1

Cell-Free Massive MIMO-OFDM Transmission over

Frequency-Selective Fading Channels
Wei Jiang, Senior Member, IEEE, and Hans Dieter Schotten, Member, IEEE

Abstract—This letter presents and analyzes orthogonal TABLE I

frequency-division multiplexing (OFDM)-based multi-carrier Definitions of Main Mathematical Symbols.
transmission for cell-free massive multi-input multi-output (CFm-
MIMO) over frequency-selective fading channels. Frequency- Symbol Definition
arXiv:2107.01402v1 [cs.IT] 3 Jul 2021

domain conjugate beamforming, pilot assignment, and user- βmk Large-scale fading between AP antenna m and user k
specific resource allocation are proposed. CFmMIMO-OFDM is hmk Small-scale fading between AP antenna m and user k
scalable to serve a massive number of users and is flexible to gmk Channel vector between AP antenna m and user k
N
offer diverse data rates for heterogeneous applications. gmk N-point channel vector by padding zeros at the tail of gmk
N
g̃mk Frequency-domain channel vector, i.e., the DFT of gmk
Index Terms—Cell-free massive MIMO, cell edge, frequency- g̃rmk and ĝrmk Frequency-domain channel response and its estimate
selective fading channel, massive MIMO-OFDM, OFDM.
xm and x̃m Time- and frequency-domain transmitted symbol vectors
xcpm and yk
cp
Transmitted and received symbol vectors with cyclic prefix
yk and ỹk Time- and frequency-domain received symbol vectors
I. Introduction U and Us The set of all users and the sth user group

C ELL-FREE massive multi-input multi-output (CFm- B and B s The set of all RBs and the RBs assigned to Us
MIMO) has recently received much attention from both K and Kr The number of all users and the number of users over Br
N and NRB The number of subcarriers and the number of RBs
academia and industry [1]. Different aspects of CFmMIMO
such as resource allocation [2], power control [3], pilot as-
signment [4], energy efficiency [5], backhaul constraint [6],
(·)H express conjugate, transpose, and Hermitian transpose. For
and scalability [7] have been studied. However, previous
ease reference, main mathematical symbols and notation are
works merely considered frequency-flat fading (narrow-band)
listed in Table I.
channels. Most of wireless communications nowadays are
broadband with signal bandwidths far wider than the co-
herence bandwidth, leading to frequency selectivity. To the II. System Model
authors’ best knowledge, only [8] discussed CFmMIMO over Consider a geographical area where M randomly distributed
frequency-selective fading channels. But it applied single- access points (APs) are connected to a central processing
carrier transmission that does not support high data rate due unit (CPU) via a fronthaul network and serve K users. We
to prohibitively high complexity of signal equalization [9]. first assume that each AP and user is equipped with a single
Moreover, the conventional CFmMIMO systems generally aim antenna as [1] for simple analysis but will demonstrate its
to offer uniform service by maximizing the minimum of per- adaptability to multi-antenna APs afterwards. In contrast to the
user rate [3], whereas it neglects the fact that heterogeneous conventional CFmMIMO such as [1]–[8] that requires K ≪ M,
users have differentiated demands on data throughput. the number of users in CFmMIMO-OFDM is scalable, ranging
To fill this gap, this letter will present and analyze orthog- from small K ≪ M to very large K ≫ M. Users are divided
onal frequency-division multiplexing (OFDM)-based multi- into groups and each group is assigned to different resource
carrier transmission for CFmMIMO, coined cell-free mas- blocks (RBs). Thus, the constraint that the number of users is
sive MIMO-OFDM (CFmMIMO-OFDM), over frequency- far smaller than M is still satisfied on each RB.
selective fading channels. Contamination-free pilot assignment The previous works assume that the small-scale fading is
and channel estimation in the uplink time-frequency grid, frequency-flat, as modelled by a circularly-symmetric complex
and frequency-domain conjugate beamforming in the down- Gaussian random variable with zero mean and unit variance,
link data transmission are provided. A user-specific resource i.e., h[t]∼CN(0, 1). This assumption is only valid for narrow-
allocation method that enables the scalability to accommodate band communications. Nevertheless, most of the current and
massive number of users and the flexibility of offering diverse future mobile communications [10] are broadband, suffering
data rates is proposed. Per-user and sum data throughput, con- from severe frequency selectivity. This letter goes beyond
sidering APs equipped with either single or multiple antennas, the current state of the art by studying CFmMIMO over a
are evaluated. Notations:: Bold lower- and upper-case letters frequency-selective fading channel [11]. It can be modeled as a
denote vectors and matrices, respectively, while (·)∗ , (·)T , and linear time-varying filter h[t]= [h0 [t], . . . , hL−1 [t]]T , where the
filter length L should be no less than multi-path delay spread l m
Corresponding author: Wei Jiang (e-mail: [email protected])) T d normalized by the sampling interval T s , namely L > TTds .
W. Jiang and H. D. Schotten are with German Research Centre for Artificial P
Intelligence (DFKI), Kaiserslautern, Germany, and are also with the University The tap gain hl [t] = i ai (tT s )e− j2π fc τi (tT s ) sinc[l − τi (tT s )Bw]
of Kaiserslautern, Germany. for l = 0, . . . , L − 1, with carrier frequency fc , attenuation
 2

AP m
Pilot
cp
x m [t ] x [t ]
m

x m [t ] g mk [t ] g1k [t ]

g Mk [t ]
y k [t ]
y cp
k [t ]
y k [t ]
User k
(a) (b)

Fig. 1. Schematic diagram of a CFmMIMO-OFDM system where M AP antennas serve K users: Part (a) elaborates the OFDM transmitter at the mth AP
antenna and the receiver at the kth user. A radio frame consists of NT OFDM symbols in the axis of time and N subcarriers in the frequency domain, which
are grouped into NRB resource blocks. Part (b) illustrates the time-frequency resource grid of a single RB containing λRB consecutive subcarriers. UL/DL
switch denotes the guard interval between the uplink and downlink transmission.

ai (tT s ) and delay τi (tT s ) of the ith signal path, signal bandwidth where ⊗ stands for the cyclic convolution [9] and gmk N
[t] is
Bw = 1/T s, and sinc(x) , sin(x) x for x , 0. The fading channel an N-point channel filter formed by padding zeros at the tail
[t]= gmk,0 [t], . . . , gmk,Lmk −1 [t], 0, . . . , 0 T . The
N  
between AP m and user k is given by of gmk [t], i.e., gmk
  p
gmk [t] = gmk,0 [t], . . . , gmk,Lmk −1 [t] T = βmk [t]hmk [t], (1) DFT demodulator outputs
p
where gmk,l [t] = βmk [t]hmk,l [t] and βmk [t] indicates large-scale ỹk [t] = Fyk [t]. (5)
fading, which is frequency independent and varies slowly. Substituting (3) and (4) into (5), and applying the convolution
III. Cell-Free Massive MIMO-OFDM system theorem for DFT [9], we have
M
The data transmission in an OFDM system [9] is orga- X 
N

nized in block-wise, as shown in Fig.1a. We write x̃m [t] = ỹk [t]= F gmk [t] ⊗ xm [t] + Fzk [t]
 
x̃m,0 [t], . . . , x̃m,n [t], . . . , x̃m,N−1 [t] T to denote the frequency- m=1
M
domain transmission block of AP m on the tth OFDM symbol. X
= g̃mk [t] ⊙ x̃m [t] + z̃k [t], (6)
(The tilde ˜ marks frequency-domain variables throughout m=1
this letter.) Transform x̃m [t] into a time-domain sequence
 
xm [t] = xm,0 [t], . . . , xm,n′ [t], . . . , xm,N−1 [t] T through an N- where ⊙ represents the Hadamard product (element-wise
N
point inverse discrete Fourier transform (IDFT), i.e., xm,n′ [t] = multiplication), g̃mk [t] = Fgmk [t] and z̃k [t] = Fzk [t] denote
1 P N−1 2π jn′ n/N
for n′ = 0, 1, . . . , N − 1 and j2 = −1. frequency-domain channel response and noise, respectively.
N n=0 x̃m,n [t]e
Defining the discrete Fourier transform (DFT) matrix At last a frequency-selective channel is transformed into a
set of N independent frequency-flat subcarriers. The signal
ω0·(N−1)
 
 ω0·0 N ··· N  transmission in the downlink on the nth subcarrier is given by
 . . 
F =  .. . . . .. (2) M
X
  ỹk,n [t] = g̃mk,n [t] x̃m,n [t] + z̃k,n [t], k ∈ {1, . . . , K}, (7)
(N−1)·0 (N−1)·(N−1) 
ωN · · · ωN m=1
′ ′
2π jnn /N
with a primitive N th root of unity ωn·n N =e , the OFDM where g̃mk,n [t] is the nth element of g̃mk [t]. Similarly, the uplink
modulation can be written in matrix form as transmission is expressed by
1
xm [t] = F−1 x̃m [t] = F∗ x̃m [t]. (3) K
X
N ỹm,n [t] = g̃mk,n [t] x̃k,n [t] + z̃m,n [t], m ∈ {1, . . . , M}. (8)
A guard interval known as cyclic prefix (CP) is added between k=1
two consecutive blocks to avoid inter-symbol interference
(ISI) and preserve orthogonality of subcarriers. Thus, we IV. The communication process
h iT get
xcp
m [t] = x m,N−Lcp [t], . . . , x m,N−1 [t], x m,0 [t], . . . , x m,N−1 [t] as The downlink (DL) transmission from the APs to the users
the transmitted signal. The ISI can be eliminated if the length and the uplink (UL) from the users to the APs are separated
of CP is larger than the length of any channel filter, i.e., by time-division multiplexing (TDD) with the assumption of
Lcp > max (Lmk ), for all m ∈ {1, . . . , M} and k ∈ {1, . . . , K}. perfect channel reciprocity. A radio frame is mainly divided
The signal xcp m [t] goes through the channel gmk [t] to reach into three phases: UL training, UL payload transmission, and
cp
the typical user k, resulting in xm [t] ⋆ gmk [t], where ⋆ denotes DL payload transmission.
the linear convolution. Thus, the overall received signal at user
k is ycp
P M cp
k [t] = m=1 xm [t] ⋆ gmk [t] + zk [t], where zk [t] is a vector A. Uplink Training
of additive noise. Removing the CP, we get
We write Rht, ni to denote a resource unit (RU) on the
M
X
N nth subcarrier of the tth OFDM symbol. The time-frequency
yk [t] = gmk [t] ⊗ xm [t] + zk [t], (4)
m=1
resource of a radio frame is divided into NRB RBs, each
 3

   
2 2
of which contains λRB = N/NRB (assumed to be an in- which applies Rgg = E g̃rmk = βmk and Rnn = E z̃m,n [t] =
teger) consecutive subcarriers. The rth RB is defined as σ2z . Compute the variance of ĝrmk as
Br , {Rht, ni|1 ≤ t ≤ NT and (r − 1)λRB ≤ n < rλRB}, for any  
r ∈ {1, . . . , NRB }, as illustrated in Fig.1b. The transmission of  β2 |P |2 √ p g̃r P + z̃ k [tk ] 2 
h i  mk k u mk k m,n p p 
a radio frame in CFmMIMO is carried out within the coherent E ĝrmk (ĝrmk )∗ =E 
 
 2
(βmk |Pk | + σz )2 2 
time and the width of one RB is smaller than the coherence
bandwidth. Consequently, this letter simply adopts the block 
√ r 2
fading model where the channel coefficients for all RUs within β2mk E pu g̃mk Pk + z̃m,nkp [tkp ] pu β2mk
one RB are assumed to be identical, we have = = .
(βmk + σ2z )2 pu βmk + σ2z

g̃mk,n [t] = g̃rmk ⇐= Rht, ni ∈ Br . (9) Consequently, we know that ĝrmk ∈ CN(0, αmk ) with αmk =
pu β2mk
pu βmk +σ2z
, in comparison with g̃rmk ∈ CN(0, βmk ).
The channel estimation of a conventional CFmMIMO sys-
tem relies on time-domain pilot sequences, where the maximal B. Uplink Payload Data Transmission
number of orthogonal sequences is τ p by using τ p pilot sym-
Suppose τu OFDM symbols are used for uplink transmis-
bols. If K 6 τ p as assumed in [3], pilot contamination can be
sion, on the RU Rht, ni ∈ Br , τ p < t ≤ τ p + τu , all Kr
avoided. Owing to the limitation of the frame length, however,
users simultaneously transmit their signals to hthe APs.i The
some users need to share the same sequence when K > τ p ,
kth user weights its transmit symbol qk,n [t], E |qk,n[t]|2 = 1,
leading to pilot contamination [4]. In contrast, CFmMIMO- √
OFDM is able to provide more orthogonal pilots by means of by a power-control
√ coefficient ψk , 0 ≤ ψk ≤ 1. Substituting
frequency-division multiplexing thanks to the extra degree of x̃k,n [t] = ψk pu qk,n [t] into (8) yields
freedom gained from the frequency domain. To estimate g̃rmk , √
Kr
X p
each user on Br needs only one pilot symbol. Suppose the ỹm,n [t] = pu g̃mk,n [t] ψk qk,n [t] + z̃m,n [t]. (13)
first τ p OFDM symbols are dedicated to the uplink training, k=1
one RB has N p = τ p λRB orthogonal pilots. The number of
users allocated to Br is denoted by Kr , there is no pilot C. Downlink Payload Data Transmission
contamination if Kr 6 N p , which is a very relaxed condition. Inspired by [1]–[5] that apply conjugate beamforming in the
We write Rhtkp , nkp i with 1 6 tkp 6 τ p and (r−1)λRB 6 nkp < rλRB downlink, CFmMIMO-OFDM employs conjugate beamform-
to denote the RU reserved for the pilot symbol of user k, ing in the frequency domain. On Rht, ni ∈ Br , τ p +τu < t ≤ NT ,
k ∈ {1, . . . , Kr }. Other users keep silence (null) on this RU each AP multiplexes a total of Kr symbols, i.e., sk,n [t] intended
to achieve orthogonality. As illustrated in Fig.1b, the pilot to user k, k = 1, . . . , Kr , before transmission. With a power-
assignment is specified by √
control coefficient ηmk , 0 ≤ ηmk ≤ 1, the transmitted signal
√ of the mth AP is
 pu Pk ,

 if t = tkp ∧ n = nkp Kr
x̃k,n [t] =  , 1 ≤ t ≤ τp , (10) √ X √

0,
 otherwise x̃m,n [t] = pd ηmk ĝmk,n [t] ∗ sk,n [t]. (14)
k=1
where ∧ represents logical AND, Pk is the known pilot symbol Substituting (14) into (7) to get the received signal at user k
with E[|Pk |2 ] = 1, and pu denotes the UL transmit power limit.
M r K
Substituting (9) and (10) into (8), we get the received signal √ X X √
ηmk′ ĝmk′ ,n [t] ∗ sk′ ,n [t] + z̃k,n [t]


ỹk,n [t]= pd g̃mk,n [t]
of the mth AP on Rhtkp , nkp i as m=1 k′ =1
M
Kr
√ X √ 2
X = pd ηmk ĝmk,n [t] sk,n [t]
ỹm,nkp [tkp ]= g̃mk,nkp [tkp ] x̃k,nkp [tkp ] + z̃m,nkp [tkp ] m=1
k=1
| {z }
Desired S ignal
Kr
X M K
=g̃mk,nkp [tkp ] x̃k,nkp [tkp ] + g̃mk′ ,nkp [tkp ] x̃k′ ,nkp [tkp ] + z̃m,nkp [tkp ]
r
√ X X √

+ pd ĝmk,n [t] ηmk′ ĝmk′ ,n [t] ∗ sk′ ,n [t] (15)
k′ ,k
√ m=1 k′ ,k
= pu g̃mk,nkp [tkp ]Pk + z̃m,nkp [tkp ] | {z }
√ Multi−User Inter f erence
= pu g̃rmk Pk + z̃m,nkp [tkp ]. (11) M r K
√ X X √

+ pd ξmk,n [t] ηmk′ ĝmk′ ,n [t] ∗ sk′ ,n [t] + z̃k,n [t] .
r
Let ĝrmk be an estimate of g̃rmk , we have ĝrmk = g̃rmk − ξmk with m=1 k′ =1
|{z}
r | {z } Noise
estimation error ξmk raised by additive noise. Applying the Channel−Estimate Error
minimum mean-square error (MMSE) estimation gets
Each userP is assumed to have the knowledge of channel statis-
! ! M √ 2
Rgg P∗k βmk P∗k tics E m=1 ηmk ĝmk,n [t] rather than channel realizations
ĝrmk = ỹm,nkp [tkp ] = ỹm,nkp [tkp ], ĝmk,n [t] since there is no pilot and channel estimation in the
Rgg |Pk |2 + Rnn βmk |Pk |2 + σ2z
(12) downlink. Applying a similar method used in Theorem 1 of
 4

[1], we can derive that the spectral efficiency

 of user k on which applies Rgg = βqk INt and Rnn = σ2z INt and we get ĝrqk ∈
Rht, ni ∈ Br is lower bounded by log2 1 + γkht,ni with pu β2
qk
CN(0, αqk INt ) with αqk = p β +σ 2 . The transmitted signal can
u qk z
P
M √ 2
also be expressed by (14) since each antenna independently
ht,ni
pd m=1 ηmk αmk
γk = PM P , (16) generates its signal using local channel estimates, regardless
σ2z + pd m=1 βmk kK′r=1 ηmk′ αmk′ of how many antennas per AP. With conjugate beamforming,
implying that the effect of small-scale fading is vanished (a.k.a. the power control is generally decided by βqk [3], we have
channel hardening), such that γkht,ni = γkr , for all Rht, ni ∈ Br . ηmk = ηqk for m = (q − 1)Nt + 1, . . . , qNt . Then, (16) can be
rewritten for the case of multi-antenna AP as
V. User-specific Resource Allocation P N √ 2
ht,ni
pd Nt2 q=1 AP
ηqk αqk
The conventional CFmMIMO systems [1]–[8] can support γk = PNAP P . (19)
only very few K ≪ M users with uniform quality of service. σ2z + pd Nt q=1 βqk kK′r=1 ηqk′ αqk′
By exploiting the frequency domain, CFmMIMO-OFDM is
adaptive to different numbers of users from a few K ≪ M to
VII. Numerical Results
massive K ≫ M and is flexible to offer diverse data rates for
heterogeneous users. Classify all users U = {u1 , u2 , · · · , uK } The performance of CFmMIMO-OFDM in terms of per-
into different groups U s , s = 1, 2, . . . , S in terms of their user and sum rates is evaluated. We consider a square area
S
demands on data throughput, subjecting to Ss=1 U s = U of 1 × 1km2 where a total of M AP antennas serve K users.
T The COST-Hata model in [1] is applied for the path loss,
and U s U s′ = ∅, ∀s′ , s. The number of users in U s
P
satisfies |U s | ≪ M and Ss=1 |U s | = K, where |·| stands for the where the three-slope breakpoints take values d0 = 10m
cardinality of a set. The resource pool is B = {Br |1 ≤ r ≤ NRB } and d1 = 50m while L = 140.72dB with carrier frequency
where the granularity for allocation is one RB. Using B s to fc = 1.9GHz, the height of AP antenna hAP = 15m, and
S PLmk +Xmk
denote the RBs allocated to U s , we have Ss=1 B s ∈ B (when the height of user antenna hu = 1.65m. Use βmk = 10 10
SS T
all RBs are used, s=1 B s = B) and B s B s′ = ∅, ∀s′ , s. If to get large-scale fading, where the shadowing fading Xmk ∼
Br ∈ B s , the number of users served by this RB is Kr = |U s |. N(0, σ2sd ) with σ sd = 8dB. 3GPP Extended Typical Urban
For any user k ∈ U s , its per-user data rate in the downlink is (ETU) model specified by delay profile in millisecond of
!
τ p + τu X   {0, 0.05, 0.12, 0.2, 0.23, 0.5, 1.6, 2.3, 5} with relative power as
Rk = 1 − λRB △ f log2 1 + γkr , (17) {−1, −1, −1, 0, 0, 0, −3, −5, −7} is applied for small-scale fad-
NT B ∈B r s ing1 . The maximum transmit power of AP antenna and user are
where △ f is the subcarrier spacing. Then, the downlink sum pd = 0.2W and pu = 0.1W, respectively. The APssimply adopt −1
P P P Kr
data throughput of the system is Rd = Ss=1 k∈Us Rk . the full-power transmission strategy, i.e., ηmk = k=1 αmk .
The white noise power density is −174dBm/Hz with a noise
VI. Adaptability to Multi-Antenna AP figure of 9dB. Deducting a guard band of 1MHz at both sides,
Inspired by [12], which argues that channel hardening and the signal bandwidth of Bw = 20MHz is divided into N = 1200
favorable propagation can be more easily obtained by deploy- subcarriers with △ f = 15kHz. Each RB consists of λRB = 12
ing multiple antennas per AP, we further study CFmMIMO- subcarriers and thus NRB = 100. We apply NT = 10 for a frame
OFDM with multi-antenna APs [5]. Generalizing the number including τ p pilot symbols and NT − τ p downlink symbols,
of antennas per AP to Nt > 1, there are NAP = M/Nt neglecting uplink transmission (i.e., τu = 0) in our simulation.
(assumed to be an integer) homogeneous APs indexed by Fig.2a compares cumulative distribution functions (CDFs)
q = 1, . . . , NAP . The only difference raised by multi-antenna of per-user and sum rates over a single RB (i.e., 180kHz)
APs is that collocated antennas have identical large-scale achieved by CFmMIMO-OFDM with M = 128 single-antenna
fading, i.e., βmk = βqk for any m = (q − 1)Nt + 1, . . . , qNt . As APs and Kr = 6, 12, 24, or 36. As we expected, per-user rate
[5] and [12], small-scale fading from each collocated antenna decreases when more users share the same amount of resource.
to the typical user is assumed to be independent. As a result, But the more users are served together, the higher sum data
the channel model with multi-antenna APs is simply a special throughput is achieved. In contrast to a single RB, the system
case of gmk in (1), which is transparent to the communication can offer diverse transmission rates through the user-specific
process and does not need any modifications. When the users resource allocation. Table II shows the achievable capacity of
transmit pilot symbols, see (11), the observation at AP q is the whole system using 100 RBs. We are interested in two
√ extreme cases: mobile broadband (MBB) that requires very
ỹq,nkp [tkp ] = pu g̃rqk Pk + z̃q,nkp [tkp ], high rate and machine-type communications (MTC) with a
where ỹq and z̃q are Nt × 1 received symbol and noise vectors, massive number of low-rate users. Let K = 6 MBB users (only
h iT
respectively, and g̃rqk = g̃r(q−1)Nt +1,k , . . . , g̃rqNt ,k ∈ CNt ×1 stands 1 The near-far effect raised by distributed APs not only generates different
for the channel vector between AP q and user k. Similar to distance-dependent large-scale fading but also causes diverse delay spreads
(12), the MMSE estimate of g̃rqk can be obtained by in small-scale fading. Due to channel hardening, however, small-scale fading
does not affect the final performance of conjugate beamforming, which relies
! !
r
Rgg P∗k k
βqk P∗k only on large-scale fading. This argument can be proved by (16) and (19),
ĝqk = ỹ k [t ] = ỹ k [tk ], and is observed also in simulations. However, it is necessary to consider
Rgg |Pk |2 + Rnn q,n p p βqk |Pk |2 + σ2z q,n p p heterogeneous delay spreads raised by cell free anywhere small-scale fading
(18) can affect performance.
 5

TABLE II
95%-likely and 50%-likely (median) per-user data rates, and averaged sum data throughput for MBB and MTC.

M = 128 M = 256 M = 512

System
MBB MTC MBB MTC MBB MTC
Rate
K=6 K = 12 K = 24 K = 36 K = 1200 K = 2400 K = 3600 K=6 K = 3600 K=6 K = 3600
95%-likely [Mbps] 28.20 19.96 10.34 5.66 0.199 0.103 0.057 40.47 0.137 54.40 0.229
Median [Mbps] 46.45 38.19 27.64 20.61 0.381 0.276 0.206 59.55 0.297 74.16 0.401
Sum [Mbps] 273.70 448.21 644.92 720.70 447.83 644.50 721.23 351.57 1034.97 437.44 1397.28

Single-antenna AP, M=128

1
K =6
r

K r =12
CDF

0.5
K =24
r

K =36
r

0
0 100 200 300 400 500 600 700 800
Per-user rate per RB [Kbps]
1
CDF

0.5

0
1 2 3 4 5 6 7 8 9
Sum rate per RB [Mbps]

(a) (b)

Fig. 2. Performance comparison with respect to the CDFs of per-user and sum data rates over a single RB. Part (a) utilizes M = 128 single-antenna APs
and the number of users varies from Kr = 6, 12, 24, to 36. In part (b), a total of M = 256 antennas are distributed over a set of multi-antenna APs (with
Nt = 1, 2, 4, 8 or 16) to serve Kr = 12 users.

one group) consume all 100 RBs, the 95%-likely per-user rate different numbers of users ranging from a few to thousands
can reach 28.2Mbps and the 50%-likely or median rate is and can offer data rates from tens of Kbps to tens of Mbps.
46.45Mbps with M = 128. Increased to M = 512, an MBB
user can get a 95%-likely rate of 54.4Mbps and a median rate References
of 74.16Mbps. In contrast, we equally divide K = 3600 users [1] H. Q. Ngo et al., “Cell-free massive MIMO versus small cells,” IEEE
Trans. Wireless Commun., vol. 16, no. 3, pp. 1834–1850, Mar. 2017.
into S = 100 groups to accommodate as many MTC users as [2] S. Buzzi et al., “User-centric 5G cellular networks: Resource allocation
possible. Each group with Kr = 36 users is allocated with a and comparison with the cell-free massive MIMO approach,” IEEE
single RB. With M = 128, each MTC user gets a 95%-likely Trans. Wireless Commun., vol. 19, no. 2, pp. 1250–1264, Feb. 2020.
[3] E. Nayebi et al., “Precoding and power optimization in cell-free massive
rate of 57Kbps and a median rate of 206Kbps. Using M = 512, MIMO systems,” IEEE Trans. Wireless Commun., vol. 16, no. 7, pp.
the system can support 3600 MTC users with an experienced 4445–4459, Jul. 2017.
rate of 229Kbps. In addition, the performance of multi-antenna [4] W. Zeng et al., “Pilot assignment for cell-free massive MIMO systems
using a weighted graphic framework,” IEEE Trans. Veh. Technol., Apr.
APs is also evaluated where a total of M = 256 antennas serve 2021, doi:10.1109/TVT.2021.3076440.
Kr = 12 users. When the number of antennas per AP increases [5] H. Q. Ngo et al., “On the total energy efficiency of cell-free massive
from Nt = 1 to 2, both per-user and sum rates boost, as shown MIMO,” IEEE Trans. Green Commun. Netw., vol. 2, no. 1, pp. 25–39,
Mar. 2018.
in Fig.2b. Increased to Nt = 4, although the sum rate can [6] H. Masoumi and M. J. Emadi, “Performance analysis of cell-free
grow slightly, the 95%-likely per-user rate drops. Remarkable massive MIMO system with limited fronthaul capacity and hardware
performance loss on both per-user and sum rates is observed impairments,” IEEE Trans. Wireless Commun., vol. 19, no. 2, pp. 1038–
1052, Feb. 2020.
if Nt = 16. That is because some users may locate far away [7] E. Björnson and L. Sanguinetti, “Scalable cell-free massive MIMO
from any APs when there are only a few APs, implying the systems,” IEEE Trans. Commun., vol. 68, no. 7, pp. 4247–4261, Jul.
Marco-diversity stemming from CFmMIMO vanishes. 2020.
[8] S.-N. Jin, D.-W. Yue, and H. H. Nguyen, “Spectral efficiency of a
frequency-selective cell-free massive MIMO system with phase noise,”
IEEE Wireless Commun. Lett., Dec. 2020.
VIII. Conclusion [9] W. Jiang and T. Kaiser, “From OFDM to FBMC: Principles and Com-
parisons,” in Signal Processing for 5G: Algorithms and Implementations,
The proposed CFmMIMO-OFDM system is effective to F. L. Luo and C. Zhang, Eds. United Kindom: John Wiley&Sons and
IEEE Press, 2016, ch. 3.
combat frequency selectivity and pilot contamination can be [10] W. Jiang et al., “The road towards 6G: A comprehensive survey,” IEEE
eliminated by exploiting frequency-domain orthogonal pilots. Open J. Commun. Society, vol. 2, pp. 334–366, Feb. 2021.
In contrast to CFmMIMO that supports only very few users [11] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
Cambridge, UK: Cambridge Univ. Press, 2005, ch. 2.
with uniform service quality, it is scalable to the scale of users [12] Z. Chen and E. Björnson, “Channel hardening and favorable propagation
and is adaptive to diverse service demands. It was verified by in cell-free massive MIMO with stochastic geometry,” IEEE Trans.
numerical results that CFmMIMO-OFDM can accommodate Commun., vol. 66, no. 11, pp. 5205–5219, Nov. 2018.