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

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
1

Joint Power Control and LSFD for

Wireless-Powered Cell-Free Massive MIMO
Özlem Tuğfe Demir, Member, IEEE, and Emil Björnson, Senior Member, IEEE

Abstract—This paper considers wireless uplink information and several network architectures have been proposed [13]–
and downlink power transfer in cell-free massive multiple-input [21].
multiple-output systems. The single-antenna user equipments Communication and positioning are the main use cases
(UEs) utilize the energy harvested in the downlink to transmit
uplink pilot and information signals to the multiple-antenna for radio frequency (RF) in current wireless systems. While
access points (APs). We consider Rician fading and maximum we are in the era of 5G for mobile communication, some
ratio processing based on either linear minimum mean-squared emerging technologies have potential to be integrated into
error (LMMSE) or least-squares (LS) channel estimation. We future generation standards. Wireless power transfer (WPT)
derive the average harvested energy by using a practical non- via RF signals is one of these technologies to exploit the
linear energy harvesting circuit model for both coherent and non-
coherent transmission schemes. Furthermore, the uplink spectral RF energy for battery-limited devices and there has been
efficiency (SE) is derived for all the considered methods and extensive research conducted in this area to charge mobile
the max-min fairness problem is cast where the optimization battery-powered devices via ambient and dedicated RF signals
variables are the AP and UE power control coefficients together [22], [23]. WPT would reduce the battery requirements (size,
with the large-scale fading decoding vectors. The objective is to wiring, etc.) of the mobile devices and provide more consistent
maximize the minimum SE of the UEs’ under APs’ and UEs’
transmission power constraints. A novel alternating optimization and ubiquitous service to energy-hungry devices by supplying
algorithm with guaranteed convergence and improvement at each reliable energy. In particular, future autonomous low-power
step is proposed to solve the highly-coupled non-convex problem. networks and Internet of Things (IoT) are expected to benefit
from this technology [22].
Index Terms—Cell-free massive MIMO, max-min fair power Simultaneous wireless information and power transfer
control, wireless power transfer, spectral efficiency, Rician fading. (SWIPT), which is an interesting paradigm in WPT, has been
considered for cellular massive MIMO systems [24]–[27]. In
these works, the UEs have either a power splitting or time
switching circuit to utilize the downlink RF signals for both
I. I NTRODUCTION information reception and energy harvesting. In [28], [29], a
Massive MIMO (multiple-input multiple-output) has re- base station (BS) assists the UEs for their uplink pilot and data
ceived great interest in the last decade and has been exten- transfer by energy beamforming in the downlink. In this paper,
sively analyzed for cellular systems due to its high spectral we adopt this setup but consider a cell-free massive MIMO
efficiency (SE) achieved by spatial multiplexing of many user system with several transmission schemes and a practical
equipments (UEs) on the same time-frequency resource [2]– non-linear energy harvesting model. The motivation behind
[7]. Now, it has reached its mature stage and is one of the integrating cell-free massive MIMO with WPT is that each UE
key technologies in 5G, and commercial deployments began is expected to have a much higher channel gain to at least one
in 2018 [2]. Although 5G cellular technology with massive of the APs with larger probability compared to cellular massive
MIMO is expected to provide higher data rates compared to MIMO. Hence, one of the main limitations in RF WPT, which
the previous technologies, the inter-cell interference is still is path loss, is overcome to some extent. Furthermore, not only
an important issue, particularly for the cell-edge UEs [8]. sensor networks and IoT devices for which changing batteries
Recently, an alternative network infrastructure is considered is infeasible, but also for the battery-limited mobile UEs, WPT
in [9], [10], which uses the name cell-free massive MIMO is a promising technology as long as the transmission range is
since a large number of access points (APs) is distributed over not too long [22], [28]–[30]. Hence, cell-free networks are
a large geographic area to serve all the UEs in a coherent advantageous compared to cellular systems in this respect.
manner without any cell boundaries. Cell-free massive MIMO In this paper, we assume all the UEs in the network benefit
was shown to improve the minimum SE achieved in the from WPT and as the simulations show, employing denser APs
network and total energy efficiency [10]–[12]. Recent works improves the minimum guaranteed uplink SE.
have focused on different aspects of cell-free massive MIMO The prior works on WPT in cell-free systems are few.
In [18], the total harvested energy throughput is maximized
The authors are with the Department of Electrical Engineering together with the AP selection under transmission power
(ISY), Linköping University, 581 83 Linköping, Sweden (e-mail: constraints for each AP. This work assumes perfect channel
[email protected], [email protected]). This work was partially state information (CSI) and does not take into account the
supported by ELLIIT and the Wallenberg AI, Autonomous Systems and
Software Program (WASP) funded by the Knut and Alice Wallenberg uplink communications. In [19], SWIPT is considered in the
Foundation. A part of this paper was presented in WCNC 2020 [1]. context of cell-free massive MIMO where information and

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
2

energy UEs are located separately. Similarly, [20] studied vesting, or max-min fairness. Furthermore, we consider
cell-free massive MIMO where the information UEs do not a different pilot signaling scheme than the random pilot
harvest energy and there is a single energy-harvesting UE that signaling [21]. These factors make our problem unique
actively eavesdrops. In [21], the authors consider minimization and more challenging compared to the case with a linear
of the total transmitted energy for wirelessly-powered cell-free energy harvesting model and no LSFD.
IoT by considering only Rayleigh fading and a linear energy • We propose an alternating optimization algorithm to
harvesting model. In this paper, we adopt a more general achieve a convergent solution to the proposed non-convex
Rician fading channel model with a common random phase problem. The resulting non-convex sub-problems are
shift to all the antennas at each AP. This leads to the channel solved efficiently after some novel transformations. The
coefficients for different AP antennas being dependent on each simulation results show that the solution found by this
other. Furthermore, the channels are not Gaussian unlike most algorithm improve the minimum guaranteed SE of the
of the previous works. Hence, the statistical results presented network compared to simpler power control scheme in
in previous works cannot be used here. We derive the exact [16].
closed-form expressions for SE and average harvested energy Note that the conference version of this paper, [1], only
for the first time for Rician fading with random phase shifts considers LMMSE-based channel estimation and non-coherent
by using two different channel estimation schemes. energy transmission using a linear energy harvesting model.
Different from the existing works, this paper is the first one
that considers power control for maximizing the minimum up- II. S YSTEM M ODEL
link SE for downlink WPT-assisted cell-free massive MIMO.
We consider a cell-free massive MIMO system where L
Max-min fairness is one of the important optimization criteria
multiple-antenna APs are distributed over a large area to serve
since it maximizes the minimum guaranteed SE to all the UEs,
K single-antenna UEs with energy harvesting capability. Each
which is highly in accordance with the uniformly great service
AP is equipped with N antennas and connected to a central
goal of cell-free systems. Furthermore, max-min fairness may
processing unit (CPU) via an error-free fronthaul link. In
be effective to reduce the traffic congestion mainly resulting
this paper, we assume time division duplex (TDD) operation
from the UEs in bad channel conditions, by increasing the
and, hence, channel reciprocity holds. Let τc denote the total
%95-likely SE of the whole network. The resulting joint
number of samples per coherence interval. Each coherence
optimization problem in terms of the uplink/downlink power
interval is divided into three phases: uplink training, downlink
coefficients and large-scale fading decoding (LSFD) weights
WPT, and uplink wireless information transfer (WIT). In
is non-convex and more challenging compared to the previous
the uplink training phase, all the UEs send pilot sequences
works due to the non-linear energy harvesting model and
of length τp to the APs, which estimate the channels to
highly-coupled variables. After some mathematical manipula-
design precoding vectors for effective energy transfer and data
tions, we come up with a problem structure where an efficient
reception. While τd samples are used for downlink WPT, the
modified bisection search-based alternating optimization can
remaining τu samples are used for the uplink WIT, hence,
be applied. We show that semidefinite programming with rank
we have τp + τd + τu = τc . In accordance with the existing
relaxation guarantees a rank one solution for the subproblems.
literature on cell-free massive MIMO, the CSI is not shared
Overall, the main contributions of this paper are:
between the APs [10], [13].
• We derive the average harvested energy and the uplink Let gkl ∈ CN denote the channel between the k th UE and
SE in closed-form when the channels are estimated the lth AP. The channels are constant in each time-frequency
using a linear minimum mean-squared error (LMMSE) coherence interval. We consider spatially uncorrelated Rician
and least squares (LS) estimators for practical Rician fading channels with unknown phase shifts, which is the
fading channels with unknown phase shifts. We derive first novelty of this paper in the context of cell-free massive
the SE expressions for the multi-antenna APs that are MIMO with multiple-antenna APs. This means each channel
generalizations of the SE for single-antenna APs in [15]. realization can be expressed as
Note that the results in [15] cannot be used for multiple- gkl = ejθkl ḡkl + g̃kl , (1)
antenna APs due to the common phase shifts of the
channels, which are the same for each BS antenna. where ejθkl ḡkl ∈ CN denotes the line-of-sight (LOS) com-
• We consider both coherent and non-coherent downlink ponent. Moreover, g̃kl is the non-line-of-sight (NLOS) com-
WPT schemes where the same or independent energy ponent and the small-scale fading is modeled as g̃kl ∼
symbol for each UE is transmitted from the APs, respec- NC (0N , βkl IN ), where βkl is the large-scale fading coefficient
tively. Furthermore, a practical non-linear energy harvest- which accounts for path-loss and shadowing. Note that the vec-
ing model [32] is utilized in the closed-form results and tors {ḡkl } and large-scale fading coefficients {βkl } describe
the optimization algorithm. the long-term channel effects and change more slowly than
• We formulate the max-min fair joint AP and UE power the small-scale fading realizations. We assume that the APs
control and LSFD design problem under the constraints have perfect knowledge of {ḡkl , βkl } corresponding to the
on harvested and transmitted power at the APs and UEs. channels between them and the UEs, in accordance with prior
Note that this problem has a different structure than [21], literature [5], [7]. We consider a realistic scenario where the
which does not consider LSFD, non-linear energy har- phase shifts {θkl } in the LOS components are unknown due

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
3

to user mobility and assume they are uniformly distributed where

in the interval [0, 2π) [15]. Note that most of the previous H H
Rkl , E{gkl gkl } = ḡkl ḡkl + βkl IN , (5)
works that consider Rician fading neglect the phase shifts X
H H 2


{θkl } since they do not affect the distribution of the channel Ψkl , E{zkl zkl } = τp ρp ḡil ḡil + βil IN + σ IN .
gain. However, in practical systems where channel estimation i∈Pk
is required at the receiver, we should take these random phase (6)
shifts into account. When the transmitter and receiver move The channel estimate ĝkl and the estimation error ekl =
over distances at order of the wavelength, a small random gkl − ĝkl are zero-mean uncorrelated random vectors with
phase shift is induced on the LOS component as well as covariance matrices
the individual paths constructing the NLOS component of a
H
= τp ρp Rkl Ψ−1

channel. The effect of the random phase shifts on the large R̂kl , E ĝkl ĝkl kl Rkl , (7)
H −1

number of scattered paths is already taken into account by Ckl , E ekl ekl = Rkl − τp ρp Rkl Ψkl Rkl . (8)
modeling the NLOS component of the channel as Gaussian.
However, the direct path leading to the LOS component is Note that neither the channel estimate nor the estimation error
usually much stronger than the NLOS part and we should treat is Gaussian. As a result, although they are uncorrelated, they
its phase shift separately. These phase shifts vary at the same are not independent.
pace as the NLOS component and from coherence block to Note that the LMMSE-based channel estimator presented
coherence block. As a result, the phase shifts are not known above requires the computation of an inverse matrix which
in advance at the BS as NLOS components and we should can be computationally complex when N is large. A simpler
consider the unknown phase shifts on the direct path separately estimation scheme is the least squares (LS) estimator that does
in channel estimation. not make use of the channel statistics. The LS estimate of the
channel gkl is a scaled version of zkl in (3). In the following
parts of the paper, we will use directly zkl for maximum
III. C HANNEL E STIMATION ratio (MR) processing since power control optimization will
Let ϕk ∈ Cτp denote the pilot sequence that is assigned be implemented and, hence, the scaling factor in front of zkl
to the k th UE where ||ϕk ||2 = τp . If τp ≥ K, we can will not affect the result.
use an orthogonal set of pilot sequences. However, this is
generally not the case for cell-free massive MIMO systems IV. D OWNLINK E NERGY H ARVESTING
since the number of UEs can be much larger than the pilot In the WPT phase, each AP transmits energy to the UEs
sequence length. Hence, so-called pilot contamination occurs. by using the CSI for downlink precoding. In this paper, we
The second novelty of this paper is to take pilot contamination will first analyze coherent energy transmission where the APs
into account in the analysis of cell-free massive MIMO based transmit the same energy symbol for each UE in a synchronous
WPT. manner in order to increase the harvested energy at the UEs.
Deriving the MMSE estimator is non-trivial since we do not ∗
Let wkl ∈ CN denote the downlink precoding vector for
have a linear Gaussian signal model. We will therefore restrict the energy harvesting phase. Then, the signal transmitted by
ourselves to the LMMSE estimator as in [15], which is the the lth AP is
conventional benchmark in the massive MIMO literature. To K
obtain the LMMSE channel estimator in a simple form, let us
X √ ∗
xE
l = pkl wkl sk , (9)
assume that the pilot sequences are either identical or mutually k=1
orthogonal and call Pk the subset of UEs which are assigned
where sk is the zero-mean unit-variance energy signal for the
the same pilot sequence as the k th UE, including itself. Then,
k th UE. The energy signals for different UEs are assumed
the received pilot signal Zl ∈ CN ×τp at the lth AP is given by
independent for the ease of analysis. pkl is the power control
K coefficient of the lth AP corresponding to the k th UE. The
X √
Zl = ρp gkl ϕTk + Nl , (2) transmission power for each AP should satisfy the maximum
k=1 power limit that is ρd in the long-term, i.e.,
where ρp is the pilot transmit power and the additive noise
n 2
o
PlE , E xE l ≤ ρd . (10)
matrix Nl ∈ CN ×τp has i.i.d. NC (0, σ 2 ) random variables. A
sufficient statistics for the estimation of the k th UE’s channel The average transmitted power PlE for the lth AP is
is  2

K K
Zl ϕ∗ √
 X √ ∗
 (a) X n
2
o
PlE = E
X
zkl = √ k = τp ρp gil + nkl , (3) pkl wkl sk = pkl E kwkl k
τp 
k=1

k=1
i∈Pk
K
√  
where nkl , Nl ϕ∗k / τp ∼ NC (0N , σ 2 IN ). Note that nil
P


 pkl tr R̂kl , if LMMSE with wkl = ĝkl in (4),
(b)
is independent of nkl for ∀i ∈
/ Pk . Then, the phase-unaware = k=1 K
LMMSE estimate of gkl , based on (3), is
P
pkl tr (Ψkl ) , if LS with wkl = zkl in (3),



√ k=1
ĝkl = τp ρp Rkl Ψ−1
kl zkl , (4) (11)

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
4

L X
K L X
where (a) is the result of the independence of the zero-mean X X
Ik = pil tr (Ψil Rkl ) + pil τp ρp ×
signals {sk }. We used (6) and (7) in (b) for the MR precoders
l=1 i=1 l=1 i∈Pk
based on the LS and LMMSE-based channel estimation. The
L X
L X √ √
received signal in the energy harvesting phase at the k th UE
  X
H
2N βkl ḡkl ḡkl + N 2 βkl
2
+ pil pil0 τp ρp ×
is
l=1 l0 =1, i∈PK
l0 6=l
L L X
K ∗
√
 
X X H H
rkE = T E
gkl xl + nE
k =
H
pil wil gkl si + nE
k , (12)
ḡkl ḡkl + N βkl ḡkl 0 ḡkl0 + N βkl0 with LS. (15)
l=1 l=1 i=1
Proof: Please see Appendix B.

where nE 2 th Note that all the terms in (14) and (15) are positive. All
k ∼ NC (0, σ ) is the additive noise at the k UE.
Since the noise floor is too low for energy harvesting, we UEs’ intended signals from all the APs make a contribution
simply neglect the effect of nE to the input power of the rectifier circuit for each UE. The
k in the average harvested energy
expression in accordance with the existing literature [18], [19], value is also affected by the power control coefficients {pil }.
[29]. Then, the average input power at the energy harvesting In addition to the first summation, having pilot contaminated
rectifier circuit of the k th UE is channel estimates brings some additional power terms into the
second summation. However, at the same time, the pilot con-

L X
K 2
 tamination reduces the channel estimation quality, which leads
X √ H

to a reduction in some of the expressions related to correlation
Ik =E pil wil gkl si

l=1 i=1
 between the actual and estimated channels. Hence, it is not
L X
L X
K
easily seen from the formula whether the pilot contamination
(a) X √ √  H H increases the harvested energy or not. As expected, the input
= pil pil0 E wil gkl gkl 0 wil0

l=1 l0 =1 i=1
power for energy harvesting increases with the increase in the
L X K large-scale fading coefficients {βkl } and the norm of the LOS
(b)
parts of the channels {kḡkl k}. From (15), it is clearly seen
X  H H
= pil E wil gkl gkl wil
l=1 i=1 that using larger number of antennas, N , at the APs increase
L X L X
K the input power of the rectifier.
X √ √  H  H
+ pil pil0 E wil gkl E gkl 0 wil0 In the prior conference version of this work, non-coherent
l=1 l0 =1, i=1 energy transmission that allows each AP to transmit their
l0 6=l choice of energy symbols is analyzed. Hence, the non-coherent
(13) scheme does not require any synchronization among APs in
the downlink. For this case, the average input power at the
where (a) and (b) follow from the independence of the zero- energy harvesting circuit of the k th UE is given by
mean energy signals and the channels to different APs. L X
K n o
X
H 2
The following lemma presents the average input power to Ik = pil E wil gkl , (16)
the harvester, Ik , analytically for coherent energy transmission l=1 i=1
with the two MR precoders. which is the first term of Ik for the coherent energy transmis-
sion in (13). Hence, using Lemma 1, the average input power
Lemma 1. The average input power at the energy harvesting
Ik for the non-coherent energy transmission case is given by
circuit of the k th UE for coherent energy transmission is given
L X
K L X
by X   X
Ik = pil tr R̂il Rkl + pil τp2 ρ2p ×
l=1 i=1 l=1 i∈Pk
L X
X K   L X
X
Ik = pil tr R̂il Rkl + pil τp2 ρ2p ×  H −1
Ψil Ril ḡkl tr Ril Ψ−1


2βkl < ḡkl il
l=1 i=1 l=1 i∈Pk

 H −1 !
Ψil Ril ḡkl tr Ril Ψ−1


2βkl < ḡkl + 2
βkl tr Ψ−1

2
il Ril with LMMSE, (17)
il
! L X
L
2 X X √ √ L X
K
2
tr Ψ−1 pil pil0 τp2 ρ2p ×


+ βkl il Ril + Ik =
X
pil tr (Ψil Rkl )
l=1 l0 =1, i∈PK
l=1 i=1
l0 6=l
! XL X  
H
H −1
ḡkl Ψil Ril ḡkl + βkl tr Ψ−1


× + pil τp ρp 2N βkl ḡkl ḡkl + N 2 βkl
2
with LS.
il Ril
l=1 i∈Pk
!∗ (18)
−1
H
Ψ−1


ḡkl 0 Ψ 0 Ril0 ḡkl0
il + βkl0 tr il0 Ril
0 with LMMSE, The average input power for the non-coherent transmission
is always less than that of the coherent transmission. Hence,
(14) on the average, the coherent scheme allows more energy

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
5

L
harvesting at each UE. In the simulation results, we will √ X ∗  H
DSk = ηk akl E vkl gkl ,
quantify this benefit.
l=1
We will use the following non-linear energy harvesting L
model in accordance with [32]. This model highly correlates √ X
a∗kl vkl
H
 H

BUk = ηk gkl − E vkl gkl , (23)
with real measured data. The total harvested energy at the k th l=1
UE in τd channel uses is L L
√ X X
τd Ak Ik UIki = ηi a∗kl vkl
H
gil , ñk = a∗kl vkl
H I
nl . (24)
Ek = (19) l=1 l=1
Bk Ik + Ck
where Ak > 0, Bk ≥ 0, and Ck are constants determined by Let us define the following vectors and matrices for ease of
curve fitting of the rectifier circuit of the k th UE [32]. If we notation:
set Bk to zero, we obtain the classical linear energy harvesting
model. ak , [ ak1 . . . akL ]T ∈ CL ,
bk , [ bk1 . . . bkL ]T ∈ CL , bkl , E vkl
 H
Note that, the harvested energy is proportional to the number gkl , (25)
of downlink energy symbols, τd . However, increasing τd is L×L ll0
 H H
Cki ∈ C , cki , E vkl gil gil0 vkl0 ,
expected to increase the SE up to some extent since for a n
H o
fixed coherence block length, τc , an increase in τd necessitates Dk ∈ CL×L , dkl , E vkl H I
nl nIl vkl , (26)
a decrease in τu that is proportional to the SE of each UE as
0
we consider in the next section. 0
where cllki is the (l, l )th element of the matrix Cki . Dk is a
diagonal matrix with the lth diagonal element being dkl .
V. U PLINK W IRELESS I NFORMATION T RANSFER Utilizing the use-and-then-forget capacity bounding tech-
nique in [6], the uplink SE for the k th UE with LSFD for any
In the uplink information transmission phase, all the UEs finite value of L, K, and N is given by
simultaneously send their data signals to the APs. Let qk
th
 2the symbol of the k UE, which is zero-mean with
denote
Rk =
τu
log2 (1 + SINRk ) , (27)
E |qk | = 1, and ηk ≥ 0 is the corresponding transmission τc
power. The received signal at the lth AP is
where the effective signal-to-noise-plus-ratio is
K
X √
rIl = ηk gkl qk + nIl , l = 1, . . . , L, (20) 2
k=1
ηk aH
k bk
SINRk = P 
2
.
K
where nIl 2


∼ NC 0N , σ IN is the additive white Gaussian aH
k
H
i=1 ηi Cki ak − ηk ak bk + aH
k Dk ak
noise. Each AP applies local decoding for each UE’s infor- (28)
∗
mation symbol before sending it to the CPU. Let vkl ∈ CN
th
denote the decoding weight vector for the k UE’s signal at In the following two lemmas, we present the uplink SE for the
the lth AP. Hence, r̃kl = vklH I
rl is the locally decoded signal two MR-based decoding vectors, which is another novelty of
th th
for the k UE at the l AP. We consider two MR decoding this paper in the context of multiple antenna cell-free massive
methods based on the LMMSE or LS-based channel estimates, MIMO with unknown phase-shifted Rician fading and LSFD.
i.e., vkl = ĝkl or vkl = zkl , respectively.
Lemma 2. The uplink SE for the k th UE with MR decoding
The CPU computes a weighted sum of the locally decoded vkl = ĝkl in (4) is given in (27) with the effective SINR as in
signals using the large-scale fading decoding (LSFD) method (28), where the elements of bk , Cki , and Dk are given as
[9]:
H −1
Ψkl Rkl ḡkl + τp ρp βkl tr Ψ−1


L
X bkl = τp ρp ḡkl kl Rkl , (29)
q̂k = a∗kl r̃kl , (21)  
l=1
cll
ki = tr R̂kl Ril + Ii∈Pk τp ρp ×
2 2

  H −1
Ψkl Rkl ḡil tr Rkl Ψ−1


for the detection of the k th UE’s information signal where 2βil < ḡil kl
{a∗kl } are the LSFD weights. We assume the CPU uses only
2
the statistical knowledge of the channels so that no CSI sharing + βil2 tr Ψ−1 kl R kl , (30)
 
is needed [9], [13], [15]. Using the SE analysis technique in cll
0
2 2 H −1 −1


ki = Ii∈Pk τp ρp ḡil Ψkl Rkl ḡil + βil tr Ψkl Rkl ×
[6], we can express q̂k for the k th UE data detection as 
−1 −1

∗

X H
ḡil 0 Ψ 0 Rkl0 ḡil0 + βil0 tr Ψ 0 Rkl0
kl kl , l0 6= l, (31)
q̂k = DSk qk + BUk qk + UIki qi + ñk , (22)  
i6=k dkl = σ 2 tr R̂kl , (32)
where DSk , BUk , UIki denote the strengths of the desired
signal (DS), beamforming gain uncertainty (BU), and the where I(.) is the indicator function, i.e., Ii∈Pk is equal to one
interference of the ith UE on the k th UE, while ñk is the if i ∈ Pk , otherwise it is equal to zero.
total noise at the CPU. DSk , BUk , UIki , and ñk are given by Proof: Please see the Appendix C.

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
6

Lemma 3. The uplink SE for the k th UE with MR decoding following part, if we can write Ik in (14)-(15) as a linear
vkl = zkl in (3) is given in (27) with the effective SINR as in function of optimization variables, the constraint in (40) can
(28), where the elements of bk , Cki , and Dk are given as be expressed in terms of second-order cone constraints that
√ H √ are convex. To do this, let us define the following positive
bkl = τp ρp ḡkl ḡkl + N τp ρp βkl , (33)
  semi-definite matrices that have also rank one.
cll H 2 2 √ 
ki = tr (Ψkl Ril ) + Ii∈Pk τp ρp 2N βil ḡil ḡil + N βil , pi1
(34)  ..  √ √ 
0
Pi ,  .  pi1 . . . piL  0, i = 1, . . . , K.
cll H H
l0 6= l, √



ki = Ii∈Pk τp ρp ḡil ḡil + N βil ḡil0 ḡil0 + N βil0 , piL
(35) (42)
dkl = σ 2 tr (Ψkl ) . (36)
Ik can be expressed as a linear function of the elements of
Proof: The proof follows similar steps in the Appendix C the matrices {Pi }. The problem in (37)-(41) for some given
and hence omitted. ak and t becomes a feasibility problem. We can include
the summation of data transmission powers of UEs in the
Note that we have more expectation terms to compute
objective to this problem to obtain good feasible solutions that
compared to the conventional cellular massive MIMO where
will improve the next stages of the alternating optimization
a base station only serves the UEs in its cell. In the cellular
algorithm. After introducing the new optimization variables
case, the UEs in a cell usually are assigned orthogonal pilot
{ek }, the equivalent version of the problem is
sequences. Hence, pilot contamination results from the other
cells’ signals. However, in cell-free massive MIMO, each AP K
X
serves all the UEs subsets of which are sharing the same pilot minimize ηk (43)
{ηk , Pk , ek }
and this in turn results in more terms that appear in both the k=1
average harvested energy and SE expressions. K
2 X
subject to (1 + t) ηk aH
k bk −t ηi aH
k Cki ak
i=1
VI. M AX -M IN FAIR J OINT LSFD AND P OWER C ONTROL
− taH
k Dk ak ≥ 0, k = 1, . . . , K, (44)
We want to maximize the minimum SE among the UEs
K
by adjusting both the downlink WPT and uplink WIT power X
Pkll E kwkl k2 ≤ ρd ,

l = 1, . . . , L, (45)
control coefficients and the LSFD weights.
k=1
For the considered optimization, the transmission power  n 0 o 
of the lth AP during the downlink WPT phase, PlE in (11) ek Bk Ik Pill + Ck ≥ 1, k = 1, . . . , K, (46)
cannot exceed the long-term maximum power limit ρd in τd Ak τd Ak Ck
(10). Furthermore, we require that the k th UE’s total uplink τu ηk + τp ρp ≤ − ek , k = 1, . . . , K,
Bk Bk
transmission energy, τu ηk + τp ρp is upper bounded by the (47)
harvested energy Ek in (19). Then, the max-min fairness SE 0
0
Pkll≥ 0, l = 1, . . . , L, l = 1, . . . , L,
optimization problem is cast as
ηk ≥ 0, k = 1, . . . , K, (48)
maximize t (37)
{ak ,ηk ,pkl },t Pk  0, rank (Pk ) = 1, k = 1, . . . , K, (49)
subject to SINRk (ak , {ηi }) ≥ t, k = 1, . . . , K, (38) 0
0
where Pill denote nthe (l,ol )th element of the matrix Pi . Note
PlE ({pil }) ≤ ρd , l = 1, . . . , L, (39) that since Bk Ik Pill
0
+ Ck is a positive affine function
τu ηk + τp ρp ≤ Ek ({pil0 }) , k = 1, . . . , K, n 0o
(40) of the optimization variables Pill , the constraints in (46)
can be written as second-order cone constraints. Note that
pkl ≥ 0, l = 1, . . . , L, ηk ≥ 0, k = 1, . . . , K,
(47) is now linear with respect to the optimization variables
(41)
and the two optimization problems (37)-(41) and (43)-(49) are
where SINRk is from (28) and t is the SINR that all equivalent in the sense that they have the same global optimum
UEs achieve. Note that this problem is neither convex nor solution. To show this, suppose at least one of the inequalities
manageable in terms of finding the global optimum solution (46) is not satisfied with equality for the optimum solution
due to the highly-coupled variables. However, an alternating of (43)-(49). In this case, the corresponding ek > 0 can
optimization approach can be developed in an efficient manner be reduced until the constraint becomes an equality without
where an improved solution is obtained at each step with affecting the feasibility and the optimality of the solution. This
guaranteed convergence. The motivation for the alternating modified solution is also a global optimum nfor (43)-(49), and
 o 
approach is explained as follows. Note that PlE is a linear hence for (37)-(41) since ek = 1 Bk Ik Pi ll0
+ Ck .
function of {pil }. Furthermore, the numerator and denominator We note that the only constraints that destroy the convexity
of SINRk are linear in {ηi }, given the LSFD vectors ak , for in (43)-(49) are the rank one constraints in (49). The following
k = 1, . . . , K. lemma shows that we can obtain an optimum solution to (43)-
The harvested energy for the k th UE, Ek in (19), is a (49) by solving it without those constraints.
concave function of the input power Ik . As shown in the

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
7

Lemma 4. Let {Pk? } denote the optimum matrices for the for the power scaling in Step 7 of the Algorithm 1 is to
problem (43)-(49) with dropped rank one constraints. Then, increase the minimum SINR that every UE attains. Since the
the rank-one matrices {Pk?? } defined as problem (43)-(49) does not take into account the max-min
q  SINR (constant t), the optimum solution of it may be scaled for
?
Pk11 a potential increase in t, which is useful in the next iterations
of the Algorithm 1.
  hq i
..
q
Pk?? , 
 
P 11 ? . . . P LL ? , k = 1, . . . , K,
q . k k

Note that in Step 9 of Algorithm 1, we change tmax to λt?

LL ?
Pk
(Here, λ > 1 is a scaling parameter to extend the bisection
(50) interval). The reason for this update is that after LSFD, it
constitute another optimum solution to the original problem may be possible to obtain feasible solution with t larger than
(43)-(49). the tmax that is set at the previous infeasible iterations. This
Proof: Please see the Appendix D. potential increase in the minimum SINR is shown with a green
arrow in Fig. 1. Due to this dynamic interval adaptation in
Using Lemma 4, we can obtain the global optimum solution the bisection search owing to LSFD, we call our algorithm a
of the problem (43)-(49) by removing the non-convex rank modified bisection search. We note that the objective function
constraints and solving it with convex programming. Hence, of the problem (37)-(41) is upper bounded as shown above
for a given set of the LSFD vectors, the original problem and an improved solution is obtained at each iteration. Hence,
in (37)-(41) can be shown to be quasi-convex and its global Algorithm 1 converges.
optimum solution can be found using bisection search over
t by solving a series of convex programming problems [10]. We note that the problem in (43)-(49) can be expressed
Furthermore, the LSFD vector ak only affects the SINR of in standard semidefinite programming form by expressing the
the k th UE and can be found in closed form for the given second-order cone constraints in terms of positive semidefinite
uplink power coefficients {ηi } by maximizing a generalized matrices and arranging the inequalities. Solving this problem
Rayleigh quotient [15]. Using these observations, we propose dominates the computational complexity of Algorithm 1. A
the alternating optimization algorithm which combines the semidefinite program can be solved by using polynomial time
closed-form LSFD vectors with the bisection search over algorithms and the interior point techniques are among the
minimum SINR as stated in Algorithm 1. most widely used methods.
It has a
worst-case complexity
Algorithm 1 is a modified bisection search algorithm over of Õ m m2 + nω + mns log (1/δ) where n, m, and s are
the minimum SINR. The reason we call it modified will be ex- the size of the positive semidefinite matrices, the number of
plained as follows. Fig. 1 shows an example search procedure inequality constraints, and the maximum number of non-zero
of Algorithm 1 for the first 8 iterations. The algorithm starts entries in each row of the weighting matrices, respectively, in
by setting the lower and upper bound for the minimum SINR, the standard formulation [31]. The parameter δ is the solution
i.e., tmin = 0 and tmax . The initial value of tmax in Algorithm accuracy and ω is the exponent of matrix multiplication.
1 can be taken as an upper bound on t for the problem (37)- Note that Õ notation corresponds to the big-O notation while
(41). A simple upper bound can be obtained by supposing excluding the logarithmic factors. After arranging the problem
there is only one UE in the setup and maximizing the SINR in (43)-(49) in a standard semidefinite programming

form,
of that UE. If we focus on the k th UE, the harvested  energy, we have n = K(L + 3), m = K L2 − L /2 + L + 4K,
Ek in (19) is maximized by setting pkl = ρd /E kwkl k2 and s = KL, that may lead to huge complexity as the
and pil = 0, ∀i 6= k by (10)-(11). Let Ek? denote the value number of APs and UEs in the network increases. The
of harvested energy for this setting. To maximize the SINRk , computational complexity of semidefinite programming can be
we equate the total uplink transmission energy for the k th UE reduced by the newly proposed methods [31] and by exploiting
to the harvested energy Ek? in (40) and obtain the data power the highly sparse structure of the most constraints such as
control coefficient as ηk? . We set all other uplink power control simple bound constraints. However, this is outside the scope
coefficients to zero, i.e., ηi = 0, ∀i 6= k. After maximizing of this paper. The main purpose of this paper is to demonstrate
the obtained generalized Rayleigh quotient for the k th UE, we the achievable performance in terms of max-min fairness
obtain SINR?k . If we repeat this procedure for each UE, we can with several channel estimation, transmission and reception
obtain a proper upper bound for the initialization of Algorithm strategies. We emphasize that the obtained closed-form results
1 as follows: can be exploited for simpler power control schemes as well.

tmax = min SINR?k . (51) Remark: We note that the steps of Algorithm 1 are intended
k
for coherent energy transmission and the non-linear energy
Then, the convex problem in (43)-(49) without rank constraints harvesting model in (19). For the linear energy harvesting
is solved for the fixed value of the LSFD vectors {ak } model, there is no need for introducing the optimization
and minimum SINR t. If the problem is not feasible, which variables {ek }. Similarly, the non-diagonal elements of the
is denoted by red double lines in Fig. 1, the conventional matrices Pk are not used for non-coherent energy harvesting
bisection search procedure is applied. However, if the problem model since the input power to the harvester is simply a linear
is feasible, which is denoted by blue double lines, then Steps function of {pkl }. Hence, for other scenarios, the optimization
6-9 in Algorithm 1 are applied consecutively. The motivation problem in (43)-(49) can be simplified accordingly.

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
8

0.03
(0)
tmin t(1) t(0)
max
0.025
(1) (2)
tmin t t(1)
max 0.02

(2)
tmin t(3) t(2)
max 0.015

(3) 0.01
tmin = t? t(4) t(3)
max = λt
?

(4)
tmin = t? t(5) t(4) ? 0.005
max = λt

(5) 0
tmin t(6) t(5)
max 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

(6)
tmin t(7) t(6)
max
Fig. 2. The non-linear energy harvesting model (M1) from [34] and the
(7)
tmin = t? t(8) t(7) ? corresponding linear energy harvesting model (L).
max = λt

Fig. 1. An example of the search procedure of the modified bisection search

(r−1)
in Algorithm 1. The leftmost and the rightmost black lines denote tmin
(r−1) th
and tmax , respectively, at the r iteration. The double lines show t (r) and square.1 For each setup, the UEs are randomly dropped and
its color is blue if the problem in Step 4 of Algorithm 1 is feasible and red a 4 m height difference between APs and UEs is taken into
otherwise. The green arrows denote the possible increase in the minimum
SINR after applying LSFD.
account when calculating distances. The noise variance is
σ 2 = −96 dBm. The uplink pilot transmission power is
−40 dBm. The total number of samples per coherence interval
Algorithm 1 Modified Bisection Search for Max-Min Fair
is τc = 200 with τp = 5, τd = 25, and τu = 170 unless
LSFD and Power Control
otherwise stated. The constant part of the LOS components
1: Initialization: Set tmin = 0 and tmax as in (51), respec-
are generated by assuming a uniform linear array in the far
tively. Initialize ak as all ones vector for k = 1, . . . , K.
field of the users with half wavelength antenna spacing. For
2: while tmax − tmin >  do .  > 0 determines the solution
each scenario, 500 random setups corresponding to different
accuracy.
user locations are considered where the number of UEs is
3: Set t = tmin +t
2
max
.
K = 20 unless otherwise stated.
4: Solve the convex problem in (43)-(49) without rank
constraints and by taking {ak } and t as constant. In the simulations, we consider both coherent and non-
5: if feasible then coherent energy transmission schemes, which are labeled
6: • Set the power control coefficients as the solution as C and NC in the figures, respectively. We consider the
of this problem. . the diagonal elements of the matrices linear and non-linear energy harvesting models in [32]. The
{Pk }. parameters for the non-linear energy harvesting model in (19)
7: • Scale all the power control coefficients {pkl } and are set to be the same for each UE. Two different non-linear
{ηk } so that at least one of the constraints in (39) and (40), models are considered to describe the saturation effect of a
respectively, are satisfied with equality. practical energy harvesting circuit. The parameters for the first
8: • Obtain the optimum {ak } by maximizing each model, which is labeled as M1 in the figures, are given as
UE’s SINR as a generalized Rayleigh quotient. Ak = 103 (ac − b), Bk = 106 c, and Ck = 103 c2 where
9: • Set tmin = t? and tmax = λt? where t? is the a = 0.3929, b = 0.01675, and c = 0.04401 are obtained
minimum of the SINRs after applying LSFD. in [32] by fitting to measurement data in [34]. As shown
10: else in Fig. 2, the harvested power starts to saturate at around
11: Set tmax = t. 0.1 mW. For most of the simulations, we consider this model.
12: end if To obtain a linear energy harvesting model as a benchmark
13: end while
to M1, we set Bk to zero and this selection always results in
14: Output: Downlink power coefficients {pkl }, uplink power
more harvested power as shown in Fig. 2. It is labeled as L in
coefficients {ηk }, LSFD vectors {ak }, minimum SINR t. the remaining figures. The second non-linear energy harvesting
model parameters are a = 2.463, b = 1.635, and c = 0.826,
which are obtained by fitting to measurement data in [35].
This model is labeled as M2 and the harvested power saturates
VII. N UMERICAL R ESULTS at approximately 10 times higher input power compared to
M1, as shown in Fig. 3. Hence, M2 behaves approximately
In this section, we will quantify the SE for different energy as a linear energy harvesting model around 0.1 mW where the
harvesting models and transmission schemes together with harvested power for M1 shows a high non-linear distortion.
various setups. The 3GPP indoor hotspot (InH) model in In the last part of the simulations, we will compare the SE
[33] is used with a 3.4 GHz carrier frequency and 20 MHz obtained by M1 and M2.
bandwidth. The large-scale fading coefficients, shadowing In the first scenario, we consider L = 36 APs, each
parameters, probability of LOS, and the Rician factors are
simulated based on [33, Table B.1.2.1-1, B.1.2.1-2, B.1.2.2.1- 1 The correlations between shadowing, terminal positions and Rician factors
4]. The APs are uniformly distributed in a 100 m×100 m in [33] are neglected for simplicity.

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
9

0.3

0.25

0.2

0.15

0.1

0.05

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 3. The non-linear energy harvesting model (M2) from [35]. Fig. 4. The CDF of SE per user for the proposed MMF and FPC in [16],
L = 36, N = 8, and LMMSE-based channel estimation.

with N = 8 antennas. The maximum power of each AP 1

is ρd = 10/36 W corresponding to 10 W of total maximum
power for the considered cell-free network. MMF stands for 0.8

the proposed max-min fairness optimization. We compare the 0.6

proposed max-min fairness optimization with a simpler power
control that is inspired by the fractional power control (FPC) 0.4
scheme for the downlink information transmission in [16]. For
this scheme, control coefficient pkl is proportional 0.2
r n the power o
2
to 1 / E |wkl | and they are scaled such that the total 0
0 0.5 1 1.5 2 2.5 3
transmission power is ρd for each AP in accordance with
the power control scheme [16]. Each UE’s power control
Fig. 5. The CDF of minimum SE per setup for the same scenario as Fig. 4.
coefficient ηk is adjusted such that the total uplink transmission
energy is equal to the harvested energy in the downlink.
The MR precoding and decoding vectors are obtained by the
LMMSE-based channel estimation scheme. quantify the impact of several parameters on the SE for the
In Fig. 4, we plot the cumulative distribution function (CDF) proposed MMF design.
of the individual SE per UE. We notice that the 90% and 95% In Fig. 6, we quantify the impact of the number of APs,
likely SE (i.e., where the CDF is 0.1 and 0.05, respectively) for L, and antennas per AP, N for the first non-linear energy
coherent energy transmission is 42% and 76% higher for the harvesting model in Fig. 2, M1. The maximum transmission
proposed MMF design in comparison to FPC for linear energy power for each AP is ρd = 10 / L W. Hence, maximum total
harvesting model. The 90% and 95% likely SE improvement transmit power for the whole AP network is 10 W for a fair
over FPC is around 28% and 54%, respectively for non-linear comparison. For the first six lines in Fig. 6, the number of total
energy harvesting. The higher improvement with the linear antennas throughout all the area is LN = 288. We notice
energy harvesting model can be explained as follows. Since that the 90% likely SE is improved by 35% and 16% for
more energy can be harvested with the linear model, the MMF coherent and non-coherent energy transmission, respectively,
optimization results in higher minimum SE among the users. by increasing the number of APs from L = 9 to L = 36.
However, for FPC, the UEs with good channel conditions However, there is a slight performance decrease in some
(corresponding to the upper tail of the CDF curves) are able regions of the CDF curve when we increase the number
to obtain higher SE compared to non-linear energy harvesting of APs by keeping the total number of antennas the same
since these users are expected to operate the saturation region especially for NC case. That the improvement is not visible
of M1 in Fig. 2 where the gap between M1 and L is higher as in increasing L = 9 to L = 16 is most probably due to the
compared to the left region of the graph. Hence, with FPC adverse effect of the increased local power constraints in (39).
that does not consider max-min fairness, the UEs with high However, if we increase the number of antennas per AP to
signal-to-noise ratio (SNR) benefit from the linear energy N = 16 for L = 36, we now see the positive impact of jointly
harvesting whereas the low-SNR UEs attain less SE in order to increasing the number of APs and total number of antennas,
compensate. On the other hand, coherent energy transmission LN , where each UE’s SE is significantly improved. Another
results in significantly higher SE compared to the non-coherent important observation from Fig. 6 is that the SE gap between
one due to the reasoning explained in Section IV. coherent and non-coherent energy transmission increases with
In order to see the fairness improvement of the proposed the number of APs since coherent combining of energy signals
algorithm, we plot the CDF of the minimum SE of all the transmitted from different APs supplies more power to be
UEs per setup in Fig. 5 for the same scenario. The minimum harvested at each UE under the same total power constraint.
SE of the network improves substantially and larger SE is In Fig. 7, we repeat the previous simulation with LS-
guaranteed for all the UEs. In the following experiments, we based channel estimation. Since LS-based estimation utilizes

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
10

of them for the LMMSE channel estimator. Hence, the closed-

form expressions will be different than the ones we have
derived in this paper. To quantify the difference between the
shared mutually orthogonal and random pilot sequences, we
use Monte Carlo estimation for the SE and average harvested
energy expressions.
In Fig. 8, we consider the non-linear energy harvesting
model M1 with coherent energy transmission and LMMSE-
based channel estimation. As the figure shows, there is a neg-
ligible performance difference between the two pilot signaling
schemes for the two considered pilot lengths: τp = 5 and
Fig. 6. The CDF of SE per user for LMMSE-based channel estimation. τp = 10. We note that the pilot contamination is not totally
undesirable thing in wireless-powered cell-free networks since
it also contributes to the harvested energy as can be seen from
(14)-(15) and thus there is an implicit trade-off. As a result, we
do not observe a substantial performance gap that may result
from the diversity of the random pilot scheme. However, the
shared pilot scheme has lower computational complexity since
we can apply the pilot de-spreading in (3) and this makes it
preferable compared to the case of random pilots.
In Fig. 9, we evaluate the SE per user for the scenario with
τp = 5 and different number of UEs, namely K = 40 and K =
60. We consider the non-linear energy harvesting model M1
and LMMSE-based channel estimation as before. As expected,
the SE that can be provided to the worst UE in the network
Fig. 7. The CDF of SE per user for LS-based channel estimation. decreases with the increase in the number of UEs. In very rare
scenarios, the SE is almost zero as can be observed from the
lower tail of CDF curves. This is because there is a potential
no channel statistics on the contrary to the LMMSE-based es- risk that the most unfortunate UE in the coverage area may
timation, significantly less SE is achieved compared to Fig. 6. have very bad channel conditions due to shadowing and the
However, it is possible now to improve the SE significantly other UEs are forced to reduce their power to maximize its SE.
by increasing the number of APs to L = 36 by keeping the However, it is possible to attain a practically reasonable SE
total number of antennas in the network fixed at LN = 288. for most of the cases and UEs. In the least favorable scenario
The 90% likely SE is 6 and 5 times greater compared to the with NC and K = 60 UEs, the median SE is 1 b/s/Hz and this
case of L = 9 APs for coherent and non-coherent energy value can be increased further by coherent energy transmission
transmission, respectively. In contrast to the LMMSE-based and deploying more antennas in the network. In fact, L = 36
channel estimation, increasing the number of APs results in APs with N = 8 antennas each (i.e., LN = 288 antennas
significantly higher SE for all the UEs. This may be due to in total) can provide satisfactorily high SE to K = 60 UEs,
that placing APs in a denser manner improves the channel which is a relatively high number of UEs for conventional
estimation quality of LS and the SE is continued to improve massive MIMO setups.
although the number of local power constraints has been In Fig 10, we plot the CDF curves of the SE per user for
increased. different number of downlink energy symbols, τd , by keeping
Throughout the following experiments, we consider the τp = 5 and τc = 200 constant as before. The number of
setup with L = 36 and N = 8. In Fig. 8, we compare the SE uplink symbols changes with τd as τu = 200 − 5 − τd . We
obtained when the UEs are either sharing a set of mutually consider only the non-linear energy harvesting model M1 and
orthogonal pilots or have randomly generated partially over- the coherent energy transmission. As can be seen from Fig. 10,
lapping pilots. The former case is the one we have used in the there is not a large performance difference for τd = 15, τd =
analytical part of this paper and it coincides with most previous 25, and τd = 45 for both channel estimation methods. This
works in the cell-free massive MIMO literature. In case of is due to the trade-off between harvested energy duration and
random pilot sequences [21], each UE’s pilot sequence is duration of uplink information transmission. As τd increases,
randomly generated from i.i.d. standard Gaussian distribution the harvested energy increases linearly with τd , however, τu
and is normalized such that its squared norm is τp . Unlike the decreases by reducing the pre-log factor in the SE formula.
i.i.d. Rayleigh fading channels that enable estimation of the Hence, an increase in harvested energy allows each UE to
channel gain of each AP antenna in an independent manner transmit with a higher uplink power but with a reduced uplink
[21], there is a common random phase shift that affects each information duration.
AP antenna jointly in the more general channel model in (1). In Fig. 11, we compare the SE performance of the two
The randomly generated pilot sequences require in this case non-linear energy harvesting models, i.e., M1 and M2, which
construction of N τp × N τp statistical matrices and inversion are presented in Figs. 2 and 3, respectively. For both channel

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
11

Fig. 8. The CDF of SE per user for shared and random pilot sequences with Fig. 11. The CDF of SE per user for different non-linear energy harvesting
different lengths. models.

Fig. 9. The CDF of SE per user for different number of UEs.

Fig. 12. The CDF of uplink data power for L = 36 and N = 8 with
non-linear energy harvesting model M1.

Fig. 10. The CDF of SE per user for non-linear energy harvesting model M1
and coherent energy transmission.

Fig. 13. The CDF of downlink power for L = 36 and N = 8 with non-linear
energy harvesting model M1.
estimation methods and energy transmission schemes, M2
that has a higher input power range before saturation effect,
provides a higher SE compared to M1. In the 0-0.1 mW
input power region in Fig. 2, M2 almost behaves as a linear to NC one since the maximum available power for the UEs
energy harvester with a greater harvested power, hence the are expected to increase with C downlink transmission. In
improvement in SE is not surprising. This shows us the addition, having a better channel estimate (LMMSE compared
importance of selecting the right energy harvesting circuit to LS) increases the uplink data power consistently. When it
parameters when analyzing these systems. comes to downlink power distribution, Fig. 13 plots the per
As a final simulation, we plot the CDF of uplink and AP and per UE symbol downlink transmission powers. For
downlink data transmission powers for the non-linear energy L = 36 APs, the maximum available power for each AP is
harvesting model M1 in Fig. 12 and Fig. 13, respectively. Note around 278 mW and for LMMSE-based channel estimation,
that the maximum uplink transmission power for each UE is there are some scenarios where some APs use its almost full
determined by the corresponding harvested power. As can be power for only one UE. However, this case does not happen
seen from Fig. 12, more uplink power is utilized to increase for LS-based channel estimation, probably to poorer channel
the minimum SE of the network for C transmission compared estimation quality.

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
12


VIII. C ONCLUSION Proof: Compute E yH Bx as
In this paper, the uplink SE of the wireless-powered cell- E yH Bx =αE xH Bx + E zH Bx
  
free massive MIMO has been derived for MR processing
(a)
with LMMSE and LS-based channel estimations and LSFD = αx̄H Bx̄ + ασx2 tr (B) (55)
at the CPU. The channels were assumed to follow a practical
Rician fading distribution with unknown phase-shifted LOS  we used the independence of zero-mean z and x, and
where
components in each coherence block. The UEs harvest energy E xxH = x̄x̄H + σx2 INnin (a). o
2
from the downlink RF signals and use a portion of it for the Let us compute now E yH Bx as
uplink data transmission. For a non-linear energy harvesting n o
2
E yH Bx = E αxH + zH BxxH BH (αx + z)


model, whose parameters can be fitted to different real data
measurements with saturation effects, the average harvested (a) 2  H
= α E x BxxH BH x + E zH BxxH BH z

energy has been derived for the LMMSE and LS-based chan-
nel estimations and two different energy transmission schemes: (b) 2 H
= α x̄ Bx̄x̄H BH x̄ + α2 σx2 E x̄H Bx̄wH BH w

coherent and non-coherent. Using the derived uplink SE and
+ α2 σx2 E x̄H BwwH BH x̄ + α2 σx2 E wH Bx̄x̄H BH w
 
harvested energy in the downlink, we optimized both the
+ α2 σx2 E wH Bwx̄H BH x̄ + α2 σx4 E wH BwwH BH w
 
downlink WPT and uplink WIT power control coefficients
together with the LSFD weights to maximize the minimum + tr B x̄x̄H + σx2 IN BH E zzH



guaranteed SE for all the UEs. An alternating optimization
(c) 2 H
= α x̄ Bx̄x̄H BH x̄ + α2 σx2 x̄H Bx̄ tr BH


algorithm is proposed for solving the non-convex problem. To
solve the resulting non-convex sub-problems efficiently, the + α2 σx2 x̄H BBH x̄ + α2 σx2 x̄H BH Bx̄
problem has been transformed into a new form with additional 2
+ α2 σx2 x̄H BH x̄ tr (B) + α2 σx4 |tr (B)| + α2 σx4 tr BBH


variables and constraints.
+ tr B x̄x̄H + σx2 IN BH Cy − α2 x̄x̄H + σx2 IN





Several simulations were carried out with practical non- ,
linear and conventional linear energy harvesting models. The (56)
results show that the proposed MMF algorithm significantly
where we used the independence of zero-mean z and x in
improves the fairness by providing greater SE to the weak-
(a) and (b).
 We have written all the non-zero individual
est UEs compared to another state-of-the-art power control
terms of E xH BxxH BH x separately by noting that θx is
scheme that was originally proposed for downlink informa-
independent of w and w is circularly symmetric in (b). We
tion transmission. Furthermore, coherent energy transmission
have used the cyclic shift property of trace and Lemma

5 in
increases the SE of each UE in comparison to its non-
(c) together with E zzH = Cy − α2 x̄x̄H + σx2 IN . After
coherent counterpart with an additional burden of downlink
arranging the terms in (56), we obtain the result in (54).
synchronization among the APs. The performance improve-
ment becomes more visible with the increase in the number
of APs that further improves the SE. We also note that the A PPENDIX B
effect of the downlink energy symbol length and the pilot P ROOF OF L EMMA 1
signaling scheme on the SE is not as significant as the other Let us calculate the average input power with LMMSE for
system parameters. the MR precoder wkl = ĝkl in (4). By using Lemma 6 with
√
x = gkl , x̄ = ḡkl , σx2 = βkl , y = zil , α = τp ρp , B =
√ −1
A PPENDIX A τp ρp Ψil Ril , and Cy = Ψil , we obtain
U SEFUL L EMMAS  H H
E ĝil gkl gkl ĝil =
Lemma 5. [36, Lemma 2]. Consider the random vector u ∈
2τp ρp βkl < ḡkl Ψil Ril ḡkl tr Ril Ψ−1
2 2
 H −1

CN that is distributed as u ∼ NC (0N , A). For a deterministic il
matrix B ∈ CN ×N , it holds that

2
+ τp2 ρ2p βkl
2
tr Ψ−1 + τp ρp tr Ψ−1


n o il Ril il Ril Rkl Ril ,
2 2
E uH Bu = |tr (AB)| + tr ABABH . i ∈ Pk .


(52) (57)
 H H
Lemma 6. Consider the vectors x = ejθx x̄ + σx w ∈ CN and For the expectation E ĝil gkl gkl ĝil for i ∈
/ Pk , ĝil and gkl
y = αx + z ∈ CN , where x̄ ∈ CN is deterministic and θx are independent and we have
is uniformly distributed in the interval [0, 2π). σx ≥ 0 and α  H
E ĝil H
gkl gkl
 H
ĝil = tr E ĝil ĝil

E gkl gklH


are some real deterministic scalars and w ∼ NC (0N , IN ) is  
(a)
independent of θx . z ∈ CN is a random vector independent of = tr R̂il Rkl , i ∈ / Pk , (58)
x and has zero-mean.  For a deterministic matrix B ∈ CN ×N
H
and given Cy , E yy , it holds that where we used (7) in (a). By using Lemma 6 with x = gkl ,
√ √
x̄ = ḡkl , σx2 = βkl , y = zil , α = τp ρp , B = τp ρp Ψ−1il Ril ,
E yH Bx =αx̄H Bx̄ + ασx2 tr (B) ,

(53) we obtain
n 2
o
H 2 2
 H H


E y Bx =2α σx < x̄ Bx̄ tr B  H
E ĝil gkl = τp ρp ḡkl H −1
Ψil Ril ḡkl + τp ρp βkl tr Ψ−1


il Ril ,
2
+ α2 σx4 |tr (B)| + tr B x̄x̄H + σx2 IN BH Cy . (54)



i ∈ Pk . (59)

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
13

 H
The expectation E ĝil gkl = 0 for i ∈ / Pk since ĝil and gkl A PPENDIX D
are independent and have zero mean. Let us now evaluate the P ROOF OF L EMMA 4
expectations for the MR precoder wkl = zkl in (3) for the LS- Let R denote the rank of Pk? for some k. In this case, Pk
based channel estimation. By using Lemma 6 with x = gkl , can be expressed as
√
x̄ = ḡkl , σx2 = βkl , y = zil , α = τp ρp , B = IN , and
R
Cy = Ψil , we obtain X
Pk? = xr xTr , (67)
r=1
E zH H H 2 2

il gkl gkl zil =2N τp ρp βkl ḡkl ḡkl + N τp ρp βkl
+ tr (Rkl Ψil ) , i ∈ Pk . (60) for some vectors xr ∈ RL . Let us now consider another Pk??
with rank one and equal diagonal elements with Pk? in (50).
 H
For the expectation E zil gkl gkl H
zil for i ∈ / Pk , zil and gkl Since Pk?? has rank one, it can be expressed as Pk?? = yy T
are independent and we have for some vector y ∈ RL such that
R
X
E zH H H H
  

il gkl gkl zil = tr E zil zil E gkl gkl x2rl = yl2 , l = 1, . . . , L, (68)
= tr (Ψil Rkl ) , i ∈ / Pk . (61) r=1

where xrl and yl denote the lth element of the vectors xr

By using Lemma 6 with x = gkl , x̄ = ḡkl , σx2 = βkl , y = zil , and y, respectively. The equality in (68) follows from the
√
α = τp ρp , B = IN , we obtain diagonal elements of Pk? and Pk?? being the same. Using
√ √ Cauchy-Schwarz inequality, it can be shown that
E zH H

il gkl = τp ρp ḡkl ḡkl + N τp ρp βkl , i ∈ Pk . (62) v v
R u R u R
ll0 ? 0 ??
X u X uX
2 x2 0 = y y 0 = P ll ,

The expectation E zH il gkl = 0 for i ∈
/ Pk since zil and P k = x x ≤0
t
rl rl x t rl rl l l k
gkl are independent and have zero mean. If we insert the r=1 r=1 r=1
expectations calculated above into (13), we obtain the results l 6= l0 . (69)
in (14) and (15).
Hence, all the off-diagonal elements of Pk?? and the harvested
energy for each UE is larger when Pk? is replaced by Pk??
A PPENDIX C without affecting other constraints and the objective function.
P ROOF OF L EMMA 2 Hence, this new solution with Pk?? in (50) is also optimum to
the considered problem.
Let us compute the expectations in the claim of Lemma 2
for the MR decoder vkl = ĝkl in (4). By using the result in R EFERENCES
(59), we obtain [1] Ö. T. Demir and E. Björnson, “Max-min fair wireless-powered cell-
 H free massive MIMO for uncorrelated Rician fading channels,” in IEEE
bkl = E ĝkl gkl Wireless Commun. Netw. Conf. (WCNC), May 2020, pp. 1–6.
[2] E. Björnson, L. Sanguinetti, H. Wymeersch, J. Hoydis, and T. L.
= τp ρp ḡkl Ψkl Rkl ḡkl + τp ρp βkl tr Ψ−1
H −1


kl Rkl . (63) Marzetta, “Massive MIMO is a reality—What is next?: Five promising
research directions for antenna arrays,” Digital Signal Processing, vol.
By using the results in (57) and (58) we have 94, pp. 3–20, Nov. 2019.
[3] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
MIMO for next generation wireless systems,” IEEE Commun. Mag., vol.
cll
 H H
ki =E ĝkl gil gil ĝkl 52, no. 2, pp. 186–195, Feb. 2014.
 H −1 [4] E. Björnson, J. Hoydis, and L. Sanguinetti, “Massive MIMO has
=2τp2 ρ2p βil < ḡil Ψkl Rkl ḡil tr Rkl Ψ−1


kl unlimited capacity,” IEEE Trans. Wirel. Commun., vol. 17, no. 1, pp.

2
+ τp2 ρ2p βil2 tr Ψ−1 kl Rkl
574–590, Jan. 2018.
[5] E. Björnson, J. Hoydis, and L. Sanguinetti, “Massive MIMO networks:
+ τp ρp tr Ψ−1


kl Rkl Ril Rkl , i ∈ Pk . (64) Spectral, energy, and hardware efficiency,” Found. Trends Signal Pro-
  cess., vol. 11, no. 3-4, pp. 154–655, 2017.
cll
ki = tr R̂ kl R il , i∈ / Pk . (65) [6] T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo, Fundamentals
of Massive MIMO. Cambridge: Cambridge University Press, 2016.
0
[7] Ö. Özdogan, E. Björnson, and E. G. Larsson, “Massive MIMO with
for l0 6= l is given
 H
The expectation cll H
ki = E ĝkl gil gil0 ĝkl0
spatially correlated Rician fading channels,” IEEE Trans. Commun.,
by using (59) as vol. 67, no. 5, pp. 3234–3250, May 2019.
[8] J. Zhang, E. Björnson, M. Matthaiou, D. W. K. Ng, H. Yang, D.
0

 J. Love, “Multiple antenna technologies for beyond 5G,” unpublished
H −1 −1
cll 2 2
ki = τp ρp ḡil Ψkl Rkl ḡil + βil tr Ψkl Rkl × paper, 2019. [Online]. Available: https://arxiv.org/abs/1910.00092
 ∗ [9] E. Nayebi, A. Ashikhmin, T. L. Marzetta and B. D. Rao, “Performance of
−1 −1 cell-free massive MIMO systems with MMSE and LSFD receivers,” in
H
, i ∈ Pk , l0 6= l,


ḡil 0 Ψ 0 Rkl0 ḡil0 + βil0 tr Ψ 0 Rkl0
kl kl 50th Asilomar Conference on Signals, Systems and Computers, Pacific
0 Grove, CA, 2016, pp. 203–207.
cll / Pk , l0 6= l.
ki = 0, i ∈ (66) [10] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta,
n o “Cell-free massive MIMO versus small cells,” IEEE Transac. Wirel.
H I

H Commun., vol. 16, no. 3, pp. 1834–1850, Mar. 2017.
The expectation dkl = E ĝkl nl nIl ĝkl is given by [11] E. Nayebi, A. Ashikhmin, T. L. Marzetta, H. Yang, and B. D. Rao, “Pre-
 
coding and power optimization in cell-free massive MIMO systems,”
dkl = σ 2 tr R̂kl using the independence of data noise and IEEE Trans. Wireless Commun., vol. 16, no. 7, pp. 4445–4459, Jul.
the channel estimate. 2017.

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.3036281, IEEE
Transactions on Wireless Communications
14

[12] H. Q. Ngo, L. Tran, T. Q. Duong, M. Matthaiou, and E. G. Larsson, “On [36] E. Björnson, M. Matthaiou, and M. Debbah, “Massive MIMO with non-
the total energy efficiency of cell-free massive MIMO,” IEEE Trans. ideal arbitrary arrays: Hardware scaling laws and circuit-aware design,”
Green Commun. Net., vol. 2, no. 1, pp. 25-–39, Mar. 2018. IEEE Tran. Wirel. Commun., vol. 14, no. 8, pp. 4353–4368, Aug. 2015.
[13] E. Björnson and L. Sanguinetti, “Making cell-free massive MIMO
competitive with MMSE processing and centralized implementation,”
IEEE Trans. Wirel. Commun., vol. 19, no. 1, pp. 77–90, Jan. 2020.
[14] J. Zhang, Y. Wei, E. Björnson, Y. Han, and S. Jin, “Performance analysis
and power control of cell-free massive MIMO Systems with hardware
impairments,” IEEE Access, vol. 6, pp. 55302–55314, 2018.
[15] Ö. Özdogan, E. Björnson, and J. Zhang, “Performance of cell-free
massive MIMO with Rician fading and phase shifts,” IEEE Trans. Wirel.
Commun., vol. 18, no. 11, pp. 5299–5315, Nov. 2019.
[16] G. Interdonato, P. Frenger, and E. G. Larsson, “Scalability aspects of
cell-free massive MIMO,” in IEEE Int. Conf. Commun. (ICC), Shanghai,
China, 2019.
[17] E. Björnson and L. Sanguinetti, “Scalable cell-free massive MIMO
systems,” IEEE Trans. Commun., vol. 68, no. 7, pp. 4247–4261, Jul.
2020.
[18] H. Tran and G. Kaddoum, “Green cell-less design for RF-wireless power
transfer networks,” in IEEE Wirel. Commun. Network. Conf. (WCNC),
Barcelona, 2018.
[19] R. Shrestha and G. Amarasuriya, “SWIPT in cell-free massive MIMO,”
in IEEE Glob. Commun. Conf. (GLOBECOM), Abu Dhabi, UAE, 2018.
[20] M. Alageli, A. Ikhlef, F. Alsifiany, M. A. M. Abdullah, G. Chen, and J.
Chambers, “Optimal downlink transmission for cell-free SWIPT massive
MIMO systems with active eavesdropping,” IEEE Trans. Inf. Forensics
Security, vol. 15, pp. 1983–1998, 2020.
[21] X. Wang, A. Ashikhmin, and X. Wang, “Wirelessly powered cell-free
IoT: Analysis and optimization, ” IEEE Internet of Things Journal, doi:
10.1109/JIOT.2020.2990378.
[22] B. Clerckx, A. Costanzo, A. Georgiadis, and N. Borges Carvalho,
“Toward 1G mobile power networks: RF, signal, and system designs
to make smart objects autonomous,” IEEE Microwave Mag., vol. 19,
no. 6, pp. 69–82, Sept.-Oct. 2018.
[23] Y. Zeng, B. Clerckx and R. Zhang, “Communications and signals design
for wireless power transmission,” IEEE Trans. Commun., vol. 65, no.
5, pp. 2264–2290, May 2017.
[24] G. Dong, H. Zhang, and D. Yuan, “Downlink achievable rate of massive
MIMO enabled SWIPT systems over Rician channels,” IEEE Commun.
Lett., vol. 22, no. 3, pp. 578–581, Mar. 2018.
[25] D. Kudathanthirige, R. Shrestha, and G. A. Aruma Baduge, “Max–min
fairness optimal rate-energy trade-off of SWIPT for massive MIMO
downlink,” IEEE Commun. Lett., vol. 23, no. 4, pp. 688–691, Apr. 2019.
[26] L. Zhao and X. Wang, “Massive MIMO downlink for wireless informa-
tion and energy transfer with energy harvesting receivers,” IEEE Trans.
Commun., vol. 67, no. 5, pp. 3309–3322, May 2019.
[27] X. Wang and C. Zhai, “Simultaneous wireless information and power
transfer for downlink multi-user massive antenna-array systems,” IEEE
Trans. Commun., vol. 65, no. 9, pp. 4039–4048, Sep. 2017.
[28] Z. Chang, Z. Wang, X. Guo, Z. Han, and T. Ristaniemi, “Energy-efficient
resource allocation for wireless powered massive MIMO system with
imperfect CSI,” IEEE Trans. Green Commun. Network., vol. 1, no. 2,
pp. 121–130, Jun. 2017.
[29] T. A. Khan, A. Yazdan, and R. W. Heath, “Optimization of power
transfer efficiency and energy efficiency for wireless-powered systems
with massive MIMO,” IEEE Trans. Wirel. Commun., vol. 17, no. 11, pp.
7159–7172, Nov. 2018.
[30] B. Khalfi, B. Hamdaoui, M. B. Ghorbel, M. Guizani, X. Zhang, and N.
Zorba, “Optimizing joint data and power transfer in energy harvesting
multiuser wireless networks,” IEEE Trans. Vehic. Tech., vol. 66, no. 12,
pp. 10989–11000, Dec. 2017.
[31] F. G. S. L. Brandao and K. M. Svore, “Quantum speed-ups for
solving semidefinite programs,” in IEEE 58th Annual Symposium on
Foundations of Computer Science (FOCS), Berkeley, CA, 2017, pp. 415-
426.
[32] Y. Chen, N. Zhao, and M. Alouini, “Wireless energy harvesting using
signals from multiple fading channels,” IEEE Trans. Commun., vol. 65,
no. 11, pp. 5027–5039, Nov. 2017.
[33] 3GPP, Further advancements for E-UTRA physical layer aspects (Re-
lease 9). 3GPP TS 36.814, Mar. 2017.
[34] M. Stoopman, S. Keyrouz, H. J. Visser, K. Philips, and W. A. Serdijn,
“A self-calibrating RF energy harvester generating 1V at -26.3 dBm,”
in Proc. Symp. VLSI Circuits (VLSIC), 2013, pp. C226–C227.
[35] T. Le, K. Mayaram, and T. Fiez, “Efficient far-field radio frequency
energy harvesting for passively powered sensor networks,” IEEE J.
Solid-State Circuits, vol. 43, no. 5, pp. 1287–1302, May 2008.

1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Gothenburg. Downloaded on December 21,2020 at 14:30:40 UTC from IEEE Xplore. Restrictions apply.