3rd Paper
3rd Paper
3rd Paper
Abstract—This paper studies a cell-free massive multiple-input reap all benefits of massive MIMO (favorable propagation, and
arXiv:2004.11833v1 [cs.IT] 24 Apr 2020
multiple-output (MIMO) system where its access points (APs) channel hardening when using multiple antennas at APs [12])
and users are equipped with multiple antennas. Two transmission and network MIMO (increased macro-diversity gain), and
protocols are considered. In the first transmission protocol, there
are no downlink pilots, while in the second transmission protocol, hence, it can offer very high SE, EE, and coverage probability.
downlink pilots are proposed in order to improve the system These benefits can be achieved with simple signal processing
performance. In both transmission protocols, the users use the and local channel acquisition at each AP. With the cell-free
minimum mean-squared error-based successive interference can- topology, the excessive handover issue in small-cell systems
cellation (MMSE-SIC) scheme to detect the desired signals. For can be resolved. Moreover, poor cell-edge performance, which
the analysis, we first derive a general spectral efficiency formula
with arbitrary side information at the users. Then analytical is typical in small-cell networks, can be resolved in cell-free
expressions for the spectral efficiency of different transmission massive MIMO network by geographically distributing the
protocols are derived. To improve the spectral efficiency (SE) of APs [11]. Compared with small cells, cell-free massive MIMO
the system, max-min fairness power control (PC) is applied for can provide up to a ten-fold improvement in 95%-likely SE
the first protocol by using the closed-form expression of its SE. [8], [13]. Thus, cell-free massive MIMO has attracted a lot of
Due to the computation complexity of deriving the closed-form
performance expression of SE for the second protocol, we apply research interest recently [14]–[18].
the optimal power coefficients of the first protocol to the second Most of previous works exploit the performance of cell-
protocol. Numerical results show that two protocols combining free massive MIMO with single-antenna users. However, in
with multi-antenna users are prerequisites to achieve the sub- practice, many user’s devices of moderate physical size (e.g.
optimal SE regardless of the number of user in the system. laptops, tablets, and smart vehicles) can be equipped with
Index terms— Cell-free massive MIMO, massive MIMO, several antennas to increase the multiplexing gain, and to
spectral efficiency, MMSE-SIC, power control. improve system reliability due to the diversity gain. Thus, it
is important to evaluate the performance of cell-free massive
I. I NTRODUCTION MIMO with multiple antennas at the users. Moreover, the
effect of equipping multiple antennas at the users needs to
Cellular massive multiple-input multiple-output (MIMO) is
be well understood to design the systems. Downlink channel
currently considered as a key wireless access technology for
estimation has already investigated in [19]. But [19] considered
5G because it can provide high spectral efficiency (SE) and
collocated massive MIMO systems with single-antenna users,
high energy efficiency (EE) with simple signal processing [2],
orthogonal pilot sequences, and no power control. In [20],
[3]. In cellular massive MIMO, the BS with massive antenna
[21], the authors studied downlink channel estimation of cell-
arrays simultaneously serves all users in its cell on the same
free massive MIMO. But in [20], [21], each AP has only one
time-frequency resource [4]–[7].
antenna, and matched filtering detection is used.
Since cellular massive MIMO is based on cellular topology,
Inspired by the above discussion, in this paper, we analyze
its inherent limitation is inter-cell interference. To overcome
the performance of cell-free massive MIMO systems with
this limitation, cell-free massive MIMO is introduced [8].
multiple antennas at both APs and users. In cell-free massive
Cell-free massive MIMO can be considered as a useful and
MIMO, each users can be close to several APs, and thus Rician
scalable version of network MIMO [9], [10] (much in the same
channel model is more reasonable in many scenarios [22].
way as cellular Massive MIMO is scalable version of multi-
However, in rich scattering environments, the Rayleigh fading
user MIMO). In cell-free massive MIMO, a large number
model is still reasonable [8], [12]–[18]. In addition, Rayleigh
of access points (APs), which are geographically distributed
fading model is analytically tractable which helps us to obtain
over a large area, coherently serve all users on same time-
initial and important insights. Therefore, in this paper, we
frequency resource [8], [11]. Cell-free massive MIMO can
consider cell-free massive MIMO systems using independent,
T. C. Mai, H. Q. Ngo, and T. Q. Duong are with the Institute of Electronics, identically distributed (i.i.d.) Rayleigh fading channels. The
Communications and Information Technology in Queen’s University Belfast, closed-form expression of downlink SE is derived with taking
Belfast, U.K. (email: {trang.mai, hien.ngo, trung.q.duong}@qub.ac.uk)
The work of Trang C. Mai and Hien Quoc Ngo was supported by account of non-orthogonal pilot sequences. As the space
the UK Research and Innovation Future Leaders Fellowships under Grant between adjacent antennas at the same user is very small, it
MR/S017666/1. The work of T. Q. Duong was supported in part by the may causes huge interference to each others. Therefore, in this
U.K. Royal Academy of Engineering Research Fellowship under Grant
RF1415\14\22. paper, orthogonal pilot sequences are assigned for antennas at
Parts of this work were presented at the 2018 IEEE GlobalSIP Conf. [1] the same user, and those pilot sequences can be reused at
2
antennas of other users. Moreover, the effects of the number the power control coefficients of the first protocol to the
of antennas at APs and users on the SE are analyzed through second protocol. Numerical results show that, with those
the use of max-min fairness power control. We evaluate power control coefficients, the SE improves significantly.
the system performance of two protocols for the downlink • We investigate the effects of number antennas at both the
data transmission of cell-free massive MIMO. In the first APs and users.
protocol, the system operates with two phases: uplink channel • We propose the framework for achieving the sub-optimal
estimation and downlink data transmission. As a result, only system performance regardless of the number of users.
statistical channel state information (CSI) is available at the The rest of this paper is organized as follows. Section II
users. Note that the statistical CSI depends on large-scale defines the system model for the downlink cell-free massive
fading which changes very slowly, and it may stay constant for MIMO for both data transmission and channel estimation.
a duration of some 40 small-scale fading coherence intervals Next, Section III derives the achievable downlink SE of
[8]. Whereas, in the second protocol, as the level of channel cell-free massive MIMO. Then, Section IV derives max-
hardening in cell-free massive MIMO is lesser than the one in min fairness power control for SE. Section V evaluates the
collocated massive MIMO [12], we use the downlink channel system performance by using numerical results. Finally, the
estimation to improve the system performance. Therefore, in conclusion is drawn in Section VI.
the second protocol, the system operates with three phases: Notation: The superscripts ()∗ , ()T , and ()H stand for
uplink channel estimation, downlink channel estimation and the conjugate, transpose, and conjugate-transpose, respectively.
downlink data transmission. As a result, estimated CSI is avail- The Euclidean norm, the expectation operators, and the deter-
able at the users. To improve the system performance, both minant of matrix are denoted by k · k, E {·}, and |.|, respec-
protocols use minimum mean-squared error-based successive tively. var(.) denotes variance. In addition, z ∼ CN 0, σ 2
interference cancellation (MMSE-SIC) detectors at the users. denotes a circularly symmetric complex Gaussian random
The computational complexity of MMSE-SIC detectors relates variable (RV) z with zero mean and variance σ 2 , and CL×N
to the inverse operations of the N × N effective channel gain denotes the L × N matrix. Finally, z ∼ N (0, σ 2 ) denotes a
matrix. Since N is small (several antennas per user), this real-valued Gaussian RV.
complexity is low. The second protocol is a generalization
of that in previous work on cell-free massive MIMO [21], II. S YSTEM M ODEL
where we consider multi-antennas at users. In both protocols,
we also compare system performances using MMSE detectors We consider a cell-free massive MIMO system operating in
with the ones using MMSE-SIC detectors. In this paper, to time division duplex (TDD) mode with M APs and K users
reduce the fronthaul and backhaul requirements, the conjugate randomly located within a large area. Each AP has L antennas,
beamforming technique is used for both the protocols since it whereas each user has N antennas. Let Gmk ∈ CL×N be the
can be implemented in a distributed manner [8]. Other linear channel response matrix between the k-th user and the m-th
processing techniques such as MMSE and zero-forcing (ZF) AP. Then,
1/2
are better than the conjugate beamforming technique in terms Gmk = βmk Hmk , (1)
of the system performance [23]. However, MMSE and ZF where βmk is large-scale fading between the k-th user and
need huge fronthaul and backhaul requirements, as we need the m-th AP, and Hmk is the L × N small-scale fading
to send the channel state information to the CPUs, and signal matrix whose elements are assumed to be independent and
processing is mainly done at the CPUs [23]–[25]. So it is identically distributed (i.i.d.) CN (0, 1) RVs. In this work, we
hard to implement MMSE or ZF in large cell-free massive focus on the downlink transmission, and hence, the uplink
MIMO networks. Recently, [26] proposed a method which data transmission is neglected. Specifically, we consider two
can implement ZF in a distributed manner. But this scheme transmission protocols. The first protocol has two phases: the
requires a very large number of antennas at the access points. uplink channel estimation and the downlink data transmission.
Thus, cell-free massive MIMO with conjugate beamforming The second protocol has three phases: the uplink channel
techniques has still received a lot of research attention recently estimation, the downlink channel estimation and the downlink
[27]–[29]. The main contributions of this paper are as follows. data transmission. System model of the two protocols will be
presented in detail in the rest of this section.
• The details of two transmission protocols with and with-
out downlink pilots are presented and analyzed. The
channel estimation with non-orthogonal pilot sequences A. Transmission Protocol 1 - No Downlink Pilots
and MMSE-SIC detectors are taken into account. This transmission protocol is commonly used in previous
• We derive a general formula for the SE with arbitrary side studies of cell-free massive MIMO systems. Each user relies
information at the users. Based on this result, analytical on the channel hardening property of massive MIMO tech-
expressions for the SE of different transmission protocol nology to detect the desired signals. So there is no downlink
are derived. channel estimation phase [8].
• Max-min fairness power control (PC) is applied for the 1) Uplink Channel Estimation: In this phase, all users will
first protocol to improve the SE of the system. For the send pilot signals to the APs. Then each AP will estimate its
second protocol, due to the high complexity associated channels to all users using the received pilot signals. Let τu
with computation for the closed-form of SE, we apply be the length of the uplink training duration per coherence
3
interval, and Φ u,k ∈ Cτu ×N , where its n-th column satisfies constraint at the m-th AP, and ηmk is power control coefficient
kφ
φu,k,n k = 1, ∀n ∈ N , be a pilot matrix of the k-th user. corresponding to the k-th user. The power control coefficients
Then, the received signal at the m-th AP is ηmk are chosen to satisfy the power constraint at each AP:
K E{kxm k2 } ≤ ρ which is equivalent to
X √
Yu,m = τu ρu GmkΦ H
u,k + Wu,m , (2) K
k=1 √ X 1
τu ρu ηmk βmk tr (Amk ) ≤ . (9)
where ρu is the normalized signal-to-noise ratio (SNR) of each L
k=1
uplink pilot symbol and Wu,m is the L × τu matrix of additive
noise at the m-th AP. We assume that the elements of Wu,m The received signal at the k-th user is
are i.i.d. CN (0, 1) RVs. A projection of Yu,m onto Φ u,k is M
X
Yu,mk = Yu,mΦ u,k rk = GH
mk xm + nk
m=1
K
X √ K
= τu ρu GmiΦ u,ik + Wu,mk , (3) √ X
= ρ Dkk′ qk′ + nk , (10)
i=1
k′ =1
where Φ u,ik , Φ H
u,iΦ u,k , and Wu,mk , Wu,mΦ u,k . By stacking PM 1/2 H
all columns of Yu,mk on top of each other, we have where Dkk′ , m=1 ηmk′ Gmk Ĝmk′ denotes the effective
K
downlink channel for the k ′ -th user and nk is the noise vector.
√ X The elements of nk are assumed to be i.i.d. CN (0, 1).
vec(Yu,mk ) = τu ρu Φ Tu,ik ⊗ IL )vec(Gmi ) + vec(Wu,mk )
(Φ
i=1
K
√ X B. Transmission Protocol 2 - With Downlink Pilots
= τu ρu Φik vec(Gmi ) + vec(Wu,mk ),
Φ̃ (4)
i=1 Different from cellular massive MIMO, cell-free massive
where vec(.) is the vectorization operation, and Φ̃ Φu,ik , MIMO offers lesser channel hardening. Therefore it is good
Φ Tu,ik ⊗IL . Since the distance between adjacent antennas at the to estimate channel at the users via the downlink pilots. Here
same user is very small, non-orthogonal pilots may cause huge for the analysis simplicity, we assume that orthogonal pilot
interference to each others. To mitigate interference between sequences are used for the downlink channel estimation phase.
antennas of the same user, we assume that orthogonal pilot 1) Uplink Channel Estimation: This phase is the same as
sequences are assigned for antennas of each user, but these that of Transmission Protocol 1. See Section II-A1.
pilot sequences can be reused in other users. Then, MMSE 2) Downlink Channel Estimation: From the received signal
estimation of vec(Gmk ) given vec(Yu,mk ) is expressed by [30] (10), to detect the desired signal, each user does not need
!−1 to estimate all channel matrices Gmk . Instead, it needs to
K
√ X H estimate only the effective channel gain matrices Dkk′ which
vec(Ĝmk ) = τu ρu βmk τu ρu Φu,ik βmiΦ̃
Φ̃ Φu,ik + ILN
have much lower dimension. To do this, the pilot sequences
i=1
will be precoded before being sent to all users [21]. Let τd
× vec(Yu,mk ). (5) be the length of the downlink training duration per coherence
interval, and Φ d,k ∈ Cτd ×N be a pilot matrix for the k-th user
Lemma 1: The estimate of the channel matrix Gmk is that satisfies
(
Ĝmk = Yu,mk Amk , (6) H IN if k = k ′ ,
Φ d,kΦ d,k′ =
where 0 if k 6= k ′ .
K
!−1
√ X Then, precoded pilot matrix which is transmitted from the m-
Amk , τu ρu βmk τu ρu βmiΦ H
u,ikΦ u,ik + IN . (7) th AP is
i=1
K
Proof: See Appendix A. √ X 1/2
Xd,m = τd ρd ηmk′ Ĝmk′ Φ H
d,k′ , (11)
2) Downlink data transmission: In this phase, each AP uses
k′ =1
the channel estimates in the uplink channel estimation phase
together with the conventional conjugate beamforming tech- where ρd is the normalized signal-to-noise ratio (SNR) of the
nique to precode the desired symbols [8]. Then the precoded m-th AP. Then, the received pilot signal at the k-th user is
signal will be sent to all users. The L × 1 transmitted signal
M X
K
from the m-th AP is √ X 1/2
Yd,k = τd ρd ηmk′ GH H
mk Ĝmk′ Φ d,k′ + Wd,k ,
K
√ X 1/2 m=1 k′ =1
xm = ρ ηmk Ĝmk qk , (8) K
k=1 √ X
= τd ρd Dkk′ Φ H
d,k′ + Wd,k . (12)
where qk , with E{qk qH
k } = IN , is the vector of symbols k′ =1
intended for the k-th user, ρ is the normalized transmit SNR
4
d d ′
kk ,j d,k,ij
otherwise, By deriving all expectations, we obtain the closed-form expres-
τd ρd ξkk′ ,j +1
sion for the achievable SE (16) as in the following theorem.
(13)
Theorem 2: Given statistic CSI, says Θk = D̄kk , and using
where yd,k,ij is ij-element of the matrix Ydp,k , MMSE-SIC detectors, the achievable downlink SE for the k-th
PM user can be represented in closed-form as
ξkk,i = L m=1 ηmk βmk γmk,i , γ =
H
−1 mk,i −1
K H
2
RkSt-SIC = (1 − τu /τc ) log2 IN + ρD̄kk Ψbkk
P
τu ρu βmk Φu,ik βmiΦ̃
τu ρu i=1 Φ̃ Φu,ik + ILN , D̄kk , (16)
(i−1)L+l
PM 1/2 where
κkk,i = L m=1 ηmk γmk,i , and ξkk′ ,i =
PM
L m=1 ηmk′ βmk γmk′ ,i . M
√ X 1/2
Proof: See Appendix B. D̄kk = L τu ρu ηmk βmk Amk , (17)
Note that Dkk′ is an N × N matrix. Since N is small, the m=1
Dkk′ are not Gaussian distributed, the elements of their MMSE Fig. 1 and Fig. 2 verify the Gaussian approximation in
estimate D̂kk′ and corresponding elements of estimation error Lemma 3. Fig. 1 shows that the probability density functions
D̃kk′ are uncorrelated, but not independent. This makes (19) (pdfs) of the empirical and the Gaussian distribution of dkk,ij
hard to be computed in closed form. However, elements of (∀i, j = 1, 2, . . . , N ) are very close. Moreover, with the high
Dkk′ are very close to Gaussian, especially when M is large. probability, the imaginary part of dkk,ii is much smaller than
This is shown in the Lemma 3 as follows. the real part, so it can be neglected. In Fig. 2, it is clear that
Lemma 3: The elements of downlink effective channels the pdfs of both the real part and the imaginary part of dkk′ ,ij
Dkk′ , (∀k, k ′ = 1, 2, . . . , K) converge in distribution to Gaus- (k ′ 6= k, ∀i, j = 1, 2, . . . , N ) are very close to the pdfs of
sian distribution as follows: their Gaussian approximations. In both Fig. 1 and Fig. 2, apart
from the imaginary part of dkk,ii which can be neglected, the
d
dkk,ij → CN (0, ξkk,j ), as M → ∞, and ∀i 6= j, probability of mismatch between approximated and enprirical
M M
! Gaussian pdfs is very small.
d
X √ X
2
dkk,ii → N L ηmk γmk,i , L ηmk γmk,i , By using Gaussian approximation of Dkk , we obtain the
m=1 m=1 following approximating closed-form expression of the SE.
as M → ∞, and Theorem 3: Given estimated CSI, says Θk =
d {D̂k1 , D̂k2 , . . . , D̂kK }, and using MMSE-SIC detectors,
dkk′ ,ij → CN (0, ξkk′ ,i ), as M → ∞, ∀k ′ 6= k, and ∀i, j ∈ N,
the achievable downlink SE of the k-th user in (14) can be
d rewritten as
where → denotes convergence in distribution.
τu +τd H
′ −1
Proof: See Appendix F. RkEt-SIC = 1 − E log2 IN +ρD̂kk Ψckk D̂kk ,
τc
(20)
×10 5
10 Empirical (Real part)
Empirical (Imagine part)
where
Gausian K K
8
′ X H X var
Ψckk = ρ D̂kk′ D̂kk′ + ρ D̃kk′ + IN , (21)
6
k′ 6=k k′ =1
pdf
4 with
N
2
var(d˜kk′ ,1j )
P
0 ... 0
j=1
0 var PN ˜
= 0 j=1 var(dkk′ ,2j ) ... 0 ,
10 -7 10 -6 d 10 -5 10 -4 D̃kk′
.. ..
kk,ii
.
.
Gaussian PN ˜
Empirical
Gaussian
0 0 ... j=1 var(dkk′ ,N j )
10 4 10 4 Empirical
(22)
10 3 10 3
where
pdf
10 2 10 2 ξkk,i +κ2kk,i
τd ρd,p (ξkk,i +κ2kk,i )+1
if k = k ′ , and i = j
var(d˜kk′ ,ij ) = ξkk′ ,j
10 1 10 1
τd ρd,p ξkk′ ,j +1 otherwise .
(23)
10 0 10 0
-1 -0.5 0
Re(dkk,ij)
0.5 1 -1 -0.5 0
Im(dkk,ij )
0.5 1 Proof: See appendix G.
×10 -4 ×10 -4
Fig. 1. True (Empirical) v.s. approximated (Gaussian) pdf of dkk,ij ), ∀i, j = C. Achievable Downlink SE with Linear MMSE Detectors
1, 2, . . . , N .
In this section, we derive the downlink SE using linear
MMSE detectors instead of MMSE-SIC detectors to evaluate
the difference of the system performance between them. Note
Gaussian Gaussian
10 4 Empirical 10 4 Empirical that compare with the MMSE-SIC detection, the linear MMSE
detection is simpler for the implementation.
10 3 10 3
1) Protocol 1 with Linear MMSE Detectors: At each user,
linear MMSE detectors detect N data streams, which are in-
pdf
10 2 10 2
tended for N antennas, independently. Then, the linear MMSE
10 1 10 1
detector for the n-th data stream of the k-th user is given by
[2], [31]
10 0 10 0 ′
-1 -0.5 0
Re(dkk',ij )
0.5
×10 -4
1 -1 -0.5 0
Im(dkk',ij )
0.5 1
×10 -4
fk,n = Ψbkk d̄kk,n , (24)
′ H
Fig. 2. True (Empirical) vs approximated (Gaussian) pdf of dkk′ ,ij ), ∀k ′ 6= where Ψbkk = (Ψbkk )−1 + D̄kk D̄kk and d̄kk,n denotes the n-th
k, and ∀i, j = 1, 2, . . . , N . column of D̄kk .
6
Given MMSE detectors at the users, the achievable down- A. Max-min Power Control for Protocol 1
link SE of the k-th user can be calculated as
N In this part, we consider that mutual orthogonal pilot
St-MMSE τu X St-MMSE sequences are used in the uplink channel estimation phase.
Rk = 1− E log2 (1 + ζk,n ) , (25)
τc n=1 To achieve the fairness good SE for all users in the system,
max-min power control is applied for all users to optimize the
where downlink SE. The max-min fairness optimization problem can
2
fk,n d̄kk,n be written as
St-MMSE
ζk,n = 2. (26)
fH b ′
k,n Ψkk fk,n − fk,n d̄kk,n max min RkSt-SIC (31)
{ηmk } k=1,··· ,K
2) Protocol 2 with Linear MMSE Detectors: Let us denote
K
d̂kk,n as n-th column of D̂kk . Then, the linear MMSE detector X 1
s.t. ηmk γmk ≤ , m = 1, . . . , M
for the n-th data stream of the k-th user is given by LN
k=1
′
pk,n = Ψckk d̂kk,n , (27) ηmk ≥ 0, k = 1, . . . , K, m = 1, . . . , M,
′ H
where Ψckk = (Ψckk )−1 + D̂kk D̂kk . where the first constraint of (31) is the power constraint in
Given MMSE detectors at the users, the achievable down- (9) when using the mutual orthogonal pilot sequences for the
link SE of the k-th user can be calculated as τ ρ β2
uplink channel estimation phase, and γmk = τu ρuu βumkmk+1 . An
N
τu + τd X equivalent form of (31) is
Et-MMSE Et-MMSE
Rk = 1− E log2 (1 + ζk,n ) ,
τc n=1
P
M
2
(28) m=1 γmk ςmk
max min M K
{ηmk } k=1,··· ,K N 1
where 2
P P
L βmk γmk′ ςmk ′ + ρL2
2 m=1 k′ =1
pk,n d̂kk,n K
Et-MMSE
ζk,n = 1
2.
X
(29) s.t. 2
ηmk ςmk ≤ , m = 1, . . . , M
pH c′ − pk,n d̂kk,n LN
k,n Ψkk pk,n k=1
0.9
In this section, similarly to the section IV-A, we also con-
0.8
sider the max-min fairness optimization problem by applying
0.7
PC. However, it is very difficult (may be impossible) and
0.6
complicated to obtain the optimal solution for max-min power
Cdf
control of protocol 2, due to the intractable form of the spectral 0.5
0.5
0.4
that the optimal PC coefficients of protocol 1 can be applied
0.3 Protocol 2 Protocol 2 (Simulation) w/o PC
effectively to PC problem of protocol 2.
0.2
Protocol 2 (Simulation) w/ PC
Protocol 1 (Simulation) w/o PC
Figure 4 shows that the performance of protocol 2 is very
0.1
Protocol (Simulation) 1 w/ PC
Protocol 1 (Analytical) w/o PC
close to its upper bound, when perfect channel information is
0
Protocol (Analytical) 1 w/ PC
available at the users. This implies that our PC method applies
0 2 4 6 8 10 12 14 16 18 20
per-user SE (bits/s/Hz)
effectively on protocol 2. Moreover, this figure also compares
the performance of two protocols using different detection
Fig. 3. Cummulative distribution function (cdf) of per-user SE with M = 50, techniques at the users, i.e., MMSE-SIC and MMSE. In pro-
K = 10, N = 2, L = 4. tocol 1, there is no difference in SE between the system using
MMSE-SIC detectors and the one using MMSE detectors.
8
Therefore, the corresponding max-min power control can Fig. 8. 95%-likely per-user downlink SE with M = 50, L = 4, N = 4.
be efficiently solved by using the successive approximation
technique [34].
10
Figure 5 shows the per-user SE of protocol 2 with sub-
95%-likely Per-User Downlink SE (bits/s/Hz)
Protocol 1
optimal power control in section IV-B, and the one with perfect 9 Protocol 2
CSI at the user and optimal power control. We can see that 8
the performance gap is quite small. This verifies that the sub- 7
optimal power control works very well. 6
2
14
Protocol 1, M=20, L=2, K=10
95%-likely Per-User Downlink SE (bits/s/Hz)
systems. However, in scenario 2, the per-user downlink SE first numbers of APs, users and arbitrary numbers of antennas at
increases when the number of antennas per user increases with the APs and the users. There are no downlink pilots in the
increasing independent channels (or degrees of freedom) per first transmission protocol, while the downlink pilots were
user. Then, this SE will reach the maximum value and then it beamformed to the used for the CSI acquisition in the second
will decrease when the number of antennas per user increases. transmission protocol. Compared with the first protocol, the
Especially, when the number of users is large, single-antenna second protocol has higher channel estimation quality at the
user setups outperform multi-antenna user setups. The reason users, but higher channel estimation overhead. Numerical
is that although the number of independent channels per user results show that no protocol always shows the advantage
increases, the channel estimation overhead also increases. This over the other in the performance as they both depend on
channel estimation overhead largely dominates when N is the number of users and number of antennas per user in the
large. system. Finally, by using two protocols and combining with
Next, we evaluate the effect of AP antennas. In Fig. 8, multi-antenna users, the system can achieve the sub-optimal
95%-likely per-user achievable downlink SE of two protocols performance regardless of number of users in the system.
with PC are shown with arbitrary number of antennas per
APs. As expected, the performances of all protocols increase
A PPENDIX
proportionally with the number of antennas per AP, especially
when the total number of antennas at the users is small in A. Proof of Lemma 1
comparison with the total number of antennas at the APs
From (5), we have
(here, M × L = 200). This comes from the fact that when
L increases, the channel is more favorable, and hence, the K
!−1
√ X
inter-user interference reduces. At the same time, the array vec(Ĝmk ) = τu ρu βmk τu ρu βmiΦ Tu,ikΦ ∗u,ik + IN
gain increases. i=1
Finally, 95%-likely per-user achievable downlink SE of two ⊗ IL vec(Yu,mk ). (35)
protocols are shown in Fig. 9 with arbitrary number of users
and N = 2. At first, when the number of user is small, Finally, (6) is obtained by applying the following identity
protocol 2 achieves higher SE than protocol 1. However, when vec(ABC) = C T ⊗ Avec(B) on (35).
the number of users increases, protocol 1 achieves higher SE
than protocol 2, as a large proportion of coherence duration is
assigned for the downlink training. These above insights are B. Proof of Lemma 2
very important for us to design the framework for achieving We denote i-th column of Gmk is gmk,i , j-th column
the sub-optimal SE of the system, in the next section. of Ĝmk is ĝmk,j , and Gmk can be written as Gmk =
[gmk,1 gmk,2 . . . gmk,N ]. Then the downlink matrix channel of
D. Framework for achieving the sub-optimal SE
k-th user can be written as
Normally, the number of APs and number of antennas
dkk,11 dkk,12 ... dkk,1N
per APs are fixed. However, the number of users in the dkk,21 dkk,22 ... dkk,2N
system, which strongly affects on the system performance, Dkk = .
.. .. , (36)
.
is unknown. Therefore, we propose the framework that au- dkk,N 1 dkk,N N
tomatically chooses suitable protocol as well as the number PM 1/2 H
of active antennas per users, based on the number of users in where dkk,ij , m=1 ηmk gmk,i ĝmk,j . To estimate the el-
the systems. The framework is as follows ement dkk,ij of the matrix channel Dkk , firstly yTdp,ki is
• In the setup session, based on the real number of users, projected onto φd,kj to obtain
K, and their locations, the system can calculate its SE
for each protocol and for each number of active antennas ydp,k,ij = yTdp,k,i φd,k,j . (37)
per user, n, ∀n = 1, 2, . . . , N .
Then, MMSE estimation of dkk,ij given ydp,k,ij is calculated
• Choose the protocol with the number of active antennas
as follows
per user, n, which achieves the highest SE.
The framework only need to be updated with the time frame dˆkk,ij = E{dkk,ij } + Cdkk,ij ,ydp,k,ij ×
of PC, i.e., updated infrequently, as it just depends on the
× C−1
ydp,k,ij ,ydp,k,ij (ydp,k,ij − E{ydp,k,ij }) , (38)
large-scale fading and number of active users.
Remark 3: The SE, which is a result of the framework, is where Cx,y denotes the covariance of x and y.
sub-optimal, as it is considered under following conditions:
1) Compute dˆkk,ij with i = j:
using mutual orthogonal pilot sequences, using sub-optimal
power control for protocol 2, and using the same number of
( M )
X 1/2
H
active antennas per users. E{dkk,ii } = E ηmk gmk,i ĝmk,i
m=1
VI. C ONCLUSION M
1/2
X
In this paper, we evaluated the downlink SE of two transmis- =L ηmk γmk,i , (39)
sion protocols of cell-free massive MIMO system with finite m=1
10
h n oi
where γmk,i , E ĝmk,i ĝH mk,i , ∀n ∈ N . Note that Substituting (43) and (44) into (42), we have
nn
γmk,i not depends on the index of antenna in the m-th AP. M
√ X
Then, from (5 ), γmk,i can be calculated as Cdkk,ii ,ydp,k,ii = τd ρd,p L ηmk βmk γmk,i +
m=1
!2 !
γmk,i M
1/2
X
h n oi + L ηmk γmk,i . (45)
= E vec(Ĝmk )vec(Ĝmk )H , ∀l ∈ L
(i−1)L+l m=1
K
!−1 Next, we have
2
X H
= τu ρu βmk τu ρu Φu,ik βmiΦ̃
Φ̃ Φu,ik + ILN . ( M
!2)
i=1 √ X 1/2 H
(i−1)L+l Cydp,k,ii ,ydp,k,ii = E τd ρd,p ηmk gmk,i ĝmk,i + wk,i
(40) m=1
( M
!2 )
1/2
X
Furthermore, we have = E τd ρd,p ηmk gH
mk,i ĝmk,i
2
+ wk,i .
m=1
M
( )
√ X 1/2
(46)
E{ydp,k,ii } = E τd ρd,p ηmk gH
mk,i ĝmk,i + wk,i
m=1
Then, following the similar method on computing
M Cdkk,ii ,ydp,k,ii , we have
√ X 1/2
= L τd ρd,p ηmk γmk,i , (41)
M
!2
m=1
1/2
X
Cydp,k,ii ,ydp,k,ii = τd ρd,p ξkk,i + L ηmk γmk,i + 1.
( M !
X 1/2
H
Cdkk,ii ,ydp,k,ii = E ηmk gmk,i ĝmk,i × m=1
m=1 (47)
M
!)
√ X 1/2 Finally, (18) is derived by the substitution of (48), (49) and
× τd ρd,p ηmk gHmk,i ĝmk,i + wk,i
m=1 (50) into (38).
( M !2 ) 2) Compute dˆkk,ij with i 6= j:
√ X 1/2
= τd ρd,p E ηmk gH
mk,i ĝmk,i E{dkk,ij } = E{ydp,k,ij } = 0, (48)
m=1 ( M !
√
= τd ρd,p (P1 + P2 ),
X 1/2
(42) Cdkk,ij ,ydp,k,ij = E H
ηmk gmk,i ĝmk,j ×
nP o m=1
1/2 M
where smk , ηmk gH , E 2 M
!)
mk,i ĝmk,i , P1 m=1 smk and √ X 1/2
nP
M PM o × τd ρd,p ηmk gH
mk,i ĝmk,j + wk,i
P2 , E m=1 n6=m s mk s nk . m=1
Compute P1 : M
(a1)√ X
= L τd ρd,p ηmk βmk γmk,j , (49)
M
X n o m=1
P1 = E s2mk ( M
!2 )
√ X 1/2
m=1 Cydp,k,ij ,ydp,k,ij = E τd ρd,p ηmk gH
mk,i ĝmk,j + wk,i
M n L 2 o m=1
1/2
X X
= E ηmk gmk,il ĝmk,il M
(a2) X
m=1 l=1 = Lτd ρd,p ηmk βmk γmk,j + 1, (50)
M
X nXL o m=1
4 2 2
= ηmk E (ĝmk,il + g̃mk,il ĝmk,il )
where (a1) and (a2) follow the similar method on computing
m=1 l=1
L XL
! Cdkk,ii ,ydp,k,ii . The substitution of (48), (49) and (50) into (38)
yields (18).
nX o
+E gmk,il ĝmk,il gmk,il′ ĝmk,il′
3) Compute dˆkk′ ,ij with k 6= k ′ and ∀i, j : The computa-
l=1 l′ 6=l
M
tion is similar to that in Section B2.
(a) X
= ηmk (Lβmk γmk,i + L2 γmk,i
2
), (43)
m=1 C. Proof of Theorem 1
Mutual information is defined as [35]
where (a) follows the fact that gmk,il ĝmk,il and gmk,il′ ĝmk,il′
are independent with ∀l′ 6= l. I(qk ; rk , Θk ) = h(qk |Θk ) − h(qk |rk , Θk ), (51)
Compute P2 : where h(.) is the differential entropy, and Θk is the channel
information at k-th user. Choose suboptimal qk as CN (0, IN ),
M X
M
X 1/2 1/2 then
P2 = L2 ηmk ηnk γmk,i γnk,i . (44)
m=1 n6=m h(qk |Θk ) = log2 |πeIN |. (52)
11
K as M → ∞, and ∀i 6= j, (72)
(b1) X
= L βmi βmk tr(Φik′ Amk′ AH H
mk′ Φik′ )IN , (67) M
1/2
X
i6=k dkk,ii = ηmk gH
mk,i ĝmk,i ,
m=1
n o
T12 , E GH H H H
mk Gmk Φkk′ Amk′ Amk′ Φkk′ Gmk Gmk M M
1/2 1/2
ηmk ĝH
X X
(b2) 2 = mk,i ĝmk,i + ηmk g̃H
mk,i ĝmk,i ,
= Lβmk Cmkk′ , (68) m=1 m=1
M
1/2
ηmk ĝH
X
≈ mk,i ĝmk,i ,
Cmkk′ is a N × N diagonal matrix that the element
where P m=1
N M M
!
cii = n=1 bnn + Lbii with b is an element of Bmkk′ = d
X √ X
2
Φkk′ Amk′ AH H →N L ηmk γmk,i , L ηmk γmk,i ,
mk′ Φkk′ . and
m=1 m=1
as M → ∞, (73)
n o M
T13 , E GH H H
mk Wnk′ Amk′ Amk′ Wnk′ Gmk
X 1/2
ηmk′ gH
d
dkk′ ,ij = mk,i ĝmk′ ,j → CN (0, ξkk′ ,j ),
(b3) m=1
= Lβmk tr(Amk′ AH
mk′ )IN , (69)
as M → ∞, ∀k ′ 6= k, and i, j = 1, 2, . . . , N. (74)
The Gaussian approximation is verified by numerical results
where (b1), (b2) and (b3) are derived by Lemma 3. in the section V.
Substituting (67) (68) (69) into (66), we have
G. Proof of Theorem 3
Denote D̂kk′ are MMSE estimated of Dkk′ , and D̃kk′ ,
M X
K K
X
X Dkk′ − D̂kk′ . With assumption that the elements of the effective
T1 = Lρ βmk ηmk′ τu ρu βmi ×
channels Dkk′ are Gaussian distributed, the corresponding ele-
m=1 k′ =1 i6=k
ments of D̂kk′ and D̃kk′ are independent in distribution. Then,
tr(Φik′ Amk′ AH H
mk′ Φik′ ) + βmk Cmkk′ + tr(Amk′ AH
mk′ ) IN . with Θk = {D̂k1 , D̂k2 , . . . , D̂kK }, the achievable downlink SE
of the k-th user in (14) when using MMSE-SIC detectors at
(70) the receivers can be rewritten as
′ −1
Et-SIC τu + τd H c
Rk = 1− E log2 IN + ρD̂kk Ψkk D̂kk ,
To calculate T2 , we have τc
(75)
where
T2 ′
K
X H
K
X n H
o
M X K
M X Ψckk = ρ D̂kk′ D̂kk′ + ρ E D̃kk′ D̃kk′ + IN . (76)
1/2 1/2
X
= τu ρu ρ ηmk′ ηnk′ k′ 6=k k′ =1
m=1 n6=m H
k′ =1
n o
n o Next, we compute E D̃kk′ D̃kk′ :
× E GH
mk Gmk Φ kk ′ Amk′ A
H
nk ′ Φ H
kk ′ G H
nk G nk
H
n o
M X
M X
K E D̃kk′ D̃kk′
1/2 1/2
X
= τu ρu ρ ηmk′ ηnk′ n T ∗
o n T ∗
o n T ∗
o
E d̃ ′ d̃ ′ E d̃kk′ ,1 d̃kk′ ,2 ... E d̃kk′ ,1 d̃kk′ ,N
m=1 n6=m k′ =1 n Tkk ,1 ∗kk ,1 o n T ∗
o n T ∗
o
E d̃kk′ ,2 d̃kk′ ,1 E d̃kk′ ,2 d̃kk′ ,2 . . . E d̃kk′ ,2 d̃kk′ ,N
n o n o
× E GH
mk Gmk Φ kk ′ Amk′ A
H
nk ′ Φ H
kk ′ E G H
nk Gnk
=
.. ..
n T . ∗ o
. n T
M X K
M X ∗
o
X 1/2 1/2 E d̃kk′ ,N d̃kk′ ,1 E d̃kk′ ,N d̃kk′ ,N
= L2 τu ρu ρ ηmk′ηnk′ βmk βnk Φkk′ Amk′ AH H
nk′ Φkk′ . PN
˜
m=1n6=mk′ =1 j=1 var(dkk′ ,1j ) 0 ... 0
PN
0 var(d˜kk′ ,2j ) ... 0
(71) =
j=1 .
.. ..
.
.
PN ˜kk′ ,N j )
0 0 ... j=1 var(d
The substitution of (70) (71) and (61) into (62) yields (18). (77)
13
Then, the variance of the MMSE estimation error of dkk′ ,ij [19] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Massive MU-MIMO
can be calculated as [30] downlink TDD systems with linear precoding and downlink pilots,”
in 2013 51st Annual Allerton Conf. Commun., Control, and Comput.
var(d˜kk′ ,ij ) = Cdkk′ ,ij ,dkk,ij − Cdkk′ ,ij ,ydp,k,ij (Allerton), Oct. 2013, pp. 293–298.
[20] G. Interdonato, H. Q. Ngo, P. Frenger, and E. G. Larsson, “Downlink
× C−1
ydp,k,ij ,ydp,k,ij Cydp,k,ij ,dkk′ ,ij . (78) training in cell-free massive MIMO: A blessing in disguise,” IEEE Trans.
Wireless Commun., vol. 18, no. 11, pp. 5153–5169, 2019.
Following the similar method on section B1, we have [21] G. Interdonato, H. Q. Ngo, E. G. Larsson, and P. Frenger, “How much
do downlink pilots improve cell-free massive MIMO?” in 2016 IEEE
ξkk,i +κ2kk,i Glob. Commun. Conf. (GLOBECOM), Dec. 2016, pp. 1–7.
τd ρd,p (ξkk,i +κ2kk,i )+1
if k = k ′ , and i = j
˜
var(dkk′ ,ij ) = [22] Ö. Özdogan, E. Björnson, and J. Zhang, “Performance of cell-free
ξkk′ ,j otherwise . massive MIMO with Rician fading and phase shifts,” IEEE Trans.
τd ρd,p ξkk′ ,j +1 Wireless Commun., vol. 18, no. 11, pp. 5299–5315, 2019.
(79) [23] E. Björnson and L. Sanguinetti, “Making cell-free massive mimo com-
Finally, substituting (79) into (77), and then plugging (77) into petitive with mmse processing and centralized implementation,” arXiv
preprint arXiv:1903.10611, 2019.
(76), we obtain (21). [24] S. Buzzi and C. D’Andrea, “Cell-free massive MIMO: User-centric
approach,” IEEE Wireless Commun. Lett., vol. 6, no. 6, pp. 706–709,
R EFERENCES Aug. 2017.
[25] P. Liu, K. Luo, D. Chen, and T. Jiang, “Spectral efficiency analysis
[1] T. C. Mai, H. Quoc Ngo, and T. Q. Duong, “Cell-free massive MIMO of cell-free massive MIMO systems with zero-forcing detector,” IEEE
systems with multi-antenna users,” in 2018 IEEE Glob. Conf. on Sig. Trans. Wireless Commun., 2019.
and Inf. Process. (GlobalSIP), Nov 2018, pp. 828–832. [26] G. Interdonato, M. Karlsson, E. Björnson, and E. G. Larsson, “Local
[2] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral effi- partial Zero-forcing precoding for cell-free massive MIMO,” arXiv
ciency of very large multiuser MIMO systems,” IEEE Trans. Commun., preprint arXiv:1909.01034, 2019.
vol. 61, no. 4, pp. 1436–1449, 2013. [27] T. Van Chien, E. Björnson, and E. G. Larsson, “Joint power allocation
[3] E. Björnson, J. Hoydis, L. Sanguinetti, “Massive MIMO networks: and load balancing optimization for energy-efficient cell-free massive
Spectral, energy, and hardware efficiency,” Founda. and Trends R in MIMO networks,” arXiv preprint arXiv:2002.01504, 2020.
Sig. Process., vol. 11, no. 3-4, pp. 154–655, 2017. [28] A. Papazafeiropoulos, P. Kourtessis, M. Di Renzo, S. Chatzinotas, and
[4] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num- J. M. Senior, “Performance analysis of cell-free massive mimo systems:
bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9, A stochastic geometry approach,” IEEE Trans. Veh. Techno., 2020.
no. 11, pp. 3590–3600, 2010. [29] R. Nikbakht, R. Mosayebi, and A. Lozano, “Uplink fractional power
[5] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive control and downlink power allocation for cell-free networks,” IEEE
MIMO for next generation wireless systems,” IEEE Commun. Mag., vol. Wireless Commun. Let., 2020.
52, no. 2, pp. 186195, Feb. 2014. [30] S. M. Kay, “Fundamentals of statistical signal processing. Vol 1,
[6] E. Björnson, M. Matthaiou, and M. Debbah, “Massive MIMO with non- Estimation theory,” 1993.
ideal arbitrary arrays: Hardware scaling laws and circuit-aware design,” [31] X. Li, E. Björnson, S. Zhou, and J. Wang, “Massive MIMO with multi-
IEEE Trans. Wireless Commun., vol. 14, no. 8, pp. 4353–4368, 2015. antenna users: When are additional user antennas beneficial?” in 2016
[7] T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo, Fundamentals 23rd Int. Conf. on Telecommun. (ICT), May 2016, pp. 1–6.
of Massive MIMO. Cambridge Uni. Press, 2016. [32] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge
[8] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta, university press, 2004.
“Cell-free massive MIMO versus small cells,” IEEE Trans. Wireless [33] Q. Zhang, S. Jin, K.-K. Wong, H. Zhu, and M. Matthaiou, “Power
Commun., vol. 16, no. 3, pp. 1834–1850, Mar. 2017. scaling of uplink massive MIMO systems with arbitrary-rank channel
[9] S. Shamai and B. M. Zaidel, “Enhancing the cellular downlink capacity means,” IEEE J. Sel. Topics in Signal Processing, vol. 8, no. 5, pp.
via co-processing at the transmitting end,” in IEEE VTS 53rd Vehic. 966–981, Oct. 2014.
Tech. Conf., Spring 2001. Proceedings (Cat. No.01CH37202), vol. 3, [34] L.-N. Tran, M. F. Hanif, A. Tolli, and M. Juntti, “Fast converging
May 2001, pp. 1745–1749 vol.3. algorithm for weighted sum rate maximization in multicell MISO
[10] S. Venkatesan, A. Lozano, and R. Valenzuela, “Network MIMO: Over- downlink,” IEEE Signal Process. Lett., vol. 19, no. 12, pp. 872–875,
coming intercell interference in indoor wireless systems,” in 2007 Conf. 2012.
Record of the Forty-First Asilomar Conf. on Signals, Syst. and Comput., [35] T. M. Cover and J. A. Thomas, Elements of information theory. John
Nov 2007, pp. 83–87. Wiley & Sons, 2012.
[11] G. Interdonato, E. Björnson, H. Q. Ngo, P. Frenger, and E. G. Larsson, [36] T. Yoo and A. Goldsmith, “Capacity and power allocation for fading
“Ubiquitous cell-free massive MIMO communications,” EURASIP J. MIMO channels with channel estimation error,” IEEE Trans. on Inform.
Wireless Commun. Netw., no. 197, 2019. Theory, vol. 52, no. 5, pp. 2203–2214, 2006.
[12] Z. Chen and E. Björnson, “Channel hardening and favorable propagation
in cell-free massive MIMO with stochastic geometry,” IEEE Trans.
Commun., vol. 66, no. 11, pp. 5205–5219, Nov 2018.
[13] E. Nayebi, A. Ashikhmin, T. L. Marzetta, H. Yang, and B. D. Rao, “Pre-
coding and power optimization in cell-free massive MIMO systems,”
IEEE Trans. Wireless Commun., vol. 16, no. 7, pp. 4445–4459, 2017.
[14] H. Q. Ngo, L. N. Tran, T. Q. Duong, M. Matthaiou, and E. G. Larsson,
“On the total energy efficiency of cell-free massive MIMO,” IEEE Trans.
Green Commun. and Netw., vol. 2, no. 1, pp. 25–39, Mar. 2018.
[15] J. Zhang, Y. Wei, E. Björnson, Y. Han, and X. Li, “Spectral and
energy efficiency of cell-free massive MIMO systems with hardware
impairments,” in 2017 9th Int. Conf. Wireless Commun. and Signal
Process. (WCSP), Oct. 2017, pp. 1–6.
[16] M. Bashar, K. Cumanan, A. G. Burr, H. Q. Ngo, and M. Debbah, “Cell-
free massive MIMO with limited backhaul,” in 2018 IEEE Int. Conf. on
Commun. (ICC). IEEE, 2018, pp. 1–7.
[17] T. C. Mai, H. Q. Ngo, M. Egan, and T. Q. Duong, “Pilot power control
for cell-free massive MIMO,” IEEE Trans. Veh. Technol., vol. 67, no. 11,
pp. 11 264–11 268, Nov. 2018.
[18] T. C. Mai, H. Q. Ngo, and T. Q. Duong, “Uplink spectral efficiency
of cell-free massive MIMO with multi-antenna users,” in 2019 3rd
Inter. Conf. on Recent Advances in Sig. Process., Telecommun. Comput.
(SigTelCom), Mar. 2019, pp. 126–129.