1

Reconfigurable Intelligent Surfaces Relying on

Non-Diagonal Phase Shift Matrices
Qingchao Li, Mohammed El-Hajjar, Senior Member, IEEE, Ibrahim Hemadeh, Member, IEEE, Arman Shojaeifard,
Senior Member, IEEE, Alain A. M. Mourad, Bruno Clerckx, Senior Member, IEEE, Lajos Hanzo, Fellow, IEEE

Abstract—Reconfigurable intelligent surfaces (RIS) have been (IoT) [1], [2]. Hence Boccardi et al. [3] identified a range of
actively researched as a potential technique for future wire- sophisticated enabling techniques, including massive multiple-
less communications, which intelligently ameliorate the signal input-multiple-output (MIMO) solutions and millimeter wave
propagation environment. In the conventional design, each RIS
element configures and reflects its received signal independently communications. As an additional promising component, re-
of all other RIS elements, which results in a diagonal phase shift configurable intelligent surfaces (RIS) have also been pro-
matrix. By contrast, we propose a novel RIS architecture, where posed for future wireless systems to intelligently reconfigure
the incident signal impinging on one element can be reflected the propagation environment [4]–[11]. Explicitly, in RIS, a
from another element after an appropriate phase shift adjust- large number of passive scattering elements are employed
ment, which increases the flexibility in the design of RIS phase
shifts, hence, potentially improving the system performance. The for creating additional signal propagation paths between the
resultant RIS phase shift matrix also has off-diagonal elements, base station (BS) and the mobile terminal users, which can
as opposed to the pure diagonal structure of the convention- substantially enhance the performance, especially when the
al design. Compared to the state-of-art fully-connected/group- direct link between the BS and the users is blocked.
connected RIS structures, our proposed RIS architecture has Previous contributions on RIS are mainly focused on maxi-
lower complexity, while attaining a higher channel gain than the
group-connected RIS structure, and approaching that of the fully- mizing the spectral efficiency/achievable rate or minimizing
connected RIS structure. We formulate and solve the problem of the transmission power [12]–[21]. In [12], Wu and Zhang
maximizing the achievable rate of our proposed RIS architecture minimized the transmission power in the downlink of RIS-
by jointly optimizing the transmit beamforming and the non- aided multi-user MIMO systems, where the popular alternating
diagonal phase shift matrix based on alternating optimization and optimization and semi-define relaxation (SDR) methods were
semi-define relaxation (SDR) methods. Moreover, the closed-form
expressions of the channel gain, the outage probability and bit employed for jointly optimizing the active transmission beam-
error ratio (BER) are derived. Simulation results demonstrate forming (TBF) of the BS and the passive beamforming, rep-
that our proposed RIS architecture results in an improved resented by the RIS phase shift matrix. This was achieved by
performance in terms of the achievable rate compared to the approximately configuring the RIS reflecting elements. Ning
conventional architecture, both in single-user as well as in multi- et al. [13] maximized the sum-path-gain of RIS-assisted point-
user scenarios.
to-point MIMO systems, where the low-complexity alternating
Index Terms—Reconfigurable intelligent surfaces (RIS), chan- direction method of multipliers (ADMM) was employed for
nel gain, outage probability, average bit error ratio (BER), joint configuring the RIS phase shift matrix, while the classic sin-
beamforming.
gular value decomposition (SVD) was employed for designing
the TBF. In [14], the optimal closed-form solution of the
I. I NTRODUCTION phase shift matrix and TBF were derived by Wang et al.
for single-user multiple-input-single-output (MISO) millimeter
I N future wireless networks an ultra-high data rate, ultra-
low latency, ultra-high reliability and ubiquitous connec-
tivity is required for communication, computation, sensing
wave systems.
Additionally, in order to reduce the overhead of channel
and location awareness, especially in the Internet of Things estimation, Han et al. [15] maximized the ergodic spectral
efficiency of RIS-assisted systems communicating over Rician
L. Hanzo would like to acknowledge the financial support of the Engi- fading channels, relying on the angle of arrival (AoA) and
neering and Physical Sciences Research Council projects EP/P034284/1 and angle of departure (AoD) information. The problem of maxi-
EP/P003990/1 (COALESCE) as well as of the European Research Council’s
Advanced Fellow Grant QuantCom (Grant No. 789028). The support of mizing the ergodic spectral efficiency based on statistical CSI
Interdigital is also gratefully acknowledged. (Corresponding author: Lajos in Rician fading channels was studied in [16], where Wang et
Hanzo.) al. considered the effect of channel correlation on the ergodic
Qingchao Li, Mohammed El-Hajjar and Lajos Hanzo are with the Elec-
tronics and Computer Science, University of Southampton, Southampton spectral efficiency.
SO17 1BJ, U.K. (e-mail: [email protected]; [email protected]; While considering quantized RIS phase shifts, Wu and
[email protected]). Zhang [17] employed the popular branch-and-bound method
Ibrahim Hemadeh, Arman Shojaeifard and Alain A. M.
Mourad are with InterDigital, London EC2A 3QR, U.K. (e-mail: and an exhaustive search method for single-user and multi-
[email protected]; [email protected]; user RIS-assisted systems, respectively. The branch-and-bound
[email protected]). algorithm was also employed by Zhang et al. [18] to design
Bruno Clerckx is with the Communications and Signal Processing Group,
Department of Electrical and Electronic Engineering, Imperial College Lon- a discrete phase shift matrix, where the RIS has the dual
don, London SW7 2AZ, U.K. (email: [email protected]). functions of both reflection and refraction. In [19], the local
 2

search (LS) method and cross-entropy (CE) method were non-reciprocal connections, in which the signal impinging
proposed for optimizing the RIS phase shift matrices having on a specific element can be reflected from another element
discrete entries. In [20], Xu et al. designed their discrete after phase shift adjustment, so the phase shift matrix can
phase shift matrix based on low resolution digital-to-analog be non-symmetric and of non-diagonal nature. These pro-
converters, and derived the lower bound of the asymptotic rate. vide flexibility in terms of configuring the RIS structure for
Furthermore, the problem of maximizing the achievable rate enhancing the system performance. We employ alternating
of RIS users was studied by Lin et al. [21], where the novel optimization and SDR methods for jointly optimizing the
concept of reflection pattern modulation was employed. TBF and phase shift matrix for single-user MISO systems
The theoretical performance analysis of RIS-assisted single- and multi-user MIMO systems, respectively. The theoretical
input-single-output (SISO) systems was also investigated. In analysis and simulation results demonstrate that our proposed
[22], the theoretical channel gain of RIS-aided systems was RIS architecture achieves better channel gain, outage proba-
characterized by Basar et al., compared to that of conventional bility, average bit error ratio (BER) and throughput than the
SISO systems operating without RIS. Furthermore, the instan- conventional RIS architecture. Furthermore, in our proposed
taneous signal-noise-ratio (SNR) has been derived based on the RIS architecture, there are only N non-zero entries in the
central-limit-theorem (CLT). In [23], Yang et al. derived the phase shift matrix, which is the same as that in the con-
accurate closed-form theoretical instantaneous SNR expression ventional RIS architecture. Additionally, the position of the
for a dual-hop RIS-aided scheme. non-zero entries in our proposed RIS architecture has to be
However, the RIS structures of [12]–[23], assumed that the updated, which requires N values of information. Hence the
incident signal impinging on a specific element can be only total information to be exchanged over the BS-RIS control
reflected from the same element after phase shift adjustment. link in each coherence time duration of our proposed RIS
In other words, there was no controlled relationship among architecture is 2N , i.e. significantly lower than that of the
the RIS elements. We refer to this RIS architecture as the fully-connected RIS architecture. Against this background, the
conventional RIS architecture. Therefore, the phase shift ma- novel contributions of this paper are summarized as follows:
trix in these designs has a diagonal structure, which does not • We propose a novel RIS architecture having a non-
exploit the full potential of RIS for enhancing the system diagonal phase shift matrix, and jointly design the TBF
performance. To the best of our knowledge, only Shen et and phase shift matrix by alternating optimization and
al. [24] studied the cooperation among RIS elements, where SDR methods for maximizing the achievable rate.
fully-connected/group-connected network architectures were • We provide both theoretical and simulation results for
proposed. The associated theoretical analysis and simulation characterizing the performance of our proposed RIS ar-
results demonstrated that their architectures are capable of chitecture, which is better than the conventional RIS ar-
significantly increasing the received signal power, compared chitecture in terms of its channel gain, outage probability,
to the conventional RIS structure, when considering SISO average BER and achievable rate.
systems. However, the performance enhancement reported in • We show that the performance of our proposed ar-
[24] is attained at the cost of increased optimization complex- chitecture approaches that of the state-of-the-art fully-
ity. For example, N (N2+1) entries are available in the fully- connected RIS architecture, while providing better perfor-
connected phase shift matrix and N (G+1) 2 entries are in the mance than that of the group-connected RIS architecture,
group-connected phase shift matrix. By contrast, there are only when the number of RIS elements increases. Additionally,
N non-zero entries in the conventional RIS case, where N is this is attained while requiring a reduced information ex-
the number of RIS elements and G is the group size of the change over the BS-RIS control link and fewer optimized
group-connected architecture. This increases the number of phase shift entries than the fully-connected architecture.
entries to optimize in the phase shift matrix and the amount Finally, Table I explicitly contrasts our contributions to the
of information to be transferred over the BS-RIS control literature.
link. Furthermore, the fully-connected/group-connected phase The rest of this paper is organized as follows. In Section
shift matrix of [24] has the additional constraint of symmetry II, we present the system model. The beamforming design
because all the proposed architectures are reciprocal. methods are formulated in Section III. Our theoretical analysis
By contrast, we propose a novel RIS structure relying on and simulation results are presented in Section IV and Section

TABLE I
N OVELTY COMPARISON WITH THE LITERATURE .

Our paper [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24]
√ √ √ √ √ √ √ √ √ √ √ √
Beamforming design
√ √
Outage performance analysis
√ √ √
Average BER analysis
√ √ √ √
Multi-user
√ √
Cooperation among RIS elements
 3

Nx RIS
V, respectively. Finally, we conclude in Section VI.








Notations: Vectors and matrices are denoted by boldface 
RIS controller 
 Ny
lower and upper case letters, respectively. (·)T , (·)H represent G  h1H

the operation of transpose and Hermitian transpose, respec- h2H User 1
1
tively. Cm×n denotes the space of m × n complex-valued 2
hK H User 2
matrix. [a]i represents the ith element in vector a, and [A]i,j M

represents the (i, j)th element in matrix A. diag{a} denotes

User K
a diagonal matrix with each element being the elements in Base station
Blocks
vector a, I represents the identity matrix. Tr(A), Rank(A) Fig. 1. The employed RIS-assisted wireless communication system model,
and |A| represent the trace, rank and determinant of matrix A, including a M -antenna base station, K single-antenna users and a RIS with
respectively. A  0 indicates that A is a positive semi-define N = Nx × Ny elements.
matrix. |a| (or |a|) and ∠a (or ∠a) represent the amplitude
and phase of the complex scalar a (or complex vector a), where δa is the distance between adjacent BS antennas, λ
respectively. kak denotes the 2-norm of vector a. fX (x) is carrier wavelength, ψ D is the angle

of departure (AoD)
A
and FX (x) are the probability density function (PDF) and of signals from the BS. fRIS φA , ϕA is the response of an
cumulative distribution function (CDF) of random variables N = Nx × Ny uniform rectangular planar array (URPA) at
X. A circularly symmetric complex Gaussian random vector the RIS, given by [26]

h δ0
with mean µ and covariance matrix Σ is denoted as CN (µ, Σ). A A A
f A φA , ϕA = 1, · · · , e−j2π λ (nx sin φ cos ϕ +ny cos φ ) ,
RIS
E(X) represents the mean of the random variable X. iT
δ0 A
cos ϕA +(Ny −1) cos φA )
· · · , e−j2π λ ((Nx −1) sin φ , (4)
II. S YSTEM M ODEL where 0 ≤ nx ≤ Nx − 1, 0 ≤ ny ≤ Ny − 1, δ0 is the distance
A. Channel Model
between adjacent RIS elements, φA and ϕA are the elevation
The RIS-assisted system model is illustrated in Fig. 1, and azimuth angle of arrival (AoA) of signals to the RIS,
including a BS having M transmit antennas, K single-antenna respectively.
mobile receivers, and a RIS with N elements. The direct link
The NLoS component G e = [e e2 , · · · , g
g1 , g em ∈ CN ×1
eM ], g
between the BS and the users is blocked, while the RIS creates
is given by
additional communication links arriving from the BS to the
users. The link spanning from the BS to the RIS is denoted em ∼ CN (0, %t I) , m = 1, 2, · · · , M.
g (5)
as G = [g1 , g2 , · · · , gM ] ∈ CN ×M , where gm ∈ CN ×1 the RIS to the kth user
The Rician channel model from s
(m = 1, 2, · · · , M ) represents the channel vector from the mth r
κ H 1
hk
BS antenna to the RIS. The links impinging from the RIS to hHk = hk + eH,
h (6)
1 + κhk 1 + κhk k
the users are denoted as H = [h1 , h2 , · · · , hK ]H ∈ CK×N ,
H
where hH k ∈ C
1×N
(k ∈ 1, 2, · · · , K) represents the channel where κhk is the Rician factor, hk ∈ C1×N and h
e H ∈ C1×N
k
vector from the RIS to the kth single-antenna receiver. We represent the LoS and NLoS components, respectively.
employ the far field RIS channel model, since the size of H
The LoS component hk is expressed as [26]
the RIS is negligible compared to both the BS-RIS distance H D,k H
φD,k , ϕD,k ,


and to the RIS-user distance [25]. Additionally, we consider a hk = fRIS (7)
distance-dependent path loss model, and Rician channel model D,k H

where fRIS φD,k , ϕD,k is the response of N -element URPA
for the small scaling fading [12]. at the RIS, given by [26]
The Rician channel model from the BS to the RIS is given
D,k H
φD,k , ϕD,k


be fRIS
√
r r h δ0
κG 1 D,k D,k D,k
= %r,k 1, · · · , e−j2π λ (nx sin φ cos ϕ +ny cos φ ) , · · · ,
G= G+ G,
e (1)
1 + κG 1 + κG δ0 D,k D,k D,k
i
where κG is the Rician factor, G ∈ CN ×M and G e ∈ CN ×M e−j2π λ ((Nx −1) sin φ cos ϕ +(Ny −1) cos φ ) , (8)
represent the line-of-sight (LoS) and non-line-of-sight (NLoS) −αr
where %r,k = C0 dr,k denotes the path loss from the RIS to
components, respectively. the kth user, in which dr,k denotes the distance between the
The LoS component G is expressed as RIS and the kth user, αr is the RIS-user path loss exponent,
√ A
D H D

and φD,k and ϕD,k are the elevation and azimuth AoD of
G = %t fRIS φA , ϕA fBS ψ , (2)
signals from the RIS to the kth user, respectively.
where %t = C0 d−αt
t
denotes the path loss of the BS-RIS link,
in which dt denotes the distance between the BS and the RIS. The NLoS component h e k is given by
C0 is the path loss at the reference distance of
1 meter, and αt e k ∼ CN (0, %r,k I) , k = 1, 2, · · · , K.
h (9)
D H
is the BS-RIS path loss exponent. fBS ψ D is the response In this paper, we assume that instantaneous CSI knowledge
of the M -antenna uniform linear array (ULA) at the BS, based can be attained at the BS.
on [26]

h i In our RIS-assisted system, the signal is precoded by the
D H δa D δa D
fBS ψ D = 1, e−j2π λ sin ψ , · · · , e−j2π λ (M −1) sin ψ , TBF at the BS and transmitted to the RIS. The RIS configures
(3) the phase shifts of the impinging signals and then reflects them
 4

x1 y1 x2 y2 x3 y3 x4 y4 x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 x6 y6
θi ∈ [0, 2π) can be configured for maximizing the channel gain
e j1 e j2 e j3 e j4 e j1 e j2 e j3 e j4 e j5 e j6 [25]. Since the phase shift of each RIS element is configured
element 2 element 3 element 4
independently, the phase shift matrix, denoted as Θ̄ 1 , is
element 1 element 2 element 3 element 4 element 1 element 5 element 6
(a1) (b1)

x1 y1 x2 y2 x3 y3 x4 y4 x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 x6 y6
diagonal and can be represented as
Θ̄ = diag{ejθ1 , ejθ2 , · · · , ejθN }. (12)
j 2,1 j 4,2 j3,3 j1,4 j3,1 j 4,2 j1,3 j5,4 j 2,5 j 6,6
e e e e e e e e e e
Hence, the equivalent channel spanning from the mth BS
element 1 element 2 element 3 element 4 element 1 element 2 element 3 element 4 element 5 element 6 transmit antenna to the kth user can be represented as
(a2) (b2)
N
Fig. 2. An example of illustration for the conventional RIS architecture X
having N = 4 elements and N = 6 elements in (a1) and (b1) respectively, hHk Θ̄g m = [hH
k ]i e
jθi
[gm ]i
and the proposed RIS architecture having N = 4 elements with the bijection
function M : {1, 2, 3, 4} → {2, 4, 3, 1} and N = 6 elements with the i=1
bijection function M : {1, 2, 3, 4, 5, 6} → {3, 4, 1, 5, 2, 6} in (a2) and (b2) N
H
respectively. j (θi +∠[hk ]i +∠[gm ]i )
X
= |[hH
k ]i | · |[gm ]i | · e , (13)
i=1
to the users. Therefore, the system model is represented as
p √ which includes the BS-RIS channel, the phase shift applied at
y = Pt HΘGW Λx + n, (10) the RIS and the RIS-user channel.
K×1 K×1
where x ∈ C is the transmitted signal vector, y ∈ C On the other hand, if there is a connection between the RIS
is the received signal vector, n ∼ CN (0, σn2 I) ∈ CK×1 elements, i.e. the incident signal impinging on the ith element
is the circularly symmetric complex Gaussian noise, W = can be reflected from other elements, then we will have more
[w1 , w2 , · · · , wK ] ∈ CM ×K represents the active TBF matrix flexibility in the design of the RIS phase shift matrix, which
at the BS, Pt is the total transmitted power of the BS, and can provide an improved performance. The one and only
Λ = diag{λ1 , λ2 , · · · , λK } is a diagonal power allocation contribution on RIS element cooperation, which was termed
matrix, where λk represents the power allocated to the signal as the fully-connected and group-connected RIS architecture,
transmitted to the kth user. Hence, in order to normalize the was disseminated by Shen et al. [24]. In their solution, the
transmit power, we have the following constraints: kxk = 1, signal impinging on each RIS element was divided into N
kwk k = 1, and λ1 + λ2 + · · · + λK = 1. Still referring to (10), components, and these N signal components are reflected from
Θ represents the RIS phase shift matrix, which is diagonal N RIS elements after phase shift configuration. This fully
in the conventional RIS architecture, while it is non-diagonal connected RIS architecture attains a substantial channel gain.
in our proposed RIS architecture. Our objective is to jointly However, its performance enhancement is achieved at the cost
optimize the RIS phase shift matrix Θ, the TBF matrix W and of having more entries in the phase shift matrix to optimize,
the power allocation matrix Λ for maximizing the achievable which includes N × N elements, as opposed to having only
rate. N non-zero elements in the conventional design. Further-
In a practical RIS-assisted wireless system as shown in more, extra information has to be transmitted over the BS-
Fig. 1, the phase shift matrix Θ, the TBF matrix W and RIS controller link. On the other hand, the group-connected
the power allocation matrix Λ are jointly optimized at the architecture has significantly lower complexity than the fully-
BS by exploiting the CSI available, i.e. the BS-RIS channel connected architecture, which imposes a modest performance
matrix G and the RIS-users channel matrix H. Then, using loss. Hence, we propose a novel RIS architecture, which
the BS-RIS controller link, the optimized phase shift matrix approaches the fully-connected performance at a significantly
Θ is transmitted to the RIS controller, which is responsible reduced complexity.
for reconfiguring the phase applied to the RIS elements. 2) The proposed RIS Architecture: Explicitly, we design a
In the following, we will briefly highlight the conventional novel RIS architecture, where the signal impinging on the ith
RIS architecture, where the phase shift matrix is diagonal. element can be reflected from another one element, denoted as
Then, our proposed non-diagonal RIS architecture will be the i0 th element, after phase shift adjustment. The relationship
presented. between the incident signals and the reflected signals can be
B. RIS Architecture represented as
1) Conventional RIS Architecture: Fig. 2 (a1) shows an yi0 = xi ejθi0 ,i , (14)
example of a conventional 4-element RIS architecture, where where i belongs to the RIS element index set of incident
each RIS element is ‘single-connected’, i.e. the signal im- signals I : {1, 2, · · · , N }, and i0 belongs to the RIS element
pinging on the ith element is only reflected from the ith index set of reflected signals I 0 : {1, 2, · · · , N }. There is a
element after phase shift adjustment. Similarly, a conventional bijective function M : I → I 0 , and i0 = M(i), where the
6-element RIS architecture is showed in Fig. 2 (b1). In bijection M is a function between the RIS element indices
conventional RIS-assisted wireless communication systems, of incident signals and that of the reflected signals. For
each RIS element changes the phase of the impinging signals example, in Fig. 2 (a2), we consider an example using four
independently, that is [5] elements, where the signal impinging on the first element is
yi = βi xi ejθi , i = 1, 2, · · · , N, (11)
1 In this paper, we use Θ̄ to represent the diagonal phase shift matrix in
where xi and yi represent the incident signal and reflected
the conventional RIS architecture, and use Θe to represent the non-diagonal
signal of the ith RIS element, respectively. The amplitude phase shift matrix in our proposed RIS architecture. Θ is used when it is not
gain βi is set to 1 to realize full reflection, and the phase shift specified which kind of RIS architecture is employed.
 5

Configurable impedance Z1
reflected from the second element, thus M(1) = 2. Similarly, Incident wave 1
C1 R L I1
M(2) = 4, M(3) = 3, and M(4) = 1. In Fig. 2 (b2), Lbot top

we consider an example using six elements, where the signal Transmission Line/Air

Reflected wave 1
impinging on the first element is reflected from the third
R1 Switch-(1,1) Switch-(1,2) Switch-(1,3) Switch-(1,4)
element, thus M(1) = 3. Similarly, M(2) = 4, M(3) = 1, Configurable impedance Z 2
M(4) = 5, M(5) = 2, and M(6) = 6. Incident wave 2
R L
C2 Lbot top I2
Therefore, in our proposed method the phase shift matrix,
Transmission Line/Air
denoted as Θ,e is non-diagonal, and there is only a single non-
Reflected wave 2
zero element in each row and each column. The phase shift R2
Switch-(2,1) Switch-(2,2) Switch-(2,3) Switch-(2,4)
matrix in Fig. 2 (a2) can be represented as Configurable impedance Z 3
Incident wave 3
ejθ1,4
 
0 0 0 C3 R L I3
Lbot top
 jθ2,1
Θe = e 0 0 0  . (15) Transmission Line/Air
 0 0 ejθ3,3 0  Reflected wave 3
jθ4,2
0 e 0 0 R3 Switch-(3,1) Switch-(3,2) Switch-(3,3) Switch-(3,4)
Configurable impedance Z 4
Similarly, the phase shift matrix in Fig. 2 (b2) can be repre- Incident wave 4
sented as C4 R L
Lbot top I4

ejθ1,3
 
0 0 0 0 0 Transmission Line/Air
 0
 jθ 0 0 0 ejθ2,5 0   Reflected wave 4
 e 3,1 0 0 0 0 0  R4 Switch-(4,1) Switch-(4,2) Switch-(4,3) Switch-(4,4)
Θ=
e  .
 0 ejθ4,2 0 0 0 0   Switch array
 0 0 0 ejθ5,4 0 0  Fig. 3. The transmission line model of our proposed non-diagonal phase
shift matrix architecture with N = 4 RIS elements.
0 0 0 0 0 ejθ6,6
(16) shifts of the non-diagonal phase shift matrix are appropriately
Since only N non-zero entries of the phase shift matrix designed. The details of optimizing our proposed non-diagonal
have to be optimized and N position information values of RIS phase shift matrix will be discussed in Section III, where
these non-zero entries have to be recorded in each coherence the bijection function M and the values of each element’s
time for our RIS-aided systems, this only modestly increases phase shift can be obtained in the non-diagonal phase shift
the optimisation complexity and the amount of information matrix Θ.
e
transmitted over the BS-RIS controller link, compared to the C. Implementation Circuits
conventional architecture.
Although the implementational specifics of our proposed
Since the phase shift matrix in our proposed method is non-
RIS architecture are beyond the scope of this paper, in the
diagonal, we may refer to it as the RIS architecture with non-
following we briefly highlight a potential implementation
diagonal phase shift matrix. Compared to the conventional RIS
suitable for our proposed architecture.
architecture with diagonal phase shift matrix, the proposed
In [24], the authors employed scattering parameter network
non-diagonal phase shift matrix method has the potential of
models based on reciprocal architectures for describing the im-
attaching higher channel gain. Let us consider an example
plementation of the conventional RIS structure and the fully-
using a 4-element RIS employed in a SISO system, and
connected/group-connected RIS structure. The corresponding
assume that the channel vector of the link from the iBS
h 3π 5π 7π π
T phase shift matrices are symmetric in [24]. On the other
to the RIS is g = 1.4e−j 4 , 0.2ej 6 , 0.4e−j 8 , 0.8e−j 6 , hand, our proposed architecture relies on a potentially non-
hwhile that of the link from thei RIS to the user is hH = symmetric matrix structure due to the fact that our design
−j π 2π π π
0.6e 4 , 1ej 3 , 0.3ej 3 , 0.1ej 8 . In the conventional phase requires non-reciprocal connections.
shift matrix based method, the channel gain can be maximized In the following, we employ the classic transmission line
when the RIS phase
shifts are designed coherently,
i.e., θ1 = model [27] for highlighting design the implementation of our
− ∠[g]1 + ∠[hH ]1 = π,
θ2 = − ∠[g]2 + ∠[hH ]2 = −
3π 2 , proposed RIS architecture. The circuits of the conventional
θ3 = − ∠[g]3 + ∠[hH ]3 = 5π 8 , θ 4 = − ∠[g] 4 + ∠[h H
]4 = RIS architecture have been presented in [27], [28]. When it
π
24 . Then, the corresponding channel gain is given by 1.4 × comes to using the transmission line model based design of our
0.6 + 0.2 × 1 + 0.4 × 0.3 + 0.8 × 0.1 = 12.4. By contrast, RIS architecture, we can employ switch arrays for connecting
in our proposed non-diagonal phase shift matrix method, if the different RIS elements. Fig. 3 illustrates an example of the
the phase shift matrix is designed as the structure in (15), transmission line model of our RIS architecture with N = 4
i.e. when the bijective function is M(1) = 2, M(2) = 4, elements. In each RIS element, the reconfigurable impedance
M(3) = 3, M(4) = 1, and the RIS phase shifts
are designed includes a bottom layer inductance Lbot , a top layer inductance
H 5π
coherently, i.e., θ1,4
π = − ∠[g] 4 + ∠[h ]1 = 12
θ2,1 =
, Ltop , an effective resistance R, and a variable capacitance
− ∠[g]1 + ∠[hH ]2 = 12 ,
θ3,3 = − ∠[g]3 + ∠[hH ]3 = 13π 24 , Cn , where n = 1, 2, · · · , N [28]. The phase shift of the
θ4,2 = − ∠[g]2 + ∠[hH ]4 = − 23π 24 , then the corresponding reconfigurable impedance Zn is controlled by its variable
channel gain is given by 0.8 × 0.6 + 1.4 × 1 + 0.4 × 0.3 + capacitance Cn . To realize an N -element non-diagonal phase
0.2 × 0.1 = 20.2. Therefore, higher channel gain can be shift matrix, an array of N × N switches is required. The
achieved when the bijective function M and the RIS phase ON/OFF state of these switches is determined by the positions
 6

of non-zero entries in the RIS phase shift matrix. Specifically, Similar to [25], by ignoring the constant terms, the achievable
the switches are turned on if the corresponding element in the rate optimization problem of SISO systems is equivalent to
RIS phase shift matrix is non zero, while they are turned off, maximizing the channel gain as follows
if the corresponding elements are zero. For example, to realize (P1.b) max |hH Θg|2 (20)
the non-diagonal phase shift matrix of (15), Switch-(2,1), Θ

Switch-(4,2), Switch-(3,3) and Switch-(1,4) are turned on, s.t. 0 ≤ θM(i),i < 2π, i = 1, 2, · · · , N.
while the other switches are turned off in Fig. 3. In this case,
1) Conventional RIS Architecture: Again, in the conven-
the signal impinging on the first RIS element is reflected from
tional RIS architecture, the phase shift matrix is diagonal, so
the second RIS element after phase shift configuration, while
the bijection f : I → I 0 is essentially M(i) = i. Therefore,
the signal impinging on the second RIS element is reflected
the channel gain |hH Θ̄g|2 is given by
from the fourth RIS element after phase shift configuration, !2
N
etc. A potential implementation for the switches relies on using X
RF micro-electromechanical systems (MEMS) [29], which |hH Θ̄g|2 = [hH ]i [g]i ejθi , (21)
i=1
have been widely used in wireless communication systems as
a benefit of their near-zero power consumption, high isolation, where the optimal solution for Θ̄ can be obtained as [25]
low insertion loss, low intermodulation products and low cost. θi = −(∠[hH ]i + ∠[g]i ), i = 1, 2, · · · , N, (22)
There are also other potential implementations for our pro- which essentially aligns all the signals reflected by the RIS
posed architecture. Explicitly, since we have to route the signal with the impinging signals to arrange for their coherent
between the different elements depending on the bijection combination, and the maximum channel gain based on (22)
function M, radio frequency couplers and isolators can be can be expressed as [25]
employed [30]–[33]. Additionally, passive phase shifters way N
!2
also be employed in conjunction with these couplers for
X
H 2
max |h Θ̄g| = %t %r ai bi , (23)
attaining accurate phase shifts [34]. Finally, the authors of [35], i=1
[36] provided comprehensive discussions of metasurfaces, in- where ai and bi are the amplitude of √1%t [g]i and √1%r [hH ]i ,
cluding their operation and functionalities. Hence the proposed respectively. Observe from (23) that in conventional RIS
architecture is viable and has several existing implementations, architectures, the maximum channel gain is proportional to the
which can be further optimized in our future researches. square of the sum of ai bi , in which ai and bi are amalgamated
based on the Equal Gain Combining (EGC) criterion.
III. B EAMFORMING D ESIGN
In the previous section, we presented a novel RIS archi- 2) The proposed RIS Architecture: Since in our proposed
tecture, while here we present our joint beamforming design RIS architecture the signal impinging on the ith element can
maximizing the attainable rate. This is achieved by jointly be reflected from the i0 th element after phase shift adjustment,
optimizing the passive beamforming matrix of the RIS and the channel gain can be written as
!2 !2
the TBF matrix of the BS. We start with deriving the closed- XN N
X
He 2 H jθi0 ,i
form solution of the SISO case, and then extend our analysis |h Θg| = [h ]i0 [g]i e = %t %r ai bi0 ,
to more general single-user MISO case and multi-user MIMO i=1 i=1
case by employing the alternating optimization method and (24)
semi-definite relaxation technique [37], respectively. where
A. Beamforming Design for SISO Systems θi0 ,i = −(∠[hH ]i0 + ∠[g]i ), i0 = M(i).
i = 1, 2, · · · , N,
In SISO systems, both the BS and the single user are (25)
equipped with a single antenna, while the RIS has N reflecting Hence, first we should aim for finding the function M : I →
elements. The system model in (10) can be written as I 0 to maximize the channel gain in (24). According to the
p Maximum Ratio Combining (MRC) criterion, the maximum
y = Pt hH Θgx + n, (17)
1×1 N ×1
of the channel gain in (24) is given by
where x ∈ C is the transmitted signal, g ∈ C is the N
!2
channel vector arriving from the BS to the RIS, hH ∈ C1×N He 2
X
max |h Θg| = %t %r a(i) b(i) , (26)
is the channel vector of the link spanning from the RIS to i=1
the user, and n ∈ C1×1 is the circularly symmetric complex where a(1) , a(2) , · · · , a(N ) represents the sequence of
Gaussian noise. The achievable rate of this SISO link is given a1 , a2 , · · · , aN sorted in an ascending order, and similarly,
by b(1) , b(2) , · · · , b(N ) is the sequence of b1 , b2 , · · · , bN sorted
 
Pt H 2 in an ascending order. According to [38], when a permutation
RSISO = log2 1 + 2 |h Θg| . (18)
σn matrix is multiplied from the left by a column vector, it
Our aim is to find the phase shift matrix Θ that maximizes will permute the elements of the column vector, while when
the rate, which can be formulated
 as  a permutation matrix is multiplied from the right by a
Pt H row vector, it will permute the elements of the row vector.
(P1.a) max log2 1 + 2 |h Θg|2 (19)
Θ σn Therefore, the channel gain in (26) can be written as
s.t. 0 ≤ θM(i),i < 2π, i = 1, 2, · · · , N. max |hH Θg|
e 2 = |hH Jr Θ̄Jt g|2 , (27)
 7

Im{•}
large.
b2 [h ]2
H

Finally, when the optimal phase shift matrix Θ

e opt is at-
b3[h H ]3 tained, the achievable rate is given by
 
a2 [g ]2 Pt H e 2
b4 [h H ]4 RSISO = log2 1 + 2 |h Θopt g| . (32)
Re{•}
σn
a3[g ]3

a4 [g ]4 B. Beamforming Design for Single-user MISO Systems

b1[h H ]1

In single-user MISO systems, the BS is equipped with M

a1[g ]1
downlink transmit antennas and the single user is equipped
Fig. 4. The illustration of channel vectors in the example of a 4-element with a single receiver antenna, while the RIS has N elements.
RIS-assisted system.
The system model in (10) can be written as
p
where Jt is a permutation matrix, which sorts the element y = Pt hH ΘGwx
e + n, (33)
amplitude in the column vector g in an ascending order,
and the achievable rate is given by
yielding √1%t |Jt g| = [a(1) , a(2) , · · · , aN ]T . Still referring to  
Pt H e 2
(27), Jr is a permutation matrix, which sorts the element RMISO = log2 1 + 2 |h ΘGw| . (34)
σn
amplitude in the row vector hH in an ascending order, leading
to √1%r |hH Jr | = [b(1) , b(2) , · · · , b(N ) ]. In this case, the channel The problem of maximizing the achievable rate can be formu-
lated as
vectors g and hH can be combined based on the MRC 
Pt H e

2
criterion. Furthermore, Θ̄ in (27) is a diagonal phase shift (P2.a) max log2 1 + 2 |h ΘGw| (35)
Θ,w
e σn
matrix, in which the ith diagonal element is given by
s.t. kwk = 1,
θi = −(∠[hH Jr ]i + ∠[Jt g]i ), i = 1, 2, · · · , N. (28)
0 ≤ θM(i),i < 2π, i = 1, 2, · · · , N.
Therefore, the optimal non-diagonal phase matrix is formulat-
ed as Θ e opt = Jr Θ̄Jt . Similar to SISO cases, the achievable rate optimization prob-
To expound further, in Fig. 4, we present an example of lem of single-user MISO systems is equivalent to maximizing
a 4-element RIS-assisted system, where the ’99K’ vectors the channel gain as follows
represent the BS-RIS channel vector g, while the ’→’ vectors (P2.b) max |hH ΘGw|
e 2
(36)
Θ,w
correspond to the RIS-user channel vector hH . It can be
e

observed that a2 < a3 < a4 < a1 and b4 < b3 < b1 < b2 , so s.t. kwk = 1,
a(1) = a2 , a(2) = a3 , a(3) = a4 , a(4) = a1 , and b(1) = b4 , b(2) = 0 ≤ θM(i),i < 2π, i = 1, 2, · · · , N.
b3 , b(3) = b1 , b(4) = b2 . Therefore, in our proposed architecture, Since (P2.b) represents a non-convex problem, we employ
the permutation matrices Jt and Jr are derived as the popular alternating optimization method for solving it
   
0 1 0 0 0 0 1 0 iteratively.
 0 0 1 0   0 0 0 1 
Jt =  Firstly, when the TBF vector w is given, Gw becomes a
 0 0 0 1  , Jr =  0 1 0 0  , (29)
  
column vector, and the phase shift matrix Θ e can be designed
1 0 0 0 1 0 0 0 similarly as in the SISO case.
and the bijective function is given by M(1) = 2, M(2) = Secondly, when the phase shift matrix Θe is given, the equiv-
4, M(3) = 3, M(4) = 1. Hence, the non-diagonal phase shift He
alent channel can be obtained as h ΘG. Then the TBF vector
matrix Θ e is designed as can be designed based on the maximum ratio transmission
He
ΘG)H
Θe = Jr Θ̄Jt (MRT) method, yielding w = (h H
kh ΘGk
e . The detailed process
= Jr · diag{ejθ1 , ejθ2 , · · · , ejθN } · Jt of the alternating optimization method conceived for RIS-

0 0 0 ejθ3
 assisted single-user MISO systems is shown in Algorithm 1.
 ejθ4 0 0 0  When the optimal phase shift matrix Θ e opt and TBF vector
= 0 jθ2
, (30) wopt are obtained, the achievable rate is given by
0 e 0   
0 ejθ1 0 0 Pt
RMISO = log2 1 + 2 |hH Θ e opt Gwopt |2 . (37)
where e jθ1
=e jθ4,2
= −(∠[h ]4 + ∠[g]2 ), ejθ2 = ejθ3,3 =
H σn
−(∠[h ]3 + ∠[g]3 ), ejθ3 = ejθ1,4 = −(∠[hH ]1 + ∠[g]4 ),
H

ejθ4 = ejθ2,1 = −(∠[hH ]2 + ∠[g]1 ). Then, the corresponding C. Beamforming Design for Multi-user MIMO Systems
channel gain is optimized as In multi-user MIMO systems, including a BS having M
e 2 = %t %r a(1) b(1) + a(2) b(2) + a(3) b(3) + a(4) b(4) 2
|hH Θg| transmit antennas and K single-antenna users, the system


2 model is given by
= %t %r (a2 b4 + a3 b3 + a4 b1 + a1 b2 ) . (31) p √
y = Pt HΘGW e Λx + n, (38)
We observe from (26) that in our proposed RIS architecture,
the maximum channel gain is proportional to the square of the and the achievable rate is formulated as
sum of a(i) b(i) , in which a(i) and b(i) are combined by obeying Pt e H HeH H
RMIMO = log2 I + 2 HΘGWΛW G Θ H . (39)
the MRC criterion, when the number of RIS elements N is σn
 8

K
!
Algorithm 1 Alternating optimization method for RIS-assisted H
X
single-user MISO systems =q Φk ΦH
k q, (42)
Input: BS-RIS channel matrix G, and RIS-user channel k=1
vector hH . where Φk = diag{hH k Jr }Jt G. Since the phase shift
Output: The optimal phase shift matrix Θ e opt , and the opti-
matrix Θ is determined
e by q, Jt and Jr , the problem
mal transmission beamforming vector wopt . PK
1: Choose the proper permutation matrix Jr which sorts the
of maximizing k=1 khH ΘGk e 2
is formulated as
!
element amplitude of row vector hH in an ascending order. K
X
2: Set an initial w satisfying kwk = 1. (P3.b2) max qH Φk ΦHk q (43)
3: Repeat q,Jt ,Jr
k=1
4: Choose the proper permutation matrix Jt which can
sort the element amplitude of column vector Gw in an s.t. |[q]i | = 1, i = 1, 2, · · · , N.
ascending order. In Φk , since the BS-RIS channel matrix G =
5: According to the row vector hH Jr and column vector [g1 , g2 , . . . , gM ] has M columns, and the RIS-users’
Jt Gw, the elements in diagonal
matrix Θ̄ is designed as
channel matrix h = [h1 , h2 , · · · , hK ]H has K rows,
θi = − ∠[hH Jr ]i + ∠[Jt Gw]i , i = 1, 2, · · · , N .
6: Design the transmission beamforming vector based the permutation matrices Jt and Jr cannot be designed
H
Jr Θ̄Jt G)H similarly to the SISO case by using a sorting method.
on MRT criterion as w = (h khH Jr Θ̄Jt Gk
.
Although the optimal permutation matrices Jt and Jr can
7: Until reaching the maximal number of iterations or the
increment of the objective value is smaller than threshold be found by exhaustive search, we resort to the following
. sub-optimal method for reducing the search
8: Return the optimal phase shift matrix Θ e opt = Jr Θ̄Jt PM complexity.
and the optimal transmission beamforming vector wopt = We introduce the notation of g0 = M 1
m=1 |gm |, and
0 1
PK
(hH ΘG)
e H h = K k=1 |hk |. Then, we choose the permutation
.
khH ΘGk
e matrix Jt which sorts the column vector g0 in an ascend-
ing order, and the permutation matrix Jr , which sorts the
H
row vector h0 in an ascending order.
The problem of maximizing the achievable rate can be formu- After the permutation matrices Jt and Jr are determined,
lated as (P3.b2) becomes a non-convex quadratically constrained
Pt e H HeH H quadratic program (QCQP), and it can be solved by
(P3.a) max log2 I+ 2 HΘGWΛW G Θ H (40) defining Q = qqH , which needs to satisfy that Q  0 and
W,Θ,Λ
e σn
Rank(Qopt ) = 1. Since the rank-one constraint is non-
s.t. kwk k = 1, k = 1, 2, · · · , K,
convex, we relax this constraint [12] and (P3.b2) can be
λ1 + λ2 + · · · + λK = 1, translated into a standard convex SDR problem as follows
0 ≤ θM(i),i < 2π, i = 1, 2, · · · , N. K
! !
X
H
According to [12], the two-stage algorithm, which decouples (P3.b3) max Tr Φk Φk Q (44)
Q
the joint beamforming design problem (P3.a) into two sub- k=1

problems, has lower computational complexity, while suffer- s.t. Q  0,

ing from a slight performance erosion, when compared to |[Q]i,i | = 1, i = 1, 2, · · · , N.
alternating optimization algorithm. Therefore, we employ the The optimal solution, denoted as Qopt , in (P3.b3) can
two-stage algorithm for maximizing the achievable rate of our be found by using CVX [39]. If Rank(Qopt ) > 1,
proposed RIS architecture as follows. the optimal solution of q, denoted as qopt , can be
• Stage I: The RIS phase shift matrix is optimized by recovered from Qopt by eigenvalue decomposition as
√
maximizing the sum of the combined channel gain of qopt = ν1 q1 , in which ν1 is the largest eigenvalue of
all users, which can be expressed as the matrix Qopt , and q1 is the corresponding eigenvector.
XK After obtaining the permutation matrix Jt , Jr and qopt ,
(P3.b1) max khH k ΘGk
e 2
(41) the optimal phase shift matrix, denoted as Θ e opt , can be
Θ
e
H
k=1 derived as Θopt = Jr diag{qopt }Jt .
e
s.t. 0 ≤ θM(i),i < 2π, i = 1, 2, · · · , N. • Stage II: When the optimal phase shift matrix is ob-
We employ the SDR method for solving the problem tained in the first stage, the equivalent channel can be
(P3.b1). Specifically, let us define PK a column vector q = represented as Hequ = HH Θ e opt G. Upon using the
[ejθ1 , ejθ2 , · · · , ejθN ]H . Then k=1 kh He
k ΘGk 2
can be SVD method, the equivalent channel can be expressed
represented as as Hequ = UΣVH . The TBF matrix W is designed as
XK XK Wopt = V(1:K) , where V(1:K) represents the first K
khH k ΘGk =
e 2
khHk Jr Θ̄Jt Gk
2
columns of the right singular matrix V. Afterwards, the
k=1 k=1 power allocation matrix Λ is designed by the popular
XK water-filling method [40], based on the optimized equiv-
= kqH diag{hH
k Jr }Jt Gk
2
alent channel Hequ and the optimized TBF matrix Wopt .
k=1
K
Finally, the achievable rate can be represented as
 
Pt
X
H
= q Φk ΦH
kq RMIMO = log2 I + 2 Hequ CHequ , H
(45)
k=1 σn
 9

H
where C = Wopt ΛWopt is the transmit covariance matrix. we can get that [41]
π
r
E (ai ) = L 1 (−κg ) , (52)
4 (1 + κg ) 2
IV. T HEORETICAL A NALYSIS r
π
In this section, to highlight the channel gain enhancement E (bi ) = L 1 (−κh ) , (53)
4 (1 + κh ) 2
in our proposed method compared with the conventional
E a2i = E b2i = 1,



diagonal phase shift matrix method, we present the theoretical (54)
analysis of our proposed RIS architecture designed for the
SISO systems. Firstly, we analyze the scaling law of our 1) Average channel gain in the conventional RIS archi-
proposed RIS architecture relying on a non-diagonal phase tecture: In SISO systems, the average channel gain of the
matrix to derive the channel gain in Rician channels. Then, we conventional RIS architecture relying on the diagonal phase
discuss some special cases for different Rician factor values. shift matrix, denoted as gdiag , when considering Rician fading
Afterwards, we present the instantaneous SNR, as well as the channel, is given by
gdiag = E |hH Θ̄g|2


outage probability and the average BER, of our proposed non-
diagonal phase shift matrix based RIS systems in Rayleigh 
N
!2 
fading channels. Finally, the complexity comparison of the
X
= E |[g]i ||[hH ]i | 
conventional RIS architecture and our proposed RIS architec- i=1
ture is presented.  !2 
N
X
A. Channel Gain Analysis = %t %r E  ai bi 
i=1
To compare the fundamental limit of the conventional RIS N
architecture and of our proposed RIS architecture, we quantify
X
E a2i E b2i



= %t %r
the channel gain as a function of the number of RIS reflecting i=1
elements N . N
X −1 X N
!
In the SISO system, the BS-RIS channel vector and RIS- +2 E (ai ) E (bi ) E (aj ) E (bj ) . (55)
user channel vector are given by i=1 j=i+1
s
r
κg 1 According to (52), (53), (54) and (55), we can get
g= g+ g
e, (46) L21 (−κg ) L21 (−κh )
!
1 + κg 1 + κg π2
gdiag = %t %r N + N(N −1) 2 2
. (56)
r r 16 (κg + 1) (κh + 1)
H κh H 1 eH
h = hr + h , (47)
1 + κh 1 + κh r 2) Average channel gain analysis in the proposed RIS ar-
where κg and κh represent the Rician factors of BS-RIS path chitecture: Since a(1) , a(2) , · · · , a(N ) and b(1) , b(2) , · · · , b(N )
and RIS-user path, respectively. Since ai is the amplitudes of represent the sequences of a1 , a2 , · · · , aN and b1 , b2 , · · · , bN
√1 [g]i , ai follows the Rice distribution with noncentrality sorted in an ascending order, the average channel gain of the
%t q q
κg 1 proposed RIS architecture having a non-diagonal phase shift
parameter ν = 1+κ g
and scale parameter σ = 2(1+κg ) ,
matrix, denoted as gnond , can be expressed as
with the PDF and CDF as [41]
gnond = E |hH Jr Θ̄Jt g|2
 q 

2
fai (x) = 2 (1 + κg ) xe−(1+κg )x −κg I0 2 κg (1 + κg )x ,  !2 
X N
(48) = %t %r E  a(i) b(i) 
 q  i=1
p
Fai (x) = 1 − Q1 2κg , 2 (1 + κg )x , (49) N
X    
= %t %r E a2(i) E b2(i)
where I0 (·) is the modified Bessel function of the first kind i=1
with order zero, and Q1 (·) is the Marcum Q-function [41]. N −1 N
!
Similarly, since bi is the amplitude of √1%r [hH ]i , bi follows
X X




q +2 E a(i) E b(i) E a(j) E b(j) . (57)
κh
Rice distribution with noncentrality parameter ν = 1+κh
i=1 j=i+1
q
1
According to the order statistic theory [42], we can get the
and scale parameter σ = 2(1+κ h)
, with the PDF and CDF PDF of a(i) and b(i) as
as i−1 n−i
2
 p  fa(i) (x) = Ci,N [Fai (x)] [1 − Fai (x)] fai (x), (58)
fbi (x) = 2 (1 + κh ) xe−(1+κh )x −κh I0 2 κh (1 + κh )x ,
i−1 n−i
(50) fb(i) (x) = Ci,N [Fbi (x)] [1 − Fbi (x)] fbi (x), (59)
√ p  where fai (x), Fai (x), fbi (x) and Fbi (x) are given in (48),
Fbi (x) = 1 − Q1 2κh , 2 (1 + κh )x . (51) (49), (50) and (51), respectively, and Ci,N = (i−1)!(N N!
−i)! .
Since in a Rice distribution Rice (ν, σ), the
first moment and Therefore, the first moment of a(i) is given by
p Z ∞
second moment are σ π/2L1/2 −ν 2 /2σ 2 and ν 2 + 2σ 2 re-

E a(i) = xfa(i) (x)dx
spectively [41], in which L1/2 (·) is the Laguerre polynomial, 0
 10

Z ∞
i−1 n−i distribution having the rate parameter of λ = 1. There-
= xCi,N [Fai (x)] [1 − Fai (x)] fai (x)dx.
0 fore,
(60) √
π
Substituting (48) and (49) into (60), we can get E (ai ) = E (bi ) = , (67)
Z ∞ 2
Ci,N e−κg x2
x2 e− 2 I0 E a2i = E b2i = 1.

p



E a(i) = p 2κg x × (68)
2 (1 + κg ) 0
 p
i−1  p
N −i Substituting (67) and (68) into (55), we can get
1 − Q1 2κg , x Q1 2κg , x dx. (61) π2
 
gdiag = %t %r N + N (N − 1) . (69)
Similarly, we can get the second moment of a(i) , the first 16
moment of b(i) , and the second moment of b(i) as • Average channel gain in the proposed RIS architecture:
  Z ∞ Since ai and bi both follow Rayleigh distribution asso-
√
E a2(i) = x2 fa(i) (x)dx
0
ciated with the scaling parameter σ = 22 , a(i) and b(i)
Ci,N e−κg
Z ∞
x2
are also identically distributed. Therefore, (57) can be
x3 e− 2 I0
p

= 2κg x × simplified as
2 (1 + κg ) 0
 p
i−1  p
N −i XN   2
1 − Q1 2κg , x Q1 2κg , x dx, (62) gnond = %t %r E a2(i)
Z ∞ i=1


E b(i) = xfb(i) (x)dx −1
N N
!
X X

2

2
0 +2 E a(i) E a(j) . (70)
Z ∞
Ci,N e−κh x2 √
x2 e− 2 I0


=p 2κh x × i=1 j=i+1
2 (1 + κh ) 0 Note that a21 , a22 , · · · , a2N
independently follow the expo-
 √
i−1  √
N −i nential distribution having the rate parameter λ = 1, and
1 − Q1 2κh , x Q1 2κh , x dx, (63)
  Z ∞ given the lemma in [42] that
E b2(i) = x2 fb(i) (x)dx Xi
Zk
0 a2(i) = , i = 1, 2, · · · , N, (71)
Ci,N e−κh ∞ 3 − x2
Z
√ N −k+1

k=1
= x e 2 I0 2κh x × where the random variable Zk follows the exponential
2 (1 + κh ) 0
 √
i−1  √
N −i distribution with the rate parameter λ = 1, we arrive at:
1 − Q1 2κh , x Q1 2κh , x dx. (64) !2
N   2 X N i
Substituting (61), (62), (63) and (64) into (57), we can get the
X X 1
E a2(i) = . (72)
average channel gain of the proposed RIS architecture gnond . i=1 i=1 k=1
N −k+1
B. Effect of the Value of the Rician Factor on the Channel PN −1 PN

2

2
To derive i=1 j=i+1 E a(i) E a(j) , firstly,
Gain we present the following Theorem 1.
To get deep insights on the effects of the Rician factors
κg and κh on the channel gain, we investigate the following Theorem 1. The PDF of a(i) is given by a linear com-
cases. bination of the PDF of i Rayleigh distributions having
1
Case I, κg → ∞ (or κh → ∞): In this case, the BS-RIS the scaling parameter σ = √ . Specifically,
2(N −i+k+1)
channel (or RIS-user channel) is fully dominated by the LoS the PDF of a(i) can be written as
path, the conventional RIS architecture and our proposed RIS i−1   
X N N −i+k
architecture get the same average channel gain as fa(i) (x) = (−1)k
i−k−1 k
gdiag = gnond . (65) k=0
!
Proof: See Appendix A. 1
fX x; σ = p , (73)
Furthermore, when κg → ∞ and κh → ∞ simultaneously, 2(N − i + k + 1)
both the conventional RIS architecture and our proposed RIS where fX (x; σ) is the PDF of a random variable X
architecture gets the channel gain upper bound as following the Rayleigh distribution having the scaling
gdiag = gnond = %t %r N 2 . (66) parameter σ.
This can be easily observed since ai = 1 and bi = 1 for all Proof: See Appendix B.
i = 1, 2, · · · , N when κg → ∞ and κh → ∞. Since the mean of fX (x; σ) equals σ π2 , then
p
Case II, κg = 0 and κh = 0: In this case, the BS-RIS !
channel and RIS-user channel experience Rayleigh fading. 1
E fX (x; σ = p )
The average channel gain of the conventional RIS architecture 2(N − i + k + 1)
gdiag and the proposed non-diagonal RIS architecture gnond √
π 1
are derived as follows. = √ . (74)
2 N −i+k+1
• Average channel gain in the conventional RIS architec-
Upon combining (73) and (74), we arrive at:
ture: When κg = 0 and κh = 0, ai and bi both follow √ X i−1   
the Rayleigh distribution associated with the scaling   π N N −i+k
√ E fa(i) (x) =
parameter σ = 22 , and a2i and b2i obey the exponential 2 i−k−1 k
k=0
 11

1 1
(−1)k √ . (75)
N −i+k+1

2 
2

Normalized channel gain

PN −1 PN 
Therefore, i=1 j=i+1 E a(i) E a(j) is ob- 0.9

tained as
N −1 X N N −1 N
X

2

2 π2 X X 0.8
E a(i) E a(j) =
i=1 j=i+1
16 i=1 j=i+1
" i−1    #2 0.7
X
k N N −i+k 1
(−1) √
i−k−1 k N −i+k+1
k=0 0.6
"j−1 #2 0 4 8 12 16 20 24 28 32 36 40 44 48 52
  
X
k N N −j +k 1 Fig. 5. Comparison of normalized channel gain versus the number of RIS
(−1) √ . elements N for the different RIS architectures.
j −k−1 k N −j +k+1 25 25
k=0
(76) 20 20

According to (70), (72) and (76), we get the theoretical 15 15

result of the channel gain gnond as 10 10

κg (dB)

κg (dB)
N i
!2 N −1 N
5 5
X X 1 π2 X X 0 0
gnond = %t %r +%t %r
i=1 k=1
N −k+1 8 i=1 j=i+1 -5 -5

" i−1  #2 -10 -10

 
X N N −i+k 1 -15 -15
(−1)k √ -15 -10 -5 0 5
κh (dB)
10 15 20 25 -15 -10 -5 0 5
κh (dB)
10 15 20 25

i−k−1 k N −i+k+1 Normalized channel gain of conventional RIS architecture Normalized channel gain of non-diagonal RIS architecture
k=0
"j−1    #2 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
X N N −j +k k 1 Fig. 6. Comparison of normalized channel gain versus BS-RIS Rician factors
(−1) √ . κg and RIS-user Rician factors κh for the conventional RIS architecture and
j −k−1 k N −j +k+1 our proposed RIS architectures.
k=0
(77)
Upon comparing (55) to (57), we can find that the channel on the MRC criterion, when the number of elements N
gain of the conventional RIS architecture is proportional is large.
to the expectation of the square of a1 b1 + a2 b2 + · · · + C. Channel Gain Analysis when N → ∞
aN bN , while the channel gain of our proposed RIS archi- In the conventional RIS architecture, when N → ∞,
tecture is proportional to the expectation of the square of according to (56), we can get the average channel gain as
a(1) b(1) +a(2) b(2) +· · ·+a(N ) b(N ) . Therefore, in conven-
π2 L21 (−κg ) L21 (−κh )
tional RIS architectures the channel gain is proportional gdiag = %t %r N 2 2 2
. (78)
to the equal gain combination of the channel parameters, 16 (κg + 1) (κh + 1)
while in our proposed RIS architecture the channel gain In our proposed RIS architecture, when N → ∞, the
is proportional to the maximum ratio combination of the average channel gain is given by
channel parameters, when the number of RIS elements gnond = %t %r N 2 , (79)
N is large.
which is equivalent to the channel gain performance of the
We define the normalized channel gain b g as the average
fully-connected RIS architecture of [24].
channel gain g normalized by the upper bound of %t %r N 2
Proof: See Appendix C.
in LoS channels, i.e. b g = %t %gr N 2 . In Fig. 5, we present the
Fig. 6 compares the normalized channel gain versus BS-
normalized power gain of our proposed RIS architecture,
RIS Rician factors κg and RIS-user Rician factors κh for the
compared to that of the fully-connected/group-connected
conventional RIS architecture and our proposed RIS architec-
RIS architecture of [24] in Rayleigh fading channels, i.e.
tures. It shows that when N → ∞, the normalized channel
κg = κh = 0, where G represents the group size of
gain of the conventional RIS architecture degrades with the
a group-connected RIS architecture. Fig. 5 shows that
decrease of the Rician factors, while that of our proposed RIS
when the number of RIS elements is small, the power
architecture remains at 1 in all Rician factor ranges, which
gain of our proposed RIS architecture is worse than that
means that our proposed RIS architecture is more robust over
of the fully connected RIS architecture of [24]. How-
a wider range of propagation conditions and especially shows
ever, the performance of our proposed RIS architecture
advantages in NLoS-dominated channel environments.
becomes better than that of the group-connected RIS
architecture and approaches that of the fully-connected D. Outage Probability and BER Performance Analysis
RIS architecture upon increasing the number of RIS Since the above analysis demonstrates that our proposed
reflecting elements N . However, regardless of the number RIS architecture shows advantages in NLoS-dominated chan-
of RIS elements N , our RIS architecture outperforms the nel environments, in this section we derive the distribution
conventional RIS architecture. This gain is due to the fact of the received SNR of our proposed RIS architecture in
that in our RIS architecture, the BS-RIS channel vector g Rayleigh fading channels, then the outage probability and BER
and the RIS-user channel vector hH are combined based performance are analyzed.
 12

According to (88), we can show that the upper bound of the while for the fully-connected RIS architecture it is N (N2+1) and
instantaneous SNR at the receiver side of our proposed RIS for the group-connected RIS architecture it is N (G+1)
2 . In our
architecture can be expressed as proposed non-diagonal phase shift matrix scheme, the number
N
! N !
X X of information values transmitted in BS-RIS control link is
Ω=ρ a2i b2i , (80) 2N , i.e., N optimized non-zeros values and their positions in
i=1 i=1 the non-diagonal phase shift matrix. Fig. 7 (b) compares the
where we have ρ = σP2t %t %r . Since a2i and b2i both obey the BS-RIS control link load for the conventional RIS structure,
n
exponential distribution
PN having
2
PN 2 of λ = 1,
the rate parameter the fully-connected/group-connected RIS structure and our
then both Ya = i=1 ai and Y b = i=1 bi follow the RIS structure, which shows that the BS-RIS control link
gamma distribution associated with the shape parameter N load of our RIS architecture is lower than that of the fully-
and the scale parameter 1. Let us introduce Z = Ya Yb , with connected and of the group-connected RIS architecture for
the following PDF [43] G ≥ 4. However, it has been presented in Fig. 5 that the
2 √ performance of our proposed RIS structure is better than that
fZ (z) = 2 z N −1 K0 (2 z), (81)
Γ (N ) of the group-connected method and approaches that of the
where Γ(·) is the gamma function, and K0 (·) is the modified fully-connected RIS structure, as the number of RIS elements
Bessel function of the second kind. Therefore, the CDF of the increases.
received SNR Ω is given by
Z ωρ  √
2 V. P ERFORMANCE RESULTS AND ANALYSIS
FΩ (ω) = 2
tN −1 K0 2 t dt. (82)
0 Γ (N ) In this section, firstly the outage probability and average
1) Outage probability: If the SNR threshold is ωth , then the BER of our proposed RIS architecture are presented. Then,
outage probability, denoted as Pout (ωth ), can be calculated as we characterize the achievable rate of our proposed RIS
[40] architecture having non-diagonal phase shift matrices.
Z ωρth  √
2 N −1 In Fig. 8 (a), we show the theoretical lower bound of the
Pout (ωth ) = Pr(Ω ≤ ωth ) = t K0 2 t dt.
0 Γ2 (N ) outage probability of our proposed RIS architecture in (83)
(83) versus the number of RIS elements, where the SNR threshold
2) Average BER: Based on the CDF of the received SNR ωth is set to 25dB. Furthermore, the simulation results of
in (82), the average BER is given by [44] our proposed RIS architecture and of the conventional RIS
Z ∞
qp architecture are presented. Note that the theoretical lower
Pe = e−qω ω p−1 FΩ (ω)dω bound is very tight, and the performance of our proposed RIS
2Γ(p) 0
Z ∞ Z ωρ  √ architecture is significantly better than that of the conventional
qp −qω p−1
= 2
e ω dω tN −1 K0 2 t dt, RIS architecture for all values of N . Then, in Fig. 8 (b) we
Γ(p)Γ (N ) 0 0 show that our simulation results and the theoretical average
(84)
BER of our proposed RIS architecture based on (84) match
where the parameters p and q are different for different well. The results are contrasted to those using the conventional
modulation schemes. For example, we have p = 21 and q = 1 RIS architecture and to those of the LoS channel, where
for binary phase shift keying (BPSK) modulation [44]. BPSK modulation is employed. Observe that the average BER
performance of our proposed RIS architecture is better than
E. Complexity Analysis
that of the conventional RIS architecture, and tends to that of
In this section, we analyze the complexity of the RIS ar- the LoS channels upon increasing the number of RIS elements.
chitecture in terms of the number of configurable impedances In the following simulation, the number of antennas at the
required and the BS-RIS control link load. BS is M = 4, the BS-RIS distance dt is fixed as dt = 50m,
Firstly, the conventional RIS architecture requires N con- the path loss at the reference distance 1 meter is C0 = −30dB,
figurable impedances, while for the fully-connected RIS ar- the total transmit power is Pt = 50mW, and the noise power
chitecture we need N (N2+1) and for the group-connected RIS is σn2 = −90dBm. The Rician factors are κG = κh1 = κh2 =
architecture we need N (G+1)
2 . By contrast, our architecture · · · = κhK = −10dB, unless otherwise specified.
only requires N configurable impedances, which is the same as Firstly, the achievable rate of single-user MISO systems is
that in the conventional RIS architecture. Fig. 7 (a) compares presented based on (37) in Fig. 9 (a), where we compare
the number of configurable impedances required for the con- the achievable rate versus RIS-user distance dr of both the
ventional RIS structure, the fully-connected/group-connected conventional RIS and of our proposed RIS structure. The
RIS structure and our RIS structure, which shows that the path loss exponent is αt = αr = 2.2. Observe that the
number of configurable impedances required by our proposed achievable rate decreases upon the increasing the RIS-user
RIS architecture is lower than that of the fully-connected RIS distance due to the increased path loss. We can also find
architecture and of the group-connected RIS architecture. that the achievable rate of our proposed RIS architecture is
Secondly, in terms of the BS-RIS control link load, the num- higher than that of the conventional RIS architecture for all
ber of information values transmitted on the BS-RIS control RIS-user distances. In Fig. 9 (b), we compare the achievable
link is N for the conventional RIS architecture, since only rate versus the number of RIS elements N in the conventional
the diagonal values of the phase shift matrix are optimized, RIS architecture and our proposed RIS architecture, where the
 13

250 250
Conventional phase shift matrix Conventional phase shift matrix
Proposed non-diagonal phase shift matrix Proposed non-diagonal phase shift matrix
Number of configurable impedances
Fully-connected phase shift matrix Fully-connected phase shift matrix
200 Group-connected phase shift matrix (G=2) 200 Group-connected phase shift matrix (G=2)

BS-RIS control link load

Group-connected phase shift matrix (G=4) Group-connected phase shift matrix (G=4)
Group-connected phase shift matrix (G=6) Group-connected phase shift matrix (G=6)
150 Group-connected phase shift matrix (G=8) 150 Group-connected phase shift matrix (G=8)

100 100

50 50

0 0
0 10 20 30 40 50 0 10 20 30 40 50
N N
(a) (b)
Fig. 7. (a) Comparison of the number of configurable impedances versus the number of RIS elements N for the different RIS architectures. (b) Comparison
of the BS-RIS control link load versus the number of RIS elements N for the different RIS architectures.
100 100 Conventional, theory
Conventional, simumation
Non-diagonal, theoretical lower bound
Non-diagonal, simulation
ρ = −10dB LoS channel, theory
ρ = −5dB
10-1 10-1
Outage Probability

Average BER
ρ = 0dB

10-2 10-2
ρ = 5dB N =8
N = 16

ρ = 10dB N = 32
10-3 Conventional, simulation 10-3
Non-diagonal, simulation
Non-diagonal, theoretical lower bound

10-4 10-4
0 8 16 24 32 40 48 56 64 -36 -32 -28 -24 -20 -16 -12 -8
N ρ (dB)
(a) (b)
Fig. 8. (a) Theoretical lower bound and simulation results of outage probability versus the number of RIS elements N in conventional RIS structure and
our proposed RIS structure. (b) Theoretical analysis and simulation results of average BER versus ρ of conventional RIS architecture, our proposed RIS
architecture and LoS channel.
14 5 5
Conventional, N =16 Conventional, αt = αr = 2.2
Non-diagonal, N =16 Non-diagonal, αt = αr = 2.2
12 Conventional, N =32 Conventional, αt = αr = 2.4
Non-diagonal, N =32 4 Non-diagonal, αt = αr = 2.4 4
Achievable rate (bps/Hz)

Achievable rate (bps/Hz)

Conventional, αt = αr = 2.6
Achievable rate (bps/Hz)

10 Conventional, N =64
Non-diagonal, N =64 Non-diagonal, αt = αr = 2.6
3 3
8

6
2 2
4

1 1
2

0 0
0 6 12 18 24 30 0
0 8 16 24 32 40 48 56 64 0 5 10 15 20
dr (meter) N Number of iterations
(a) (b) (c)
Fig. 9. Comparison of the achievable rate in conventional RIS architecture and our proposed RIS architecture in single-user MISO systems: (a) achievable rate
versus RIS-user distance dr , (b) achievable rate versus the number of RIS elements N , and (c) achievable rate versus the number of iterations in alternating
optimization method.

RIS-user distance is dr = 30m. Observe that the performance RIS with radius RD . Explicitly, we compare the achievable
of our proposed RIS architecture is better than that of the rate versus the number of RIS elements N of the conventional
conventional RIS architecture, especially when the number and of our proposed RIS architecture. In Fig. 10 (a), the path
of RIS elements N is large. This is due to the fact that loss exponent is αt = αr = 2.2, and the number of users is
our proposed RIS architecture benefits from using the MRC K = 2. It can be seen that compared to the conventional RIS
criterion upon increasing the number of RIS elements. In architecture, our proposed RIS architecture has better coverage
Fig. 9 (c), the achievable rate versus the number of iterations under the same data rate requirement. In Fig. 10 (b), the path
in alternating optimization algorithm is presented, where the loss exponent is αt = αr = 2.2, and the radius is RD =
path loss exponent αt = αr = 2.2 and RIS-user distance 30m, where it can be seen that the achievable rate increases
is dr = 30m. It shows that the convergence speed in our near linearly with the number of RIS elements. Compared to
proposed RIS architecture is almost the same as that of the the conventional RIS architecture, the achievable rate of our
conventional RIS architecture. proposed method is significantly high for any number of users.
In Fig. 10 (c), the number of users is K = 2, and the radius
Then, the simulation based on achievable rate of the multi- is RD = 30m. We can observe the performance enhancement
user MIMO systems is presented in Fig. 10, where the users of our proposed RIS architecture when the number of RIS
are distributed uniformly within a half circle centered at the
 14

9 9 8
Conventional, αt = αr = 2.2
8 Non-diagonal, αt = αr = 2.2
8 7 Conventional, αt = αr = 2.4
7 Non-diagonal, αt = αr = 2.4

Achievable rate (bps/Hz)

7

Achievable rate (bps/Hz)

Achievable rate (bps/Hz)

6 Conventional, αt = αr = 2.6
6 Non-diagonal, αt = αr = 2.6
6
5
5 5
4
4 4
3
3 Conventional, RD = 20m 3 Conventional, K = 1
Non-diagonal, RD = 20m Non-diagonal, K = 1
2 Conventional, RD = 30m 2
2 Conventional, K = 2
Non-diagonal, RD = 30m Non-diagonal, K = 2
1 Conventional, RD = 40m 1 Conventional, K = 4 1
Non-diagonal, RD = 40m Non-diagonal, K = 4
0 0 0
0 8 16 24 32 40 48 56 64 0 8 16 24 32 40 48 56 64 0 8 16 24 32 40 48 56 64
N N N
(a) (b) (c)
6 8 9

7 8
5
7

Achievable rate (bps/Hz)

Achievable rate (bps/Hz)

Achievable rate (bps/Hz)

6
4 6
5
5
3 4
Conventional, σh2 = 0 4 Conventional, κG = −10dB
Non-diagonal, σh2 = 0 3 Non-diagonal, κG = −10dB
2 Conventional, σh2 = 0.1 Conventional, δ0 = 4λ 3 Conventional, κG = −3dB
Non-diagonal, σh2 = 0.1 2 Non-diagonal, δ0 = 4λ Non-diagonal, κG = −3dB
Conventional, σh2 = 0.2 Conventional, δ0 = 41 λ 2 Conventional, κG = 0dB
1 Non-diagonal, σh2 = 0.2 Non-diagonal, δ0 = 14 λ Non-diagonal, κG = 0dB
1 Conventional, δ0 = 641
λ 1
Conventional, σh2 = 0.4 Conventional, κG = 10dB
1
Non-diagonal, σh2 = 0.4 Non-diagonal, δ0 = 64 λ Non-diagonal, κG = 10dB
0 0 0
0 8 16 24 32 40 48 56 64 0 8 16 24 32 40 48 56 64 0 8 16 24 32 40 48 56 64
N N N
(d) (e) (f)
Fig. 10. Comparison of the achievable rate versus the number of RIS elements N in conventional RIS structure and our proposed RIS structure in multi-user
MIMO systems: (a) with the user radius RD = 20m, 30m, 40m, (b) with the number of users K = 1, 2, 4, (c) with the path loss exponent α = 2.2, 2.4, 2.6,
2 = 0, 0.1, 0.2, 0.4, (e) with the distance between adjacent RIS elements δ = 4λ, 1 λ, 1 λ, and (f) with Rician
(d) with channel estimation error variance σh 0 4 64
factor αG = −10dB, −3dB, 0dB, 10dB.
1
elements N is large for any pathloss exponent value. distance δ0 = 64 λ is almost the same as that in conventional
When considering imperfect CSI, referring to [45], the RIS architecture with adjacent RIS element distance δ0 = 14 λ,
estimated BS-RIS and RIS-user channels are given by ĝm = and the achievable rate in our proposed RIS architecture with
gm + gm,e and ĥH H H adjacent RIS element distance δ0 = 41 λ is even higher than that
k = hk + hk,e , respectively, where m =
1, 2, · · · , M , k = 1, 2, · · · , K. The channel estimation error in conventional RIS architecture with adjacent RIS element
distance δ0 = 4λ.



component obeys gm,e ∼ CN 0, σh2 I , hH 2
k,e ∼ CN 0, σh I ,
where σh2 represents the estimation error variance. In the Finally, we investigate the impact of the Rician factor on
linear minimum mean squared error (LMMSE) method, we the achievable rate in Fig. 10 (f), where the path loss exponent
have σh2 = 1+τ1p ρt , where ρt is the pilot transmission power is αt = αr = 2.2, the number of users is K = 2, the radius
and τp represents the pilot symbols transmission duration is RD = 30m. We also assume the Rician factor of RIS-user
[46]. In Fig. 10 (d), the achievable rate is presented when path is κhr = −20dB for k = 1, 2, · · · , K, while the Rician
considering imperfect CSI, where the path loss exponent is factor of BS-RIS path κG can range between −10dB to 10dB.
αt = αr = 2.2, the number of users is K = 2 and the With the increase of the Rician factor κG , the achievable rate
radius is RD = 30m. As shown in Fig. 10 (d), our proposed of the conventional RIS architecture and that of our proposed
RIS architecture outperforms the conventional RIS architecture RIS architecture can both achieve better performance, which is
under any channel estimation error variance σh2 , which shows due to the reduced impact of channel fading with the increase
the robustness of our proposed architecture compared to the of LoS component. Besides, it shows that the achievable rate
conventional one. enhancement of our proposed RIS architecture is not obvious
in the high range of Rician factor, while it is significant when
In most of the previous contributions [12]–[15], [17]–
the channel is dominated by NLoS component, which agrees
[24], it is assumed that the distance between the adjacent
with the above theoretical analysis.
RIS elements is large enough to ensure that the BS-RIS
and RIS-user channels are uncorrelated. To make our work
more realistic, in Fig. 10 (e), we employ the exponential VI. C ONCLUSIONS
correlation channel model of [47] to investigate the effect of In this paper, we proposed a novel RIS architecture, where
channel correlation on the achievable rate, where the path the signal impinging on a specific element can be reflected by
loss exponent is αt = αr = 2.2, the number of users is another element after phase shift adjustment. Compared to the
K = 2, the radius is RD = 30m and the reference correlation conventional RIS architecture, our proposed RIS architecture
distance dref = λ. As shown in Fig. 10 (e), our proposed provides better channel gain as a benefit of using the MRC
RIS architecture attains a better performance than that of the criterion. The theoretical analysis showed that our system
conventional RIS architecture. For example, the achievable rate performs better than the conventional RIS system both in
in our proposed RIS architecture with adjacent RIS element terms of its average channel gain, outage probability and
 15

!
average BER. Furthermore, the performance of our proposed 1
fXx; σ = p . (87)
RIS architecture is better than that of the group-connected 2(N − i + k + 1)
RIS architecture and approaches that of the fully-connected The proof is thus completed.
RIS architecture, with increasing the number of RIS elements,
despite its considerably reduced complexity compared to the A PPENDIX C
fully-connected architecture. Furthermore, we formulated and P ROOF OF FORMULA (79)
solved the problem of maximizing the achievable rate of our In (57), according to Cauchy-Schwarz inequality, we can
proposed RIS architecture by jointly optimizing the active TBF get
and the non-diagonal phase shift matrix of both single-user !2 

XN
MISO systems and of multi-user MIMO systems based on gnond = %t %r E  a(i) b(i) 
alternating optimization and on SDR methods, respectively. i=1
The simulation results showed that our proposed technique N
! N
!!
is capacity of enhancing the achievable rate of RIS-assisted
X X
≤ %t %r E a2(i) b2(i) , (88)
wireless communications. i=1 i=1
where the equality in (88) is established only when a(i) = b(i) .
A PPENDIX A When N → ∞, we can get
   
P ROOF OF FORMULA (65) i −1 i
a(i) = b(i) = Fa−1 = F bi , (89)
When κg → ∞, according to (46), we can get that ai = 1
i
N N
(i = 1, 2, · · · , N ). Therefore, where Fa−1 i
(·) and Fb−1
i
(·) represent the inverse CDF of a(i)
XN N
X N
X and b(i) , respectively. Based on (88) and (89), we can show
ai bi = a(i) b(i) = bi . (85) that when N → ∞, the channel gain of our proposed RIS
i=1 i=1 i=1 architecture is
According to (55), (57) and (85), we can show that gdiag = N
! N !!
X X
gnond . When we consider the case that κh → ∞, we can gnond = %t %r E a2(i) b2(i)
similarly get that gdiag = gnond . The proof is thus completed. i=1
!
i=1
!
N
X N
X
= %t %r E a2i E b2i
A PPENDIX B i=1 i=1
P ROOF OF Theorem 1
= %t %r N 2 . (90)
We denote the CDF of the random variable X following the The proof is thus completed.
Rayleigh distribution with the scale parameter σ as FX (x; σ).
The PDF of a(i) is given by [42] R EFERENCES
√ !" √ !#i−1 [1] Z. Ma, M. Xiao, Y. Xiao, Z. Pang, H. V. Poor, and B. Vucetic,
2 2 “High-reliability and low-latency wireless communication for Internet
fa(i) (x) = Ci,N fX x; σ = FX x; σ = × of Things: challenges, fundamentals, and enabling technologies,” IEEE
2 2
Internet of Things Journal, vol. 6, no. 5, pp. 7946–7970, 2019.
" √ !#N −i [2] H. Yetgin, K. T. K. Cheung, M. El-Hajjar, and L. H. Hanzo, “A survey of
2 network lifetime maximization techniques in wireless sensor networks,”
1 − FX x; σ =
2 IEEE Communications Surveys & Tutorials, vol. 19, no. 2, pp. 828–854,
2017.
2 i−1 2 N −i
   
2 [3] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski,
= 2Ci,N xe−x 1 − e−x e−x “Five disruptive technology directions for 5G,” IEEE communications
magazine, vol. 52, no. 2, pp. 74–80, 2014.
2 i−1 2 N −(i−1)
   
= 2Ci,N x 1 − e−x e−x . (86) [4] M. A. ElMossallamy, H. Zhang, L. Song, K. G. Seddik, Z. Han, and
G. Y. Li, “Reconfigurable intelligent surfaces for wireless communica-
2 i−1
 −x2 k tions: Principles, challenges, and opportunities,” IEEE Transactions on
  Pi−1
In (86), 1 − e−x = k=0 i−1 k −e , so Cognitive Communications and Networking, vol. 6, no. 3, pp. 990–1002,
2020.
i−1  
2 N −(i−1) X i − 1  −x2 k [5] M. Di Renzo, A. Zappone, M. Debbah, M.-S. Alouini, C. Yuen,
 
fa(i) (x) = 2Ci,N x e−x −e J. de Rosny, and S. Tretyakov, “Smart radio environments empowered by
k reconfigurable intelligent surfaces: How it works, state of research, and
k=0
i−1   the road ahead,” IEEE Journal on Selected Areas in Communications,
X i − 1  −x2 N −(i−1)+k vol. 38, no. 11, pp. 2450–2525, 2020.
= 2Ci,N x −e [6] S. Gong, X. Lu, D. T. Hoang, D. Niyato, L. Shu, D. I. Kim, and Y.-C.
k Liang, “Toward smart wireless communications via intelligent reflecting
k=0
i−1   surfaces: A contemporary survey,” IEEE Communications Surveys &
X i−1 1 Tutorials, vol. 22, no. 4, pp. 2283–2314, 2020.
= Ci,N (−1)k × [7] C. Pan, H. Ren, K. Wang, W. Xu, M. Elkashlan, A. Nallanathan,
k N −i+k+1
k=0 and L. Hanzo, “Multicell MIMO communications relying on intelligent
!
reflecting surfaces,” IEEE Transactions on Wireless Communications,
1 vol. 19, no. 8, pp. 5218–5233, 2020.
fX x; σ = p
2(N − i + k + 1) [8] L. Dai, B. Wang, M. Wang, X. Yang, J. Tan, S. Bi, S. Xu, F. Yang,
Z. Chen, M. Di Renzo, C.-B. Chae, and L. Hanzo, “Reconfigurable
i−1   
X N N −i+k intelligent surface-based wireless communications: Antenna design, pro-
= (−1)k × totyping, and experimental results,” IEEE Access, vol. 8, pp. 45 913–
i−k−1 k 45 923, 2020.
k=0
 16

[9] T. Hou, Y. Liu, Z. Song, X. Sun, Y. Chen, and L. Hanzo, “Reconfigurable rectangular array based massive MIMO systems with low complexity,”
intelligent surface aided NOMA networks,” IEEE Journal on Selected IEEE Transactions on Vehicular Technology, vol. 68, no. 3, pp. 2545–
Areas in Communications, vol. 38, no. 11, pp. 2575–2588, 2020. 2556, 2019.
[10] Z. Zhou, N. Ge, Z. Wang, and L. Hanzo, “Joint transmit precoding and [27] S. Abeywickrama, R. Zhang, Q. Wu, and C. Yuen, “Intelligent reflecting
reconfigurable intelligent surface phase adjustment: A decomposition- surface: Practical phase shift model and beamforming optimization,”
aided channel estimation approach,” IEEE Transactions on Communi- IEEE Transactions on Communications, vol. 68, no. 9, pp. 5849–5863,
cations, vol. 69, no. 2, pp. 1228–1243, 2020. 2020.
[11] X. Cao, B. Yang, C. Huang, C. Yuen, M. Di Renzo, Z. Han, D. Niyato, [28] Y. Liu, X. Liu, X. Mu, T. Hou, J. Xu, M. Di Renzo, and N. Al-Dhahir,
H. V. Poor, and L. Hanzo, “AI-assisted MAC for reconfigurable intel- “Reconfigurable intelligent surfaces: Principles and opportunities,” IEEE
ligent surface-aided wireless networks: Challenges and opportunities,” Communications Surveys & Tutorials, 2021.
IEEE Communications Magazine, vol. 59, no. 6, pp. 21–27, 2021. [29] G. M. Rebeiz and J. B. Muldavin, “RF MEMS switches and switch
[12] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless circuits,” IEEE Microwave magazine, vol. 2, no. 4, pp. 59–71, 2001.
network via joint active and passive beamforming,” IEEE Transactions
[30] V. Laur, J. Gouavogui, and B. Balde, “C-band hybrid 3-D-printed
on Wireless Communications, vol. 18, no. 11, pp. 5394–5409, 2019.
microwave isolator,” IEEE Transactions on Microwave Theory and
[13] B. Ning, Z. Chen, W. Chen, and J. Fang, “Beamforming optimization Techniques, vol. 69, no. 3, pp. 1579–1585, 2021.
for intelligent reflecting surface assisted MIMO: A sum-path-gain max-
[31] S. Yang, D. Vincent, J. R. Bray, and L. Roy, “Study of a ferrite LTCC
imization approach,” IEEE Wireless Communications Letters, vol. 9,
multifunctional circulator with integrated winding,” IEEE Transactions
no. 7, pp. 1105–1109, 2020.
on Components, Packaging and Manufacturing Technology, vol. 5, no. 7,
[14] P. Wang, J. Fang, X. Yuan, Z. Chen, and H. Li, “Intelligent reflect- pp. 879–886, 2015.
ing surface-assisted millimeter wave communications: Joint active and
[32] J. Wang, A. Yang, Y. Chen, Z. Chen, A. Geiler, S. M. Gillette, V. G.
passive precoding design,” IEEE Transactions on Vehicular Technology,
Harris, and C. Vittoria, “Self-biased Y-junction circulator at Ku band,”
vol. 69, no. 12, pp. 14 960–14 973, 2020.
IEEE Microwave and Wireless Components Letters, vol. 21, no. 6, pp.
[15] Y. Han, W. Tang, S. Jin, C.-K. Wen, and X. Ma, “Large intelligen- 292–294, 2011.
t surface-assisted wireless communication exploiting statistical CSI,”
[33] W. D’orazio, K. Wu, and J. Helszajn, “A substrate integrated waveguide
IEEE Transactions on Vehicular Technology, vol. 68, no. 8, pp. 8238–
degree-2 circulator,” IEEE Microwave and Wireless Components Letters,
8242, 2019.
vol. 14, no. 5, pp. 207–209, 2004.
[16] J. Wang, H. Wang, Y. Han, S. Jin, and X. Li, “Joint transmit beamform-
[34] S. Adhikari, A. Ghiotto, S. Hemour, and K. Wu, “Tunable non-reciprocal
ing and phase shift design for reconfigurable intelligent surface assisted
ferrite loaded SIW phase shifter,” in 2013 IEEE MTT-S International
MIMO systems,” IEEE Transactions on Cognitive Communications and
Microwave Symposium Digest (MTT). IEEE, 2013, pp. 1–3.
Networking, vol. 7, no. 2, pp. 354–368, 2021.
[35] A. Li, S. Singh, and D. Sievenpiper, “Metasurfaces and their applica-
[17] Q. Wu and R. Zhang, “Beamforming optimization for wireless network
tions,” Nanophotonics, vol. 7, no. 6, pp. 989–1011, 2018.
aided by intelligent reflecting surface with discrete phase shifts,” IEEE
[36] C. Liaskos, G. G. Pyrialakos, A. Pitilakis, A. Tsioliaridou,
Transactions on Communications, vol. 68, no. 3, pp. 1838–1851, 2019.
M. Christodoulou, N. Kantartzis, S. Ioannidis, A. Pitsillides, and I. F.
[18] S. Zhang, H. Zhang, B. Di, Y. Tan, Z. Han, and L. Song, “Beyond Akyildiz, “The internet of metamaterial things and their software en-
intelligent reflecting surfaces: Reflective-transmissive metasurface aided ablers,” ITU Journal on Future and Evolving Technologies, 2020.
communications for full-dimensional coverage extension,” IEEE Trans-
[37] S. Boyd, S. P. Boyd, and L. Vandenberghe, Convex optimization.
actions on Vehicular Technology, vol. 69, no. 11, pp. 13 905–13 909,
Cambridge university press, 2004.
2020.
[38] X.-D. Zhang, Matrix analysis and applications. Cambridge University
[19] W. Chen, X. Ma, Z. Li, and N. Kuang, “Sum-rate maximization for
Press, 2017.
intelligent reflecting surface based Terahertz communication systems,”
[39] M. Grant, S. Boyd, and Y. Ye, “CVX: Matlab software for disciplined
in 2019 IEEE/CIC International Conference on Communications Work-
convex programming,” 2008.
shops in China (ICCC Workshops). IEEE, 2019, pp. 153–157.
[40] D. Tse and P. Viswanath, Fundamentals of wireless communication.
[20] J. Xu, W. Xu, and A. L. Swindlehurst, “Discrete phase shift design for
Cambridge university press, 2005.
practical large intelligent surface communication,” in 2019 IEEE Pacific
[41] A. Abdi, C. Tepedelenlioglu, M. Kaveh, and G. Giannakis, “On the
Rim Conference on Communications, Computers and Signal Processing
estimation of the K parameter for the rice fading distribution,” IEEE
(PACRIM). IEEE, 2019, pp. 1–5.
Communications letters, vol. 5, no. 3, pp. 92–94, 2001.
[21] S. Lin, B. Zheng, G. C. Alexandropoulos, M. Wen, M. Di Renzo,
[42] A. Rényi, “On the theory of order statistics,” Acta Mathematica A-
and F. Chen, “Reconfigurable intelligent surfaces with reflection pattern
cademiae Scientiarum Hungarica, vol. 4, no. 3-4, pp. 191–231, 1953.
modulation: Beamforming design and performance analysis,” IEEE
Transactions on Wireless Communications, vol. 20, no. 2, pp. 741–754, [43] C. S. Withers and S. Nadarajah, “On the product of gamma random
2020. variables,” Quality & Quantity, vol. 47, no. 1, pp. 545–552, 2013.
[22] E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M.-S. Alouini, and [44] I. S. Ansari, S. Al-Ahmadi, F. Yilmaz, M.-S. Alouini, and
R. Zhang, “Wireless communications through reconfigurable intelligent H. Yanikomeroglu, “A new formula for the BER of binary modula-
surfaces,” IEEE Access, vol. 7, pp. 116 753–116 773, 2019. tions with dual-branch selection over generalized-K composite fading
channels,” IEEE Transactions on Communications, vol. 59, no. 10, pp.
[23] L. Yang, F. Meng, Q. Wu, D. B. da Costa, and M.-S. Alouini, “Accurate
2654–2658, 2011.
closed-form approximations to channel distributions of RIS-aided wire-
less systems,” IEEE Wireless Communications Letters, vol. 9, no. 11, [45] T. Yoo and A. Goldsmith, “Capacity of fading MIMO channels with
pp. 1985–1989, 2020. channel estimation error,” in 2004 IEEE International Conference on
Communications (IEEE Cat. No. 04CH37577), vol. 2. IEEE, 2004, pp.
[24] S. Shen, B. Clerckx, and R. Murch, “Modeling and architecture design
808–813.
of intelligent reflecting surfaces using scattering parameter network
analysis,” IEEE Transactions on Wireless Communications, 2021. [46] S. Katla, L. Xiang, Y. Zhang, M. El-Hajjar, A. A. Mourad, and L. Hanzo,
“Deep learning assisted detection for index modulation aided mmwave
[25] Q. Wu, S. Zhang, B. Zheng, C. You, and R. Zhang, “Intelligent reflecting
systems,” IEEE Access, vol. 8, pp. 202 738–202 754, 2020.
surface aided wireless communications: A tutorial,” IEEE Transactions
[47] J. R. Hampton, Introduction to MIMO communications. Cambridge
on Communications, 2021.
university press, 2013.
[26] A. Wang, R. Yin, and C. Zhong, “Channel estimation for uniform

View publication stats