IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO.

2, FEBRUARY 2022 1229

Modeling and Architecture Design of

Reconfigurable Intelligent Surfaces
Using Scattering Parameter
Network Analysis
Shanpu Shen , Member, IEEE, Bruno Clerckx , Senior Member, IEEE, and Ross Murch , Fellow, IEEE

Abstract— Reconfigurable intelligent surfaces (RISs) are an popularity as a revolutionary technology for achieving spec-
emerging technology for future wireless communication. The vast trum efficient, energy efficient, and cost effective wireless
majority of recent research on RIS has focused on system level communication [1]–[3]. RISs can alter or reconfigure the
optimizations. However, developing straightforward and tractable
electromagnetic models that are suitable for RIS aided communi- propagation environment so that the performance of wireless
cation modeling remains an open issue. In this paper, we address communications can be significantly improved [4]. For these
this issue and derive communication models by using rigorous reasons RIS is considered a potentially important technology
scattering parameter network analysis. We also propose new RIS for use in future 6G communications [5].
architectures based on group and fully connected reconfigurable An RIS consists of a large number of reconfigurable passive
impedance networks that can adjust not only the phases but
also the magnitudes of the impinging waves, which are more elements, where each element is able to introduce phase
general and more efficient than conventional single connected shift to the scattered signal. By collaboratively adjusting the
reconfigurable impedance network that only adjusts the phases phase shifts of all passive elements of the RIS, the scattered
of the impinging waves. In addition, the scaling law of the signals can add coherently with the signals from other paths
received signal power of an RIS aided system with reconfigurable at the desired receiver to boost the received signal power.
impedance networks is also derived. Compared with the single
connected reconfigurable impedance network, our group and Alternatively the signal paths can also be made to destructively
fully connected reconfigurable impedance network can increase add at non-intended receivers to suppress interference as well
the received signal power by up to 62%, or maintain the same as enhance security and privacy. In contrast with amplify-and-
received signal power with a number of RIS elements reduced forward (AF) relay technology [6], RIS has several advantages
by up to 21%. We also investigate the proposed architecture including low cost, low power consumption, contributing no
in deployments with distance-dependent pathloss and Rician
fading channel, and show that the proposed group and fully active additive thermal noise or self-interference enabling
connected reconfigurable impedance networks outperform the full-duplexing operation. In addition, RIS exhibits poten-
single connected case by up to 34% and 48%, respectively. tial features such as being low profile, light weight, and
Index Terms— Network analysis, reconfigurable intelligent having conformal geometry, making them straightforward to
surface, reflection coefficient, scattering parameter/matrix. deploy.
Due to the potential advantages, RIS has been investi-
I. I NTRODUCTION gated in various wireless communication systems including
multiple-input single-output (MISO) [7], [8], multiple-input
R ECONFIGURABLE intelligent surfaces (RISs), also
known as intelligent reflecting surfaces, have gained
multiple-output (MIMO) [9], multicell [10], and multigroup
multicast [11]. Specifically, for RIS aided MISO systems,
Manuscript received November 23, 2020; revised April 19, 2021; accepted in [7] energy efficiency is maximized by optimizing phase
August 2, 2021. Date of publication August 16, 2021; date of current version shifts while in [8] the transmit power is minimized by
February 14, 2022. This work was supported by Hong Kong Research Grants jointly optimizing active and passive beamforming. For RIS
Council through the Collaborative Research Fund under Grant C6012-20G.
The associate editor coordinating the review of this article and approving it aided MIMO systems, in [9] spectral efficiency is maxi-
for publication was Z. Zhang. (Corresponding author: Shanpu Shen.) mized based on a sum-path-gain maximization criterion. For
Shanpu Shen is with the Department of Electronic and Computer Engi- RIS aided multicell MIMO systems, the weighted sum rate
neering, The Hong Kong University of Science and Technology, Hong Kong
(e-mail: [email protected]). maximization problem is considered in [10] and it is shown
Bruno Clerckx is with the Department of Electrical and Electronic that the cell-edge performance can be significantly enhanced
Engineering, Imperial College London, London SW7 2AZ, U.K. (e-mail: by RIS. For RIS aided multigroup multicast communication
[email protected]).
Ross Murch is with the Department of Electronic and Computer Engineering systems, the sum rate of multiple multicasting groups is
and the Institute of Advanced Study, The Hong Kong University of Science maximized in [11]. Furthermore, RIS has also been inves-
and Technology, Hong Kong (e-mail: [email protected]). tigated to provide performance enhancement in orthogonal
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TWC.2021.3103256. frequency division multiplexing (OFDM) [12], [13], non-
Digital Object Identifier 10.1109/TWC.2021.3103256 orthogonal multiple access (NOMA) [14], and emerging areas
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
1230 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 2, FEBRUARY 2022

such as secure wireless communication [15], backscatter it accounts for the impedance mismatch and mutual coupling
communication [16], simultaneous wireless information and at the transmitter, RIS, and receiver. Additionally, assuming
power transfer (SWIPT) [17], [18], spectrum sharing [19], perfect matching and no mutual coupling, we can simplify
cognitive radio [20], unmanned aerial vehicle (UAV) commu- the model and achieve a straightforward and tractable RIS
nication [21], millimeter wave [22], [23], and mobile edge aided communication model. The conventional RIS model
computing [24]. In addition, research on optimizing RIS aided used in [7]–[30] for example is a particular instance of the
wireless systems with discrete phase shifts [25], [26], statistic proposed model.
and imperfect channel state information (CSI) [27], [28] and Second, we investigate the RIS architecture and propose
deep learning [29], [30] has also been conducted. two new architectures based on fully connected and group
While the vast majority of research on RIS has been devoted connected reconfigurable impedance networks, that are respec-
to system level optimization [7]–[30], developing models that tively modeled using complex symmetric unitary and block
satisfy the necessary electromagnetic (EM) equations while diagonal matrices with each block being complex symmetric
providing tractable and useful RIS aided communication mod- unitary. Those two architectures are more general than the
els still remains an open problem. There are only a few conventional single connected reconfigurable impedance net-
published results analyzing the physical and EM properties of work used in [7]–[30], which is modeled using a diagonal
RIS. In [31], physical optics is utilized to obtain expressions matrix with each entry having a unit modulus. In sharp
for the scattered field from a passive metallic surface and contrast with the conventional single connected architecture
accordingly an RIS pathloss model is derived. In addition, that only adjusts the phases of the impinging waves, our
in [32] a free-space pathloss model for an RIS aided wireless proposed fully and group connected architectures can adjust
communication is introduced from the perspective of EM not only the phases but also the magnitudes of the impinging
theory and experimentally verified in a microwave anechoic waves. This is the first paper to introduce fully connected and
chamber. In addition to the pathloss models, practical phase group connected networks and show the benefit over a single
shift models of RIS accounting for lumped inductance and connected network.
capacitance are proposed in [33], [34], and the optimization Third, we derive the scaling law of the received signal power
based on the practical phase shift models are also provided. of a single-input single-output (SISO) RIS aided system as a
However, the limitation of [31]–[34] is that they only focus function of the number of RIS elements. Both line-of-sight
on very specific physical and EM properties of RIS. There- (LoS) and Rayleigh fading channels have been considered.
fore, how to derive a straightforward and tractable yet EM It shows the power gain of the fully connected and group
based RIS aided communication model remains an open connected reconfigurable impedance network over the single
problem. connected reconfigurable impedance network in Rayleigh fad-
In addition to the RIS aided communication modeling issue, ing channels can be up to 1.62. Given the same received
another challenge in enabling the promise of RIS is that the signal power, it is shown that using fully connected and group
signal power received from the RIS is limited (or equivalently connected reconfigurable impedance networks can reduce the
the composite transmitter-RIS-receiver channel gain is very number of RIS elements by up to 21%, which is beneficial
low). As shown in [31], the signal power received from an for reducing the cost and area of RIS, especially when the
2
RIS is proportional to the square of RIS area and to 1/ (di r) number of RIS elements is large. In addition, it shows that the
where di is the distance between the transmitter and RIS and r group connected reconfigurable impedance network with small
is the distance between the RIS and receiver. In addition, com- group size can provide most of the performance enhancement
parisons with massive MIMO [35] and decode-and-forward and come close to the fully connected case while maintaining
relays [36] indicate that RIS needs a large number of elements low complexity.
to be competitive. Therefore, it remains a challenge to develop Fourth, we optimize the scattering matrix of the reconfig-
an efficient RIS architecture to improve the received signal urable impedance network in RIS to maximize the received
power. signal power in the SISO RIS aided system. We also eval-
In this paper, we derive a straightforward and tractable uate the received signal power in channel models with
yet EM based RIS aided communication model using a distance-dependent pathloss and Rician fading channels, which
rigorous scattering parameter network analysis. It has been is more general than the channel model used in previous
inspired by previous results on MIMO antennas [37] and scaling law analysis. The numerical results show that the fully
here it is extended to RIS. We also propose efficient RIS connected and group connected reconfigurable impedance net-
architectures, namely fully connected and group connected works can increase the received signal power by up to 48%
reconfigurable impedance networks, to improve the received and 34%, respectively.
signal power. The contributions of the paper are summarized as Organization: Section II provides the scattering parameter
follows. network analysis and proposes the fully connected and group
First, we derive a physical and EM compliant RIS aided connected reconfigurable impedance network. Section III
communication model using scattering parameter network provides the RIS aided communication model. Section IV
analysis. This is the first paper to characterize and model RIS provides the scaling laws for the SISO RIS aided system.
from the perspective of scattering parameters. Using scattering Section V evaluates the performance of the proposed RIS and
parameters is beneficial for accounting for the scattering Section VI provides conclusions and details possible future
mechanism of RIS. The derived model is general enough that work.
SHEN et al.: MODELING AND ARCHITECTURE DESIGN OF RISs 1231

by a scattering matrix S where S ∈ CN ×N and S = ST due

to the reciprocity, so that we have

b = Sa, (1)

where a, b ∈ CN ×1 refer to the incident and reflected waves

at the N ports, respectively. Particularly, we can partition a
and b as
⎡ ⎤ ⎡ ⎤
aT bT
⎢ ⎥ ⎢ ⎥
a = ⎣ aI ⎦ , b = ⎣ bI ⎦ , (2)
aR bR
where the subscripts T , I, and R refer to the transmitter, RIS,
T
and receiver, respectively, and ai = [ai,1 , ai,2 , . . . , ai,Ni ] ∈
Ni ×1 T Ni ×1
C and bi = [bi,1 , bi,2 , . . . , bi,Ni ] ∈ C for i ∈
{T, I, R} refer to the incident and reflected waves of the
antennas at the transmitter/RIS/receiver. Accordingly, we can
partition S as
⎡ ⎤
ST T ST I ST R
⎢ ⎥
S = ⎣ SIT SII SIR ⎦ , (3)
SRT SRI SRR
Fig. 1. System diagram of RIS aided wireless communication system.
where ST T ∈ CNT ×NT , SII ∈ CNI ×NI , SRR ∈ CNR ×NR
refer to the scattering matrices of the antenna arrays at
Notation: Bold lower and upper letters denote vectors the transmitter, RIS, and receiver, respectively. The diag-
and matrices, respectively. Letters not in bold font represent onal entries of ST T , SII , and SRR refer to the antenna
scalars.  {a}, |a|, and arg (a) refer to the real part, modulus, reflection coefficients while the off-diagonal entries refer
and phase of a complex scalar a, respectively. [a]i and a to antenna mutual coupling. SRT ∈ CNR ×NT , SIT ∈
refer to the ith element and l2 −norm of vector a, respectively. CNI ×NT , and SRI ∈ CNR ×NI refer to the transmis-
AT , AH , and [A]i,j refer to the transpose, conjugate trans- sion scattering matrices from the transmitter to receiver,
pose, and (i, j)th element a matrix A, respectively. R and from the transmitter to RIS, and from RIS to the receiver,
√ C respectively.
denote real and complex number set, respectively. j = −1
denotes imaginary unit. 0 and I denote an all-zero matrix and
an identity matrix, respectively, with appropriate dimensions. A. Transmitter and Receiver
χk denotes the chi distribution with k degrees of freedom.
CN (0, I) denotes the distribution of a circularly symmetric At the transmitter, for nT = 1, . . . , NT , the nT th transmit
complex Gaussian random vector with mean vector 0 and antenna is connected in series with a voltage source, denoted as
covariance matrix I and ∼ stands for “distributed as”. A  B vs,nT , and a source impedance, denoted as ZT,nT . Therefore,
means that B − A is positive semi-definite. diag(a1 , . . . , aN ) aT and bT are related by
refers to a diagonal matrix with diagonal elements being aT = bs,T + ΓT bT , (4)
a1 , . . . , aN . diag(A1 , . . . , AN ) refers to a block diagonal
matrix with blocks being A1 , . . . , AN . T
where bs,T = 12 [vs,1 , vs,2 , . . . , vs,NT ] ∈ CNT ×1 refers to the
wave source vector and ΓT ∈ CNT ×NT is a diagonal matrix
with its (nT , nT )th entry referring to the reflection coefficient
II. N ETWORK A NALYSIS
of the nT th source impedance, i.e.
In this section, we use scattering parameter network the-
ZT,nT − Z0
ory to analyze RIS aided wireless communication systems. [ΓT ]nT ,nT = , (5)
A brief review of the basic concepts of reflection coefficient ZT,nT + Z0
and scattering parameter network theory is provided in the where Z0 refers to the reference impedance used for
Appendix. computing the scattering parameter, and usually set as
A system diagram of an RIS aided wireless communication Z0 = 50 Ω.
system is shown in Fig. 1. There are NT antennas at the At the receiver, for nR = 1, . . . , NR , the nR th receive
transmitter, NR antennas at the receiver, and NI antennas at antenna is connected in series with a load impedance, denoted
the RIS. In total, there are N = NT + NI + NR antennas as ZR,nR . Therefore, aR and bR are related by
embedded in the wireless channel, which can be modeled as
an N -port network. The N -port network can be characterized aR = ΓR bR , (6)
1232 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 2, FEBRUARY 2022

where the (nI , nI )th entry of Θ, denoted as [Θ]nI ,nI , is the

reflection coefficient of the reconfigurable impedance ZnI , i.e.
ZnI − Z0
[Θ]nI ,nI = , (11)
ZnI + Z0

so that we have an equivalent constraint that [Θ]nI ,nI ≤ 1.

Furthermore, to increase the power scattered by RIS, ZnI is
purely reactive for nI = 1, . . . , NI , i.e. ZnI = jXnI where
XnI denotes the reconfigurable reactance. Therefore
jXnI − Z0
[Θ]nI ,nI = = ejθnI , (12)
jXnI + Z0
where 0 ≤ θnI ≤ 2π denotes phase shift. Hence, we have
Fig. 2. (a) 4-element RIS with single connected reconfigurable impedance
network and (b) 4-element RIS with fully connected reconfigurable impedance
an equivalent unit modulus constraint of [Θ]nI ,nI = 1. The
network. single connected reconfigurable impedance network and the
corresponding constraints (10), (12) have been widely adopted
where ΓR ∈ CNR ×NR is a diagonal matrix with its (nR , nR )th in RIS aided wireless communication system designs and
entry referring to the reflection coefficient of the nR th load optimizations [7]–[30].
impedance, i.e. 2) Fully Connected Reconfigurable Impedance Network:
In this paper, we propose a more general reconfigurable
ZR,nR − Z0
[ΓR ]nR ,nR = . (7) impedance network, which is denoted as the fully connected
ZR,nR + Z0 reconfigurable impedance network, to further improve the
signal power received from RIS and enhance the performance
B. Reconfigurable Intelligent Surface of RIS. In this category, each port of the NI ports of the
At the RIS, the NI antennas are connected to a NI -port reconfigurable impedance network is connected to other ports.
reconfigurable impedance network. Therefore, aI and bI are An illustrative example, for a 4-element RIS with a 4-port
related by fully connected reconfigurable impedance network, is shown
aI = ΘbI , (8) in Fig. 2(b). Generally, for nI = 1, . . . , NI , the nI th port is
connected to ground with a reconfigurable impedance ZnI and
where Θ ∈ CNI ×NI refers to the scattering matrix of the NI - the nI th port is connected to the mI th port, for mI = nI +
port reconfigurable impedance network. According to [38], Θ 1, . . . , NI , with a reconfigurable impedance ZnI ,mI , so that
can be expressed as in total there are NI (NI + 1) /2 reconfigurable impedance
Θ = (ZI + Z0 I)−1 (ZI − Z0 I) , (9) components in the network. Therefore, ZI is a full matrix
and, following [39], ZI can be obtained from the following
where ZI ∈ CNI ×NI refers to the impedance matrix of relationship
the NI -port reconfigurable impedance network. The NI -port
reconfigurable impedance network is constructed with recon- −Zn−1
I ,mI
, n I = mI ,
Z−1
I nI ,mI
= −1
 −1 (13)
figurable and passive elements so that it can reflect the incident ZnI + k=nI ZnI ,k , nI = mI ,
signal with a reconfiguration that can be adapted to the
channel. The NI -port reconfigurable impedance network is where ZnI ,mI = ZmI ,nI due to the symmetric ZI . According
also reciprocal so that we have symmetry where ZI = ZTI to (13), we can implement an arbitrary impedance matrix ZI
by selecting proper impedance ZnI and ZnI ,mI . Subsequently,
and Θ = ΘT . According to the circuit network topology, the
Θ can be found by (9). According to network theory [38], Θ
NI -port reconfigurable impedance network can be classified
is a full matrix satisfying the constraints
into three categories.
1) Single Connected Reconfigurable Impedance Network: Θ = ΘT , ΘH Θ  I. (14)
In this category, each port of the NI ports of the recon-
figurable impedance network are not connected to the other Furthermore, to increase the power scattered by RIS, ZnI and
ports. An illustrative example, for a 4-element RIS with a ZnI ,mI are purely reactive. That is ZnI = jXnI and ZnI ,mI =
4-port single connected reconfigurable impedance network, jXnI ,mI where XnI and XnI ,mI denote the reconfigurable
is shown in Fig. 2(a). Generally, for nI = 1, . . . , NI , the nI th reactances, so that we have ZI = jXI where XI ∈ RNI ×NI
port is connected to ground with a reconfigurable impedance denotes the reactance matrix of the NI -port reconfigurable
ZnI , so that in total there are NI reconfigurable impedance impedance network and XI = XTI . Hence, Θ is given by
components in the network. Hence, ZI is a diagonal matrix
Θ = (jXI + Z0 I)−1 (jXI − Z0 I) , (15)
given by ZI = diag (Z1 , Z2 , . . . , ZNI ) and according to (9),
Θ is also a diagonal matrix given by so that according to [38] we have equivalent constraints

Θ = diag [Θ]1,1 , [Θ]2,2 , . . . , [Θ]NI ,NI , (10) Θ = ΘT , ΘH Θ = I, (16)
SHEN et al.: MODELING AND ARCHITECTURE DESIGN OF RISs 1233

the constraints
Θ = diag (Θ1 , Θ2 , . . . , ΘG ) , (18)
Θg = ΘTg , ΘH
g Θg  I, ∀g. (19)
Furthermore, to increase the power scattered by RIS, ZI,g are
purely reactive, i.e. ZI,g = jXI,g where XI,g ∈ RNG ×NG
denotes the reactance matrix of the NG -port reconfigurable
impedance network and XI,g = XTI,g . Accordingly, Θg can
be found as
−1
Θg = (jXI,g + Z0 I) (jXI,g − Z0 I) , (20)
so that according to [38] we have equivalent constraints
Θg = ΘTg , ΘH
g Θg = I, ∀g. (21)
Therefore, Θ is a block diagonal matrix with each block being
a complex symmetric unitary matrix. For an NI -port group
connected reconfigurable impedance network, there are in
total NI (NG + 1) /2 reconfigurable impedance components.
When the group size NG = 1, it becomes the single con-
nected reconfigurable impedance network. When the group
Fig. 3. 8-element RIS with group connected reconfigurable impedance size NG = NI , it becomes the fully connected reconfigurable
network (a) having 4 groups with group size of 2 and (b) having 2 groups
with group size of 4. impedance network.
Comparisons between the single connected, fully connected,
and group connected reconfigurable impedance networks will
which shows that Θ is a complex symmetric unitary matrix. be shown in the following sections.
The single connected reconfigurable impedance network
(10), (12) is a special case of the fully connected reconfig- III. RIS A IDED C OMMUNICATION M ODEL
urable impedance network (16), so that the fully connected We have analyzed the fundamental relationships between
reconfigurable impedance network is more general and is the incident and reflected waves of the antennas at the trans-
expected to provide better RIS performance. mitter, receiver, and RIS. In this section, we establish the RIS
3) Group Connected Reconfigurable Impedance Network: aided communication model based on these relationships.
For the fully connected reconfigurable impedance network,
when NI becomes large, the number of reconfigurable A. General RIS Aided Communication Model
impedance components becomes huge (increasing quadrati-
cally with NI ) and the circuit topology will become intricate. We first consider a general RIS aided communication model.
This limits its practical use and therefore we also propose a Combining (4), (6), and (8), we can relate a and b in a
group connected reconfigurable impedance network to achieve compact form as
a good tradeoff between performance enhancement and com- a = bs + Γb, (22)
plexity. Two illustrative examples, for an 8-element RIS with
a group connected reconfigurable impedance network having where bs and Γ are respectively given by
4 groups and 2 groups, are shown in Fig. 3(a) and (b), ⎡ ⎤ ⎡ ⎤
bs,T ΓT 0 0
respectively. In Fig. 3(a), the 8 elements in RIS are divided into bs = ⎣ 0 ⎦ , Γ = ⎣ 0 Θ 0 ⎦. (23)
4 groups and each group has 2 elements and uses a 2-port fully 0 0 0 ΓR
connected reconfigurable impedance network. In Fig. 3(b), the
8 elements in RIS are divided into 2 groups and each group Substituting (1) into (22), we have that
has 4 elements and uses a 4-port fully connect reconfigurable b = S (I − ΓS)−1 bs . (24)
impedance network. Generally, for NI -element RIS, we can
−1
divide it into G groups with each group having NG = NGI We define T  S (I − ΓS) ∈ CN ×N and partition T as
elements. We refer to NG as the group size. For the gth group, ⎡ ⎤
TT T TT I TT R
a NG -port fully connected reconfigurable impedance network T = ⎣ TIT TII TIR ⎦ , (25)
with impedance matrix of ZI,g ∈ CNG ×NG is used. Therefore, TRT TRI TRR
ZI is a block diagonal matrix given by
so that we can find bT and bR as
ZI = diag (ZI,1 , ZI,2 , . . . , ZI,G ) . (17) bT = TT T bs,T , bR = TRT bs,T . (26)
Subsequently Θ can be found by (9). According to We define the voltage vector at the transmitter as
T
(9) and [38], Θ is a block diagonal matrix satisfying vT  [vT,1 , vT,2 , . . . , vT,NT ] ∈ CNT ×1 where vT,nT
1234 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 2, FEBRUARY 2022

refers to the voltage across the nT th transmit antenna. Making use of (30) and (31), we can simplify the expression
−1
We also define the voltage vector at the receiver as vR  of T = S (I − ΓS) and then write that
[vR,1 , vR,2 , . . . , vR,NR ]T ∈ CNR ×1 where vR,nR refers to the
TT T = ST I ΘSIT , (32)
voltage across the nR th receive antenna. With the incident and
reflected waves, we can find vT and vR as TRT = SRT + SRI ΘSIT . (33)

vT = aT + bT , vR = aR + bR . (27) Substituting (32) and (33) into (29) and making use of ΓT = 0
and ΓR = 0, we can simplify the channel matrix H as
More details about the relationship between the voltages and
−1
the incident and reflected waves of an N -port network can be H = (SRT + SRI ΘSIT ) (I + ST I ΘSIT ) . (34)
found in the Appendix. Utilizing (4), (6), and (26)-(27), we The term ST I ΘSIT refers to the second order reflections
can relate vT and vR by between the transmitter and RIS and back to the transmitter.
−1
vR = (ΓR + I) TRT (I + ΓT TT T + TT T ) vT . (28) However, in most applications, there is no need to consider
these second order reflections because the power of the sec-
Defining vT as the transmit signal x and vR as the receive ond reflection ST I ΘSIT is extremely small, and is propor-
signal y, we can find the channel matrix of the RIS aided tional to the square of the pathloss between the transmitter
wireless communication system as to RIS. Hence, in most applications, we can approximate
−1
H = (ΓR + I) TRT (I + ΓT TT T + TT T )
−1
, (29) (I + ST I ΘSIT ) as I without affecting the accuracy, and
then simplify H as
so that we have y = Hx (ignoring the additive white Gaussian
noise (AWGN) at the receiver). According to (23) and T  H = SRT + SRI ΘSIT . (35)
S (I − ΓS)−1 , the submatrice TT T and TRT are functions of The transmission scattering matrices SRT , SIT , and SRI
Θ, denoted as TT T (Θ) and TRT (Θ), so that from (29) the are equivalently the channel matrices from the transmitter to
channel matrix H is also a function of Θ, denoted as H (Θ). receiver, from the transmitter to RIS, and from the RIS to
Hence, we can optimize Θ to intelligently control the channel receiver, respectively. To show the equivalence, we consider
H (Θ) and enhance the wireless system performance. the impedance matrix of the N -port network
The general communication model (29) includes the effects ⎡ ⎤
of impedance mismatching and mutual coupling at the trans- ZT T ZT I ZT R
mitter, RIS, and receiver. However, generally it is difficult Z = ⎣ ZIT ZII ZIR ⎦ . (36)
to find expressions for TT T (Θ) and TRT (Θ) due to the ZRT ZRI ZRR
matrix inversion operation. Subsequently, it is difficult to Since we assume perfect matching and no mutual coupling,
find the expressions of H (Θ), which makes it difficult to we have ZT T = ZII = ZRR = Z0 I and following [38] we
obtain insight into the role of RIS in the communication can derive
model and to optimize Θ of RIS. Considering this issue, Zij
in the following, we consider a special case to simplify the Sij = , (37)
2Z0
expression of H (Θ).
for ij ∈ {RT, RI, IT }. We take ZRT as an example to see the
details. The (nR , nT )th entry of ZRT , denoted as [ZRT ]nR ,nT ,
B. RIS Aided Communication Model With Perfect
refers to the trans-impedance between the nT th transmit
Matching and No Mutual Coupling
antenna and nR th receive antenna. To find [ZRT ]nR ,nT , we
We consider a special case that assumes the antenna arrays excite the nT th transmit antenna with current iT,nT and keep
at the transmitter, RIS, and receiver are perfectly matched all the other antennas open circuited, and then measure the
and have no mutual coupling, i.e. ST T = 0, SII = 0, open
open-circuit voltage vR,n at the nR th receive antenna. Using
and SRR = 0. In practice, this assumption can be approx- R
the multipath propagation based model, we have that
imately achieved by individually matching each antenna to L 
the reference impedance Z0 and keeping the antenna spacing open
RT

vR,nR = eR,nR (ΩR,l ) βRT,l eT,nT (ΩT,l ) iT,nT , (38)

larger than half-wavelength. It is also assumed that the source
l=1
impedance at the transmitter ZT,nT and the load impedance   
at the receiver ZR,nR are all reference impedances Z0 so that [ZRT ]n
R ,nT
ΓT = 0 and ΓR = 0. With these two assumptions, we can
where eT,nT (·) and eR,nR (·) denote the open-circuit radiation
simplify
⎡ ⎤ ⎡ ⎤ pattern of the nT th transmit antenna and the nR th receive
0 ST I ST R 0 0 0 antenna, respectively. We assume a channel with LRT paths
S = ⎣ SIT 0 SIR ⎦ , Γ = ⎣ 0 Θ 0 ⎦ , (30) for propagation from transmitter to receiver with the lth path
SRT SRI 0 0 0 0 characterized by departure and arrival angles ΩT,l and ΩR,l ,
respectively, and a complex channel gain βRT,l . Therefore,
so that accordingly we can simplify (I − ΓS)−1 as from (37) and (38), we show that SRT is equivalently the chan-
⎡ ⎤
I 0 0 nel matrix from the transmitter to receiver. Similarly, we can
−1
(I − ΓS) = ⎣ ΘSIT I ΘSIR ⎦ . (31) show that SIT and SRI are the channel matrices from the
0 0 I transmitter to RIS and from the RIS to receiver, respectively.
SHEN et al.: MODELING AND ARCHITECTURE DESIGN OF RISs 1235

We use auxiliary notations HRT = SRT , HIT = SIT , and IV. S CALING L AW
HRI = SRI to facilitate understanding, so that we can rewrite In order to obtain insights into the fundamental limits of
(35) as single, fully, and group connected reconfigurable impedance
networks in RIS, we quantify how the received signal power
H = HRT + HRI ΘHIT . (39)
scales as a function of the number of RIS elements NI . For
Furthermore, assuming there are LoS and non-LoS (NLoS) simplicity, we consider a SISO RIS aided system (NT = 1,
paths in (38), we can model HRT , HIT , and HRI as Rician NR = 1) with perfect matching and no mutual coupling
  in the
fading, i.e. following. The transmit signal is x ∈ C with E |x|2 = PT .
   According to (39), the received signal y can be expressed as
 Kij LoS 1 NLoS
Hij = Lij H + H , (40) y = (hRT + hRI ΘhIT ) x + n, (41)
1 + Kij ij 1 + Kij ij
where hRT ∈ C, hIT ∈ CNI ×1 , and hRI ∈ C1×NI denote
for ij ∈ {RT, RI, IT } where Lij refers to the pathloss, the channel from the transmitter to the receiver, from the
Kij refers to the Rician factor, HLoS ij and HNLoS
ij represent transmitter to the RIS, and from the RIS to the receiver,
the small-scale LoS and NLoS (Rayleigh fading) components, respectively, and n is the AWGN. For simplicity, we assume
respectively. the transmit power PT = 1 and omit the direct channel
The simplified H (39) is a linear function of Θ and together hRT , so that we can express the received signal power as
with 2
PR = |hRI ΘhIT | .
1) the single connected reconfigurable impedance net-
work which satisfies the  constraint that Θ = A. Single Connected Reconfigurable Impedance Network
diag ejθ1 , ejθ2 , . . . , ejθNI ,
2) the fully connected reconfigurable impedance network For the single connected reconfigurable impedance network
which satisfies the constraints that Θ = ΘT , ΘH Θ = I, (10), (12), it is obvious that the optimal Θ is
3) the group connected reconfigurable impedance net-    
Θ = diag ejθ1 , ejθ2 , . . . , ejθNI , (42)
work which satisfies the constraints that Θ =  
diag (Θ1 , Θ2 , . . . , ΘG ), Θg = ΘTg , ΘHg Θg = I, ∀g, θn I = − arg [hRI ]nI [hIT ]nI , ∀nI , (43)
make up our proposed RIS aided communication model. Based
which achieves the maximum received signal power
on this model, we can optimize Θ to intelligently control the
 N 2
channel H (Θ) and enhance the wireless system performance.  I
Single
Note that, the conventional RIS aided communication model PR = [hRI ]nI [hIT ]nI . (44)
used in [7]–[30] is a special case of our proposed model that nI =1
corresponds to the single connected reconfigurable impedance
network. Importantly, in sharp contrast with the conventional
single connected architecture [7]–[30] that only adjusts the B. Fully Connected Reconfigurable Impedance Network
phases of the impinging waves using a diagonal scattering For the fully connected reconfigurable impedance net-
matrix, our proposed group and fully connected architectures work (16), using the Cauchy-Schwarz inequality and that
enable scattering matrices to be block diagonal or full and ΘH Θ = I, we can find an upper bound for the maximum
can consequently adjust not only the phases but also the received signal power PRFully as
magnitudes of the impinging waves. This leads to significant 2 2
PRFully ≤ P̄RFully = hRI  hIT  . (45)
performance gains in fading channels as it will appear in
Sections IV and V. The key to achieve the upper bound P̄RFully is that we need to
Remark 1: It is worthwhile to clarify the differences find a complex symmetric unitary matrix Θ satisfying
between our proposed model with the RIS aided communi-
cation model proposed in recent work [40]. There are four hH
RI hIT
=Θ . (46)
differences. First, compared with using impedance parameter hRI  hIT 
in [40], we have found it is more natural to use the reflection However, it is difficult to derive a closed-form solution
coefficient and scattering parameter to account for the scat- for the optimal Θ which satisfies the equation (46) and
tering mechanism of RIS and derive the RIS aided communi- achieves the upper bound. Hence, we directly optimize Θ to
cation model. Second, our general RIS aided communication approach the upper bound using the quasi-Newton method as
model (29) includes the effects of impedance mismatching detailed in the next section. Numerical results using the Monte
and mutual coupling at the transmitter, RIS, and receiver, Carlo method confirms that the upper bound (45) is tight.
which is more general than [40]. Third, we clearly explain It is also worthwhile to compare the maximum received
the physical significance of the phase shifts and the unit signal power of the single connected and fully connected
modulus constraint. Fourth, we go beyond the single connected reconfigurable impedance networks. From (44) and (45), using
reconfigurable impedance network (the main focus in [40]) and Cauchy-Schwarz inequality we can deduce that
propose more general fully connected and group connected
reconfigurable impedance networks. PRSingle ≤ P̄RFully , (47)
1236 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 2, FEBRUARY 2022

and the equality is achieved if and only if it is straightforward to show that PRSingle ≤ P̄RGroup ≤ P̄RFully .
[hRI ]nI = α [hIT ]nI , ∀nI , (48) In the next subsections, we investigate PRSingle , PRFully , and
PRGroup with LoS and Rayleigh fading channels.
where α can be any positive scalar. In other words, when
the channel gains (the modulus) of hRI and hIT are linearly
D. Line-of-Sight Channel
independent, the fully connected reconfigurable impedance
network can achieve a higher received signal power than the Assuming hRI and hIT are both LoS channels, we have
T
single connected case. that hRI = ejφ1 , . . . , ejφNI and hIT = ejψ1 , . . . , ejψNI .
We also provide physical explanations to account for the In this case, it is obvious that the optimal Θ for the single,
better performance of the fully connected case. For the single fully, and group connected reconfigurable impedance networks
connected case, each port of the reconfigurable impedance are the same and given by
network is not connected to other ports. As a result, only the 
Θ = diag e−j(φ1 +ψ1 ) , . . . , e−j (φNI +ψNI ) . (52)
phase of the elements of the vector ΘhIT can be adjusted.
Therefore the best that RIS can achieve is to make the two Therefore, the single, fully, and group connected reconfig-
channel vectors hRI and hIT element-wise in phase. However, urable impedance networks have the same performance with
for the fully connected case, each port of the reconfigurable LoS channel such that
impedance network is connected to each other. As a result,
the phase and magnitude of the elements of the vector ΘhIT PRSingle = PRFully = PRGroup = NI2 . (53)
can be jointly adjusted so that the RIS can align the two This is consistent with that the equality in (47) that can be
channel vectors hRI and hIT in the same direction to achieve a achieved when (48) is satisfied.
better performance. Intuitively speaking, the single connected
case is analogous to the equal-gain combining while the fully
connected case is analogous to the maximum ratio combining. E. Rayleigh Fading Channel
Namely, instead of adjusting only the phase of the impinging Assuming hRI and hIT are independent and identically
wave as in the single connected architecture, the fully con- distributed (i.i.d.) Rayleigh fading channels, we have hRI ∼
nected architecture can adjust not only the phases but also the CN (0, I) and hIT ∼ CN (0, I). We first consider the group
magnitudes of the impinging waves. connected case. We rewrite P̄RGroup as
G
 2 2
C. Group Connected Reconfigurable Impedance Network P̄RGroup = hRI,g  hIT,g 
For the group connected reconfigurable impedance network g=1

(18), (21), we can rewrite the received signal power as + hRI,g1  hIT,g1  hRI,g2  hIT,g2  ,
2 g1 =g2
G

PR = hRI,g Θg hIT,g , (49) (54)
g=1 Taking the expectation and making use of the i.i.d. Rayleigh
where hRI = [hRI,1 , hRI,2 , . . . , hRI,G ] with hRI,g ∈ C1×NG fading assumption of hRI and hIT , we can find the average
T
and hIT = [hIT,1 , hIT,2 , . . . , hIT,G ] with hIT,g ∈ CNG ×1 . P̄RGroup as
Using the triangle inequality, Cauchy-Schwarz inequality, and    2
2 4
that ΘH E P̄RGroup = GE hRI,1  + G (G − 1) E [hRI,1 ] .
g Θg = I ∀g, we can find an upper bound for the
maximum received signal power PRGroup as (55)
 G 2 Making use of the moment of χ2NG distribution, we have
     that
Group Group 1 2
PR ≤ P̄R = hRI,g  hIT,g  . (50) E [hRI,1 ] = Γ NG + 2 /Γ (NG ) and E hRI,1  = NG
g=1
where Γ (x) refers to the gamma function.
 With
 G = NI /NG
The key to achieve the upper bound P̄RGroup is that we need and the expressions of moments, E P̄RGroup is given by
to find a complex symmetric unitary matrix Θg satisfying
      4
hH Group NI NI Γ NG + 12
RI,g hIT,g E P̄R = NI NG + −1 .
= Θg , ∀g. (51) NG NG Γ (NG )
hRI,g  hIT,g 
(56)
However, it is difficult to derive a closed-form solution for the
optimal Θg achieving the upper bound. Similar to the fully The single and fully connected reconfigurable impedance
connected case, we directly optimize Θg ∀g to approach the networks can be viewed as two special cases of the group
upper bound using the quasi-Newton method and numerical connected reconfigurable impedance network, i.e. the group
results using the Monte Carlo method confirm that the upper size NG = 1 and NG = NI . Therefore, from (56), we can
bound (50) is tight. straightforwardly derive that
The group connected case can be viewed as a tradeoff  
4
between the single connected and fully connected cases so that E PRSingle = NI + NI (NI − 1) Γ (1.5) , (57)
SHEN et al.: MODELING AND ARCHITECTURE DESIGN OF RISs 1237

Fig. 4. Average received signal power versus the number of RIS elements. Fig. 5. Power gain of the group connected and fully connected reconfigurable
impedance networks over the single connected reconfigurable impedance
network.
 
E P̄RFully = NI2 . (58)
When NI → ∞, the limits of the power gain are
     4
4
From (57), we can deduce that E PRSingle → NI2 Γ (1.5) = 1 Γ N G + 1
lim G Group = 2 2
(61)
NI2 π 2 /16 when NI → ∞, which is consistent with existing NI →∞ NG Γ (NG ) Γ (1.5)
results [8] indicating that the RIS has a squared power gain. 1 16
For the single, group, and fully connected reconfigurable lim G Fully = 4 = . (62)
NI →∞ Γ (1.5) π2
impedance networks, we directly optimize Θ to maximize
the received signal power using the quasi-Newton method as The power gains G Group and G Fully versus NI are shown
detailed in the next section. Using the Monte Carlo method, in Fig. 5. We can observe that the power gain increases with
we optimize Θ for each channel realization and find the the group size. For NG = 2, 3, 4, 6, 8, the power gain is
average received signal power. The comparisons between the around 1.26, 1.37, 1.43, 1.49, 1.52, respectively. However,
optimized result and the upper bound (or the closed-form increasing the group size cannot increase the power gain
solution) for the single, group, and fully connected recon- without limit. For the fully connected case (the maximum
figurable impedance networks are shown in Fig. 4. We can group size), we can find that the limit of the power gain
observe that the optimized average received signal power is is around 1.62, which is consistent with (61). Comparing
the same as the upper bound for the group and fully connected G Group and G Fully , we find that small group size NG , such
cases, which shows the upper bounds  (45) and
 (50)are tight. as 2, 3, 4, achieves satisfactory power gain while maintaining
Group
Therefore, we can conclude that E PR = E P̄RGroup low complexity. Therefore, the group connected reconfigurable
    impedance network with small group size is more useful in
and E PRFully = E P̄RFully . practice.
Additionally, we can observe that 1) the group connected Given the same average received signal power, the number
case achieves higher average received signal power than the of RIS elements required by the group connected or fully con-
single connected case, 2) the larger the group size is, the higher nected reconfigurable impedance network (denoted as NIGroup
the average received signal power is, and 3) the fully connected and NIFully ) is less than that required by the single connected
case achieves the highest power. The higher received power of reconfigurable impedance network (denoted as NISingle ). When
the fully and group connected architectures in Rayleigh fading the number of RIS elements is large, from (59), we can deduce
channel comes from their ability to adjust both the phases that
and the magnitudes of the impinging waves. To quantify the  2  2
increase in the received signal power, we can find the power NISingle = G Group NIGroup . (63)
gain of the group and fully connected cases over the single
The percentage decrease in the number of RIS elements is
connected case, which are respectively given by
given by
 4
(NI −NG ) Γ(NG + 12 ) NISingle − NIGroup 1
NG + δ= =1− √ , (64)
NISingle
NG2 Γ(NG ) G Group
Group
G = 4 , (59)
1 + (NI − 1) Γ (1.5) which can also be applied to the fully connected
NI case by replacing G Group with G Fully . Therefore, for
G Fully = . (60)
1 + (NI − 1) Γ (1.5)4 NG = 2, 3, 4, 6, 8, we have δ = 11%, 14%, 16%, 18%, 19%,
1238 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 2, FEBRUARY 2022

respectively. For the fully connected case, we have δ = 21%.

Such reduction in the number of RIS elements is beneficial
for reducing the cost and area of RIS, especially when the
number of RIS elements is large. Similarly, we find for small
group sizes NG , such as 2, 3, 4, they achieve the greatest
relative RIS element reduction while maintaining low com-
plexity, demonstrating that the group connected reconfigurable
impedance network with small group size is more useful in Fig. 6. 2D coordinate system for the SISO RIS aided system.
practice.
To conclude, comparing the performance of the three RIS
architectures in LoS and Rayleigh fading channels, we show problem (69)-(72) to an unconstrained problem which opti-
the gains of fully and group connected cases appear when mizes NI (NG + 1) /2 unconstrained variables [XI,g ]i,j for
the channel gains (the modulus) of hRI and hIT are linear i ≤ j and all g. To solve the unconstrained optimization
independent. problem, we can use the quasi-Newton method in MATLAB
to directly optimize [XI,g ]i,j for i ≤ j and all g, and then
V. P ERFORMANCE E VALUATION find a stationary point of the problem (69)-(72). Using the
quasi-Newton method with BFGS update, the computational
In this section, we formulate the received signal power 2
maximization in an RIS aided SISO system with the fully complexity for each iteration is O NI2 (NG + 1) /4 [41].
connected and group connected reconfigurable impedance The single and fully connected reconfigurable impedance
networks, and evaluate the performance in a realistic networks can be viewed as two special cases of the group
channel model. The received signal power is given by connected reconfigurable impedance network, i.e. the group
PR = PT hRT + hRI ΘhIT 2 . We first consider the size NG = 1 and NG = NI . Therefore, we can follow the
group connected reconfigurable impedance network in the same approach to solve the received signal power maximiza-
RIS, corresponding to the constraints (18), (21). Maximiz- tion problem with single and fully connected cases.
ing PT hRT + hRI ΘhIT 2 is equivalent to maximizing We now evaluate the performance of the RIS aided SISO
hRI ΘhIT 2 since hRI ΘhIT can always be made in phase system with the single, fully, and group connected reconfig-
with hRT . Therefore, we can equivalently formulate the urable impedance networks. We consider a two-dimensional
received signal power maximization problem with the group (2D) coordinate system as shown in Fig. 6. A single-antenna
connected reconfigurable impedance network as transmitter is located at (0, 0) and a single-antenna receiver
2
is located at (52, 0). A uniform linear array (ULA) at the
max hRI ΘhIT  (65) RIS are located in x-axis. The antenna spacing is half wave-
Θ,Θg
length and the center of the array is located at (50, 2). The
s.t. Θ = diag (Θ1 , Θ2 , . . . , ΘG ) , (66)
distance-dependent pathloss model is given by
ΘH
g Θg = I, ∀g, (67)  −αij
Θg = ΘTg , ∀g. (68) dij
Lij (dij ) = C0 (73)
D0
The constraints (66)-(68) indicate that Θ is a block diagonal
matrix with each block being a complex symmetric unitary where C0 refers to the pathloss at the reference distance
matrix, which makes the optimization difficult. To handle that, D0 = 1 meter (m), dij refers to the distance, and αij refers
we leverage the relationship between the scattering matrix to the pathloss exponent for ij ∈ {RT, RI, IT }. For the
Θg and the reactance matrix XI,g , as provided in (20), to small-scale fading, we assume that the transmitter-receiver and
equivalently rewrite the problem (65)-(68) as RIS-receiver channels are both Rayleigh fading channels, and
assume that the transmitter-RIS channel is Rician fading as
2
max hRI ΘhIT  (69) provided in (40). We set αRT = 3.5, αIT = 2, αRI = 2.8,
Θ,XI,g
C0 = −30 dB, and PT = 10 W.
s.t. Θ = diag (Θ1 , Θ2 , . . . , ΘG ) , (70) Using the Monte Carlo method, we compute the average
−1
Θg = (jXI,g + Z0 I) (jXI,g − Z0 I) , ∀g, (71) received signal power achieved after optimizing the single,
XI,g = XTI,g , ∀g, (72) fully, and group connected reconfigurable impedance net-
works. The average received signal power versus the num-
which can be transformed to an unconstrained optimization ber of RIS elements with different Rician factors of the
problem. Specifically, substituting (70) and (71) into the transmitter-RIS channel is shown in Fig. 7. We can make the
objective (69), we can express the objective (69) as a function following observations. First, the group connected reconfig-
of XI,g ∀g. Since XI,g can be an arbitrary NG × NG real urable impedance network achieves a higher received signal
symmetric matrix, XI,g is a function of the NG (NG + 1) /2 power than the single connected reconfigurable impedance net-
entries in the upper triangular part, i.e. [XI,g ]i,j for i ≤ j, work. The fully connected reconfigurable impedance network
and there is no constraint for [XI,g ]i,j for i ≤ j. Therefore, achieves the highest received signal power. This demonstrates
we can express the objective (69) as a function of [XI,g ]i,j the benefit of the group and fully connected reconfigurable
for i ≤ j and all g. Subsequently we can transform the impedance networks. Second, the larger the group size is,
SHEN et al.: MODELING AND ARCHITECTURE DESIGN OF RISs 1239

TABLE I
C OMPREHENSIVE C OMPARISON OF S INGLE C ONNECTED , G ROUP C ONNECTED , AND F ULLY C ONNECTED R ECONFIGURABLE I MPEDANCE N ETWORK

Fig. 8. Power gains of the group connected and fully connected recon-
figurable impedance networks over the single connected reconfigurable
Fig. 7. Average received signal power versus the number of RIS elements impedance network with different Rician factors of the transmitter-RIS
with different Rician factors of the transmitter-RIS channel. channel.

the higher the received signal power is, which indicates that 1) the circuit topology complexity, which refers to the number
we can trade complexity for signal power enhancement. Last, of reconfigurable impedance components in the reconfigurable
the received signal power of the fully connected reconfig- impedance network, and 2) the optimization computational
urable impedance network does not change with the Rician complexity, which refers to the computational complexity
factor. Indeed, it can be deduced from the scaling laws for optimizing the reconfigurable impedance networks with
(53) and (58) in that the received signal power is always NI2 . different constraints of Θ. A comprehensive comparison of the
However, the received signal power of the single connected single, group, and fully connected reconfigurable impedance
case increases with the Rician factor. Indeed, we can deduce networks in terms of the power gain at different Rician factors,
from the scaling laws (53) and (57) in that the single connected the circuit topology complexity, and the optimization computa-
case achieves a higher power in the LoS channel compared to tional complexity is summarized in Table. I. We can conclude
Rayleigh fading channels. that the power gain can be enhanced by introducing more
We also plot the power gains of the group and fully reconfigurable components and more optimization computa-
connected reconfigurable impedance networks over the single tions, i.e. trading circuit topology complexity and optimiza-
connected reconfigurable impedance network with different tion computational complexity for performance enhancement.
Rician factors of the transmitter-RIS channel in Fig. 8. We find Particularly, the group connected reconfigurable impedance
that the power gain decreases with the Rician factor. To show network with group size of 2 achieves a good tradeoff between
the tradeoff between performance and complexity for the complexity and performance enhancement, which uses half
single, group, and fully connected reconfigurable impedance more reconfigurable impedance components and 1.25 times
networks, we quantify the complexity from two perspectives, more computations to improve the received signal power by
1240 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 2, FEBRUARY 2022

around 20%. In addition, given the same received signal power,

using the group connected reconfigurable impedance network
can reduce the number of RIS elements. For example, for the
group size of 2 and 4 with Rician factor of 0 dB, the number of
RIS elements can be reduced by 9.5% and 13.6%, respectively.
Such reduction of the number of RIS elements is beneficial
for reducing the cost and area of RIS, especially when the
number of RIS elements is large.
Fig. 9. An arbitrary 1-port network.

VI. C ONCLUSION AND F UTURE W ORK

We use scattering parameter network analysis to derive a communication model, from the channel matrix expression
physical and EM compliant yet straightforward and tractable (29), we need to first measure the impedance mismatch ΓT
RIS aided communication model. The proposed general RIS and ΓR and mutual coupling ST T , SII , and SRR by a
aided communication model fully considers the effects of vector network analyzer and then use the channel estimation
impedance mismatching and mutual coupling at the transmit- methods [8], [42] to estimate the channel matrix. In the future,
ter, RIS, and receiver, and thus is more comprehensive than we can develop more efficient channel estimation methods for
the conventional RIS aided communication model [4], [5], the proposed RIS aided communication models.
[7]–[30] which does not consider these effects. Furthermore, 2) Extending to multi-user and multi-cell scenarios. Pre-
the proposed general model can be reduced to the conventional vious work on multi-user [8] and multi-cell scenarios [10]
RIS aided communication model [4], [7]–[30] under special only use RIS with the single connected architecture to mini-
conditions. mize the transmit power and maximize the weight sum rate,
Using the proposed RIS model we also develop new RIS respectively. In the future, we can consider using the fully and
architectures based on group and fully connected reconfig- group connected architectures, which are more general than the
urable impedance networks, which are more general and effi- single connected architecture, to further decrease the transmit
cient than previously utilized single connected architecture [4], power in multi-user scenario and increase the weighted sum
[5], [7]–[30]. In sharp contrast with the single connected rate in multi-cell scenario.
architecture that only adjusts the phases of the impinging 3) Optimizing with discrete values of Θ. For the single con-
waves using a diagonal scattering matrix, our proposed group nected architecture, we have Θ = diag ejθ1 , ejθ2 , . . . , ejθNI
and fully connected architectures enable scattering matrices to and we can restrict the continuous θnI ∈ [0, 2π] to discrete
be block diagonal or full and can adjust not only the phases but values, which is called discrete phase shifts. The optimization
also the magnitudes of the impinging waves, so as to provide with discrete phase shifts has been well studied in [25], [26].
better performance in RIS aided systems. Inspired by the discrete phase shifts, in the future we can
We derive the scaling law of the received signal power of consider the design and optimization of discrete values of
a SISO RIS aided system as a function of the number of RIS matrix Θ for the fully and group connected architectures.
elements in both LoS and Rayleigh fading channels. It shows
that using fully and group connected reconfigurable impedance A PPENDIX
networks can increase the received signal power by up to 62% Scattering parameter network theory [38] is useful to model
compared with the single connected case. It also indicates that and analyze wireless systems. This theory has been used to
given the same received signal power, using fully connected accurately characterize MIMO wireless systems in previous
and group connected reconfigurable impedance networks can work [37], [43], [44] for example. In this appendix, we briefly
reduce the number of RIS elements by up to 21%. We also review the concept of reflection coefficient and scattering
formulate the received signal power maximization problem in parameters to help provide some background [38].
the SISO RIS aided system and evaluate the received signal
power in a realistic model with distance-dependent pathloss
and Rician fading channel. The numerical results show that the A. Reflection Coefficient
fully and group connected reconfigurable impedance networks Consider an arbitrary 1-port network as shown in Fig. 9.
can increase the received signal power by up to 48% and 34%, The 1-port network can be a source impedance, or a load
respectively. impedance, or an antenna impedance, and it can be constructed
Future research avenues include, but are not limited to, from wires, transmission lines, circuits, antennas, or more
the following areas: generally it can be any linear electromagnetic system [38].
1) Developing efficient channel estimation methods. For Assume that an incident voltage wave, denoted as a1 , is input
the proposed RIS aided communication model with perfect into the 1-port network. The incident voltage wave will be
matching and no mutual coupling, the channel matrix (39) is reflected by the 1-port network and subsequently a reflected
exactly the same as the conventional RIS aided communication voltage wave, denoted as b1 , is generated. Therefore, the volt-
model [4], [7]–[30], so that we can use the channel estimation age across the port, denoted as v1 , is the sum of the incident
methods for conventional RIS aided communication model [8], and reflected voltage waves, i.e. v1 = a1 + b1 , and the
[42] to estimate the channel matrix. For the general RIS aided current through the port, denoted as i1 , is described by
SHEN et al.: MODELING AND ARCHITECTURE DESIGN OF RISs 1241

in = (an − bn ) /Z0 . Therefore, the scattering matrix S can be

one-to-one converted to the impedance matrix Z through
−1
S = (Z + Z0 I) (Z − Z0 I) . (75)
For a 1-port network, the scattering matrix S reduces to
a scalar, which is essentially the reflection coefficient as
introduced in Appendix A.
It should be noted that scattering parameters can char-
acterize any N -port network. It does not matter if the
N -port network is constructed from wires, transmission lines,
circuits, antennas, or more generally it can be any linear
electromagnetic system. Particularly, for SISO wireless sys-
tems, the single transmit antenna and single receive antenna
embedded in the wireless channel can be viewed as a 2-port
Fig. 10. An arbitrary N -port network. network, where the [S]2,1 parameter in the 2 × 2 scattering
matrix S is the channel gain between two antennas. More-
over, the scattering parameter network analysis can accurately
i1 = (a1 − b1 ) /Z0 where Z0 is a chosen reference impedance characterize MIMO wireless systems, as shown in [37], [43],
and usually it is set as Z0 = 50 Ω. We define the ratio [44]. In practice, scattering parameters can be measured by a
of the reflected and incident voltage waves as the reflection vector network analyzer [38] and have been widely used in the
coefficient of the 1-port network, which is denoted as Γ and measurements of microwave circuit and component, antennas,
given by and wireless systems. To conclude, reflection coefficient and
b1 Z − Z0 scattering parameters are suitable and accurate for modeling
Γ= = , (74) wireless systems.
a1 Z + Z0
where Z = v1 /i1 denotes the input impedance of the R EFERENCES
1-port network. Γ and Z have a one-to-one correspondence [1] E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M. Alouini,
relationship so that Γ can completely characterize any input and R. Zhang, “Wireless communications through reconfigurable
impedance of a 1-port network. For a passive input impedance intelligent surfaces,” IEEE Access, vol. 7, pp. 116753–116773,
2019.
 {Z} ≥ 0, we have that |Γ| ≤ 1. Particularly, for a pure [2] S. Gong et al., “Towards smart wireless communications via intelligent
reactive input impedance  {Z} = 0, we have that Γ = ejθ reflecting surfaces: A contemporary survey,” 2019, arXiv:1912.07794.
and |Γ| = 1, which is helpful to increase the power of scattered [Online]. Available: http://arxiv.org/abs/1912.07794
[3] Q. Wu, S. Zhang, B. Zheng, C. You, and R. Zhang, “Intelli-
wave and is the key property of RIS (the phase shift and the gent reflecting surface aided wireless communications: A tutorial,”
unit modulus constraint). 2020, arXiv:2007.02759. [Online]. Available: http://arxiv.org/abs/2007.
It should also be noted that the reflection coefficient can 02759
[4] W. Qingqing and Z. Rui, “Towards smart and reconfigurable envi-
characterize any 1-port network no matter how it is con- ronment: Intelligent reflecting surface aided wireless network,” IEEE
structed. Particularly, the reflection coefficient of an antenna Commun. Mag., vol. 58, no. 1, pp. 106–112, Jan. 2019.
characterizes how much of the incident wave the antenna can [5] N. Rajatheva et al., “White paper on broadband connec-
tivity in 6G,” 2020, arXiv:2004.14247. [Online]. Available:
radiate, which is an important parameter in antenna design for https://arxiv.org/abs/2004.14247
wireless systems. [6] S. Jin, R. McKay, C. Zhong, and K.-K. Wong, “Ergodic capacity analysis
of amplify-and-forward MIMO dual-hop systems,” IEEE Trans. Inf.
Theory, vol. 56, no. 5, pp. 2204–2224, May 2010.
[7] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and
B. Scattering Parameters C. Yuen, “Reconfigurable intelligent surfaces for energy efficiency in
Generalizing the concept of the 1-port network, we consider wireless communication,” IEEE Trans. Wireless Commun., vol. 18, no. 8,
pp. 4157–4170, Aug. 2019.
an arbitrary N -port network as shown in Fig. 10, where [8] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless
an refers to the wave incident on the nth port and bn network via joint active and passive beamforming,” IEEE Trans. Wireless
refers to the wave reflected from the nth port. Denote a = Commun., vol. 18, no. 11, pp. 5394–5409, Nov. 2019.
T T [9] B. Ning, Z. Chen, W. Chen, and J. Fang, “Beamforming optimiza-
[a1 , a2 , . . . , aN ] and b = [b1 , b2 , . . . , bn ] . The scattering tion for intelligent reflecting surface assisted MIMO: A sum-path-gain
parameter matrix S is defined in relation to these incident maximization approach,” IEEE Wireless Commun. Lett., vol. 9, no. 7,
and reflected waves as b = Sa. Similar to the impedance pp. 1105–1109, Jul. 2020.
[10] C. Pan et al., “Multicell MIMO communications relying on intelligent
or admittance parameter matrix for an N -port network, the reflecting surfaces,” IEEE Trans. Wireless Commun., vol. 19, no. 8,
scattering parameter matrix can completely characterize the pp. 5218–5233, Aug. 2020.
network as seen at its N ports. While the impedance and [11] G. Zhou, C. Pan, H. Ren, K. Wang, and A. Nallanathan, “Intelligent
reflecting surface aided multigroup multicast MISO communication
admittance parameter matrices relate the total voltages and systems,” IEEE Trans. Signal Process., vol. 68, pp. 3236–3251, 2020,
currents at the ports, the scattering parameter matrix relates the doi: 10.1109/TSP.2020.2990098.
waves incident on the ports to those reflected from the ports. [12] Y. Yang, B. Zheng, S. Zhang, and R. Zhang, “Intelligent reflect-
ing surface meets OFDM: Protocol design and rate maximiza-
In particular, at the nth port, the voltage and current are related tion,” IEEE Trans. Commun., vol. 68, no. 7, pp. 4522–4535,
with the incident and reflected waves by vn = an + bn and Mar. 2020.
1242 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 2, FEBRUARY 2022

[13] B. Zheng and R. Zhang, “Intelligent reflecting surface-enhanced [33] S. Abeywickrama, R. Zhang, and C. Yuen, “Intelligent reflecting surface:
OFDM: Channel estimation and reflection optimization,” Practical phase shift model and beamforming optimization,” in Proc.
IEEE Wireless Commun. Lett., vol. 9, no. 4, pp. 518–522, IEEE Int. Conf. Commun. (ICC), Jun. 2020, pp. 1–6.
Apr. 2020. [34] W. Cai, H. Li, M. Li, and Q. Liu, “Practical modeling and beamforming
[14] B. Zheng, Q. Wu, and R. Zhang, “Intelligent reflecting surface-assisted for intelligent reflecting surface aided wideband systems,” IEEE Com-
multiple access with user pairing: NOMA or OMA?” IEEE Commun. mun. Lett., vol. 24, no. 7, pp. 1568–1571, Jul. 2020.
Lett., vol. 24, no. 4, pp. 753–757, Apr. 2020. [35] E. Björnson and L. Sanguinetti, “Demystifying the power scaling law
[15] X. Yu, D. Xu, Y. Sun, D. W. K. Ng, and R. Schober, “Robust of intelligent reflecting surfaces and metasurfaces,” in Proc. IEEE 8th
and secure wireless communications via intelligent reflecting sur- Int. Workshop Comput. Adv. Multi-Sensor Adapt. Process. (CAMSAP),
faces,” IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2637–2652, Dec. 2019, pp. 549–553.
Nov. 2020. [36] E. Björnson, O. Özdogan, and E. G. Larsson, “Intelligent reflecting
[16] W. Zhao, G. Wang, S. Atapattu, T. A. Tsiftsis, and X. Ma, “Performance surface versus decode-and-forward: How large surfaces are needed to
analysis of large intelligent surface aided backscatter communication beat relaying?” IEEE Wireless Commun. Lett., vol. 9, no. 2, pp. 244–248,
systems,” IEEE Wireless Commun. Lett., vol. 9, no. 7, pp. 962–966, Feb. 2020.
Jul. 2020. [37] J. W. Wallace and M. A. Jensen, “Mutual coupling in MIMO wireless
[17] Q. Wu and R. Zhang, “Joint active and passive beamforming opti- systems: A rigorous network theory analysis,” IEEE Trans. Wireless
mization for intelligent reflecting surface assisted SWIPT under QoS Commun., vol. 3, no. 4, pp. 1317–1325, Jul. 2004.
constraints,” IEEE J. Sel. Areas Commun., vol. 38, no. 8, pp. 1735–1748, [38] D. M. Pozar, Microwave Engineering. Hoboken, NJ, USA: Wiley,
Aug. 2020. 2009.
[39] S. Shen and R. D. Murch, “Impedance matching for compact multiple
[18] Q. Wu and R. Zhang, “Weighted sum power maximization for intelligent
antenna systems in random RF fields,” IEEE Trans. Antennas Propag.,
reflecting surface aided SWIPT,” IEEE Wireless Commun. Lett., vol. 9,
vol. 64, no. 2, pp. 820–825, Feb. 2016.
no. 5, pp. 586–590, May 2020.
[40] G. Gradoni and M. Di Renzo, “End-to-end mutual coupling aware
[19] X. Guan, Q. Wu, and R. Zhang, “Joint power control and passive communication model for reconfigurable intelligent surfaces: An
beamforming in IRS-assisted spectrum sharing,” IEEE Commun. Lett., electromagnetic-compliant approach based on mutual impedances,”
vol. 24, no. 7, pp. 1553–1557, Jul. 2020. 2020, arXiv:2009.02694. [Online]. Available: http://arxiv.org/
[20] J. Yuan, Y.-C. Liang, J. Joung, G. Feng, and E. G. Larsson, “Intelligent abs/2009.02694
reflecting surface-assisted cognitive radio system,” IEEE Trans. Com- [41] J. Nocedal and S. Wright, Numerical Optimization. New York, NY, USA:
mun., vol. 69, no. 1, pp. 675–687, Jan. 2021. Springer, 2006.
[21] L. Ge, P. Dong, H. Zhang, J.-B. Wang, and X. You, “Joint beam- [42] D. Mishra and H. Johansson, “Channel estimation and low-complexity
forming and trajectory optimization for intelligent reflecting surfaces- beamforming design for passive intelligent surface assisted MISO wire-
assisted UAV communications,” IEEE Access, vol. 8, pp. 78702–78712, less energy transfer,” in Proc. IEEE Int. Conf. Acoust., Speech Signal
2020. Process. (ICASSP), May 2019, pp. 4659–4663.
[22] X. Lu, W. Yang, X. Guan, Q. Wu, and Y. Cai, “Robust and secure [43] J. W. Wallace and M. A. Jensen, “Termination-dependent diver-
beamforming for intelligent reflecting surface aided mmWave MISO sity performance of coupled antennas: Network theory analy-
systems,” IEEE Wireless Commun. Lett., vol. 9, no. 12, pp. 2068–2072, sis,” IEEE Trans. Antennas Propag., vol. 52, no. 1, pp. 98–105,
Dec. 2020. Jan. 2004.
[23] J. Qiao and M.-S. Alouini, “Secure transmission for intelligent reflecting [44] B. K. Lau, J. B. Andersen, G. Kristenson, and A. F. Molisch,
surface-assisted mmWave and terahertz systems,” IEEE Wireless Com- “Impact of matching network on bandwidth of compact antenna
mun. Lett., vol. 9, no. 10, pp. 1743–1747, Oct. 2020. arrays,” IEEE Trans. Antennas Propag., vol. 54, no. 11, pp. 3225–3238,
[24] T. Bai, C. Pan, Y. Deng, M. Elkashlan, A. Nallanathan, and L. Hanzo, Nov. 2006.
“Latency minimization for intelligent reflecting surface aided mobile
edge computing,” IEEE J. Sel. Areas Commun., vol. 38, no. 11,
pp. 2666–2682, Nov. 2020.
[25] J. Xu, W. Xu, and A. L. Swindlehurst, “Discrete phase shift design for
practical large intelligent surface communication,” in Proc. IEEE Pacific
Rim Conf. Commun., Comput. Signal Process. (PACRIM), Aug. 2019,
pp. 1–5.
[26] Q. Q. Wu and R. Zhang, “Beamforming optimization for wire-
less network aided by intelligent reflecting surface with discrete
phase shifts,” IEEE Trans. Commun., vol. 68, no. 3, pp. 1838–1851,
May 2020.
[27] Y. Han, W. Tang, S. Jin, C. Wen, and X. Ma, “Large intel-
ligent surface-assisted wireless communication exploiting statistical
CSI,” IEEE Trans. Veh. Technol., vol. 68, no. 8, pp. 8238–8242,
Jun. 2019.
[28] G. Zhou, C. Pan, H. Ren, K. Wang, and A. Nallanathan, “A framework
of robust transmission design for IRS-aided MISO communications with
imperfect cascaded channels,” IEEE Trans. Signal Process., vol. 68,
pp. 5092–5106, 2020, doi: 10.1109/TSP.2020.3019666.
Shanpu Shen (Member, IEEE) received the bach-
[29] C. Huang, R. Mo, and Y. Yuen, “Reconfigurable intelligent surface
elor’s degree in communication engineering from
assisted multiuser MISO systems exploiting deep reinforcement learn-
Nanjing University of Science and Technology,
ing,” IEEE J. Sel. Areas Commun., vol. 38, no. 8, pp. 1839–1850,
Nanjing, China, in 2013, and the Ph.D. degree
Jun. 2020.
in electronic and computer engineering from The
[30] K. Feng, Q. Wang, X. Li, and C. Wen, “Deep reinforcement learning Hong Kong University of Science and Technology
based intelligent reflecting surface optimization for MISO communica- (HKUST), Hong Kong, in 2017.
tion systems,” IEEE Wireless Commun. Lett., vol. 9, no. 5, pp. 745–749, He was a Visiting Ph.D. Student with the
May 2020. Microsystems Technology Laboratories, Massa-
[31] O. Özdogan, E. Björnson, and E. G. Larsson, “Intelligent chusetts Institute of Technology, Cambridge, MA,
reflecting surfaces: Physics, propagation, and pathloss modeling,” USA, in 2016. He was a Post-Doctoral Fellow with
IEEE Wireless Commun. Lett., vol. 9, no. 5, pp. 581–585, HKUST from 2017 to 2018, and a Post-Doctoral Research Associate with
May 2020. the Communications and Signal Processing Group, Imperial College London,
[32] W. Tang et al., “Wireless communications with reconfigurable London, U.K., from 2018 to 2020. He is currently a Research Assistant
intelligent surface: Path loss modeling and experimental measure- Professor with HKUST. His current research interests include RF energy
ment,” IEEE Trans. Wireless Commun., vol. 20, no. 1, pp. 421–439, harvesting, wireless power transfer, the Internet-of-Things, MIMO systems,
Jan. 2021. and antenna design and optimization.
SHEN et al.: MODELING AND ARCHITECTURE DESIGN OF RISs 1243

Bruno Clerckx (Senior Member, IEEE) received Ross Murch (Fellow, IEEE) received the bache-
the M.S. and Ph.D. degrees in electrical engi- lor’s and Ph.D. degrees in electrical and electronic
neering from the Université Catholique de Lou- engineering from the University of Canterbury, New
vain, Louvain-la-Neuve, Belgium, in 2000 and Zealand.
2005, respectively. He is currently a Full Pro- He was the Department Head with The Hong Kong
fessor, the Head of the Wireless Communications University of Science and Technology (HKUST)
and Signal Processing Laboratory, and the Deputy from 2009 to 2015. He spent sabbaticals at MIT,
Head of the Communications and Signal Process- USA; AT&T, USA; Allgon Mobile Communica-
ing Group, within the Electrical and Electronic tions, Sweden; and Imperial College London. He is
Engineering Department, Imperial College London, currently the Chair Professor with the Department of
London, U.K. From 2006 to 2011, he was with Electronic and Computer Engineering and a Senior
Samsung Electronics, Suwon, South Korea, where he actively contributed to Fellow with the Institute for Advanced Study, HKUST. He has successfully
4G (3GPP LTE/LTE-A and IEEE 802.16m) and acted as a Rapporteur for the supervised over 50 research graduate student. His research contributions
3GPP Coordinated Multi-Point (CoMP) Study Item. Since 2011, he has been include more than 20 patents. His current research interests include the
with Imperial College London, first as a Lecturer from 2011 to 2015, a Senior Internet-of-Things, RF imaging, ambient RF systems, energy harvesting,
Lecturer from 2015 to 2017, a Reader from 2017 to 2020. From 2014 to multiport antenna systems, and reconfigurable intelligent surfaces. His unique
2016, he was an Associate Professor with Korea University, Seoul, South expertise lies in his combination of knowledge from both wireless communi-
Korea. He also held various long or short-term visiting research appointments cation systems and electromagnetic areas.
at Stanford University, EURECOM, the National University of Singapore, He is an IET and HKIE Fellow. He has been a David Bensted Fellow with
The University of Hong Kong, Princeton University, The University of Simon Fraser University, Canada, and an HKTIIT Fellow with Southamp-
Edinburgh, the University of New South Wales, and Tsinghua University. ton University, U.K. He received several awards, including the Computer
He has authored two books on MIMO Wireless Communications and MIMO Simulation Technology (CST) University Publication Award. He served at
Wireless Networks: Channels, Techniques and Standards for Multi-Antenna, IEEE in various positions, including an IEEE Area Editor, the Technical
Multi-User and Multi-Cell Systems, 200 peer-reviewed international research Program Chair, a Distinguished Lecturer, and a Fellow Evaluation Committee
articles, and 150 standards contributions, and an Inventor of 80 issued or Member. He also enjoys teaching and received two teaching awards.
pending patents among which 15 have been adopted in the specifications of 4G
standards and are used by billions of devices worldwide. His research area
is communication theory and signal processing for wireless networks. He has
been a TPC member, a symposium chair, or a TPC chair of many symposia
on communication theory and signal processing for communication, and
wireless communication for several leading international IEEE conferences.
He was an Elected Member of the IEEE Signal Processing Society SPCOM
Technical Committee. He served as an Editor for the IEEE T RANSACTIONS
ON C OMMUNICATIONS, the IEEE T RANSACTIONS ON W IRELESS C OM -
MUNICATIONS , and the IEEE T RANSACTIONS ON S IGNAL P ROCESSING .
He was an Editor of the 3GPP LTE-Advanced Standard Technical Report
on CoMP. He has been a Lead Guest Editor of special issues, such as
the EURASIP Journal on Wireless Communications and Networking, IEEE
A CCESS , the IEEE J OURNAL ON S ELECTED A REAS IN C OMMUNICATIONS ,
the IEEE J OURNAL OF S ELECTED T OPICS IN S IGNAL P ROCESSING, and the
P ROCEEDINGS OF THE IEEE. He is also an IEEE Communications Society
Distinguished Lecturer for the term 2021–2022.