Systems & Control Letters 66 (2014) 94–103

Contents lists available at ScienceDirect

Systems & Control Letters

journal homepage: www.elsevier.com/locate/sysconle

On the consensus and bipartite consensus in high-order multi-agent

dynamical systems with antagonistic interactions
Maria Elena Valcher a,∗ , Pradeep Misra b
a
Dipartimento di Ingegneria dell’Informazione, Università di Padova, via Gradenigo 6/B, 35131 Padova, Italy
b
Department of Electrical Engineering, Wright State University, Dayton, OH, United States

article info abstract

Article history: The aim of this paper is to address consensus and bipartite consensus for a group of homogeneous agents,
Received 7 June 2013 under the assumption that their mutual interactions can be described by a weighted, signed, connected
Received in revised form and structurally balanced communication graph. This amounts to assuming that the agents can be
13 October 2013
split into two antagonistic groups such that interactions between agents belonging to the same group
Accepted 8 January 2014
Available online 25 February 2014
are cooperative, and hence represented by nonnegative weights, while interactions between agents
belonging to opposite groups are antagonistic, and hence represented by nonpositive weights. In this
Keywords:
framework, bipartite consensus can always be reached under the stabilizability assumption on the state-
Consensus space model describing the dynamics of each agent. On the other hand, (nontrivial) standard consensus
Bipartite consensus may be achieved only under very demanding requirements, both on the Laplacian associated with the
Multi-agents communication graph and on the agents’ description. In particular, consensus may be achieved only if
Antagonistic interactions there is a sort of ‘‘equilibrium’’ between the two groups, both in terms of cardinality and in terms of the
Communication graph weights of the ‘‘conflicting interactions’’ amongst agents.
© 2014 Elsevier B.V. All rights reserved.

1. Introduction A common assumption in most of the literature about consen-

sus is that agents achieve this goal through collaboration. How-
ever, in a number of contexts where the consensus is meaningful,
Mathematical formulation of multi-agents systems and consen-
the interactions among agents are not necessarily cooperative. On
sus problems has been of interest for a considerable length of time.
the contrary, in contexts like markets or social networks [23], for
Some of the pioneering works are reported in [1–5] and references
instance, agents may also display non-cooperative or antagonistic
therein. However, a decade ago, thanks to milestone contributions
interactions with some of their neighboring agents. In social net-
such as [6–12], a wide stream of literature on these topics started
works, mutual relationships between pairs of individuals may be
and flourished. The driving force behind considerable research ac-
friendly or hostile, and this may create two antagonistic groups. In
tivity on this topic is the wide variety of areas where the consen-
economic systems, duopolistic regimes arise quite frequently: all
sus problem lends itself in a natural manner. In applications such
the companies producing a certain product, or providing a certain
as sensor networks, coordination of mobile robots or UAVs, flock-
service, are split into two competing cartels. But this is also the
ing and swarming in animal groups, dynamics of opinion forming,
case when modeling two competing teams, as it happens, for in-
etc., the problem can be formulated as that of a group of agents stance, in sport disciplines, or in robot competitions like RoboCup.
exchanging information with the objective of reaching a common Each individual or robot collects information regarding both the
decision, a consensus, by resorting to distributed algorithms that team mates and the antagonists, and processes this data in order
make use of the information that each agent collects from neigh- to take decisions (position, speed, behavior, elevation, . . . ) that are
boring agents (see, e.g. [13–19,11,20,21]). The interested reader is in agreement with those of their team mates. Game theory pro-
referred to [22,10] for a more complete list of references. vides several examples where players are divided into two compet-
ing teams, and antagonistic interactions between the two groups
are crucial when modeling the overall system dynamics. Finally,
∗
Corresponding author. Tel.: +39 049 827 7795; fax: +39 049 827 7614.
in biological systems, interactions between genes or chemical el-
E-mail addresses: [email protected] (M.E. Valcher), ements may be cooperative or antagonistic in the form of activa-
[email protected] (P. Misra). tors/inhibitors.
http://dx.doi.org/10.1016/j.sysconle.2014.01.006
0167-6911/© 2014 Elsevier B.V. All rights reserved.
 M.E. Valcher, P. Misra / Systems & Control Letters 66 (2014) 94–103 95

In graph theoretic terms, antagonistic interactions can be Notation. R+ is the semiring of nonnegative real numbers. For any
taken into account by replacing the standard communication pair of positive integers k and n with k ≤ n, [k, n] is the set of
graph, characterized by nonnegative weights, with a signed graph integers {k, k + 1, . . . , n}. The (i, j)th entry of a matrix A will be
[24,25] displaying both positive and negative weights. Positive denoted by [A]ij , the ith entry of a vector v by [v]i . A matrix (in
arcs correspond to cooperative interactions between agents, while particular, a vector) A with entries in R+ is called nonnegative,
negative arcs describe interactions between antagonistic agents. and denoted by A ≥ 0. The symbol 1N denotes the N-dimensional
While there is considerable literature about consensus in cooperat- vector with all entries equal to 1. The spectrum of a square matrix
ing multi-agent systems, research results pertaining to consensus A is denoted by σ (A).
without cooperation are relatively few [26–28]. In a recent paper
[22], Altafini developed the concept of bipartite consensus among
2. Consensus and bipartite consensus problems: statements
agents with antagonistic interactions. Specifically, based on defi-
nitions and properties of signed weighted (directed or undirected)
We consider a multi-agent system consisting of N agents, each
graphs and of the associated Laplacians (see [24,29]), he introduced
of them described by the same single-input continuous-time state-
the concept of bipartite consensus (or agreed dissensus). This is the
space model. Specifically, xi (t ), the ith agent state, i ∈ [1, N ],
situation when agents are split into two groups such that within
evolves according to the first-order differential equation
each group all the agents converge to a unique decision, but the
decisions of the two groups are opposite. ẋi (t ) = Axi (t ) + bui (t ), (1)
By addressing the classical example of homogeneous agents
modeled as simple scalar integrators, he proved that if the signed, where xi (t ) ∈ R , ui (t ) ∈ R, A ∈ R
n n×n n
, and b ∈ R . The communi-
weighted and connected communication graph describing the cation among the N agents is described by an undirected, weighted
agents’ interactions is structurally balanced, then the agents reach and signed, communication graph [22] G = (V , E , A), where V =
bipartite consensus. On the other hand, if the interactions are {1, 2, . . . , N } is the set of vertices, E ⊆ V × V is the set of arcs,
antagonistic, but not structurally balanced, the only agreement and A is the matrix of the signed weights of G. The (i, j)th entry
that can be achieved among the agents is the trivial one, where all of A, [A]ij , is nonzero if and only if the arc (j, i) belongs to E ,
the agents’ states converge to zero. namely information about the status of the jth agent is available
The main objective of this paper is to extend the results reported to the ith agent. We assume that the interactions between pairs
in [22], by addressing the general case of N homogeneous agents of agents are symmetric and hence A = A⊤ . The interaction
described by a generic n-dimensional linear state-space model. between the ith and the jth agents is cooperative if [A]ij > 0 and
This represents a natural extension of the case when agents are antagonistic if [A]ij < 0. Also, we assume that [A]ii = 0, for all
modeled as simple integrators. In certain situations, agents’ status i ∈ [1, N ]. The graph G is connected if, for every pair of vertices
may require more than a single variable for accurate representa- j and i, there is path, namely an ordered sequence of arcs
tion. These variables may include, e.g., position and velocity, price (j, i1 ), (i1 , i2 ), . . . , (ik−1 , ik ), (ik , i) ∈ E , connecting them.
and production levels, etc. These decision variables are updated The Laplacian matrix associated with the adjacency matrix A is
based on the information collected from the neighboring agents, defined as in [22,24,29], namely:
and consensus must be achieved on all of them. Specifically, we
establish conditions for consensus and bipartite consensus for a
L := C − A, (2)
group of N homogeneous agents under the assumption that their where C is the (diagonal) connectivity matrix, whose diagonal
mutual interactions can be described by a weighted, signed, con- entries are the sums of the absolute values of the corresponding
nected and structurally balanced communication graph. row entries of A, namely
We show that, in this set-up, bipartite consensus can always 
be reached under the fairly weak assumption of stabilizability [C ]ii = |[A]ij |, ∀i ∈ [1, N ].
on the state-space model describing the dynamics of each (j,i)∈E
agent. However, nontrivial consensus to a common decision for
Therefore
the two antagonistic groups can be achieved only under more 
restrictive requirements, both on the Laplacian associated with the  |[A]ij |, if i = j
communication graph and on the agents’ description. In particular, [L]ij = (j,i)∈E (3)
consensus may be achieved only if there is a sort of ‘‘equilibrium’’ −[A] , if i ̸= j.
ij
between the two groups, both in terms of cardinality and in terms
of the weights of the ‘‘conflicting interactions’’ among agents. Throughout the paper, we assume that the weighted and signed
Briefly, Section 2 introduces the problem formulation and for- graph G, describing the interactions among agents, is connected
malizes the consensus and bipartite consensus problems. In ad- and structurally balanced. This latter property means [22,24,25]
dition, basic definitions and results regarding weighted signed that the set of vertices V can be partitioned into two disjoint
graphs and their Laplacians are reviewed, and a new technical subsets V1 and V2 such that for every i, j ∈ Vp , p ∈ [1, 2], [A]ij ≥
result regarding the Laplacian of structurally balanced graphs is 0, while for every i ∈ Vp , j ∈ Vq , p, q ∈ [1, 2], p ̸= q, [A]ij ≤ 0.
presented. Section 3 investigates the bipartite consensus prob- This amounts to saying that the agents can be split into two groups,
lem, and it is shown there that, under the structural balance and interactions between pairs of agents belonging to the same
assumption, it is possible to extend to the case of antagonistic in- group are cooperative, while interactions between pairs of agents
teractions and bipartite consensus the results presented in [30,31] belonging to different groups are antagonistic. Therefore, after a
(see, also, [6,32,21,33]) for the consensus of high-order cooperating suitable reordering of the agents, we can always assume that
agents. Section 4 explores the consensus problem, by focusing on  
the case when the common trajectory that the agents converge to A11 A12
A= , (4)
is bounded, but not converging to zero. In this section, conditions A⊤
12 A22
under which consensus may be achieved are investigated, and an
11 ∈ R+ , A22 = A22 ∈
k×k
algorithm to design the control law that makes this possible is pre- where k := |V1 |, N − k := |V2 |, A11 = A⊤ ⊤
(N −k)×(N −k) k×(N −k)
sented. Finally, it is shown that when the previous conditions are R+ , while −A12 ∈ R+ . Under this fundamental
not met, nontrivial consensus can never be achieved. requirement on the mutual interactions among the agents, we
96 M.E. Valcher, P. Misra / Systems & Control Letters 66 (2014) 94–103

assume that the ith agent adopts the static state-feedback control Proposition 1 ([22]). Given a connected, weighted and signed graph
algorithm: G = (V , E , A), let L denote the Laplacian matrix associated with A.
 The following facts are equivalent:
ui (t ) = −K |[A]ij | · [xi (t ) − sign([A]ij )xj (t )],
(j,i)∈E 1. G is structurally balanced;
i ∈ [1, N ], (5) 2. A is equivalent, through a gauge transformation, to a nonnegative
matrix;
where K ∈ R1×n is a feedback matrix to be designed, and sign(·) is
3. λ1 (L) = 0.
the sign function.
Upon defining the state and input vectors Note that, when G is structurally balanced and connected, 0 is
⊤ a simple eigenvalue of L. Next we establish a key technical result
x(t ) := x1 (t )⊤
x2 (t )
⊤
xN (t )
⊤
,

··· that will be used in subsequent sections.
⊤
u(t ) := u1 (t ) u2 (t ) uN (t ) ,

···
Lemma 1. Given a connected, weighted and signed graph G = (V ,
the overall dynamics of the multi-agent system are henceforth E , A), with A having at least one negative entry, let L denote the
described by the equations Laplacian matrix associated with A. If G is structurally balanced, and
ẋ(t ) = (IN ⊗ A)x(t ) + (IN ⊗ b)u(t ), (6) V1 and V2 denote the two disjoint sets of vertices that partition the
set V , then
u(t ) = −(L ⊗ K )x(t ). (7)
[L]ij , i ∈ [1, N ], are nonzero,
N
(i) at least two of the row sums
The aim of this paper is to investigate both the standard (state) j =1
and hence positive;
consensus and the bipartite (state) consensus problems for the
(ii) if all row sums j=1 [L]ij , i ∈ [1, N ], are identical, then the two
N
group of agents described by (6)–(7). Specifically, we will first
search for conditions ensuring that bipartite consensus is possible, sets of vertices V1 and V2 have the same cardinality.
namely that a matrix K can be found so that, for every choice of the
initial conditions, the agents’ states evolve as follows: Proof. We first observe that the signed weight matrix A can
be expressed as A = A+ + A− , where A+ and A− are the
ζ (t ),

∀ i ∈ V1 ; matrices that represent the positive and the negative entries of
lim xi (t ) =
t →+∞ −ζ (t ), ∀ i ∈ V2 , A, respectively. Accordingly, if we denote by L+ and L− the
Laplacians corresponding to the matrices A+ and A− , respectively,
for some function ζ (t ), t ∈ R+ , that depends on the initial condi- then L = L+ + L− . Since all row sums of L+ are zero,
tions. Subsequently, we will search for conditions ensuring that all the row sums of L ordinately coincide with the row sums of
the agents converge to the same decision (so, that standard consen- the nonnegative matrix L− . Moreover, L− 1N = −2A− 1N . By
sus among agents is possible), and hence a matrix K can be found
assumption, at least one element [A]ik is negative (and hence, by
such that, for every choice of the initial conditions,
symmetry, [A]ki < 0), therefore condition (i) holds (for the ith and
lim xi (t ) = ζ (t ), ∀ i ∈ V, kth rows).
t →+∞
To prove (ii), we need to show that if all the row sums of A−
for some function ζ (t ), t ∈ R+ , that depends on the initial condi- take the same value, say α < 0, then |V1 | = |V2 |. By the structural
tions. balance assumption, the matrix A− can be reduced, by means of
Before proceeding, it is convenient to recall from [22] some permutations, to the following form
fundamental results about the Laplacian of a weighted, signed,  
0 A12
undirected graph G, in general, and about the Laplacian of a A− = ⊤ , −A12 ∈ Rk+×(N −k) , A12 ̸= 0.
A12 0
structurally balanced graph, in particular. First of all, the Laplacian
matrix L, defined as in (2), is a symmetric and positive semidefinite Next, assume that k = |V1 | ̸= |V2 | = N − k. If both conditions
matrix, whose nonnegative real eigenvalues can always be sorted
A12 1N −k = α 1k ,

in such a way that
k A12 = α 1N −k ,
1⊤ ⊤
0 ≤ λ1 (L) ≤ λ2 (L) ≤ · · · ≤ λN (L).
were true, the fact that α ̸= 0 would lead to the contradiction
Note that, L being symmetric and hence diagonalizable, there
exists a basis of RN consisting of eigenvectors of L. Clearly, one can α(N − k) = 1⊤
k A12 1N −k ̸= 1k (A12 1N −k ) = α k.
⊤
 
always find (at least) one (left or right) eigenvector v of L whose
Hence, k = N − k.
N 
entries sum up to a nonzero value (i.e., i=1 [v]i ̸= 0).1 When so,
N
we can assume without loss of generality that i=1 [v]i = 1.
Remark 1. It is worthwhile noticing that, in the special case when
We also recall the concept of gauge transformation [22] (also
known as change of orthant order). We denote by D the set of all all row sums of L take the same value, say λ̂ ∈ R, namely L1N =
N × N diagonal matrices whose diagonal entries are either 1 or −1. λ̂1N , then there is a special relationship between the Laplacian L
Two N × N matrices A1 and A2 are said to be equivalent through a and the unsigned Laplacian Lu of A, this latter defined as
gauge transformation, if there exists D ∈ D such that A2 = DA1 D. 
In the sequel, we will make use of the following fundamental result.
 [A]ij , if i = j
[Lu ]ij = (j,i)∈E
−[A] , if i ̸= j.
ij

1 Indeed, condition N [v] = 0 is equivalent to v⊤ 1 = 0. So, if every eigen- Indeed, by following the same lines as in the proof of Lemma 1, it
i=1 i N
vector v would be orthogonal to 1N , the vector space generated by all the eigen- is easily seen that
⊥
vectors would be a vector subspace of ⟨1N ⟩ , and hence the eigenvectors could not
generate RN , a contradiction. Lu = L − 2C− ,
 M.E. Valcher, P. Misra / Systems & Control Letters 66 (2014) 94–103 97

where C− is the (diagonal) connectivity matrix of A− . On the other and consider the coordinate transformation
hand, as Lu 1N = 0, condition L1N = λ̂1N is equivalent to  ζ (t ) 

λ̂  δ2 (t ) 
..
C− 1N = 1N ,
 
2
 . 
 δk (t ) 
 
 δk+1 
 
and since C− is diagonal, this means C− = λ̂2 I. As a result, 
.

 .. 
Lu = L − λ̂I , δN (t )
 
p1 In p2 In · · · pk In pk+1 In · · · pN In
which ensures that
:=  − 1k−1 ⊗ In In(k−1) 0 
σ (L) = (λ1 (L), λ2 (L), . . . , λN (L)) − 1N −k ⊗ In 0 −In(N −k)
 x1 (t ) 
= (µ1 + λ̂, µ2 + λ̂, . . . , µN + λ̂),
 x2 (t ) 
..
where (µ1 , µ2 , . . . , µN ) is the spectrum of Lu .
 
 . 
( )
 
×  kx t 
 xk+1 

3. Bipartite consensus ..
 
.
 
xN (t )
As a first step we want to derive necessary and sufficient condi-
= (S ⊗ In )x(t ).
tions for bipartite consensus of non-cooperating multi-agent sys-
tems described as in (6), under the assumption that the feedback Note that δi (t ) = xi (t ) − x1 (t ), for i ∈ [2, k], while δi (t ) =
control algorithm is described as in (7) and the communication −xi (t ) − x1 (t ), for i ∈ [k + 1, N ]. So, convergence to zero of
graph G is structurally balanced. all δi ’s corresponds to bipartite consensus. It is a matter of simple
It will be shown that the results regarding consensus of coop- computation to see that
erating multi-agent systems described by (6), first reported in [31]  
(see, also, [30,32] for additional information), can be extended to −1 0 0
S LS =
the case of non-cooperating agents and bipartite consensus under 0 L2
the assumption that the control algorithm is described as in (7), and
provided that the communication graph G is structurally balanced. and hence the N − 1 eigenvalues of L2 are 0 < λ2 (L) ≤ · · · ≤
λN (L). Also, upon setting
Theorem 1. Consider the multi-agent system (6) with control û1 (t )
 
algorithm (7), where K ∈ R1×N is a matrix to be determined and the  û2 (t ) 
communication graph G is connected and structurally balanced. The û(t ) :=  ..   = Su(t ),
agents asymptotically reach bipartite consensus if and only if the N − 1 .

polynomials ûN (t )

ψi (s) := det(sIn − A) + λi (L) · [K adj(sIn − A)b] the multi-agent system dynamics are described, with respect to the
= det(sIn − A + λi (L)b, K ), i ∈ [2, N ], (8) new state and input coordinates, by

where adj(sIn − A) denotes the adjoint matrix of sIn − A, are Hurwitz. ζ̇ (t ) = Aζ (t ) + bû1 (t ), (9)
δ˙2 (t ) δ2 (t ) û2 (t )
     
Proof. As G is structurally balanced, we can assume without any
loss of generality that V1 = [1, k] and V2 = [k + 1, N ], and hence  ..   .   . 
 .  = (IN −1 ⊗ A)  ..  + (IN −1 ⊗ b)  ..  , (10)
(N −k)×(N −k)
A is described as in (4), where A11 ∈ Rk+×k , A22 ∈ R+ , δN (t ) ûN (t )
k×N
δ˙N (t )
while −A12 ∈ R+ . It is easily seen that the gauge transformation
  û1 (t ) = 0, (11)
I 0
D= k û2 (t ) δ2 ( t )
   
∈D
0 −I N − k  ..   .. 
  .  = −(L2 ⊗ K )  .  . (12)
is such that DAD is a nonnegative matrix. Let p := p1 p2 · · · ûN (t ) δN (t )
⊤
pN ∈ RN be a left eigenvector of L corresponding to λ1 (L) = 0.
Since Dp is the left eigenvector of the DLD, corresponding to the These equations can be equivalently rewritten as
zero eigenvalue, and DLD is the Laplacian matrix associated with
the nonnegative matrix DAD, we can assume ζ̇ (t ) = Aζ (t ), (13)
δ˙2 (t ) δ2 (t )
    
1
 ,
 i ∈ [1, k];  ..   . 
pi = N  .  = [(IN −1 ⊗ A) − (IN −1 ⊗ b)(L2 ⊗ K )]  ..  . (14)
1
− ,
 i ∈ [k + 1, N ]. δ˙N (t ) δN (t )
N
Introduce the nonsingular matrix Due to the way the variables δi (t )’s are defined, it is clear
  that bipartite consensus is achieved if and only if the system
p1 p2 ··· pk pk+1 ··· pN (14) is asymptotically stable. This is the case if and only if the
S :=  −1k−1 Ik−1 0  characteristic polynomial of [(IN −1 ⊗ A) + (IN −1 ⊗ b)(L2 ⊗ K )] is
−1N −k 0 −I N − k Hurwitz. By following the same reasoning adopted in [31], it can
98 M.E. Valcher, P. Misra / Systems & Control Letters 66 (2014) 94–103

be shown that Proof. By the previous theorem, we have to show that the sta-
bilizability of the pair (A, b) is a necessary and sufficient condi-
det sIn(N −1) − [(IN −1 ⊗ A) − (IN −1 ⊗ b)(L2 ⊗ K )]
 
tion for the existence of a matrix K such that all the polynomials
N
 ψi (s), i ∈ [2, N ], in (8) are Hurwitz.
= (det(sIn − A) + λi (L) · [K adj(sIn − A)b]) , (Necessity) If the pair (A, b) is not stabilizable, there exists a
i=2 similarity transformation T such that
and hence bipartite consensus is achieved if and only if all the 
A11 A12
  
b1
polynomials ψi (s), i ∈ [2, N ], are Hurwitz.  TAT −1 = , Tb = , (15)
0 A22 0
Remark 2. By referring to the notation adopted within the proof, where the pair (A11 , b1 ) ∈ Rr ×r × Rr is controllable, and the (n − r )-
it is easy to see that, when the polynomials ψi (s), i ∈ [2, N ], in (8) dimensional matrix A22 is not Hurwitz. Consequently, for every
are Hurwitz, and hence the new state variables δi (t ), i ∈ [2, N ], choice of K both det(sIn − A) and K adj(sIn − A)b are multiple of
converge to zero, then det(sIn−r − A22 ). Therefore all ψi (s) are multiple of det(sIn−r − A22 )
N
and hence are not Hurwitz.
Conversely, assume that the pair (A, b) is stabilizable, and hence

ζ (t ) = pi xi (t )
there exists a similarity transformation T such that TAT −1 and Tb
i=1
are described as in (15), with det(sIn−r − A22 ) Hurwitz and
k
 N

= pi (x1 (t ) + δi (t )) + pi (−x1 (t ) − δi (t )) 
0 1 0 ··· 0

0
 
i=1 i=k+1  0 0 1 ··· 0  0
. .. .. .. ..  .
   
 ..  ..  . (16)
k N k N
. ,

A11 =  . . . b1 = 
   
pi x1 (t ) + pi δi (t ) − pi δi ( t )
 
= pi −
i=1 i=k+1 i =1 i=k+1
 0 0 0 ··· 1
 0

k N
 −a 0 −a 1 −a 2 ··· −ar −1 1
 
= x1 (t ) + pi δi (t ) − pi δi (t ) If we accordingly partition K as K =

K1

K2 , with K1 =
i=1 i=k+1  
k0 k1 · · · kr −1 , then
tends to x1 (t ), as t tends to +∞, and hence describes the behavior
det(sIn − A) = det(sIr − A11 ) det(sIn−r − A22 )
of the first group of agents, while −ζ (t ) describes the behavior of
the second group. As for standard consensus, the evolution of ζ (t ) K adj(sIn − A)b = [kr −1 sr −1 + kr −2 sr −2
depends on the initial conditions and on the properties of the state + · · · + k1 s + k0 ] det(sIn−r − A22 ).
matrix A, in the agents’ description. Indeed, if A is Hurwitz, the
Clearly if det(sIr − A11 ) is also Hurwitz, we can trivially make all
only bipartite consensus we achieve is the trivial one to the zero
the polynomials ψi (s), i ∈ [2, N ], Hurwitz by choosing K = 0. On
trajectory, while if A is simply stable or not stable, corresponding to
the other hand, assume that
certain initial conditions, ζ (t ) may be bounded or even diverging.
det(sIr − A11 ) = [sr + ar −1 sr −1 + · · · + a1 s + a0 ] = d− (s)d0+ (s),
Remark 3. The Hurwitz property of the polynomials ψi (s), i ∈
where d0+ (s) ∈ R[s] is a monic polynomial, with degree ℓ ≥ 1,
[2, N ], is the same condition derived in [31] (see also [30]) for
whose roots have nonnegative real part, while d− (s) is monic, Hur-
the solvability of the standard consensus problem, in case of
witz and has degree r − ℓ. Since all ki ’s are arbitrary, we can choose
cooperation. So, the assumption on the communication graph does
them in such a way that kr −1 sr −1 + kr −2 sr −2 + · · · + k1 s + k0 is a
not change the solvability condition but only the outcome.
multiple of d− (s). In this way
Next, we show that the stabilizability of the pair (A, b) is
kr −1 sr −1 + kr −2 sr −2 + · · · + k1 s + k0
a necessary and sufficient condition for the solvability of the
bipartite consensus problem. The result is analogous to the one = d− (s)[βr −ℓ−1 sr −ℓ−1 + · · · + β1 s + β0 ],
obtained for the standard consensus problem, since it reduces to where βi , i = 0, . . . , r − ℓ − 1, are arbitrary. To prove that for
the problem of imposing that the real polynomials ψi (s), i ∈ [2, N ], a suitable choice of the matrix K , all ψi (s), i ∈ [2, N ], can be-
are all Hurwitz, a result obtained in [30] (see also [6]). Specifically, come Hurwitz, is equivalent to show that it is possible to choose
upon defining the polynomial β0 , β1 , . . . , βr −ℓ−1 ∈ R in such a way that the N − 1 polynomials
ψ(s, λ) := det(sIn − A) + λK adj(sIn − A)b, λ ∈ R, pi (s) := d0+ (s) + λi (L)[βr −ℓ−1 sr −ℓ−1 + · · · + β1 s + β0 ],
our goal is to show that, under the stabilizability assumption on i ∈ [2, N ],
(A, b), it is always possible to choose the polynomial K adj(sIn − A)b,
of degree smaller than n = deg(det(sIn − A)), in such a way are Hurwitz. Introduce any monic Hurwitz polynomial β̂(s) ∈ R[s]
that ψ(s, λ) is Hurwitz for every λ belonging to some half-line of degree r − ℓ − 1. By referring to the positive root locus of the
β̂(s)
[λ̂, +∞) ( [0, +∞) including all the positive eigenvalues of L. function d0+ (s)
, namely to the set of zeros of the expression
For completeness, we provide a constructive proof of the above
statement. The proof is completely independent from the one d0+ (s) + γ β̂(s), γ ∈ R+ ,
given in [30] (based on the work of Tuna [33]) and is essential
it is clear that all the root locus branches start from the r − ℓ unsta-
in demonstrating why the analogous algebraic conditions for
ble zeros of d0+ (s) and end in the open left half plane: r − ℓ − 1 end
reaching standard consensus are more difficult to achieve.
at the finite zeros of β̂(s), while the last branch goes to −∞ along
Theorem 2. Consider the multi-agent system (6) with control the negative real halfline. This proves that there exists γ̄ > 0 such
algorithm (7), and suppose that the communication graph G is that for every γ ≥ γ̄ the polynomial d0+ (s) + γ β̂(s) is Hurwitz.
connected and structurally balanced. There exists K ∈ R1×N such that But then, by choosing
agents asymptotically reach bipartite consensus if and only if the pair γ̄
(A, b) is stabilizable. βr −ℓ−1 sr −ℓ−1 + · · · + β1 s + β0 = β̂(s),
λ2 (L)
 M.E. Valcher, P. Misra / Systems & Control Letters 66 (2014) 94–103 99

we can ensure that all polynomials pi (s), i ∈ [2, N ], and hence may lead to a common decision for some suitable choice of K . This
all polynomials ψi (s), i ∈ [2, N ], are Hurwitz. This proves that if is equivalent to investigating whether it is possible to impose that
(A, b) is stabilizable, then some matrix K can be found such that all for every index i ∈ [1, N ] we have limt →+∞ xi (t ) = ζ (t ), for some
ψi (s), i ∈ [2, N ], are Hurwitz.  function ζ (t ). In the following, to avoid considering either trivial or
undesirable cases,2 we will search for conditions ensuring that the
Example 1. Consider the system of N = 3 agents, each of them state space evolution of the agents is always bounded, but not al-
described by the 3-dimensional (n = 3) (controllable, and hence ways converging to zero. In this way, we rule out the case when all
stabilizable) state-space model the agents dynamics converge to zero, independently of the initial
0

1 0

0
  conditions, but also the case when all agents may have diverging
ẋi (t ) = Axi (t ) + bui (t ) = 0 0 1 xi (t ) + 0 ui (t ), trajectories.
1 −3 3 1 To investigate this problem, we need to distinguish between the
following two cases: (A) the case when all row sums of L take the
i ∈ [1, 3].
same value (if so, we know from Lemma 1 that the two sets of the
partition, V1 and V2 , have the same cardinality, say k = N /2); and
⊤
Note that det(sI3 − A) = (s − 1)3 and adj(sI3 − A)b = 1 s s2 .

We assume that the signed and weighted communication matrix (B) the case when not all row sums of L are the same.
A is

0 −1 −2
 4.1. All row sums of L take the same value
A= −1 0 3 ,
−2 3 0 Suppose that ∃ λ̂ ∈ R such that L1N = λ̂1N . Clearly, by
Lemma 1, λ̂ > 0 and obviously λ̂ ∈ σ (L). The vector p :=
which means that the corresponding graph is structurally balanced ⊤
1/N 1/N 1/N

with V1 = {1} and V2 = {2, 3}. The associated Laplacian is ··· ∈ RN is hence a left eigenvector of L
corresponding to λ̂. We introduce the nonsingular matrix
3 1 2
 
L= 1 4 −3 , 1/N 1/N 1/N
 
···
2 −3 5 S := .
−1N −1 IN − 1
√
its eigenvalues are λ1 (L) = 0, λ2 (L) = 6 −
and √ 3, λ3 (L) = Consider the coordinate transformation
6 + 3. The matrix
ζ (t )
 
1/3 −1/3 −1/3  δ2 ( t ) 
 
(1/N )In (1/N )In · · · (1/N )In
 
S= −1 −1 0  .  :=
 .  − 1 ⊗ I In(N −1)
−1 0 −1 . N − 1 n

δN (t )
is such that
x1 (t )
 
0 0 0
   
0  x (t ) 
S LS −1
= = 0 5 −1 .  2
×  .  = (S ⊗ In )x(t ).

L2
0 −2 7  .. 
Introduce the state and input transformation: xN (t )

ζ (t ) It is a matter of simple computation to see that in this case, due to

 
δ2 (t ) = (S ⊗ I3 )x(t ), û(t ) = Su(t ). the fact that j=1 [L]ij = λ̂, ∀ i ∈ [1, N ],
N
δ3 (t )  
S LS −1 =
λ̂ 0
,
The state dynamics with respect to the new basis and the new input
0 L2
are described as in (13)–(14) for N = 3. As det(sI3 − A) = (s − 1)3
has all roots in the closed right half-plane, we assume β̂(s) = for some matrix L2 , whose eigenvalues are the N − 1 (possibly
(s + 1)2 = s2 + 2s + 1. It is easily seen (for instance, through not distinct) eigenvalues of L, obtained by removing λ̂ from the
Routh criterion), that N-tuple (λ1 (L), λ2 (L), . . . , λN (L)). Note that there is no guaran-
tee that λ̂ is a simple eigenvalue of L and hence λ̂ can belong also
det(sI3 − A) + γ β̂(s) = (s − 1)3 + γ (s + 1)2 ,
√ to σ (L2 ). By following the same steps as in the proof of Theorem 1
(but in this case δi (t ) = xi (t ) − x1 (t ) for all i ∈ [2, N ]), we can
√ for every γ > 1 √
is Hurwitz + 5. Choose, then, for simplicity,
γ̄ = 6 − 3 = λ2 (L) > 1 + 5 and K adj(sI2 − A)b = β̂(s). Then obtain the following description:

ψi (s) = det(sI3 − A) + λi (L)β̂(s) = (s − 1)3 + λi (L)(s + 1)2 , ζ̇ (t ) = (A − λ̂bK )ζ (t ),

δ˙2 (t ) δ2 (t )
   
i = 2, 3,
 ..   . 
are two Hurwitz polynomials. The choice K adj(sI2 −  .  = [(IN −1 ⊗ A) − (IN −1 ⊗ b)(L2 ⊗ K )]  ..  .
 A)b = β̂(s) =
s2 + 2s + 1 corresponds to setting K = 1 2 1 , and it is just δ˙N (t ) δN ( t )
a matter of straightforward computation to verify that the matrix
[(I2 ⊗ A) − (I2 ⊗ b)(L2 ⊗ K )] is Hurwitz. Therefore, consensus to some common bounded trajectory ζ (t ) is
achieved if and only if the following conditions hold:
4. Consensus

In this section, we investigate whether for multi-agent systems 2 In addition, notice that the trivial case, when all the agents’ states asymptoti-
described in (6), with a weighted, signed, connected and struc- cally converge to zero, has already been regarded as a special case of the bipartite
turally balanced communication graph G, the control algorithm (7) consensus case.
100 M.E. Valcher, P. Misra / Systems & Control Letters 66 (2014) 94–103

(i) λ̂ is a simple eigenvalue of L; and its eigenvalues are λ1 (L) = 0, λ2 (L) = λ3 (L) = λ4 (L) =
(ii) all polynomials ψi (s) = det(sIn − A) + λi (L)[K adj(sIn − 4. Moreover all row sums are equal to 4. In this case it is clear
A)b], λi ∈ σ (L), λi ̸= λ̂, are Hurwitz; that, independently of the agents’ state-space description (A, b),
(iii) A − λ̂bK is simply (but not asymptotically) stable. no consensus can be reached except possibly for the one that leads
all the agents to the zero state, since condition (i) is not satisfied.
Remark 4. It is worth noticing that while the conditions for bipar-
At this point the following natural question arises: assuming
tite consensus given in Theorem 2 pertain only to the properties
of the agents’ description, the possibility of reaching a consensus that condition (i) holds, namely that λ̂ > 0 is a simple eigenvalue of
among agents with antagonistic interactions depends not only on L, under what conditions some feedback matrix K can be found so
the model adopted for the agents, but also on the algebraic prop- that conditions (ii) and (iii) hold? It is worthwhile noticing that this
erties of the associated Laplacian. problem is more complicated than the analogous one we addressed
for the bipartite consensus in Section 3, because in this case, to
We illustrate the above distinction by means of two examples. satisfy (ii) and (iii), a matrix K must be found such that ψ(s, λ)
is an Hurwitz polynomial except on a (possibly infinite) interval
Example 2. Consider the system of four agents (N = 4), each of
having λ̂ on one of its extremes. We need to address the following
them described by the 3-dimensional (n = 3) state-space model
two situations:
ẋi (t ) = Axi (t ) + bui (t )
1. If λ̂ = λN (L), namely λ̂ is the largest of the eigenvalues of L,
0 1 0 0
   
we have to find K such that ψ(s, λ) is Hurwitz in [0, λ̂)3 and it
= 0 0 1 xi (t ) + 0 ui (t ), i ∈ [1, 4]. has a simple zero at 0 (or a pair of simple imaginary conjugate
−1 −3 −3 1
zeros) for λ = λ̂.
Note that det(sI3 − A) = (s + 1)3 . We assume that the signed 2. If, on the other hand, λ̂ is not the largest eigenvalue of L, we
communication matrix A is have to find K such that4 ψ(s, λ) is Hurwitz in [0, λ̂) ∪ (λ̂ +
∆, +∞), and is not Hurwitz in [λ̂, λ̂ + ∆], where ∆ > 0 is
 
0 2 −1 −1
2 0 −1 −1  chosen in such a way that no eigenvalues of L lie in [λ̂, λ̂ + ∆]
A= ,
−1 −1 0 3  except for λ̂.
−1 −1 3 0
It is clear that the latter case imposes strong constraints on the
which means that the corresponding graph is structurally balanced polynomial ψ(s, λ) that, in general, cannot be achieved for an ar-
with two antagonistic groups of agents of the same cardinality bitrary choice of L, and hence for an arbitrary choice of the eigen-
V1 = {1, 2} and V2 = {3, 4}. The associated Laplacian is values λi ∈ σ (L), unless the degree of ψ(s, λ) is sufficiently large.
  More precisely, as we know that the characteristic polynomial of
4 −2 1 1
the non-controllable part of the pair (A, b) is a common divisor of
−2 4 1 1 
L= , det(sIn − A) and K adj(sIn − A)b for every choice of K , and hence
1 1 5 −3
also a divisor of ψ(s, λ) for every λ ∈ R+ , in order to be able to
1 1 −3 5
meet the previous constraints in any possible situation, we have
and its eigenvalues are λ1 (L) = 0, λ2 (L) = 4, λ3 (L) = 6, to assure that the part of the polynomial ψ(s, λ) we can actually
λ4 (L) = 8. We note that L14 = 414 , and hence all row sums are ‘‘shape’’ according to the needs has at least degree 3. The preced-
the same and they are equal to λ̂ = 4. Assume
 that the
 agents adopt ing discussion leads to the following result.
the feedback control law (7), with K = 12 1 1 . The consensus
Proposition 2. Consider the multi-agent system (6) with control
is possible if and only if conditions (i)–(iii) above hold. Clearly, λ̂ =
algorithm (7). Suppose that the communication graph G is connected
4 is a simple eigenvalue of L, and hence condition (i) holds. If we
and structurally balanced with respect to the partition in two subsets
consider the polynomial ψ(s, λ) = det(sIn −A)+λ[K adj(sIn −A)b],
as λ varies over the nonnegative real numbers, it is easily seen,
V1 and V2 of equal cardinality. Suppose, also, that the row sums of L
by resorting to the Routh criterion, for instance, that ψ(s, λ) is all equal to λ̂ > 0, and λ̂ is a simple eigenvalue of L. If the following
Hurwitz for every λ ∈ [0, 2) ∪ (4, +∞]. This implies, in particular, conditions hold true:
that ψ(s, λ) is Hurwitz for λ ∈ {0, 6, 8}. So, condition (ii) holds. (a) A is Hurwitz;
Finally, for λ = λ̂ = 4, ψ(s, λ) = (s + 7)(s2 + 7) which proves that (b) the dimension of the controllable subspace of the pair (A, b) is at
A−λ̂bK is simply (but not asymptotically) stable. So, for certain sets least 3;
of initial conditions, there is convergence of the state trajectories then there exists K ∈ R1×N such that the N agents asymptotically
of all agents to a sinusoidal trajectory. reach consensus.
Example 3. Consider a system of N = 4 agents, with signed Proof. Choose any ∆ > 0 such that no eigenvalue of L lie in
communication matrix [λ̂, λ̂ + ∆], except for λ̂ itself. By assumptions (a) and (b),
and following the same process and notation as in the proof of
 
0 1 −1 −1
1 0 −1 −1 Theorem 2, we can factorize the Hurwitz polynomial det(sIn − A)
A= , as det(sIn − A) = d− (s)d(s), where both d− (s) and d(s) are Hurwitz
−1 −1 0 1 
−1 −1 1 0 monic polynomials, the characteristic polynomial det(sIn−r − A22 )

which ensures that the corresponding graph is structurally

balanced with V1 = {1, 2} and V2 = {3, 4}. The associated
3 Recall that 0 is always an eigenvalue of L. Note, also, that this is not the
Laplacian is
only possible solution, as we could introduce more complicated conditions on the
behavior of ψ(s, λ), as λ varies in R+ , but it is the simplest.
 
3 −1 1 1
−1 3 −1 −1 4 Also in this case the algebraic conditions imposed on ψ(s, λ) could be slightly
L= ,
1 1 3 −1 different. For instance we could replace the interval [λ̂, λ̂ + ∆] with [λ̂ − ∆, λ̂], but
1 1 −1 3 this would not affect the solvability conditions.
 M.E. Valcher, P. Misra / Systems & Control Letters 66 (2014) 94–103 101

of the non-controllable part of (A, b) (see the decomposition (15)) signed communication matrix A is
is a factor of d− (s), while deg d(s) = 3.  
0 1 −2 −2
Since for every choice of a polynomial p(s) of degree n−1, which
1 0 −2 −2 
is a multiple of det(sIn−r − A22 ), we can always find a matrix K such A= ,
−2 −2 0 1 
that K adj(sIn − A)b = p(s), we want to show that by choosing
−2 −2 1 0
p(s) = d− (s)n(s) for a suitable second order Hurwitz polynomial
n(s), we can always ensure that which means that the corresponding graph is structurally balanced
with two antagonistic groups of agents of the same cardinality
ψ(s, λ) = d− (s)[d(s) + λn(s)] V1 = {1, 2} and V2 = {3, 4}. The associated Laplacian is
is Hurwitz for λ ∈ [0, λ̂)∪(λ̂+ ∆, +∞), and is not Hurwitz for λ ∈
 
5 −1 2 2
[λ̂, λ̂ + ∆]. Clearly, this amounts to showing that a second order −1 5 2 2 
L= ,
Hurwitz polynomial n(s) can be found such that the third order 2 2 5 −1
monic polynomial d(s) + λn(s) satisfies the previous constraints. 2 2 −1 5
Assume, w.l.o.g., d(s) = s3 + d2 s2 + d1 s + d0 , and note that,
and its eigenvalues are λ1 (L) = 0, λ2 (L) = 6, λ3 (L) = 6,
by the Routh criterion, the Hurwitz property of this polynomial is
λ4 (L) = 8. We note that L14 = 814 , and hence all row sums
equivalent to the following conditions on its coefficients:
are the same and they are equal to λ̂ = 8 = λ4 (L). In this case we
d0 , d1 , d2 > 0 and d1 d2 − d0 > 0. (17) have to choose K in such a way that K adj(sI2 − A)b = (s + 1)n0
makes the polynomial ψ(s, λ) = (s + 1)[s + 1 + λn0 ] Hurwitz for
We want to prove that, by choosing
λ < 8 and simply stable for λ = λ̂ = 8. To this end, it is sufficient
to choose K = n0 = −1/8 and the result is obtained.
 
d1 d2 − d0 (d2 + 2λ̂ + ∆)(d1 d2 − d0 )
n(s) = s + 2
s+ d1 + ,
λ̂(λ̂ + ∆) λ̂(λ̂ + ∆) 4.2. Not all row sums of L take the same value

we can ensure that d(s) + λn(s) is Hurwitz for λ ∈ [0, λ̂) ∪ (λ̂ + We first observe that a necessary condition for the overall
∆, +∞), and is not Hurwitz for λ ∈ [λ̂, λ̂ + ∆]. Note that by the state x(t ) to converge to some bounded trajectory ztot (t ) :=
assumptions on λ̂, ∆ and (17), the second order polynomial n(s) is
⊤
ζ (t )⊤ ζ (t )⊤· · · ζ (t )⊤ is that there exist a vector vtot :=

necessarily Hurwitz, as all its coefficients are positive. By applying ⊤
v⊤ · · · v⊤ , v ̸= 0, and some α ∈ C, Re(α) = 0, such
 ⊤
v
the Routh criterion to d(s) + λn(s) we obtain
that [(IN ⊗ A) − (IN ⊗ b)(L ⊗ K )] vtot = α vtot . This condition can
d1 + λ
d1 d2 −d0
3 1 be easily seen to be equivalent to the set of conditions:
λ̂(λ̂+∆) 
(d2 +2λ̂+∆)(d1 d2 −d0 )
d2 + λ d0 + λ d1 +  
2 N
λ̂(λ̂+∆) 
(λ−λ̂)(λ−λ̂−∆) A− [L]1j bK v = α v
d1 d2 −d0

1 · 0


λ̂(λ̂+∆)  d2 +λ 
 j =1

(d +2λ̂+∆)(d1 d2 −d0 )
d0 + λ

0 d1 + 2 0

λ̂(λ̂+∆)

 

  N
[L]2j bK v = α v

 A−
So, it is easy to see that the coefficients in the first column of the
(18)
table are all positive for λ ∈ [0, λ̂) ∪ (λ̂ + ∆, +∞), and the
j =1
..


coefficient in row ‘‘1’’ annihilates for λ = λ̂ and λ = λ̂ + ∆, and is .



negative in (λ̂, λ̂ + ∆). 

 

  N
[L]Nj bK v = α v.

 A−


Remark 5. From the previous proof it is clear that if the degree j =1
of d(s) were lower than 3, a polynomial n(s) of degree smaller If not all row sums of L take the same value, the above set of
than deg d(s) such that d(s) + λn(s) is Hurwitz for λ ∈ [0, λ̂) ∪ conditions is satisfied if and only if Av = α v and K v = 0. But
(λ̂ + ∆, +∞), and is not Hurwitz for λ ∈ [λ̂, λ̂ + ∆] would this implies that (A − λbK )v = α v for every λ ∈ R, and hence that
not necessarily exist. It is also clear that when λ̂ is not the largest ψ(s, λ) = det(sIn − A) + λK adj(sIn − A)b has a zero in α for every
choice of λ.
eigenvalue of L, ψ(s, λ̂) has necessarily two imaginary conjugate  ⊤
Now we select a left eigenvector p := p1 p2 · · · pN ∈
zeros (and not a zero in the origin), and hence the consensus is
achieved on a periodic trajectory. On the other hand, when λ̂ = RN of L corresponding to any eigenvalue λ̂ ∈ σ (L) and endowed
N
λn (L), the problem of finding n(s) of degree smaller than deg d(s) with the property that j=1 pj ̸= 0. As noted earlier, such an
N
such that d(s) + λn(s) is Hurwitz for λ ∈ [0, λ̂), and is not Hurwitz eigenvector always exists, and we can always assume j =1 pj = 1.
for λ ∈ [λ̂, +∞), can be trivially solved even for deg d(s) = 1, and Introduce the matrix
 
hence with a controllable subspace of dimension 1. This result is p1 p2 · · · pN
quite immediate and hence we omit the proof, but to better clarify
S := ,
−1N −1 IN − 1
it we provide the following example.
and consider the coordinate transformation
ζ (t )
 
Example 4. Consider the system of N = 4 agents, each of them
 δ2 ( t ) 
 
described by the 2-dimensional (n = 2) state-space model p I p I
1 n ··· 2 n pN In
 ..   := −1N −1 ⊗ In In(N −1)
.
    
−1 1 1
ẋi (t ) = Axi (t ) + bui (t ) = x (t ) + u (t ), δN (t )
0 −1 i 0 i
x1 (t )
 
i ∈ [1, 4].
 x (t ) 
Note that A is Hurwitz and the pair (A, b) has 1-dimensional  2
× .  = (S ⊗ In )x(t ).

controllable subspace. The characteristic polynomial of the non-  .. 
controllable part is det(sI − A22 ) = (s + 1). We assume that the xN (t )
102 M.E. Valcher, P. Misra / Systems & Control Letters 66 (2014) 94–103

It is a matter of simple computation to see that in this case to the ones derived for standard consensus in case of cooperative
interactions [30].
λ̂
 
0 In the second part of the paper we have investigated the
N N
 j =1 [ L ]2j − j=1 [L]1j
 possibility of achieving consensus to a common bounded trajectory
S L S −1 = ..
 
L2 
 even in case of antagonistic interactions, provided that there is
 . structural balance in the agents’ communications. In this case
N N
j =1 [ L] Nj − j=1 [L]1j conditions allowing for a nontrivial agreement between the two
groups are very stringent, and require in particular that the two
and the eigenvalues of L2 are the remaining N − 1 eigenvalues of
teams have the same cardinality.
L (note that, as before, λ̂ can also be an eigenvalue of L2 , in case An interesting open problem is that of investigating what kind
its multiplicity as eigenvalue of L is greater than 1). By proceeding
of nontrivial agreements may be reached in case of antagonistic
as in the previous two cases, we obtain the following description:
interactions, when the structural balance assumption is not
ζ̇ (t ) = (A − λ̂bK )ζ (t ), satisfied.
δ˙2 (t ) δ2 ( t )
   
Acknowledgments
 ..   . 
 .  = [(IN −1 ⊗ A) − (IN −1 ⊗ b)(L2 ⊗ K )]  .. 
δN (t ) We are indebted to the Reviewers for their fruitful comments
δ˙N (t )
 N N
  and, in particular, to one of them for suggesting Remark 1.
 
[L]2j − [L]1j bK 
References

 j =1 j =1

..
 
+  ζ (t ).
 
. [1] J.A. Benediktsson, P.H. Swain, Consensus theoretic classification methods, IEEE

  N N
 
 Trans. Syst. Man Cybern. 22 (4) (1992) 688–704.
 [2] S. Chatterjee, E. Seneta, Towards consensus: some convergence theorems on
[L]Nj − [L]1j bK
 
repeated averaging, J. Appl. Probab. 14 (1) (1977) 89–97.
j =1 j =1 [3] M.H. DeGroot, Reaching a consensus, J. Amer. Statist. Assoc. 69 (345) (1974)
118–121.
For every choice of λ̂, the characteristic polynomial of (IN −1 ⊗ A) − [4] H.L. Smith, Monotone Dynamical Systems: An Introduction to the Theory
of Competitive and Cooperative Systems, in: Mathematical Surveys and
(IN −1 ⊗ b)(L2 ⊗ K ) is the product of N − 1 polynomials ψi (s), each Monographs, vol. 41, AMS, Providence, RI, 1995.
of them having (at least) one not asymptotically stable zero in α . [5] J.N. Tsitsiklis, Problems in Decentralized Decision Making and Computation,
Ph.D. Thesis, Department of EECS, MIT, 1984.
This prevents consensus among agents.
[6] J.A. Fax, R.M. Murray, Information flow and cooperative control of vehicle
By putting together the results of these two subsections, we can formations, IEEE Trans. Automat. Control 49 (9) (2004) 1465–1476.
formalize the following result. [7] A. Jadbabaie, J. Lin, A.S. Morse, Coordination of groups of mobile autonomous
agents using nearest neighbor rules, IEEE Trans. Automat. Control 48 (6) (2003)
988–1001.
Theorem 3. Consider the multi-agent system (6) with control [8] J. Lin, A.S. Morse, B.D.O. Anderson, The multi-agent rendezvous problem,
algorithm (7), and assume that the communication graph G is in: Proceedings of the 42nd IEEE Conf. Decision and Control, Maui, Hawaii,
2003, pp. 1508–1513.
connected and structurally balanced, with respect to the partition in [9] L. Moreau, Stability of multiagent systems with time-dependent communica-
two subsets V1 and V2 . The consensus problem is solvable only if the tion links, IEEE Trans. Automat. Control 50 (2) (2005) 169–182.
following conditions are satisfied: [10] R. Olfati-Saber, J.A. Fax, R.M. Murray, Consensus and cooperation in networked
multi-agent systems, Proc. IEEE 95 (1) (2007) 215–233.
(i) the two sets V1 and V2 have the same cardinality; [11] W. Ren, R.W. Beard, Consensus seeking in multiagent systems under
(ii) all row sums of the Laplacian L are equal to λ̂ > 0, and λ̂ is a dynamically changing interaction topologies, IEEE Trans. Automat. Control 50
(5) (2005) 655–661.
simple eigenvalue of L; [12] R.O. Saber, R.M. Murray, Consensus problems in networks of agents with
(iii) A is Hurwitz. switching topology and time-delays, IEEE Trans. Automat. Control 49 (9)
(2004) 1520–1533.
If (i)–(iii) hold, then a necessary and sufficient condition for the [13] D. Bauso, L. Giarrè, R. Pesenti, Nonlinear protocols for optimal distributed
consensus problem to be solvable is that there exists K ∈ R1×N consensus in networks of dynamic agents, Systems Control Lett. 55 (11) (2006)
918–928.
such that ψ(s, λ̂) is simply stable, while ψ(s, λi ) is Hurwitz for every [14] S. Hara, T. Hayakawa, H. Sugata, Stability analysis of linear systems with
generalized frequency variables and its applications to formation control,
λi ∈ σ (L), λi ̸= λ̂. Such a matrix K always exists if and only if either in: Proceedings of the 46th IEEE Conference on Decision and Control, New
one of the following cases apply: (1) λ̂ is the largest eigenvalue of Orleans, LA, 2007, pp. 1459–1466.
[15] S. Hara, T. Hayakawa, H. Sugata, LTI systems with generalized frequency
L and b ̸= 0; or (2) λ̂ is not the largest eigenvalue of L and the variables: a unified approach for homogeneous multi-agent dynamical
dimension of the controllable subspace of (A, b) is at least three. systems, SICE J. Control, Meas., Sys. Integr. 2 (5) (2009) 299–306.
[16] Q. Hui, W.M. Haddad, Distributed nonlinear control algorithms for network
consensus, Automatica 44 (9) (2008) 2375–2381.
5. Conclusions [17] Z. Lin, M. Broucke, B. Francis, Local control strategies for groups of mobile
autonomous agents, IEEE Trans. Automat. Control 49 (4) (2004) 622–629.
In this paper we have investigated the bipartite consensus [18] Z. Lin, B. Francis, M. Maggiore, State agreement for continuous time coupled
nonlinear systems, SIAM J. Control 46 (1) (2007) 288–307.
problem and the standard consensus problem for a group of [19] M. Mesbahi, M. Egerstedt, Graph-Theoretic Methods in Multiagent Networks,
N homogeneous agents, each of them described by a generic Princeton Univ. Press, Princeton, NJ, 2010.
n-dimensional state-space model. Under the assumption that [20] W. Ren, R.W. Beard, E.M. Atkins, Information consensus in multivehicle
cooperative control, IEEE Control Syst. Mag. 27 (2) (2007) 71–82.
the communication graph describing the cooperative/antagonistic [21] L. Scardovi, R. Sepulchre, Synchronization in networks of identical linear
relationships among agents is structurally balanced, we have seen systems, Automatica 45 (11) (2009) 2557–2562.
that bipartite consensus can be achieved if and only if the state- [22] C. Altafini, Consensus problems on networks with antagonistic interactions,
IEEE Trans. Automat. Control 58 (4) (2013) 935–946.
space model describing each agent is stabilizable. In addition, [23] D. Easley, J. Kleinberg, Networks, Crowds, and Markets. Reasoning About a
the common agents’ dynamics is regulated by the state matrix A Highly Connected World, Cambridge Univ. Press, Cambridge, UK, 2010.
involved in the agents’ description. In fact, the common agents’ [24] Y. Hou, J. Li, Y. Pan, On the Laplacian eigenvalues of signed graphs, Linear
Multilinear Algebra 51 (1) (2003) 21–30.
dynamics converge to zero, to a bounded or diverging common [25] T. Zaslavsky, Signed graphs, Discrete Appl. Math. 4 (1) (1982) 47–74.
state trajectory depending on the spectrum of A (and on the [26] D. Bauso, L. Giarrè, R. Pesenti, Dissensus, death and division, in: Proceedings of
agents’ initial conditions). These results turn out to be analogous the American Control Conference, Seattle, WA, 2009, pp. 2307–2312.
 M.E. Valcher, P. Misra / Systems & Control Letters 66 (2014) 94–103 103

[27] D. Bauso, L. Giarrè, R. Pesenti, Quantized dissensus in networks of agents [30] P. Wieland, J.-S. Kim, F. Allgover, On topology and dynamics of consensus
subject to death and duplication, IEEE Trans. Automat. Control 57 (2012) among linear high-order agents, Internat. J. Systems Sci. 42 (10) (2011)
783–788. 1831–1842.
[28] G. Facchetti, G. Iacono, C. Altafini, Computing global structural balance in [31] P. Wieland, J.-S. Kim, H. Scheu, F. Allgower, On consensus in multi-agent
large-scale signed social networks, Proc. Natl. Acad. Sci. USA 108 (2011) systems with linear high-order agents, in: Proceedings of the 17th IFAC World
20953–20958. Congress, Seoul, Korea, 2008, pp. 1541–1546.
[29] J. Kunegis, S. Schmidt, A. Lommatzsch, J. Lerner, E.W. De Luca, S. Albayrak, [32] Y. Hu, P. Li, J. Lam, Consensus of multi-agent systems: a simultaneous
Spectral analysis of signed graphs for clustering, prediction and visualization, stabilisation approach, IET Control Theory Appl. 6 (11) (2012) 1758–1765.
in: Proceedings of the Tenth SIAM International Conference on Data Mining, [33] S.E. Tuna, LQR-based coupling gain for synchronization of linear systems, Jan
SDM10, Columbus, OH, USA, 2010, pp. 559–570. 2008. arXiv:0801.3390v1 [math.OC].