1

Beyond Diagonal Reconfigurable Intelligent

Surfaces Utilizing Graph Theory: Modeling,
Architecture Design, and Optimization
Matteo Nerini, Graduate Student Member, IEEE, Shanpu Shen, Senior Member, IEEE,
Hongyu Li, Graduate Student Member, IEEE, Bruno Clerckx, Fellow, IEEE

Abstract—Recently, beyond diagonal reconfigurable intelligent In a conventional RIS architecture, each element is indepen-
arXiv:2305.05013v1 [cs.IT] 8 May 2023

surface (BD-RIS) has been proposed to generalize conventional dently controlled by a tunable impedance connected to ground
RIS. BD-RIS has a scattering matrix that is not restricted to [4]. Such architecture, referred to as the single-connected
being diagonal and thus brings a performance improvement over
conventional RIS. While different BD-RIS architectures have RIS, results in a diagonal scattering matrix, which is also
been proposed, it still remains an open problem to develop a commonly known as the phase shift matrix. Conventional
systematic approach to design BD-RIS architectures achieving the RIS with diagonal scattering matrix has been widely used
optimal trade-off between performance and circuit complexity. In to enhance different wireless systems, such as single-cell
this work, we propose novel modeling, architecture design, and [5], [6], [7] and multi-cell communications [8], orthogonal
optimization for BD-RIS based on graph theory. This graph the-
oretical modeling allows us to develop two new efficient BD-RIS frequency division multiplexing (OFDM) [9], non-orthogonal
architectures, denoted as tree-connected and forest-connected multiple access (NOMA) [10], rate splitting multiple access
RIS. Tree-connected RIS, whose corresponding graph is a tree, (RSMA) [11], radio frequency (RF) sensing systems [12],
is proven to be the least complex BD-RIS architecture able to wireless power transfer (WPT) systems [13], and simultaneous
achieve the performance upper bound in multiple-input single- wireless information and power transfer (SWIPT) systems
output (MISO) systems. Besides, forest-connected RIS allows
us to strike a balance between performance and complexity, [14]. In addition, research has been conducted to optimize
further decreasing the complexity over tree-connected RIS. To RIS based on imperfect channel state information (CSI) [15],
optimize tree-connected RIS, we derive a closed-form global [16], in multi-RIS scenarios [17], [18], and with discrete phase
optimal solution, while forest-connected RIS is optimized through shifts to account for practical hardware impairments [19], [20].
a low-complexity iterative algorithm. Numerical results confirm Furthermore, RIS prototypes have been realized in [21], [22].
that tree-connected (resp. forest-connected) RIS achieves the
same performance as fully-connected (resp. group-connected) Recently, the conventional RIS architecture considered in
RIS, while reducing the complexity by up to 16.4 times. [5]-[22] has been generalized with the introduction of beyond
diagonal RIS (BD-RIS). BD-RIS has a scattering matrix that is
Index Terms—Beyond diagonal reconfigurable intelligent sur-
face (BD-RIS), forest-connected, graph theoretical modeling, tree- not restricted to being diagonal [23], with various architectures
connected. and modes available as depicted in the classification tree
in Fig. 1. Specifically, group-connected and fully-connected
architectures, whereby some or all RIS elements are connected
I. I NTRODUCTION with each other through tunable impedance components, are
Reconfigurable intelligent surface (RIS) is expected to be a proposed in [4], generalizing the single-connected architecture
key technology in 6G to enhance the performance of wireless to enhance the RIS performance. Subsequently, group- and
systems in an efficient and cost-effective manner [1], [2], [3]. fully-connected RIS with discrete values are efficiently opti-
An RIS consists of a large number of reconfigurable scattering mized in [24], while the closed-form global optimal solutions
elements, each with the capability to manipulate the phase for group- and fully-connected RIS with continuous values
of an incident electromagnetic wave, so that the phase shifts are derived in [25]. In addition, the simultaneously trans-
of these elements can be coordinated to direct the scattered mitting and reflecting RIS (STAR-RIS) or intelligent omni-
electromagnetic signal toward the intended receiver. Owing surface (IOS) is presented in [26], [27], which differs from
to the benefits of low power consumption and cost-effective conventional RIS as signals impinging on this RIS can be both
architecture, RIS has gained widespread attention. reflected and transmitted through the RIS for a full space cov-
erage. STAR-RIS is generalized in [28] which unifies different
Corresponding author: Shanpu Shen. BD-RIS modes (reflective/transmissive/hybrid) and different
M. Nerini and H. Li are with the Department of Electrical and Electronic BD-RIS architectures (single-/group-/fully-connected). To fur-
Engineering, Imperial College London, London SW7 2AZ, U.K. (e-mail: ther enhance RIS performance while maintaining a full-space
{m.nerini20, c.li21}@imperial.ac.uk).
S. Shen is with the Department of Electronic and Computer Engineering, coverage, multi-sector BD-RIS is proposed in [29], where the
The Hong Kong University of Science and Technology, Clear Water Bay, elements are split into multiple sectors with each covering a
Kowloon, Hong Kong (e-mail: [email protected]). sector of space. Further, the synergy of multi-sector BD-RIS
B. Clerckx is with the Department of Electrical and Electronic Engineering,
Imperial College London, London SW7 2AZ, U.K., and with Silicon Austria and RSMA is proved to enlarge the coverage, improve the
Labs (SAL), Graz A-8010, Austria (e-mail: [email protected]). performance, and save on antennas in multi-user systems [30].
 2

Fig. 1. RIS classification tree.

Departing from existing works on BD-RIS with fixed archi- circuit complexity significantly reduced by 16.4 times, when
tectures [4], [23]-[30], dynamically group-connected RIS has 64 RIS elements are considered.
been proposed in [31], outperforming fixed group-connected Third, we prove that there always exists one and only one
RIS by dynamically dividing the RIS elements into groups global optimal solution to optimize tree-connected RIS. To
depending on the channel realization. Besides, a BD-RIS with find this solution, we derive a closed-form algorithm valid
a non-diagonal phase shift matrix has been proposed in [32], for any tree-connected architecture. With this algorithm, tree-
which is able to achieve a higher rate than conventional RIS. connected RIS achieves a performance improvement of 51.7%
Among all the BD-RISs [4], [23]-[32], the fully-connected over single-connected RIS.
architecture provides the highest flexibility and best perfor- Fourth, we propose forest-connected RIS as an additional
mance, which is however at the expense of high circuit BD-RIS family to reach a trade-off between performance
complexity since it has a large number of tunable impedance and circuit complexity. Forest-connected RIS serves as a
components. Although the group-connected architecture can bridge between single- and tree-connected RIS. We pro-
simplify the circuit complexity, it still remains an open pose a low-complexity iterative algorithm to optimize forest-
problem to develop a systematic approach to design BD- connected RIS. Numerical results show that forest-connected
RIS architecture achieving the optimal trade-off between per- RIS achieves the same performance as group-connected RIS,
formance and circuit complexity, that is achieving the best while reducing the circuit complexity by 2.4 times.
performance with the least complexity. To address this issue, Organization: In Section II, we introduce the graph theoret-
in this work, we propose novel modeling, architecture design, ical modeling of BD-RIS. In Section III, we derive tree- and
and optimization for BD-RIS by utilizing graph theory. The forest-connected RIS architectures and provide two examples
contributions of this paper are summarized as follows. of tree-connected RIS. In Section IV we provide a global
First, we propose a novel modeling of BD-RIS architectures optimal closed-form solution to optimize tree-connected RIS
utilizing graph theory. To the authors’ best knowledge, it is and an iterative approach to optimize forest-connected RIS.
the first time that BD-RIS (or any other RIS) architectures are In Section V, we evaluate the performance of the proposed
modeled through graph theory. Specifically, we model each BD-RISs. Finally, Section VI concludes this work.
BD-RIS architecture by a graph, whose vertices character- Notation: Vectors and matrices are denoted with bold lower
ize the RIS ports and whose edges characterize the tunable and bold upper letters, respectively. Scalars are represented
impedance components connecting the RIS ports. This graph with letters not in bold font. < {a}, = {a}, |a|, arg (a), and
theoretical modeling lays down the foundations for developing a∗ refer to the real part, imaginary part, modulus, phase, and
new efficient BD-RIS architectures. the complex conjugate of a complex scalar a, respectively. [a]i
Second, we derive a necessary and sufficient condition for a and kak refer to the ith element and l2 -norm of a vector a,
BD-RIS architecture to achieve the performance upper bound respectively. AT , AH , [A]i,j , and kAk refer to the transpose,
in multiple-input single-output (MISO) systems. Consequently, conjugate transpose, (i, j)th element, and l2 -norm of a matrix
we characterize the least complex BD-RISs achieving such an A, respectively. A ∼ B means that the matrices A and B
upper bound. The resulting least-complexity BD-RIS architec- are equivalent. R and√C denote real and complex number
tures are referred to as tree-connected RIS since their corre- set, respectively. j = −1 denotes imaginary unit. 0 and I
sponding graph is a tree. In addition, two specific examples denote an all-zero matrix and an identity matrix, respectively.
of tree-connected RIS are proposed, namely tridiagonal RIS CN (0, I) denotes the distribution of a circularly symmetric
and arrowhead RIS. Tree-connected RIS achieves the same complex Gaussian random vector with mean vector 0 and
performance in MISO systems as fully-connected RIS with a covariance matrix I and ∼ stands for “distributed as”. diag(a)
 3

refers to a diagonal matrix with diagonal elements being the where Yn and θn are respectively the tunable admittance and
vector a. diag(A1 , . . . , AN ) refers to a block diagonal matrix phase shift for the nth RIS port for n = 1, . . . , N . Such circuit
with blocks being A1 , . . . , AN . topology is also referred to as single-connected RIS [4] and
has been widely adopted [5]-[22].
II. BD-RIS M ODELING U TILIZING G RAPH T HEORY B. Beyond Diagonal RIS
Consider a MISO system with an M -antenna transmitter In BD-RIS, the RIS ports can also be connected to each
and a single-antenna receiver, which is aided by an N -element other through additional tunable admittance components. We
RIS. The N -element RIS can be modeled as N antennas denote the tunable admittance connecting the nth port to the
connected to an N -port reconfigurable impedance network mth port as Yn,m . According to [34], given the admittance
that is characterized by the scattering matrix Θ ∈ CN ×N components Yn and Yn,m , the (n, m)th entry of the admittance
[4]. We assume the direct channel between the transmitter matrix Y is given by
and receiver is blocked and thus is negligible compared to (
the indirect channel provided by RIS. Therefore, the overall −Yn,m n 6= m
[Y]n,m = P . (6)
channel h ∈ C1×M between the transmitter and receiver can Yn + k6=n Yn,k n = m
be written as P
Thus, by selecting Yn,m = − [Y]n,m and Yn = k [Y]n,k ,
h = hRI ΘHIT , (1)
we can implement an arbitrary symmetric Y so that the
where hRI ∈ C1×N and HIT ∈ CN ×M are the channel matri- admittance matrix Y and the scattering matrix Θ are beyond
ces from the RIS to the receiver, and from the transmitter to the diagonal. Particularly, there are two special categories of BD-
RIS, respectively. We denote the transmit signal as x = ws, RIS, namely fully-connected and group-connected RIS [4].
where w ∈ CM ×1 is the precoder satisfying kwk = 1 and Due to the enhanced flexibility in Y and Θ, BD-RIS is proven
2
s ∈ C is the transmit symbol with power PT = E[|s| ]. Hence, to outperform single-connected RIS [4], [23]-[32].
the received signal is given by y = hx + n, where n is the
additive white Gaussian noise (AWGN). C. Graph Theoretical Modeling
The N -port reconfigurable impedance network consists of To model the general circuit topology of BD-RIS, we resort
tunable passive impedance components. In addition to S- for the first time to graph theoretical tools [35]. Specifically,
parameters (the scattering matrix), the N -port reconfigurable we represent the circuit topology of BD-RIS through a graph
impedance network can be also characterized by Y-parameters
G = (V, E) , (7)
[33], that is the admittance matrix denoted as Y ∈ CN ×N .
According to microwave network theory [33], the scattering where V represents the vertex set of G and is given by the set
matrix Θ and admittance matrix Y are related by of indices of RIS ports, i.e.
−1
Θ = (I + Z0 Y) (I − Z0 Y) , (2) V = {1, 2, . . . , N } , (8)
where Z0 denotes the reference impedance used for computing and E represents the edge set of G and is given by
the scattering parameter, and usually set as Z0 = 50 Ω.
Furthermore, we have Y = YT and Θ = ΘT because of E = {(n` , m` ) | n` , m` ∈ V, Yn` ,m` 6= 0, n` 6= m` } . (9)
the reciprocity of the reconfigurable impedance network. To Thus, there exists an edge between vertex n` and vertex m`
maximize the power scattered by the RIS, Y should be purely if and only if there is a tunable admittance connecting port
susceptive and thus writes as Y = jB, where B ∈ RN ×N n` and port m` , namely the `th admittance. Accordingly,
denotes the susceptance matrix of the N -port reconfigurable given a graph G with edge set E = {(n` , m` )}`=1 , the
L
impedance network. Hence, Θ is given by corresponding BD-RIS admittance matrix has 2L non-zero off-
Θ = (I + jZ0 B)
−1
(I − jZ0 B) . (3) diagonal entries, i.e. [Y] n` ,m` and [Y] m` ,n` for ` = 1, . . . , L,
while the other off-diagonal entries are zero.
Depending on the circuit topology of the N -port recon- To better clarify our graph theoretical modeling of BD-RIS,
figurable impedance network, the scattering matrix Θ and we show a 4-element single-connected RIS with its associated
admittance matrix Y satisfy different constraints, which results graph in Fig. 2(a), and four examples of 4-element BD-RISs
in diagonal RIS and beyond diagonal RIS as follows. with their associated graphs in Fig. 2(b)-(e). We also use Fig. 2
to illustrate the graph theoretical definitions [35] that are used
in this work, as briefly summarized in the following.
A. Diagonal RIS
• A graph is empty when there is no edge.
In diagonal RIS, each RIS port is not connected to the other • A graph is connected when there is a path, i.e., a finite
ports and is connected to ground through a tunable admittance sequence of distinct edges joining a sequence of distinct
[4]. Accordingly, the admittance matrix Y and the scattering vertices, from any vertex to any other vertex.
matrix Θ are diagonal, written as • A graph that is not connected is called disconnected.
• A cyclic graph is a graph containing at least one cycle,
Y = diag (Y1 , Y2 , . . . , YN ) , (4)
i.e., a finite sequence of distinct edges joining a sequence
Θ = diag ejθ1 , ejθ2 , . . . , e jθN


, (5) of vertices, where only the first and last vertices are equal.
 4

ple1 . Therefore, graph theory is effective to characterize and

model the circuit topology of BD-RIS and the corresponding
admittance matrix Y.
Utilizing graph theory, we propose two novel families of
BD-RIS with remarkable properties, which are referred to as
tree-connected RIS and forest-connected RIS, in the following
section.

III. BD-RIS A RCHITECTURE D ESIGN U TILIZING G RAPH

T HEORY
A. Tree-Connected RIS
As a starting point, we investigate how the graph of BD-RIS
influences the system performance. To that end, we formulate
the BD-RIS optimization problem for MISO systems utilizing
graph theory. Specifically, we assume the BD-RIS is lossless
and purely susceptive and aim to maximize the received signal
2
power PR = PT |hRI ΘHIT w| by jointly optimize w and Θ.
Thus, we can formulate the problem as
2
max PT |hRI ΘHIT w| (10)
Θ,w
−1
s.t. Θ = (I + jZ0 B) (I − jZ0 B) , (11)
T
B=B , (12)
B ∈ BG , (13)
where BG represents the set of possible susceptance matrices
of BD-RISs characterized by the graph G = (V, E) and is
written as
BG = {B| [B] n,m = 0, n 6= m, n, m ∈ V, (n, m) ∈
/ E} (14)
which means that an off-diagonal entry is zero if there exists
no corresponding edge in the graph.
From (10)-(13), we can deduce that the maximum received
signal power depends on the graph of the BD-RIS. Specifically,
when the graph G has more edges, the susceptance matrix B
is more flexible so that the maximum received signal power
is higher. However, the enhanced performance is achieved at
the expense of higher circuit complexity (more connections
Fig. 2. Examples of 4-port BD-RIS architectures (left) and their corresponding from the increased edges). Thus, it is worth designing the
graphs G (right), where G is (a) empty, (b) disconnected and acyclic, (c)
connected and acyclic, i.e., a tree, (d) connected and cyclic, (e) complete.
graph of BD-RIS to achieve the best performance with the
least complexity.
To that end, we first consider the case that the graph G is
complete, that is the fully-connected RIS. As shown in [4] and
• A graph that is not cyclic is called acyclic, or a forest. [25], maximum ratio transmission (MRT) gives the optimal
• A tree is defined as a connected and acyclic graph. w and subsequently the fully-connected RIS can achieve the
• A graph is complete when each pair of distinct vertices upper bound of the received signal power given by
is joined by an edge. 2 2
• A graph is simple when there is no edge with identical P̄R = PT khRI k kHIT k , (15)
ends (called loops) and no multiple edges joining the following the sub-multiplicativity of the norm [25]. Leveraging
same pair of vertices. this upper bound, herein we define a BD-RIS that achieves
the performance upper bound P̄R for any channel realization
Given these definitions, we recognize that the graph for the as MISO optimal. Thus, the fully-connected RIS is MISO
single-connected RIS in Fig. 2(a) is empty, the graph in optimal. More generally, the following lemma provides a
Fig. 2(b) is disconnected and acyclic, the graph in Fig. 2(c)
is connected and acyclic, i.e., a tree, the graph in Fig. 2(d) is 1 In this work, we consider graphs where the edges are not oriented since

connected and cyclic as it presents the cycle 1 − 2 − 3, and we assume the RIS to have a reciprocal reconfigurable impedance network.
BD-RIS architectures with non-reciprocal reconfigurable impedance networks
the graph for the fully-connected RIS in Fig. 2(e) is complete. could be modeled through directed graphs, i.e., graphs with oriented edges,
In addition, graphs associated with BD-RIS are always sim- but are beyond the scope of this work.
 5

Fig. 3. Tridiagonal RIS with N = 8.

sufficient and necessary condition for a BD-RIS to be MISO

optimal.
Lemma 1. A BD-RIS with associated graph G is MISO
optimal if and only if G is a connected graph. Fig. 4. Arrowhead RIS with N = 5 where port 1 is the central vertex.
Proof. Please refer to Appendix A.
Using Lemma 1, we have the following proposition provid- illustrative example of tridiagonal RIS with N = 8. As a
ing the least complexity for MISO optimal BD-RIS. consequence of such architecture, the susceptance matrix of
Proposition 1. The minimum number of edges in the graph tridiagonal RIS is symmetric and tridiagonal ([B]i,j = 0 if
of a MISO optimal BD-RIS is N − 1, with N denoting the |i − j| > 1), written as
number of RIS elements. 
[B]1,1 [B]1,2 ··· 0

.. ..
Proof. Please refer to Appendix B.
 
[B]1,2 [B]2,2 . . 
B=  .
. (18)
Remarkably, a connected graph G on N vertices with N −1  .. .. .. 
. . [B]N −1,N 
edges is a tree. To show this, we observe that removing 0 ··· [B]N −1,N [B]N,N
any edge from G makes the graph disconnected because of
Proposition 1. Thus, G cannot have a cycle and must be a tree Note that if the port indexes in V Tri are permuted, B is a
(connected and acyclic). Therefore, to minimize the BD-RIS permuted tridiagonal matrix.
circuit complexity while maintaining optimal performance, we 2) Arrowhead RIS: We define arrowhead RIS as tree-
propose tree-connected RIS as BD-RIS whose corresponding connected RIS whose graph G is a star graph. In graph theory,
graph is a tree. According to this definition, tree-connected a star graph is defined as a graph where at most one vertex,
RIS includes N admittance components connecting each port called central, has degree greater than one and all the others
to ground and N − 1 admittance components interconnecting have degree one [35]. Accordingly, the vertex set of the graph
the ports to each other, yielding a total of 2N − 1 admittance associated with arrowhead RIS is defined as
components, which is much less than the N (N + 1)/2 admit-
tance components in fully-connected RIS. V Arrow = {c, n1 , n2 , . . . , nN −1 } , (19)
where c denotes the central vertex, while the edge set is
B. Two Examples of Tree-Connected RIS consequently expressed as
Tree-connected RIS refers to a family of BD-RISs, includ- E Arrow = {(c, n` ) | ` = 1, . . . , N − 1} . (20)
ing multiple possible architectures. To clarify the concept of
In Fig. 4, we provide an illustrative example of arrowhead
tree-connected RIS, in this subsection we provide two specific
RIS with N = 5 where port 1 is the central vertex. As
examples of tree-connected RIS, referred to as tridiagonal and
a consequence, the susceptance matrix of arrowhead RIS is
arrowhead RIS.
symmetric and arrowhead ([B]i,j = 0 if i 6= c, j 6= c,
1) Tridiagonal RIS: We define tridiagonal RIS as tree-
and i 6= j) [36]. Assuming port 1 is the central vertex, the
connected RIS whose graph G is a path graph. In graph theory,
susceptance matrix of arrowhead RIS can be written as
the degree of a vertex is defined as the number of edges
[B]1,1 [B]1,2 · · · [B]1,N
 
incident with it, and a path graph is defined as a graph where
two vertices have degree one and all the others (if any) have  [B]1,2 [B]2,2 · · · 0 
B= . ..  . (21)
 
degree two [35]. The vertex set of the graph associated with . .
 .. .. .. . 
tridiagonal RIS is defined as
[B]1,N 0 · · · [B]N,N
V Tri = {n1 , n2 , . . . , nN } , (16)
Note that if the central vertex is not port 1, B is a permuted
and the edge set can be expressed accordingly as arrowhead matrix.
It should be noted that all tree-connected RISs satisfy the
E Tri = {(n` , n`+1 ) | ` = 1, . . . , N − 1} . (17)
necessary condition to be MISO optimal and contain the same
Interestingly, the tridiagonal architectures have also been pro- number of tunable admittance components. Thus, tridiagonal
posed to realize low-complexity impedance matching networks and arrowhead RIS have the same circuit complexity and
for multiple antenna systems [34]. In Fig. 3, we provide an MISO optimality. In practice, we can select the suitable tree-
 6

connected RIS architecture according to the use case. Tridiago- admittance components in forest-connected RIS with group
nal RIS can be easily implemented with cables or tapes, thanks size NG is N (2 − 1/NG ).
to the path graph circuit topology, so it enables the practical
development of tree-connected RIS with uniform linear arrays
(ULAs) or radio stripes [37]. Conversely, arrowhead BD-RIS IV. O PTIMIZATION OF T REE - AND F OREST-C ONNECTED
is more suitable to develop tree-connected RIS with uniform RIS FOR MISO S YSTEMS
planar arrays (UPAs).
In this section, we optimize tree-connected and forest-
connected RIS to maximize the received signal power in
C. Forest-Connected RIS MISO systems. For tree-connected RIS, we show that any
tree-connected RIS is MISO optimal and provide a closed-
To further reduce the circuit complexity of tree-connected
form global optimal solution. For forest-connected RIS, we
RIS, especially for the large-scale case, we propose forest-
provide a low-complexity iterative solution.
connected RIS. A forest-connected RIS is defined as a BD-
RIS where the N RIS elements are divided into G = N/NG
groups with group size NG and each group utilizes the tree-
A. Tree-Connected RIS Optimization
connected architecture (i.e. the graph for each group is a tree
on NG vertices)2 . Therefore, the graph G associated with the The tree-connected RIS optimization to maximize the re-
forest-connected RIS is a forest, and the corresponding vertex ceived signal power in MISO systems can be formulated as
and edge sets can be represented as (10)-(13) plus a constraint that graph G is a tree. Recall that
the received signal power in BD-RIS-aided MISO systems is
V = ∪G G
g=1 Vg , E = ∪g=1 Eg , (22) 2 2
upper bounded by P̄R = PT khRI k kHIT k . In the follow-
where Vg and Eg denote the vertex and edge sets of the graph ing, we prove that there is always one and only one solution
associated with the gth group for g = 1, . . . , G, with Vg1 ∩ for the optimal susceptance matrix B of tree-connected RIS
Vg2 = ∅ for all g1 6= g2 . to achieve the upper bound P̄R , and propose an algorithm to
More specifically, the tree-connected architecture for each find such optimal B.
group in the forest-connected RIS can be either tridiagonal The key to achieve the upper bound P̄R given by (15) is to
or arrowhead. When the tridiagonal architecture is used, the find a complex symmetric unitary matrix Θ satisfying
vertex and edge sets of the graph associated with forest-
ĥH
RI = ΘuIT , (27)
connected RIS are defined as
where ĥRI = hRI / khRI k and uIT = umax (HIT ) is the
Tri
VForest = ∪G
g=1 {ng,1 , ng,2 , . . . , ng,NG } , (23)
dominant left singular vector of HIT [25]. Substituting (3)
Tri
EForest = ∪G
g=1 {(ng,` , ng,`+1 ) | ` = 1, . . . , NG − 1} , (24) into (27), we can equivalently rewrite (27) as

similarly to (16) and (17). Accordingly, the susceptance matrix (I + jZ0 B) ĥH
RI = (I − jZ0 B) uIT , (28)
of forest-connected RIS with tridiagonal architecture for each which can be expressed in a compact form as
group is block diagonal with each block being symmetric
and tridiagonal. On the other hand, when the arrowhead Bα = β, (29)
architecture is used, the vertex and edge sets of the graph  
N ×1
associated with forest-connected RIS are defined as where α = jZ0 uIT + ĥH RI ∈ C and β = uIT −ĥHRI ∈
CN ×1 . The equation in (29) is composed of N linear equations
Arrow
VForest = ∪G
g=1 {cg , ng,1 , ng,2 , . . . , ng,NG −1 } , (25) with 2N − 1 unknowns. The linear equation coefficients α
Arrow
EForest = ∪G {(cg , ng,` ) | ` = 1, . . . , NG − 1} , (26) and β are complex vectors, while the 2N − 1 unknowns are
g=1
real since they are the entries of the susceptance matrix B of
where cg is the index of the central vertex for the gth group, the tree-connected RIS, which are not constrained to be zero
similarly to (19) and (20). Accordingly, the susceptance matrix as shown in (18) and (21). We aim to solve (29) as it is a
of forest-connected RIS with arrowhead architecture for each sufficient and necessary condition to achieve the performance
group is block diagonal with each block being symmetric and upper bound P̄R . To that end, we rewrite (29) as 2N linear
arrowhead. equations with real coefficients and 2N − 1 real unknowns,
The forest-connected RIS achieves a good performance- namely  
complexity trade-off between the single-connected and the B
a = b, (30)
tree-connected RIS. Note that the single-connected and tree- B
connected RIS can be viewed as two special cases of the
where a ∈ R2N ×1 and b ∈ R2N ×1 are defined as
forest-connected RIS, with NG = 1 and NG = N , respec-    
tively. In addition, forest-connected RIS is equivalent to group- < {α} < {β}
a= , b= . (31)
connected RIS when NG = 2 [4]. The number of tunable = {α} = {β}

2 Note
To solve (30), we further rewrite the equation such that the
that the tree-connected RIS is a special case of the forest-connected
RIS, i.e., with G = 1. This is consistent with the fact that in graph theory a 2N − 1 real unknowns are explicitly collected in a vector x ∈
tree is a special case of forest. R(2N −1)×1 . Specifically, given a tree-connected RIS whose
 7

N −1 Algorithm 1 Closed-form global optimization of tree-

graph has an edge set E = {(n` , m` )}`=1 such as (17) or
(20), the 2N − 1 real unknowns are collected as connected RIS for MISO systems
h iT Input: hRI ∈ C1×N , HIT ∈ CN ×M .
x = [B]1,1 , . . . , [B]N,N , [B]n1 ,m1 , . . . , [B]nN −1 ,mN −1 , Output: B.
(32) h
1: Obtain ĥRI = khRI k , uIT = umax (HIT ).
where [B]n,n for n = 1, . . . , N denote the N diagonal entries RI 
of B and [B]n` ,m` for ` = 1, . . . , N − 1 denote the N − 1 2: Obtain α = jZ0 uIT + ĥH H
RI , β = uIT − ĥRI .
off-diagonal entries of B specified by the edge set of graph. 3: Obtain A1 and A2 by (35) and (36), respectively.
Accordingly, (30) can be equivalently rewritten as 4: Obtain A and b by (34) and (31), respectively.

−1 T
Ax = b, (33) 5: Obtain x = AT A A b.
6: Obtain B by (32).
where A ∈ R2N ×(2N −1) is the coefficient matrix given by
  7: Return B.
< {A1 } < {A2 }
A= , (34)
= {A1 } = {A2 }
where A1 ∈ CN ×N and A2 ∈ CN ×(N −1) . Because the real A, and subsequently the optimal susceptance matrix B of the
and imaginary parts of A1 are multiplied by the N diagonal tree-connected RIS can be found through (32).
entries of B in (33), to build the equivalence between (30) To conclude, we prove that for any tree-connected RIS there
and (33), A1 should be given by is always one and only one global optimal solution for the
susceptance matrix B to achieve the upper bound P̄R for any
A1 = diag (α) . (35) channel realization. In Alg. 1, we summarize the steps required
to optimize tree-connected RIS for MISO systems.
On the other hand, the real and imaginary parts of A2 are
multiplied by the N − 1 off-diagonal entries of B including
[B]n1 ,m1 , . . . , [B]nN −1 ,mN −1 . For this reason, the structure of B. Forest-Connected RIS Optimization
A2 depends on how the RIS ports are connected to each other, The forest-connected RIS optimization to maximize the
that is the specific graph associated with the tree-connected received signal power in MISO systems is formulated as
RIS. Accordingly, the entry of the `th column of A2 for ` =
1, . . . , N − 1 is given by
2
 max PT |hRI ΘHIT w| (37)
[α]n` if k = m`
 Θ,w
−1
[A2 ] k,` = [α]m` if k = n` . (36) s.t. Θ = (I + jZ0 B) (I − jZ0 B) , (38)

0 otherwise B = diag (B1 , . . . , BG ) , (39)


The following proposition provides a useful property of Bg = BTg , ∀g, (40)

tree-connected RIS, which we can leverage to solve the linear Bg ∈ BG,g , ∀g, (41)
equations (33).
where the susceptance matrix B of the forest-connected RIS
Proposition 2. For any tree-connected RIS with N elements, is a block diagonal matrix with Bg being the gth block and
the coefficient matrix A ∈ R2N ×(2N −1) has full column rank, BG,g represents the set of the gth block written as
i.e., r (A) = 2N − 1.
BG,g = {Bg | [Bg ] n,m = 0, n 6= m, n, m ∈ Vg , (n, m) ∈ / Eg } ,
Proof. Please refer to Appendix C. (42)
where Vg and Eg are the vertex and edge sets of the graph for
According to Proposition 2, the linear equations (33) has the gth group for g = 1, . . . , G. Specifically, Vg and Eg can
one solution or no solution [38]. Furthermore, we show that it be found by (23) and (24) when the tridiagonal architecture is
has exactly one solution by using the following proposition. used for the gth group, or by (25) and (26) when the arrowhead
architecture is used.
Proposition 3. For any tree-connected RIS with N elements,
Different from optimizing the tree-connected RIS, it is hard
the augmented matrix [A|b] ∈ R2N ×2N has rank r ([A|b]) =
to find a tight upper bound on the received signal power in
2N − 1.
forest-connected RIS-aided MISO systems due to block diag-
Proof. Please refer to Appendix D. onal characteristics of the susceptance matrix, which makes it
hard to derive a closed-form global optimal solution. To solve
According to the Rouche-Capelli theorem [38], a system the challenging forest-connected RIS optimization problem,
of linear equations is consistent if the rank of the coefficient we propose an iterative optimization approach, where the
matrix A is equal to the rank of the augmented matrix precoder w and the RIS susceptance matrix B are alternatively
[A|b]. Thus, according to Proposition 3, the system of linear optimized.
equations (33) is consistent and has exactly one solution. As 1) Optimizing B with Fixed w: When the precoder w is
†

−1findT the solution of (33) as x = A b, where
a result, we can fixed, we introduce heffIT = HIT w as the effective channel
† T
A = A A A is the Moore-Penrose pseudo-inverse of between the transmitter and RIS so that the forest-connected
 8

Algorithm 2 Optimization of forest-connected RIS for MISO

systems
Input: hRI ∈ C1×N , HIT ∈ CN ×M .
Output: B and w.
1: Initialize w.
2: while no convergence of objective (37) do
3: Update Bg by Alg. 1 with input hRI,g , heffIT,g , ∀g.
4: Obtain Θ from B = diag (B1 , . . . , BG ) by (3).
H
5: Update w = (hRI ΘHIT ) /khRI ΘHIT k. Fig. 5. Two-dimensional coordinate system for the RIS-aided MISO system.
6: end while
7: Return B and w. computed by (3) as a function of B.
The precoder w and the RIS susceptance matrix B are
alternatively optimized until convergence of the objective (37).
We summarize the steps to optimize forest-connected RIS in
RIS-aided MISO system is equivalently converted to a forest-
MISO systems in Alg. 2. The iterative algorithm is initialized
connected RIS-aided SISO system. Thus, we can equivalently
by randomly setting w to a feasible value. In addition, the
rewrite the problem (37)-(41) as
convergence of Alg. 2 is guaranteed by the following two facts.
2
G
X First, at each iteration, the objective given by the received
max PT hRI,g Θg heff
IT,g (43) signal power PR is non-decreasing. Second, the objective
Θg 2 2
g=1 function is upper bounded by PT khRI k kHIT k because of
−1
s.t. Θg = (I + jZ0 Bg ) (I − jZ0 Bg ) , (44) the sub-multiplicativity of the spectral norm.
Bg = BTg , ∀g, (45) V. P ERFORMANCE E VALUATION
Bg ∈ BG,g , ∀g, (46) In this section, we provide the performance evaluation for
where hRI,g ∈ C1×NG and heff NG ×1 the proposed tree- and forest-connected RIS. We consider a
IT,g ∈ C consist of the NG
eff
entries of hRI and hIT corresponding to the RIS elements of two-dimensional coordinate system, as shown in Fig. 5. The
the gth group, respectively [4]. To solve the problem (43)- transmitter and receiver are located at (0, 0) and (52, 0) in
(46), we first consider the case that the graph for each group meters (m), respectively. The RIS is located at (50, 2) and
is complete, that is the group-connected RIS where Bg can it is equipped with N elements. For the large-scale path
be arbitrary symmetric matrix. As shown in [4] and [25], the loss, we use the distance-dependent path loss modeled as
received signal power in the group-connected RIS aided SISO Lij (dij ) = L0 (dij /D0 )−αij , where L0 is the reference path
PG 2 loss at distance D0 = 1 m, dij is the distance, and αij is
eff
system, i.e. PT g=1 hRI,g Θg hIT,g , is upper bounded by the path loss exponent for ij ∈ {RI, IT }. We set L0 = −30
G
!2 dB, αRI = 2.8, αIT = 2, and PT = 10 mW. For the small-
scale fading, we assume that the channel from the RIS to the
X
P̄Reff = PT khRI,g k heff
IT,g . (47)
g=1
receiver is Rayleigh fading. The channel from the transmitter
to the RIS is modeled with Rician fading, given by
The key to achieve such upper bound (47) is to find complex r r !
symmetric unitary matrices Θg for g = 1, . . . , G satisfying p K LoS 1 NLoS
HIT = LIT H + H , (49)
1 + K IT 1 + K IT
ĥH eff
RI,g = Θg ĥIT,g , ∀g, (48)
where K refers to the Rician factor, while HLoS and
where ĥRI,g = hRI,g /khRI,g k and ĥeff eff eff
IT,g = hIT,g /khIT,g k.
IT
vec(HNLoS
IT ) ∼ CN (0, I) represent the small-scale line-of-
Recall that each group in the forest-connected RIS utilizes
sight (LoS) and non-line-of-sight (NLoS) (Rayleigh fading)
the tree-connected architecture. Thus, referring to the tree-
components, respectively. We consider two scenarios, with
connected RIS optimization shown in Section IV.A, it is
Rician factor K = 0 dB and K = 10 dB.
straightforward to show that for each group of the forest-
connected RIS there is always one and only one solution for
A. Tree-Connected RIS-Aided MISO Systems
Bg whose corresponding Θg satisfying (48), so that the upper
bound (47) can be achieved. Moreover, the optimal Bg for We first evaluate the performance of the tree-connected
each group can be found by Alg. 1 with the input of the RIS to verify that it is MISO optimal. Specifically, the tree-
channels hRI,g and heff connected RIS is optimized through Alg. 1 and compared
IT,g . Therefore, the susceptance matrix
B of the forest-connected RIS can be globally optimized to with the performance upper bound (15) achieved by the fully-
achieve the upper bound (47) when w is fixed. connected RIS. We recall that in the fully-connected RIS, all
the RIS ports are connected with each other, so that B is an
2) Optimizing w with Fixed B: When the RIS susceptance
arbitrary symmetric matrix and Θ satisfies that
matrix B is fixed, it is obvious that the optimal w is the MRT
given by w = (hRI ΘHIT )H / khRI ΘHIT k, where Θ can be ΘH Θ = I, Θ = ΘT , (50)
 9

Fig. 6. Received signal power in MISO systems aided by fully- and tree-connected RIS.

Fig. 7. Received signal power in MISO systems aided by tree-, group-, forest-, and single-connected RIS, with M = 2.

as shown in [4]. In Fig. 6, we provide the received signal as shown in [4]. In Figs. 7 and 8, we provide the received
power achieved in MISO systems aided by fully- and tree- signal power in MISO systems aided by tree-, group-, forest-
connected RIS. We can make the following observations. First, , and single-connected RIS, with M = 2 and M = 8,
as expected, the tree-connected RIS can always achieve the respectively. Group-connected and single-connected RIS are
performance upper bound. However, it has much lower circuit optimized as proposed in [25, Alg. 2]. We can make the
complexity than the fully-connected RIS, which will be quan- following observations.
titatively shown in Section V-C. Second, higher received signal First, the forest-connected RIS can always achieve the
power can be obtained by increasing the number of transmit same received signal power as the group-connected RIS with
antennas M for both fully- and tree-connected RIS. Last, the same group size. However, it has much lower circuit
Rician fading channels with a lower Rician factor offer richer complexity than the group-connected RIS with the same group
scattering, allowing to reach a slightly higher performance. size, as quantitatively shown in Section V-C.
Second, the forest-connected RIS achieves a higher received
B. Forest-Connected RIS-Aided MISO Systems
signal power than the single-connected RIS. For example,
We next evaluate the performance of the forest-connected forest-connected RIS with group size 8 brings an improvement
RIS. Specifically, the forest-connected RIS is optimized in the received signal power of 44.6%, when M = 2, N = 64,
through Alg. 2 and it is compared with tree-, group-, and and K = 0 dB. On the other hand, it achieves a lower
single-connected RIS. For the group-connected RIS, we recall received signal power than the tree-connected RIS due to the
that the N element are divided into G groups and all the RIS simplified circuit complexity. Therefore, it is shown that the
ports within the same group are connected with each other, forest-connected RIS achieves a good performance-complexity
so that B is arbitrary symmetric block diagonal matrix and Θ trade-off between the single-connected and the tree-connected
satisfies RIS.
Θ = diag (Θ1 , Θ2 , . . . , ΘG ) , (51)
Third, the received signal power increases with the group
Θg = ΘTg , ΘH
g Θg = I, ∀g, (52) size in the forest-connected RIS, which is because increasing
 10

Fig. 8. Received signal power in MISO systems aided by tree-, group-, forest-, and single-connected RIS, with M = 8.

the group size can provide more flexibility for RIS optimiza- To conclude, we demonstrate that the benefit of the
tion. tree-connected (resp. forest-connected) RIS over the fully-
Fourth, higher received signal power can be obtained by connected (resp. group-connected) RIS lies in their highly sim-
increasing the number of transmit antennas M for all the BD- plified circuit complexity while maintaining optimal perfor-
RIS architectures. mance. Therefore, the proposed tree- and forest-connected ar-
Last, BD-RIS is particularly beneficial over single- chitectures significantly improve the performance-complexity
connected RIS in the presence of fading channels with lower trade-off over existing BD-RIS architectures.
Rician factors, in agreement with [4].
VI. C ONCLUSION
C. Circuit Topology Complexity We propose novel modeling, architecture design, and op-
We finally evaluate the circuit topology complexity of the timization for BD-RIS by utilizing graph theory. In partic-
proposed tree- and forest-connected RIS. As analyzed in ular, we model BD-RIS architectures as graphs, capturing
Section III, the circuit topology complexity in terms of the the presence of interconnections between the RIS elements.
number of the tunable admittance components of the tree- Through this modeling, we prove that a BD-RIS achieves the
and forest-connected RIS is 2N − 1 and N (2 − 1/NG ), performance upper bound in MISO systems if and only if its
respectively. For comparison, we also consider the circuit associated graph is connected. This remarkable result allows
topology complexity of fully-, group-, and single-connected us to characterize the least complex BD-RIS architectures
RIS, given by N (N +1)/2, N (NG +1)/2, and N , respectively able to achieve the performance upper bound in MISO sys-
[4]. In Fig. 9, we provide the number of tunable admittance tems, denoted as tree-connected RIS. We also propose forest-
components in the fully-, group, tree-, forest-, and single- connected RIS to bridge between the single-connected and the
connected RIS. We can make the following observations. tree-connected architectures. To optimize these novel BD-RIS
First, compared with the fully-connected RIS, the tree- architectures, we derive a closed-form global optimal solution
connected RIS has much lower circuit topology complexity. for tree-connected RIS, and an iterative algorithm for forest-
For example, the number of tunable admittance components connected RIS. Numerical results confirm that tree-connected
is decreased by 16.4 times when N = 64 in the tree-connected (resp. forest-connected) RIS achieve the same performance as
RIS compared to the fully-connected RIS, but they achieve the fully-connected (resp. group-connected) RIS, with a signifi-
same performance. cantly reduced circuit complexity by up to 16.4 times.
Second, compared with the group-connected RIS with same The proposed graph theoretical modeling of BD-RIS is
group size, the forest-connected RIS has much lower circuit expected to promote significant advancements in the growth
topology complexity. For example, the number of tunable of BD-RIS. This modeling can be used to explore the vast
admittance components is reduced by 2.4 times when N = 64 design space of possible architectures and evaluate them in
in the forest-connected RIS with group size 8 compared to the terms of achievable performance and circuit complexity. This
group-connected RIS with group size 8, but they achieve the will enable the systematic development of new BD-RIS archi-
same performance. tectures with a more favorable balance between performance
Third, increasing the group size in the forest-connected RIS and complexity.
will increase the number of tunable admittance components
A PPENDIX
while also enhancing the performance.
Fourth, compared with the single-connected RIS, the tree- A. Proof of Lemma 1
and forest-connected RIS introduce appropriate complexity but We prove the necessary condition by showing that if G
achieves an improvement in the received power. is not a connected graph, the corresponding BD-RIS is not
 11

for any RIS whose associated graph is a tree to achieve the

performance upper bound (i.e., to be MISO optimal) as shown
in Section IV. A.

B. Proof of Proposition 1
Applying Lemma 1, we need to prove that a connected
graph with N vertices has at least N −1 edges. This is achieved
by induction. The base case is easily verified: a connected
graph with a single vertex has at least zero edges. As the
induction step, we consider a connected graph with N vertices.
From this graph, we remove edges (at least one) until we
obtain a disconnected graph with two connected components.
Assuming the two components have K and N − K vertices,
they have at least K − 1 and N − K − 1 edges by the
Fig. 9. Circuit topology complexity given by the number of tunable admit-
tance components in fully-, group-, tree-, forest-, and single-connected RIS. induction hypothesis, respectively. Since we removed at least
one edge to disconnect the graph, it had originally at least
(K − 1) + (N − K − 1) + 1 = N − 1 edges.
MISO optimal. If G is disconnected, it has C ≥ 2 con-
nected components, where the cth PCcomponent includes N (c) C. Proof of Proposition 2
(c)
ports, for c = 1, . . . , C, with c=1 N = N . With no
This proof is conducted by induction. As the base case,
loss of generality, we assume that each component includes
we consider the only tree-connected RIS that can be realized
adjacent ports. Thus, the admittance matrix Y given by (6)
with N = 2 ports. This RIS includes two tunable admittance
is block diagonal, with the cth block having dimensions
components connecting the two ports to ground and a further
N (c) × N (c) , for c = 1, . . . , C. As a consequence of (2),
tunable admittance connecting the two ports to each other.
also the scattering matrix Θ is block diagonal, and writes
(c) (c) Based on the susceptance matrix B ∈ R2×2 , the left-
as Θ = diag (Θ1 , . . . , ΘC ), where Θc ∈ CN ×N , for
hand side of the system (33) is built by setting x =
c = 1, . . . , C. Thus, the received signal power (10) in the
[[B]1,1 , [B]2,2 , [B]1,2 ]T . Accordingly, A ∈ R4×3 is given by
case of a disconnected graph G is
(34) with A1 ∈ C2×2 as in (35) and A2 ∈ C2×1 given by
2
C
X A2 = [[α]2 , [α]1 ]T . Thus, it is easy to recognize that A has
PRDisC = PT hRI,c Θc HIT,c w , (53) in general full column rank, i.e., r (A) = 3. The proposition
c=1 is hence verified for the case N = 2.
(c) (c) As the induction step, we prove that if the proposition is
where hRI,c ∈ C1×N and HIT,c ∈ CN ×M contain the
valid for RISs with N − 1 ports, it also holds for RISs with N
N (c) elements of hRI and rows of HIT corresponding to the
ports. We consider a tree-connected RIS with N − 1 elements,
N (c) RIS elements grouped into the cth component, respec-
whose coefficient matrix is given by
tively. The received signal power PRDisC is upper bounded by  n o n o 
(N −1) (N −1)
C
!2 < A1 < A2
X A(N −1) =  n (N −1) o o . (56)
P̄RDisC = PT
n
khRI,c k kHIT,c wk , (54) = A1 = A2
(N −1)
c=1

following the Cauchy-Schwarz inequality and that ΘH To this (N − 1)-port tree-connected RIS, we connect an
c Θc = I,
for c = 1, . . . , C. Furthermore, it holds that additional port creating an N -port tree-connected RIS. With
no loss of generality, we assume that the additional port is
(a) (b)
2
P̄RDisC ≤ PT khRI k kHIT wk ≤ P̄R ,
2
(55) connected to the (N − 1)th port through a tunable admittance.
In this way, the resulting N -port BD-RIS has the coefficient
where (a) follows by the Cauchy-Schwarz inequality and (b) matrix
follows by the definition of the spectral norm. Note that (b)  n
(N )
o n
(N )
o 
is an equality if and only if w is the dominant right singular < A1 < A2
A(N ) =  n (N ) o n
(N )
o , (57)
vector of HIT . However, (a) is in general a strict inequality = A1 = A2
if w is the dominant right singular vector of HIT , which thus
(N ) (N )
proves the necessary condition of the lemma. where A1 ∈ CN ×N and A2 ∈ CN ×(N −1) are given by
On the other hand, we prove the sufficient condition by  (N −1)

(N ) A1 0(N −1)×1
showing that if G is a connected graph, the corresponding BD- A1 = , (58)
RIS is MISO optimal. This is straightforward to prove because 01×(N −1) [α]N
1) for a connected graph, we can always remove some edges
 
(N −1) 0(N −2)×1
to make it a tree, i.e. setting the corresponding off-diagonal (N ) A 2
A2 =  [α]N , (59)
entries of B zero, 2) there exist one and only one solution 01×(N −2) [α]N −1
 12

with 0R×C denoting an R × C all-zero matrix. To prove the for n = 1, . . . , N − 1. The resulting matrix is given by
induction step, we need to show that r(A(N −1) ) = 2(N −1)−  
< {A1 } < {A2 } < {β}
1 implies r(A(N ) ) = 2N − 1. To this end, A(N ) is rewritten [A|b] ∼ , (67)
0N ×N A00 b00
as

0(N −2)×1
 where we introduced A00 ∈ RN ×N −1 and b00 ∈ RN ×1 .
< {[α]N }  We now show that (67) has rank 2N − 1 since its 2N th row
 A(N −1)

0(2N −2)×1 
is all-zero. Indeed, we have
(N )
 0 (N −2)×1

A ∼
  , (60)
 = {[α] N } 
 [A00 ]N,` = < [α]n` [A0 ]n` ,` + < [α]m` [A0 ]m` ,` = 0
 
 < {[α]N } < [α]N −1

(68)
02×(2N −3) 
= {[α]N } = [α]N −1 for ` = 1, . . . , N − 1, with n` and m` being the row indexes
by applying appropriate row and column swapping op- of the two non-zero elements in the `th column of A2 .
erations. From (60), we notice that if A(N −1) has full Furthermore, the N th entry of vector b00 is given by
column
 rank and
 the vectors
 [< {[α]N } , = {[α]N }] and N
X
< [α]N −1 , = [α]N −1 are linearly independent, then [b00 ]N = < {[α]n } [b0 ]n (69)
(N )
A  has full column
 rank.Since [< {[α]N } , = {[α]N }] and n=1
< [α]N −1 , = [α]N −1 are in practice linearly indepen- XN

dent with probability 1, the induction step is proven and the = < {[α]n } = {[β]n } − = {[α]n } < {[β]n } , (70)
proof is concluded. n=1

where
< {[α]n } = {[β]n } − = {[α]n } < {[β]n } (71)
D. Proof of Proposition 3  nh i o2
2
=jZ0 = ĥH RI − = {[uIT ]n } (72)
To prove that r ([A|b]) = 2N − 1, we show that it is n
i o2 
always possible to obtain a zero row in [A|b] by applying
nh
H 2
+< ĥRI − < {[uIT ]n } (73)
appropriate row operations. Firstly, we execute on [A|b] the n
N row operations
h i 2 
2
=jZ0 ĥH
RI − |[u ]
IT n | . (74)
= {[α]n } n
rN +n = rN +n − rn , (61)
< {[α]n } Thus, recalling that kĥH 2 2
RI k = 1 and kuIT k = 1, we have
for n = 1, . . . , N , where rn denotes the nth row of [A|b]. 
2

00 H 2
The resulting matrix is given by [b ]N = jZ0 ĥRI − kuIT k = 0. (75)
 
< {A1 } < {A2 } < {β}
[A|b] ∼ , (62)
0N ×N A0 b0 R EFERENCES
[1] E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M.-S. Alouini,
where we introduced A0 ∈ RN ×(N −1) and b0 ∈ RN ×1 . and R. Zhang, “Wireless Communications Through Reconfigurable
The matrix A0 has exactly two non-zero elements in each Intelligent Surfaces,” IEEE Access, vol. 7, pp. 116 753–116 773, 2019.
column, in the same positions as the non-zero elements in [2] Q. Wu and R. Zhang, “Towards Smart and Reconfigurable Environment:
Intelligent Reflecting Surface Aided Wireless Network,” IEEE Commun.
A2 . Specifically, Mag., vol. 58, no. 1, pp. 106–112, 2020.
 n o [3] Q. Wu, S. Zhang, B. Zheng, C. You, and R. Zhang, “Intelligent
= [α]n
[A0 ]
  Reflecting Surface-Aided Wireless Communications: A Tutorial,” IEEE
= = [α] − `o
< [α]m` ,
 n


 n ` ,` m` < [α]n Trans. Commun., vol. 69, no. 5, pp. 3313–3351, 2021.
`o
[4] S. Shen, B. Clerckx, and R. Murch, “Modeling and Architecture De-
 n
= [α]m (63)
0 sign of Reconfigurable Intelligent Surfaces Using Scattering Parameter
 
[A ]m` ,` = = [α]n` − n ` o < [α]
n` , Network Analysis,” IEEE Trans. Wireless Commun., pp. 1–1, 2021.


 < [α]m
`
[5] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless

 0

[A ]p,` = 0, ∀p 6= m` , n` , network via joint active and passive beamforming,” IEEE Trans. Wireless
Commun., vol. 18, no. 11, pp. 5394–5409, 2019.
for ` = 1, . . . , N − 1, with n` and m` being the row indexes [6] H. Guo, Y.-C. Liang, J. Chen, and E. G. Larsson, “Weighted sum-
of the two non-zero elements in the `th column of A2 . The rate maximization for reconfigurable intelligent surface aided wireless
vector b0 writes as networks,” IEEE Trans. Wireless Commun., vol. 19, no. 5, pp. 3064–
3076, 2020.
= {[α]n } [7] Y. Liu, J. Zhao, M. Li, and Q. Wu, “Intelligent reflecting surface
[b0 ]n = = {[β]n } − < {[β]n } , (64) aided MISO uplink communication network: Feasibility and power
< {[α]n } minimization for perfect and imperfect CSI,” IEEE Trans. Commun.,
for n = 1, . . . , N . Secondly, we execute on the resulting [A|b] vol. 69, no. 3, pp. 1975–1989, 2021.
[8] C. Pan, H. Ren, K. Wang, W. Xu, M. Elkashlan, A. Nallanathan,
the N row operations and L. Hanzo, “Multicell MIMO communications relying on intelligent
reflecting surfaces,” IEEE Trans. Wireless Commun., vol. 19, no. 8, pp.
rN +n = < {[α]n } rN +n , (65) 5218–5233, 2020.
[9] H. Li, W. Cai, Y. Liu, M. Li, Q. Liu, and Q. Wu, “Intelligent reflecting
for n = 1, . . . , N , followed by the N − 1 row operations surface enhanced wideband MIMO-OFDM communications: From prac-
tical model to reflection optimization,” IEEE Trans. Commun., vol. 69,
r2N = r2N + rN +n , (66) no. 7, pp. 4807–4820, 2021.
 13

[10] Y. Xiu, J. Zhao, W. Sun, M. D. Renzo, G. Gui, Z. Zhang, and N. Wei, [31] ——, “A dynamic grouping strategy for beyond diagonal reconfigurable
“Reconfigurable intelligent surfaces aided mmwave NOMA: Joint power intelligent surfaces with hybrid transmitting and reflecting mode,” arXiv
allocation, phase shifts, and hybrid beamforming optimization,” IEEE preprint arXiv:2210.02499, 2022.
Trans. Wireless Commun., vol. 20, no. 12, pp. 8393–8409, 2021. [32] Q. Li, M. El-Hajjar, I. Hemadeh, A. Shojaeifard, A. A. M. Mourad,
[11] A. Bansal, K. Singh, B. Clerckx, C.-P. Li, and M.-S. Alouini, “Rate- B. Clerckx, and L. Hanzo, “Reconfigurable intelligent surfaces relying
splitting multiple access for intelligent reflecting surface aided multi- on non-diagonal phase shift matrices,” IEEE Trans. Veh. Technol.,
user communications,” IEEE Trans. Veh. Technol., vol. 70, no. 9, pp. vol. 71, no. 6, pp. 6367–6383, 2022.
9217–9229, 2021. [33] D. M. Pozar, Microwave engineering. John wiley & sons, 2011.
[12] J. Hu, H. Zhang, B. Di, L. Li, K. Bian, L. Song, Y. Li, Z. Han, and H. V. [34] S. Shen and R. D. Murch, “Impedance Matching for Compact Multiple
Poor, “Reconfigurable intelligent surface based RF sensing: Design, Antenna Systems in Random RF Fields,” IEEE Trans. Antennas Propag.,
optimization, and implementation,” IEEE J. Sel. Areas Commun., vol. 38, vol. 64, no. 2, pp. 820–825, 2016.
no. 11, pp. 2700–2716, 2020. [35] J. A. Bondy and U. S. R. Murty, Graph theory with applications.
[13] Z. Feng, B. Clerckx, and Y. Zhao, “Waveform and beamforming design Macmillan London, 1976, vol. 290.
for intelligent reflecting surface aided wireless power transfer: Single- [36] D. O’Leary and G. Stewart, “Computing the eigenvalues and eigenvec-
user and multi-user solutions,” IEEE Trans. Wireless Commun., vol. 21, tors of symmetric arrowhead matrices,” J. Comput. Phys., vol. 90, no. 2,
no. 7, pp. 5346–5361, 2022. pp. 497–505, 1990.
[14] Y. Zhao, B. Clerckx, and Z. Feng, “IRS-aided SWIPT: Joint wave- [37] Z. H. Shaik, E. Björnson, and E. G. Larsson, “MMSE-optimal sequential
form, active and passive beamforming design under nonlinear harvester processing for cell-free massive MIMO with radio stripes,” IEEE Trans.
model,” IEEE Trans. Commun., vol. 70, no. 2, pp. 1345–1359, 2022. Commun., vol. 69, no. 11, pp. 7775–7789, 2021.
[15] Z. Peng, Z. Chen, C. Pan, G. Zhou, and H. Ren, “Robust transmission [38] G. Strang, Introduction to linear algebra. Wellesley-Cambridge Press
design for RIS-aided communications with both transceiver hardware Wellesley, MA, 1993, vol. 3.
impairments and imperfect CSI,” IEEE Wireless Commun. Lett., vol. 11,
no. 3, pp. 528–532, 2022.
[16] Y. Chen, Y. Wang, and L. Jiao, “Robust transmission for reconfigurable
intelligent surface aided millimeter wave vehicular communications with
statistical CSI,” IEEE Trans. Wireless Commun., vol. 21, no. 2, pp. 928–
944, 2022.
[17] B. Zheng, C. You, and R. Zhang, “Double-IRS assisted multi-user
MIMO: Cooperative passive beamforming design,” IEEE Trans. Wireless
Commun., vol. 20, no. 7, pp. 4513–4526, 2021.
[18] W. Mei and R. Zhang, “Intelligent reflecting surface for multi-path
beam routing with active/passive beam splitting and combining,” IEEE
Commun. Lett., vol. 26, no. 5, pp. 1165–1169, 2022.
[19] Q. Wu and R. Zhang, “Beamforming optimization for intelligent reflect-
ing surface with discrete phase shifts,” in ICASSP 2019 - 2019 IEEE
International Conference on Acoustics, Speech and Signal Processing
(ICASSP), 2019, pp. 7830–7833.
[20] B. Di, H. Zhang, L. Song, Y. Li, Z. Han, and H. V. Poor, “Hybrid
beamforming for reconfigurable intelligent surface based multi-user
communications: Achievable rates with limited discrete phase shifts,”
IEEE J. Sel. Areas Commun., vol. 38, no. 8, pp. 1809–1822, 2020.
[21] L. Dai, B. Wang, M. Wang, X. Yang, J. Tan, S. Bi, S. Xu, F. Yang,
Z. Chen, M. D. Renzo, C.-B. Chae, and L. Hanzo, “Reconfigurable
intelligent surface-based wireless communications: Antenna design, pro-
totyping, and experimental results,” IEEE Access, vol. 8, pp. 45 913–
45 923, 2020.
[22] J. Rao, Y. Zhang, S. Tang, Z. Li, S. Shen, C.-Y. Chiu, and R. Murch,
“A novel reconfigurable intelligent surface for wide-angle passive beam-
forming,” IEEE Trans. Microw. Theory Tech., vol. 70, no. 12, pp. 5427–
5439, 2022.
[23] H. Li, S. Shen, M. Nerini, and B. Clerckx, “Reconfigurable intelligent
surfaces 2.0: Beyond diagonal phase shift matrices,” arXiv preprint
arXiv:2301.03288, 2023.
[24] M. Nerini, S. Shen, and B. Clerckx, “Discrete-value group and fully
connected architectures for beyond diagonal reconfigurable intelligent
surfaces,” arXiv preprint arXiv:2110.00077, 2021.
[25] ——, “Optimal group and fully connected design for beyond diagonal
reconfigurable intelligent surfaces,” arXiv preprint arXiv:2211.06117,
2022.
[26] J. Xu, Y. Liu, X. Mu, and O. A. Dobre, “STAR-RISs: Simultaneous
transmitting and reflecting reconfigurable intelligent surfaces,” IEEE
Commun. Lett., vol. 25, no. 9, pp. 3134–3138, 2021.
[27] H. Zhang, S. Zeng, B. Di, Y. Tan, M. Di Renzo, M. Debbah, Z. Han,
H. V. Poor, and L. Song, “Intelligent omni-surfaces for full-dimensional
wireless communications: Principles, technology, and implementation,”
IEEE Commun. Mag., vol. 60, no. 2, pp. 39–45, 2022.
[28] H. Li, S. Shen, and B. Clerckx, “Beyond diagonal reconfigurable intelli-
gent surfaces: From transmitting and reflecting modes to single-, group-
, and fully-connected architectures,” IEEE Trans. Wireless Commun.,
vol. 22, no. 4, pp. 2311–2324, 2023.
[29] ——, “Beyond diagonal reconfigurable intelligent surfaces: A multi-
sector mode enabling highly directional full-space wireless coverage,”
arXiv preprint arXiv:2209.00301, 2022.
[30] ——, “Synergizing beyond diagonal reconfigurable intelligent surface
and rate-splitting multiple access,” arXiv preprint arXiv:2303.06912,
2023.