Automatica ( ) –

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Technical communique

Matrix-weighted consensus and its applications✩

Minh Hoang Trinh a , Chuong Van Nguyen a , Young-Hun Lim b , Hyo-Sung Ahn a, *
a
School of Mechanical Engineering, Gwangju Institute of Science and Technology (GIST), Gwangju, Republic of Korea
b
Department of Electronic Engineering, Gyeongnam National University of Science and Technology, Jinju, Republic of Korea

article info a b s t r a c t
Article history: This paper proposes the matrix-weighted consensus algorithm, which is a generalization of the con-
Received 11 March 2017 sensus algorithm. Given a networked dynamical system where the interconnections between agents
Received in revised form 8 August 2017 are weighted by nonnegative definite matrices, it is shown that consensus and clustering phenomena
Accepted 9 October 2017
naturally exist. We examine algebraic and algebraic graph conditions for achieving a consensus.
Available online xxxx
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction conditions for synchronization. The concept of deviated cyclic pur-

Given a system of n single-integrator agents whose intercon-
nections between agents are described by a weighted undirected
graph G , the consensus algorithm (Olfati-Saber & Murray, 2004) is
defined as:
n matrix-weighted consensus protocol proposed in this paper. In the
∑
ẋi = aij (xj − xi ), ∀i = 1, . . . , n,
j=1
where xi , xj ∈ Rd are the state vectors of agents i and j, and aij is
a positive scalar (or zero) if i and j are connected (or disconnected,
respectively). It is well-known that under the consensus protocol
(1), an average consensus is globally achieved if and only if G is
connected (Olfati-Saber & Murray, 2004). This paper generalizes
the consensus algorithm (1) by using matrix weight Aij instead
of the scalar weight aij to describe the interconnection between
two agents i and j. A matrix weight could be a positive definite
matrix (full connection), a positive semidefinite matrix (partial
connection), or a zero matrix (no direct connection). Thus, the
matrix-weight consensus covers a larger set of problems in multi-
agent systems.
In the literature, matrix-weights have been often used in net-
worked systems to describe the interconnections between agents.
For example, the author of Tuna (2016) used matrix weights to de-
scribe interconnections between coupled oscillators and provided
✩ This work was supported by the National Research Foundation (NRF) of Korea
under the grant NRF-2017R1A2B3007034. The material in this paper was not pre-
sented at any conference. This paper was recommended for publication in revised
form by Associate Editor Carlo Fischione under the direction of Editor André L. Tits.
*
Corresponding author.
E-mail addresses: trinhhoangminh@gist.ac.kr (M.H. Trinh),
chuongnguyen@gist.ac.kr (C.V. Nguyen), yhunlim@gntech.ac.kr (Y.-H. Lim),
hyosung@gist.ac.kr, hyosung@ieee.org (H.-S. Ahn).
https://doi.org/10.1016/j.automatica.2017.12.024
0005-1098/© 2017 Elsevier Ltd. All rights reserved.
suit introduced in Ramirez, Pavone, Frazzoli, and Miller (2009) and
orientation estimation in Lee and Ahn (2016) can be considered
as consensus protocols with rotation matrix weights (Ahn, Trinh,
& Lee, 2017). Also, the bearing-based network localization setup
in Zhao and Zelazo (2016) can be considered as a special case of the
context of social networks, when a group of people are discussing
(1)
multiple topics, matrix weights can be used to describe the logical
inter-dependency of the topics (Friedkin, Proskurnikov, Tempo, &
Parsegov, 2016).
This paper studies the matrix-weighted consensus algorithm with
fixed undirected graphs. First, we define several terminologies
and propose the matrix-weighted consensus algorithm. Second,
a necessary and sufficient condition for globally exponentially
reaching an average consensus based on the nullspace of the
matrix-weighted Laplacian is provided. Third, due to the exis-
tence of semidefinite matrix weights, clustered consensus happens
even when the graph is connected. We examine the algebraic
graph conditions for consensus and clustered consensus. Finally,
an algorithm to determine all clusters in the network is provided.
The algorithm initially partitions the graph into a set of clusters
associated with the positive trees in the graph. If two clusters
satisfy several algebraic conditions on their connections, they will
be merged together at each iteration. Consequently, the algorithm
gradually reduces the number of clusters in the graph, and it ends
when no two clusters can be further merged together. If there is a
cluster containing all vertices in the graph, a consensus is globally
achieved.
2. Preliminaries and problem formulation
A fixed undirected graph with matrix weights is denoted by
G . The graph G is characterized by a triple (V , E , A). Here, V =
{1, . . . , n} denotes the set of |V | = n vertices, E = {eij = (i, j)| i, j ∈

Please cite this article in press as: Trinh, M.H., et al., Matrix-weighted consensus and its applications. Automatica (2018), https://doi.org/10.1016/j.automatica.2017.12.024.
2 M.H. Trinh et al. / Automatica ( ) –

agents update their states under the following matrix-weighted

consensus protocol:
∑
ẋi = Aij (xj − xi ), ∀i = 1, . . . , n. (2)
j∈Ni

The n-agent system’s dynamics can be expressed in the following

Fig. 1. T is a positive spanning tree of G .
matrix form:

ẋ = −Lx, (3)
V , i ̸ = j} denotes the set of |E | = m edges, and A = {Aij ∈
where L is the matrix-weighted Laplacian.
Rd×d | (i, j) ∈ E , Aij = ATij ≥ 0} denotes the set of matrix weights.
The dimension d (d ≥ 1) of the matrix weights in A depends on the
Definition 4 (Consensus). The n-agent system is said to achieve a
problem. If d = 1, G is an usual undirected scalar-weighted graph.
consensus if and only if xi = xj , for all i, j ∈ V , i ̸ = j.
Depending on the matrix weights, the interconnections between
vertices in G are classified into two types. If the weight matrix Aij Define R = Range{1n ⊗ Id } as the consensus space. A consensus
corresponding to an edge (i, j) is positive definite (positive semi- of the n-agent system is globally/locally asymptotically achieved
definite), we say that (i, j) is a positive (semi-positive) definite edge if and only if x globally/locally asymptotically approaches R. Al-
and two vertices i and j are connected via a positive (semipositive) though consensus is an important objective, in some applications,
definite edge. If i and j are disconnected, Aij = 0. We also assume the agents’ states are desired to converge to some different values.
⋂given by C1 , . . . , Cl (1 ≤ ⋃
that the interconnections between any two vertices are symmetric, A partition of V is l ≤ n) satisfying two
i.e., Aij = Aji , ∀(i, j) ∈ E . A path is a sequence of vertices in properties: (i) Ci
l
Cj = ∅, for i ̸ = j, and (ii) k=1 Ck = V .
G , denoted by P = i1 i2 . . . il , such that each edge (ik , ik+1 ), k =
1, . . . , l − 1, is a positive/semipositive definite edge. The graph G Definition 5 (Cluster Consensus). The n-agent system is said to
is called positive semiconnected if and only if there exists a path achieve a cluster consensus if there exists a partition C1 , . . . , Cl ,
between any two vertices in G . Otherwise, G is disconnected. such that all agents belonging to the same partition achieve con-
sensus, while for any two agents i and j belonging to two different
Definition 1 (Positive Path). A positive path is a sequence of ver- partitions, xi ̸ = xj . Each Ci , i = 1, . . . , l, is referred to as a cluster.
tices in G , denoted by P = i1 i2 . . . il , such that each edge (ik , ik+1 ),
k = 1, . . . , l − 1, is a positive definite edge.
3. Algebraic condition for reaching a consensus
A tree is a graph (containing at least one vertex) in which any
two vertices are connected by exactly one path. This section aims to find an algebraic condition of the matrix-
weighted Laplacian for reaching a consensus.
Definition 2 (Positive Tree). A positive tree T is a tree contained
in G having all positive connections. Equivalently, for all i, j ∈ T , Lemma 2. The matrix-weighted Laplacian L is symmetric, posi-
there exists a positive path in T connecting i and j. tive semidefinite, and has the nullspace N (L) = span{R, {v =
[vT1 , . . . , vTn ]T ∈ Rdn | (vj − vi ) ∈ N (Aij ), ∀(i, j) ∈ E }}.
Definition 3 (Positive Spanning Tree). A positive spanning tree T of
G is a positive tree containing all vertices in V . Proof. From Lemma 1, we can write
Note that a tree of k vertices (k ≥ 1) contains exactly k − 1
L = H̄T blkdiag(Ak )H̄ = MT M, (4)
edges. Thus, a positive spanning tree of G contains exactly n − 1
positive edges (see Fig. 1). Next, we define several algebraic struc- 1/2
where M = blkdiag(Ak )H̄.
Eq. (4) shows that L is positive
tures corresponding to the matrix weighted graph G . The matrix- semidefinite, and N (L) = N (MT M) = N (M). As a result, N (L) =
weighted adjacency matrix of G is defined as A = [Aij ]i,j=1,...,n . Since
∑ [v1 , . . . , vn ] ̸∈ R such that
T T T
N (M) ⊇ N (H̄) ⊇ R. Consider v =
G is undirected and Aij = Aji , the matrix A is symmetric. For each Lv = 0. It follows vT Lv = 0, or i.e., (i,j)∈E (vi − vj )T Aij (vi − vj ) = 0,
vertex i, the∑ neighbor set of i is defined as Ni = {j ∈ V | eij ∈ E }.
which implies that (vi − vj ) ∈ N (Aij ), for all (i, j) ∈ E . ■
Let Di = j∈Ni Aij . We define D = blkdiag(Di ) as the degree
matrix of the graph G . The matrix-weighted Laplacian is defined as
Remark 1. According to Lemma 2, dim(N (L)) ≥ dim(R). The
L = D − A (Tuna, 2016).
matrix-weighted Laplacian L has at least d zero eigenvalues. Let
Consider an arbitrary index of the edges of G , we can write the
{λi }i=1,...,dn be L’s eigenspectra, we have 0 = λ1 = · · · = λd ≤
edge set and the matrix-weight set as E = {ekij }k=1,...,m and A =
λd+1 ≤ · · · ≤ λdn .
{Akij }k=1,...,m , correspondingly.1 Let H ∈ Rm×n be the incidence
matrix corresponding to an arbitrary orientation of the edges in E , Lemma 3. Under
and let H̄ = H ⊗ Id , where ⊗ denotes the Kronecker product and Id ∑n the matrix-weighted consensus protocol (2), the
average x̄ = 1n i=1 xi is invariant.
is the d × d identity matrix. We have the following lemma whose
proof can be found in Trinh and Ahn (2017).
Proof. We can write x̄ = 1n (1Tn ⊗ Id )x. Taking the derivative of x̄
along the trajectory of (3) yields
Lemma 1. The matrix-weighted Laplacian can be written in the
following form L = H̄T blkdiag(Ak )H̄. x̄˙ = 1/n(1Tn ⊗ Id )ẋ = −1/n(1Tn ⊗ Id )Lx.
Consider a system consisting of n single integrator agents. Sup- Since L is symmetric, if v ∈ N (L), then vT belongs to the left
pose an agent i in the system has a state vector xi = [xi1 , . . . , xid ]T nullspace of L. Thus, x̄˙ = −1/n(1Tn ⊗ Id )L = 0, or i.e. x̄ is
∈ Rd , where d ≥ 1. Denote x = [xT1 , . . . , xTn ]T ∈ Rdn . The invariant. ■

1 If it is unimportant to specify the end-vertices explicitly, we will dropout the Theorem 1 (Stability). Assume that G is positive semiconnected. Any
subscript ij and write ek and Ak without ambiguity. trajectory of (3) globally asymptotically approaches the set N (L).

Please cite this article in press as: Trinh, M.H., et al., Matrix-weighted consensus and its applications. Automatica (2018), https://doi.org/10.1016/j.automatica.2017.12.024.
 M.H. Trinh et al. / Automatica ( ) – 3

Proof. Consider the potential function V = 12 ∥x∥2 , which is posi- 4. Theory of consensus and clustered consensus phenomena
tive definite, radially unbounded,
∑ and continuously differentiable.
Further, V̇ = −xT Lx = − (i,j)∈E (G ) (xi − xj )T Aij (xi − xj ) ≤ 0. In the previous section, Theorem 2 provides an algebraic con-
It follows ∥x(t)∥ ≤ ∥x(0)∥, or i.e., x(t) is bounded. Further, V̇ is dition for reaching a consensus. However, that condition requires
negative semidefinite and V̇ = 0 if and only if x ∈ N (L). Based finding the nullspace of L. Further, the requirement for achieving
on LaSalle’s invariance principle, any trajectory of (3) globally consensus in Corollary 1 could be conservative. This section aims
asymptotically approaches N (L). ■ to find conditions for consensus/cluster consensus related with the
matrix-weighted graph G .
Lemma 4. If N (L) = R, the system (3) has a unique equilibrium
point x∗ = 1n ⊗ x̄. Lemma 5. If G has a positive spanning tree T , the n-agent system
globally exponentially achieves an average consensus under the con-
Proof. We prove that 1n ⊗ x̄ is the unique equilibrium of (3) if sensus protocol (2).
N (L) = R by contradiction. Let {ei }i=1,...,d be a basis of Rd , where
ei = [0, . . . , 1, . . . , 0]T is a vector with all zero entries except for
′ Proof. We can label the edges of G such that the n − 1 edges
an 1 on the ith row. Suppose ∑dthat′ there exists x ∈′ R such that in T are e1 , e2 , . . . , en−1 and the m − n + 1 edges in E \ E (T )
x′ ̸ = x∗ . We can write x′ = x̄ (1 ⊗ e ) = 1 ⊗ x̄ = (1n ⊗ Id )x̄′ ,
i=1 i n i n
are en , en+1 , . . . , em . The incidence matrix corresponding to this
where x̄ = [x̄1 , . . . , x̄d ] . It follows from Lemma 3 that
′ ′ ′ T
labeling can be written as H = [HTE (T ) , HTE \E (T ) ]T , where HE (T ) ∈
1 1 R(n−1)×n represents n − 1 edges of T and HE \E (T ) ∈ R(m−n+1)×n
x̄ = (1Tn ⊗ Id )x′ = (1Tn ⊗ Id )(1n ⊗ Id )x̄′
n n represents the remaining edges in the graph. Note that the rows
1 1 of HE \E (T ) are linearly dependent on the rows of HE (T ) (Zelazo
= (1Tn 1n ⊗ Id )x̄′ = (n ⊗ Id )x̄′ = x̄′ .
n n & Mesbahi, 2011). Specifically, there exists T ∈ R(m−n+1)×(n−1)
As a result, x′ = 1n ⊗ x̄ = x∗ , which is a contradiction. Thus, such that THE (T ) = HE \E (T ) . Any equilibrium point of (3) satisfies
x∗ = 1n ⊗ x̄ is the unique equilibrium of (3). ■ ẋ = −H̄T blkdiag(Ak )H̄x = 0. Thus, xT H̄T blkdiag(Ak )H̄x = 0,
1/2
which in turn implies ∥blkdiag(Ak )H̄x∥2 = ∥Mx∥2 = 0. Denoting
Theorem 2 (Average Consensus). The system (3) globally exponen- T̄ = T ⊗ Id , this condition is equivalent to
tially converges to the system’s average x∗ = 1n ⊗ x̄ if and only if [ 1/2
]
blkdiag(Ak )nk=
−1
1 H̄E (T ) x
N (L) = R. Mx = 1/2
= 0. (6)
blkdiag(Ak )m
k=n T̄H̄E (T ) x
Proof (Necessity). We prove by contradiction. Assume that (3) 1/2
globally asymptotically converges to x∗ = 1n ⊗ x̄ but N (L) ̸ = R. Observe that blkdiag(Ak )H̄E (T ) x = 0 is equivalent to H̄E (T ) x =
From Lemma 2, there exists x′ ∈ Rdn such that Lx′ = 0 and 0. Since HE (T ) is the incidence matrix corresponding to a tree,
x′ ̸ ∈ R. Thus, x = x′ is also an equilibrium point of (3), and any N (HE (T ) ) = span{1n }, which means N (H̄E (T ) ) = R. Based on
trajectory with x(0) = x′ stays at x′ for all t ≥ 0. Thus, x∗ is not Theorem 2, the equilibrium x∗ = 1n ⊗ x̄ is unique and globally
globally asymptotically stable, which contradicts the assumption. exponentially stable. ■
(Sufficiency): Suppose that N (L) = R. Following the proof of
Theorem 1, any trajectory of (3) converges to R. From Lemma 4, Lemma 6. Suppose there exists a positive tree T ⊂ G of l vertices,
x∗ = 1n ⊗ x̄ ∈ N (L) is the unique equilibrium point of (3). under consensus protocol (2), xi (t) → xj (t), ∀i, j ∈ T , as t → ∞.
Consider the Lyapunov function V = 12 δT δ, where δ = x − 1n ⊗ x̄
[H ]
is the disagreement vector. Then, V is positive definite, radially 1 0
unbounded, continuously differentiable, and V̇ = δT δ̇ = −δT Lx = Proof. By writing x = [ xTT , xTV \V (T ) T ,
] we express H = H2 0 ,
H3 H4
−δT Lδ ≤ 0, where in the third equality, we have used the fact that where [H1 0] ∈ R (l−1)×n
associates with the edges belonging
Lδ = Lx − L(1n ⊗ x̄) = Lx − L(1n ⊗ Id )x̄ = Lx. Moreover, δ ⊥ R to the tree T , [H2 0] associates with the other edges between
since (1n ⊗ Id )T δ = (1n ⊗ Id )T x − (1Tn 1n ⊗ Id )x̄ = nx̄ − nx̄ = 0.
vertices in V (T ) which do not belong to the tree, and [H3 H4 ]
Hence,
associates with the remaining edges in E . Observe that H2 is lin-
V̇ = −δT Lδ ≤ −λd+1 (L)δT δ ≤ −α V ≤ 0, (5) early dependent on H1 and this dependency is characterized by
H2 = TH1 . The equilibria of (3) must satisfy H̄x = 0, which implies
where α = 2λd+1 (L) > 0. Further, V̇ = 0 if and only if δ = 0, that
or x = x∗ = 1n ⊗ x̄. Thus, x∗ is a globally exponentially stable
1/2
equilibrium of (3). ■ blkdiag(Ak )lk−=11 H̄1 xT = 0. (7)

Since T is a positive tree, H̄1 xT = 0. Thus, any equilibrium x∗ of

Remark 2. Eq. (5) shows that λd+1 , the smallest positive eigenvalue
of L, determines the convergence rate of the matrix-weighted
consensus protocol (3). Thus, λd+1 is a performance index of the
network, and this index is analogous to the algebraic connectiv-
ity of G in the scalar-weighted consensus algorithm (Mesbahi &
Egerstedt, 2010; Olfati-Saber, Fax, & Murray, 2007).
Corollary 1. Under the consensus protocol (2), if Aij > 0 for all
(i, j) ∈ E , all agents globally exponentially achieve a consensus.
Proof. Since Ak > 0, ∀k = 1, . . . , m, we have N (L) = N (M) =
1/2
N (diag(Ak )H̄) = N (H̄) = R. Thus, the claim follows immediately
from Theorem 2. ■
(3) must have x∗T ∈ N (H̄1 ) = Range{1l ⊗ Id }, i.e., all states of l
agents belonging to the positive tree T are the same. On the basis
of Theorem 1, all agents in T asymptotically reach a consensus. ■
The following result provides a condition to determine whether
or not two vertices belong to a same cluster.
Theorem 3. Given a positive tree T , let the cluster C (T ) generated
from T be containing:
(i) all vertices in T ,
(ii) any vertex i ̸ ∈ T , which defines the set Si = {Pk =
{v1k . . . v|kPk | }| v1k = i, v|kPk | ∈ T , and ∀j = 1, . . . , |Pk |− 1, vjk ̸∈ T },
satisfying

Please cite this article in press as: Trinh, M.H., et al., Matrix-weighted consensus and its applications. Automatica (2018), https://doi.org/10.1016/j.automatica.2017.12.024.
4 M.H. Trinh et al. / Automatica ( ) –
⋃|Pk |−1
(a) For each path Pk , let N (Pk ) = j=1 N (Av k v k ), it holds
j j+1
( |S | )
⋂i
dim N (P k ) = 0. (8)
k=1

(b) Each Pk ∈ Si has no loop, i.e. vl ̸ = vm , ∀vl , vm ∈ Pk .

Then, under the consensus protocol (2), all agents in the cluster
C (T ) have the same equilibrium state. Furthermore, in algorithmic
perspective, the set Si is finite. Fig. 2. Illustration of the four-agent system in Example 1.

Proof. From Lemma 6, all xj (j ∈ V (T )) converge to a common

value x∗T . Consider a vertex i ̸ ∈ V (T ) satisfying the condition (ii). Let C (Tm ) be the cluster generated from the positive tree Tm .
Let x∗i be the equilibrium state of agent i. Then, according to the If there exists a vertex i ∈ C (Tm ) satisfying the condition (ii) in
definition of N (Pk ), there holds Theorem 3 with a cluster C (Tl ), we can form a new cluster C (Tm ) ∪
C (Tl ) by merging C (Tm ) and C (Tl ). By this procedure, we can extend
x∗i − x∗T ∈ N (Pk ), ∀Pk ∈ Si . (9) the positive trees in the graph. All vertices in the new cluster reach
It follows from (8) that the only solution for (9) is xi − xT = 0, ∗ ∗ a consensus under (2). To check whether two clusters C (Tm ) and
C (Tl ) can be merged or not, it is sufficient to check condition (ii)
or x∗i = x∗T . Thus, the cluster C (T ) reaches a consensus. Suppose
of Theorem 3 for only one vertex i ∈ C (Tm ) with regard to C (Tl ).
P1 and P2 are two paths, and P2 is obtained by adding loops to P1 .
Algorithm 1 provides a method for finding all clusters in the graph
Then, N (P1 ) ⊆ N (P2 ). It follows that N (P1 ) ∩ N (P2 ) = N (P1 ).
by iteratively checking condition (ii) in Theorem 3 and merging
Thus, it is unnecessary to consider loops when checking condition
clusters together. Initially, there are p clusters Cm (m = 1, . . . , p)
(8), which means that |Si | is finite. Finally, consider a vertex i which
containing only vertices of Tm . After each iteration, a cluster may
does not satisfy
⋂|S | both (i) and (ii). Let Si be the set of paths from i to T remain or be merged into another cluster. The number of clusters
with dim( k=i 1 N (Pk )) ≥ 1. Clearly, there exists nontrivial solution at each step i could be reduced by 1 or equal to which of step i − 1.
satisfying x∗i − x∗T ̸ = 0. ■ Thus, Algorithm 1 must terminate after finite iterations and the
clusters CG = {C1 , . . . , Cq } (1 ≤ q ≤ p) satisfying:
output⋂is a set of ⋃
Corollary 2. Consider two positive ∑ trees T1 , T2 connected via the edge (i) Cm
q
Cl = ∅, m=1 Cm = V (G ) and (ii) For 1 ≤ m ̸ = l ≤ q, Cm
set S = {(i, j)|i ∈ T1 , j ∈ T2 }. If (i,j)∈S Aij is positive definite, the and Cl cannot be merged together. We can now state the following
equilibrium states of all agents in T1 and T2 are the same under the theorem.
consensus protocol (2).
Theorem 4. Under the consensus protocol (2), an average consensus
Lemma 7 (Partitioning a Graph into Positive Trees). Given a positive is achieved if and only if G is spanned by a cluster.
semiconnected graph G , consider a set of positive trees {T1 , . . . , Tp }
(1 ≤ p ≤ n) ⋂ satisfying ⋃p Proof. If G is spanned by a cluster, based on Theorem 3 the
(i) V (Tm ) V (Tl ) = ∅, m=1 V (Tm ) = V (G ), equilibrium state of all agents in the graph is the same, i.e. x∗ is
(ii) For each Tk (1 ≤ k ≤ p), i, j ∈ V (Tk ) if and only if there exists the only equilibrium point of (3). Thus, the average consensus is
a positive path from i to j. globally asymptotically achieved based on Theorem 1. However, if
Then, the partition {V (T1 ), . . . , V (Tp )} of G is unique. there is no cluster spanning G , the agents belonging to two clusters
of G may not agree and an average consensus cannot be globally
achieved. ■
Algorithm 1 Finding clusters of a matrix-weighted graph G
Example 1. Consider a 4-agent system in R3 with the inter-
Require: G (V , E , A)
1: i ← 0;
action
[ 0
graph
0
] depicted [in Fig. 2 ]and the matrix
0 1 0 0
] A12 =
[ weights: 1 0 0
2: Find the set of positive trees {T1 , . . . , Tp } in G ; 0 1 0 , A13 = 0 0 0 , A23 = 0 0 0 , A14 =
3: CG (0) ← {Cm = {V (Tm )}, m = 1, . . . , p}; 0 0 1 0 0 0 0 0 1
4: repeat
[1 0 0
]
5: CG (i + 1) ← CG (i); 0 2 0 . There is a positive tree T containing vertices 1 and
0 0 1
6: check ← false;
7: for all Cm ∈ CG (i) do 4. Further, we have N (A12 ) = span{[1, 0, 0]T }, N (A13 ) =
8: for all Cl ∈ CG (i), l ̸ = m do span{[0, 1, 0]T , [0, 0, 1]T }, and N (A23 ) = span{[0, 1, 0]T }. There
9: if ∃ vertex k ∈ Cl satisfies Theorem 3(ii) then are two paths from vertex 2 to vertex 1: P1 = 21, and
⋃N (P1 ) = N (A12 ) = span{[1, 0, 0] }, and
T
10: Vtemp ← V (Tm ) ∪ V (Tl ); P2 = 231, where
N (P2 ) = N⋂ (A13 ) N (A23 ) = span{[0, 1, 0]T , [0, 0, 1]T }. It fol-
11: Etemp ← E (Tm ) ∪ E (Tl ) ∪ S ;
lows N (P1 ) N (P2 ) = {0}, which implies that agent 2 is in
12: Ctemp ← Cm ∪ Cl ;
C (T ) due to Theorem 3(ii). Consider the vertex 3, which has two
13: CG (i + 1) ← (CG (i + 1) \ {Cm , Cl }) ∪ {Ctemp };
14: check ← true;
paths to⋂the cluster C = {1, 2, 4}: P3 = 31Tand P4 = 32. Since
N (P3 ) N (P4 ) = N (A23 ) = span{[0, 1, 0] }, vertex 3 does not
15: break;
16: end if belong to C . From numerical simulation, as depicted in Fig. 3, we
17: end for have x∗1 = x∗2 = x∗4 ̸ = x∗3 , which is as expected from the above
18: if check == true then discussion.
19: break;
20: end if The next result follows from Theorems 2 and 4.
21: end for
22: i ← i + 1; Theorem 5. Given a matrix-weighted graph G , there is a cluster
23: until CG (i) == CG (i − 1) spanning all vertices of G if and only if N (L) = R.

Please cite this article in press as: Trinh, M.H., et al., Matrix-weighted consensus and its applications. Automatica (2018), https://doi.org/10.1016/j.automatica.2017.12.024.
 M.H. Trinh et al. / Automatica (
(a) x-axis.
Fig. 3. The states along 3 axes of four agents under (2).
5. Applications
The matrix-weighted consensus proposed in this paper can be
applied to various networked systems. For examples, it can be
utilized to formulate distributed formation control in multi-agent
systems. Also, as mentioned in introduction, it is natural to model
the logical inter-dependency of opinions of social agents or states
of biological agents by matrix-weighted edges. It is noticeable
that uncertainties of inter-dependency between agents or errors
in sensing/control variables in distributed agents can be handled
by matrix-weighted edge sets. Due to the space limitation, we
omit the detailed applications. For more technical details, it is
recommended to refer to Trinh and Ahn (2017) and Trinh, Ye, Ahn,
and Anderson (2017).
6. Conclusion
In this paper, the matrix-weighted consensus algorithm was
proposed. Due to the existence of semipositive definite connec-
tions, cluster phenomenon may easily happen in the network.
A global average consensus can be achieved if and only if the
nullspace of the matrix-weighted Laplacian is spanned by the
consensus space R, or equivalently, the graph is spanned by a
cluster. Also, an algorithm for finding all clusters in a given matrix-
weighted graph was provided. For future works, we would like to
investigate a consensus problem of directed or switching/random
networks with matrix weights.
Acknowledgments
Due to the space limitation, we refer readers to the techni-
cal report (Trinh & Ahn, 2017) for several omitted proofs and
Please cite this article in press as: Trinh, M.H., et al., Matrix-weighted consensus and its applications. Automatica (2018), https://doi.org/10.1016/j.automatica.2017.12.024.
) – 5
(b) y-axis. (c) z-axis.
simulations. Also, we would like to thank Mengbin Ye for his
discussion on social network.
References
Ahn, H.-S., Trinh, M. H., & Lee, B.-H. (2017). Consensus under misaligned orien-
tations. In Proc. of the 17th International conference on control, automation and
systems (pp. 283–288).
Friedkin, N. E., Proskurnikov, A. V., Tempo, R., & Parsegov, S. E. (2016). Network
science on belief system dynamics under logic constraints. Science, 354(6310),
321–326.
Lee, B.-H., & Ahn, H.-S. (2016). Distributed formation control via global orientation
estimation. Automatica, 73, 125–129.
Mesbahi, M., & Egerstedt, M. (2010). Graph theoretic methods in multiagent networks.
Princeton NJ: Princeton University Press.
Olfati-Saber, R., Fax, J. A., & Murray, R. M. (2007). Consensus and cooperation in
networked multi-agent systems. Proceedings of the IEEE, 95(1), 215–233.
Olfati-Saber, R., & Murray, R. M. (2004). Consensus problems in networks of agents
with switching topology and time-delays. IEEE Transactions on Automatic Con-
trol, 49(9), 1520–1533.
Ramirez, J. L., Pavone, M., Frazzoli, E., & Miller, D. W. (2009). Distributed control
of spacecraft formation via cyclic pursuit: Theory and experiments. In Prof. of
American control conference (pp. 4811–4817).
Trinh, M. H., & Ahn, H.-S. (2017). Theory and applications of matrix-weighted
consensus. https://arxiv.org/abs/1703.00129.
Trinh, M. H., Ye, M., Ahn, H.-S., & Anderson, B. D. O. (2017). Matrix-weighted con-
sensus with leader-following topologies. In Proc. of the Asian control conference
(pp. 1795–1800).
Tuna, S. E. (2016). Synchronization under matrix-weighted Laplacian. Automatica,
73, 76–81.
Zelazo, D., & Mesbahi, M. (2011). Edge agreement: Graph-theoretic performance
bounds and passivity analysis. IEEE Transactions on Automatic Control, 56(3),
554–555.
Zhao, S., & Zelazo, D. (2016). Localizability and distributed protocols for bearing-
based network localization in arbitrary dimensions. Automatica, 69,
334–341.