Quantized Power Allocation Algorithms in Cell-
Free Massive MIMO Systems
Bassant Amin1, Bassant Abdelhamid2, Salwa El-Ramly2
1 [Link]. student at Faculty of Engineering, Ain Shams University, Cairo, Egypt
2 Faculty of Engineering, Ain Shams University, Cairo, Egypt
the perspective of pre-coding methods such as Conjugate
Abstract—Cell-Free Massive Multiple Input Multiple Output Beamforming (CB), Zero-Forcing (ZF), in uncorrelated
(M-MIMO) is a system where a large number of low-cost single shadowing conditions without pilot contamination. On the other
antenna Access Points (APs) are distributed over a specific area.
hand, in [6], CB pre-coder with pilot contamination is studied
The APs are connected to a Network Controller (NC) and serve all
the existing users simultaneously. This work proposes two power under correlated shadowing channel model. In [7], the authors
allocation algorithms for Conjugate Beamforming (CB) and Zero- analyzed the performance of Cell-Free M-MIMO considering
Forcing (ZF) pre-coders in downlink scenario. They are based on the effects of channel estimation, non-orthogonality of pilot
quantizing the transmitted powers of APs to reduce the system sequences, power control and backhaul power consumption. In
complexity. The performance of the proposed algorithms is [8], for collocated M-MIMO system, they propose Zadoff-Chu
evaluated under the most common practical scenarios such as
Pilot Sequences (ZCPS) as the orthogonal uplink pilot
channel errors, pilot contamination and correlated shadowing
channel model. The results show that the proposed algorithms sequences for multi-cell TDD M-MIMO systems.
have closer solution to the optimal one for both pre-coders. The main contribution of this paper is to propose two power
Moreover, ZF pre-coder significantly outperforms CB pre-coder allocation algorithms for both pre-coders to maintain a good
with and without pilot contamination effect under both correlated downlink rate per-user (close to the optimal one). Unlike [4],
and uncorrelated shadowing channel models. the proposed algorithms do not need to involve convex
programming, which means they are more simplified i.e., low
Index Terms—Massive MIMO, Cell-Free, Conjugate
hardware and low cost.
Beamforming, Zero-Forcing, Power Allocation.
The paper is organized as follows: Section II presents the
I. INTRODUCTION Cell-Free system model. Section III studies the CB and ZF pre-
coders. Section IV illustrates the proposed power allocation
M assive Multi-Input Multi-Output (M-MIMO) has been
considered a promising 5G wireless access technology
[1]. In M-MIMO, a Base Station (BS) is mounted with a large
algorithms for both pre-coders. Finally, Sections V and VI
represent the numerical results and the conclusion.
number of antennas, which can be either collocated in the BS II. SYSTEM MODEL
or distributed all over the covered area. The collocated M-
MIMO systems provide less coverage at the advantage of less Consider ܯsingle antenna APs be distributed randomly over
backhaul requirements. While the distributed M-MIMO is more a specific area serving ܭsingle antenna randomly distributed
users where ܭبܯ. The APs are connected to one NC, these
efficient due to space diversity. Thus, the distributed M-MIMO
connections are used to exchange the payload data and power
systems can provide much higher coverage at the cost of higher
coefficients which slowly change. Instantaneous Channel State
backhaul requirements. M-MIMO with the distributed antenna-
Information (CSI) is not shared between the APs and the NC.
setup is called a Cell-Free M-MIMO system. The uplink pilots are sent by the users to the APs to estimate all
In Cell-Free M-MIMO, the selected pre-coders and power the channels in the training phase. The APs use these channel
allocation algorithms are the most important parameters which estimates to pre-code the transmitted data in the downlink and
influence the system performance. In [2], the authors detect the data in the uplink. Moreover, APs do not co-operate
investigated the uplink rate performance when the quantized and no channel estimates are shared between them. The Cell-
version of the estimated channel and the quantized received
signal are available at the NC using Maximum Ratio
Combining (MRC) detector. The authors proposed to
decompose the original max-min non convex problem into two
sub-problems (the receiver filter coefficient design, and power
allocation). Based on these sub-problems, an iterative algorithm
was developed, which solves the original problem using
singular value decomposition and convex programming. In [2-
3], the authors considered power allocation algorithms using
max-min criterion for cell-free M-MIMO. In [4,5], the authors
analyzed the performance of cell-free M-MIMO systems from Fig. 1 Cell-Free System Model
978-1-5386-9230-1/18/$31.00 2018
c IEEE 35
Free system model is shown in Fig. 1. The channel coefficient A. Conjugate Beamforming (CB) Pre-coder
between the m-th AP and user k is as follows [4]: The m-th AP transmits the signal ݔ as follows [6]:
݄ ൌ ඥߚ ݄ത ሺͳሻ
where ߚ is the large-scale fading coefficient which expresses ݔ ൌ ඥ ඥߟ ݄
כ
ݑ ሺͷሻ
the path loss and the shadowing effects of the surrounding ୀଵ
environment depending on the distance between m-th AP and where ݑ is the transmitted data signal to user k, (ܧሺȁݑ ȁଶ ሻ ൌ
user k ሺ݀ ሻ. ݄ത ̱ܰܥሺͲǡͳሻ is the small-scale fading ͳ), is the maximum transmission power of each AP. The
coefficient. It is assumed to be identically independent factor ߟ is the power coefficient which the m-th AP has
distributed (i.i.d) complex random variables which are assigned to user k. The power coefficients ߟ are the factors
constants over a coherence time interval duration and to be controlled to raise the system performance according to
independent of each other in different coherence intervals. ߚ the power constraint equation (ܧሺȁݔ ȁଶ ሻ ሻ. Then, the
is considered frequency independent, while ݄ത has frequency
received signal at the k-th user can be written as [6]:
dependence. It is considered that ࡴ ܥ אெൈ ǡ ݄ is the ெ
element of this matrix at row m and column k.
ݕ ൌ ඥ ඥߟ ᇲ ݄ ݄ כᇱ ݑ ᇲ ݖ ሺሻ
The used protocol in channel estimation is as follows [4,6]:
ୀଵ ᇲ ୀଵ
First, all K users synchronously and simultaneously send their
where ݖ is the additive noise ̱ܰܥሺͲǡͳሻ at user k. In general,
pilot sequences ࣐ଵ ǡ ǥ ǡ ࣐ of length ߬ to all APs. Second, all
the power constraint equation for CB pre-coder is
APs calculate the channel coefficient estimates ሺ݄ ሻ of ݄ .
σୀଵ ߟ ߛ ͳǡ ሺ݉ሻ. The downlink rate for CB pre-coder
Finally, APs beamform the data back to all users using these
estimates. Pilot sequences are assumed to be orthonormal. can be written as:
Two pilot scenarios are considered, the first scenario assumes ܴ ൌ ݈݃ଶ ሺͳ ܴܵܰܫ ሻ ሺሻ
that the pilot sequences are generated using Zadoff-Chu method where ܴܵܰܫ of CB pre-coder [6] is in (8).
[8] with length ߬ ൌ ܭ, so there exist ܭperfectly-orthogonal
pilot sequences and no pilot contamination occurs. The second B. Zero Forcing (ZF) Pre-coder
scenario assumes that some of the pilot sequences are generated
The main advantage of ZF pre-coder (ࢂி ) is cancelling the
using Zadoff-Chu method but with length ߬ ൏ ܭ, so pilot
interference caused by a given user to the rest of the users. The
contamination exists. The ߬ ൈ ͳ received pilot vector in the
training phase at the m-th AP can be written as [4,6]: interference cancellation is done by introducing the pseudo
inverse matrixࢂ ൌ ࡴ ்ࡴ
כ൫ࡴ כ൯ିଵ , where ൣࡴ
൧ ൌ ݄ is the
࢟ǡ ൌ ඥ ߬ ݄ ࣐ ࢠ ሺʹሻ estimation of the channel matrix H. By using the pseudo inverse
ୀଵ matrix V ܥ אெൈ , the interference is completely canceled
where is the normalized signal-to-noise ratio of each uplink ் ࢂ ൌ ࡵ ). This can be done by making ࡴ
் ࢂி diagonal, thus
(ࡴ
pilot symbol, ߬ is the pilot sequence length, and ࢠ ̱ܰܥሺͲǡ ࡵ߬ ሻ the power coefficients are equal i.e.,ߟଵ ൌ ڮൌ ߟெ ൌ ߟ ,
is the vector of the additive noise received at the m-th AP. Based meaning that ߟ should be only function of user k, and we
on the pilot signal ࢟ǡ , AP m computes the estimate of the ଵ
have:ࢂி ൌ ࢂࡼ ൗ, where P is a diagonal power allocation
channel coefficient ݄ . Let ݕǡ be the projection of ݕǡ
matrix with ߟଵ ǡ ǥ ǡ ߟ on its diagonal. The received signal at the
onto ߮ு [6]:
k-th user is [4]:
ݕ ൌ ඥ ߟ ݑ ඥ ࢋ் ࡴ כ൫ࡴ כ൯ିଵ ࡼ࢛ ݖ ሺͻሻ
்ࡴ
ݕǡ ൌ ࣐ு ு
࢟ǡ ൌ ࣐ ቌඥ ߬ ݄ ࣐ ࢠ ቍሺ͵ሻ ೖ
ୀଵ where ࢎ is the actual channel column vector for user k and the
The MMSE estimate of ݄ given ݕǡ is ݄ ൌ ݎ ݕǡ channel error vector ࢋ்ೖ ൌ ࢎ െ ࢎ , where ࢎ is the estimated
where [6]: channel vector between user k and all APsǤ ࢛ is the column data
ඥ ߬ߚ vector of all users. The achievable rate per user can be
ݎ ሺͶሻ calculated using (7), but for ZF pre-coder, the SINR can be
߬ σ ᇲୀଵ ߚ ᇲ ȁ࣐ு ଶ
࣐ ᇲ ȁ ͳ
calculated as [4]:
ߟ
Let ݁ೖ ൌ ݄ െ ݄ be the channel estimation error with ܴܵܰܫ ൌ ሺͳͲሻ
ͳ σ ୀଵ ߟ ߱
zero mean and variance ߚ െ ߛ . ݄ has zero mean with where ߱ is the i-th element of the following vector:
variance ߛ [6], where ߛ ൌ ඥ ߬ߚ ݎ Ǥ Due to channel
࣓ ൌ ݀݅ܽ݃ ቄ ܧቀ൫ࡴ כ൯ିଵ ࡴ
்ࡴ ் ܧ൫ࢋכ ࢋ் ൯ࡴ כ൫ࡴ
்ࡴ כ൯ିଵ ቁቅ ሺͳͳሻ
hardening [9], user ݇ is assumed to know only the statistics of ೖ ೖ
its own estimated channels. where ࣓ ܥ אଵൈ Ǥ Let the vector ࢻ ܥ אଵൈ be [4]:
ࢻ ൌ ݀݅ܽ݃ ቄ ܧቀ൫ࡴ ்ࡴ כ൯ିଵ ࢎ ሾሿ ு ൫ࡴ
ሾሿ ࢎ ்ࡴ כ൯ିଵ ቁቅ ሺͳʹሻ
III. PRECODING IN CELL-FREE M-MIMO ்
where ࢎሾሿ is the m-th column of the matrix [ࡴ], let ࢜ሾሿ be the
In this section, the downlink performance of a cell-free M- m-th row of the matrix ࢂி then the transmitted power by the m
MIMO system with two pre-coders: Conjugate Beamforming -th AP is where ൌ σ ் ு
ୀଵ ߟ ߙ ǡ ݉. ሺǤ ሻ ǡ ሺǤ ሻ and
and Zero-Forcing are studied under uncorrelated / correlated ሺǤ ሻ כare the transpose, hermitian and complex conjugate
shadowing conditions with / without pilot contamination. operations respectively.
36 International Japan-Africa Conference on Electronics, Communications and Computations (JAC-ECC)
ଶ
൫σெ
ୀଵ ඥߟ ߛ ൯
ܴܵܰܫ ൌ ሺͺሻ
ߚ ଶ ு ଶ
ͳ σ ெ ெ
ᇲ ୀଵ σୀଵ ߟᇱ ߛᇱ ߚ σᇱஷ ൬σୀଵ ඥߟᇱ ߛᇱ ൰ ࣐
ߚ ᇲ ห࣐ ᇲ ห
of quantization levels. The optimization problem to be solved
IV. PROPOSED POWER ALLOCATION (PA) ALGORITHMS using the Heuristic Quantized algorithm is:
The world is heading into reducing the power consumption in
the network systems. Thus, APs should not transmit the signal ܴܵܰܫி ሺͳͶܽሻ
ఎೖ
in full power mode all the time unless it is needed. When the
estimated channel is strong between a user and an AP, this AP ݏǤ ݐ ߟ ߙ ൌ ݍǢ ݉ ൌ ͳǡ ǥ ǡ ܯ ሺͳͶܾሻ
should not send much power to that user. From this perspective, ୀଵ
the idea of quantizing the AP power with the knowledge of the where ݍis a constant ሺͲ ൏ ݍ ͳሻ. First, all APs starts to
channel statistics was proposed in the present work. For transmit the signal in 50% power mode (= ݍ0.5). It’s inspired
example, the user with a high ߚ does not need high power, from [4] that the user can improve and maximize its rate by
so the algorithm assigns him a low quantization level and vice selecting the AP with the strongest channel and calculate the
versa. power coefficient as follows:
ݍ
A. Fixed Quantized PA for CB Precoding (FQPA-CB) ߟ ൌ Ǣ Ͳ ൏ ݍ ͳ ሺͳͷሻ
σୀଵ ߙ
The proposed algorithm computes the power coefficients ߟ
The algorithm is described as follows:
as function of ߛ (to take the effect of channel estimation) and
Algorithm (2): Heuristic Quantized Power Allocation
a quantization vectorሺிொ ሻ, which is a vector containing
ଵ 1: Initialize: quantization level ுொ ൌ ሾͲǤͷǡͲǤͷǡ ǥ ǡͲǤͷሿ,
elements between and 1 in an ascending order. The values of with length M, (50% power mode), i=0, ߤ ൌ
ெ
these elements were chosen to let the difference in powers for ܿݐ݊ܽݐݏ݊.
ଵ
APs to be integers of . The quantization vector depends on 2: Compute ߟ using (15).
ெ
3: Compute ܴܵܰܫ using (10) for all users.
the strength of the channel between the m-th AP and user k. This
4: Find the user ݇ כwith the minimum SINR, where:
vector represents the portion of the maximum power that will ݇ כൌ ܽ ݃ݎ୩
be assigned to the AP. The algorithm can be illustrated as
ಹೂ ሺାଵሻ
follows: 5: Letுொ ሺ݅ ͳሻ ൌ ͳ െ ߤ݅, ߟ כൌ , i = i+1.
௫ σ಼
ೕసభ ఈೕ
Algorithm (1): Fixed Quantized Power Allocation
ଵ ଶ 6: Repeat steps 3 to 5 until all APs assigned a power
1: Initialize: Quantization vector ிொ ൌ ሾ ǡ ǡ ǥ ǡͳሿ. coefficient.
ெ ெ
2: For each AP ݉ǡ It is clear that Algorithm (1) and (2) are more simplified
௫
Calculate ߚ ൌ ߚ Ǣ ݇ ൌ ͳǡ ǥ ǡ ܭ alternative for solving the optimal power allocation using
Let ࢼ௫ ={ߚଵ௫ ǡ ǥ ǡ ߚெ
௫
ሽ convex programming, with the guarantee of good performance
3: ௫
Sort ࢼ in descending order to get ࢼ௫ with close to the optimal one, as will be shown in the next section.
ordered AP index vector ᇱ ൌ ሾ݉ଵᇱ ǡ ǥ ǡ ݉ெ
ᇱ
ሿ.
4: For i=1:M V. NUMERICAL EVALUATION AND DISCUSSION
Compute ߟᇲ as follows: Consider the designated area be 4 ଶ . For the large-scale
ிொ ሺ݅ሻ fading coefficientsߚ , Cost Hata propagation model is used
Ʉᇲ ൌ ሺͳ͵ሻ whereߚ ൌ ͳͲǤଵሺିଵଷିଷହభబሺௗೖ ሻାௌೖ ሻ , where ݀ is in
σୀଵ ߛᇲ ݆
kilometers. In case of the uncorrelated shadowing scenario,
End For
ܵ ̱ܰሺͲǡ ߪ௦ௗ ଶ ሻ where ߪ௦ௗ ൌ ͺ dB. The maximum
5: Compute ܴܵܰܫ using (8) for all users.
transmission power of each AP antenna and user is and
It is worthy noted that the power coefficients ሺɄ ǡ ݉ሻ is
ǡrespectively where
തതത ൌ തതത
ൌ ʹͲͲ mW. തതത and
തതത are
computed at the NC, which has the information of the sorted
large-scale fading coefficients. normalized over the noise variance which is ߪଶ ൌ ܶ ൈ ݇ ൈ
ܹܤൈ ܰܨ, where ܶ ǡ ݇ ǡ ܨܹܰ݀݊ܽܤ, are the noise
B. Heuristic Quantized PA for ZF Pre-coder (HQPA-ZF) temperature (ܶ ൌ ʹͻͲ Kelvin), Boltzmann constant,
This work proposes an algorithm which maximizes the bandwidth ( ܹܤൌ ʹͲ MHz), and noise figure (ܰ ܨൌ ͻdB).
minimum downlink rate. It predefines a set of quantization Number of APs is (M=128), number of users is (K=16) and
levels in vector (ܙୌ୕ ) where its elements change from 0 to 1. pilot sequence length in case of pilot contamination is (τ=8).
For example if the element in this vector equals 0.1, this means Moreover, two components are used to form the correlated
that the total transmitted power from the AP to all users equals shadowing coefficients as in [6], ܵ ൌ ξߜܽ
to 10% of its maximum power. If this element equals 1, this
ξͳ െ ߜܾ Ǣ ݉ǡ ݇, where ܽ ̱ܰሺͲǡͳሻ and ܾ ̱ܰሺͲǡͳሻ are
means that the AP operates in its full power mode. Every AP is
independent random variables and Ͳ ߜ ͳ is a parameter
constrained by its power budget, which is specified by these set
International Japan-Africa Conference on Electronics, Communications and Computations (JAC-ECC) 37
Fig. 2 CDF of the downlink rate per user for CB pre-coder with/without pilot Fig. 3 CDF of the downlink rate per user for ZF pre-coder with/without pilot
contamination for correlated/uncorrelated shadow fading. contamination for uncorrelated/correlated shadow fading.
expressing the domination of both components on the was proposed. Both algorithms have low complexity where
shadowing coefficients. ܽ models the contribution to the they do not solve any optimization problems. Moreover, it is
shadow fading that comes from the neighboring objects to observed that CB pre-coder is more robust in the correlated
the݉-th AP. ܾ models the contribution to the shadow fading shadowing conditions than the ZF pre-coder.
that comes from the near objects to the k-th user. The orthogonal
pilot sequences are generated using Zadoff-Chu equation [8]. REFERENCES
For performance criteria, the 5%-outage rate is used, which is [1] T. Marzetta, E. G. Larsson, H. Yang and H. Q. Ngo,
the smallest rate among 95% of the best users [4]. Algorithms Fundamentals of Massive MIMO, Cambridge University
(1) and (2) take 0.05 and 0.2 second per realization respectively. Press, 2016.
It is worthy noting that Algorithm (2) takes more time than [2] M. Bashar, K. Cumanan, A. G. Burr, H. Q. Ngo, and M.
Algorithm (1) due to its iterative nature. Debbah ,“Cell-Free Massive MIMO with Limited
In Fig. 2, Cumulative Distribution Function (CDF) of the Backhaul.” Department of Electronic Engineering,
average downlink rate per user of Algorithm (1) is shown in University of York, Heslington, York, UK.
uncorrelated / correlated shadowing channel conditions with / arXiv:1801.10190 [[Link]], Jan 2018.
without pilot contamination for M=128, K=16,߬=8. It is [3] M. Bashar, K. Cumanan, A. G. Burr, M. Debbah, and H.
observed that the proposed algorithm can achieve 80% of the Q. Ngo, “Enhanced Max-Min SINR for Uplink Cell-Free
optimal performance in case of no Pilot Contamination (PC). Massive MIMO Systems.” Department of Electronic
The difference between the performance of the uncorrelated Engineering, University of York, Heslington, York, UK.
and correlated shadowing rates is not much. This might be Jan 2018.
explained as the SINR Eq.(8) of the CB pre-coder is less [4] E. Nayebi, A. Ashikhmin, T. L. Marzetta, H. Yang and B.
dependent on the large-scale fading coefficients (ߚ ሻ, which D. Rao, "Precoding and Power Optimization in Cell-Free
is directly affected by the correlated shadowing channel model. Massive MIMO Systems," in IEEE Transactions on
It is shown that, in uncorrelated shadowing condition, the 5% Wireless Communications, vol. 16, no. 7, pp. 4445-4459,
outage rate is degraded from 2.3 bits (in case of no pilot July 2017.
contamination) to 0.7 bit (with pilot contamination). [5] Q. Huang and A. Burr, "Compute-and-forward in cell-free
Fig. 3 presents the CDF of the downlink rate per user of massive MIMO: Great performance with low backhaul
Algorithm (2). The used step is Ɋ ൌ ͲǤͲͷ. It can be shown that load," 2017 IEEE International Conference on
ZF pre-coder achieves a 5% outage rate of 5.8 bits, hence it Communications Workshops (ICC Workshops), Paris,
significantly-outperforms the CB pre-coder. Moreover, in 2017, pp. 601-606, Nov 2017.
correlated shadowing conditions, the performance of ZF pre- [6] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson and T.
coder was degraded to 3.6 bits, which means that CB pre-coder L. Marzetta, "Cell-Free Massive MIMO Versus Small
is more robust in correlated shadowing conditions than ZF pre- Cells," in IEEE Transactions on Wireless
coder. In uncorrelated shadowing conditions, pilot Communications, vol. 16, no. 3, pp. 1834-1850, March
contamination caused performance degradation from 5.8 to 4.5 2017.
bits, while in correlated shadowing, the degradation was from [7] H. Q. Ngo, L. N. Tran, T. Q. Duong, M. Matthaiou and E.
3.6 to 2.8 bits. The ZF pre-coder is not robust in correlated G. Larsson, "On the Total Energy Efficiency of Cell-Free
shadowing as its SINR in (10) is more dependent on the large- Massive MIMO," in IEEE Transactions on Green
scale channel coefficients than that in CB pre-coder. Communications and Networking, vol. 2, no. 1, pp. 25-39,
Mar 2018.
VI. CONCLUSION [8] S. Ali, Z. Chen, and F. Yin, “Design of Orthogonal
Uplink Pilot Sequences for TDD Massive MIMO under
Two power allocation algorithms were proposed for CB and Pilot Contamination.” Journal of Communications Vol. 12,
ZF pre-coders achieving performance close to the optimal one. No. 1, pp. 40-48, Jan 2017.
The results show that the ZF pre-coder outperforms the CB one [9] T. L. Marzetta, “Noncooperative cellular wireless with
with/without pilot contamination. The (FQPA) algorithm was unlimited numbers of base station antennas,” IEEE Trans.
proposed for CB pre-coder, its 5% outage rate is 80% of the Wireless Commun., vol. 9, no. 11, pp. 3590–3600, Nov.
optimal performance. For ZF pre-coder, the (HQPA) algorithm 2010.
38 International Japan-Africa Conference on Electronics, Communications and Computations (JAC-ECC)
