Cell-Free Massive MIMO Versus Small Cells

NGO et al.
: CELL-FREE MASSIVE MIMO VERSUS SMALL

1
Cell-Free Massive MIMO Versus Small Cells

Hien Quoc Ngo, Alexei Ashikhmin, Hong Yang, Erik G. Larsson, Fellow, IEEE,
and Thomas L. Marzetta, Fellow, IEEE,
Abstract— A Cell-Free Massive MIMO (multiple-input

multiple-output) system comprises a very large number of arrays at the base stations can be deployed in collocated or
distributed access points (APs), which simultaneously serve a distributed setups. Collocated Massive MIMO architectures,
much smaller number of users over the same time/frequency where all service antennas are located in a compact area,
resources based on directly measured channel characteristics. have the advantage of low backhaul requirements. In contrast,
The APs and users have only one antenna each. The APs in distributed Massive MIMO systems, the service antennas
acquire channel state information through time-division duplex
operation and the reception of uplink pilot signals transmitted are spread out over a large area. Owing to their ability
by the users. The APs perform multiplexing/de-multiplexing to more efficiently exploit diversity against the shadow
through conjugate beamforming on the downlink and matched fading, distributed systems can potentially offer much higher
filtering on the uplink. Closed-form expressions for individual probability of coverage than collocated Massive MIMO [4],
user uplink and downlink throughputs lead to max–min power at the cost of increased backhaul requirements.
control algorithms. Max–min power control ensures uniformly
good service throughout the area of coverage. A pilot assignment In this work, we consider a distributed Massive MIMO
algorithm helps to mitigate the effects of pilot contamination, but system where a large number of service antennas, called
power control is far more important in that regard. Cell-Free access points (APs), serve a much smaller number of
Massive MIMO has considerably improved performance with autonomous users distributed over a wide area [1]. All
respect to a conventional small-cell scheme, whereby each user is
APs cooperate phase-coherently via a backhaul network,
served by a dedicated AP, in terms of both 95%-likely per-user
throughput and immunity to shadow fading spatial correlation. and serve all users in the same time-frequency resource via
Under uncorrelated shadow fading conditions, the cell-free time-division duplex (TDD) operation. There are no cells or
scheme provides nearly fivefold improvement in 95%-likely per- cell boundaries. Therefore, we call this system “Cell-Free
user throughput over the small-cell scheme, and tenfold Massive MIMO”. Since Cell-Free Massive MIMO combines
improvement when shadow fading is correlated.
the distributed MIMO and Massive MIMO concepts, it is
Index Terms— Cell-Free Massive MIMO system, conjugate expected to reap all benefits from these two systems. In
beamforming, massive MIMO, network MIMO, small cell.
addition, since the users now are close to the APs, Cell-Free
Massive MIMO can offer a high coverage probability.
I. INTRODUCTION Conjugate beamforming/matched filtering techniques, also
ASSIVE multiple-input multiple-output (MIMO), known as maximum-ratio processing, are used both on
M where a base station with many antennas
simultaneously serves many users in the same time-frequency
uplink and downlink. These techniques are computationally
simple and can be implemented in a distributed manner, that
resource, is a promising 5G wireless access technology that is, with most processing done locally at the APs.1
can provide high throughput, reliability, and energy efficiency In Cell-Free Massive MIMO, there is a central processing
with simple signal processing [2], [3]. Massive antenna unit (CPU), but the information exchange between the APs
and this CPU is limited to the payload data, and power
Manuscript received August 3, 2015; revised February 22, 2016, control coefficients that change slowly. There is no sharing
August 25, 2016, and December 16, 2016; accepted January 5, 2017. Date
of publication January 19, 2017; date of current version March 8, 2017.
of instantaneous channel state information (CSI) among the
The work of H. Q. Ngo and E. G. Larsson was supported in part by APs or the central unit. All channels are estimated at the
the Swedish Research Council and in part by the ELLIIT. Part of this APs through uplink pilots. The so-obtained channel estimates
work was presented at the 16th IEEE International Workshop on Signal are used to precode the transmitted data in the downlink
Processing Advances in Wireless Communications, Stockholm, Sweden,
June 2015 [1]. Portions of this work were performed while H. Q. Ngo and to perform data detection in the uplink. Throughout we
was with the Bell Laboratories, Murray Hill, NJ 07974 USA. The associate emphasize per-user throughput rather than sum-throughput.
editor coordinating the review of this paper and approving it for publication To that end we employ max-min power control.
was M. Vu.
H. Q. Ngo is with the Department of Electrical Engineering (ISY), In principle, Cell-Free Massive MIMO is an incarnation of
Linköping University, 581 83 Linköping, Sweden, and also with the School of general ideas known as “virtual MIMO”, “network MIMO”,
Electronics, Electrical Engineering and Computer Science, Queen’s “distributed MIMO”, “(coherent) cooperative multipoint joint
University Belfast, Belfast BT3 9DT, U.K. (e-mail: hien.ngo@liu.se).
A. Ashikhmin, H. Yang, and T. L. Marzetta are with the Nokia Bell Labs, processing” (CoMP) and “distributed antenna systems”
Murray Hill, NJ 07974 USA (e-mail: alexei.ashikhmin@nokia-belllabs.com; (DAS). The objective is to use advanced backhaul to achieve
h.yang@nokia-bell-labs.com; tom.marzetta@nokia-bell-labs.com). coherent processing across geographically distributed base
E. G. Larsson is with the Department of Electrical Engineer-
ing (ISY), Linköping University, 581 83 Linköping, Sweden (e-mail: 1
Other linear processing techniques (e.g. zero-forcing) may improve the
erik.g.larsson@liu.se).
Color versions of one or more of the figures in this paper are available system performance, but they require more backhaul than maximum-ratio
online at http://ieeexplore.ieee.org. processing does. The tradeoff between the implementation complexity and
Digital Object Identifier 10.1109/TWC.2017.2655515 the system performance for these techniques is of interest and needs to be
studied in future work.
1536-1276 © 2017 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: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 01,2024 at 11:37:56 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 3, MARCH 2017
2
station antennas, in order to provide uniformly good ser- systems. Most previous work compares collocated Massive
vice for all users in the network. The outstanding aspect of MIMO and small-cell systems [25], [26]. In [25], the authors
Cell-Free Massive MIMO is its operating regime: many show that, when the number of cells is large, a small-cell sys-
single- antenna access points simultaneously serve a much tem is more energy-efficient than a collocated Massive MIMO
smaller number of users, using computationally simple system. By taking into account a specific transceiver hardware
(conjugate beamforming) signal processing. This facilitates impairment and power consumption model, paper [26] shows
the exploita- tion of phenomena such as favorable propagation that reducing the cell size (or increasing the base station
and channel hardening – which are also key characteristics of density) is the way to increase the energy efficiency. However
cellular Mas- sive MIMO [5]. In turn, this enables the use of when the circuit power dominates over the transmission
computationally efficient and globally optimal algorithms for power, this benefit saturates. Energy efficiency comparisons
power control, and simple schemes for pilot assignment (as between collocated massive MIMO and small-cell systems are
shown later in this paper). In summary, Cell-Free Massive also studied in [27] and [28]. There has however been little
MIMO is a useful and scalable implementation of the network work that compares distributed Massive MIMO and small-
MIMO and DAS concepts – much in the same way as cell systems. A comparison between small-cell and distributed
cellular Massive MIMO is a useful and scalable form of the Massive MIMO systems is reported in [12], assuming perfect
original multiuser MIMO concept (see, e.g., [5, Chap. 1] for CSI at both the APs and the users. Yet, a comprehensive
an extended discussion of the latter). performance comparison between small-cell and distributed
A. Related Work Massive MIMO systems that takes into account the effects
of imperfect CSI, pilot assignment, and power control is not
Many papers have studied network MIMO [6], [8], [9] and available in the existing literature.
DAS [7], [10], [11], and indicated that network MIMO and
DAS may offer higher rates than colocated MIMO. However, B. Specific Contributions of the Paper
these works did not consider the case of very large numbers of • We consider a Cell-Free Massive MIMO with conjugate
service antennas. Related works which use a similar system beamforming on the downlink and matched filtering on
model as in our paper are [12]–[18]. In these works, DAS the uplink. We show that, as in the case of collocated
with the use of many antennas, called large-scale DAS or systems, when the number of APs goes to infinity, the
distributed massive MIMO, was exploited. However, in all effects of non-coherent interference, small-scale fading,
those papers, perfect CSI was assumed at both the APs and and noise disappear.
the users, and in addition, the analysis in [18] was asymptotic • We derive rigorous closed-form capacity lower bounds
in the number of antennas and the number of users. A realistic for the Cell-Free Massive MIMO downlink and uplink
analysis must account for imperfect CSI, which is an with finite numbers of APs and users. Our analysis takes
inevitable consequence of the finite channel coherence in a into account the effects of channel estimation errors,
mobile system and which typically limits the performance of power control, and non-orthogonality of pilot sequences.
any wireless system severely [19]. Large-scale DAS with • We compare two pilot assignment schemes: random
imperfect CSI was considered in [20]–[23] for the special case assignment and greedy assignment.
of orthogonal pilots or the reuse of orthogonal pilots, and in • We devise max-min fairness power control algorithms
[24] assuming frequency-division duplex (FDD) operation. that maximize the smallest of all user rates. Globally
In addition, in [20], the authors exploited the low-rank optimal solutions can be computed by solving a
structure of users’ channel covariance matrices, and examined sequence of second-order cone programs (SOCPs) for the
the performance of uplink transmission with matched- downlink, and a sequence of linear programs for the
filtering detection, under the assumption that all users use uplink.
the same pilot sequence. By contrast, in the current paper, • We quantitatively compare the performance of Cell-Free
we assume TDD operation, hence rely on reciprocity to Massive MIMO to that of small-cell systems, under
acquire CSI, and we assume the use of arbitrary pilot uncorrelated and correlated shadow fading models.
sequences in the network – resulting in pilot contamination,
The rest of paper is organized as follows. In Section II,
which was not studied in previous work. We derive rigorous
we describe the Cell-Free Massive MIMO system model.
capacity lower bounds valid for any finite number of APs and
In Section III, we present the achievable downlink and uplink
users, and give algorithms for optimal power control (to
rates. The pilot assignment and power control schemes are
global optimality) and pilot assignment.
developed in Section IV. The small-cell system is discussed
The papers cited above compare the performance between
in Section V. We provide numerical results and discussions in
distributed and collocated Massive MIMO systems. An alter-
Section VI and finally conclude the paper in Section VII.
native to (distributed) MIMO systems is to deploy small Notation: Boldface letters denote column vectors. The
cells, consisting of APs that do not cooperate. Small-cell sys- superscripts ()∗, ()T , and ()H stand for the conjugate, trans-
tems are considerably simpler than Cell-Free Massive MIMO, pose, and conjugate-transpose, respectively. The Euclidean
since only data and power control coefficients are exchanged norm and the expectation operator s are d enoted by ∗ · ∗
between the CPU and the APs. It is expected that Cell- and 0,
denotes
E {·}, respectively. Finally, z ∼
a
Massive MIMO systems perform better than small-cell sys-

NGO et al.: CELL-FREE MASSIVE MIMO VERSUS SMALL
symmetric complex Gaussian random variable (RV) z with 3
tems. However it is not clear, quantitatively, how much Cell- zero mean and variance σ 2, and z N (0,σ 2) denotes a real-
Free Massive MIMO systems can gain compared to small-cell ∼
valued Gaussian RV.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 3, MARCH 2017
4
perfect calibration of the hardware chains. The feasibility

of the latter is demonstrated for example in [31] for
collocated Massive MIMO and it is conceivable that the
problem can be similarly somehow for Cell-Free Massive
MIMO. Investigating the effect of imperfect calibration is
an important topic for future work.
• We let gmk denote the channel coefficient between the
kth user and the mth AP. The channel gmk is modelled
as follows:
gmk = β 1/2 h mk , (1)
m
where hmk represents the small-scale fading, and βmk
represents the large-scale fading. We assume that h mk ,
m 1, . . . , M, K 1, . . . K , are independent and iden-
Fig. 1. Cell-Free Massive MIMO system. = =
tically distributed (i.i.d.) CN (0, 1) RVs. The justification
of the assumption of independent small-scale fading is
II. CELL-FREE MASSIVE MIMO SYSTEM MODEL
that the APs and the users are distributed over a wide
We consider a Cell-Free Massive MIMO system with M area, and hence, the set of scatterers is likely to be
APs and K users. All APs and users are equipped with a different for each AP and each user.
single antenna, and they are randomly located in a large area. • We assume that all APs are connected via perfect back-
Furthermore, all APs connect to a central processing unit via haul that offers error-free and infinite capacity to the
a backhaul network, see Figure 1. We assume that all M APs CPU. In practice, backhaul will be subject to signif-
simultaneously serve all K users in the same time-frequency icant practical constraints [32], [33]. Future work is
resource. The transmission from the APs to the users (down- needed to quantify the impact of backhaul constraints on
link transmission) and the transmission from the users to the performance.
APs (uplink transmission) proceed by TDD operation. Each • In all scenarios, we let qk denote the symbol asso-
coherence interval is divided into three phases: uplink ciated with the kth user. These symbols are mutually
training, downlink payload data transmission, and uplink independent, and independent of all noise and channel
payload data transmission. In the uplink training phase, the coefficients.
users send pilot sequences to the APs and each AP estimates
the channel to all users. The so-obtained channel estimates are
used to precode the transmit signals in the downlink, and to A. Uplink Training
detect the signals transmitted from the users in the uplink. In The Cell-Free Massive MIMO system employs a wide
this work, to avoid sharing of channel state information spectral bandwidth, and the quantities gmk and hmk are
between the APs, we consider conjugate beamforming in the dependent on frequency; however βmk is constant with
downlink and matched filtering in the uplink. respect to frequency. The propagation channels are assumed
No pilots are transmitted in the downlink of Cell-Free to be piece-wise constant over a coherence time interval
Massive MIMO. The users do not need to estimate their and a frequency coherence interval. It is necessary to per-
effective channel gain, but instead rely on channel hardening, form training within each such time/frequency coherence
which makes this gain close to its expected value, a known block. We assume that βmk is known, a priori, wherever
deterministic constant. Our capacity bounds account for the required.
error incurred when the users use the average effective Let τc be the length of the coherence interval (in
channel gain instead of the actual effective gain. Channel samples), which is equal to the product of the coherence
hardening in Massive MIMO is discussed, for example, in time and the coherence bandwidth, and let τ cf be the uplink
[2]. training duration (in samples) per coherence interval, where
Notation is adopted and assumptions are made as follows: the superscript cf stands for “cell-free”. It is required that τ cf <
• The channel model incorporates the effects of small-scale τc. During the training phase, all K users simultaneously send
fading and large-scale fading (that latter includes path pilot sequences √
cf
loss and shadowing). The small-scale fading is assumed of length τ samples to the APs. Let τ cf ϕk
cf
Cτ
×1 ∈
to be static during each coherence interval, and change , where ϕk 2 1, be the pilot
∗ ∗
independently from one coherence interval to the next. sequence used by the kth user,
The large-scale fading changes much more slowly, and k 1, 2, , K . Then, the τ cf 1 received pilot vector at the
= ··· ×
stays constant for several coherence intervals. Depending mth AP is given by
on the user mobility, the large-scale fading may stay
constant for a duration of at least some 40 small-scale √ K
cf cf
fading coherence intervals [29], [30]. y = p τ ρ g
k=1
• We assume that the channel is reciprocal, i.e., the channel ϕ + w,signal-to-noise
where ρcfp is the normalized (2)p,m ratio (SNR)mkof
gains on the uplink and on the downlink are the same. This reciprocity assumption requires TDD operation and
5
each pilot symbol and wp,m is a vector of additive noise at
the
mth AP. The elements of wp,m are i.i.d. CN (0, 1) RVs.
1
Based on the received pilot signal yp,m, the mth AP esti-
where
mates the channel gmk , k = 1, . . . , K . Denote by yˇp,mk
√
the projection of yp,m onto ϕk :
H γmk E .gˆm τ ρcf βmkcmk. (8)
.2 p
= cf
yˇp,mk = ϕ H kyp,m k
The received signal at the kth user is given by
K
√ √
= τ cf ρcf gmk + τ cf ρcf gmkrϕ H ϕkr + ϕ H wp,m.
p p k k √M M
Although, for arbitrary pilot sequences,
k
r yˇ is not a sufficient rd,k gmkxm + 1/2 ∗
(
statistic for the estimation of gmk, one can still use this quantity =m d mk
to obtain suboptimal estimates. In the special case when any m=1 kr=1
K
two pilot sequences are either identical p,or orthogonal, then ρ CN (0, 1)η noise
where wd,k is additive r gmk at
gˆ mk
the k r +user.
r qkth wd,k ,Then
yˇp,mk is a sufficient statistic, and estimates based on yˇp,mk qk will be detected from rd,k.
are optimal. The MMSE∗ estimate of gmk given yˇp,mk is
E yˇ g
p ,mk mk
gˆ =
=c (4) C. Uplink Payload Data Transmission
yˇ yˇ
mk p,mk mk p,mk,
E . p,mk 2 In the uplink, all K users simultaneously send their data
. to the APs. Before sending the data, the kth user weights its
yˇ √
where symbol qk, E |qk|2 = 1, by a power control coefficient ηk,
√
cmk cf ρcf β
p mk . 0 ≤ ηk ≤ 1. The received signal at the mth AP is given by
τ
H
τ cf ρcf βmk r .ϕ ϕ k r . + 1
K p kr k 1 √ K √
yu,m g ηkqk + wu,m,
so Remark
that they1:are τ ≥c K , then
If pairwisely we can choose
orthogonal, and hence, , 2· ·second
ϕ , ϕ the · ,ϕK ρu k=1
term in (3) disappears. Then the channel estimate gmk is inde-
ˆ where ρcf is the normalized uplink SNR and wu,m is additive
pendent of gmkr , kr k. However, owing to the limited length u
/ noise at the mth AP. We assume that wu,m CN (0, 1).
cf
of the coherence interval, in general, τ < K , and mutually ∼
To detect the symbol transmitted from the kth user, qk, the
non-orthogonal pilot sequences must be used throughout the
mth AP multiplies the received signal yu,m with the conjugate
network. The channel estimate gmk is degraded by pilot signals
ˆ of its (locally obtained) channel estimate gmk. Then the so-
transmitted from other users, owing to the second term in (3). ∗ ˆ
This causes the so-called pilot contamination effect. obtained quantity gm kˆ yu,m is sent to the CPU via a backhaul
Remark 2: The channel estimation is performed in a decen- network. The CPU sees
tralized fashion. Each AP autonomously estimates the M
∗
channels to the K users. The APs do not cooperate on the ru,k = gˆ m k yu,m
channel estimation, and no channel estimates are m=1
interchanged
the APs. among √
K
M M
∗ ∗
= ρ cf ηk r gˆ gmk r qk r + gˆ wu,m . (11)
B. Downlink Payload Data Transmission u mk mk
kr=1 m=1 m=1
The APs treat the channel estimates as the true channels,

Then, qk is detected from ru,k.
and use conjugate beamforming to transmit signals to the K
users. The transmitted signal from the mth AP is given by
K III. PERFORMANCE ANALYSIS
√ cf
xm = ρ 1/2 η∗ gˆ mk qk , (5)
d m A. Large-M Analysis
k k=1
where qk, which satisfies E qk 2 1, is the symbol intended
| | =
for the kth user, and ηmk, m 1, . . . , M, k 1, . . . K ,
= =
are power control coefficients chosen to satisfy the following
power constraint at each AP: {
2 cf
E |x d | →
≤ ρ(1),
With the channel model in . the power constraint
2 cf
E |x d| ≤ρ
K
can be rewritten as:m
ηmkγmk ≤ 1, for all m, (7)
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 3, MARCH
1
In this section, we provide some insights into the perfor-
mance of Cell-Free Massive MIMO systems when M is very
large. The convergence analysis is done conditioned on a set
of deterministic large-scale fading coefficients βmk . We show
that, as in the case of Collocated Massive MIMO, when
M , the channels between the users and the APs
become orthogonal. Therefore, with conjugate beamforming
respectively matched filtering, non-coherent interference,
small-scale fading, and noise disappear. The only remaining
impairment is pilot contamination, which consists of
interference from users using same pilot sequences as the
user of interest in the training phase.
1
On downlink, from (9), the received signal at the kth user non-orthogonality:
can be written as: √
τ cf ρcf ρcf M
r p d 1
−
d √ ρ η m m k
1/2 ∗ M M ηmk cmkβmkq⎞k
r = cf gˆ m
g q M K
m 1= ⎠ →P
z ¸, s + η1/2 c r β ϕ T ϕ∗q r 0. (17)
D m m m k k
1/2 ∗ k
cf
d ρ M K η r gmk gˆ
m k r qk r +wd,k , (12) m=1 kr /=k
M
m=1 kr /=k If the H pilot sequences are pairwisely orthogonal,
√z i.e., ϕ ϕk = 0 for k /= kr, then the received signal
+ ¸ , s becomesr free of interference and noise:
M U Ik √
r ,k τ cf ρpcf ρcfd M 1/2 P
where DSk and MUIk represent the desired signal and d ηm cmkβmk
M M → 0.
multiuser interference, respectively. m M →∞
By using the channel estimates in (4), we have Similar results hold on the uplink.
M
1/2
r
mk B. Achievable Rate for Finite M
m=1
∗ ∗ In this section, we derive closed-form expressions for the
ηM1/2η g K
= ϕ ϕ + w˜ downlink
technique and
fromuplink achievable
[21]. rates, using the analysis
rr rr r
c rg
τ cf ρcf
√ [35] and
g H
m=1 mk
r mk mk p mk kr k p,mk
krr =1 1) Achievable Downlink Rate: We assume that each user
√
1/2 has knowledge of the channel statistics but not of the channel
= τ cf ρcf η r |g ϕ ϕ∗ k
2 T
p m m|
m k realizations. The received signal rd,k in (9) can be written
c
√ mK K
M
+ τ cf ρ cf η1/2 c r
g∗ ϕT ϕ∗ rd,k = DSk · qk + BUk · qk + UIkkr · qkr + wd,k, (19)
g
p mkr mk mk mkrr kr krr
r
krr /=k m=1 k /=k
M where
∗
+ η1/2 c r g w˜ , (13) √ M
m mk mk p
, mk r
,DSk m
m ρcf
d 1/2
where w˜ r ϕHw . Then by Tchebyshev’s m=1 η gmk
p, 2 kp 1/2∗ ∗ 1/2 ∗
theorem [34], we have √ ηMm gkmk gˆ
−Egˆ ηM gmk gˆ ,
cf
BUk ρ
√
M 1
gˆ
∗ — τ cf ρ cf η1/2 c r mk mk
1 1/2 β d m=1 m=1 mk mk
η mkgr ϕT r ϕ∗ (21)
M p cf M
gˆ
m=1 mk mkr
M
r
m=1 mk mk
mk k k UI r
√ρ η g ∗ , (22)
1/2
P kk d mkr mk mkr
m=1
→ 0. (14)
M represent the strength of desired signal (DS), the beamform-
Using (14),→∞
we obtain the following results:
ing gain uncertainty (BU), and the interference caused by
M the krth user (UI), respectively.
√DSk − cf cf ρcf
1 τ→ ρ1 η cmkβmkqk P 0, (15) We treat the sum of the second, third, and fourth terms
M M 1/2 p d mk
M →∞ in (19) as “effective noise”. Since qk is independent of DSk
m=1 and BUk, we have
1
MUI − √ M
τ cf ρpcf ρdcf ηK1/2 c r β ϕT ϕ∗q r ∗ ∗ 2
k
M M
1 mk kr k k E DSk · qk × (BUk · qk) = E DSk × E |qk| = 0.
mkr mk
BUk
m=1 kr/=k The above expressions show that when M ,
→P → the received signal includes only the desired signal
0. (16)
M →∞ plus interference originating from the pilot sequence
1
Thus, the first and the second terms of (19) are uncorrelated. uncorrelated. By using the fact that uncorrelated Gaussian
A similar calculation shows that the third and fourth noise represents the worst case, we obtain the following
terms of (19) are uncorrelated with the first term of (19). achievable rate of the kth user for Cell-Free (cf) operation:
Therefore, the effective noise and the desired signal are
2
Tchebyshev’s theorem: Let X1, X2, ...Xn be independent RVs such that
E {Xi } = μi and Var {Xi } ≤ c < ∞, ∀i. Then |DSk|2
Rcfd = log 1+
1 1 P
(2(X1 + X2 + ... + Xn) − (μ1 + μ2 + ...μn ) →
n n
1
We next provide a new exact closed-form expression for the
achievable rate (23), for a finite M.
Theorem 1: An achievable downlink rate of the transmis-
sion from the APs to the kth user in the Cell-Free Massive
MIMO system with conjugate beamforming, for any finite
M and K , is given by (24), shown at the top of the next
page.
Proof: See Appendix A.
Remark 3: The main differences between the capacity
bound expressions for Cell-Free and collocated Massive
MIMO systems [3] are: i) in Cell-Free systems, in general
βmk /= βmr k , for m /= mr, whereas in collocated Massive
MIMO, βmk βmr k ; and ii) in Cell-Free systems, a power
=
constraint is applied at each AP individually, whereas in
collocated systems, a total power constraint is applied at each
base station. Consider the special case in which all APs are
collocated and the power constraint for each AP is replaced
by a total power constraint over all APs. In this case, we have
βmk = βmr k βk, γmk = γmrk γk, and the power control
Fig. 2. Achievable rate versus the number of APs for different K . Here,
coefficient is ηmk = ηk /(Mγmk). If, furthermore, the K pilot ρdsc = 10 dB, ρpcf = 0 dB, τ cf = K , βmk = 1, ηmk = 1/(K γmk ), and pilot
sequences are pairwisely orthogonal, then, (24) becomes
sequences are pairwisely orthogonal.
R cf
= log2 cf
Mρ γkηk
k +1 d
, (25)
d
k
K 11 η + Remark 5: In the special case that all APs are collocated
r
ρcfd β k
which is identical to the rate expression for collocated Massive and all K pilot sequences are pairwisely orthogonal, then
MIMO systems in [3]. βmk = βmrk βk, γmk = γmr k γk, and ϕ H ϕkr = 0, ∀kr /= k.
k
Remark 4: The achievable rate (24) is obtained under the Equation (27) then reduces to
assumption that the users only know the channel statistics. Mρcf ηkγk
u
+ K
However, this achievable rate is close to that in the case Rcfk = log2 1 , (28)
u
where the users know the actual channel realizations. This is a ρcf η k r β kr + 1
u kr=1
consequence of channel hardening, as discussed in Section II. which is precisely the uplink capacity lower bound of obtained
a single-
To see this more quantitatively, we compare the achievable cell Massive MIMO system with a collocated array
rate (24) with the following expression, in [21], and a variation on that in [36].
. 12 IV. PILOT ASSIGNMENT AND POWER CONTROL
⎧ ⎛
⎪ .⎞⎫
⎪
K M
M
R ˜ cf = log2⎜1 + ρcfd
2
⎪⎬
E k
. m=1 2 To obtain good system performance, the available radio
η/ gmk ⎟
. ∗ ,
gˆ
mk . mk
η g . ⎟
. ⎠⎪
ρ
d, resources must be efficiently managed. In this section, we
⎪ ⎝ .
⎪⎭
1 . ∗ IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 3, MARCH

m=1 mk gˆ m k r . 6) (2
+1
will present methods for pilot sequence assignment and power

cf 1/2
r
mk
≥
control. Importantly, pilot assignment and power control can
kr /=k
be performed independently, because the pilots are not power
1
that knows the instantaneous channel gain. Figure 2 shows

which represents an achievable rate for a genie-aided user
1
only know the channel statistics, and the genie-aided rate (26),
a comparison between (24), which assumes that the users
1
figure, the gap is small, which means that downlink training

which assumes knowledge of the realizations. As seen in the
1
2) Achievable Uplink Rate: The central processing unit

is not necessary.
1
that the central processing unit uses only statistical knowl-

detects the desired signal qk from ru,k in (11). We assume
1
a similar methodology as in Section III-B.1, we obtain a

edge of the channel when performing the detection. Using
1
as follows.
rigorous closed-form expression for the achievable uplink rate
1
the Cell-Free Massive MIMO system with matched filtering

Theorem 2: An achievable uplink rate for the kth user in
1
top of the next page.

detection, for any M and K , is given by (27), shown at the
1
A. Greedy Pilot Assignment
controlled.
1
sequences, due to the limited length of the coherence interval.

Typically, different users must use non-orthogonal pilot
1
orthogonal pilot sequences. Here we focus on the case that

Since the length of the pilot sequences is τ cf , there exist τ cf
1
sequences to the K users.

τ cf < K . If τ cf K , we simply assign K orthogonal pilot
1
length τ cf samples to the K users is random pilot assign-

A simple baseline method for assigning pilot sequences of
1
randomly assigned one pilot sequence from a predetermined

ment [37]. With random pilot assignment, each user will be
1
ment could alternatively be done by letting each user choose

set Sϕ of τ cf orthogonal pilot sequences. Random pilot assign-
1
of pilots). However, it appears from simulations that the latter

an arbitrary unit-norm vector (i.e. not from a predetermined set
1
1
⎛ !M "2 ⎞
cf 1/2
ρd m=1 η m γmk ⎟
c ⎜
⎟
d
K k dK M m m m
mk m
k
Rd,k = log2 ⎜1 +
! "2
M ⎠, (24)
⎝ βmk +1
ρcf η
1/2 r
r
k m γ |ϕ H ϕ |2 + η r γ r β
r
⎛ ρcf k =1
! "2 ⎞
M
⎜ ρucf ηk
cf
⎜ K
"
mk
γ
⎟
K M M ⎟
u m=1 u
β k
+u,k = log2
R 1
ρcf k
r
! γmk
βmkr k
r
m
⎝ M
ηkr ηkr γmkβmkr +

2 , (27)
⎠
γmk
m
m
|ϕ H ϕkr |2 + ρcf
is a useful baseline, occasionally two users in close vicinity

scheme does not work well. While random pilot assignment
∗ . ∗
. mkr
1
which iteratively refines the pilot assignment. The K users
of each other will use the same pilot sequence, which results
are first randomly assigned K pilot sequences. Then the
in strong pilot contamination.
user that has the lowest downlink rate, say user k∗, updates
its pilot sequence so that its pilot contamination effect
Optimal pilot assignment is a difficult combinatorial
is minimized.3 The pilot contamination effect at the k∗th
user is quantified by the second term in (3) which has

problem. We propose to use a simple greedy algorithm,
variance
1
B. Power Control
unit significantly. Furthermore, since ϕk∗ is chosen from Sϕ,
We next show that Cell-Free Massive MIMO can provide
to inform the users about their assigned pilots, the CPU only
uniformly good service to all users, regardless of their geo-
graphical location, by using max-min power control. While
needs to send an index to each user.
power control in general is a well studied topic, the max-
1
Free Massive MIMO are entirely new. The power control is

min power control problems that arise when optimizing Cell-
performed at the CPU, and importantly, is done on the large-

⎫
. K .2 ⎪⎬ K
.k r /=k ∗ . = βmkr ϕ H ϕkr
⎪ ⎭ r ∗
k /=k
# $
ϕ H M K2 Rcf
β H1 /=k ϕ rϕ H ϕ
r∗
ϕk∗ . .
k∗ k m
2
where we used M theK fact that ϕk 1. The algorithm
∗
then arg
proceeds
min iteratively
β for
ϕk∗ a predetermined number of $2
∗
iterations.ϕk m=1 m k k { } k=1,···
d
"2 γ ς
k
,K γ β #ς K
∗
! M
3 k m k m βmkmr
= arg
In principle, this “worst user” could be taken to be the user that has either kr/=k M kr=1 mk cf
the lowest uplink or the lowest downlink rate. In our numerical experiments, , K
d
we4Hence
reassign
thistherecomputation
pilot of the user having theeven
is infrequent lowest downlink
in high rate,For
mobility. hence giving
example,
at user mobility of v 100 km/h, and a∗ ∗ =
carrier frequency of f c 2 GHz,
the channel coherence time is on the order of a millisecond. The large-scale
d m
fading changes much more slowly, at least some 40 times slower according
to [29], [30]. As a result, the greedy pilot assignment method must only be
H 2
m=1 mk mk
The pilot assignment is recomputed on the large-scale
K M
mk
ξ m
+ β γmkr ς 2 r + ρ
= =
where ξkkr |ϕkr ϕ k | . 1
1
1
Algorithm 1 Greedy Pilot Assignment Algorithm 2 Bisection Algorithm for Solving (34)
1) Initialization: choose K pilot sequences ϕ1, · , ϕ K 1) Initialization: choose the initial values of tmin and tmax,
using the random pilot assignment method. Choose the where tmin and tmax define a range of relevant values of
number of iterations, N , and set n 1. the objective function in (34). Choose a tolerance γ > 0.
= + tmax
2) Compute Rcf , using (24). Find the user with the lowest 2) Set t tmin
. Solve the following convex feasibility
rate: d : 2
program:
k∗ = arg min Rcf . (33)
∗v ∗ ≤ √1 M
k = 1, . . . , K,
k
⎪γmk ςmk ,
k t
k ϕ K
m
∗
3) Update theS pilot
ϕ ∗ from whichsequence for the k th user by choosing
minimizes ⎪ m=1
M K ⎨⎪ βmk ς 2 r ≤ ϑ2 , rm = 1, . . . , M,
γmk r
2
βmkr .ϕ.kH∗ ϕkr . . ⎪ kr γmkr βmkςr mkr ≤ αkr k , ∀k /= k,
⎪0 ≤
⎪ = ϑm ≤ 1, m = 1, . . . , M,
m=1 (3
m=1 kr /=k∗
⎩ςmk ≥ 0, k = 1, . . . , K, m = 1, . . . , M,
4) Set n := n + 1. Stop if n > N . Otherwise, go to Step % &T 1
2.
where v k
, and where
k √ v vT I−k
k k1
d
ρ
By introducing slack variables αkr k and ϑm , we reformu- . .T

ϕ H ϕ k α1k ... ϕ H ϕ K α K k , I−k is a K × (K − 1)
late (32) as matrix
1 K
.√
obtained from the K ×.K √ identity matrix with the kth
# $2 k2 β1kϑT1column
... β Mk ϑ M
M
γ ς 3) If problem (35) is feasible, then set tmin := t, else set
m m removed, and v .
M K
m=1 mk mk tmax :=
{ςmk ,αkr k,ϑm } K
k=1,··· , |ϕ ϕ H
| 2 2
+ 1 4) Stop if tmax − tmin < γ. Otherwise, go to Step 2.
2
+
r
α kr
k /=k ϕkr ϕk k ρ
K βmkϑm cf
m=1 d
subject to γ r ς 2
≤
M ϑ , m = 1, . . . ,
2
where Rcfk is given by (27). Problem (36) can be equivalently
mk r u,
mk m
kr=1 reformulated as
M
γ r βmk ς r ≤α r , ∀kr /= max t
k
m=1 mk mk kk {ηk },t
βmkr subject to t ≤ Rcf , k = 1, . . . , K
0 ≤ ϑm ≤ 1, m = 1, . . . , u,k
M
ςmk 0 ≤ ηk ≤ 1, k = 1, . . . , K. (37)
≥ 0, k = 1, . . . , K, m = 1, . . . , M. (34)
Proposition 2: The optimization problem (37) is quasi-
linear.
The equivalence between (32) and (34) follows directly from where the K users are weighted according to priority:
the fact that the first and second constraints in (34) hold with max min wk Rk , where wk > 0 is the weighting factor of
equality at the optimum. the kth user. A user with higher priority will be assigned a
Proposition 1: The objective function of (34) is quasi- smaller weighting factor.
concave, and the problem (34) is quasi-concave. 2) Uplink: In the uplink, the max-min power control
Proof: See Appendix B. problem can be formulated as follows:
Consequently, (34) can be solved efficiently by a bisection
search, in each step solving a sequence of convex feasibility
problem [38]. Specifically, Algorithm 2 solves (34).
Remark 7: The max-min power control problem can be
directly extended to a max-min weighted rate problem,
{
1
Proof: From (27), for a given t, all inequalities involved rate expressions, and max-min power control for small-cell
in (37) are linear, and hence, the program (37) is quasi- systems. These will be used in Section VI where we compare
linear. the performance of Cell-Free Massive MIMO and small-cell
systems.
Consequently, Problem (37) can be efficiently solved by using For small-cell systems, we assume that each user is served
bisection and solving a sequence of linear feasibility problems. by only one AP. For each user, the available AP with the
largest average received useful signal power is selected. If an
AP has already been chosen by another user, this AP becomes
V. SMALL-CELL S YSTEM
unavailable. The AP selection is done user by user in a
In this section, we give the system model, achievable random
max order. Let mk be the AP chosen by the kth user. Then,
{ηk } min Rcf
u,k
k=1,··· , mk arg max βmk. (38)
K
subject to 0 ≤ ηk ≤ 1, k = 1, . . . , K, (36) m∈{available APs}
1
We consider a short enough time scale that handovers between constraint. Channel estimation at the user is required in order
APs do not occur. This modeling choice was made to enable to demodulate, as there is no channel hardening (see
a rigorous performance analysis. While there is precedent for discussion above).
this assumption in other literature [12], [39], future work may 1) Achievable Downlink Rate: Treating the last three terms
address the issue of handovers. As a result of this assumption, of (41) as uncorrelated effective noise, we obtain the achiev-
the performance figures we obtain for small-cell systems may able downlink rate for the kth user as in (42), shown at the
be overoptimistic. top of the next page.
In contrast to Cell-Massive MIMO, in the small-cell sys- Since the channel does not harden, applying the bound-
tems, the channel does not harden. Specifically, while in Cell- ing techniques in Section III, while not impossible in
Free Massive MIMO the effective channel is an inner product principle, would yield very pessimistic capacity bounds.
between two M-vectors—hence close to its mean, in the 2
small-cell case the effective channel is a single Rayleigh . .However, since
mkk
gˆ is
mean μmk k , the achievable rate in (42) can be expressed
fading scalar coefficient. Consequently, both the users and
inexponentially distributed
closed form in withexponential integral function
terms of the
the APs must estimate their effective channel gain in order
to demodulate the symbols, which requires both uplink and Ei(·) [40, Eq. (8.211.1)] as:
downlink training. The detailed transmission protocols for the 1/μ¯ m k k
1
sc
Rd,k = −(log2 e)e Ei , (43)
uplink and downlink of small-cell systems are as follows. ! "
μ¯ mk k
where
A. Downlink Transmission ρscαd,kμmkk
In the downlink, the users first estimate their channels μ¯ mk k

d
K .
based ρ αd,k(βmkk − μmkk) +
sc
ρdsc
αd,kr βm r k +
on pilots sent from the APs. The so-obtained channel estimates d
r
k
are used to detect the desired signals. 1 k /
=k
(44)
Let τ sc be the downlink training duration in samples,
√ d
sc τ sc ×1 2 2) Max-Min Power Control: As in the Cell-Free Massive
τd k ∈ , where ∗φk = 1, is the pilot
transmitted from the mkth AP, and is the transmit MIMO systems, we consider max-min power control
per downlink pilot symbol. The MMSE destimate of gmkk can can be formulated as follows:
be expressed as
max min sc
{αd,k } k=1,··· ,K Rd,k
gˆ mk k = gm k k − εm k k , (39)
subject to 0 ≤ αd,k ≤ 1, k = 1, · · · , K. (45)
where ε mkk is the channel estimation
ˆ error, which is inde- Since Rsc is a monotonically increasing function
pendent of the channel estimate gmkk . Furthermore, we have d,k
gˆ mk k ∼ CN 0, μm k k and εm k k ∼ CN 0, βm k k − μm k k , of μ¯ mk k , (45) is equivalent to
where
max min μ¯ mk k
τ scρsc β2 {αd,k k=1,··· ,
} K
μm k d d,p mkk . (40) subject to 0 ≤ α ≤ 1, k = 1, · · · , K. (46)
k K
sc sc 2 d,k
τd r.φ H +
k =1k
√ Problem (46) is a quasi-linear program, which can be
chosensending
the K After thethepilots
APs send for the
data. Let αd k qchannel
k , E |qk|
2
= 1, by using bisection.
be the symbol transmitted from the m ,
kth AP, destined for
the kth user, where αd,k is a power control coefficient, B. Uplink Transmission
0 ≤ αd,k ≤ 1. The kth user receives In the uplink, the APs first estimate the channels based on
√ √ pilots sent from the users. The so-obtained channel estimates
y = ρsc g α rq r + w are used to detect the desired signals. Let ρsc and 0 ≤ αu,k ≤ 1
k d r m d k k u
k =1 √ √ be the normalized SNR and the power control coefficient at the
= √ρ gˆ m k αd k qk + √ρ εm k αd
sc sc
kth user, respectively. Then, following the same methodology
k qk
, ,
√d k d k as
theinfollowing
the derivation of theuplink
achievable downlink transmission,
rate for we obtain
the kth user:
K √ ! "
+ ρ sc gm r k αd,k r qk r + wk , (41) 1
−
sc R=
Ei −(log2 e)e
d 1/ω¯ m k k
k
r u,k ω¯ , (47)
k /=k mk k
where ρsc is the normalized downlink transmit SNR and
d
wk ∼ CN (0, 1) is additive Gaussian noise. where
Remark 8: In small-cell systems, since only one single- antenna AP is involved in transmission to a given user, the
1
concept of “conjugate beamforming” becomes void. Downlink ρscαu,kωm k
ω¯ m k u k ,
k ρuscαu,k(βm kk
− ωm k k ) + ρusc K αu,kr βm kkr + 1
transmission entails only transmitting the symbol destined for kr /
the kth user, appropriately scaled to meet the transmit power =k (48)
1
⎛ . ⎞ ⎫
.α2 gˆ
R d,k m k
s ⎜ ⎟ ⎪
= ρ ,
d
log2 ⎜1
+ K
⎪ ⎨
⎝ sc . . ⎬⎠
⎪⎩ ⎜ρscαd,k(βm k k − μm k k ) + ρsc k αd,k r βm r k +⎟1⎟ ⎪⎭
d d r k
k
and where ωmk k is given by

usc u,p
sc m2 kk
where
ωm k τ ρ β . (49)
L 46.3 + 33.9 ( f ) − 13.82 AP)
2
log10 log10(h
K .ψ H
ψ . +1 — (1.1 log10( f ) − 0.7)hu + (1.56 log10( f ) − 0.8),
τ scu ρscu
β kr m k
k (5
√ In (49), scis× the uplink training duration in
τ scψ k ∈ C
uτ 1
, where ∗ψ k∗2 = 1, is the pilot and where f is the carrier frequency (in MHz), h is the AP
sequence
u
u
A
transmitted from the kth user, and ρsc isu the transmit power antenna height (in m), and hu denotes the user antenna height
per uplink pilot symbol. (in m). The path loss PLmk is a continuous function of dmk.
Similarly to in the downlink, the max-min power control Note that when dmk d1, there is no shadowing.
≤
problem for the uplink can be formulated as a quasi-linear 2) Shadowing Correlation Model: Most previous work
program: assumed that the shadowing coefficients (and therefore zmk)
max are uncorrelated. However, in practice, transmitters/receivers
{αu,k min ω¯ mk
} k=1,··· , that are in close vicinity of each other may be surrounded by
K k
subject to 0 ≤ αu,k ≤ 1, k = 1, · · · , K, (50) common obstacles, and hence, the shadowing coefficients are
correlated. This correlation may significantly affect the system
which can be solved by using bisection. performance.
For the shadow fading coefficients, we will use a model
VI. NUMERICAL RESULTS AND DISCUSSIONS with two components [42]:
√ √
We quantitatively study the performance of Cell-Free Mas- zmk = δam + 1 − δbk, m = 1, . . . , M, K = 1, . . . , K,
sive MIMO, and compare it to that of small-cell systems. (54)
We specifically demonstrate the effects of shadow fading
correlation. The M APs and K users are uniformly distributed where am N (0, 1) and bk N (0, 1) are independent
at random within a square of size D × D km2. ∼ ∼
random variables, and δ , 0 δ 1, is a parameter. The
≤ ≤
variable am models contributions to the shadow fading that
result from obstructing objects in the vicinity of the mth AP,
A. Large-Scale Fading Model
and which affects the channel from that AP to all users in the
We describe the path loss and shadow fading correlation same way. The variable bk models contributions to the shadow
models, which are used in the performance evaluation. The fading that result from objects in the vicinity of the kth user,
large-scale fading coefficient βmk in (1) models the path loss and which affects the channels from that user to all APs in
and shadow fading, according to the same way. When δ 0, the shadow fading from a given
σsh zmk = but different users are affected by
user is the same to all APs,
βmk = PLmk · 10 10 , (51) different shadow fading. Conversely, when δ 1, the shadow
σshzmk fading from a given AP is the same to all users;= however,
where PLmk represents the path loss, and 10 10 repre- different APs are affected by different shadow fading. Varying
sents the shadow fading with the standard deviation σsh, and δ between 0 and 1 trades off between these two extremes.
zmk N (0, 1).
∼ The covariance functions of am and bk are given by:
1) Path Loss Model: We use a three-slope model for the r r
da (m,m ) du (k,k )
path loss [41]: the path loss exponent equals 3.5 if distance
between the mth AP and the kth user (denoted by dmk) is E {amamr } = 2 − ddecorr , E {bkbkr } = − decorr
d , (55)
2
greater than d1, equals 2 if d1 dmk > d0, and equals 0 if r
≥ where da(m, m ) is the geographical distance between the mth
dmk d0 for some d0 and d1. When dmk > d1, we employ the and mrth APs, du(k, kr) is the geographical distance between
≤
Hata-COST231 propagation model. More precisely, the path the kth and krth users, and ddecorr is a decorrelation distance
loss in dB is given by
1
which depends on the environment. Typically, the decorre-
−L − 35 log10(dmk), if dmk > d1 lation distance is on the order of 20Â—200 m. A shorter
decorrelation distance corresponds to an environment with
−L − 15 log10(d1) − 20 log10(dmk),
PLmk = (52) a lower degree of stationarity. This model for correlation
if d0 <dmk ≤ d1
⎨⎪ between different geographical locations has been validated
−L − 15 log10(d1) − 20 log10(d0), if dmk ≤ d0 both in theory and by practical experiments [42], [43].
1
TABLE I
are the same for all users; 3) a cumulative distribution is
SYSTEM PARAMETERS FOR THE SIMULATION
generated over the so-obtained per-user net throughputs.
• For the case without power control: same procedure, but
in 2) no power control is performed. Without power
control, for Cell-Free Massive MIMO, in the downlink
transmission, all APs transmit with full power, and at the
mth AP, the power control coefficients ηmk , k = 1, ... K ,
= ∀
1
are the same,
i.e., ηmk =
# NGO et al.: CELL-FREE MASSIVE MIMO VERSUS SMALL
1
K $−1 r
=1 γmk
r , ∀k = 1, . . . K ,
(this directly comes from (7)), while in the uplink, all
users transmit with full power, i.e., ηk 1, k 1, ... K .
For the small-cell system, in the downlink, all chosen
APs transmit with full power, i.e. αd,k = 1, and in the
B. Parameters and Setup
uplink, all users transmit with full power, i.e. αu,k = 1,
In all examples, we choose the parameters summarized
in Table I. The quantities ρ¯ cf , ρ¯ cf , and ρ¯ cf in this k = 1, . . . K .
table are u d p • For the correlated shadow fading scenario, we
the transmit powers of downlink data, uplink data, and pilot use the shadowing correlation model discussed
symbols,cfrespectively. The corresponding normalized transmit in Section VI-A.2, and we choose ddecorr = 0.1 km and
SNRs ρ , ρcf , and ρcf can be computed by dividing these
d u p δ = 0.5.
powers by the noise power, where the noise power is given by • For the small-cell systems, the greedy pilot assignment
noise power = bandwidth × kB × T0 × noise figure (W), works in the same way as the scheme for Cell-Free
−23
Massive MIMO discussed in Section IV-A, except for
where kB 1.381
= 10× (Joule per Kelvin) is the Boltzmann that in the small-cell systems, since the chosen APs do
constant, and T0 290 (Kelvin) is the noise temperature. not cooperate, the worst user will find a new pilot which
=
To avoid boundary effects, and to imitate a network with an minimizes the pilot contamination corresponding to its
infinite area, the square area is wrapped around at the edges, AP (rather than summed over all APs as in the case of
and hence, the simulation area has eight neighbors. Cell-Free systems).
We consider the per-user net throughputs which take into
account the channel estimation overhead, and are defined as
follows: C. Results and Discussions
cf 1 − τ cf /τc cf We first compare the performance of Cell-Free Massive
SA,k = B RA,k, (56)
MIMO with that of small-cell systems with greedy pilot
2 sc sc assignment and max-min power control. Figure 3 compares
d u
s = B 1 − (τ 2+ τ
sc
R , (57) the cumulative distribution of the per-user downlink net
SA,k A throughput for Cell-Free Massive MIMO and small-cell
systems, with
where A ∈ {d, u} correspond to downlink respectively M = 100, K = 40, and τ cf = τ sc = τ sc = 20, with and
uplink
d u
transmission, B is the spectral bandwidth, and τc is again
without shadow fading correlation.
the coherence interval in samples. The terms τ cf /τc and
Cell-Free Massive MIMO significantly outperforms small-
(τ sc τ sc)/τc in (56) and (57) reflect the fact that, for each
cell in both median and in 95%-likely performance. The
d + u interval of length τc samples, in the Cell-Free
coherence
net throughput of Cell-Free Massive MIMO is much more
Massive MIMO systems, we spend τ cf samples for the uplink
concentrated around its median, compared with the small-cell
training, while in the small-cell systems, we spend τ sc τ sc
systems. Without shadow fading correlation, the 95%-likely
d + u
samples for the uplink and downlink training. In all examples,
net throughput of the Cell-Free downlink is about 14 Mbits/s
we take τc 200 samples, corresponding to a coherence
which is 7 times higher than that of the small-cell downlink
bandwidth of=200 KHz and a coherence time of 1 ms, and
(about 2.1 Mbits/s). In particular, we can see that the small-
choose B 20 MHz.
To ensure = a fair comparison between Cell-Free Massive cell systems are much more affected by shadow fading
MIMO and small-cell systems, we choose ρsc = M ρcf , correlation than Cell-Free Massive MIMO is. This is due
to the fact
that when the shadowing coefficients are highly correlated,
d d
ρsc = ρcf , and ρsc K
u u u = ρscd ρcf
= p , which makes the total the gain from choosing the best APs in a small-cell
ated power equal in all cases. The cumulative distributions of is reduced. With shadowing correlation, the 95%-likely net
the per-user downlink/uplink net throughput in our examples throughput of the Cell-Free downlink is about 10 times higher
are generated as follows: than that of the small-cell system. The same insights can be
• For the case with max-min power control: 1) 200 random obtained for the uplink, see Figure 4. In addition, owing to the
realizations of the AP/user locations and shadow fading fact that the downlink uses more power (since M > K and
profiles are generated; 2) for each realization, the per- ρcf > ρcf ) and has more power control coefficients to choose
d u
user net throughputs of K users are computed by using than the uplink does, the downlink performance is better than
max-min power control as discussed in Section IV-B for the uplink performance.
Cell-Free Massive MIMO and in Section V for small-cell Next we compare Cell-Free Massive MIMO and small-
systems—with max-min power control these throughputs cell systems, assuming that no power control is performed.
1
Fig. 3. Cumulative distribution of the per-user downlink net throughput for

correlated and uncorrelated shadow fading, with the greedy pilot assignment Fig. 5. Cumulative distribution of the per-user downlink net throughput for
correlated and uncorrelated shadow fading, with the greedy pilot assignment
and max-min power control. Here, M = 100, K = 40, and τ cf = τ sc = 20.
and without power control. Here, M = 100, K = 40, and τ cf = τ sc = 20.
d
1
Fig. 6. Same as Figure 5 but for the uplink, and τ cf = τ sc = 20.

Fig. 4. Same as Figure 3 but for the uplink, and τ = τ sc = 20. u
cf u
control. For the small-cell system, power control improves the

Figures 5 and 6 show the cumulative distributions of the per- 95%-likely throughput but not the median throughputd (recall
user net throughput for the downlink and the uplink, that the power control policy explicitly aims at improving the
cf sc
d performance of the worst user).
respectively,
sc with
20, and = 100,
M the
with = 40,
greedyK pilot and τ =
assignment m τhod.=
τu = e In Figures 7 and 8, we consider the same setting as in
In both uncorrelated and correlated shadowing scenarios, Cell- Figures 3 and 4, but here we use the random pilot assign-
Free Massive MIMO outperforms the small-cell approach ment scheme. These figures provide the same insights as
in terms of 95%-likely per-user net throughput. In addition, Figures 3 and 4. Furthermore, by comparing these figures
a comparison of Figure 3 (or 4) and Figure 5 (or 6) shows with Figures 3 and 4, we can see that with greedy pilot
that with power control, the performance of Cell-Free Massive assignment, the 95%-likely net throughputs can be improved
MIMO improves significantly in terms of both median and by about 20% compared with when random pilot assignment
95%-likely throughput. In the uncorrelated shadow fading sce- is used.
nario, the power allocation can improve the 95%-likely Cell- In addition, we study how the M APs assign powers to
Free throughput by a factor of 2.5 for the downlink and a a given user in the downlink of Cell-Free Massive MIMO.
factor of 2.3 for the uplink, compared with the case without From (5), the average transmit power expended by the mth AP
power
1
Fig. 7. Cumulative distribution of the per-user downlink net throughput for Fig. 8. Same as Figure 7 but for the uplink, and τ cf = τ sc = 20.
correlated and uncorrelated shadow fading, with the random pilot assignment
and max-min power control. Here, M = 100, K = 40, and u τ cf = τ sc =
20. d
=
1
Fig. 9. Cumulative distribution of the effective number of APs serving each
user. Here, M = 100, K = 40, and τ cf = 5 and 20.
Fig. 10. Average downlink net throughput versus the number of users for
different τ cf . Here, M = 100.
K 40 users; 2) for each user k in each realization, we found

on the kth user is ρcf ηmkγmk . Then the minimum number of APs, say n, such that the n largest
d
ηmkγmk values of p(m,
{ k) sum up to at least 95% (k is arbitrary
p(m, k) (58) here, since all users have the same statistics); 3) a cumulative
Mmr
r r
distribution was generated over the 200 realizations. We can
is the ratio between the power spent by the mth AP on the see that, on average, only about 10–20 of the 100 APs really
kth user and the total power collectively spent by all APs on participate in serving a given user. The larger τ cf , the less
the kth user. Figure 9 shows the cumulative distribution of the pilot contamination and the more accurate channel estimates
effective number of APs serving each user, for τ cf 5 and 20, — hence, more AP points can usefully serve each user.
=
and uncorrelated shadow fading. The effective number of Finally, we investigate the effect of the number of users K ,
APs serving each user is defined as the minimum number number of APs M, and the training duration τ cf on the per-
of APs that contribute at least 95% of the power allocated formance of Cell-Free Massive MIMO and small-cell systems.
to a given user. This plot was generated as follows: 1) 200 Figure 10 shows the average downlink net throughput versus
random realizations of the AP/user locations and shadow K for different τ cf , at M 100 and uncorrelated shadow fad-
=
fading profiles were generated, each with M = 100 APs and ing. The average is taken over the large-scale fading. We can
1
TABLE II
THE 95%-LIKELY PER-USER NET THROUGHPUT (MBITS/s) OF THE CELL-FREE AND SMALL-CELL DOWNLINK,
cf
FOR M = 100, K = 40, AND τ = τ sc = 20
d
TABLE III
THE 95%-LIKELY PER-USER NET THROUGHPUT (MBITS/s) OF THE CELL-FREE AND SMALL-CELL UPLINK,
cf
FOR M = 100, K = 40, AND τ = τ sc = 20
u
small-cell systems, under uncorrelated and correlated shadow

fading.
VII. CONCLUSION
We analyzed the performance of Cell-Free Massive MIMO,
taking into account the effects of channel estimation, non-
orthogonality of pilot sequences, and power control. A
comparison between Cell-Free Massive MIMO systems and
small-cell systems was also performed, under uncorrelated
and correlated shadow fading.
The results show that Cell-Free Massive MIMO systems
can significantly outperform small-cell systems in terms
of throughput. In particular, Cell-Free systems are much
more robust to shadow fading correlation than small-cell
systems. The 95%-likely per-user throughputs of Cell-Free
Massive MIMO with shadowing correlation are an order
of magnitude higher than those of the small-cell systems.
In terms of implementation complexity, however, small-cell
Fig. 11. Average downlink net throughput versus the number of APs for systems require much less backhaul than Cell-Free Massive
different τ cf . Here, K = 20.
MIMO.
see that when reducing K or τ cf , the effect of pilot con- A PPENDIX

tamination increases, and hence, the performance decreases.
As expected, Cell-Free Massive MIMO systems outperform A. Proof of Theorem 1
small-cell systems. Cell-Free Massive MIMO benefits from To derive the closed-form expression for the achievable
favorable propagation, and therefore, it suffers less from rate given in (23), we need to compute DSk, E |BUk|2 , and
interference than the small-cell system does. As a result, E|UIkkr .
a fixed τ cf , the relative performance gap between Cell-Free 1) Compute DSk: Let εmk gmk g be the channel
Massive MIMO and small-cell systems increases with K . — mk
estimation error. Owing to the properties of MMSE estimation,
Figure 11 shows the average downlink net throughput versus
εmk and gˆ mk are independent. Thus, we have
M for different τ cf , at K = 20. Owing to the array
gain
(for Cell-Free Massive MIMO systems) and diversity gain (for √ M
1/2 ∗
small-cell systems), the system performances of both Cell- DSk = ρcf η (gˆ mk + εmk )gˆ m k
d m
Free Massive MIMO and small-cell systems increase when M m=1
increases. Again, for all M, Cell-Free Massive MIMO systems E
are significantly better than small-cell systems. √ 1/2
Tables II and III summarize the downlink respectively = ρcfd ηmk γmk. (59)
uplink performances of the Cell-Free Massive MIMO and M m=1
1
2) Compute E BUk 2 : Since the variance of a sum of
where
independent RVs is equal to the sum of the variances, ⎧ 2
we have | | ⎨.
T 1/2 r |g |2ϕ H ϕ .. , (64)
M η ⎫
c
. .
⎭
mkr mk mk kr k
E |BUk|2 .⎫
M 1 i m m m
m k
H
= m=1
.
. i ⎞∗ 2
2⎪
⎨. ⎪ M . ⎪⎬
.2 T .
E ⎩ η1/2 c r g ⎝⎛ g ϕ ϕ ⎠ . ⎭ . (65)
ηmk .gmk − gmk . K
d .m=
m ∗
gˆ m k
M
ρ η . m m m m
gˆ . − | E g
∗ 2 ∗
cf gˆ
= # $
2
E .g |
We first compute T1. We have
M M
2
⎧ 1/2
M M 1/2 2 2
2
M
⎩
M
2
d m m m m m
M $
d m — m
m=1 . H .
E | = m and | = βmk −
1 ! + "
gˆ
∗ 2
2 T=1
*E | . —γ2 .ϕ
kϕr k
.ε .E
M
= ⎩
. K
× τ cf p
i
= ρcf η +| gˆ m=1n=1
{
M 2
ηmkr ηnkr cmkr cnkr | gmk | |gnk | 1
d m=1 mk mk mk mk mk
2
1
. H .
1
1 ( a) ∗
ρ cf # gˆ
η= .
E .ε
= 1
|4 $ − γ
2 .ϕ
kr
ϕ
.
k
1 η|gmkr c| mkr
mk
⎫
⎬#
1
2 2 γmk(βmk − γmk) + 2γ 2 2 +
( b) ηmk . .ϕ k r ϕ k ηmkr ηnkr cmkr cnkr |gmk | |gnk |
=
ρ cf ρcf = 2 .ϕ k r ϕk . ηmkMr cmkr βmk
=
+. .ϕH .
2
kr 2
ρ E ηmkr cmkr gmk M βmk
ϕk . 1/2
η r γ rβ r
⎫ kk d kηr mkkr ηnkr nkr βmkβnkmk
cmkr cmk .
∗ 2 mean
where (a) follows that fact that εmk has zero ⎪⎬ and m=1
. M K cf
mk kr p,m T2
2
. kr
ηmkr cmkr βmkβmi ϕ .
m=1
H
gmi ϕ k r ϕ i + ,
. 2
⎪
⎭ M
+ ηmkr γmkr βmk.
m
.
⎭
#
.
K
(6
m ρc
#
p d
γ
K
ςmk
k
kr k k
M
= S : ∗vk∗ ≤
m
1
1
1E |UIkkr |2 Denote by S ςmk, αkr k , ϑm the set of variables, and f (S)
the objective function of (34):
.
1
1/2 ∗
1 m=1
$ VOL. 16, NO. 3, MARCH
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS,
γmkςmk 2
⎩. VERSUS SMALL
NGO et al.: CELL-FREE MASSIVE MIMO
ηmk r cmk r gmk 1
m w˜ mk r
=1
1
k
1
1
=1,··· ,K
1
1
1
1
1 1
ϑ2 +
1
1
1 r
r
kr k kk M
k /=k mk m cf
∗ .2 m=1 d
ρ
1
cf cf cf η1/2 c g ϕHϕ .
+τ ρ ρ
E . r g
1
1
1
1
1
1
1
⎩.m=1 mkr
mk
mk
.
mi kr i
1
⎭
(
62
)
1
U( f, t)
= {S :
f (S) ≥
t}
1
⎧
1
E {X } = =
⎪⎨
0, then E
|X + Y |2 S:
= E |X |
2
+E |
Y |2 , (62)
NGO et al.: CELL-FREE MASSIVE MIMO VERSUS SMALL $2 1
m=1 M
mk
m=1
γmkςmk, ,
1
⎫
⎪⎪ 1
⎬
∀k
1
1
E |UIkkr |2 = ρcf
ηmkr c2
1 |ϕ H ϕ |2α2
⎩⎪
1
NGO et al.: CELL-FREE MASSIVE MIMO VERSUS1 SMALL
ϑ2 + ⎪⎭ 1
1
1
% &T
[20] H. Yin, D. Gesbert, and L. Cottatellucci, “Dealing with interference
1
where vk vkT I−k k √ , and where vk1 in distributed large-scale MIMO systems: A statistical
. .T d Oct.IEEE
2014.
J. Select. Topics Signal Process., vol. 8, no. 5, pp. 942–
ϕ H ϕ k α1k ... ϕ H ϕ K α K k , I−k ρ is a K × (K − 1) matrix
1 K [21] H. Yang and T. L. Marzetta, “Capacity performance of
.√
obtained from the K × .K√identity matrix with the kth column cell large-scale antenna systems,” in Proc. 51st Allerton Conf.
k2 β1kϑ1 ... Tisβ Mk ϑ M Commun., Control, Comput., Monticello, IL, USA, Oct. 2013,
pp. 668–675.
Since the upper-level set U( f, t) can be represented as [22] K. T. Truong and R. W. Heath, Jr., “The viability of distributed
aremoved,
SOC, it and
is av convex. set. Thus, f (S) is quasi-concave. antennas for massive MIMO systems,” in Proc. 47th Asilomar Conf.
Furthermore, the optimization problem (34) is a quasi-concave Signals, Syst. Comput., Monterey, CA, USA, Nov. 2013, pp. 1318–
1323.
optimization problem since the constraint set in (34) is also [23] E. Nayebi, A. Ashikhmin, T. L. Marzetta, and H. Yang, “Cell-
convex. free massive MIMO systems,” in Proc. 49th Asilomar Conf. Sig-
nals, Syst. Comput. (ACSSC), Pacific Grove, CA, USA, Nov. 2015,
REFERENCES pp. 695–699.
[24] H. Huh, A. M. Tulino, and G. Caire, “Network MIMO with linear zero-
[1] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta, forcing beamforming: Large system analysis, impact of channel esti-
“Cell-free massive MIMO: Uniformly great service for everyone,” in mation, and reduced-complexity scheduling,” IEEE Trans. Inf. Theory,
Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun. vol. 58, no. 5, pp. 2911–2934, May 2012.
(SPAWC), Stockholm, Sweden, Jun. 2015, pp. 201–205. [25] W. Liu, S. Han, and C. Yang, “Energy efficiency comparison of
[2] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num- massive MIMO and small cell network,” in Proc. IEEE Global Conf.
bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9, Signal Inf. Process. (GlobalSIP), Dec. 2014, pp. 617–621.
no. 11, pp. 3590–3600, Nov. 2010. [26] E. Bj “ornson, L. Sanguinetti, and M. Kountouris, “Deploying dense
[3] H. Yang and T. L. Marzetta, “A macro cellular wireless network with networks for maximal energy efficiency: Small cells meet massive
uniformly high user throughputs,” in Proc. IEEE Veh. Technol. Conf. MIMO,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 832–847,
(VTC), Sep. 2014, pp. 1–5. Apr. 2016.
[4] S. Zhou, M. Zhao, X. Xu, J. Wang, and Y. Yao, “Distributed wireless [27] H. S. Dhillon, M. Kountouris, and J. G. Andrews, “Downlink
communication system: A new architecture for future public wireless MIMO HetNets: Modeling, ordering results and performance analy-
access,” IEEE Commun. Mag., vol. 41, no. 3, pp. 108–113, Mar. 2003. sis,” IEEE Trans. Wireless Commun., vol. 12, no. 10, pp. 5208–5222,
[5] T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo, Fundamentals Oct. 2013.
Massive MIMO. Cambridge, U.K.: Cambridge Univ. Press, 2016. [28] L. W. A. He, M. Elkashlan, Y. Chen, and K. K. Wong, “Spec-
[6] M. K. Karakayali, G. J. Foschini, and R. A. Valenzuela, “Network trum and energy efficiency in massive mimo enabled HetNets: A
coordination for spectrally efficient communications in cellular stochastic geometry approach,” IEEE Commun. Lett., vol. 19, no. 12,
systems,” IEEE Wireless Commun., vol. 13, no. 4, pp. 56–61, Aug. pp. 2294–2297, Dec. 2015.
2006. [29] T. S. Rappaport, Wireless Communications: Principles and Practice.
[7] W. Choi and J. G. Andrews, “Downlink performance and capacity of Upper Saddle River, NJ, USA: Prentice-Hall, 1956.
distributed antenna systems in a multicell environment,” IEEE Trans. [30] A. Ashikhmin, T. L. Marzetta, and L. Li, “Interference reduction in
Wireless Commun., vol. 6, no. 1, pp. 69–73, Jan. 2007. multi-cell massive MIMO systems I: Large-scale fading precoding and
[8] R. Irmer et al., “Coordinated multipoint: Concepts, performance, and decoding,” IEEE Trans. Inf. Theory, to be published.
field trial results,” IEEE Commun. Mag., vol. 49, no. 2, pp. 102–111, [31] F. Kaltenberger, H. Jiang, M. Guillaud, and R. Knopp, “Relative
Feb. 2011. channel reciprocity calibration in MIMO/TDD systems,” in Proc.
[9] X. Hong, Y. Jie, C.-X. Wang, J. Shi, and X. Ge, “Energy-spectral Future Netw. Mobile Summit, Florence, Italy, Jun. 2010, pp. 1–10.
efficiency trade-off in virtual MIMO cellular systems,” IEEE J. Sel. [32] P. Marsch and G. Fettweis, “Uplink CoMP under a constrained
Areas Commun., vol. 31, no. 10, pp. 2128–2140, Oct. 2013. backhaul and imperfect channel knowledge,” IEEE Trans. Wireless
[10] D. Castanheira and A. Gameiro, “Distributed antenna system capacity Commun., vol. 10, no. 6, pp. 1730–1742, Jun. 2011.
scaling,” IEEE Wireless Commun. Mag., vol. 17, no. 3, pp. 68–75, [33] O. Simeone, O. Somekh, E. Erkip, H. V. Poor, and S. Shamai,
Jun. 2010. “A broadcast approach to robust communications over unreliable multi-
[11] R. Heath, S. Peters, Y. Wang, and J. Zhang, “A current per spective on relay networks,” in Proc. Inf. Theory Appl. Workshop (ITW), Taormina,
distributed antenna systems for the downlink of cellular systems,” IEEE Italy, Feb. 2009, pp. 334–340.
Commun. Mag., vol. 51, no. 4, pp. 161–167, Apr. 2013. [34] H. Cramer, Random Variables and Probability Distributions.
[12] Z. Liu and L. Dai, “A comparative study of downlink MIMO cellular Cambridge, U.K.: Cambridge Univ. Press, 1970.
networks with co-located and distributed base-station antennas,” IEEE [35] B. Hassibi and B. M. Hochwald, “How much training is needed in
Trans. Wireless Commun., vol. 13, no. 11, pp. 6259–6274, Nov. 2014. multiple-antenna wireless links?” IEEE Trans. Inf. Theory, vol. 49, no.
[13] Q. Sun, S. Jin, J. Wang, Y. Zhang, X. Gao, and K.-K. Wong, “On 4, pp. 951–963, Apr. 2003.
scheduling for massive distributed MIMO downlink,” in Proc. IEEE [36] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral effi-
Global Commun. Conf. (GLOBECOM), Atlanta, GA, USA, Dec. 2013, ciency of very large multiuser MIMO systems,” IEEE Trans. Commun.,
pp. 4151–4156. vol. 61, no. 4, pp. 1436–1449, Apr. 2013.
[14] M. Matthaiou, C. Zhong, M. R. McKay, and T. Ratnarajah, “Sum rate [37] H. Ahmadi, A. Farhang, N. Marchetti, and A. MacKenzie, “A game
analysis of ZF receivers in distributed MIMO systems,” IEEE J. Sel. theoretic approach for pilot contamination avoidance in massive
Areas Commun., vol. 31, no. 2, pp. 180–191, Feb. 2013. MIMO,” IEEE Wireless Commun. Lett., vol. 5, no. 1, pp. 12–15, Feb.
[15] K. Hosseini, W. Yu, and R. S. Adve, “Large-scale MIMO versus network 2016.
MIMO for multicell interference mitigation,” IEEE J. Select. Topics [38] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
Signal Process., vol. 8, no. 5, pp. 930–941, Oct. 2014. Cambridge Univ. Press, 2004.
[16] J. Joung, Y. K. Chia, and S. Sun, “Energy-efficient, large-scale [39] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to
distributed-antenna system (L-DAS) for multiple users,” IEEE J. Select. coverage and rate in cellular networks,” IEEE Trans. Commun., vol. 59,
Topics Signal Process., vol. 8, no. 5, pp. 930–941, Oct. 2014. no. 11, pp. 3122–3134, Nov. 2011.
[17] Q. Ye, O. Y. Bursalioglu, and H. C. Papadopoulos, “Harmonized cellular [40] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
and distributed massive MIMO: Load balancing and scheduling,” in Products, 7th ed. San Diego, CA, USA: Academic, 2007.
Proc. IEEE Global Commun. Conf. (GLOBECOM), San Diego, CA, [41] A. Tang, J. Sun, and K. Gong, “Mobile propagation loss with a low
USA, Dec. 2015, pp. 1–6. base station antenna for NLOS street microcells in urban area,” in Proc.
[18] J. Wang and L. Dai, “Asymptotic rate analysis of downlink multi-user IEEE Veh. Technol. Conf. (VTC), May 2001, pp. 333–336.
systems with co-located and distributed antennas,” IEEE Trans. [42] Z. Wang, E. K. Tameh, and A. R. Nix, “Joint shadowing process in
Wireless Commun., vol. 14, no. 6, pp. 3046–3058, Jun. 2015. urban peer-to-peer radio channels,” IEEE Trans. Veh. Technol., vol. 57,
[19] A. Lozano, R. W. Heath, Jr., and J. G. Andrews “Fundamental limits of no. 1, pp. 52–64, Jan. 2008.
cooperation,” IEEE Trans. Inf. Theory, vol. 59, no. 9, pp. 5213–5226, [43] M. Gudmundson, “Correlation model for shadow fading in mobile
Sep. 2013. radio systems,” Electron. Lett., vol. 27, no. 23, pp. 2145–2146,
Nov. 1991.
1
Hien Quoc Ngo received the B.S. degree in elec-
Erik G. Larsson (S’99–M’03–SM’10–F’16)
trical engineering from the Ho Chi Minh City
received the Ph.D. degree from Uppsala University,
University of Technology, Vietnam, in 2007, the
M.S. degree in electronics and radio engineering Sweden, in 2002. He was with the Royal Institute
from Kyung Hee University, South Korea, in 2010, of Technology (KTH), Stockholm, Sweden, the
and the Ph.D. degree in communication systems University of Florida, USA, The George
from Linköping University (LiU), Sweden, in 2015. Washington University, USA, and Ericsson
In 2014, he visited the Nokia Bell Labs, Murray Research, Sweden. In 2015, he was a Visiting
Hill, NJ, USA. Fellow with Princeton University, Princeton, NJ,
He is currently a Post-Doctoral Researcher with USA, for four months. He is currently a Professor of
the Division for Communication Systems, Depart- communication systems with Linköping University,
ment of Electrical Engineering, LiU. He is also a Visiting Research Fel- Linköping, Sweden.
His main professional interests are within the
low with the School of Electronics, Electrical Engineering and Com-
areas of wireless communications and signal processing. He has co-authored
puter Science, Queen’s University Belfast, U.K. His current research
some 130 journal papers on these topics and is a co-author of the two
interests include massive (large-scale) MIMO systems and cooperative
Cambridge University Press textbooks Space-Time Block Coding for Wireless
communications.
Dr. Ngo has been a member of the technical program committees for several Communications (2003) and Fundamentals of Massive MIMO (2016). He is a
IEEE conferences such as the ICC, the Globecom, the WCNC, the VTC, co-inventor on 16 issued and many pending patents on wireless technology.
the WCSP, the ISWCS, the ATC, and the ComManTel. He was a recipient Dr. Larsson is a member of the IEEE Signal Processing Society Awards
of the IEEE ComSoc Stephen O. Rice Prize in Communications Theory in Board during 2017–2019. He served as an Associate Editor of the IEEE
2015. He also received the IEEE Sweden VT-COM-IT Joint Chapter Best TRANSACTIONS ON COMMUNICATIONS from 2010 to 2014 and the IEEE
Student Journal Paper Award in 2015. He was an IEEE C OMMUNICATIONS TRANSACTIONS ON SIGNAL PROCESSING from 2006 to 2010. From 2015 to
LETTERS Exemplary Reviewer for 2014, an IEEE TRANSACTIONS ON 2016, he served as the Chair of the IEEE Signal Processing Society SPCOM
COMMUNICATIONS Exemplary Reviewer for 2015, and an IEEE WIRELESS Technical Committee, and in 2017 he is the past chair of this committee.
COMMUNICATIONS LETTERS Exemplary Reviewer for 2016. He served as the Chair of the Steering Committee for the IEEE W IRELESS
COMMUNICATIONS LETTERS from 2014 to 2015. He was the General Chair
of the Asilomar Conference on Signals, Systems and Computers in 2015,
and the Technical Chair in 2012.
He received the IEEE Signal Processing Magazine Best Column Award,
Alexei Ashikhmin is currently a Distinguished in 2012 and 2014, and the IEEE ComSoc Stephen O. Rice Prize in
Member of the Technical Staff of the Communi- Communications Theory in 2015.
cations and Statistical Sciences Research Depart-
ment, Nokia Bell Labs, Murray Hill, NJ, USA. His
research interests include communications theory,
massive MIMO, classical and quantum information
theory, and error correcting codes.
In 2014, he received the Thomas Edison Patent
Award for Patent on Massive MIMO System with
Decentralized Antennas. In 2004, he received the
Stephen O. Rice Prize for the best paper of the
IEEE TRANSACTIONS ON COMMUNICATIONS. In 2002, 2010, and 2011 he
received the Bell Laboratories President Awards for breakthrough research in
communication projects.
Dr. Ashikhmin served as an Associate Editor of the IEEE TRANSACTIONS
ON INFORMATION THEORY from 2003 to 2006 and from 2011 to 2014.
Thomas L. Marzetta (F’13) was born in

Washington, D.C. He received the PhD and SB
in Electrical Engineering from Massachusetts
Hong Yang received the Ph.D. degree in applied Institute of Technology in 1978 and 1972, and the
MS in Systems Engineering from University of
mathematics from Princeton University, Princeton,
Pennsylvania in 1973. After careers in petroleum
NJ, USA. He was with the Systems Engineering
exploration at Schlumberger-Doll Research and
Department and the Wireless Design Center, Lucent
defense research at Nichols Research Corporation,
Technologies and Alcatel-Lucent, and has worked
he joined Bell Labs in 1995 where he is currently
for a start-up network technology company. He is
a Bell Labs Fellow. Previously he directed
currently a Member of the Technical Staff with the Communications and Statistical Sciences
the Mathematics of Networks and Communications Department within the former Mathematical Sciences Research Center.
Research Department, Nokia Bell Labs, Murray Dr. Marzetta is on the Advisory Board of MAMMOET (Massive MIMO
Hill, NJ, where he is involved in research in for Efficient Transmission), an EU-sponsored FP7 project, and he was
communications networks. He has co-authored Coordinator of the GreenTouch Consortium’s Large Scale Antenna Systems
many research Project. He has received awards including the 2015 IEEE Stephen O. Rice
papers in wireless communications, applied mathematics, and financial eco- Prize, the 2015 IEEE W. R. G. Baker Award, and the 2013 IEEE Guglielmo
nomics, co-invented many U.S. and international patents, and co-authored the Marconi Prize Paper Award. He was elected a Fellow of the IEEE in 2003,
book Fundamentals of Massive MIMO (Cambridge Univ. Press, 2016). and he received an Honorary Doctorate from Linköping University in 2015.

Cell-Free Massive MIMO Versus Small Cells

Uploaded by

Copyright:

Available Formats

Cell-Free Massive MIMO Versus Small Cells

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Cell-Free Massive MIMO Versus Small Cells

Uploaded by

Copyright:

Available Formats

NGO et al.

: CELL-FREE MASSIVE MIMO VERSUS SMALL

Cell-Free Massive MIMO Versus Small Cells

Abstract— A Cell-Free Massive MIMO (multiple-input

Massive MIMO systems perform better than small-cell sys-

perfect calibration of the hardware chains. The feasibility

The APs treat the channel estimates as the true channels,

d, resources must be efficiently managed. In this section, we

1 . ∗ IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 3, MARCH

will present methods for pilot sequence assignment and power

be performed independently, because the pilots are not power

that knows the instantaneous channel gain. Figure 2 shows

figure, the gap is small, which means that downlink training

2) Achievable Uplink Rate: The central processing unit

that the central processing unit uses only statistical knowl-

a similar methodology as in Section III-B.1, we obtain a

the Cell-Free Massive MIMO system with matched filtering

top of the next page.

A. Greedy Pilot Assignment

sequences, due to the limited length of the coherence interval.

orthogonal pilot sequences. Here we focus on the case that

sequences to the K users.

length τ cf samples to the K users is random pilot assign-

randomly assigned one pilot sequence from a predetermined

ment could alternatively be done by letting each user choose

of pilots). However, it appears from simulations that the latter

ηkr ηkr γmkβmkr +

is a useful baseline, occasionally two users in close vicinity

which iteratively refines the pilot assignment. The K users

are first randomly assigned K pilot sequences. Then the

in strong pilot contamination.

its pilot sequence so that its pilot contamination effect

Optimal pilot assignment is a difficult combinatorial

is minimized.3 The pilot contamination effect at the k∗th

user is quantified by the second term in (3) which has

unit significantly. Furthermore, since ϕk∗ is chosen from Sϕ,

We next show that Cell-Free Massive MIMO can provide

uniformly good service to all users, regardless of their geo-

graphical location, by using max-min power control. While

needs to send an index to each user.

power control in general is a well studied topic, the max-

Free Massive MIMO are entirely new. The power control is

performed at the CPU, and importantly, is done on the large-

By introducing slack variables αkr k and ϑm , we reformu- . .T

In the downlink, the users first estimate their channels μ¯ mk k

and where ωmk k is given by

Fig. 3. Cumulative distribution of the per-user downlink net throughput for

Fig. 6. Same as Figure 5 but for the uplink, and τ cf = τ sc = 20.

control. For the small-cell system, power control improves the

K 40 users; 2) for each user k in each realization, we found

small-cell systems, under uncorrelated and correlated shadow

see that when reducing K or τ cf , the effect of pilot con- A PPENDIX

We first compute T1. We have

Thomas L. Marzetta (F’13) was born in

You might also like