This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2018.2872197, IEEE
Transactions on Automatic Control
MANUSCRIPT TO IEEE TRANS. ON AUTOMATIC CONTROL 1

Consensus Control of General Linear Multi-agent

Systems with Antagonistic Interactions and
Communication Noises
Jiangping Hu, Senior Member, IEEE, Yanzhi Wu, Tao Li, Senior Member, IEEE, and Bijoy K.
Ghosh, Fellow, IEEE

Abstract—This technical note considers a consensus problem representing the noisy intensities are positive definite diagonal
of high-order multi-agent systems with antagonistic interactions matrices. It is assumed that ρij = ρji .
and communication noises. The interaction network associated Our objective is to design a distributed control for the sys-
with the multi-agent system is modeled by a signed graph (called
coopetition network) and the agent dynamics is described by a tem (1) by using the relative state measurements to guarantee
general linear system. A novel stochastic-approximation based the mean square bipartite consensus, that is,
control strategy is designed for each agent by using the relative N
state information from its neighbors. Additionally, convergence
n 1 X 2 o
lim E xi (t) − si sj xj (t) = 0, (2)

of a consensus error system is analyzed by the stochastic stability t→∞ N j=1
theory. Finally, the effectiveness of our results is demonstrated
by simulation.
where E{·} denotes the mathematical expectation, and si
Index Terms—Mean square bipartite consensus, linear multi- equals 1 or −1. It is noted that the equation (2) is equivalent
agent systems, communication noise, stochastic-approximation
to the following two equations
gain, coopetition network.
N 2 o
n 1 X
I. I NTRODUCTION lim E xi (t) − sj xj (t) = 0 for si = 1,

t→∞ N j=1

R ECENTLY, distributed consensus problems of multi-

agent systems have attracted much attention in the
control community due to their widespread applications in
n
lim E xi (t) +
t→∞
1 X
N

N j=1
2 o
sj xj (t) = 0 for si = −1.

systems such as unmanned air vehicles, satellite formation

flying and distributed reconfigurable sensor networks, etc. In
most cases, communication between agents is degraded due
to thermal, fading channel and quantization noises during
encoding and decoding. The effect of these noises should
therefore be carefully considered while investigating realistic
problems with multi-agent systems.
In this technical note, we consider a multi-agent system with
a general linear system as follows,
ẋi = Axi + Bui , i = 1, · · · , N,
n uses the relative state information in the neighborhood.
where xi = col(xi1 , xi2 , · · · , xin ) ∈ R represents the state
of agent i, and ui ∈ Rm is the control input. It is assumed
that (A, B) is a stabilizable pair. For agent i ∈ V, a relative
state measurement from its neighbor j is defined by
 
[1]
zij = |Aij | xi − sign(Aij )(xj + ρij nij ) ,
 cooperative-competitive multi-agent systems with com-
where nij (t) ∈ Rn , i, j = 1, 2, · · · , N are independent
standard white noises and ρij ∈ Rn×n , i, j = 1, 2, · · · , N
J. Hu and Y. Wu are with the School of Automation Engineering, University
of Electronic Science and Technology of China, Chengdu 611731, China.
T. Li is with the Department of Mathematics, East China Normal University,
Shanghai 200241, China.
B. K. Ghosh is with the Department of Mathematics and Statistics, Texas
Tech University, Lubbock, TX, 79409-1042, USA.
This work was supported by the National Natural Science Foundation of
China under Grant nos. 61473061, 61522310, 71503206, 61104104, and the
Program for New Century Excellent Talents in University under Grant no.
NCET-13-0091.
To the best of our knowledge, very few results have been
obtained for consensus control of high-order multi-agent sys-
tems with communication noises due to great challenges in
the controller design and the convergence analysis.
The main contributions of this note are listed as follows.
(i) This note proposes a distributed control strategy which
is not dependent on the structure of the state matrix of
the agent dynamics. Moreover, the strategy does not need
(1) the absolute state information of the agents own and only
(ii) A novel stochastic-approximation type protocol is de-
veloped to attenuate communication noises. Through the
protocol, the primary term in the closed-loop system is
noise-free, which is beneficial for convergence analysis.
(iii) This note investigates a bipartite consensus control of
munication noises, which is analyzed from a new per-
spective.
The remainder of this note is organized as follows. Some
helpful preliminary results are presented in Section II. A mean
square bipartite consensus problem is solved in Section III
with a dynamic output-feedback control strategy. Simulation
results are presented in Section IV, and concluding remarks
are given in Section V.
Notations: Throughout this note, the following notations are
used. In ∈ Rn×n represents the identity matrix, 1 denotes a

0018-9286 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2018.2872197, IEEE
Transactions on Automatic Control
MANUSCRIPT TO IEEE TRANS. ON AUTOMATIC CONTROL 2

column vector with elements all being 1. λmin (·) and λmax (·) B. Algebraic graph theory
denote the minimum and maximum eigenvalues, respectively. It is convenient to use a signed graph to model the coope-
P > 0 (P ≥ 0) denotes a positive definite (positive semi- tition network. An undirected signed graph G = (V, E, A) is
definite) matrix. χs denotes the indicator function. col(·) introduced, where V = {1, 2, · · · , N } is the set of vertices
denotes a concatenation. diag(·) represents a diagonal matrix. representing agents, E ⊆ V × V denotes the set of edges
sign(·) denotes a sign function. and A = [Aij ] ∈ RN ×N is the adjacency matrix. When
the interaction weight Aij 6= 0, it means that there is
a connection between agent i and agent j. Specifically, if
II. P RELIMINARY RESULTS Aij > 0, the interaction between agent i and agent j is
cooperative and competitive if Aij < 0. It is assumed that
A. Motivations the signed graph G is simple, i. e. Aii = 0 for i ∈ V.
The undirected signed graph G is structurally balanced, if
Distributed control of multi-agent systems with communi- the set of vertices V can be divided into two disjoint subsets
cation noises was firstly concerned in [1], [2], [3], where the V1 = {1, 2,S · · · , N0 }Tand V2 = {N0 +1, · · · , N }, which satis-
agent dynamics is first-order. A stochastic delayed consensus fy V = V1 V2 , V1 V2 = ∅, Aij ≥ 0, ∀i, j ∈ Vp (p ∈ 1, 2),
problem was considered in [4]. Furthermore, necessary and Aij ≤ 0, ∀i ∈ Vp , j ∈ Vq , p 6= q (p, q ∈ {1, 2}).
sufficient conditions were provided in [5] for an average For a structurally balanced graph, a gauge transformation
consensus problem of continuous-time second-order multi- is defined by S = diag(s1 , s2 , · · · , sN ) ∈ RN ×N , where
agent systems with communication noises and fixed topol- si = 1 when vertex i ∈ V1 and si = −1 when vertex
ogy. Although some distributed control strategies have been i ∈ V2 . Obviously, the gauge transformation is an orthogonal
developed for first-order or second-order multi-agent systems matrix. Additionally, the adjacency matrix A is similar to a
with communication noises in the past few years, the results nonnegative matrix through the gauge transformation.
are scarce in the case when the agent dynamics is high- A signed Laplacian matrix is defined by
order. A mean square consensus problem was studied in
[6] for continuous-time high-order multi-agent systems with L = Cr − A,
communication noises and fixed topology. Furthermore, a
where Cr is a diagonal matrix with the diagonal elements
consensus problem was investigated for discrete-time single- N
P
input multi-agent systems in [7] and for multi-input multi- being cr,ii = |Aij | for i = 1, 2, · · · , N . The signed
j=1
output continuous-time multi-agent systems in [8]. However,
Laplacian matrix has an important spectral property as follows,
the control strategies proposed in [6]-[8] need the absolute
Lemma 1: ([11], [12]) Let L be the signed Laplacian matrix
state information of the agents own, which motivates us to
of the undirected signed graph G. If G is structurally balanced
further consider a distributed control design with only relative
and connected, then the following statements are true.
state information in the neighborhood.
• L has a simple zero eigenvalue with s =
Another motivation is to consider high-order multi-agent
col(s1 , s2 , · · · , sN ) ∈ RN as its zero eigenvector,
systems with cooperative-competitive interactions. It is well
and all the non-zero eigenvalues are strictly positive.
known that cooperation and competition coexist in many
• The smallest nonzero eigenvalue λ2 (L) of the matrix L
complex systems where two teams compete, for example,
satisfies
activators-inhibitors in biological systems, markets and social xT Lx
networks with two competing cartels, and competing robots. λ2 (L) = min .
xT s=0,x6=0 xT x
As a specific social dynamics, opinion formation was inves-
tigated for trust-mistrust networks through DeGroot-type or III. M EAN SQUARE AVERAGE BIPARTITE CONSENSUS
Laplacian-type dynamics in [9], [10]. Generally, consensus,
In this note, a dynamic output-feedback controller is de-
polarization and fragmentation can emerge on coopetition
signed for agent i,
networks even though they have connected topologies [11].  [1] [2]
βij ρ−1 βij %−1

P
Particularly, polarization or bipartite consensus, where all
P
ui = a(t) + c K
 ij zij − cK ij zij ,
agents reach a final state with identical magnitude but with j∈Ni
P [2]
j∈Ni
P [1]
opposite sign, was investigated extensively in the control θ̇i = (A + BK1 )θi − r1 a(t)
 zij − r2 a(t) zij ,
j∈Ni j∈Ni
community [12], [13], for coopetition networks with first-
(3)
order agent dynamics. An interval bipartite consensus prob- [2]
where θi is the internal state of the controller, zij =
lem was solved in [14] for first-order multi-agent systems.
Some adaptive control strategies were proposed for bipartite |Aij | θi − sign(Aij )(θj + %ij nij ) is the relative internal
consensus of reaction-diffusion neural networks in [15] and state, βij = βji ∈ Rn×n is a positive definite diagonal matrix
high-order multi-agent systems in [16], [17]. Cooperative satisfying βij ρ−1
ij = βρ In , βρ is a positive constant. c, r1 and
output synchronization and bipartite output synchronization r2 are positive constants. K ∈ Rm×n and K1 ∈ Rm×n are
were investigated by using the absolute state information in feedback gain matrices. The time-varying gain a(t) > 0 is a
[18] for heterogenous linear multi-agent systems. However, till uniformly continuous function.
now, there is no result investing high-order coopetitive multi- Remark 1: The controller (3) uses only the relative in-
agent systems with communication noises. formation of agent i from its neighbors, which is different

0018-9286 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2018.2872197, IEEE
Transactions on Automatic Control
MANUSCRIPT TO IEEE TRANS. ON AUTOMATIC CONTROL 3

from the existing control strategies in the references [6]- noted that the matrix M satisfies M L = LM = L. Since
[8]. Additionally, the appropriate gains a(t) and βij can be (sT ⊗ In )Π = Π(s ⊗ In ) = 0, one has
selected such that the term, which plays an essential role  for    
βij ρ−1 (M ⊗ In ) IN ⊗ (BK) Π = IN ⊗ (BK) Π(M ⊗ In ).
P
consensus, in the controller (3) is cK ij |Aij | xi −
j∈Ni
   (5)
βij %−1
P
sign(Aij )xj − cK ij |Aij | θi − sign(Aij )θj . Let xθ = col(x, θ) ∈ R2N n . A variable of change is
j∈Ni
2: If the time-varying gain a(t) satisfies given by ξ = col(ξ1 , ξ2 ) = Mξ xθ . From (5), the system
R ∞Remark
2 (4) is transformed to the following Itô stochastic differential
0
a (s)ds < ∞, from Barbalat’s Lemma, it is not difficult
to show that a(t) is bounded in [0, +∞), i.e., |a(t)| < ā for equation
a positive constant ā.
dξ = Ac ξdt + a(t)Bc ξdt + a(t)Dc dW (t), (6)
Some assumptions are listed as follows,
(A1 ) G = (V, E, A) is structurally balanced.
where
(A2 ) G R∞= (V, E, A) is connected.
(A3 ) R0 a(s)ds = ∞.
  !
∞ 2 IN ⊗ A + cβρ L ⊗ (BK) −c IN ⊗ (BK) Π
(A4 ) 0 a (s)ds < ∞. Ac = ,
0 IN ⊗ (A + BK1 )
Applying the controller (3) to the multi-agent system (1)
results in the following closed-loop system  
βρ L ⊗ (BK) 0 
  i Bc = and Dc = col −
−r2 L ⊗ In −r1 L ⊗ In
    h


 ẋ = IN ⊗ A x + cβρ L ⊗ (BK) x − c IN ⊗ (BK) Π θ


  
+a(t)βρ L ⊗ (BK) x − a(t) IN ⊗ (BK) D η,

h  i M ⊗ BK D, r 1 (M ⊗ I n )D 1 + r2 (M ⊗ I n )D 2 . W (t) =
2
    col(w1 , w2 , · · · , wN ) is an N n dimensional standard Browni-
θ̇ = IN ⊗ (A + BK1 ) θ − r1 a(t) L ⊗ In θ + r1 a(t)D1 η an motion, wi = col(wi1 , wi2 , · · · , wiN ), and t nij ds = wij .
 R


   0
Now a main result is given to provide a sufficient condition

−r2 a(t) L ⊗ In x + r2 a(t)D2 η,

(4) for the mean square average bipartite consensus.
where x = col(x1 , x2 , · · · , xN ) ∈ R Nn
and θ = Theorem 1: Consider the multi-agent system (1). Suppose
Nn
col(θ1 , θ2 , · · · , θN ) ∈ R . D = diag(d1 , d2 , · · · , dN ) ∈ that Assumptions (A1 )−(A4 ) hold. Let P > 0 and P1 > 0 be
R N n×N 2 n
, di = (Ai1 βi1 , Ai2 βi2 , · · · , AiN βiN ) ∈ R n×N n
, the solutions of the algebraic Riccati equation (ARE) f (P) =
N 0 with κ = k and κ = k1 , respectively, where
Π = (Πij )N ×N ∈ RN n×N n , Πii = |Aik |βik %−1
P
ik ,
k=1
ˆ ˆ ˆ f (P) = AT P + PA − PBB T P + κIn = 0. (7)
Πij = −Aij βij %−1 ij for i 6= j, D1 = diag(d1 , d2 , · · · , dN ) ∈
2
RN n×N n , dˆi = (Ai1 %i1 , Ai2 %i2 , · · · , AiN %iN ) ∈ Rn×N n , Then the mean square average bipartite consensus can be
2
D2 = diag(d¯1 , d¯2 , · · · , d¯N ) ∈ RN n×N n , achieved for all the agents under the distributed controller
d¯i = (Ai1 ρi1 , Ai2 ρi2 , · · · , AiN ρiN ) ∈ Rn×N n , η = (3) with the feedback gain matrices K = −B T P and
2
col(n1 , n2 , · · · , nN ) ∈ RN n , ni = col(ni1 , ni2 , · · · , niN ) ∈ K1 = − 1 B T P1 .
2
RN n . Proof: Consider a Lyapunov function candidate
N n×n
Define s̄ = s ⊗ In ∈ R . Then we have the following
lemma for the matrix Π. V = ξ T P̄ ξ, (8)
Lemma 2: If G is structurally balanced and connected, Π  
has a simple zero eigenvalue satisfying Πs̄ = 0, and all the L⊗P 0
where P̄ = . According to Lemma 1, one
non-zero eigenvalues are strictly positive. 0 L ⊗ P1
Proof:
 N From the definition, one has Π = has
|A1k |β1k %−1 −A12 β12 %−1 −A1N β1N %−1
P
···
 k=1
1k 12 1N
 ξ T P̄ ξ = ξ1T L ⊗ P ξ1 + ξ2T L ⊗ P1 ξ2
N
−A21 β21 %−1 |A2k |β2k %−1 −A2N β2N %−1 ≥ λ2 (L)λmin (P )ξ1T ξ1 + λ2 (L)λmin (P1 )ξ2T ξ2
 P 
 21 2k ··· 2N  (9)
 k=1 .
. . .
≥ c2 ξ T ξ > 0, ∀ξ 6= 0,
 
. . .

 . . ··· . 

N
−AN 1 βN 1 %−1 −AN 2 βN 2 %−1 |AN k |βN k %−1
P
N1 N2 ··· Nk
 
k=1
It is not difficult to see that Πs̄ = 0. Additionally, Π is similar where c2 = min λ2 (L)λmin (P ), λ2 (L)λmin (P1 ) . Therefore,
to a Laplacian matrix associated with an unsigned undirected the Lyapunov function candidate V satisfies
graph through an orthogonal transformation S ⊗ In . Thus the
conclusion follows with a similar proof of Lemma 1.  c2 ξ T ξ ≤ V ≤ λmax (P̄ )ξ T ξ. (10)
An average bipartite consensus error is defined for agent i
N The derivative of V along the trajectory determined by (6)
by ξ1i = xi − N1 si sj xj ∈ Rn , or, in a compact form,
P
is given by
j=1  
ξ1 = col(ξ11 , ξ12 , · · · , ξ1N ) = (IN − N1 ssT ) ⊗ In x. Let dV =ξ T P̄ Ac + ATc P̄ + a(t)(P̄ Bc + BcT P̄ ) ξdt
 
(11)
M = IN − N1 ssT and Mξ = diag(M ⊗ In , M ⊗ In ). It is + a2 (t)tr(DcT P̄ Dc )dt + 2a(t)ξ T P̄ Dc dW (t).

0018-9286 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2018.2872197, IEEE
Transactions on Automatic Control
MANUSCRIPT TO IEEE TRANS. ON AUTOMATIC CONTROL 4

When the gain matrices are selected respectively as K = From (10), (11) and (13), one has
−B T P and K1 = − 21 B T P1 , one has dV =ξ T (P̄ Ac + ATc P̄ )ξdt + a(t)ξ T (P̄ Bc + BcT P̄ )ξdt
+ a2 (t)tr(DcT P̄ Dc )dt + 2a(t)ξ T P̄ Dc dW
ξ T P̄ Ac + ATc P̄ + a(t)(P̄ Bc + BcT P̄ ) ξ
 
h  i ≤ − ξ T Qξ + a2 (t)tr(DcT P̄ Dc )dt + 2a(t)ξ T P̄ Dc dW
≤ξ1T L ⊗ (P A + AT P ) − 2cβρ L2 ⊗ (P BB T P ) ξ1
λmin (Q)
h   i ≤− a(t)V dt + a2 (t)tr(DcT P̄ Dc )dt
+ 2ξ1T c L ⊗ (P BB T P ) Π − r2 a(t)L2 ⊗ P1 ξ2 λmax (P̄ )ā
h i + 2a(t)ξ T P̄ Dc dW (t).
+ ξ2T L ⊗ (AT P1 + P1 A − P1 BB T P1 ) ξ2 . (14)
(12) Next, we will show
For the signed Laplacian  L , there exists a u-
matrix hZ t i
a(s)ξ T P̄ Dc dW (t) = 0,

nitary matrix U = √s U 1 such that U T LU = E (15)
N
t0
Λ = diag λ1 (L), λ2 (L), · · · , λN (L) , where λi (L) are t0
for t ≥ t0 ≥ 0. Define a stopping time τm̄ , inf{t ∈ [t0 , T ] :
the eigenvalues of L. Define the variable of changes by
V (t) ≥ m̄}, where t0 ≥ 0, T ≥ t0 , m̄ > 0. Then, from (14),
ξˆ1 = col(ξˆ11 , ξˆ12 , · · · , ξˆ1N ) = (U T ⊗ In )ξ1 and ξˆ2 =
one has
col(ξˆ21 , ξˆ22 , · · · , ξˆ2N ) = (U T ⊗ In )ξ2 . By properly choosing t0
positive constants c, βρ such that cβρ λ2 (L) ≥ 21 and according E[V (t ∧ τm̄ )χt≤τ t0 ] − E[V (t0 )]
to the algebraic Riccati equations (7), the inequality (12) can Z t m̄ Z t
λmin (Q)
be transformed to ≤− a(s)V (s)ds + tr(DcT P̄ Dc ) a2 (s)ds
λmax (P̄ )ā t0 t0
Z t
ξ T (P̄ Ac + ATc P̄ )ξ + a(t)ξ T (P̄ Bc + BcT P̄ )ξ ≤tr(DcT P̄ Dc ) a2 (s)ds,
h  i
t0
≤ξˆ1T Λ ⊗ (P A + AT P ) − 2cβρ Λ2 ⊗ (P BB T P ) ξˆ1
t0
h   i which means that E[V (t∧τm̄ )χt≤τ t0 ] ≤ M1 < ∞, where M1
m̄
+ 2ξ1T c L ⊗ (P BB T P ) Π − r2 a(t)L2 ⊗ P1 ξ2 is a positive constant and dependent on t0 and T . It is noted
t0
h i that lim t ∧ τm̄ = t, then one has sup E[V (t)] ≤ M1
+ ξˆT Λ ⊗ (AT P1 + P1 A − P1 BB T P1 ) ξˆ2
2
M1 →∞ t0 ≤t≤T
by Fatou lemma. Thus, ∀t ∈ [t0 , T ], one has
≤ − kλ2 (L)ξ1T ξ1 − k1 λ2 (L)ξ2T ξ2 hZ t i Z T
h  
i
+ 2ξ1T c L ⊗ (P BB T P ) Π − r2 a(t) L2 ⊗ P1 ξ2 , E a2 (s)V (s)ds ≤ sup E[V (t)] a2 (s)ds < ∞.
t0 t0 ≤t≤T t0

or, Since T can be arbitrary, thus one has

hZ t i
E a2 (s)V (s)ds < ∞, ∀t ≥ t0 .
ξ T (P̄ Ac + ATc P̄ )ξ + a(t)ξ T (P̄ Bc + BcT P̄ )ξ ≤ −ξ T Qξ, (13) t0
  Furthermore, one has
Q11 Q12
where Q = , Q11 = kλ2 (L)IN n , Q12 = hZ t i hZ t i
 Q21 Q22    E a2 (s)kξ T P̄ Dc k2 ds ≤ kP̄ kkDc k2 E a2 (s)V (s)ds ,
−c L⊗(P BB T P ) Π+r2 a(t) L2 ⊗P1 , Q21 = −cΠT L⊗ t0 t0
   which leads to (15).
(P BB T P ) + r2 a(t) L2 ⊗ P1 , and Q22 = k1 λ2 (L)IN n .
   From the analysis presented above, for any t ≥ 0 and h > 0,
Furthermore, let Q̄1 = r2 L2 ⊗ P1 and Q̄2 = c L ⊗ one has
 Z t+h
(P BB T P ) Π, then one has Q12 = QT21 = a(t)Q̄1 − Q̄2 . λmin (Q)
E[V (t + h)] − E[V (t)] ≤ − a(s)V (s)ds
It is not difficult to obtain λmax (P̄ )ā t
Z t+h
QT12 Q12 ≤a(t)(Q̄1 − Q̄2 )T (Q̄1 − Q̄2 ) + a2 (t)Q̄T1 Q̄1 + Q̄T2 Q̄2 + tr(DcT P̄ Dc ) a2 (s)ds , V1 (t, t + h).
t
≤ā(Q̄1 − Q̄2 )T (Q̄1 − Q̄2 ) + ā2 Q̄T1 Q̄1 + Q̄T2 Q̄2 . Thus, one has
E[V (t + h)] − E[V (t)] V1 (t, t + h)
By using Remark 2, Lemma 2, Schur Complement Lemma lim sup ≤ lim sup
h→0+ h h→0+ h
and selecting c, r2 , a(t), k, k1 properly such that following
conditions hold λmin (Q)
=− a(t)E[V (t)] + tr(DcT P̄ Dc )a2 (t).
 λmax (P̄ )ā
kλ (L) > 0, By using the comparison lemma, one has
 2

 
k1 λ2 (L)IN n − kλ21(L) ā(Q̄1 − Q̄2 )T (Q̄1 − Q̄2 ) Z t λ
− min
(Q) R t
a(τ )dτ
  E[V (t)] ≤ tr(DcT P̄ Dc ) a2 (s)e λmax (P̄ )ā s ds
+ā2 Q̄T Q̄ + Q̄T Q̄ > 0,

0

1 1 2 2
λmin (Q) Rt
−λ a(s)ds
+ E[V (0)]e max (P̄ )ā 0
.
it follows that the symmetry matrix Q is positive definite. (16)

0018-9286 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2018.2872197, IEEE
Transactions on Automatic Control
MANUSCRIPT TO IEEE TRANS. ON AUTOMATIC CONTROL 5

R∞
From Assumption (A4 ), that is, 0 a2 (s)ds < ∞, R ∞one has, for
any given  > 0, there exists T ∗ > 0 such that T ∗ a2 (s)ds <
. For the right-side term in (16), one has
Z t λ (Q) R t
T − min a(τ )dτ
0 ≤ tr(Dc P̄ Dc ) a2 (s)e λmax (P̄ )ā s ds
0
Z T ∗
λ (Q) R t
− min a(τ )dτ
= tr(DcT P̄ Dc ) a2 (s)e λmax (P̄ )ā s ds
0
Z t λ (Q) R t
− min a(τ )dτ
+ tr(DcT P̄ Dc ) a2 (s)e λmax (P̄ )ā s ds
T∗ Fig. 1. Coopetition network G
λ (Q) R t
Z ∞
− min a(τ )dτ
≤ tr(DcT P̄ Dc )e λmax (P̄ )ā T ∗ a2 (s)ds
0 The eigenvalues of the Laplacian matrix L are given by
Z ∞
T 2 λ1 (L) = 0, λ2 (L) = 0.753, λ3 (L) = 0.753, λ4 (L) =
+ tr(Dc P̄ Dc ) a (s)ds
T∗ 2.445, λ5 (L) = 2.445, λ6 (L) = 3.8019 and λ7 (L) =
≤ o(1) + tr(DcT P̄ Dc ), 3.8019. The zero eigenvector of the Laplacian matrix L
as t → ∞. By the arbitrariness of , one has is given by s = col(1, 1, 1, −1, −1, −1, −1). Take c =
Z t 4, r1 = 1, r2 = 2, k = 10,and k1 = 20. The t-
λ (Q) R t
T − min a(τ )dτ wo positive definite
 positive matrices are respectively giv-
lim tr(Dc P̄ Dc ) a2 (s)e λmax (P̄ )ā s ds = 0. 20.0277 15.0554 3.1623

t→∞ 0
en by P = 15.0554 26.9903 6.3333 and P1 =
Thus, according to (16), one has lim E[V (t)] = 0, further- 3.1623 6.3333 4.7609
t→∞
more, lim Ekξ(t)k2 = 0, which implies lim Ekξ1 (t)k2 = 0.
 
t→∞ t→∞
38.6284 27.3039 4.4721
Therefore, the mean square average bipartite consensus is 27.3039 48.2632 8.6376. The feedback gain matrices
achieved for the multi-agent system. The proof is thus com- 4.4721 8.6376 6.1053

pleted.  are selected as K = −3.1623 −6.3333
−4.7609 and
Remark 3: In the controller (3), we assume that the K1 = −2.2361 −4.3188 −3.0527 . The time-varying
1
stochastic-approximation gain a(t) is uniformly continuous in gain is designed as a(t) = t+20 . The intensity matrices of
the interval [0, ∞), which plays a key role in the proof of the communication noises ρij ∈ R3×3 and %ij ∈ R3×2 for
Theorem 1. Additionally, when the interaction weights aij are i, j = 1, 2, · · · , 7 are arbitrarily positive definite diagonal
constrained to be nonnegative numbers, the controller (3) can matrices such that βρ = 14 . Under the proposed controller
guarantee an average consensus for cooperative multi-agent (3), the average consensus errors all approach to zero in mean
systems, which is a great extension of the existing results. square, as shown in Fig. 2. Simultaneously, the seven agents
Remark 4: Structural balance is an important assumption reach an average bipartite consensus, as illustrated in Fig. 3,
7 2
on the coopetition network G, which is a sufficient condition

that is, E xi (t) − 71
P
sj xj (t) → 0 for i = 1, 2, 3, while

to ensure the bipartite consensus of cooperative-competitive j=1
multi-agent systems. In fact, the equilibrium set of the multi- 7 2
E xi (t) + 71
P
sj xj (t) → 0 for i = 4, 5, 6, 7, respectively,

agent systems is determined by {s ⊗ c(t)|c(t) ∈ Rn }, where j=1

x1 (0)
 as time t goes to infinity.
1 T
s ⊗ eAt  ...  ∈ Rn .
 
c(t) =
 
N V. C ONCLUSION
xN (0)
In this note, an average bipartite consensus problem has
Thus it is noted the linearly independent group depends on been investigated for general linear multi-agent systems with
the initial states x(0) = col(x1 (0), · · · , xN (0)) and the state antagonistic interactions and communication noises. A dis-
matrix A. tributed dynamic output-feedback controller together with a
stochastic-approximation gain has been designed for each
IV. S IMULATION agent to achieve bipartite consensus in mean square. A simu-
lation example has been presented to validate the proposed
A simulation example is presented in this section. Suppose control strategy. Some future topics are to design optimal
that the agent dynamics (1) is described by a third-order adaptive control strategies for linear multi-agent systems with
linear system and seven agents interact each other on the unknown time-varying disturbances, communication noises
coopetition network as illustrated in Fig. 1. The blue solid and optimal control parameters.
edges and the red dash edges are used to describe cooperative
and competitive relationships, respectively. From the definition
R EFERENCES
of the structural balance, G is structurally balanced and can
be divided into two competitive subnetworks V1 and V2 , [1] M. Huang and J. H. Manton, “Coordination and consensus of networked
agents with noisy measurements: Stochastic algorithms and asymptotic
where V1 = {1, 2, 3} and V2 = {4, 5, 6, 7}. The gauge behavior,” SIAM J. Control and Optimiz., vol. 48, no. 1, pp. 134-161,
transformation matrix S = diag(1, 1, 1, −1, −1, −1, −1). Feb. 2009.

0018-9286 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2018.2872197, IEEE
Transactions on Automatic Control
MANUSCRIPT TO IEEE TRANS. ON AUTOMATIC CONTROL 6

8 5000
agent1
6 agent2
agent3
si sj xj,1 (t)
4 agent4
agent5 agent1
agent6 agent2
2
agent7 agent3

xi,1 (t)
1X
7 j=1

agent4
7

0 0
agent5
−2 agent6
xi,1 (t) −

agent7
−4

−6

−8 −5000
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time Time

(a) (a)
3 150
agent1
agent2
2 agent3 100
si sj xj,2 (t)

agent4 agent1
agent5
1 50 agent2
agent6
agent3
agent7

xi,2 (t)
agent4
1X
7 j=1

0 agent5
7

0
agent6
agent7
xi,2 (t) −

−1 −50

−2 −100

−3 −150
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time Time

(b) (b)
8 15
agent1 agent1
6 agent2 agent2
agent3 agent3
si sj xj,3 (t)

10
4 agent4 agent4
agent5 agent5
agent6 agent6
2 5
agent7 agent7
xi,3 (t)
1X
7 j=1
7

0
0
−2
xi,3 (t) −

−4
−5
−6

−8 −10
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time Time

(c) (c)

Fig. 2. Evolution of the consensus errors ξ1i (t) ∈ R3 (i = 1, · · · , 7) Fig. 3. Evolution of the agent states xi (t) ∈ R3 (i = 1, · · · , 7)

[2] T. Li and J. F. Zhang, “Consensus conditions of multi-agent systems
with time-varying topologies and stochastic communication noises,” IEEE
Trans. Autom. Control, vol. 55, no. 9, pp. 2043-2057, Sep. 2010.
[3] T. Li and J. F. Zhang, “Mean square average-consensus under measure-
ment noises and fixed topologies: Necessary and sufficient conditions,”
Automatica, vol. 45, no. 8, pp. 1929-1936, Aug. 2009.
[4] X. Wu, Y. Tang, and W. Zhang, “Stability analysis of stochastic delayed
systems with an application to multi-agent systems,” IEEE Trans. Autom.
Control, vol. 61, no. 12, pp. 4143-4149, Dec. 2016.
[5] L. Cheng, Z. G. Hou, M. Tan, and X. Wang, “Necessary and sufficient
conditions for consensus of double-integrator multi-agent systems with
measurement noises,” IEEE Trans. Autom. Control, vol. 56, no. 8, pp.
1958-1963, Aug. 2011.
[6] L. Cheng, Z. G. Hou, and M. Tan, “A mean square consensus protocol
for linear multi-agent systems with communication noises and fixed
topoligies,” IEEE Trans. Autom. Control, vol. 59, no. 1, pp. 261-267,
Jan. 2014.
[7] Y. Wang, L. Cheng, Z. G. Hou, M. Tan, and M. Wang, “Consensus seeking
in a network of disctete-time linear agents with communication noises,”
Int. J. Syst. Sci., vol. 46, no. 10, pp. 1874-1888, Sep. 2015.
[8] Y. Wang, L. Cheng, W. Ren, Z. G. Hou, and M. Tan, “Seeking consensus
in networks of linear agents: Communication noises and markovian
switching topologies,” IEEE Trans. Autom. Control, vol. 60, no. 5, pp.
1374-1379, May 2015.
[9] A. V. Proskurnikov, A. S. Matveev, and M. Cao, “Opinion dynamics in
social networks with hostile camps: Consensus vs. polarization,” IEEE
Trans. Autom. Control, vol. 61, no. 6, pp. 1524-1536, June 2016.
[10] W. Xia, M. Cao, and K. H. Johansson, “Structural balance and opinion
separation in trust-mistrust social networks,” IEEE Trans. Control Netw.
Syst., vol. 3, no. 1, pp. 46-56, Mar. 2016.
[11] J. Hu and W. X. Zheng, “Emergent collective behaviors on coopetition
networks,” Physics Letter A, vol. 378, no. (26-27), pp. 1787-1796, May
2014.
[12] C. Altafini, “Consensus problems on networks with antagonistic inter-
actions,” IEEE Trans. Autom. Control, vol. 58, no. 4, pp. 935-946, Apr.
2013.
[13] Z. Meng, G. Shi, K. H. Johansson, M. Cao and Y. Hong, “Behaviors
of networks with antagonistic interactions and switching topologies,”
Automatica, vol. 73, no. 11, pp. 110-116, Nov. 2016.
[14] D. Meng, M. Du, and Y. Jia, “Interval bipartite consensus of networked
agents associated with signed digraphs,” IEEE Trans. Autom. Control.,
vol. 61, no. 12, pp. 3755-3770, Dec. 2016.
[15] Y. Wu, L. Liu, J. Hu, and G. Feng, “Adaptive antisynchronization of
multilayer reaction-diffusion neural networks,” IEEE Trans. on Neural
Networks and Learning Systems, vol. 29, no. 4, pp. 807-818, April 2018.
[16] J. Hu, and Y. Wu, “Interventional bipartite consensus on coopetition
networks with unknown dynamics,” J. Franklin Inst., vol. 354, no. 11, pp.
4438-4456, July 2017.
[17] J. Hu, Y. Wu, L. Liu, and G. Feng, “Adaptive bipartite consensus control
of high-order multi-agent systems on coopetition networks,” Int J. Robust
Nonlinear Control, vol. 28, no. 7, pp. 2868-2886, May 2018.
[18] F. A. Yagmaie, R. Su, F. L. Lewis, and S. Olaru, “Bipartite and
cooperative output synchronizations of linear heterogeneous agents: A
unified framework,” Automatica, vol. 80, no. 6, pp. 172-176, June 2017.

0018-9286 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.