization is formulated as: problem is formulated as follows:
|C|
γ
X
FRF ,FBB = argmax log2(1+SINRk ) C =argmaxlog2det(I|C| + |HC FBB,C |2) (15)
FRF ,FBB (7) C |C|ζC
k=1
2 where γ represents the Signal to Noise Ratio (SNR) and
s.t. kfRF,kk∈F,kFRF FBB kF =|C|
δT
ζC represents the ZF normalization factor. Applying ZF as the
2
|h F f
|C| k RF BB,k | digital precoding ensures that |HC FBB,C | = I|C|. Thus, the
SINRk = δT P 2 2
(8) problem can be dually approached as choosing the optimal
|C| i6=k |hk FRF fBB,i | +σ
such that F represents a codebook with all the possible set of UEs C that minimizes the normalization factor ζC .
analog beamformers based on the hardware constraints and δT The normalization factor for digital ZF precoder ζC is
represents the total transmit power. The hybrid beamforming calculated as follows [21]:
|C|
problem is decoupled into two problems, one to design the ana- X 1
ζC = (16)
log beamformer and the other is to design the digital one. The λk
k=1
analog problem is solved for fRF,k for each UE separately, but where λk represent the k−th eigen-value of the propagation
yet simultaneously and it is formulated as (here the UE index k channel HC .
is omitted for readability, since each UE is tackled separately): 2) Problem Formulation for Hybrid Beamforming (LoS
(P ) Beamsteering - ZF): Here we reformulate the problem of
X
2 H H the sum SE maximization given the analog beamformers are
f =argmaxλmax |αp| at(φp) ff at(φp)
f (9) already calculated and the equivalent channel ĥk is available at
p=1
the BS for each UE k. Then formulating the problem as in Equa-
s.t. kfk=1 tion 13, with expressing the constraint kFRF FBB kF 2
=|C| in
The steering vector is calculated based on the knowledge the objective function and solving it for the best UEs set to be
of the best path AoD φk,pLoS as follows: selected for maximizing the sum SE can be written as:
fLoS =at(φpLoS ) (10) γ
C =argmax log2det(I|C| + |ĤC FC |2) (17)
Thus, the equivalent channel for each UE k, ĥk is formed C |C|ζC
Lemma 1: The normalization factor for hybrid beamforming
before the digital layer processing as follows:
based on LoS beamsteering in the analog domain and ZF
precoder in the digital domain in a pure LoS channel, with
ĥk =hk FRF , (11)
assuming that the channel amplitudes α are normalized, ζC
where, hk ∈C1×M ,FRF ∈CM×L,ĥk ∈C1×L, then the MU is calculated as follows:
equivalent channel Ĥ can be denoted as
|C|
X 1
Ĥ= [ĥT1 ,...,ĥT|C|] (12) ζC = (18)
λk
Hence, the sum SE problem can be formulated as: k=1
The assumption that α is normalized is just taken into
|C| δT 2 account in the derivation, since in pure LoS it just applies linear
|C| |ĥk fBB,k |
X
FBB =argmax log2(1+ δ P ) scaling for the steering vectors of the UEs. Thus, it is clear
2 2
i6=k |ĥk fBB,i | +σ
T
FBB
k=1 |C| that in both digital and hybrid beamforming cases, the sum
2
s.t.kFRF FBB kF =|C| SE is controlled by the propagation channel’s eigen-values.
(13) The proof is given in Appendix A. Therefore, the sum SE
Given the fact that, the per UE SNR maximization problem maximization problem can be reformulated, for a constant
was already tackled in the analog domain, ZF is utilized in number of selected UEs |C|=L as follows:
the digital domain to suppress the MU Inter User Interference
(IUI). Thus, the digital beamformer is given as: C =argmin ζC (19)
C

FBB =FZF = ĤH (ĤĤH )−1 (14)
IV. USER GROUPING STRATEGIES
A. Problem Formulation and Analytical Analysis
In this paper we try to involve the analog domain in
considering the interference/correlation between the selected
UEs, since the analog beamforming don’t handle this task and
relies on the digital precoding stage. This is achieved through
applying spatial correlation minimization (interference aware)
UEs selection.
B. Condition Number based User Selection
The normalization factor of the hybrid beamforming ζC
is limited by the minimum eigen-value of the propagation
channel matrix HC or equivalently the minimum eigen-value
of the equivalent channel ĤC from Equations 18 and 32.
Therefore, the problem in Equation 19 can be reformulated
again as maximizing the minimum eigen-value of the
equivalent channel as follows:
C =argmax λmin (20)
C

1146
 where λmin is calculated as λmin =min{λ1,...,λ|C|}. Con- Algorithm 1 Condition Number Based Selection
sequently, this can be directly translated to minimizing the con- 1) Initialize i=1, C0 =Φ, T0 =K, β ≥1
dition number ρ of the MU equivalent channel ĤC as follows: 2) While (i≤min(|Ti−1|, L)) do
3) if (i=1)
C =argmin ρ(ĤC ) (21) 4) find
C k =argmax ĥk
such that the condition number ρ(ĤC ) is calculated as: k∈Ti−1

r 5) else
λmax find
ρ(ĤC )= (22)
λmin k =argmin ρ(ĤCi−1 ∪k )
where λmax is the maximum eigen-value of the equivalent k∈Ti−1
channel and is calculated as λmax =max{λ1,...,λ|C|} 6) end if
7) Ci =Ci−1 ∪k
1) Proposed Algorithm: After formulating the problem 8) Ti ={r ∈Ti−1, r ∈C
/ i, ρ(Ĥr∪Ci )<β }
to minimizing the condition number of the MU equivalent 9) i=i+1
channel ρ(ĤC ), here we propose an algorithm to select the 10) end
UEs based on this objective.
Initially, the first selected UE in the set C is the UE that
has the maximum equivalent channel norm ĥk . Then
H
the algorithm iterates with maximum number of iterations |ĥr Vi−1Vi−1 ĥH
k |
Ti ={r ∈Ti−1,r ∈C
/ i, }<α (24)
(maximum number of UEs selected) equals to the number of kĥr kkĥk Vi−1k
the RF chains L. where Vi−1 is a matrix, which its columns form an
Within each iteration i a filtration process is done to filter the orthonormal basis for the null space of the equivalent channel
UEs and find the candidate set Ti based on a certain correlation ĤCi−1 at the iteration i−1 and is obtained from the SVD of
threshold, defined by a threshold condition number β ≥1. This ĤCi−1 as follows:
threshold β represents a trade-off parameter between the UEs
1 0
diversity gain which can be maximized in case β →inf and the ĤCi−1 =Ui−1Di−1[Vi−1 Vi−1 ]H
(25)
low MU correlation gain maximized in case β →1. In this pa- Vi−1 =Vi−1 0
per we consider serving |C|=L UEs, thus we set β →inf. How- In case α = 0, this represents full (strict) channel
ever, we introduce the threshold β here for completeness and orthogonality scenario, and increasing α relaxes the channel
generic introduction of the algorithm. Then, in each iteration the orthogonality constraint. Again, throughout the paper we
selected UE k is selected such that to minimize the condition use only the full multiplexing case (α = 1). Then, in each
number of the MU equivalent channel together with the pre- iteration the selected UE k is the one with the maximum
viously selected UEs in the previous iterations Ci−1 as follows: channel component orthogonal on the subspace spanned by
the interference (previously selected UEs Ci−1) as follows:
k =argmin ρ(ĤCi−1 ∪k ) (23)
k∈Ti−1 k =argmax ĥk Vi−1 (26)
The proposed algorithm is summarized in Algorithm (1). k∈Ti−1
2) Complexity Analysis: In order to evaluate the proposed The SUS algorithm is summarized in Algorithm (2).
algorithm and have a fair comparison with the other UEs
selection algorithms, we analyze the worst case (upper bound) Algorithm 2 Semi-Orthogonal Users Selection [9], [10]
computational complexity. The computational complexity 1) Initialize i=1, V0 =I, C0 =Φ, T0 =K, α∈[0,1]
analysis is based on the maximum number of Singular Value 2) While (i≤min(|Ti−1|, L)) do
Decompositions (SVDs) needed in our algorithm, since SVD is 3) find
the most complex operation applied in the proposed algorithm k =argmax ĥk Vi−1
with a complexity of O(nm2 + n3) for an m × n matrix. k∈Ti−1
Henceforth, the computational complexity in the worst case (in 4) Ci =Ci−1 ∪k
case β →inf) is of the order O(|K|L3), given that K is the set 5) Vi =P⊥(ĤCi ) as in Equation 25
of all the UEs and L is the number of the transmit RF chains. H
|ĥr Vi−1 Vi−1 ĥH
k |
6) Ti ={r ∈Ti−1, r ∈C
/ i, kĥr kkĥk Vi−1 k
}<α
7) i=i+1
C. Semi-Orthogonal User Selection 8) end
1) Algorithm Description: In this subsection we briefly
describe the SUS algorithm proposed in [9], [10], which we 2) Complexity Analysis: Similar to the aforementioned
use as a baseline for our introduced UEs selection algorithms. condition number based UEs selection algorithm, here we also
Within each iteration i a filtration process is done to filter analyze the computational complexity of the SUS algorithm
the UEs and find the candidate set Ti based on a certain based on the number of SVDs needed in the worst case
orthogonality threshold α∈[0,1] as follows: scenario (upper bound - in case α = 1). In this case the

1147
 Table I
COMPUTATIONAL COMPLEXITY OF THE PROPOSED ALGORITHMS paper. In this approach, the UEs with the highest received
signal power are selected without considering the interference
CN SUS [9] Simple SUS [10] SLNR [18] (leakage/correlation) between the selected UEs. Henceforth,
it aims at sub-optimally maximizing the sum SE in a similar
O(|K|L3 ) O(|K|L3 ) O(|K|L2 ) O(|K|L) fashion to the hybrid beamforming approach in Equation 7.
Thus, decoupling the UEs selection from the interference
management problem and relying on ZF for nulling the Inter
User Interference (IUI) at the digital layer. Therefore, this
computational complexity is also of order O(|K|L3) [10].
strategy aims at maximizing the SNR of the selected UEs.
However, in [10] a simplified, yet equivalent, SUS algorithm
Thus, the maximum received signal algorithm is similar to
was introduced that avoids applying SVD decompositions
the maximum SLNR one explained in Algorithm (3) with
and requires an upper bound total number of flops expressed
replacing Step (5) to be similar to Step (4), which is:
as 12|K|L2 + 8L3 + 7|K|L − |K| − 17L2, henceforth the
computational complexity for the simplified SUS in [10] can
be represented of order O(|K|L2). k =argmax ĥk (28)
k∈Ti−1

D. Maximum Signal to Leakage and Noise Ratio Based

Selection V. SIMULATION RESULTS
Given the fact that all the previously studied algorithms In this section, we evaluate the performance of the
have high computational complexity, which limits the practical introduced UEs selection algorithms in terms of the SE based
implementation of such algorithms in a realistic mmWave on the simulation parameters summarized in Table (II).
system, specifically in the analog part which has limited We compare between the different introduced UEs selection
Digital Signal Processing (DSP) resources. Henceforth, the algorithms in terms of SE and complexity. As we observe
SLNR metric can be utilized as a sub-optimum low complexity in Table I that maximum SLNR algorithm has the lowest
approximation of the condition number UEs selection, since computational complexity in the introduced interference aware
it encounters both maximizing the signal and minimizing the UEs selection algorithms. Thus, it is the most favorable one
leakage (correlation) objectives. in terms of complexity and realistic implementation.
1) Proposed Algorithm: The SLNR metric for a given UE In Figure 1, we evaluate the normalization factor ζC in
k in the analog domain is defined as follows: a given scenario (M = 16, ∆φ = π3 ). It is shown that both
SUS and condition number based UEs selection achieve the
|hk fRF,k |2 lower bound of the normalization factor. This normalization
SLNRk = P 2 2
(27)
i6=k |hi fRF,k | +σ factor’s lower bound is calculated twice, once using the
Then, the SLNR maximization UEs selection is explained exhaustive search on the sum SE in Equation 17 and another
in Algorithm (3), and the computational complexity of this time using the exhaustive search on the eigen-values equivalent
algorithm can be determined in analogy with [18] and [19] bound in Equation 18. Also, we can observe that the SLNR
of order O(|K|L). algorithm achieves sub-optimal normalization factor with
median difference of ≈ 0.5 dB with the lower bound, while
Algorithm 3 SLNR Based Selection the maximum signal algorithm and the Round Robin (RR)
1) Initialize i=1,C0 =Φ,T0 =K ones have a significant performance degradation.
2) While (i≤L) do In Figure 2, we realize the effect of the normalization
3) if (i=1) factor on the per stream SE in the same scenario (M = 16,
4) find ∆φ= π3 ), we can observe again that, both SUS and condition
k =argmax ĥk number UEs selection achieve the upper bound of the SE,
k∈Ti−1 while maximum SLNR achieves sub-optimal SE with only 0.5
5) else bps/Hz difference to the SE upper bound at SNR γ = 19 dB.
find However, interference non-aware algorithms such as maximum
|hk fRF,k |2 signal and RR suffer from a significant SE loss.
k =argmax P 2 2 In Figures 3 and 4 , we generalize the simulation scenarios
k∈Ti−1 i∈Ci−1 |hi fRF,k | +σ
to have generic insights about the performance of the
6) end if introduced algorithms in various channel environments. In
7) Ci =Ci−1 ∪k Figure 3, the SNR is fixed γ =10 dB and the angular spread
8) Ti ={r ∈Ti−1,r ∈C
/ i} ∆φ = π3 , while the number of transmit antennas is varied
9) i=i+1 M = {16,32,64,128}. It is shown that, with increasing the
10) end number of transmit antennas, the spatial correlation/interference
decreases, due to the higher spatial resolution at the transmitter
side. Thus, enhancing the performance of non-interference
E. Maximum Received Signal Based Selection aware algorithm (Maximum received signal algorithm)
This maximum received signal UEs selection is considered significantly, from 1.5 bps/Hz difference with the SE upper
the basic base-line UEs selection approach throughout the bound at M =16 to 0.2 bps/Hz difference at M =128.

1148
 Table II 10
Round Robin
SIMULATION PARAMETERS 9 Max Signal
Max SLNR
8 Min Condition number
Bound - Sum SE

Spectral Efficiency (bps/Hz)

Parameter Value 7

Number of Propagation Paths 1 Ray (Pure LoS) 6

5
Simulation Type Monte Carlo (1000 realizations)
4
Transmit Array Architecture ULA
3

Inter-antenna spacing λ/2

2

Number of Transmit RF Chains 3 1

Total UEs, Receive Antennas 7, 1 0

16 32 64 128
Number of Transmit Antennas

Figure 3. Comparison between the per stream SE for the different user
1
selection strategies introduced in a pure LoS environment with different spatial
0.9
correlation given different number of transmit antennas with fixed angular
Min. Condition number
spread ∆φ= π3 ,γ =10dB,|K|=7,|C|=3
0.8 Max SLNR
Max Signal 9
0.7 Bound - Eigen-values Round Robin
Bound - Sum SE 8 Max Signal
0.6 SUS = 1 Max SLNR
Round Robbin 7 Min Condition number

Spectral Efficiency (bps/Hz)

CDF

Bound - Sum SE
0.5
6

0.4
5

0.3 4

0.2 3

0.1 2

0 1
-15 -10 -5 0 5 10 15 20 25 30
0
Normalization Factor (dB) π/3 π/2 3π/2 2π
Angular Spread (radians)

Figure 1. Comparison between the normalization factor for the different user
selection strategies introduced in a pure LoS environment given that M =16, Figure 4. Comparison between the per stream SE for the different user
|K|=7, |C|=3, and ∆φ= π3 . selection strategies introduced in a pure LoS environment with different spatial
correlation given different angular spread with fixed number of transmit
antennas M =16,γ =10dB,|K|=7,|C|=3
9
Bound - Eigen-values
8 Bound - Sum SE
Round Robin
∆φ = { π3 , π2 , 3π
2 ,2π}. It is shown that, with increasing the
Min. Condition number angular spread, the spatial correlation/interference decreases,
Spectral Efficiency (bps/ Hz)

7
SUS α = 1
Max SLNR
due to the higher angular separation between the UEs.
6 Max Signal Thus, enhancing the performance of non-interference aware
5
algorithm (Maximum received signal algorithm) significantly,
from 1.5 bps/Hz difference with the SE upper bound at
4 ∆φ= π3 to 0.4 bps/Hz difference at ∆φ=2π.
3
VI. CONCLUSION
2
In this paper we provided a detailed study on the UEs
1 selection topic with hybrid beamforming in mmWave systems.
We proved that in highly correlated environments, due to
0
-20 -15 -10 -5 0 5 10 15 20 clustered UEs deployments (limited angular spread) and
SNR (dB) average number of transmit antennas, interference aware
UEs selection is crucial and achieves considerable gains over
Figure 2. Comparison between the per stream SE for the different user
selection strategies introduced in a pure LoS environment given that M =16,
non aware UEs selection approach based on maximizing
|K|=7, |C|=3, and ∆φ= π3 . the received signals. However, in perfect channel conditions
(Massive transmit antennas and large angular spreads), selecting
the UEs based on their received signal strength can be sufficient.
We derived analytically an optimum interference aware UEs
In Figure 4, the SNR is fixed γ =10 dB and the number of selection based on the condition number of the MU channel and
transmit antennas M =16, while the angular spread is varied then we provided a sub-optimal low complexity (realistic) UEs

1149
 selection approach based on SLNR maximization. We proved [2] M. Xiao, S. Mumtaz, Y. Huang, L. Dai, Y. Li, M. Matthaiou,
with simulation results that maximum SLNR algorithm is a G. K. Karagiannidis, E. Björnson, K. Yang, C. Lin, and
A. Ghosh, “Millimeter wave communications for future mobile
potential candidate for the future mmWave systems achieving networks,” CoRR, vol. abs/1705.06072, 2017. [Online]. Available:
a trade-off between high SE performance and low complexity. http://arxiv.org/abs/1705.06072
[3] O. E. Ayach, R. W. Heath, S. Abu-Surra, S. Rajagopal, and Z. Pi, “Low
complexity precoding for large millimeter wave mimo systems,” in 2012
APPENDIX A IEEE International Conference on Communications (ICC), June 2012,
pp. 3724–3729.
PROOF OF LEMMA 1 [4] T. Bai and R. W. Heath, “Coverage analysis for millimeter wave cellular
The normalization factor for hybrid beamforming (LoS networks with blockage effects,” in 2013 IEEE Global Conference on
Signal and Information Processing, Dec 2013, pp. 727–730.
Beamsteering - ZF) can be analyzed as follows: [5] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath,
ζC =E(T r(FH H
RF,C FBB,C FRF,C FBB,C )) (29) “Spatially sparse precoding in millimeter wave mimo systems,” IEEE
Transactions on Wireless Communications, vol. 13, no. 3, pp. 1499–1513,
Given the fact that the ZF precoder FBB,C is applied on the March 2014.
equivalent channel Ĥ, and given that the SVD of the equivalent [6] Z. Li, S. Han, and A. F. Molisch, “Hybrid beamforming design for
channel is Ĥ=UeDeVeH . Henceforth, it can be deduced that: millimeter-wave multi-user massive mimo downlink,” in 2016 IEEE
International Conference on Communications (ICC), May 2016, pp. 1–6.
[7] J. Noh, T. Kim, J. Y. Seol, and C. Lee, “Zero-forcing based
−2
ζC =E(T r(FH RF,C FRF,C De )) (30) hybrid beamforming for multi-user millimeter wave systems,” IET
Communications, vol. 10, no. 18, pp. 2670–2677, 2016.
Again, due to the fact that the SVD of the propagation channel [8] A. Alkhateeb, G. Leus, and R. W. Heath, Jr, “Limited Feedback Hybrid
is H=UDVH , and that the RF beamforming matrix in this Precoding for Multi-User Millimeter Wave Systems,” ArXiv e-prints,
specific case (pure LoS channel with the LoS Beamsteering Sep. 2014.
H [9] T. Yoo and A. Goldsmith, “On the optimality of multiantenna broadcast
approach and normalizing α) FRF =√HN . Thus, scheduling using zero-forcing beamforming,” IEEE Journal on Selected
t
Areas in Communications, vol. 24, no. 3, pp. 528–541, March 2006.
−2 1 [10] J. Mao, J. Gao, Y. Liu, and G. Xie, “Simplified semi-orthogonal
E{T r(FH
RF FRF De )=E{T r( HHH D−2
e )} user selection for mu-mimo systems with zfbf,” IEEE Wireless
Nt Communications Letters, vol. 1, no. 1, pp. 42–45, February 2012.
1 [11] D. L. Ruyet, H. Khanfir, and B. Ozbek, “Performance of semi-orthogonal
=E{T r( D2D−2e )} user selection for multiuser miso systems,” in 2008 IEEE 9th Workshop
Nt (31)
on Signal Processing Advances in Wireless Communications, July 2008,
P|C| λHk pp. 321–325.
k=1 λĤ [12] S. Nam, J. Kim, and Y. Han, “A user selection algorithm using angle
k
= between subspaces for downlink mu-mimo systems,” IEEE Transactions
Nt on Communications, vol. 62, no. 2, pp. 616–624, February 2014.
Where λHk and λĤ
k represent the k−th eigen-value of the [13] X. Yi and E. K. S. Au, “User scheduling for heterogeneous multiuser
channel and the equivalent channel respectively. In this specific mimo systems: A subspace viewpoint,” IEEE Transactions on Vehicular
Technology, vol. 60, no. 8, pp. 4004–4013, Oct 2011.
scenario (pure LoS environment with LoS beamsteering) [14] K. Ko and J. Lee, “Multiuser mimo user selection based on chordal
the eigen-values of the equivalent channel are related to the distance,” IEEE Transactions on Communications, vol. 60, no. 3, pp.
eigen-values of the channel as follows: 649–654, March 2012.
[15] ——, “Determinant based multiuser mimo scheduling with reduced
pilot overhead,” in 2011 IEEE 73rd Vehicular Technology Conference
(λHk )
2
(VTC Spring), May 2011, pp. 1–5.
λĤ
k = (32) [16] X. Lu, J. Wu, X. Huang, W. Li, J. Lu, and Z. Gong, “An improved
Nt semi-orthogonal user selection algorithm based on condition number
Thus, substituting Equation 32 in 31, the normalization for multiuser mimo systems,” China Communications, vol. 11, no. 13,
factor is represented as follows: pp. 23–30, Supplement 2014.
[17] M. Sadek and S. Aissa, “Leakage based precoding for multi-user
|C| |C| mimo-ofdm systems,” IEEE Transactions on Wireless Communications,
X 1 X 1 vol. 10, no. 8, pp. 2428–2433, August 2011.
ζC = H
= (33) [18] L. Dong, G. Wu, H. Feng, and S. Li, “Leakage-based power allocation
k=1 k
(λ ) k=1 λk for relaxed SLNR beamforming in multi-cell MU-MIMO systems,”
in 2014 IEEE International Conference on Communication Systems
(ICCS), Nov. 2014, pp. 512–516.
ACKNOWLEDGEMENT [19] M. Shehata, M. Hélard, M. Crussière, A. Roze, and C. Langlais, “Angular
Based Beamforming and Power Allocation Framework in a Multi-User
This work has received a French state support granted to Millimeter-Wave Massive MIMO System,” in 2018 IEEE 87th Vehicular
the CominLabs excellence laboratory and managed by the Technology Conference: VTC2018-Spring, Porto, Portugal, Jun. 2018.
National Research Agency in the Investing for the Future [Online]. Available: https://hal.archives-ouvertes.fr/hal-01730747
[20] A. M. Sayeed, “Deconstructing multiantenna fading channels,” IEEE
program under reference Nb. ANR-10-LABX-07-01 and Transactions on Signal Processing, vol. 50, no. 10, pp. 2563–2579, Oct
project name M5HESTIA (mmWave Multi-user Massive 2002.
MIMO Hybrid Equipments for Sounding, Transmissions and [21] N. S. Perović, P. Liu, M. D. Renzo, and A. Springer, “Receive spatial
modulation for los mmwave communications based on tx beamforming,”
HW ImplementAtion). IEEE Communications Letters, vol. 21, no. 4, pp. 921–924, April 2017.

REFERENCES
[1] T. S. Rappaport, Y. Xing, G. R. MacCartney, A. F. Molisch, E. Mellios,
and J. Zhang, “Overview of millimeter wave communications for
fifth-generation (5g) wireless networks—with a focus on propagation
models,” IEEE Transactions on Antennas and Propagation, vol. 65,
no. 12, pp. 6213–6230, Dec 2017.

1150

View publication stats