2021 61st IEEE Conference on Decision and Control (CDC)

December 6-9, 2022. Cancún, Mexico

Matrix-Scaled Consensus
Minh Hoang Trinh∗, Dung Van Vu†, Quoc Van Tran‡, Hyo-Sung Ahn§,
Abstract—This paper proposes matrix-scaled (HUST), Hanoi 11615, Vietnam. E-mail:
consensus algorithm, which generalizes the scaled [email protected], [email protected]
§
consensus algorithm in [1]. In (scalar) scaled School of Mechanical Engineering, Gwangju Institute of
consensus algorithms, the agents’ states do not Science and Technology (GIST), Gwangju 61005, Republic of
converge to a common value, but to different points Korea. Email: [email protected]
along a straight line in the state space, which
depends on the scaling factors and the initial states 978-1-6654-6761-2/22/$31.00 ©2022 IEEE
of the agents. In the matrix-scaled consensus The author in [1] proposed a scaled consensus
algorithm, a positive/negative definite matrix weight
is assigned to each agent. Each agent updates its model, in which each agent has a scaling gain si
state based on the product of the sum of relative
matrix scaled states and the sign of the matrix and updates its state variable xi based on the
weight. Under the proposed algorithm, each agent X
asymptotically converges to a final point differing consensus law1x˙i = sign(si) (sjxj − sixi), i = 1,...,n.
with a common consensus point by the inverse of its
own scaling matrix. Thus, the final states of the (1)
agents are not restricted to a straight line but are j∈Ni
extended to an open subspace of the statespace.
Convergence analysis of matrix-scaled consensus for The system (1) achieves a scaled-consensus
single and double-integrator agents are studied in globally asymptotically, that is, sixi(t) → sjxj(t), as
detail. Simulation results are given to support the
analysis. t → ∞ and agents with the same si will converge
to the same point (or cluster). The system (1)
I. Introduction can describe a cooperative network, where
agents have different levels of consensus on a
Consensus algorithm and its variations [1]–[6] single topic. Further extensions of the scaled
have been the main model for studying consensus algorithm with consideration to
networked systems. Though simple, consensus switching graphs, time delays, disturbance
algorithms can describe intricate phenomena attenuation, or different agents’ models can be
such as bird flocking, synchronization behaviors, found in the literature, for examples, see
or how a group of people eventually reaches an [12]–[17].
agreement after discussions [7]–[10]. The This paper generalizes the consensus model
consensus algorithm is also used to coordinate (1) by assuming that each agent has a state
large-scale systems such as formation of vector and a positive or negative definite scaling
vehicles, electrical, sensor, and traffic networks matrix. The proposed model has some
[11]. interesting features. First, thanks to the matrix
Consider a network in which the interactions weights, the system still achieves clustering
between subsystems, or agents, is modeled by a behavior, but the final states are not restricted
graph. In the consensus algorithm, each agent to be distributed along a straight line. Under the
updates its state based on the sum of the matrix-scaled consensus algorithm, a virtual
relative states with its nearby agents. If the consensus point is jointly determined by the
interaction graph is connected, the agents’ initial states and the scaling matrices of all
states asymptotically converge to a common agents. The state vector of each agent converges
point in the space, and we say that the system to a point differently from the virtual consensus
asymptotically reaches a consensus. point by the inverse of its scaling matrix. As a
∗
Department of Automation Engineering, School of
result, clustering behaviors usually happen, and
Electrical and Electronic Engineering, Hanoi University of agents with the same scaling matrix converge to
Science and Technology (HUST), Hanoi 11615, Vietnam. E- a common cluster in the space. Second,
mail: [email protected]
†
although the proposed consensus law has
Unmanned Aerial Vehicle Center, Viettel High Technology
Industries Corporation, Hanoi 11209, Vietnam. E-mail: similarities with the biased consensus [18] and
[email protected] orientation estimation algorithms [19], [20], in
‡
Department of Mechatronics, School of Mechanical the proposed model, the scaling matrices are
Engineering, Hanoi University of Science and Technology

1 Notations will be defined in detail in Section II.

346
not limited to the rotation matrices. Under the transpose is given by A>. The kernel, image, rank,
assumption that the interaction graph is and determinant of A are denoted by ker(A),
undirected and connected, it is shown that if the im(A), rank(A), and det(A), respectively. For a
scaling matrices are positive definite or negative vector
definite (possibly asymmetric), then the system P
achieves a matrix-scaled consensus globally x = [x1,...,xd]>, its 2-norm is kxk = q di=1 x2i. Let
asymptotically. Finally, the proposed algorithm d
x1,...,xn ∈ R , the vectorization operator is
can be used as a multi-dimensional model for
studying clustering behaviors in a social defined as vec(x1,...,xn ,...,xn>]> ∈ Rdn. A
network. Unlike the matrix-weighted consensus matrix
algorithm [6], in which a positive
A ∈ Rd×d is positive definite (negative definite) if
definite/semidefinite matrix weight associated
with an edge in the graph characterizes the and only if ∀x ∈ Rd, x 6= 0d, then x>Ax > 0 (resp.,
degree of cooperation between the agents in x>Ax < 0).
the network, in the matrix-scaled consensus
algorithm, a positive/negative definite matrix B. Useful lemma
weight corresponds to a vertex in the graph and Lemma 1: [8] The complex-coefficient
represents a local coordinate system of each polynomial p(s) = s2 +(a+bj)s+c+jd, where j2 = −1,
agent. Each local coordinate system can be is Hurwitz if and only if abd + a2c − d2 > 0.
interpreted as the private belief system of an
individual on d logically dependent topics, and C. Algebraic graph theory
this belief system is usually not perfectly aligned An undirected graph is given by G = (V,E),
with a social norm (a global coordinate system). where V = {1,...,n} is the vertex set and E ⊂ V2 is
If each individual updates his/her opinion based the set of |E| = m edges. If there is an edge (i,j)
on his/her own belief system, due to the ∈ E connecting vertices i,j ∈ V, i 6= j, then i and
existence of negative definite scaling matrices, j are adjacent to each other. The neighbor set of
the system will be unstable. By self-realizing the a vertex i, denoted by Ni = {j ∈ V | (i,j) ∈ E},
negative/positive of his/her own belief’s system contains all adjacent vertices of i. A path in G is a
relative to a social norm, each individual adjusts sequence of edges connecting adjacent vertices
the opinion along a direction that is not contrary in the graph. A graph is connected if and only if
to the social norm. The sign of the scaling matrix there is a path between any pair of vertices in V.
in the proposed algorithm, thus, realizes the Let the edges be indexed as E = {e1,...,em}, and
readiness of each individual to compromise in oriented such that for each edge ek = (i,j), i is the
order to prevent the society from divergence. starting vertex and j is the end vertex of ek. The
Moreover, we propose a matrix-scaled incidence matrix H = [hkl] ∈ Rm×n has hkl = −1 if l =
consensus algorithm for a network of double i, hkl = 1 if l = j, and hkl = 0, otherwise. The
integrator agents. A corresponding convergence Laplacian matrix L = [lij] ∈ Rn×n of G is defined as
condition related to the damping gain and the follows:
eigenvalues of the scaled Laplacian is also given.
The remainder of this paper is organized as 
−1, i 6= j, (i,j) ∈ E,
follows. Section II provides notations and the
theoretical framework that will be used 
throughout the paper. The matrixscaled lij = 0, i 6= j, (i,j) ∈ E/
consensus algorithms for single and double ,
integrator agents are proposed and examined in
 −Pni=1,i6=j lij, i = j.
Sections III and IV, respectively. Simulation
results are given in Section V and Section VI As G is undirected, L is symmetric positive
concludes the paper. semidefinite with eigenvalues given by 0 = λ1 ≤ λ2
II. Preliminaries ≤ ... ≤ λn. We can write L = H>H. For a connected
A. Notations graph, λ2 > 0 and ker(L) = ker(H) = im(1n).
The sets of real and complex numbers are D. Matrix-scaled consensus
denoted by R and C, respectively. Scalars are Consider a multi-agent system consisting of n
denoted by lowercase letters, while bold font agents. Each agent i ∈ {1,...,n} has a state vector
normal and capital letters are used for vectors xi ∈ Rd (d ≥ 2) and a scaling matrix Si ∈ Rd×d,
and matrices, respectively. Let A ∈ Rm×n, its

2 Note that Si is not required to be symmetric.

347
which is either positive definite or negative main diagonal. Introduce the variable
definite.2 Define the signum function ¯
transformation xc = Sx and let Θ = |S|L , where
1, Si is positive definite, ¯
sign(Si) = L = L ⊗ Id and |S| = blkdiag(|Si|). We can
−1, Si is negative definite. reexpress the xc-dynamics in the matrix form as
follows:
Then |Si| , sign(Si)Si is a positve definite matrix. It
is worth noting that sign(Si) = sign(S>i ) = sign(S−i 1) x˙c = −Θxc.
and |S−i 1| = |Si|−1. Let x = vec(x1,...,xn), the Note that if the signum term is omitted, the
following definition will be used in this paper: ¯ ¯
system (6) becomes x˙c = −SLx c. Since SL is the
Definition 1: The n-agent system achieves a product of a nondefinite matrix S with a positive
matrix scaled consensus (MSC) in a state x if and ¯ ¯
semidefinite matrix L , it will be likely that SL
only if x ∈ A, where
contains eigenvalues with positive real parts,
∈ and the system is unstable.
A = {x Rdn| S1x1 = S2x2 = ... = Snxn = xa}, (2)
and xa ∈ Rd is called the virtual consensus point B. Stability analysis

of the system. We will study the system (6) in this

subsection. Since |Si|, i ∈ V, are positive
Equivalently, the n-agent system achieves a
definite, |S| = blkdiag(|Si|) is positive definite.
matrixscaled consensus if and only if ¯
Thus, rank(Θ) = rank(L ) = dn−d and ker(Θ) =
xi = S−i 1xa = S−i 1Sjxj, ∀i,j ∈ V. (3) III. ¯
ker(L ) = im(1n ⊗ Id). The following lemma
Matrix-scaled consensus of single-
characterizes the spectrum of Θ:
integrator agents Lemma 2: Suppose that G is undirected and
A. The proposed consensus law connected. The matrix Θ has d zero eigenvalues
and dn−d eigenvalues with positive real parts.
Consider a system of n single-integrator
Proof: Following the proof of [21][Lem.
modeled agents in Rd (d ≥ 2):
¯
8.2.4], we define the following matrices H = H
x˙i = ui, i = 1,...,n, ⊗ Id,
where xi,ui ∈ Rd are respectively the state I s−1|S|H¯ > I 0
X= dn dn dn×dm .
variable and the input of agent i. The matrix-
H¯ Idm , and Y = −H¯ Idm
scaled consensus is proposed as follows: ui = Then,
X
sign(Si) (Sjxj − Sixi),
Idn − s−1|S|L¯ s−1|S|H¯ >
j∈Ni
XY = ,
0dm×dn Idm
X
= |Si| (S−i 1Sjxj − xi), i = 1,...,n. Idn s−1|S|H¯ >
(4)
j∈Ni YX = 0dm×dn −1H¯ |S|H¯ >
In the consensus algorithm (4), each agent i Idm − s
measures the relative state vector S−i 1Sjxj from From the fact that det(XY) = det(YX), one has
its neighboring agents, sums up the relative ¯ ¯ ¯
P det(Idn − s−1|S|L ) = det(Idm − s−1H |S|H
vector ri = i(S−i 1Sjxj − xi), and updates its state >
)
variable j∈N ¯ ¯ ¯>
along the direction of |Si|ri. The n-agent system sd(m−n)det(sIdn − |S|L ) = det(sIdm − H |S|H
can be written in the compart form as follows: (7)

x˙ = − diag(sign(Si))L ⊗ Idblkdiag(Si)x, (5) Thus, the nonzero eigenvalues of two matrices Θ

¯ ¯ ¯>
where x = vec(x1,...,xn), ⊗ stands for the = |S|L and N = H |S|H are the same. Since |
Kronecker product and S , blkdiag(Si) is the block ¯ ¯
S| is positive definite, N + N> = H (|S| + |S|>)H >
diagonal matrix with the matrices S1,...,Sn in the is symmetric and positive semidefinite,
348
 ¯ ˙ ¯ ¯
rank(N+N>) = rank(H ) = dn−d. Therefore, Θ has V = −xc>L (|S| + |S|>)Lx c.
dn − d eigenvalues with positive real parts.
Since |S| + |S|> is symmetric positive definite, it
Let Θ be expressed in the Jordan canonical ˙
follows that V ≤ 0. It follows that limt→+∞ V exists
form Θ = WJW−1, where W = [w1,...,wdn] ∈ Cdn×dn.
¯ ¨
For brevity, the notation W[j:k] = [wj,...,wk] is and is finite and Lx c is bounded. Thus, V =
¯ ¯ ¯
adopted to denote the columns from j to k (j < k) 2xc>L (|S|+|S|>)L |S|Lx c is also bounded. It
of the matrix W. follows from Barbalat’s lemma [22] that lim t→+∞
˙
By selecting W[1:d] = 1n⊗Id, we have ((W−1)>) V = 0. Thus, xc(t) → im(1n ⊗ Id) as t → +∞. Since
[1:d] = (|S−1|)>(1n ⊗ Id)P>, where |S−1| = blkdiag(| xc = Sx, there holds
S−i 1|) and P ∈ Rd×d is included so that the
normalization condition .
This means (
(((W−1)>)[1:d])>W[1:d] = Id Pn
|S−1|Sixi(0) = Pni=1 sign(Si)xi(0), ∀t ≥ 0. i=1
is satisfied. The equation (8) is equivalent to
From ( Pni
P Id,

P Pn
n |S−i 1|−1. Since i=1 sign(Si)xi(0), it follows that
all and it follows that P = i=1 =
nonzero eigenvalues of −Θ have negative real P
P ni=1 sign(Si)xi(0) = xa. Thus, xci(t) → xa, xi(t)
parts, there holds → S−i 1xa as t → +∞, or i.e., the
lim xc(t) = lim exp(−Θt)xc(0) system
t→+∞ t→+∞ (5) globally asymptotically achieves a matrix-
scaled consensus.
n
IV. Matrix-scaled consensus of double-integrator
= (1n ⊗ Id)PX|S−i 1|Sixi(0)
agents
i=1 n
! A. Proposed consensus laws
X This section studies the matrix-scaled
= 1n ⊗ P sign(Si)xi(0) . consensus algorithm for a system of double
i=1
integrators modeled by
Thus, limt→+∞ Sx(t) = 1n ⊗ xa, where x˙ , (12a) x˙ ui, i =
X X 1,...,n, (12b)
n !−1 n x
a
= |S−i 1| sign(Si)xi(t).
i=1 i=1 where xi1, xi2 ∈ Rd are states of agent i, and ui ∈
Rd is its control input. Let xi = vec(
Therefore, the system (5) asymptotically
achieves a matrix-scaled consensus. Because x˙ vec(x11,...,xn1), and x2 = vec( ,...,xn2). The
a
(t) = objective is to make the agents’ states xi1 to
Pn asymptotically achieve a matrix scaled
i=1 |Si | d,
consensus, i.e., to make x = vec(x1,x2)
and which shows that xa(t) is time-invariant. We asymptotically converge to the set
can now state the main theorem of this section.
Theorem 1: Suppose that G is undirected and A0 = {x ∈ R2dn| S1x11 = S2x21 = ... = Snxn1, x2 = 0dn}
connected. Under the matrix-scaled consensus
algorithm (4), x(t) → S−1(1n ⊗ xa) as t → +∞. The following consensus law is proposed to
Below, another proof of Theorem 1 will be achieve the matrix-scaled consensus:
given based on Barbalat’s lemma. X
ui = −sign(Si) Sjxj1) − αxi2,
¯
Proof: Consider the function V (xc) = x Lx c
>
c j∈Ni
¯
which is positive definite with regard to Lx c and
continuously differentiable. Moreover,
349
where α > 0 is a control gain. The n-agent system the matrix N1 has d zero eigenvalues and 2dn − d
under eigenvalues with negative real parts. The right
(13) is given as follows and left eigenvectors of N1 corresponding to the
x˙ 1 = x2, zero eigenvalues,
¯ 1n ⊗ Id
x˙ 2 = −(sign(Si)) ⊗ Id)LSx 1
− αx2. are columns and rows of the matrices , and 0n ⊗
B. Stability analysis Id
The behavior of the system (14) is given in the P . Thus,
following theorem. x1 x
Theorem 2: Suppose that G is undirected and t)
connected. Under the consensus law (13), x(t) x
asymptotically converges to a point in A0. t→+∞ c t→+∞ x
Proof: Let xc1 = Sx1 and xc2 = Sx2, we can "1n ⊗ xa
rewrite the system (14) as follows a2 #
=
0dn
xx˙˙cc1 = 0dn¯ −IαdnIdnxxcc21 = N1 xxcc21.
(15) where xal = Pni=1 |S−i 1|−1 Pni=1 sign(Si)xil(0), l =
2 −|S|L
1,2. This implies that limt→+∞ Six
−1 a2
¯ x and limt→ d, = 1,...,n.
Substituting Θ = |S|L = WJW−1 into the
V. Simulation results
characteristic equation det(sI2dn − N1) = 0, one Consider a system of 16 agents having the
2
has det(s Idn + αsIdn + WJW ) = 0, −1 interaction graph as depicted in Fig. 1. We will
provide some simulations to support the results
dn
in the previous sections.
Y
or, equivalently (s2 + αs + µk) =0, A. Simulation 1: MSC of single integrators
k=1
Let the SO(2) rotation matrix of angle θ (rad)
where µ1 = ... = µd = 0 and C 3 µk = ak + jbk 6= 0, ∀k be cos(θ) −sin(θ)
= d+1,...,dn. Based on Lemma 1, each polynomial denoted by R(θ) = . Let the scaling sin(θ)
s2 +αs+µk is Hurwitz if and only if ak /α2. cos(θ)
Thus, by choosing α so that matrices be chosen as S1 = ... = S6 = R(0) = I2
(positive definite), S7 = ... = S11 = R( ), and S12 =
(maxk=d+1,...,dn Im(µk))2 ... = S16 = R( ) (negative definite). The initial
min Re(µk) > condition x(0) is randomly selected. The
, (17) simulation results of (4) depicted in Fig. 2 show
k=d+1,...,dn α2 that the agents converge to three clusters,
which are three vertices of an equilateral
triangle. Agents with the same matrix Si
converge to the same cluster. Notice that θ
should be not equal to so that the analysis
holds. See [23] for further simulation of this
algorithm.
B. Simulation 2: MSC of double-integrators
Next, let the agents be modeled by double-
integrators. The scaling matrices are Si = R( ),i =
1,...,4, Si = R( ), i = 5,...,8, Si = R(− ), i = 9,...,12,
and Si = R(− ), i = 13,...,16. We conduct three
Fig. 1: The 16-vertex graph used in the simulations of the MSC algorithm (13) with α =
simulations. 1.8,

350
Fig. 2: The agents converge to 3 clusters in the plane under the matrix-scaled consensus algorithm (4).
1.9724 and 3, respectively. The trajectories of [6] Trinh et al., “Matrix-weighted consensus and its
applications,” Automatica, vol. 89, pp. 415–419, 2018.
the agents, corresponding to these parameters [7] Jadbabaie, Lin, and Morse, “Coordination of groups of
are depicted in Fig. 3. For α = 1.8, the system is mobile autonomous agents using nearest neighbor
unstable. Correspondingly, Figs. 3 (a), (d)–(e) rules,” IEEE Transactions on Automatic Control, vol. 48,
no. 6, pp. 988– 1001, 2003.
show the state variables grow unbounded. For
[8] Li et al., “Consensus of multiagent systems and
α = 1.9724, Θ has pairs of imaginary eigenvalues synchronization of complex networks: A unified
with the corresponding independent viewpoint,” IEEE Transactions on Circuits and Systems
eigenvectors, xli,l = 1,2, are asymptotic to I: Regular Papers, vol. 57, no. 1, pp. 213–224, 2009.
[9] Proskurnikov and Tempo, “A tutorial on modeling and
sinusoidal functions (see Figs. 3 (b), (h)–(k)). analysis of dynamic social networks: Part I,” Annual
Finally, for α = 3, the condition (17) is satisfied. Reviews in Control, vol. 43, pp. 65–79, 2017.
The agents converge to 4 clusters as shown in [10] Ye et al., “Continuous-time opinion dynamics on
multiple interdependent topics,” Automatica, vol. 115,
Figs. 3 (c), (l)–(o). no. 108884, 2020.
[11] Olfati-Saber, Fax, and Murray, “Consensus and
VI. Conclusions cooperation in networked multi-agent systems,”
Proceedings of the IEEE, vol. 95, no. 1, pp. 215–233,
A novel matrix-scaled consensus model, which 2007.
can describe a multi-dimensional opinion [12] Meng and Jia, “Scaled consensus problems on
switching networks,” IEEE Transactions on Automatic
dynamics system with heterogeneous Control, vol. 61, no. 6, pp. 1664–1669, 2015.
individuals’ private belief systems was proposed. [13] Meng and Jia, “Robust consensus algorithms for
The dissimilarities on individuals’ belief systems multiscale coordination control of multivehicle systems
with disturbances,” IEEE Transactions on Industrial
cause clustering phenomenon to happen Electronics, vol. 63, no. 2, pp. 1107–1119, 2015.
frequently. The matrix scaling gains allow agents [14] Aghbolagh, Ebrahimkhani, and Hashemzadeh, “Scaled
keeping their own biased states (in both consensus tracking under constant time delay,”
IFACPapersOnLine, vol. 49, no. 22, pp. 240–243, 2016.
amplitude and direction) with regard to a virtual
[15] Y. Shang, “On the delayed scaled consensus problems,”
consensus point. Extension of the model to Applied Sciences, vol. 7, no. 7, p. 713, 2017.
double-integrator agents was also proposed. For [16] Kenta et al., “On a new class of structurally balanced
further studies, it will be of interests to combine graphs for scaled group consensus,” in 58th Annual
Conf. Soc. Instrument Control Eng. Japan (SICE), 2019,
the matrix-scaled consensus with the Altafini pp. 1671–1676.
model, and study other applications such as [17] Wu et al., “Adaptive scaled consensus control of
scaled synchronization and formation control. coopetition networks with high-order agent dynamics,”
International Journal of Control, vol. 94, no. 4, pp. 909–
922, 2021.
References [18] Ahn and Trinh, “Consensus under biased alignment,”
Automatica, vol. 110, p. 108605, 2019.
[1] S. Roy, “Scaled consensus,” Automatica, vol. 51, pp.
[19] Lee and Ahn, “Distributed formation control via global
259–262, 2015.
orientation estimation,” Automatica, vol. 73, pp. 125–
[2] Ren, Randal, and Atkins, “Information consensus in
129, 2016.
multivehicle cooperative control,” IEEE Control
[20] Tran, Trinh, and Ahn, “Surrounding formation of star
Systems Magazine, vol. 27, no. 2, pp. 71–82, 2007.
frameworks using bearing-only measurements,” in
[3] Freeman, Peng, and Lynch, “Stability and convergence
Proc. of the European Control Conference, Cyprus,
properties of dynamic average consensus estimators,”
2018, pp. 368–373.
in 45th IEEE Conf Decision Control, CA, USA, 2006, pp.
[21] Godsil and Royle, Algebraic graph theory.
338–343.
Springer, 2001.
[4] Sarlette and Sepulchre, “Consensus optimization on
[22] Slotine and Li, Applied Nonlinear Control. Prentice Hall,
manifolds,” SIAM Journal of Control and Optimization,
Englewood Cliffs, NJ, 1991, vol. 199, no. 1.
vol. 48, no. 1, pp. 56–76, 2009.
[23] Trinh et al., “Matrix-scaled consensus,” arXiv preprint
[5] C. Altafini, “Consensus problems on networks with
arXiv:2204.10723, 2022.
antagonistic interactions,” IEEE Transactions on
Automatic Control, vol. 58, no. 4, pp. 935–946, 2013.
351
 (a) α = 1.8 (b) α = 1.9724

(d) x1i1 vs t [s] (e) x1i2 vs t [s] (f) xi21 vs t [s]

(h) x1i1 vs t [s] (i) x1i2 vs t [s] (j) xi21 vs t [s]

(l) x1i1 vs t [s] (m) x1i2 vs t [s] (n) xi21 vs t [s]

Fig. 3: Simulations of double-integrator agents: (a)–(c): Trajectories of agents with α = 1.8, 1.9724,
and 3, respectively; (d)–(g): xki1 and xki2, k = 1,2, vs time t [s] corresponding to α = 1.8; (h)–(k): xik1 and
xki2, k = 1,2, vs time t [s] corresponding to α = 1.9724; (l)–(o): xki1 and xki2, k = 1,2, vs time t [s]
corresponding to α = 3.

352