See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/383022379

RIS-Assisted Cell-Free Massive MIMO Relying on Reflection Pattern

Modulation

Article in IEEE Transactions on Communications · August 2024

DOI: 10.1109/TCOMM.2024.3446589

CITATIONS READS

2 42

5 authors, including:

Zeping Sui Hien Quoc Ngo

University of Essex Queen's University Belfast
24 PUBLICATIONS 59 CITATIONS 301 PUBLICATIONS 18,813 CITATIONS

SEE PROFILE SEE PROFILE

Trinh Van Chien Michail Matthaiou

Hanoi University of Science and Technology Queen's University Belfast
154 PUBLICATIONS 2,247 CITATIONS 410 PUBLICATIONS 12,953 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Zeping Sui on 11 August 2024.

The user has requested enhancement of the downloaded file.

 1

RIS-Assisted Cell-Free Massive MIMO Relying

on Reflection Pattern Modulation
Zeping Sui, Member, IEEE, Hien Quoc Ngo, Senior Member, IEEE, Trinh Van Chien, Member, IEEE,
Michail Matthaiou, Fellow, IEEE, and Lajos Hanzo, Life Fellow, IEEE

Abstract—We propose reflection pattern modulation-aided Index Terms—Cell-free massive MIMO, energy efficiency,
reconfigurable intelligent surface (RPM-RIS)-assisted cell- iterative optimization, reconfigurable intelligent surfaces,
free massive multiple-input-multiple-output (CF-mMIMO) reflection pattern modulation, spectral efficiency.
schemes for green uplink transmission. In our RPM-RIS-
assisted CF-mMIMO system, extra information is conveyed
by the indices of the active RIS blocks, exploiting the joint I. I NTRODUCTION
benefits of both RIS-assisted CF-mMIMO transmission and
RPM. Since only part of the RIS blocks are active, our Cell-free massive multiple-input-multiple-output (CF-
proposed architecture strikes a flexible energy vs. spectral mMIMO) schemes constitute promising candidates for
efficiency (SE) trade-off. We commence with introducing the next-generation wireless communications [1]–[3]. Specif-
system model by considering spatially correlated channels. ically, a large number of access points (APs) are ran-
Moreover, we conceive a channel estimation scheme subject
domly distributed in a given area to support many user
to the linear minimum mean-square error (MMSE) con-
straint, yielding sufficient information for the subsequent sig- equipments (UEs), which are connected to one or several
nal processing steps. Then, upon exploiting a so-called large- central processing units (CPUs) via backhaul links [4],
scale fading decoding (LSFD) scheme, the uplink signal-to- [5]. Compared to conventional cellular mMIMO systems,
interference-and-noise ratio (SINR) is derived based on the where the cell-edge UEs suffer from poor quality-of-
RIS ON/OFF statistics, where both maximum ratio (MR)
service (QoS) and severe inter-cell interference [6], [7],
and local minimum mean-square error (L-MMSE) combiners
are considered. By invoking the MR combiner, the closed- a CF-mMIMO system experiences reduced interference
form expression of the uplink SE is formulated based only and excellent load-balancing, yielding improved spectral
on the channel statistics. Furthermore, we derive the total efficiency (SE), lower transmission latency and increased
energy efficiency (EE) of our proposed RPM-RIS-assisted network capacity [1], [2], [4]. However, under propagation
CF-mMIMO system. Additionally, we propose a chaotic
scenarios with poor scattering and/or high transmission
sequence-based adaptive particle swarm optimization (CSA-
PSO) algorithm to maximize the total EE by designing the path loss, the SE and EE of CF-mMIMO will be compro-
RIS phase shifts. Specifically, the initial particle diversity mised [8]. As a remedy, reconfigurable intelligent surfaces
is promoted by invoking chaotic sequences, and an adap- (RISs) have been widely integrated with various commu-
tive time-varying inertia weight is developed to improve nication scenarios [9], [10]. Specifically, a RIS includes a
its particle search performance. Furthermore, the particle
large number of passive elements coordinated by a micro-
mutation and reset steps are appropriately selected to enable
the algorithm to escape from local optima. Finally, our controller, while the reflected signals can be shaped in
simulation results demonstrate that the proposed RPM-RIS- terms of their phase and magnitude at the electromagnetic
assisted CF-mMIMO architecture strikes an attractive SE level, resulting in passive beamforming [11]. Since no
vs. EE trade-off, while the CSA-PSO algorithm is capable active beamformers and sophisticated signal processing
of attaining a significant EE performance gain compared to
techniques are needed, this results in decreased overall
conventional solutions.
energy consumption [12].
This work is a contribution by Project REASON, a UK Government More recently, RIS-assisted CF-mMIMO systems have
funded project under the Future Open Networks Research Challenge been widely investigated in the existing literature [8], [13]–
(FONRC) sponsored by the Department of Science Innovation and Tech- [17]. Specifically, in [13], both the uplink and downlink
nology (DSIT). It was also supported by the U.K. Engineering and Physi-
cal Sciences Research Council (EPSRC) (grant No. EP/X04047X/1). The SEs of RIS-assisted CF-mMIMO systems was studied
work of M. Matthaiou was supported by the European Research Council based on maximum ratio (MR) processing and the es-
(ERC) under the European Union’s Horizon 2020 research and innovation timated cascaded two-hop channels were derived, where
programme (grant agreement No. 101001331).
L. Hanzo would like to acknowledge the financial support of the Engi- single-antenna APs were considered. Shi et al. [14] con-
neering and Physical Sciences Research Council projects EP/W016605/1, sidered the electromagnetic interference (EMI) in RIS-
EP/X01228X/1, EP/Y026721/1, EP/W032635/1 and EP/X04047X/1 as assisted CF-mMIMO systems, and the uplink SE was also
well as of the European Research Council’s Advanced Fellow Grant
QuantCom (Grant No. 789028) derived. The uplink SE and EE of a RIS-assisted CF-
Zeping Sui, Hien Quoc Ngo and Michail Matthaiou are with the Centre mMIMO system having realistic hardware impairments
for Wireless Innovation (CWI), Queen’s University Belfast, Belfast BT3 were investigated in [8], while a joint max-min SE op-
9DT, U.K. (e-mail: {z.sui, hien.ngo, m.matthaiou}@qub.ac.uk).
Trinh Van Chien is with the School of Information and Communication timization paradigm of both large-scale fading decoding
Technology, Hanoi University of Science and Technology, Hanoi, 10000, (LSFD) and power control was also proposed. A two-
Vietnam (email: [email protected]). timescale transmission protocol-based RIS-assisted CF-
Lajos Hanzo is with the Department of Electronics and Computer
Science, University of Southampton, Southampton SO17 1BJ, U.K. (e- mMIMO system was proposed in [15], where the RIS-
mail: [email protected]). based passive beamforming was designed relying on the
 2

channel statistics and the MR combiner was formulated spatially correlated channels. Explicitly, by intrinsically
based on the instantaneous channels. In [16], Yao et amalgamating RPM and RIS-assisted CF-mMIMO, we
al. investigated a robust beamforming scheme of a RIS- propose the RPM-RIS-assisted CF-mMIMO philosophy
assisted CF-mMIMO system by incorporating the realistic for green communications.
channel state information (CSI) uncertainties, where a The contributions of this paper are boldly contrasted to
limited-rate backhaul was considered. In [17], a com- the existing literature in Table I, which are itemized below:
prehensive literature survey of wireless energy transfer
in RIS-aided CF-mMIMO systems was carried out. Both
• We propose an RPM-RIS-assisted CF-mMIMO
the downlink and uplink SE of RIS-aided CF-mMIMO
scheme for supporting green uplink transmission over
systems with channel aging were analyzed in [18]. Diluka
spatially correlated channels, where information is
et al. derived tight bounds of the achievable rate and
conveyed both by the classic amplitude-phase mod-
outage probability in closed-form for RIS-assisted cell-free
ulated (APM) symbols and the indices of the active
systems in [19]. Later in [20], the authors investigated RIS-
RIS elements. The RIS-reflected channel is modeled
aided CF-mMIMO systems relying on single-antenna APs.
by the correlated Rician fading distribution. Then, the
Moreover, a closed-form expression of the achievable rate
uplink linear minimum mean-square error (MMSE)
was derived. In [21], upon considering a wideband RIS-
cascaded two-hop channel estimation algorithm is
aided cell-free network, a joint precoding scheme was con-
presented, where realistic pilot contamination is con-
ceived for maximizing the network capacity. On the other
sidered and yields sufficient information for sequen-
hand, the maximum EE optimization of RIS-aided cell-
tial data processing. Moreover, the LSFD cooperation
free networks was disseminated in [22]–[25]. Explicitly,
technique is exploited to formulate the uplink input-
limited backhaul capacity was considered in [24], while
output relationship, where both the MR and local
hybrid RIS-based systems were studied in [25]. However,
MMSE (L-MMSE) combiners are invoked.
the above-mentioned studies consider RISs only as passive
• Based on the MR combiner and the RIS ON/OFF
reflectors to reshape the propagation channels. It is known
statistics, a closed-form expression is derived for
from [26] that extra information can also be transmitted by
the uplink SE for finite numbers of UEs and APs.
a RIS, leading to an improved communication data rate,
Explicitly, the effects of pilot contamination, channel
which is close to the RIS channel capacity in theory.
estimation error, and RIS phase shifts are consid-
As a parallel development, reflection modulation (RM) ered. Furthermore, the amount of information con-
has been proposed to achieve additional information trans- tained in the received signals is derived by invoking
mission [27], [28]. In [27], a passive beamforming and entropy theory. Upon utilizing the results derived,
information transfer diagram was proposed. In more de- the total EE of our proposed RPM-RIS-assisted
tails, extra bits can be mapped onto the ON/OFF states CF-mMIMO system is investigated. The simulation
of the RIS elements, while passive beamforming can be results demonstrate that the proposed RPM-RIS-
achieved by appropriately adjusting the active phase shifts. assisted CF-mMIMO system is capable of approach-
However, since the number of active RIS elements changes ing the SE of the conventional RIS-assisted CF-
among the transmitted symbols, the outage probability mMIMO. Moreover, it can also be observed from the
performance of passive beamforming and information simulations that our RPM-RIS-assisted CF-mMIMO
transfer is compromised [28]. To circument this problem, can attain better total EE than RIS-assisted CF-
Lin et al. [29] proposed reflection pattern modulation mMIMO, where all RIS elements are active.
(RPM), where the RIS reflection pattern is fixed during • Aiming for maximizing the total EE under a RIS
each symbol transmission, thereby enhancing the outage phase shift constraint and per-user EE constraint, the
probability compared to that of the conventional passive optimization problem of RIS phase shift design is
beamforming and information transfer scheme. In this formulated. Due to the non-convexity of the formu-
context, a received signal power vs. achievable rate trade- lated optimization problem, we propose a chaotic
off can also be formulated. Then, the RPM system was ex- sequence-based adaptive particle swarm optimization
tended to quadrature reflection modulation in [30], where (CSA-PSO) algorithm to solve the RIS phase shift
all RIS elements are activated. In [31], extra information design problem. In more detail, we use chaotic
bits were mapped onto the superimposed RIS phase shifts. sequences to improve the initial particle diversity.
Upon dividing the RIS into active and passive blocks, a Furthermore, a novel adaptive inertia weight factor is
novel hybrid reflection modulation scheme was proposed proposed for improving its search performance, and
in [32]. However, the above-mentioned design philosophy beneficial particle mutation and reset steps are con-
has not been harnessed for RIS-assisted CF-mMIMO ceived to help the particles escape from local optima.
systems. Moreover, by scanning the related literature, we It is shown that our proposed CSA-PSO significantly
can conjecture that the performance of RIS-assisted CF- enhances the EE compared to its conventional PSO
mMIMO systems can be further improved by leveraging and random phase shift based counterparts. Combined
RPM techniques. Against this backdrop, we activate part with the SE simulation results, we conclude that
of the RIS elements for supporting the uplink transmis- the proposed RPM-RIS-assisted CF-mMIMO system
sion of a RIS-assisted CF-mMIMO system over Ricean studies an attractive SE vs. EE trade-off.
 3

TABLE I
C ONTRASTING THE CONTRIBUTIONS TO STATE - OF - THE - ART

Contributions This paper [8] [13] [14] [15] [16] [27]–[29]

RIS-assisted CF-mMIMO X X X X X X
RPM X X
Spatially correlated Rician channels X X
LSFD cooperation X X X
SE in closed-form of RPM-RIS-assisted CF-mMIMO X
EE analysis X X X
Max EE-based RIS phase shift design X X
PSO-based EE optimization X
Complexity analysis of PSO-like algorithms X

The remainder of our paper is organized as follows: The blocks, and each block has N = L/G elements, yielding
system model is described in Section II, while the uplink overall LA = N K active elements. Now, we focus our
data transmission and SE are derived in Section III. In attention on the m-th AP-RIS group. At j the beginning of
G

k
Section IV, the uplink power consumption and EE are each channel coherence interval, L1 = log2 K bits are
investigated, while the RIS phase shift design formulated mapped onto K activated RIS block index symbols, which
for uplink EE maximization is introduced in Section V. In are then transmitted to the corresponding RIS controller
Section VI, we present our simulation results. Finally, our to activate specific RIS blocks. Briefly, there are only K
conclusions are discussed in Section VII. out of G blocks that stay activated (ON-state), and the
Notation: Henceforth, ⊗, E{·} and tr(·) are the Kro- remaining (G − K) are OFF-state blocks. Hence, there
necker product, expectation and trace operators. Matrices are a total of C = 2L1 reflection patterns (RPs), which
and vectors are denoted by upper- and lower-case boldface can be expressed as Im = {Im,1 , . . . , Im,C }. Let us
letters, respectively. A complex Gaussian distribution with denote the cth RP as Im,c = {Im,c (0), . . . , Im,c (K − 1)},
mean vector a and covariance matrix B is denoted by where Im,c (k) ∈ ZG + for k = 0, . . . , K − 1, and the
CN (a a, B ); (·)T , (·)∗ , (·)H , (·)−1 and |·|2 are the transpose, corresponding activated RIS element index set is given by
conjugate, conjugate transpose, inverse and magnitude Īm,c = {Īm,c (0), . . . , Īm,c (LA − 1)}, where Im,c (lA ) ∈
operators respectively; B(a, b) represents the (a, b)th el- ZL+ for lA = 0, . . . , LA − 1. Moreover, the transmitted
ement of the matrix B ; I N denotes the N -dimensional symbols of U UE are generated based on a Q-ary nor-
identity matrix; U[a, b] represents the uniform distribution malized quadrature amplitude modulation (QAM)/phase-
in the interval [a, b]. The determinant operator is denoted shift keying (PSK) constellation B = {b1 , . . . , bQ }, yield-
by det(·). The Euclidean norm operator and truncation ing L2 = U log2 Q bits. We consider a quasi-static
argument are denoted by k·k and b·c, respectively. Finally, block flat-fading channel model having a τc symbol-
A = diag{a a} denotes a diagonal matrix A with the length coherence interval, while the length of the uplink
elements of the vector a on the diagonal. channel estimation block is τp symbols. Consequently,
the remaining τu = τc − τp symbols are utilized for
II. S YSTEM M ODEL AND C HANNEL E STIMATION uplink data transmission. It should be noted that the uplink
Let us consider a centralized RPM-RIS-assisted CF- SE at the UE side is only associated withPthe UE bit
U
mMIMO communication system that involves U single- component L2 , yielding the sum SE as η SE = u=1 ηuSE =
antenna UEs, M APs and M RISs, as shown in Fig. (τu /τc ) log2 (1 + δu ), where δu denotes uplink signal-to-
1. Specifically, each RIS includes L elements and each interference-and-noise ratio (SINR) of UE u. On the other
AP is equipped with J antennas. Moreover, each RIS is hand, the RIS bit component L1 is taken into account
associated with an RIS controller, capable of adjusting the when we calculate the amount of information contained
phase shift and reflection amplitude of all RIS elements, at the AP side, as formulated in (15) and (16). Moreover,
while receiving information symbols from UE [29].1 We we emphasize that since the RIS RP is associated with
assume that the locations of all the APs and users are the RIS-reflected channels, the bit sequence L1 is only
randomly chosen in a specific area, and the time-division sent at the beginning of each coherence interval. This
duplexing (TDD) operation is adopted similarly to con- implies that the RP remains fixed during the coherence
ventional CF-mMIMO systems [1], yielding channel reci- interval, and the RIS elements have to switch across the
procity between the uplink and downlink transmissions. ON/OFF state, when a new coherence interval commences.
By utilizing RIS beamforming, the reflected signal of one Since the RIS controller has to detect the bit sequence
RIS can be directed to its associated AP [33], [34], i.e., L1 to obtain the RIS RP, the hardware complexity of our
each AP is served by only one RIS that is employed close proposed RPM-RIS-assisted CF-mMIMO is higher than
to the corresponding AP. All the RISs are divided into G that of the conventional schemes.2

1 Since our proposed channel estimation method is conceived for the 2 For the typical maximum likelihood detector, the complexity is on
cascaded channels only in terms of statistical CSI, it is straightforward the order of O(2L1 ). There are also other low-complexity detectors that
to extend our architecture to generalized systems in which the number can be considered, such as MMSE and approximate message passing
of APs and RISs are different, as shown in [13], [14]. (AMP) detectors [35], [36].
 4

A. Channel Model RIS m

RIS controller
The lA th phase shift of the m-th RIS can be for- Φm
mulated as θlA ,m ∈ [−π, π], ∀m, lA . Therefore, the m- z mu
Gm
th RIS phase shift matrix can be expressed as Φ m =
diag{eiθ1,m , . . . , eiθLA ,m } ∈ CLA ×LA . The channel ma-
f mu
trix between the AP m and the m-th RIS can be expressed UE u AP m
as G m ∈ CJ×L , while z mu ∈ CL denotes the channel AP-RIS group
vector between the m-th RIS and the UE u. We consider
an urban environment having a multi-scatterer distribution,
such that there is no direct path in the UE-AP link channel.
Consequently, the channel between the AP m and the UE
u is realistically modeled via the Rayleigh fading distri-
bution, which is given as f mu ∼ CN (00, R mu ), ∀m, u, Fig. 1. Illustration of the RPM-RIS-assisted CF-mMIMO system em-
where R mu ∈ CJ×J denotes the corresponding spatial ploying M AP-RIS groups and U UEs.
correlation matrix, and βmu = tr(R Rmu )/J is the large- αm ξmu
where we have bmu , (κm +1)(ι , the LoS com-
scale fading coefficient. Moreover, we consider the sce- √ mu +1)
ponent is given by h̄ hmu = bmu κm ιmuḠGmΦ mz̄z mu ,
nario that the UE-RIS and RIS-AP channels are modeled 1
the NLoS components are formulated as h̃ hmu =
using the Rician distribution with high line-of-sight (LoS) √ 2 √
probability. Explicitly, the UE-RIS and RIS-AP channels bmu κmḠ GmΦ mz̃z mu , h̃
hmu = bmu ιmG̃ GmΦ mz̄z mu and
3 √
can be respectively formulated as h
h̃ mu = G Φ z h
bmuG̃ m mz̃ mu , while h̄ mu and h̃ hmu ,
P3 c
r hmu + f mu represent the LoS and NLoS compo-
c=1 h̃
ξmu √ nents of the aggregated channel, respectively. We assume
z mu = ( ιmuz̄z mu + z̃z mu ) ,
ιmu + 1 that the LoS component h̄ hmu is known at all the APs [40].
αm √
r  Moreover, we have the following statistical result of the
Gm = Gm + G̃
κmḠ Gm , (1)
κm + 1 unknown NLoS component h̃ hmu .
where ξmu and αm are the path-loss coefficients, while Proposition 1: The covariance matrix of the NLoS
ιmu and κm denote the Rician coefficients, respectively. component h̃ hmu of (2) can be formulated as
Moreover, z̄z mu ∈ CLA and Ḡ Gm ∈ CJ×LA represent the H
n o
H
R hmu ,E h̃ hmuh̃hmu = bmu κmḠ GmΦ mR̃ RmuΦ H Gm
mḠ
deterministic LoS components, while the non-LoS (NLoS)
components can be formulated as z̃z mu ∼ CN (00, R̃ Rmu ) + R mu + Π mu + Ξ mu , (3)
and vec(G̃ Gm ) ∼ CN (00, R̃ Rmu ∈ CLA ×LA
Rm ). Moreover, R̃ n h o n h o
and R̃Rm ∈ CJLA ×JLA denote the covariance and full where Π mu , E R̃Rmu,2 , Ξ mu , E R̃
Rmu,3 , while
correlation matrices, which can be respectively formulated h h 2
R̃ hmu and
Rmu,3 are the correlation matrices of h̃
Rmu,2 and R̃
as R̃Rmu = R RIS,m = dH dV R m and R̃ Rm = (R RTAP,m ⊗ 3
R RIS,m )/JLA [14], where R AP,m and R RIS,m denote the hmu , respectively.
h̃
correlation matrices at the AP side [37] and the RIS side, Proof: The proof is offered in Appendix B. 
respectively, while dH and dV denote the horizontal width
and the vertical height of each RIS element. Moreover,
the RIS spatial correlation matrix R m ∈ CLA ×LA can B. Uplink Channel Estimation
be obtained based on [38]. For the LoS components, we
By utilizing τp pilot symbols {φ φ1 , . . . , φ τp } with
consider that the AP √ and√RIS have a (J ×1)-uniform linear
array (ULA) and ( L× L)-uniform squared planar array φt ||2 = τp transmitted by all UEs where t ∈ {1, . . . , τp },
||φ
(USPA), respectively. Therefore, the LoS components can we carry out the channel estimations independently, where
be expressed as z̄z mu = a LA (ϕamu , ϕemu ) and Ḡ Gm = we have τp < U and assume that τp pilot symbols are
a
a J (υm e a
, υm )a LA (ϕ̄am , ϕ̄em )H [39], where ϕamu and ϕemu shared by U UEs. Let Pu ⊂ {1, . . . , U } represent the UE
are the azimuth and elevation angles of arrival (AoA) of index set that harnesses the same pilot symbol as the UE u,
the signal from the UE u to RIS m. Furthermore, υm a
and including the u-th UE itself. More specifically, the pilot
e
υm denote the azimuth and elevation angles of departure symbol allocated to UE u is given by φ tu ∈ Cτp with
(AoD) from the m-th RIS to the m-th AP, while ϕ̄am as tu ∈ {1, . . . , τp }. Given UE u and Pu , the pilot reuse
well as ϕ̄em represent the azimuth and elevation AoA at pattern can be formulated as φ Ttk φ ∗tu = τp , k ∈ Pu and
the m-th AP, respectively. Moreover, the elements of the φ Ttk φ ∗tu = 0, k ∈
/ Pu . The symbol received at AP m can
PU √
steering vectors a LA (χa , χe ) ∈ CLA and a J (χa , χe ) ∈ CJ be formulated as Y pm = k=1 pkh mkφ Ttk + N m , where
can be obtained based on [15]. Given the activated RIS pk ≥ 0 represents the normalized SNR per pilot symbol of
block index subset Ic and the m-th phase shift matrix, the the kth UE during channel estimation, while the elements
RIS-assisted aggregated channel spanning from the u-th of the noise term N m ∈ CJ×τp are independent and
UE to AP m can be formulated as identically distributed (i.i.d.) Gaussian random variables
(RVs) with zero mean and unit variance. Consequently,
hmu + h̃
h mu = f mu + G mΦ mz mu , h̄ hmu , (2) the pilot-projected-version of the received training signal
 5

y pmu = Y pmφ ∗tu ∈ CJ can be obtained as the weighting factorsPcmu for m = 1, . . . , M , yielding the
M
√ X √ final estimate ŝu = m=1 c∗mu s̃mu as
y pmu = pu τph mu + pk τph mk + n mu , (4)
U
k∈Pu \{u} X
√ ŝu = c H
u g uu su + cH
u g uk sk + nu , (9)
where n mu = N mφ ∗tu / τp∼ CN (00, τpI J ). Based on (4), k6=u
and upon invoking the linear MMSE channel estimation
technique [4], the estimated aggregated channel can be where we have g uk = [vv H H
1u h 1k , . . . , v M u h M k ]
T
∈ CM
c T M
formulated as and u = [c1u , . . . , cM u ] ∈ C denotes the weighting
M
√ c∗muv H
P
factor vector, nu = mun m , while the equivalent
hmu = h̄
ĥ hmu + puR hmuΨ −1 mu (yy pmu − ȳy pmu ), (5) m=1
h channel gain can be expressed as {ccH
u g uk , k = 1, . . . , U }.
y pmu =
P
where Ψ mu = k∈Pu pk τpR mk + I J and ȳ
P √
k∈Pu hmk . The expectation and covariance ma-
pk τph̄
trices of the channel estimate ĥ hmu and the estimation B. Uplink Spectral Efficiency
hmu = h mu − ĥ
error h̃ hmu can be respectively derived as We assume that each UE is capable of obtaining only the
E{ĥhmu } = h̄ hmu , Cov{ĥ
hmu } = C̄ C mu = pu τpΓ mu and channel statistics but not the specific CSI. Upon exploiting
E{h̃hmu } = 0 , Cov{h̃hmu } = Λ mu , where we have the use-and-then-forget bounding method [1], the uplink
SE associated with UE u (in bit/s/Hz) can be expressed
Γ mu , R hmuΨ −1 h
muR mu , Λ mu , R hmu − C̄
C mu . (6) based on (9) as
hmu and h̃
It should be noted that ĥ hmu = h mu − ĥ
hmu are ηuSE = (τu /τc ) log2 (1 + δu ) , (10)
uncorrelated.
where δu is the effective uplink SINR of UE u, defined
as follows
III. U PLINK DATA T RANSMISSION AND S PECTRAL p̄u |ccH g uu }|2
u E{g
δu =
E FFICIENCY U
p̄k E{|ccH cH g uu }|2 + σ 2c H
P 2
u g uk | } − p̄u |c u E{g u V uc u
k=1
A. Uplink Data Transmission Phase
p̄u |ccH g uu }|2
u E{g
We denote the symbol transmitted by the u-th user =  U
 , (11)
√ cH
P
p̄kΩ uk − p̄uG u + σ 2V u c u
as su = p̄u s̄u , where p̄u represents the uplink power u
k=1
control coefficient, and s̄u is generated based on a Q-
ary normalized QAM/PSK constellation B = {b1 , . . . , bQ } where G u = E{gg uu }E{gg H uu } and Ω uk ∈ C
M ×M

with E{|s̄u |2 } = 1. The J-dimensional signal received at represents the inter-user interference introduced
the m-th AP is given by by pilot contamination and imperfect channel
U
estimation, whose elements can be expressed
ym =
X
h mu su + n m = H mP 1/2s̄s + n m , (7) as E{vv H H 0
 h m0 kv2m0 u }, ∀m,m , and
muh mk we have
u=1
V u = diag E ||vv 1u || , . . . , E ||vv M u ||2 .
√ √ Proposition 2: Upon exploiting the MR combining
where P 1/2 = diag{p̄p} with p̄p = [ p̄1 , . . . , p̄U ]T de- scheme associated with v mu = ĥ hmu and the LSFD
noting the uplink power control matrix, s̄s = [s̄1 , . . . , s̄U ]T technique, the uplink SINR of UE u can be expressed in
is the symbol vector, H m
= [h hm1 , . . . , h mU ] ∈ CJ×U closed-form as
2
while n m ∼ CN 0 , σ I J represents the noise term. We
assume that each AP is capable of processing the uplink CH
p̄u |tr(C u Υ u )|
2
δu = U
, (12)
received signal locally by applying a specific combining P
C uk ) +
P
Du)
p̄k tr(C̄ p̄k Luk + tr(D
vector v mu ∈ CJ to y m [4], [14]. Hence, the local symbol k=1 k∈Pu \{u}
estimates s̃mu = v H
muy m can be formulated as
where we have C̄C uk = C H H 2
u T uk C u and D u = C u (σ Υ u −
U 2 CH
X p̄u∆ u )C u . In (12), we specifically denote C u =
s̃mu = v H
muh mu su + vH H
muh mk sk + v mun m , (8)
diag{ccu } ∈ CM ×M and Υ u = diag{ξ̄ξ u } ∈ CM ×M ,
k6=u H
where ξ̄ξ u = [ξ¯1u , . . . , ξ¯M u ]T with ξ¯mu = h̄ hmuh̄
hmu +
where any combiner can be exploited in (8). In Γmu ) associated with the desired signal. Moreover,
pu τp tr(Γ
this paper, we first consider the simple MR com- the non-coherent interference can be formulated as T uk =
bining scheme associated with v mu = ĥ hmu . More- diag{µ µuk } ∈ CM ×M and µ uk = [µuk,1 , . . . , µuk,M ]T ,
over,
hP we invoke
 the
H
L-MMSE
 i−1 v mu =
combiner where
U
p̄u hmkĥ
k=1 p̄k ĥ hmk + Λ mk + σ 2I M hmu to at-
ĥ H
n o ΓmuR hmk ) + h̄
µuk,m =pk τp tr(Γ hmuR hmkh̄
hmu
tain the minimum MSE = E ksu − v H mu y m k2 h
|ĥ mk . H H
hmkΓ muh̄
+ pk τph̄ hmk + |h̄ hmk |2 .
hmuh̄ (13)
Then, we exploit the LSFD scheme to mitigate the in-
terference [4]. All the local symbol estimates {s̃mu , ∀m} Furthermore, the coherent interference Luk can be ex-
are sent to the CPU and are linearly combined by applying pressed by Luk = pu pk τp2 |tr(C
C uW uk )|2 and W uk =
 6

diag{w w uk } ∈ CM ×M , where w uk = [wuk,1 , . . . , wuk,M ]T [12]. Finally, we have PuUE = p̄u /αUE + PuUE,fix with the
with  wuk,m = tr(R RhmkΨ −1 h
muR mu ), and ∆ u = power amplifier efficiency αUE and UE-side fixed power
2 2
diag ||h̄h1u || , . . . , ||h̄
hM u || . consumption PuUE,fix . The total uplink power consumption
Proof: The proof is provided in Appendix C.  can be expressed as
It should be noted that (11) is in a generalized Rayleigh U M
quotient form. Therefore, based on the technique in [4],
X τu X
P tot = P fix + γ RIS
p̄u + Pm +B C̄m %̃m , (17)
the optimal LSFD weighting vector c opt u optimized for u=1
τc m=1
maximizing the SINR given in (11) is fix
!−1 where P denotes the fixed power consumption, %̃m =
XU %AP BH
m + %m , and the total EE measured in (bit/Joule) can
opt 2
cu = p̄kΩ uk − p̄uG u + σ V u E{gg uu }, (14) be formulated as η EE = Bη SE /P total [2].
k=1
The static power consumption of RIS includes the
yielding the corresponding maximum SINR as δuopt = adjustable electronic circuit and the number of control-
p̄u E{gg H copt
uu }c u . Consequently, the sum SE can be expressed signal-related components [43]. For the RPM-RIS-assisted
U
as η SE = u=1 ηuSE . CF-mMIMO systems, the power dissipated by the elec-
P
tronic circuit is higher than that of the conventional RIS-
IV. P OWER C ONSUMPTION AND E NERGY E FFICIENCY assisted CF-mMIMO, since the ON/OFF states of RIS
Based on the standard random coding theory of [41], elements are switched at the beginning of each coherence
the transmitted symbol vector s̄ can be modeled as s̄s ∼ interval. However, only a fraction of the RIS elements are
CN (00, I U ). Therefore, based on (7), given the effective activated in the proposed RPM-RIS-assisted CF-mMIMO
channel matrix H m and the uplink power control matrix systems, yielding lower power consumption than the RIS-
P , it can be observed that the received signal follows the assisted CF-mMIMO. Therefore, we still exploit the con-
distribution of y m ∼ CN (00, H mP H H 2 ventional power consumption model of (17) in this paper.
m +σ I J ). Therefore,
the conditional entropy of y m and the differential entropy A detailed power consumption analysis of our RPM-RIS-
n m can be respectively formulated as [41] assisted CF-mMIMO will be carried out in our future
work.
H(yy m |H H mP H H
H m ) = log2 {det[πe(H 2
m + σ I J )]},
nm ) = log2 {det(πeσ 2I J )}.
H(n (15)
V. RIS P HASE S HIFT D ESIGN FOR U PLINK E NERGY
Given the RP and AP index m, the channel matrix H m
E FFICIENCY M AXIMIZATION
is seen to be a function of the RIS phase shift matrix
Φ m . Therefore, based on the mutual information between
A. Problem Statement
H mP 1/2s̄s and y m , the amount of information contained
in y m measured in (bit/s/Hz) can be formulated as Upon denoting the collected set of RIS phase shift
C̄m = EH m {I(yy m ; H mP 1/2
s̄s|H
H m )} matrices as Φ = {Φ Φ1 , . . . , Φ M }, hence the optimization
problem of maximizing the total EE can be formulated as
= EH m {H(yy m |H
H m )} − H(n
nm )

EE
H mP H H max η
  
m
 (18a)
= EH m log2 det + I J . (16)  Φ

σ2 (P0 ) : s.t. ηuSE ≥ ηu,min
EE
, ∀u, (18b)
Consequently, the total uplink
PM power consumption can


−π ≤θ ≤ π, ∀l , m.

be formulated as P tot = P AP
+
PM
P BH
+

lA ,m A (18c)
PM PU m=1 m m=1 m
RIS UE AP
m=1 m P + P
u=1 u [2], [12], [42], where P m and It can be readily shown that the objective function (18a)
BH
Pm represent the m-th AP’s power consumption and the and the constraint of (18b) are non-convex. The EE op-
corresponding backhaul link, respectively. Furthermore, timization problem of existing works is formulated based
RIS
Pm and PuUE represent the power consumption of the on the instantaneous SINR [22]–[25]. By contrast, our cor-
AP
m-th RIS and u-th UE, respectively. Also, Pm can be responding optimization problem (P0 ) is conceived based
AP τu AP AP,fix AP,a
expressed as Pm = B τc C̄m %m + Pm + JPm , on our new closed-form expression of the ergodic SINR.
AP
where %m denotes the traffic-dependent power coefficient Consequently, it can be observed from (12) and (17) that
AP,fix
measured in (Watt/bit/s), Pm is the fixed AP power the expression of η EE is complex and the phase shift ma-
AP,a
consumption related to the traffic load, while Pm is the trices Φ m for m = 1, . . . , M are highly coupled. Hence,
traffic-agnostic uplink power consumed by each antenna it is challenging to derive the globally optimal solution of
of the m-th AP. Moreover, we have Pm BH
= B ττuc C̄m %BHm + (18) upon utilizing conventional optimization algorithms,
BH,fix BH
Pm , with bandwidth B, while %m denotes the traffic- such as majorization-minimization (MM), gradient ascent-
dependent power coefficient measured in (Watt/bit/s), and based algorithms and successive convex approximation
BH,fix
Pm represents the fixed power dissipated by the m- (SCA) algorithms. Hence we opted for the heuristic PSO
th backhaul link. Then, the RIS power consumption is technique. Therefore, we propose a CSA-PSO scheme for
RIS
given by Pm = LA P (b), ∀m, where P (b) is the power solving the phase shift design optimization problem shown
consumed by each RIS element having b-bit resolution in (18).
 7

B. Phase Shift Optimization Using CSA-PSO Algorithm 1 CSA-PSO Algorithm

Require: κi (d), µ̃, ω 0 , c1 , c2 , ωmin and ωmax .
We assume that the CSA-PSO population Q includes I 1: Preparation: Set a maximum iterations number Tmax , ti,check = 0,
individuals. Based on the system model shown in Section ni = 0 and Ni = 10
II, there are a total of M LA phase shift elements. yielding 2: Initialize Compute θ 0i using chaotic sequence.
3: while t = 0 < Tmax ∧ni < Ni do
θ = {θθ 1 , . . . , θ M } associated with θ m ∈ CLA , ∀m. 4: Calculate (ηiEE )t .
Hence, the ith initial individual can be expressed as 5: Obtain the local best particles θ ti,pbest , global best particles θ tgbest
θ 0i = {θi0 (1), . . . , θi0 (M LA )}, where each element is in and local worst particles θ tpworst based on (ηiEE )t .
the range of [−π, π], and can be formulated as θi0 (d) = 6: Mutate the local worst particle as θ tpworst = θ tgbest + θ t∆ .
 
2
−π + 2πκi (d), ∀i, d, for d = 1, . . . , M LA and i = 7: ω t = ωmin + (ωmax − ωmin )
1+e−5ζ
t −1 .

1, . . . , I. Moreover, κi (d) ∈ [0, 1] is the chaotic sequence

   
8: v t+1
i = ω tv ti + c1 r1 θ ti,pbest − θ ti + c2 r2 θ tgbest − θ ti .
that can be obtained via the Logistic function [44], which 9: Obtain the updated particles as θ t+1 = θ ti + v t+1 .
i i
can be generated as κi (d + 1) = µ̃κi (d)[1 − κi (d)], where 10: while t > 0
11: EE t
if (ηi ) = (ηi ) EE t−1 then
µ̃ = 4 and κi (1) ∈ / {0, 0.25, 0.5, 0.75, 1}. During the tth
12: ti,check = ti,check + 1.
iteration, all the I populations can be updated in parallel. 13: else ti,check = 0.
According to the fitness value that each particle has 14: end if
achieved, the CSA-PSO algorithm registers both the local 15: end while
16: If ti,check > Tcheck
and global optimal particles, which are denoted as θ ti,pbest  t t
θgbest (d)−θi,pbest (d)

and θ tgbest , respectively. We propose an update scheme of 17: θit+1 (d) ∼ N 2
, θ t
gbest (d) − θ t
i,pbest (d) .
the velocities of each particle, which 18: end if
  can be formulated  as
19: If |(ηiEE )t − (ηiEE )t−1 | < 
v it+1 = ω tv ti + c1 r1 θ ti,pbest − θ ti + c2 r2 θ tgbest − θ ti , 20: ni = ni + 1
where ω t denotes the inertial weight factor, c1 and c2 21: end if
22: end while
are the acceleration factors, while we have rk ∼ U(0, 1) 23: return θ T max +1
and the corresponding EE value η EE .
gbest
for k = 1, 2. Moreover, we have c1 = c2 = 1.496
[45]. It should be noted that the velocity is limited by given by O(M LA I). Hence, we can readily show that the
[−vmax , vmax ]. We propose adaptive strategies to adjust the overall complexity of PSO is on the order of O[M LA I +
inertial weight factor, yielding (I log I + M LA I)Tmax ] = O[(I log I + M LA I)Tmax ].
  Compared to the conventional PSO, our proposed CSA-
t 2
ω = ωmin + (ωmax − ωmin ) − 1 , (19) PSO invokes chaotic sequences to generate the initial
1 + e−5ζ t RIS phase shifts, while the complexity of each step is
where ωmin and ωmax are the preset minimum and maxi- on the order of O(M LA I). Additionally, the complexity
mum values of the inertial weight factor, ζ t = (Tmax − of updating the inertial weight factor based on (19) is
t)/Tmax , where Tmax denotes the maximum number of given by O(Tmax ). The complexity of the particle reset
iterations. Consequently, the initial weight factor can be step is on the order of O(M LA T1 ), while the best case
obtained when t = 0, i.e., we have ω 0 = ωmin . The is associated with ti,check ≤ Tcheck during the iterations,
particles are updated as θ t+1 i = θ ti + v t+1
i . To further yielding T1 = 0. The worst case occurs, when the
enhance the performance, we propose a mutation scheme, particles are reset every Tcheck times, and we have T1 =
where the worst local particle θ tpworst is mutated as θ tpworst = bTmax /Tcheck c. Finally, the complexity of the proposed
θ tgbest +θθ t∆ , where θ t∆ ∼ N (00, σt2I M LA ) with σt = 2π(1− CSA-PSO is given as O[M LA I + M LA I + (I log I +
t/Tmax ). Since the conventional PSO algorithm might get M LA I +1)Tmax +M LA T1 ] = O[(I log I +M LA I)Tmax +
trapped in a local optimum, we check the fitness values M LA (T1 + 2I)]. Since the complexity is dominated by
of all the particles based on their history. If the fitness O[(I log I+M LA I)Tmax ], it can be observed that the com-
values remain unchanged for Tcheck iterations, we reset the plexity of our CSA-PSO is only slightly higher compared
corresponding particles based on θ tgbest and θ ti,pbest , yielding to the PSO algorithm.
 θt (d)−θt (d) 
θit+1 (d) ∼ N gbest 2 i,pbest , θgbest t t
(d) − θi,pbest (d) VI. S IMULATION R ESULTS
for i ∈ {1, . . . , I}. After Tmax iterations, the best individ-
In this section, we provide simulation results for charac-
ual in a set of Q is selected as the optima. The proposed
terizing the SE and EE of our proposed RPM-RIS-assisted
CSA-PSO algorithm is summarized in Algorithm 1.
CF-mMIMO system relying on our CSA-PSO phase shift
optimization algorithm. For all simulations, we consider
C. Complexity Analysis M APs and U UEs, which are independently and uni-
In the conventional PSO algorithm, the initial phase formly distributed inside a 1 × 1 km2 area, where a wrap-
shifts θ 0i are generated randomly. Moreover, the inertial around setup is emulated for avoiding the ‘desert island’
factor is fixed [45]. Furthermore, the mutation and particle edge effects. The heights of AP, RIS and UE are set as
reset steps are saved. The complexity of phase shift 12.5 m, 30 m and 1.5 m, respectively. Based on the micro-
initialization is on the order of O(M LA I). During each cellular COST 321 Walfish-Ikegami model, the path-loss
iteration, obtaining the best local and global solutions by (PL) coefficients of the NLoS component associated with
sorting the particles based on their EEs requires O(I log I) the UE-AP, UE-RIS and RIS-AP links can be formulated
operations. The complexity of calculating the velocities is as PLNLoS [dB] = −34.53 − 38 log10 (d/1m) + F [14],
 8

[40], [46], where d denotes the communication pdistance

p 1
according to the scenario, where F = δf a + 1 − δf b
represents the shadow fading having the parameter δf [1],
0.8
[40], [46]. Moreover, a ∼ N (0, δsf2 ) and b ∼ N (0, δsf2 ) are
independent random variables at the transmitter and re- 0.51
0.6 0.5
ceiver sides, respectively. Consider the UE-AP direct link 0.49
0.48
as an example. The covariance functions can be formulated 0.47
d d
− uu0 − mm0 4.6 4.8
0.4
as E{au au0 } = 2 ddc and E{bm bm0 } = 2 ddc , where
duu0 and dmm0 are the UE u-UE u0 and AP m-AP m0
distances, respectively, while ddc denotes the decorrela- 0.2

tion distance [1]. The Rician factors are formulated as

ιmu = 101.3−0.003dmu and κm = 101.3−0.003du , where 0
0 5 10 15 20
dmu and du represent the distances of the UE u-RIS m and
RIS m-AP m pairs, respectively. Here, we set δf = 0.5,
δsf = 8 and ddc = 100 m [40], [46], while we select Fig. 2. CDF of SE for the CF-mMIMO and RPM-RIS-assisted CF-
dm = 10 m, bandwidth B = 20 MHz and noise power mMIMO systems with L-MMSE combiner (G = 4 and K = 1, 2 and
4).
σ 2 = −94 dBm. The transmit power is pu = p̄u = 200
mW, ∀u, the length of each coherence block is τc = 200 1

where τp = 2 pilots are used in each coherence block,

which corresponds to a coherence bandwidth of Bc = 200 0.8
KHz and a coherence time of Tc = 1 ms. Each RIS
contains L = 64 elements and it is uniformly divided into 0.6
0.54
0.52
G = 4 blocks with dH = dV = λ/4. Moreover, we have 0.5
H 0.48
RAP,mR̃
R AP,m = R̃ RAP,m with R̃ RAP,m = dPL/2 [A A 0 J×J−P ], 0.4 3 3.2
where A = [a aφ1 , . . . , a φP ] ∈ CJ×P . As a function of φp ,
the J-dimensional steering vector is given by 0.2

1 h −j2π dAP sin(φp ) dAP

iT
a φp = √ 1, e λ , . . . , e−j2π λ (J−1) sin(φp ) ,
P 0
0 1 2 3 4 5 6 7 8
(20)
where φp = −π/2 + (p − 1)π/P for p = 1, . . . , P .
Fig. 3. CDF of SE for the CF-mMIMO and RPM-RIS-assisted CF-
Furthermore, d and PL denote the communication distance mMIMO systems with MR combiner (G = 4 and K = 1, 2 and 4).
and path loss corresponding to the scenario, respectively,
different numbers of active RIS blocks are considered. It
and we utilize P = J/2 [37]. We leverage the Gaussian
can be observed that the RPM-RIS-assisted CF-mMIMO
local scattering channel model to generate the spatial
system significantly outperforms CF-mMIMO in terms of
correlation matrix R mu . Its (p, q)th element is given by
SE, which is consistent with [13], [14]. Moreover, RPM
[14]
Z ∞ 2
introduces a slight SE loss, as shown in Fig. 2, since the
βmu − χ2 cascaded channel gain is lower for less active RIS blocks.
Rmu (p, q) = √ ej2πdH (p−q)χ̃lA ,m e 2σχ dχ,
2πσχ −∞ However, it should be emphasized that the CDF associated
(21) with K = 1, 2 and 4 are nearly identical.
where χ̃lA ,m = sin(θlA ,m + χ), and χ ∼ N (0, σχ2 ) with The CDFs of SE per UE for the CF-mMIMO and RPM-
an angular standard deviation (ASD) of σχ , and we use RIS-assisted CF-mMIMO systems are depicted in Fig. 3.
σχ = 15◦ . Unless stated otherwise, M = 20 APs having Moreover, the results of Monte-Carlo simulations based
J = 4 antennas and U = 5 UEs are employed. The on (10) as well as (11) and the SE closed-form results
RIS and AP element spacings are set as dRIS = λ/4 and based on (10) and (12) are also illustrated. From Fig.
dAP = λ/2, respectively. Each RIS includes K = 1, 2 and 3, we have the following observations. Firstly, a perfect
4 active RIS blocks, it should be noted that the K = G = 4 overlap between the SE closed-form expression and the
scenario is equivalent to the RIS-assisted CF-mMIMO Monte-Carlo simulations can be observed, which validates
system. In the following figures, the conventional CF- the accuracy of our analytical results. Secondly, the SE
mMIMO and RPM-RIS-assisted CF-mMIMO setups are performance gaps of RPM-RIM-assisted CF-mMIMO as-
denoted as “Cell Free” and “RPM-RIS-CF”, respectively. sociated with K = 1 and 2 compared to the RIS-assisted
CF-mMIMO (K = 4) are very small. Thirdly, as seen
In Fig. 2, we investigate the cumulative distribution in Fig. 2 and Fig. 3, the L-MMSE combiner is capable
function (CDF) of the SE per UE associated with the of attaining higher SE compared to the MR combiner,
conventional CF-mMIMO and our proposed RPM-RIS- since the L-MMSE combiner can minimize the MSE at
assisted CF-mMIMO systems using simulations based the cost of increased computational complexity, which is
on (10) and (11), where the L-MMSE combiner and similar to conventional CF-mMIMO systems [4]. Since the
 9

1
observed that given the number of APs, Rician fading is
capable of striking a higher average SE than in Rayleigh
0.8 fading scenarios, while the pure LoS scenario exhibits
the highest average SE. This is because the channel
0.6 attenuation of Rician fading channels is statistically lower
than the Rayleigh fading channels. Specifically, at K = 4
0.4 and M = 35, the Rician fading and pure LoS channels can
attain about 18.1% and 25.1% SE improvements compared
0.2 0.2 to that of Rayleigh fading channels, respectively.3
0.18
0.16
1.4
0 3
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5

2.5
Fig. 4. CDF of SE per UE of the RPM-RIS-assisted CF-mMIMO
systems with U = [5, 15] and K = 1, 2 and 4.

4.5 2 2.05
2
4
1.95
1.5 10
3.5

3
1
2.5 6 8 10 12 14 16

2
Fig. 6. Average SE of CF-mMIMO and RPM-RIS-assisted CF-mMIMO
1.5 systems for K = 1, 2 and 4.

1 4.5
10 15 20 25 30 35
4

Fig. 5. Average SE of CF-mMIMO and RPM-RIS-assisted CF-mMIMO 3.5

systems with G = 4 using different numbers of activated RIS blocks
and different fading scenarios. 3

Monte-Carlo simulation results are challenging to obtain, 2.5

the closed-form SE derived based on the MR combiner 2.4
will be utilized in the following figures. 2 2.2
8 8.5
In Fig. 4, the CDF of the SE per UE for the proposed 1.5
RPM-RIM-assisted CF-mMIMO is investigated along with
1
K = 1, 2 and 4 active RIS blocks, while supporting 2 4 6 8 10 12 14 16
U = [5, 15] UEs. It can be observed that higher SE can
be achieved when employing U = 5 UEs compared to
the U = 15 case. This is because lower pilot contami- Fig. 7. Average SE of CF-mMIMO and RPM-RIS-assisted CF-mMIMO
systems with M J = 80, for different numbers of activated RIS blocks.
nation and interference are achieved when fewer UEs are
deployed and the pilot length τp is fixed. Furthermore, the
SE gap introduced by RPM remains narrow, regardless of In Fig. 6, the average SE of both the CF-mMIMO
the value of U . and RPM-RIS-assisted CF-mMIMO systems versus the
Figure 5 shows the average SE per UE of the CF- number of UEs U is evaluated. It can be observed that the
mMIMO and RPM-RIM-assisted CF-mMIMO systems as average SE is reduced as the number of UEs increases.
a function of the number of APs M for different numbers Moreover, the performance loss between K = 1 and 2
of active RIS blocks K and different fading scenarios. compared to the K = 4 case becomes lower under the
As shown in Fig. 5, the SE can be enhanced by deploying scenario of higher U values, which is consistent with the
more APs, regardless of the specific fading scenarios. This findings in Fig. 4. This observation can be explained by
observation implies that both the number of spatial degrees the fact that the interference among UEs is expected to be
of freedom and the channel hardening effect can be higher when more UEs are supported. Furthermore, given
enhanced by increasing the value of M , yielding improved the U = 10 case, our RPM-RIS-assisted CF-mMIMO can
beamforming gain. Moreover, our proposed RPM-RIS-
assisted CF-mMIMO scheme operating in Rician fading 3 We emphasize that the philosophy of our proposed architecture can be
channels is capable of achieving up to 112.5% higher SE readily extended to other fading channel models, such as Nakagami-m,
than the conventional CF-mMIMO. Furthermore, it can be κ − µ, etc [47].
 10

3
attain about 53.9% SE gain compared to the conventional
CF-mMIMO for any specific K.
2.5
In Fig. 7, the average SE of both the CF-mMIMO and
RPM-RIS-assisted CF-mMIMO systems are investigated 2
as a function of the number of antennas per AP, while
the total number of antennas is fixed as M J = 80. It can 1.5

be seen that although the proposed RPM-RIS-assisted CF-

1
mMIMO system still outperforms CF-mMIMO in terms of
its average SE per UE, there is a significant performance
0.5
erosion as J increases. This trend can be explained by
the following arguments. Firstly, since the value of M J 0
10 15 20 25 30 35
remains unchanged, a higher J implies that fewer APs are
deployed, yielding higher path losses. Furthermore, lower
macro-diversity can be obtained for a higher value of J. Fig. 8. Total EE of the RPM-RIS-assisted CF-mMIMO systems with
In addition, compared to the array gain, it can be readily K = 1, 2, and 4, when P (b) = 10 dBm and P (b) = 25 dBm.
shown that macro-diversity is the dominant factor for this
3
scenario.
Next, the EE of our proposed RPM-RIS-assisted CF-
2.5
mMIMO system is evaluated. The traffic-dependent power
coefficients are set as %AP BH
m = %m = 0.25 (W/Gbps),
2
∀m. The fixed AP power consumption and the traffic-
independent power consumed by each antenna are selected
AP,fix AP,a 1.5
as Pm = 6 W and Pm = 0.15 W, ∀m, respectively.
The power amplifier efficiency is αUE = 0.4. The fixed
BH,fix
power consumed by the backhaul links is Pm = 0.8 1

W, ∀m. Furthermore, the fixed power associated with the

UEs and RIS elements is chosen as PuUE,fix = 10 dBm and 0.5
6 8 10 12 14 16
P (b) = 10/25 dBm, ∀u, respectively.
We compare the total EE of our proposed RPM-RIS-
assisted CF-mMIMO system relying on random phase Fig. 9. Total EE of the proposed RPM-RIS-assisted CF-mMIMO systems
shifts, K = [1, 2, 4] and P (b) = [10, 25] dBm in Fig. with K = 1, 2, and 4, when P (b) = 10 dBm and P (b) = 25 dBm.
8, when different numbers of APs are employed. From as possible, while sustaining a satisfactory SE to obtain
Fig. 8, we can make the following observations. Firstly, better EE. Furthermore, the more UEs are supported, the
given a value of K, higher EE can be attained by selecting better the total EE becomes. This is because the total SE
P (b) = 10 dBm compared to the P (b) = 25 dBm case. can be enhanced while using more UEs, while the power
This is because the power consumed by the RISs is higher dissipated by the UEs is relatively low compared to the
when using P (b) = 25 dBm. Secondly, given a value of other energy components of P tot . Finally, according to Fig.
P (b), the settings of using less active RIS blocks, i.e., 2 - Fig. 9, it can be observed that our proposed RPM-
K = 1 results in better EE performance. This is because RIS-assisted CF-mMIMO system is capable of striking an
the total energy dissipation can be reduced by activating attractive SE vs. EE trade-off.
fewer RIS blocks, while the SE of the K = 1 and K = 2 In Fig. 10, we characterize the total EE of both our
cases is similar to the K = 4 case, as shown in the proposed CSA-PSO algorithm and of the conventional
figures of the SE part. In particular, compared to the RIS- PSO algorithm. In particular, M = 8 APs and U = 4
assisted CF-mMIMO (K = 4) system having M = 25, UEs are considered. The maximum number of iterations
the proposed RPM-RIS-assisted CF-mMIMO system with is Tmax = 100, the value range of the velocity is given by
K = 1 can improve the EE by 22.6% and 120% associated [−4, 4], and we set ω = 0.7298 [45] in the conventional
with P (b) = 10 dBm and P (b) = 25 dBm, respectively. PSO algorithm. For the proposed CSA-PSO, we invoke
Finally, an EE reduction can be observed as more APs are Tcheck = 2, ωmax = 0.9, ωmax = 0.4 and Ni = 10.
deployed, since the power consumption level increases and It is clear that both CSA-PSO and PSO can converge
dominates the EE. within 40 iterations. Moreover, it can be observed that our
In Fig. 9, the total EE associated random phase shifts is proposed CSA-PSO is capable of significantly improving
shown versus the number of UEs U for different numbers the EE of conventional PSO. More explicitly, under the
of active RIS blocks and different values of P (b). It can be condition of K = 1, our CSA-PSO can obtain about
observed that a lower P (b) can significantly enhance the 25% and 26.7% EE gain over PSO, for P (b) = 10 dBm
EE. Specifically, given U = 16 and K = 2, the setup with and P (b) = 25 dBm, respectively. This is because of
P (b) = 10 dBm is capable of obtaining about twice higher the following reasons. Firstly, we use a chaotic sequence
total EE compared to the P (b) = 10 dBm scenario. This to improve the particle’s initial diversity. Secondly, an
observation implies that we should reduce P (b) as much adaptive inertial weight factor is proposed to enhance the
 11

6.5 7

6
6
5.5

5
5

4.5
4

4
3
3.5

3 2
0 20 40 60 80 100 0.25 0.5 0.75 1 1.25 1.5

(a)
Fig. 11. Total EE of the proposed K = 2 RPM-RIS-assisted CF-
mMIMO systems with P (b) = 10 dBm using U = [4, 8] and different
4 RIS phase shift design schemes.
0.9
3.5

0.8
3

0.7
2.5

0.6
2

0.5
1.5

0.4
1
0 20 40 60 80 100
0.3
(b)

Fig. 10. Total EE of the conventional PSO and the proposed CSA-PSO
RIS phase shift design schemes operating at (a) P (b) = 10 dBm and Fig. 12. Average running time versus the number of active RIS blocks
(b) P (b) = 25 dBm, while K = 1, 2 and 4. operating at K = 1, 2, 4 using PSO and CSA-PSO-based phase shift
optimization schemes with P (b) = [10, 25] dBm.
algorithm’s search capability. Finally, the introduction of
14900HX processor. It can be observed that given the
particle mutation and reset steps can help the particles
values of K and P (b), PSO is more computationally
escape from local optima.
efficient than CSA-PSO, which is consistent with our
The total EEs are compared in Fig. 11, as a function complexity analysis results in Section V-C. However, the
of the traffic-dependent power coefficient %AP BH
m = %m = %, CSA-PSO method can still converge within the coherence
∀m of our proposed CSA-PSO, and of the conventional time of Tc = 1 ms, regardless of the K and P (b) settings.
PSO phase shift design and random phase shift schemes, Hence, based on Fig. 10-Fig. 12, we can readily state that
where U = [4, 8] UEs are supported. From Fig. 11, we our proposed CSA-PSO strikes an attractive complexity
have the following observations. First, compared to PSO vs. EE trade-off.
and random phase shift, the proposed CSA-PSO algorithm
is capable of attaining better EE performance. Specifically,
given U = 4 and % = 1, the CSA-PSO can achieve about VII. S UMMARY AND C ONCLUSIONS
1.3× and 2× higher EE compared to its PSO and random In this paper, an RPM-RIS-assisted CF-mMIMO system
phase shift counterparts. Secondly, the EE decreases as % was conceived, where only a part of RIS elements have
increases, since higher energy is consumed. Thirdly, we been active, and extra information has been mapped onto
can see that a higher EE can be obtained by supporting the RIS ON/OFF states. Based on the uplink LSFD cooper-
more UEs, consistent with our observations in Fig. 9. ation scheme, both the MR and L-MMSE combiners were
Finally, the widest EE performance gap between U = 4 considered. Then, a closed-form expression was derived
and U = 8 can be obtained by leveraging the CSA-PSO, for the uplink SE by using the MR combiner, where
followed by the conventional PSO and random phase shift both the channel estimation errors and pilot contamination
schemes. were considered. Our simulation results demonstrated that
In Fig. 12, we compare the complexity of our proposed the RPM-RIS-assisted CF-mMIMO system is capable of
CSA-PSO algorithm and the conventional PSO in terms attaining nearly identical SE compared to the conventional
of the average running time under different numbers of RIS-assisted CF-mMIMO. Furthermore, the EE was char-
active blocks with P (b) = [10, 25] dBm and the same acterized. For EE maximization, a CSA-PSO algorithm
parameters as in Fig. 10. We leverage an Inter Core i9- was proposed for RIS phase shift design. Moreover, the
 12

complexity of both CSA-PSO and classic PSO was ana- Now we calculate each term of (11). Since the chan-
lyzed. Our simulations illustrated that RPM-RIS-assisted nel estimate ĥhmu and the corresponding error h̃ hmu are
CF-mMIMO attains a substantially better EE than RIS- uncorrelated, for the numerator term we have E{g g uu } =
h
H H
iT 
assisted CF-mMIMO using a random phase shift. Finally,
E ĥ h1uĥ
h1u , . . . , ĥ
hM uĥhM u . Consequently, let ỹy pmu =
our proposed CSA-PSO scheme is capable of offering
significantly higher EE compared to its conventional coun- y pmu p
n H− ȳy muo
, hence the m-th term of E{gg uu }, i.e.,
terparts. E ĥ hmuĥ
hmu can be derived based on (5) as

A PPENDIX A 2
+ pu tr E{(ỹy pmu )H (Ψ
Ψ−1 Rhmu )R RhmuΨ −1 y pmu )}


hmu
h̄ mu )(R mu (ỹ
U SEFUL L EMMAS
(a) 2
RhmuΨ −1 y pmu )(ỹy pmu )H Ψ −1 h


Lemma 1 [39, Lemma 7]: Assume that the elements = h̄hmu + pu tr E{R mu (ỹ muR mu }
of the matrix Z ∈ Cm×n are i.i.d. RVs with zero mean (b) 2
hmu
= h̄ RhmuΨ −1
+ pu τp tr(R −1 h
muΨ muΨ muR mu )
and a variance of ζz with m, n ≥ 1, and A ∈ Cn×n
2
Z AZ H } =
is a deterministic matrix. Then, we have E{Z hmu
= h̄ Γmu ) = ξ¯mu ,
+ pu τp tr(Γ (26)
A}II m .
ζz tr{A
Lemma 2 [13, Lemma 4]: We consider an M × 1 where (a) invokes tr(X X Y ) = tr(Y
Y X ) along with the
RV vector a ∼ CN (00, R a ) with R a ∈ CM ×M , and appropriate dimensions of matrices X and Y , and (b) can
a deterministic matrix W ∈ CM ×M . Then, we have be obtained by virtue of (25). Hence, we have E{gg uu } =
E |a aH W a |2 = | tr(R RaW R aW H ).
RaW )|2 + tr(R [ξ¯1u , . . . , ξ¯M u ]T and the numerator term can be readily
obtained.
A PPENDIX B
Similarly, we consider the noise term of (11):
Based on (2), it can be readily shown that n n o n oo
n h h h h
o cH H
u V uc u = c u diag E ||ĥ h1u || , . . . , E ||ĥ
hM u || c u
R hmu = E R̃
Rmu,1 + R̃ Rmu,2 + R̃ Rmu,4 , (22)
Rmu,3 + R̃
= tr C H ¯ ¯


Cu .
u diag{ξ1u , . . . , ξM u }C (27)
h H
where we have R̃ GmΦ mz̃z muz̃z H
Rmu,1 = bmu κmḠ HG
muΦ mḠ m, For the second term of the denominator
h H GH h
Rmu,2
R̃ = bmu ιmuG̃ GmΦ mz̄z muz̄z H
muΦ mG̃ Rmu,3 =
m , R̃ cH
u g
E{g uu g
}E{g H c
uu }c u , it can be formulated as
H h
GmΦ mz̃znmuz̃z H
bmuG̃ ΦH
muo Rmu,4 = f muf H
Gm and R̃
mG̃ mu . Then,
( M
h i
)
h
X 2
we have E R̃Rmu,1 = bmu κmḠ GmΦ mR̃ H
RmuΦ mḠ
H
Gm and cmu tr(pu τpΓ mu ) + h̄ hmu
n h o m=1
Rmu,4 = R mu . Moreover, we have
E R̃ ( M h i
)
X 2
n h o × c∗nu tr(pu τpΓ nu ) + h̄
hnu . (28)
H H
n o
Π mu = E R̃ GmΦ mz̄z muz̄z H
Rmu,2 = bmu ιmu E G̃ Gm
muΦ mG̃ n=1

When m = n, the first component of (28) can be attained

Πmu .
= bmu ιmuΠ̃ (23)
as
Let B mu , Φ mz̄z munz̄z H H 0
muΦ m , then the (j, j )thocomponent M h
X 2 2
of Π̃Πmu , i.e., E G̃ Gm (j, :)B Gm (j , :)H
B muG̃ 0
can be |cmu |2 p2u τp2 | tr(Γ
Γmu )|2 + h̄
hmu hmu
h̄
 n o m=1
formulated as tr E B muG̃ Gm (j 0 , :)H G̃
Gm (j, :) = 2
i
  + 2pu τp tr(Γ Γmu ) h̄ hmu . (29)
tr B muR̃Rm (jLA − L + 1 : jLA , j 0 LA − LA + 1 : j 0 LA ) .
n h o In the case of m 6= n, the second component of (28) can
Moreover, we have Ξ mu = E R̃ Rmu,3 = be formulated as
H
n o
bmu E G̃GmΦ mR̃ RmuΦ H mG̃Gm . Let B̃ B mu , Φ mR̃ RmuΦ H m, XM X M h  H 
∗ 2 2
0 c c p τ tr Γ
(Γ ) Γ
tr(Γ ) + tr h
h̄ h
h̄
 the (j, j )th element of Ξ mu can be expressed as 
then mu nu u p mu nu mu mu
m=1 n=1
Rm (jLA − LA + 1 : jLA , j 0 LA − LA + 1 : j 0 LA ) .
B muR̃
tr B̃  H   H 
Consequently, the proof can be completed by combining × tr h̄ hnuh̄hnu + pu τp tr h̄ hmuh̄ hmu tr (Γ Γnu )
the above results. A PPENDIX C  H
 i
+pu τp tr h̄ hnuh̄hnu tr (Γ Γmu ) . (30)
We first offer several useful results that are utilized in
the proof. Given the set Pu and k ∈ Pu \ {u}, we have Therefore, c H g uu }E{gg H cu is calculated by
u E{g uu }c
p p p p
y mu = y mk , ȳy mu = ȳy mk , Ψ mu = Ψ mk . (24) combining (29) and (30). Furthermore, the interference
2
 
PM H
term E m=1 cmuĥ hmuh mk can be expressed as
Based on Proposition 1, it can be observed that R hmu are
H H
Hermitian matrices. Hence, we have R hmu = (R Rhmu )H and PM PM
 
H
 
∗ h h
c
n=1 mu nu c E ĥ h
mu mk ĥ h
nu nk ,
Ψ hmu = (ΨΨhmu )H , ∀m, u. Furthermore, it can be readily m=1

shown based on (5) that where all the combinations of the AP and UE
indices have to be considered. Since the estimated
y pmu − ȳy pmu ∼ CN (00, τpΨ mu ). (25) channels of different APs are independent, we have
 13


H
H  H  H
hmuh mk
E ĥ hnuh nk
ĥ = 0 when m 6= n E{|b|2 } = E{|c|2 } = E{ad} = pu τph̄ hmuh̄hmu tr (Γ
Γmu ).
Next, E{|d|2 } can be reformulated as
and k ∈/ Pu . Under the condition of m 6= n and
 H H  H  
2

k ∈ Pu \ {u}, we have E ĥ hmuh mk hnuh nk
ĥ = E{|d|2 } = p2u τp2 E w H (Ψ
Ψ−1/2
mu ) H H
R mu R mu Ψ −1/2
mu w
H
 
H H
 n o
(a)
E ĥhmuĥ
hmk E ĥhnuĥ
hnk , and the component = p2u τp2 tr(Γ


Γmu ) tr(Γ C muΓ mu , (34)
Γmu ) + pu τp tr C̄
H
 
H

E ĥ hmk
hmuĥ can be formulated as X Y ) = tr(Y
where (a) invokes (6) and tr(X YX ). Hence, by
2

H
hmuĥ
putting together all the above results, E ĥ hmu
nh H √ i
E hmu + pu (ỹy pmu )H (Ψ
h̄ Ψ−1mu )H
R
(R h
mu ) H
can be expressed as
 √ h −1 p  H H H
× h̄ hmk + pkR mkΨ mk (ỹy mk ) h̄ hmuh̄
hmuh̄ hmu + 3pu τph̄
hmuh̄ hmu tr (Γ
hmuh̄ Γmu )
(a) H
= p̃uk E (ỹy pmk )H (Ψ Ψ−1 H
Rhmu )H R hmkΨ −1 y pmk
p
Rhmu − Λ mu )) + p2u τp2 tr(Γ

mu ) (R mk ỹ hmuh̄
+ tr(h̄ hmu (R Γmu ) tr(Γ
Γmu )
(b) p
Rhmu − Λ mu )Γ


Rmu ) R mkΨ −1 y pmk (ỹy pmk )H + pu τp tr (R Γmu . (35)
 −1 H h H h

= p̃uk tr E (Ψ Ψmu ) (R mk ỹ

RhmkΨ −1 h
p
= p̃uk τp tr(R muR mu ), (31) By combining the above results associated with m = n
n H o and k = u, we can arrive at
where p̃uk = pu pk , (a) invokes E h̄ hmuh̄ hmk = 0 
H 2

2
H H
 p H −1 H h H
and E (ỹy mu ) (Ψ Ψmu ) (R Rmu ) h̄ hmk = 0, (b) uses E ĥhmuh mu hmuh̄
= h̄ hmu + 3pu τph̄ hmu tr (Γ
hmuh̄ Γmu )
(24) and tr(X X Y ) = tr(Y Y X ). Hence, it can be 2 H
 H H  n H o + p2u τp2 |tr(Γ
Γmu )| + tr(h̄ hmuR hmu )
hmuh̄
readily shown that E ĥ hmuĥhmk E ĥ hnuĥhnk =
RhmuΓmu ).
+ pu τp tr(R (36)
pu pk τp2 tr R hmkΨ −1 h
tr R hnuΨ −1 h



muR mu  nu R nk . When m 6=
 H H  H  Furthermore, we can compute the expectation term as
n and k = u, we express E ĥ hmuh mk hnuh nk
ĥ =
 n
H H
o H
n H o n H o hmuĥ
tr E h mkh mkĥ hmu hmkh̄
= pu τph̄ hmk tr(Γ
Γmu ) +
2
E ĥ hmuĥ hmk E ĥ hnuĥ
hnk since the channel estimates at H
hmuh̄
h̄ hmk + pu τp tr(R RhmkΓ mu ) + tr(R Rhmk )h̄
H
hmuh̄hmu ,
different APs are independent, while the channel esti- when m = n and k ∈ / Pu . In addition, for
mation error and the estimated
n H channel are uncorrelated. m = n and  k ∈ Pu \ {u},  we  can obtain
o H

In addition, we have E ĥ hmuĥhmk = tr h̄ hmuh̄ hmu + H 2 H 2 H 2

pu τp tr(Γ Γmu ). Therefore, we have E ĥ hmuh mk = E ĥ hmuĥ hmk + E ĥ hmuh̃ hmk
2
 
n H o n H o  H   H  H
E ĥ hmuĥ hmk E ĥ hnk = tr h̄
hnuĥ hmuh̄hmu tr h̄ hnuh̄ hnu where we have E ĥ hmuh̃
hmk ΓmuΛ mk ) +
= pu τp tr (Γ
 H   H 
+ p2u τp2 tr(Γ
Γmu ) tr(Γ
Γnu ) + pk τp tr h̄ hmuh̄hmu tr(Γ Γnu ) tr h̄hmuh̄
hmuΛ mk . Similar to the calculation
 H  process
n H of (33), and bearing
o in mind that
+ pk τp tr h̄hnuh̄
hnu tr(Γ Γmu ). (32) E h̄ hmkR hmuΨ −1 y
(y
mu mu
p
− y
ȳ p
) = 0, we can obtain
 mu 2

H
In the case of m = n and k = u, the the final result of E ĥ hmuĥ hmk as in (37), which
interference includes the following two terms

H 2
  H
  is shown at the top of this page, where we have a =
√ −1/2 √ H
E ĥ hmuh̃hmu hmuh̄
= tr pu τpΓ mu + h̄ hmu Λ mu pu τpw H Ψ mu R hmuh̄
hmk , b = pu τph̄ hmuR hmkΨ −1/2
mk w
w H Ψ −1/2 R R Ψ −1/2
w
2 and c =
 
H mu mu mk mk with
and E ĥ hmuĥhmu can be formulated as (33) w ∼ CN (00, I J ). By combining the above results
at the top of next page, where (a) is obtained when m = n and k ∈ Pu \ {u}, we can derive
by using (24) with w ∼ CN (00, I J ). By denoting 
H 2

2  
H √ H H
hmuh̄
a = h̄ Ψ−1/2
hmu , b = pu τpw H (Ψ H Rh
mu ) (R
Hh
mu ) h̄ mu , c = E ĥ hmuh mk hmuh̄
= h̄ hmk + pu τp tr h̄ hmkh̄hmkΓ mu
√ H √ 2
h −1/2 h −1/2
hmuR muΨ mu w and d = pu τpR muΨ mu w ,
pu τph̄  H

hmuR hmk + pu τp tr Γ̄


hmuh̄
+ tr h̄ Γmuk
2
 
H
(33) can be rewritten as E ĥ hmuĥ
hmu =
2
+ pk pu τp2 tr R mkΨ −1 h
muR mu , (38)
E{|a|2 } + E{|b|2 } + E{|c|2 } + E{|d|2 } + 2E{ad}.
Γ h
Consequently, we have E{|a|2 } = h̄
H
hmuh̄
hmuh̄
H
hmuh̄
hmu . where Γ̄ muk = Γ muR mk . By collecting all the above
2

2 PM H
Then, n we can also obtain E{|b| } o = cases, E hmuh mk
m=1 cmuĥ can be derived as (39),
H (a)
Ψ−1/2
pu τp E w H (Ψ H Rh
mu ) (R
Hh
mu ) h̄ muh̄
−1/2
hmuR hmuΨ mu w = which is provided at the top of this page. Finally, we
H substitute all the above-derived terms into (11), and then
hmuh̄
pu τph̄ hmu Γmu ), where (a) follows Lemma 1, and
tr (Γ
XY ) =
that the LoS components are determinstic and tr(X (12) can be obtained.
Y X ). Moreover, it can be readily shown that
tr(Y
 14

2
   
H √ p
H  √ p
2
hmu + puR hmuΨ −1 h −1

E hmuĥ
ĥ hmu =E h̄ y
(ỹ
mu mu ) h
h̄ mu + p R Ψ
u mu mu mu y
(ỹ )
( )
(a)
h √ iH h √ i2
h −1/2 h −1/2
= E hmu + pu τpR muΨ mu w
h̄ hmu + pu τpR muΨ mu w
h̄
( )
H √ √ H √ 2 2
=E hmuh̄
h̄ hmu + pu τpw H Ψ −1/2 h
mu R muh̄ hmuR hmuΨ −1/2
hmu + pu τph̄ mu w + pu τpR hmuΨ −1/2
mu w , (33)

2
 
H H 2 n o n o n o
2 2 2
E hmuĥ
ĥ hmk hmuh̄
= h̄ hmk + E |a| + E |b| + pu pk τp2 E |c|

H 2  H
  H

= h̄ hmk
hmuh̄ hmkh̄
+ pu τp tr h̄ hmkΓ mu + tr h̄ hmuh̄ C mk
hmuC̄

2
+ pk pu τp2 tr R mkΨ −1 h


muR mu C mk ,
+ pu τp tr Γ muC̄ (37)

M 
2

X H H H
2
hmuR hmkh̄ Γ muR hmk


|cmu | hmuh̄
h̄ hmk + pu τph̄ hmk
hmkΓ muh̄ + h̄ hmu + pu τp tr
m=1
 M M
cmu c∗nu pu pk τp2 tr(R RhmkΨ −1 h RhnkΨ −1 h
P P
muR mu ) tr(R nu R nu ), k ∈ Pu \ {u},



m=1 n=1


 M M
 h    
H H
cmu c∗nu tr h̄
P P
+ hmuh̄ hmu tr h̄ hnuh̄hnu + p2u τp2 tr(Γ Γmu ) tr(Γ
Γnu )
 m=1 n=1  H   H  i




 +pk τp tr h̄hmuh̄ hmu tr(Γ Γnu ) + pk τp tr h̄
hnuh̄hnu tr(Γ Γmu ) , k = u,

0, k∈ / Pu .

 M
P 2 2 −1 h

2


 |c mu | pk pu τ p tr R mk Ψ mu R mu , k ∈ Pu \ {u},
m=1

M
+ P (39)
h H
i
2 2 2 2
 |c mu | 2p u τp h
h̄ mu h
h̄ mu tr Γ
(Γ mu ) + p u τp Γ
|tr(Γ mu )| , k = u,
m=1



0, k∈/ Pu .

R EFERENCES [9] J. An, C. Xu, L. Gan, and L. Hanzo, “Low-complexity channel es-
timation and passive beamforming for RIS-assisted MIMO systems
[1] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. relying on discrete phase shifts,” IEEE Trans. Commun., vol. 70,
Marzetta, “Cell-free massive MIMO versus small cells,” IEEE no. 2, pp. 1245–1260, Feb. 2022.
Trans. Wireless Commun., vol. 16, no. 3, pp. 1834–1850, Mar. 2017. [10] Q. Li, M. El-Hajjar, I. Hemadeh, A. Shojaeifard, A. A. Mourad,
[2] H. Q. Ngo, L.-N. Tran, T. Q. Duong, M. Matthaiou, and E. G. Lars- B. Clerckx, and L. Hanzo, “Reconfigurable intelligent surfaces
son, “On the total energy efficiency of cell-free massive MIMO,” relying on non-diagonal phase shift matrices,” IEEE Trans. Veh.
IEEE Trans. Green Commun. Netw, vol. 2, no. 1, pp. 25–39, Mar. Technol., vol. 71, no. 6, pp. 6367–6383, Jun. 2022.
2018. [11] Q. Wu and R. Zhang, “Towards smart and reconfigurable environ-
[3] J. Zhang, E. Björnson, M. Matthaiou, D. W. K. Ng, H. Yang, and ment: Intelligent reflecting surface aided wireless network,” IEEE
D. J. Love, “Prospective multiple antenna technologies for beyond Commun. Mag., vol. 58, no. 1, pp. 106–112, Jan. 2020.
5G,” IEEE J. Sel. Areas Commun., vol. 38, no. 8, pp. 1637–1660, [12] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and
Aug. 2020. C. Yuen, “Reconfigurable intelligent surfaces for energy efficiency
[4] E. Björnson and L. Sanguinetti, “Making cell-free massive MIMO in wireless communication,” IEEE Trans. Wireless Commun.,
competitive with MMSE processing and centralized implementa- vol. 18, no. 8, pp. 4157–4170, Aug. 2019.
tion,” IEEE Trans. Wireless Commun., vol. 19, no. 1, pp. 77–90, [13] T. Van Chien, H. Q. Ngo, S. Chatzinotas, M. Di Renzo, and
Jan. 2019. B. Ottersten, “Reconfigurable intelligent surface-assisted cell-free
[5] H. Q. Ngo, G. Interdonato, E. G. Larsson, G. Caire, and J. G. massive MIMO systems over spatially-correlated channels,” IEEE
Andrews, “Ultradense cell-free massive MIMO for 6G: Technical Trans. Wireless Commun., vol. 21, no. 7, pp. 5106–5128, Jul. 2021.
overview and open questions,” Proceedings of the IEEE, pp. 1–27, [14] E. Shi, J. Zhang, D. W. K. Ng, and B. Ai, “Uplink performance
2024. of RIS-aided cell-free massive MIMO system with electromagnetic
[6] A. Lozano, R. W. Heath Jr., and J. G. Andrews, “Fundamental interference,” IEEE J. Sel. Areas Commun., vol. 41, no. 8, pp. 2431–
limits of cooperation,” IEEE Trans. Inf. Theory, vol. 59, no. 9, pp. 2445, Aug. 2023.
5213–5226, Sep. 2013. [15] J. Dai, J. Ge, K. Zhi, C. Pan, Z. Zhang, J. Wang, and X. You,
[7] L. Du, L. Li, H. Q. Ngo, T. C. Mai, and M. Matthaiou, “Cell-free “Two-timescale transmission design for RIS-aided cell-free massive
massive MIMO: Joint maximum-ratio and zero-forcing precoder MIMO systems,” IEEE Trans. Wireless Commun., vol. 23, no. 6,
with power control,” IEEE Trans. Commun., vol. 69, no. 6, pp. pp. 6498–6517, Jun. 2024.
3741–3756, Jun. 2021. [16] J. Yao, J. Xu, W. Xu, D. W. K. Ng, C. Yuen, and X. You,
[8] Y. Zhang, H. Zhao, W. Xia, W. Xu, C. Tang, and H. Zhu, “Robust beamforming design for RIS-aided cell-free systems with
“How much does reconfigurable intelligent surface improve cell- CSI uncertainties and capacity-limited backhaul,” IEEE Trans.
free massive MIMO uplink with hardware impairments?” IEEE Commun., vol. 71, no. 8, pp. 4636–4649, Aug. 2023.
Trans. Commun., vol. 71, no. 11, pp. 6677–6694, Nov. 2023. [17] E. Shi, J. Zhang, S. Chen, J. Zheng, Y. Zhang, D. W. K. Ng, and
 15

B. Ai, “Wireless energy transfer in RIS-aided cell-free massive IEEE J. Sel. Areas Commun., vol. 31, no. 2, pp. 160–171, Feb.
MIMO systems: Opportunities and challenges,” IEEE Commun. 2013.
Mag., vol. 60, no. 3, pp. 26–32, Mar. 2022. [38] E. Björnson and L. Sanguinetti, “Rayleigh fading modeling and
[18] E. Shi, J. Zhang, J. Zheng, B. Ai, and D. W. K. Ng, “RIS-aided channel hardening for reconfigurable intelligent surfaces,” IEEE
cell-free massive MIMO systems with channel aging,” IEEE Trans. Wireless Commun. Lett., vol. 10, no. 4, pp. 830–834, Apr. 2020.
Veh. Technol., pp. 1–16, 2024. [39] K. Zhi, C. Pan, H. Ren, K. Wang, M. Elkashlan, M. D. Renzo,
[19] D. L. Galappaththige, D. Kudathanthirige, and G. Amarasuriya, R. Schober, H. V. Poor, J. Wang, and L. Hanzo, “Two-timescale
“Performance analysis of IRS-assisted cell-free communication,” design for reconfigurable intelligent surface-aided massive MIMO
in Proc. IEEE GLOBECOM, Dec. 2021, pp. 1–6. systems with imperfect CSI,” IEEE Trans. Inform. Theory, vol. 69,
[20] B. Al-Nahhas, M. Obeed, A. Chaaban, and M. J. Hossain, “RIS- no. 5, pp. 3001–3033, 2023.
aided cell-free massive MIMO: Performance analysis and compet- [40] Ö. Özdogan, E. Björnson, and J. Zhang, “Performance of cell-free
itiveness,” in Proc. IEEE ICC, Jun. 2021, pp. 1–6. massive MIMO with Rician fading and phase shifts,” IEEE Trans.
[21] Z. Zhang and L. Dai, “A joint precoding framework for wideband Wireless Commun., vol. 18, no. 11, pp. 5299–5315, Nov. 2019.
reconfigurable intelligent surface-aided cell-free network,” IEEE [41] T. M. Cover, Elements of Information Theory. John Wiley & Sons,
Trans. Signal Process., vol. 69, pp. 4085–4101, Jun. 2021. 1999.
[22] B. Li, Y. Hu, Z. Dong, E. Panayirci, H. Jiang, and Q. Wu, “Energy- [42] G. Femenias, J. Garcı́a-Morales, and F. Riera-Palou, “SWIPT-
efficient design for reconfigurable intelligent surface aided cell-free enhanced cell-free massive MIMO networks,” IEEE Trans. Com-
ultra dense hetnets,” IEEE Trans. Veh. Technol., vol. 73, no. 3, pp. mun., vol. 69, no. 8, pp. 5593–5607, Aug. 2021.
3767–3785, Mar. 2024. [43] J. Wang et al., “Reconfigurable intelligent surface: Power consump-
tion modeling and practical measurement validation,” IEEE Trans.
[23] S.-N. Jin, D.-W. Yue, Y.-L. Chen, and Q. Hu, “Energy efficiency
Commun., pp. 1–13, 2024.
maximization in IRS-aided cell-free massive MIMO system,” IEEE
[44] R. Caponetto, L. Fortuna, S. Fazzino, and M. G. Xibilia, “Chaotic
Wireless Commun. Lett., vol. 12, no. 10, pp. 1652–1656, Oct. 2023.
sequences to improve the performance of evolutionary algorithms,”
[24] Q. N. Le, V.-D. Nguyen, O. A. Dobre, and R. Zhao, “Energy IEEE Trans. Evol. Comput., vol. 7, no. 3, pp. 289–304, Jun. 2003.
efficiency maximization in RIS-aided cell-free network with limited [45] M. Clerc and J. Kennedy, “The particle swarm: Explosion, stability,
backhaul,” IEEE Commun. Lett., vol. 25, no. 6, pp. 1974–1978, Jun. and convergence in a multidimensional complex space,” IEEE
2021. Trans. Evol. Comput., vol. 6, no. 1, pp. 58–73, Feb. 2002.
[25] W. Lyu, Y. Xiu, S. Yang, C. Yuen, and Z. Zhang, “Energy-efficient [46] Z. Wang, J. Zhang, B. Ai, C. Yuen, and M. Debbah, “Uplink per-
cell-free network assisted by hybrid RISs,” IEEE Wireless Commun. formance of cell-free massive MIMO with multi-antenna users over
Lett., vol. 12, no. 4, pp. 718–722, Apr. 2023. jointly-correlated rayleigh fading channels,” IEEE Trans. Wireless
[26] R. Karasik, O. Simeone, M. Di Renzo, and S. Shamai (Shitz), Commun., vol. 21, no. 9, pp. 7391–7406, Sep. 2022.
“Adaptive coding and channel shaping through reconfigurable in- [47] J. Zhang, M. Matthaiou, G. K. Karagiannidis, and L. Dai, “On the
telligent surfaces: An information-theoretic analysis,” IEEE Trans. multivariate gamma–gamma distribution with arbitrary correlation
Commun., vol. 69, no. 11, pp. 7320–7334, Nov. 2021. and applications in wireless communications,” IEEE Trans. Veh.
[27] W. Yan, X. Yuan, and X. Kuai, “Passive beamforming and informa- Technol., vol. 65, no. 5, pp. 3834–3840, May 2015.
tion transfer via large intelligent surface,” IEEE Wireless Commun.
Lett., vol. 9, no. 4, pp. 533–537, Apr. 2020.
[28] S. Guo, S. Lv, H. Zhang, J. Ye, and P. Zhang, “Reflecting modula-
tion,” IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2548–2561,
Nov. 2020.
[29] S. Lin, B. Zheng, G. C. Alexandropoulos, M. Wen, M. D. Renzo,
and F. Chen, “Reconfigurable intelligent surfaces with reflection
pattern modulation: Beamforming design and performance analy-
sis,” IEEE Trans. Wireless Commun., vol. 20, no. 2, pp. 741–754,
Feb. 2021.
[30] S. Lin, F. Chen, M. Wen, Y. Feng, and M. Di Renzo, “Recon-
figurable intelligent surface-aided quadrature reflection modulation
for simultaneous passive beamforming and information transfer,”
IEEE Trans. Wireless Commun., vol. 21, no. 3, pp. 1469–1481,
Mar. 2022.
[31] J. Yao, J. Xu, W. Xu, C. Yuen, and X. You, “Superimposed RIS-
phase modulation for MIMO communications: A novel paradigm
of information transfer,” IEEE Trans. Wireless Commun., vol. 23,
no. 4, pp. 2978–2993, Apr. 2024.
[32] Z. Yigit, E. Basar, M. Wen, and I. Altunbas, “Hybrid reflection
modulation,” IEEE Trans. Wireless Commun., vol. 22, no. 6, pp.
4106–4116, Jun. 2023.
[33] B. Di, H. Zhang, L. Song, Y. Li, Z. Han, and H. V. Poor, “Hybrid
beamforming for reconfigurable intelligent surface based multi-
user communications: Achievable rates with limited discrete phase
shifts,” IEEE J. Sel. Areas Commun., vol. 38, no. 8, pp. 1809–1822,
Aug. 2020.
[34] E. Shi, J. Zhang, R. He, H. Jiao, Z. Wang, B. Ai, and D. W. K. Ng, Zeping Sui (Member, IEEE) received the B.
“Spatially correlated reconfigurable intelligent surfaces-aided cell- Eng. degree in electronic information engineer-
free massive MIMO systems,” IEEE Trans. Veh. Technol., vol. 71, ing from Northwestern Polytechnical Univer-
no. 8, pp. 9073–9077, Aug. 2022. sity, Xi’ an, China, in 2017, and the Ph.D.
[35] Z. Sui, S. Yan, H. Zhang, L.-L. Yang, and L. Hanzo, “Approximate degree in signal and information processing
message passing algorithms for low complexity OFDM-IM detec- from the University of Chinese Academy of
tion,” IEEE Trans. Veh. Technol., vol. 70, no. 9, pp. 9607–9612, Sciences, Beijing, China, in 2023. From March
Sep. 2021. 2022 to Feb 2023, He was a research assistant
[36] Z. Sui, S. Yan, H. Zhang, S. Sun, Y. Zeng, L.-L. Yang, and with the Institute for Infocomm Research (I2R),
L. Hanzo, “Performance analysis and approximate message passing Agency for Science, Technology and Research
detection of orthogonal time sequency multiplexing modulation,” (A*STAR), Singapore. He is currently a re-
IEEE Trans. Wireless Commun., vol. 23, no. 3, pp. 1913–1928, search fellow with the Centre for Wireless Innovation (CWI), Queen’s
Mar. 2024. University Belfast, Belfast, UK. His research interests include 6G
[37] J. Hoydis, S. Ten Brink, and M. Debbah, “Massive MIMO in the wireless communication networks, statistical signal processing, system
UL/DL of cellular networks: How many antennas do we need?” performance analysis and optimization, and compressed sensing.
 16

Hien Quoc Ngo (Senior Member, IEEE) re- Michail Matthaiou (Fellow, IEEE) was born
ceived the B.S. degree in electrical engineering in Thessaloniki, Greece in 1981. He obtained
from the Ho Chi Minh City University of Tech- the Diploma degree (5 years) in Electrical and
nology, Vietnam, in 2007, the M.S. degree in Computer Engineering from the Aristotle Uni-
electronics and radio engineering from Kyung versity of Thessaloniki, Greece in 2004. He
Hee University, South Korea, in 2010, and the then received the M.Sc. (with distinction) in
Ph.D. degree in communication systems from Communication Systems and Signal Processing
Linköping University (LiU), Sweden, in 2015. from the University of Bristol, U.K. and Ph.D.
In 2014, he visited the Nokia Bell Labs, Murray degrees from the University of Edinburgh, U.K.
Hill, New Jersey, USA. From January 2016 in 2005 and 2008, respectively. From Septem-
to April 2017, Hien Quoc Ngo was a VR re- ber 2008 through May 2010, he was with the
searcher at the Department of Electrical Engineering (ISY), LiU. He was Institute for Circuit Theory and Signal Processing, Munich University
also a Visiting Research Fellow at the School of Electronics, Electrical of Technology (TUM), Germany working as a Postdoctoral Research
Engineering and Computer Science, Queen’s University Belfast, UK, Associate. He is currently a Professor of Communications Engineering
funded by the Swedish Research Council. and Signal Processing and Deputy Director of the Centre for Wireless
Hien Quoc Ngo is currently a Reader (Associate Professor) at Queen’s Innovation (CWI) at Queen’s University Belfast, U.K. after holding
University Belfast, UK. His main research interests include massive an Assistant Professor position at Chalmers University of Technology,
MIMO systems, cell-free massive MIMO, reconfigurable intelligent Sweden. His research interests span signal processing for wireless
surfaces, physical layer security, and cooperative communications. He communications, beyond massive MIMO, intelligent reflecting surfaces,
has co-authored many research papers in wireless communications and mm-wave/THz systems and deep learning for communications.
co-authored the Cambridge University Press textbook Fundamentals of Dr. Matthaiou and his coauthors received the IEEE Communications
Massive MIMO (2016). Society (ComSoc) Leonard G. Abraham Prize in 2017. He currently
Dr. Hien Quoc Ngo received the IEEE ComSoc Stephen O. Rice Prize holds the ERC Consolidator Grant BEATRICE (2021-2026) focused on
in 2015, the IEEE ComSoc Leonard G. Abraham Prize in 2017, the the interface between information and electromagnetic theories. To date,
Best Ph.D. Award from EURASIP in 2018, and the IEEE CTTC Early he has received the prestigious 2023 Argo Network Innovation Award, the
Achievement Award in 2023. He also received the IEEE Sweden VT- 2019 EURASIP Early Career Award and the 2018/2019 Royal Academy
COM-IT Joint Chapter Best Student Journal Paper Award in 2015. He of Engineering/The Leverhulme Trust Senior Research Fellowship. His
was an IEEE Communications Letters exemplary reviewer for 2014, an team was also the Grand Winner of the 2019 Mobile World Congress
IEEE Transactions on Communications exemplary reviewer for 2015, Challenge. He was the recipient of the 2011 IEEE ComSoc Best Young
and an IEEE Wireless Communications Letters exemplary reviewer for Researcher Award for the Europe, Middle East and Africa Region and
2016. He was awarded the UKRI Future Leaders Fellowship in 2019. a co-recipient of the 2006 IEEE Communications Chapter Project Prize
Dr. Hien Quoc Ngo currently serves as an Editor for the IEEE Transac- for the best M.Sc. dissertation in the area of communications. He has
tions on Wireless Communications, the IEEE Wireless Communications co-authored papers that received best paper awards at the 2018 IEEE
Letters, Digital Signal Processing, Elsevier Physical Communication WCSP and 2014 IEEE ICC. In 2014, he received the Research Fund
(PHYCOM). He was a Guest Editor of IET Communications, special for International Young Scientists from the National Natural Science
issue on “Recent Advances on 5G Communications” and a Guest Editor Foundation of China. He is currently the Editor-in-Chief of Elsevier
of IEEE Access, special issue on “Modelling, Analysis, and Design Physical Communication, a Senior Editor for IEEE W IRELESS C OMMU -
of 5G Ultra-Dense Networks”, in 2017. He has been a member of NICATIONS L ETTERS and IEEE S IGNAL P ROCESSING M AGAZINE , and
Technical Program Committees for many IEEE conferences such as ICC, an Associate Editor for IEEE T RANSACTIONS ON C OMMUNICATIONS.
GLOBECOM, WCNC, and VTC. He is an IEEE and AAIA Fellow.

Trinh Van Chien received the B.S. degree in

electronics and telecommunications from the
Hanoi University of Science and Technology
(HUST), Hanoi, Vietnam, in 2012, the M.S.
degree in electrical and computer engineer-
Lajos Hanzo (Life Fellow, IEEE) received
ing from Sungkyunkwan University (SKKU),
Honorary Doctorates from the Technical
Seoul, South Korea, in 2014, and the Ph.D. de-
University of Budapest (2009) and Edin-
gree in communication systems from Linköping
burgh University (2015). He is a For-
University (LiU), Linköping, Sweden, in 2020.
eign Member of the Hungarian Science-
He was a Research Associate with the Univer-
Academy, Fellow of the Royal Academy
sity of Luxembourg, Esch-sur-Alzette, Luxem-
of Engineering (FREng), of the IET, of
bourg. He is currently with the School of Information and Communica-
EURASIP and holds the IEEE Eric Sum-
tion Technology (SoICT), HUST. His interest lies in convex optimization
ner Technical Field Award. For further details
problems and machine learning applications for wireless communications
please see http://www-mobile.ecs.soton.ac.uk/,
and image & video processing. Dr. Chien received the Award of Scientific
http://en.wikipedia.org/wiki/Lajos Hanzo
Excellence in the first year of the 5G wireless project funded by European
Union Horizon 2020. He was an Exemplary Reviewer of IEEE Wireless
Communications Letters Exemplary Reviewer in 2016, 2017, and 2021,
and IEEE Transactions on Communications Exemplary Reviewer in
2022. He won the Science & Technology Award in 2023.

View publication stats