ISA Transactions 71 (2017) 68–75

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Research article

Scaled consensus seeking in multiple non-identical linear autonomous

agents
Cheng-Lin Liu
Key Laboratory of Advanced Process Control for Light Industry (Ministry of Education), Institute of Automation, Jiangnan University, Wuxi 214122, China

art ic l e i nf o a b s t r a c t

Article history: Scaled consensus problem is studied for a heterogeneous multi-agent system composed of non-identical
Received 12 November 2016 stable linear agents, and a leader-following scaled consensus algorithm is designed. By using frequency-
Received in revised form domain analysis, consensus conditions are obtained for the agents without communication delay under
21 June 2017
undirected and directed topologies, respectively. Moreover, consensus criteria are also gained for the
Accepted 22 June 2017
Available online 8 July 2017
agents suffering from communication delay under directed topology. Simulation examples show the
correctness of theoretical results.
Keywords: & 2017 ISA. Published by Elsevier Ltd. All rights reserved.
Heterogeneous linear multi-agent systems
Scaled consensus problem
Leader-following algorithm
Communication delay

1. Introduction heterogeneous linear multi-agent system with diverse communica-

As a fundamental phenomenon in flocking, swarming, and cou-
pled synchronization, consensus behavior has attracted more and
more attention of researchers in the control theory field in recent
years. Consensus problem requires several autonomous agents to
reach a common agreement on their states, and has broad en-
gineering applications, e.g., coordination control of unmanned aerial
and ground vehicles, smart grid, wireless sensor networks, etc.
Current research results on consensus problem have been
mainly focused on homogeneous multi-agent systems, of which
the agent's dynamics are identical [1–10]. With the help of alge-
braic graph theory [1], matrix theory [2–6], frequency-domain
analysis [7–9], and Lyapunov functions [10], various consensus
algorithms have been designed and analyzed for the agents with
fixed topology or switching topologies.
Over the past few years, furthermore, a lot of research effort has
also been put into the consensus analysis and synthesis of hetero-
geneous multi-agent system composed of the agents with distinct
dynamics. As a simple heterogeneous multi-agent system, mixed
system with multi-class agents has attracted many researchers’ in-
terests, such as the system of first-order agents and second-order
agents [11,12] and the system of Euler-Lagrange agents and double-
integrator agents [13,14]. For the heterogeneous linear multi-agent
systems, state-feedback and output-feedback consensus algorithms
were proposed by using the internal model principle [16,17] and the
harmonic control [18]. To solve the consensus problem of
E-mail address: liucl@jiangnan.edu.cn
tion delays, Lee and Spong [15] used the spectral radius theorem to
obtain the delay-independent consensus conditions. Meanwhile,
Muenz [19] and Tian and Zhang [20] adopted the consensus algo-
rithms with devisable self delays distinct from the corresponding
communication delays, and provided the consensus criteria based on
frequency-domain analysis. In addition, Tian and Zhang [20] de-
signed an adaptive adjustment algorithm in order to adjust the self
delays on-line if the communication delays were unknown. Liu and
Liu [21] proposed three adaptive consensus algorithms for the het-
erogeneous first-order time-invariant multi-agent systems, and ob-
tained the consensus conditions for the agents under fixed topology
without and with time delays respectively. Specially, non-identical
disturbances also lead to the heterogeneity of multi-agent systems
[22]. Based on Lyapunov functions, Kim et al. [22] adopted a general
consensus algorithm and analyzed the robust consensus con-
vergence condition For the heterogeneous first-order time-varying
multi-agent systems with distinct disturbances.
Treated as the generalized consensus problem, scaled consensus
problem defined in [23] means that the network components’ scalar
states reach the assigned proportions rather than a common value,
and has attracted some researchers’ interests for its engineering
applications, such as compartmental mass-action systems, closed
queueing networks, and water distribution systems [23]. In addition
to usual consensus problem, group consensus problem [24],
wherein the agents form several sub-groups reach corresponding
agreement values respectively, is also a special case of scaled con-
sensus problem. With a connected topology, Roy [23] proved that
the first-order multi-agent systems not only achieved a stationary
scaled consensus asymptotically, but also could track on the stable

http://dx.doi.org/10.1016/j.isatra.2017.06.022
0019-0578/& 2017 ISA. Published by Elsevier Ltd. All rights reserved.
 C.-L. Liu / ISA Transactions 71 (2017) 68–75
manifold. For the first-order multi-agent systems, Meng and Jia [25]
studied the scaled consensus problem with time-varying scaled ra-
tios, and demonstrated that the agents converged to the scaled
consensus asymptotically under jointly-connected switching topol-
ogies. By constructing Lyapunov-Krasovskii functionals, Aghbolagh
et al. [26] considered the scaled consensus of first-order multi-agent
systems with time-varying delay, and gained the consensus condi-
tions in the form of linear matrix inequalities, which were used to
calculate the allowable delay value.
In this paper, we consider the scaled consensus problem of a
class of heterogeneous multi-agent systems with the agents
modeled by single-input and single output linear stable dynamics,
and an adaptive scaled consensus algorithm is constructed in the
leader-following structure. Based on frequency-domain analysis,
firstly, we obtain the consensus conditions for the multi-agent
systems without communication delay under undirected and di-
rected topologies, respectively. For the multi-agent systems with
identical communication delay, besides, the delay-independent
and delay-dependent consensus conditions are obtained, respec-
tively, for the system under directed topology.
2. Problem description
2.1. Non-identical agents and topology
Consider a class of heterogeneous multi-agent system com-
posed of N non-identical single-input single-output linear agents
given by
xi̇ (t ) = Ai xi (t ) + Bi(ui (t ) + ni ),
yi (t ) = cxi(t ), i = 1, …, N ,
⎡ 0 1 0 ⋯ 0 ⎤ ⎡ 0⎤
⎢ ⎥ ⎢ ⎥
⎢ 0 0 1 ⋯ 0 ⎥ ⎢ 0⎥
Ai = ⎢ ⋮ ⋮ ⋮ ⋯ ⋮ ⎥, Bi = ⎢ ⋮ ⎥,
⎢ 0 0 0 ⋯ 1 ⎥ ⎢ 0⎥
⎢⎣ −α −α −α ⎢ ⎥
i0 i1 i2 ⋯ −αin − 1⎥⎦ ⎣ bi ⎦
c = ⎣⎡ 1 0 0 ⋯ 0⎤⎦.
where x i(t ) ∈ Rn , yi (t ) ∈ R and ui(t ) ∈ R are the state, the output
and the input of agent i, respectively, ni ∈ R is the constant dis-
turbance of each agent i, Ai ∈ Rn × n , Bi ∈ Rn , c ∈ R1 × n , bi ∈ R , and
αik ∈ R, k = 0, … , n − 1. Correspondingly, the frequency-domain
description of (1) is formulated as
Yi(s ) = gi (s )(Ui(s ) + Ni(s )), i = 1, …, N ,
where Yi(s), Ui(s) and Ni(s) are the Laplace transforms of yi(t), ui(t)
and ni respectively, and gi(s) is the transfer function. Since the
agents’ dynamics (1) are formulated as a controllable canonical
form, we get
bi
gi (s ) =
di(s )
bi
= .
s n + αin − 1s n − 1 + ⋯ + αi1s + αi0
Assumption 1. The system matrix Ai , i = 1, … , N is Hurwitz, i.e.,
di(s ) , i = 1, … , N is Hurwitz.
Evidently, Ai is non-singular with Assumption 1 guaranteeing
that all the eigenvalues of Ai are non-zero.
69
Assumption 2. Without loss of generality, bi is assumed to be
positive, i.e., bi > 0.
Definition 1. The multi-agent network (1) achieves Scaled Con-
sensus asymptotically, if
lim (ri yi (t ) − r j yj (t )) = 0, i, j = 1, 2, …, N ,
t →∞ (4)
where the ratios ri ∈ R, i = 1, … , N are assumed to be non-zero.
If ri = r j, ∀ i, j ∈ {1, … , N}, the output scaled consensus in De-
finition 1 becomes an output synchronization. Hence, the output
scaled consensus problem (4) is a generalized output synchroni-
zation problem.
Usually, the interconnection topology of multi-agent system (1)
is described as an N-order weighted digraph G = (V , E,  ), which
is composed of a set of vertices V = {1, … , N}, a set of edges
E ⊆ V × V and a weighted adjacency matrix  = [aij ] ∈ R N × N with
aij ≥ 0. In the diagraph G, a directed edge from the node i to the
node j is denoted by eij = (i, j ) ∈ E , and it is assumed that
aij > 0 ⇔ eij ∈ E . Moreover, we assume aii = 0 for all i ∈ V . Besides,
the digraph G is undirected graph or bidirectional if eij ⇔ eji , and
the digraph is symmetric if aij = aji . The set of node i's neighbors is
denoted by Ni = {j ∈ V : (i, j ) ∈ E}. The Laplacian matrix of the
weighted digraph G is defined as L = D −  = [lij ] ∈ Rn × n , where
N
D = diag{ ∑ j = 1 aij , i ∈ V} is the degree matrix.
In the digraph, if there is a path from one node i to another
node j, then j is said to be reachable from i, otherwise, j is said to be
not reachable from i. If a node is reachable from every other node
in the digraph, then we say it globally reachable. A globally
reachable node is precisely the degree of connectedness required
(1) of the digraph. An undirected graph is connected if it has a globally
reachable node.
2.2. Design of scaled consensus algorithm
In this paper, we focus on the leader-following coordination
control, which has been extensively studied for multi-agent sys-
tems [27–31]. For the agents (1), a usual scaled consensus algo-
rithm is given by
⎛ ⎞
1⎜
ξi(t ) = ∑ aij(r j yj (t ) − ri yi (t )) + pi ( y0 − ri yi (t ))⎟⎟, i ∈ V ,
ri ⎜⎝ ⎠
j ∈ Ni (5)
where the non-zero constants ri, i ∈ V denotes the ratios of agents’
states, Ni denotes the neighbors of agent i, aij > 0 is the adjacency
element of  in the digraph G = (V , E,  ), y0 is the static leader's
(2) state, and pi denotes the linking weight from agent i to the leader.
Assume that pi > 0 if there is a directed edge from agent i to the
leader; otherwise, pi ¼0. In the rest of this paper, the notation
P = diag{pi , i ∈ V} is used. Apparently, the algorithm (1) is similar
to that in [25].
Taking ui(t ) = κξi(t ) with κ > 0, we get the closed-loop form of
multi-agent system (1) as follows
xi̇ (t ) = Ai xi (t ) + Bi(κξi(t ) + ni ), i ∈ V . (6)
(3)
Remark 1. With the interconnection topology that has a globally
reachable node, it is easily proved that the multi-agent system (6)
converges to an asymptotic scaled consensus, only if
−1
rcA
i i Bini = y0 , ∀ i ∈ V . (7)
Condition (7) can be obtained by assuming that the agents converge
70 C.-L. Liu / ISA Transactions 71 (2017) 68–75

to the scaled consensus asymptotically, i.e., limt →∞ri yi (t ) = y0 . Hence, the scaled consensus problem of system (9) becomes a
consensus problem of system (10).
Evidently, the condition (7) is dependent on the parameters
Define y^i (t ) = y¯i (t ) − y0 , x^i(t ) = x¯ i(t ) − x0 with x0 = [y0 , 0, … , 0]T
strictly, i.e., the agents cannot achieve the asymptotic scaled con-
α
sensus when the condition (7) is not satisfied. Moreover, (7) also and z^ (t ) = − i0 y + rn + rz (t ), and we obtain the dynamics of x^
i bi 0 i i i i i
implies that the agents (1) without disturbances cannot reach the as follows
asymptotic scaled consensus with the algorithm (5).
̇
To achieve the asymptotic scaled consensus and cancel the x^ i (t ) = A i x^ i (t ) + Bi(z^i(t ) + κ ( ∑ a ij(y^j (t ) − y^i (t )) − pi y^i (t )) ,
j ∈ Ni
impact of the diversity of the disturbances ni, we design the fol-
̇
lowing adaptive algorithm for multi-agent system (1), z^i(t ) = γ ( ∑ a ij(y^j (t ) − y^i (t )) − pi y^i (t )) ,
j ∈ Ni
ui (t ) = zi(t ) + κξi(t ),
y^i (t ) = cx^ i (t ) . (11)
zi̇ (t ) = γξi(t ), (8)
Moreover, the scaled consensus problem of system (9) turns to be
where κ > 0, γ > 0, and zi(t) is an additional variable of consensus
controller ui(t). Furthermore, the algorithm (8) is a proportional- an asymptotic stability problem of system (11).
Taking the Laplace transform of system (11) yields
integral form of algorithm (5).
^ ^ ^ ^ ^ ^
sX i (s) = A i X i (s) + Bi(Z i(s) + κ ( ∑ a ij(Yj(s) − Yi (s)) − pi Yi (s)),
2.3. Useful lemmas j ∈ Ni

^ ^ ^ ^
sZ i(s) = γ ( ∑ a ij(Yj(s) − Yi (s)) − pi Yi (s)),
For the leader-following control structure, we come to the j ∈ Ni
following property from Lemma 2 in [32], Lemma 2 in [28], ^ ^
Yi (s) = cX i (s), (12)
Lemma 3 in [27] and Lemma 1 in [21].
^ ^ ^
Lemma 1. uppose that the interconnection topology of N agents (3) where Xi(s ), Zi(s ) and Yi (s ) are the Laplace transforms of x^i(t ), z^i(t )
and the static leader y0 has the leader as a globally reachable node. and y^ (t ) respectively. Thus, the characteristic equation of system
The eigenvalues of L + P are all none-zero, and D + P > μI , where (11) about y^ (t ) = [y^1(t ) , … , y^N (t )]T is given by
μ > 0 and L is the Laplacian matrix of the topology of N agents (3)
^
without leader. det(diag{sdi(s ), i ∈ V} + (κs + γ ) L ) = 0, (13)

In addition, the following two lemmas will play important roles where  = diag{bi , i ∈ V} and L^ = L + P .
in the proof of main results.

Lemma 2. [33] Gr (s ) ∈ R(s )N × N is proper, where N is a positive in-

3.1. Undirected topology
tegrator. The zeros of Δ(s ) = det(I + Gr (s )) lie on the open left half
complex plane if and only if the number of counterclockwise en-
At first, we provide the scaled consensus condition for multi-
circlements of ( − 1, j0) by the eigenloci of Gr (jω) as ω goes from −∞
agent system (9) with an undirected topology.
to +∞ equals the total number of right-half plane poles of Gr(s).

Lemma 3. [34] Let Q ∈ C n × n , Q = Q⋆ ≥ 0 and T = diag{ti, ti ∈ C}. Theorem 1. Consider multi-agent system (9) with an undirected and
Then symmetric topology that has the leader as a globally reachable node.
λ(QT ) ∈ ρ(Q )Co(0 ∪ {ti}), Under Assumptions 1 and 2, then, the agents in system (9) converge
to a scaled consensus asymptotically, if
where λ(·) denotes matrix eigenvalue, ρ(·) denotes the matrix spectral
radius, and Co(·) denotes the convex hull. 2bi (max( ∑ aij + pi ))Γ < 1,
i∈V
j ∈ Ni (14)

where Γ > 0 is the positive minimum value that satisfies

3. Scaled consensus seeking without time delay ( − Γ , j0) ∉ Co(0 ∪ gi(jω) , i ∈ V ) with ω ∈ R , and gi(jω) =
jκω + γ
.
jωdi(jω)

With (8), the dynamics of agents (1) are rewritten as ^

Proof. Let f (s ) = det(diag{sdi(s ) , i ∈ V} + (κs + γ ) L ), and
ẋi(t ) = A i x i(t ) + Bi(n i + z i(t ) + κξi(t )), n n−1 ^
f (0) = det(diag{0(0 + αin − 10 + ⋯ + α i 0 ) , i ∈ V } + (κ 0 + γ ) L )
^
z i̇ (t ) = γξi(t ), =det(γ  L ) ≠ 0 holds from Lemma 1.
1
ξi(t ) = ( ∑ a ij (r j yj (t ) − ri yi (t )) + pi ( y0 − ri yi (t )), i ∈ V ,
When s ≠ 0, the equation (13) is equivalent to
ri j ∈ Ni

yi (t ) = cx i(t ). (9) κs + γ ^
det(I + diag{ , i ∈ V} L ) = 0.
sdi(s ) (15)
Let y¯i (t ) = ri yi (t ) and x¯ i(t ) = rx
i i(t ), and we get
Denote h(s ) = det(I +
κs + γ ^
x¯ i̇ (t ) = Ai x¯ i (t ) + Bi(rn ¯ diag{ sd (s) , i ∈ V} L ). According to Assump-
i i + rz
i i(t ) + κξi(t )), i

1 tion 1 and Lemma 2, the zeros of h(s) are on the open left half
zi̇ (t ) = γ ξ¯i(t ),
complex plane, if λ(diag{ jωd ( jω) , i ∈ V} L^ ) does not enclose the
jκω + γ
ri
i

ξ¯i(t ) = ∑ aij( y¯j (t ) − y¯i (t )) + pi ( y0 − y¯i (t ), i ∈ V , point ( − 1, j0) for ω ∈ R . The definition of Laplacian matrix and
j ∈ Ni ^ ^T
symmetric weights make L = L > 0 hold. Based on Assumption 2
y¯i (t ) = cx¯ i (t ). (10) and Lemma 3, we get
 C.-L. Liu / ISA Transactions 71 (2017) 68–75 71

jκω + γ ^ matrix.
λ(diag{ , i ∈ V } L )
jωdi(jω)
Proof. From the assumption in Theorem 2, the equation (18) is
^
= λ(diag{gi (jω), i ∈ V}diag{ bi , i ∈ V}L diag{ bi , i ∈ V}) rewritten as
^ ^
∈ ρ(diag{ bi , i ∈ V}L diag{ bi , i ∈ V})Co(0 ∪ gi (jω), i ∈ V ) det(I + sΛ(s )(sd(s )I + (κs + γ ) L )−1) = 0, (21)
⊂ 2bi (max( ∑ aij + pi ))ΓΓ −1Co(0 ∪ gi (jω), i ∈ V ). where Λ(s ) = diag{Δdi(s ) , i ∈ V}. It follows from condition (20) that
i∈V
j ∈ Ni
^
Thus, gi(jω) , i ∈ V crosses the negative real axis at some certain σ¯ (jωΛ(jω)(jωd(jω)I + (jκω + γ ) L )−1)
frequencies under Assumption 1, and Γ must exist from the ^
≤ σ¯ (Λ(jω))σ¯ (jω(jωd(jω)I + (jκω + γ ) L )−1)
definition. For the condition (14), hence, ρ(diag{ bi , i ∈ V} ^
= {max|Δdi(jω)|}σ¯ (jω(jωd(jω)I + (jκω + γ ) L )−1)
^ i∈V
L diag{ bi , i ∈ V})Co(0 ∪ gi(jω) , i ∈ V ) does not enclose the point
< 1.
( − 1, j0), and the zeros of h(s) are on the open left half plane. Now,
we have proved that the zeros of f(s) all have negative real parts, Therefore, the roots of (21) all lie on the open left half complex
i.e., limt →∞x^i(t ) = 0, i ∈ V . Therefore, the agents in (9) converge to a plane, i.e., the agents in (16) converge to a scaled consensus pro-
scaled consensus asymptotically. □ blem asymptotically. □

It should be highlighted that the roots’ distribution of equation

3.2. General directed topology (19) with complex coefficients can be determined by the Routh-
like criterion [35].
Now, we analyze the scaled consensus seeking of system (9)
with general directed topology.
1 n 4. Scaled consensus with communication delay
Define Ai = A + ΔAi , where A = n ∑i = 1 Ai , and the agents’ dy-
namics (9) are reformulated as
With non-negligible communication delay, the asynchro-
x i̇ (t ) = (A + ΔA i )x i(t ) + Bi(n i + z i(t ) + κξi(t )), nously-coupled form of algorithm (8) is
z i̇ (t ) = γξi(t ),
u i (t ) = z i(t ) + κξiτ(t ),
1
ξi(t ) = ( ∑ a ij(r j yj (t ) − ri yi (t )) + pi ( y0 − ri yi (t )), i ∈ V , z i̇ (t ) = γξiτ(t ),
ri j ∈ Ni
1
yi (t ) = cx i(t ). (16) ξiτ(t ) = ∑ a ij (r j yj (t − τ ) − ri yi (t )) + pi ( y0 − ri yi (t )),
ri j ∈ Ni (22)
In (16), the characteristic polynomial of A is
d(s ) = sn + αn − 1sn − 1 + ⋯ + α1s + α0 . where τ > 0 is the communication delay. Since the leader's state
^ tracked by the agents (1) is stationary, the asynchronously-cou-
In the same way, let y^i (t ) = ry i i (t ) − y0 , x i(t ) = rx
i i(t ) − x 0 with
α
pled algorithm (22) does not change the final consensus behavior
x = [ y , 0, … , 0] and z^ (t ) = − y + rn + rz (t ), and we get
0 0
T i 0
i bi 0 i i i i [9]. With (22), the closed-loop dynamics of agents (1) are for-
̇ mulated as
x^i(t ) = (A + ΔA i )x^i(t ) + Bi(z^i(t ) + κ ( ∑ a ij(y^j (t ) − y^i (t )) − pi y^i (t )),
j ∈ Ni
ẋi (t ) = Ai xi + Bi(ni + zi(t ) + κξiτ (t )),
̇
z^i(t ) = γ ( ∑ a ij(y^j (t ) − y^i (t )) − pi y^i (t )), zi̇ (t ) = γξiτ (t ). (23)
j ∈ Ni

y^i (t ) = cx^i(t ). (17)

Then, the characteristic equation of system (17) about 4.1. Delay-independent case
y^ (t ) = [y^1(t ) , … , y^n (t )]T is given by
^
det(diag{s(d(s ) + Δdi(s )), i ∈ V} + (κs + γ ) L ) = 0, (18) Theorem 3. Consider multi-agent system (23), of which the directed
topology has the leader as a globally reachable node. Assume that the
where Δdi(s ) = di(s ) − d(s ).
roots of following equations
Theorem 2. Consider multi-agent system (16) with a directed to- ^ −1
sdi(s )bi−1di + (κs + γ ) = 0 (24)
pology that has the leader as a globally reachable node. Suppose that
the roots of following equations ^
all lie on the open left half complex plane, where di = ∑j ∈ N aij + pi .
i

sd(s ) + λ^i(κs + γ ) = 0, i = 1, …, N (19) Then, all the agents in system (23) asymptotically converge to the
scaled consensus asymptotically, if
lie on the open left half complex plane, where λ^i , i ∈ V is the eigen-
|jωκ + γ |σ¯ ( −1(D + P )−1 )
value of  L^ . Then, the agents in system (16) converge to the scaled < 1, i ∈ V
^ −1
consensus asymptotically, if |jωdi(jω)bi−1di + (κ jω + γ )| (25)

1 holds for ∀ ω ∈ R .
|Δdi(jω)| < ,∀i∈V
^
σ¯ (jω(jωd(jω)I + (jκω + γ ) L )−1) (20) Proof. Similar to the proof of Theorem 1, we take variable trans-
α
hold for ω ∈ R , where σ̄(·) denotes the largest singular value of formations as z^ (t ) = − i0 y + rn + rz (t ), y^ (t ) = ry (t ) − y , and
i bi 0 i i i i i i i 0
72 C.-L. Liu / ISA Transactions 71 (2017) 68–75

x^i(t ) = rx T 4.2. Delay-dependent case

i i(t ) − x 0 with x 0 = [ y0 , 0, … , 0] , and get

̇
x^ i(t ) = A i x^ i(t ) + Bi(z^i(t ) + κ ( ∑ a ij( y^j (t − τ ) − y^i (t )) − pi y^i (t )) , In this section, we will present a delay-dependent condition
j ∈ Ni
that is suitable for an arbitrary connected leader-following
̇
z^i(t ) = γ ( ∑ a ij( y^j (t ) − y^i (t − τ )) − pi y^i (t )) , topology.
j ∈ Ni

y^i (t ) = cx^ i(t ) .

Theorem 4. Suppose that the agents in system (23) without com-
(26)
munication delay converges to the scaled consensus asymptotically.
Taking the Laplace transform of system (26), we get the char- Let
acteristic equation on y^ (t ) = [y^1(t ) , … , y^N (t )]T as ^ (κs + γ )
M (s ) = .
−sτ ^
det(diag{sdi(s ), i ∈ V} + (κs + γ ) (D + P −  e )) = 0. (27) diag{sdi(s ), i ∈ V} + (κs + γ ) L (30)

Under the assumption in Theorem 3, (27) is equivalent to Then, all the agents in system (23) achieve an asymptotic scaled
consensus, if
det(I − M (s )) = 0, (28)
ωτ ^
2|sin( )|σ¯ (M (jω)) < 1
2 (31)
(κs + γ ) holds with ω ∈ R .
M (s ) = diag{ −1
, i ∈ V} −1(D + P )−1 e−sτ ,
^
sdi(s )bi−1di + (κs + γ ) Proof. With the supposition that the agents without commu-
nication delay converge to the scaled consensus asymptotically,
where M(s) has no poles on the open right half complex plane.
the roots of following equation
It follows from Lemma 1 and condition (25) that
^
ρ(M (jω)) det(diag{sdi(s ), i ∈ V} + (κs + γ ) L ) = 0 (32)
jωκ + γ ^
≤ σ¯ (diag{ , i ∈ V} all lie on the open left half complex plane, i.e., M (s ) has its poles on
^ −1 ^
jωdi(jω)bi−1di + (κ jω + γ ) the open left half complex plane, so rank(L ) = N .
 −1(D + P )−1 e−jωτ ) In the same way as Theorem 3, the characteristic equation (27)
is also rewritten as
jωκ + γ
≤ σ¯ (diag{ , i ∈ V})
−1 ^
^
jωdi(jω)bi−1di + (κ jω + γ ) det(diag{sdi(s ), i ∈ V} + (κs + γ ) (L +  (1 − e−sτ ))) = 0, (33)
−1 −1
σ¯ ( (D + P )  ) which equals
< 1. (29) ^
det(I + M (s )(1 − e−sτ )) = 0. (34)
Therefore, the roots of Eq. (27) lie on the open left half complex
plane, i.e., the agents in system (23) converge to the scaled con- Then, we obtain from the condition (31) that
sensus asymptotically. □ ^
ρ(M (jω)(1 − e−jωτ ))
Intuitively, if the coupling weights aij between agents them- ^
≤ σ¯ (M (jω)(1 − e−jωτ ))
selves are smaller and the coupling weights pi between the agents ^
and the leader are higher, σ̄( −1(D + P )−1 ) turns to be smaller, i.e., = |1 − e−jωτ |σ¯ (M (jω))
ωτ ^
the condition (25) in Theorem 3 hold more easily. < 2|sin( )|σ¯ (M (jω))
2
< 1.
Remark 2. Take into account a simplest case that the inter-
connection topology of N agents and a leader in system (23) has Hence, the equation (34) has its roots on the open left half com-
the leader as a globally reachable node, and each agent just has plex plane, i.e., the agents in system (23) converge to the scaled
one direct path to reach the leader. Besides, we assume that the consensus asymptotically. □
direct edge from agent i to j satisfies i > j , and we get Furthermore, we present a concrete delay bound for the
⎡ 0 0 ⋯ 0⎤ achievement of asymptotic scaled consensus seeking.
⎢ ⎥
a 0 ⋯ 0⎥
 = ⎢ 21 . Corollary 1. It is assumed that the agents in system (23) without
⎢ ⋮ ⋮ ⋯ ⋮⎥ communication delay converges to the scaled consensus asymptoti-
⎢ ⎥
⎣ a n1 a n2 ⋯ 0⎦ cally. Then, all the agents in system (23) reach an asymptotic scaled
consensus, if
Since the eigenvalues of adjacent matrix  are all equivalent to
zero, 1
τ<
jωκ + γ ^
ρ(M (jω)) = ρ(diag{ , i ∈ V} −1(D + P )−1 e−jωτ ) = 0 < 1 σ¯ ( jωM (jω)) (35)
^ −1
jωdi( jω)b −1di + (κ jω + γ )
i
^
holds with arbitrary communication delay if the assumption in holds with ω ∈ R , where M (jω) is defined in (30).
Theorem 3 holds. Thus, asymptotic scaled consensus convergence
in this case is delay-independent.
5. Numerical simulation
It should be pointed out that the delay-independent scaled
consensus condition in Theorem 3 is relatively conservative, and 5.1. Scaled consensus convergence without time delay
does not hold for arbitrary topology, coupling weights and control
parameters according to Remark 3, Remark 5 and Remark 7 in [29]. Investigate a heterogeneous second-order multi-agent network
 C.-L. Liu / ISA Transactions 71 (2017) 68–75 73

Fig. 1. Undirected topology of five agents and a leader.

given by
⎡ 0 1 ⎤ ⎡ 0⎤
ẋi (t ) = ⎢ x (t ) + ⎢ ⎥ (ui (t ) + ni ),
⎣ −αi0 −αi1⎥⎦ i ⎣ 1⎦
⎡ ⎤
y (t ) = ⎣ 1 0⎦ xi (t ), (36)
i
Fig. 3. Nyquist curves of gi(jω) .
and we consider the agents (36) driven by the algorithm (8) under
an undirected topology shown in Fig. 1 firstly. The undirected to- Fig. 3). In Theorem 1, Γ is an infinitesimal positive number, and
pology is connected and the leader is the globally reachable node. it is concluded from Theorem 1 that the agents (36) with (8)
The agents’ parameters are α10 = 1.5, α11 = 0.5, α20 = 1.6, α21 = 0.6, converge to the scaled consensus asymptotically (see Fig. 4).
α30 = 1.7, α31 = 0.7, α40 = 1.8, α41 = 0.8, α50 = 1.9, α51 = 0.9 , b1= Next, we study the scaled consensus seeking of second-order
0.8, b2 = 1.2, b3 = 0.5, b4 = 1, b5 = 1.5 , and the constant dis-
agents with a directed and asymmetric topology shown in Fig. 5.
turbances are n1 = 0.5, n2 = 0.3, n3 = 0.7, n4 = 1.0, n5 = − 0.2. The weights
The adjacent weights are chosen as: a13 = 0.4, a21 = 0.5,
with respect to the edges are a12 = a21 = 0.5, a14 = a41 = 0.5, a23
a32 = 0.4, a34 = 0.2, a45 = 0.3, a51 = 0.5, p1 = 0.2 , and the eigenva-
=a32 = 0.4, a34 = a43 = 0.2, a45 = a54 = 0.3, p1 = 0.5, p5 = 0.2. . The
scaled ratios for each agent are chosen as r1 = 0.3, r2 = 0.4, lues of ^
 L are λ1 = 1.232, λ2 = 0.7334 + j0.0991, λ3 = 0.7334
r3 = − 0.2, r4 = 0.5, r5 = − 0.4 , and the control parameters are set −j0.0991, λ4 = 0.1519, λ5 = 0.3792 . Then, we set A = ⎡⎣ −1 −2 ⎤⎦, i.e.,
0 1
as κ = 3, γ = 1.2. Besides, the static leader's output is y0 = 1. In
y
addition, we define ei = ⎡⎣ 1 0⎤⎦x i − 0 as the errors of agents’ states. d(s ) = s2 + 2 s + 1, and other parameters are same as the above,
ri
and we will design ΔAi for each agent. For simplicity, we just in-
Firstly, we take into account the usual scaled consensus algo-
⎡ 0 0⎤
rithm (5). With the above coupling weights and control para- vestigate the case that ΔAi = ⎣⎢ −Δa 0 ⎦⎥, i.e., Δdi(s ) = Δαi0 .
i0
meters of agents (36), the condition (7) is not satisfied, so the al- According to the Routh-like criterion, the roots of Eq. (19) are
gorithm (5) cannot drive the agents (36) to achieve the assigned proved to be lie on the open left half complex plane, and we obtain
scaled consensus seeking (see Fig. 2). from the numerical simulation that σ¯(jω(jωd(jω)I
Then, we study the agents (36) driven by our proposed al- ^ −1
+(jκω + γ ) L ) ) < 1.0136 hold for ω ∈ R . Based on the condition
gorithm (8). Now, we analyze the Nyquist curves of
γ 1
jκω + γ κ (jω + )
κ
(20), we get |Δdi(jω)| = |Δαi0| < 1.0136
≃ 0.9866, i = 1, 2, 3, 4, 5, so
gi( jω) = 2
= with the above assigned
jω(( jω) + αi1( jω) + αi0) jω( jω + si )( jω + 1) we choose Δα10 = − 0.2, Δα20 = 0.3, Δα30 = 0.2, Δα40 = − 0.3,
parameters, and s1 = 0.5, s2 = 0.6, s3 = 0.7, s4 = 0.8, s5 = 0.9. Since Δα50 = 0.8 . Therefore, the agents (36) with (8) converge to the
γ
κ
= 0.4 < si , i = 1, 2, … , 5, we get arg(gi(jω)) ∈ ( − 2π , − π ), i.e., the scaled consensus asymptotically (see Fig. 6).
curves of gi(jω) just lie in the third and fourth quadrants (see

Fig. 2. Convergence under usual scaled consensus algorithm with undirected topology.
74 C.-L. Liu / ISA Transactions 71 (2017) 68–75

Fig. 4. Scaled consensus convergence under our proposed algorithm with undirected topology.

τ < 0.4892(s) from condition (35), i.e., the agents in system (36)
with (22) converge to the scaled consensus asymptotically if
τ < 0.4892(s) (see Fig. 7 with τ = 0.3(s)).

6. Conclusion

In this paper, we propose an adaptive scaled consensus algo-

Fig. 5. Directed topology.
5.2. Scaled consensus convergence with time delay
Remark 2 has demonstrated a delay-independent scaled con-
sensus seeking case based on Theorem 3, so this section just cares
about the delay-dependent scaled consensus convergence in
Theorem 4 and Corollary 1.
Study second-order multi-agent system (36) with delayed
scaled consensus algorithm (22) under the same directed and
asymmetric topology shown in Fig. 5, and we choose the same
coupling weights, agents’ parameters αi0, αi1, bi , scaled gains ri, and
control parameters κ , γ as those of the directed topology case in
above subsection. By numerical computing, then, we get
^
σ¯( jωM (jω)) < 2.044 for ω ∈ R , so the time delay bound is
rithm, which is in the leader-following form, to solve the scaled
consensus problem of a class of heterogeneous linear stable mul-
tiple agents with distinct constant disturbances. Based on fre-
quency-domain analysis, we obtain the scaled consensus criteria
for the multi-agent systems under the undirected and directed
topologies, respectively. Moreover, the scaled consensus seeking is
also analyzed for the multi-agent systems with communication
delay under the general directed topology. With proper topologies
and parameters, the asymptotic scaled consensus convergence can
be delay-independent. Furthermore, the delay-dependent con-
sensus condition, which is less conservative, is also provided for
the agents converging to the scaled consensus asymptotically
under general directed topology. Compared with current results
on scaled consensus problem [23–26], we focus on a hetero-
geneous linear multi-agent system, which is relatively

Fig. 6. Scaled consensus convergence under our proposed algorithm with directed topology.
 C.-L. Liu / ISA Transactions 71 (2017) 68–75
Fig. 7. Delay-dependent scaled consensus seeking under our proposed algorithm.
complicated, and scaled consensus algorithm is designed and
analyzed for the agents with diverse disturbances and commu-
nication delay. However, the results herein are just for the agents
under fixed topology and time-invariant identical communication
delay. In our future work, thus, we will propose some proper
analysis methods to gain the convergence conditions for the het-
erogeneous multi-agent systems with time-varying disturbances,
switching topologies and time-varying communication delays.
Acknowledgements
This work was supported by the National Natural Science
Foundation of China (Grant No. 61473138 and 61104092), Natural
Science Foundation of Jiangsu Province (Grant No. BK20151130),
Six Talent Peaks Project in Jiangsu Province (Grant no. 2015-DZXX-
011), and Fundamental Research Funds for the Central Universities
(Grant No. JUSRP51407B).
References
[1] Jadbabaie A, Lin J, Morse AS. Coordination of groups of mobile autonomous
agents using nearest neighbor rules. IEEE Trans Autom Control 2003;48
(6):988–1001.
[2] Olfati-Saber R, Murray R. Consensus problems in networks of agents with
switching topology and time-delays. IEEE Trans Autom Control 2004;49
(9):1520–33.
[3] Hong Y, Gao L, Cheng D, Hu J. Lyapunov-based approach to multiagent systems
with switching jointly connected interconnection. IEEE Trans Autom Control
2007;52(5):943–8.
[4] Ren W. Multi-vehicle consensus with a time-varying reference state. Syst
Control Lett 2007;56(7–8):474–83.
[5] Ren W, Atkins E. Distributed multi-vehicle coordinated control via local in-
formation exchange. Int J Robust Nonlinear Control 2007;17(10–11):1002–33.
[6] Qin J, Gao H, Zheng W. Second-order consensus for multi-agent systems with
switching topology and communication delay. Syst Control Lett 2011;60
(6):390–7.
[7] Yu W, Chen G, Cao M. Some necessary and sufficient conditions for second-
order consensus in multi-agent dynamical systems. Automatica 2010;46
(6):1089–95.
[8] Cepeda-Gomez R, Olgac N. Stability analysis for the group dynamics consensus
with time delayed communications. Eur J Control 2012;18(5):456–68 [Sep-
tember-October].
[9] Liu C-L, Liu F. Consensus problem of delayed linear multi-agent systems:
analysis and design.Singapore: Springer; 2017. p. 32–42.
[10] Sun YG, Wang L, Xie G. Average consensus in networks of dynamic agents with
switching topologies and multiple time-varying delays. Syst Control Lett
2008;57(2):175–83.
[11] Liu C-L, Liu F. Stationary consensus of heterogeneous multi-agent systems
with bounded communication delays. Automatica 2011;47(9):2130–3.
[12] Liu C-L, Liu F. Dynamical consensus seeking of heterogeneous multi-agent
75
systems under input delays. Int J Commun Syst 2013;26(10):1243–58.
[13] Hu H, Yu W, Xuan Q, Zhang C, Xie G. Group consensus for heterogeneous
multi-agent systems with parametric uncertainties. Neurocomputing
2014;142:383–92 [October, SI].
[14] Liu Y, Min H, Wang S, Liu Z, Liao S. Distributed consensus of a class of networked
heterogeneous multi-agent systems. J Frankl Inst 2014;351(3):1700–16.
[15] Lee DJ, Spong MK. Agreement with non-uniform information delays. In: Pro-
ceedings of the 2006 American control conference. Minneapolis, MN, United
States; 2006. p. 756–61.
[16] Kim H, Shim H, Seo J. Output consensus of heterogeneous uncertain linear
multi-agent systems. IEEE Trans Autom Control 2011;56(1):200–6.
[17] Dong W, Felipe DLT, Xing Y. Distributed output tracking control of hetero-
geneous linear agents. In: Proceedings of the 2014 American control con-
ference, Portland, OR, United States; 2014. p. 4677–4682.
[18] Zhu J. Stabilization and synchronization for a heterogeneous multi-agent
system via harmonic control. Syst Control Lett 2014;66:1–7 (April).
[19] Muenz U, Papachristodoulou A, Allgoewer F. Delay robustness in non-identical
multi-agent systems. IEEE Trans Autom Control 2012;57(6):1597–603 (June).
[20] Tian Y-P, Zhang Y. High-order consensus of heterogeneous multi-agent sys-
tems with unknown communication delays. Automatica 2012;48(6):1205–12.
[21] Liu C-L, Liu F. Adaptive consensus algorithms for heterogeneousnfirst-order
multi-agent systems with diverse nominal velocities. Trans Inst Meas Control
2016;38(11):1401–8.
[22] Kim J, Yang J, Kim J, Shim H. Practical consensus for heterogeneous linear
time-varying multi-agent systems. In: Proceedings of the 12th international
conference on control, automation and systems. Jeju Island, Korea' 2012. p.
23–28.
[23] Roy S. Scaled consensus. Automatica 2015;51:259–62 (January).
[24] Yu J, Wang L. Group consensus in multi-agent systems with switching
topologies and communication delays. Syst Control Lett 2010;59(6):340–8.
[25] Meng D, Jia Y. Scaled consensus problems on switching networks. IEEE Trans
Autom Control 2016;61(6):1664–9.
[26] Aghbolagh HD, Ebrahimkhani E, Ghiasi AR. Scaled consensus of multi-agent
systems under time-varying delay. In: Proceedings of the 2nd international
conference on knowledge-based engineering and innovation. Hehran, Iran;
2015. p. 729–32.
[27] Hong Y, Hu J, Gao L. Tracking control for multi-agent consensus with an active
leader and variable topology. Automatica 2006;42(7):1177–82.
[28] Liu C-L, Tian Y-P. Formation control of multi-agent systems with hetero-
geneous communication delays. Int J Syst Sci 2009;40(6):627–36.
[29] Liu C-L, Liu F. Delayed-compensation algorithm for second-order leader-fol-
lowing consensus seeking under communication delay. Entropy 2015;17
(6):3752–65.
[30] Wen G, Yu W, Chen MZQ, Yu X, Chen G.  ∞ Pinning synchronization of di-
rected networks with aperiodic sampled-data communications. IEEE Trans
Circuits Syst I: Regul Pap 2014;61(11):3245–55.
[31] Wen G, Zhao Y, Duan Z, Yu W, Chen G. Containment of higher-order multi-
leader multi-agent systems: a dynamic output approach. IEEE Trans Autom
Control 2016;61(4):1135–40 (April).
[32] Lin Z, Francis B, Maggiore M. Necessary and sufficient graphical conditions for
formation control of unicycles. IEEE Trans Autom Control 2005;50(1):121–7.
[33] Desoer CA, Wang YT. On the generalized Nyquist stability criterion. IEEE Trans
Autom Control 1980;25(2):187–96.
[34] Lestas I, Vinnicombe G. Scalable robustness for consensus protocols with
heterogeneous dynamics. In: Proceedings of IFAC world congress, Prague,
Czechrepublic; 2005. p. 185–90.
[35] Parks PC, Hahn V. Stability theory.Englewood Cliffs, NJ: Prentice Hall; 1993. p.
19–20.