H2 suboptimal leader-follower consensus

control of multi-agent systems ⋆

Yuan Gao ∗ Junjie Jiao ∗ Sandra Hirche ∗
∗
Chair of Information-oriented Control, TUM School of Computation,
Information and Technology, Technical University of Munich, 80333
Munich, Germany (e-mail:{ge54sem, junjie.jiao, hirche}@tum.de)

Abstract: In this paper, we investigate the distributed H2 suboptimal leader-follower consensus

control problem for linear multi-agent systems using dynamic output feedback. By considering
an autonomous leader, a number of followers, and an associated H2 cost functional, we aim to
design a distributed protocol to ensure that the leader-follower consensus is achieved while the
associated H2 cost is smaller than an a priori given upper bound. To this end, we first show that
the H2 suboptimal leader-follower consensus control problem can be equivalently derived as the
H2 suboptimal control problem of a set of independent systems. Based on this, we then present
a design method for computing a distributed protocol. The computation of the feedback gains
involves two Riccati inequalities whose dimension matches the state dimension of the agents. A
simulation example is provided to demonstrate the performance of the proposed protocol.

Keywords: Distributed control, H2 optimal control, multi-agent systems, suboptimal control,

leader-follower systems.

1. INTRODUCTION lem. For linear multi-agent systems, the leader-follower

consensus control using static state feedback has been
The recent two decades have seen a significant increase in studied in Liu et al. (2015) by considering the H∞ per-
interest in distributed control for multi-agent networks due formance region. Instead of considering the performance
to their broad range of potential applications. A variety region, the case of prescribed H∞ disturbances attenuation
of distributed control scenarios have been investigated, level was considered in Zhang et al. (2017) using a static
including consensus (Olfati-Saber and Murray (2004)), state protocol to achieve leader-follower consensus of linear
containment control (Li et al. (2013)), and formation con- discrete-time multi-agent systems with the switching con-
trol (Oh et al. (2015)). As one of the fundamental research nected topologies. In contrast to only considering system-
problems of multi-agent systems, consensus control can be theoretic notions, a graph-theoretic approach using static
classified into leaderless consensus (Li et al. (2009)) and state protocol to the H∞ performance of leader following
leader-follower consensus (Ni and Cheng (2010)) according consensus dynamics was proposed in Pirani et al. (2019).
to whether a leader exists or not. However, the works discussed above focus on the H∞
performance index, which measures the robustness of sys-
In this paper, we study leader-follower consensus control, tems to external disturbances in the worst-case scenario.
which means that the states of followers should follow Furthermore, the works above consider static state feed-
the leader’s state. Leader-follower consensus control has back cases, while in practice only output measurements of
been studied in the literature for single integrator agent agents are available.
systems with undirected graphs (Jadbabaie et al. (2003))
and directed graphs (Ren and Beard (2005)), for second- Meanwhile, there are several efforts dealing with the H2
order follower-agent systems with a switching topology performance of leaderless multi-agent systems, where the
(Hong et al. (2008)), for general linear systems under H2 performance index indicates the error energy of the
switching interaction topologies (Ni and Cheng (2010)) system subjected to external disturbance. The leaderless
and directed fixed topologies (Li and Duan (2017)). consensus problem using static state protocols has been
studied with undirected graphs (Li et al. (2011)) by con-
In practice, agent dynamics are subjected to external sidering H2 performance regions and with directed com-
disturbances, potentially leading to a deterioration of the munication graphs (Wang et al. (2014)) by guaranteeing
performance of the multi-agent system. In the literature, H2 performance index. Unlike considering performance re-
many efforts have been devoted to addressing H∞ perfor- gions for the robustness of systems, suboptimal distributed
mance guarantees in the leader-follower consensus prob- protocols based on static state (Jiao et al. (2018)) and
⋆ The authors acknowledge the financial support by the European on dynamic output feedback (Jiao et al. (2020)) were
Union’s Horizon 2020 research and innovation programme under established to minimize a given H2 cost criterion while
the Marie Sklodowska-Curie grant agreement no. 899987 and by achieving consensus of multi-agent systems. However, in
the Federal Ministry of Education and Research of Germany in the the context of distributed H2 control, little attention was
programme of “Souverän. Digital. Vernetzt.”. Joint project 6G-life, paid to the leader-follower consensus with a leader.
project identification number: 16KISK002.
Motivated by the above, in this paper we study the dis- followers, and one agent indexed by N called the leader.
tributed H2 control problem of leader-follower multi-agent The dynamic of the leader is represented by
systems using dynamic output feedback. More concretely, ẋN (t) = AxN (t) , yN (t) = C1 xN (t) , zN (t) = C2 xN (t) .
this paper extends the results from Jiao et al. (2020) for (1)
leaderless systems to the leader-follower case. To this end,
for a given leader-follower system, we first introduce a The dynamics of followers are identical and denoted by
suitable performance output and, subsequently, an associ- ẋi (t) = Axi (t) + Bui (t) + Edi (t) ,
ated H2 cost functional. The goal is to design distributed yi (t) = C1 xi (t) + D1 di (t) , (2)
protocols by dynamic output feedback such that the multi- zi (t) = C2 xi (t) + D2 ui (t) , i = 1, . . . , N − 1,
agent system achieves leader-follower consensus while min-
where xi ∈ Rn , yi ∈ Rr , zi ∈ Rp , ui ∈ Rm and di ∈ Rq are,
imizing the associated H2 cost functional. Due to the com-
respectively, the state, the measured output, the output
munication constraints among the agents, this problem is
to be controlled, the coupling input and the unknown
non-convex, and up to now, a closed-form solution has
external disturbance of ith follower. The matrices A, B,
not been given in the literature. Therefore, we seek an
C1 , C2 , D1 , D2 and E are of compatible dimensions. In
alternative involving only suboptimality.
this paper, we assume that the pair (A, B) is stabilizable
The outline of this paper is as follows. Section 2 provides and the pair (C1 , A) is detectable.
some notations and graph theory. In Section 3, we formu-
Throughout this paper, we assume that each follower
late the H2 suboptimal leader-follower consensus problem.
has access to relative output measurements concerning
We then design distributed H2 suboptimal protocols in
its neighbors and consider output feedback protocols. In
Section 4. A simulation example is presented in Section 5,
particular, following Trentelman et al. (2013), we propose
followed by the conclusion in Section 6.
the observer-based distributed dynamic protocol
2. PRELIMINARIES N
X
ẇi = (A − GC1 )wi + aij [BF (wi − wj ) + G(yi − yj )] ,
2.1 Notation j=1
ui = F wi , i = 1, . . . , N − 1, (3)
The field of real numbers is denoted by R, the space of n n×r m×n
where G ∈ R and F ∈ R are local feedback gain
dimensional real vectors is denoted by Rn , and the space
matrices to be designed and the state wi takes the role of
of m × n real matrices is denoted by Rm×n . For vectors PN
and matrices, the superscript ⊤ means transposition. In estimate the relative state j=1 aij (xi −xj ) and note that
represents the identity matrix of dimension n × n. The wN = 0, where aij is the ijth entry of the adjacency matrix
trace of a square matrix A is denoted by tr(A). It is said A associated with graph G, which satisfies the following
that a matrix is Hurwitz (or stable) if all its eigenvalues standard assumption.
have negative real parts. In the case of a symmetric matrix Assumption 1. The leader receives no information from
P , we denoted P > 0 if P is positive definite and P < 0 any follower. The leader’s state is available to at least one
if P is negative definite. The n × n diagonal matrix with follower, and the communication graph between the N − 1
d1 , . . . , dn on the diagonal is denoted by diag(d1 , . . . , dn ). followers is connected, simple, and undirected.
For matrices M1 , . . . , Mm , let blockdiag(M1 , . . . , Mm ) be
the block diagonal matrix with diagonal blocks Mi . A ⊗ B Since the leader has no neighbors, the Laplacian matrix
denotes the Kronecker product of matrix A and B. associated with graph G can be partitioned as
 
L1 L2
2.2 Graph theory L= , (4)
01×(N −1) 0
A directed graph is denoted by G = (V, E) with node set where L1 ∈ R(N −1)×(N −1) , L2 ∈ R(N −1)×1 .
V = {1, . . . , N } and edge set E = {e1 , . . . , eM } satisfying Lemma 2. (Meng et al. (2010)). Under Assumption 1, L1
E ⊂ V × V. The edge from node i to node j is represented is positive definite, and subsequently, all the eigenvalues
by the pair (i, j) ∈ E. We say a graph is undirected if of L1 have positive real parts.
(i, j) ∈ E implies (j, i) ∈ E, and a graph is simple if
(i, i) ∈
/ E which means no self-loops. The adjacency matrix Foremost, we want the protocol (3) to solve the leader-
A = [aij ] ∈ RN ×N of the graph G is defined as aii = 0, follower consensus control problem for agents (1) and (2).
aij = 1 if (j, i) ∈ E, and aij = 0 otherwise. Subsequently, In the context of leader-follower consensus control, it is
the Laplacian matrix L = [Lij ] ∈ RN ×N of graph G is desired that the states of followers follow the leader’s state,
PN so we are interested in the differences between the states
defined as Lii = j=1 aij and Lij = −aij . It can also
be written into a compact form as L = D − A, where of leader and followers. Therefore, we introduce the new
D = diag(d1 , . . . , dN ) is the degree matrix of graph G error state variable for each follower as ei = xi −xN , where
PN the leader-follower consensus is achieved if ei = 0, i.e.,
with di = j=1 aij . Furthermore, the Laplacian matrix xi → xN as t → ∞ for all i = 1, . . . , N − 1.
L of an undirected graph is symmetric and only has real
nonnegative eigenvalues. Meanwhile, in the context of distributed H2 optimal con-
trol for multi-agent systems, we are interested in the
3. PROBLEM FORMULATION differences in the leader’s and followers’ output values.
Therefore, the performance output variable is defined as
We consider a leader-follower multi-agent system that ϵi = zi − zN , i = 1, . . . , N − 1, which reflects the output
consists of N − 1 agents indexed by 1, ..., N − 1, called the disagreement between leader and followers.
 Consider U ∈ RN −1×N −1 as an orthogonal matrix
⊤ ⊤
Denote e = e⊤ ⊤
, ϵ = ϵ⊤ ⊤
 
1 , . . . , eN −1 1 , . . . , ϵN −1 , ξ =
 ⊤ ⊤ ⊤ ⊤ ⊤
  ⊤ ⊤
⊤ which diagonalizes the matrix L1 , i.e., U ⊤ L1 U = Λ =
y − yN , . . . , yN −1 − yN , u = u1 , . . . , uN −1 , d = diag(λ1 , . . . , λN −1 ), where λi > 0, i = 1, . . . , N − 1 are the
 1⊤ ⊤ ⊤
d1 , . . . , d ⊤ , and w = w1⊤ , . . . , wN ⊤

N −1 −1 . Then, the eigenvalues of L1 . By using the state transformation:
dynamics of the error system can be written as    ⊤
ê U ⊗ In 0
 
e
ė = (IN −1 ⊗ A)e + (IN −1 ⊗ B)u + (IN −1 ⊗ E)d, = , (10)
ŵ 0 U ⊤ ⊗ In w
ξ = (IN −1 ⊗ C1 )e + (IN −1 ⊗ D1 )d, (5) the controlled error system (7) becomes
ϵ = (IN −1 ⊗ C2 )e + (IN −1 ⊗ D2 )u.
ê˙
    
IN −1 ⊗ A IN −1 ⊗ BF ê
Correspondingly, the protocol (3) can be written as ˙ = Λ ⊗ GC1 IN −1 ⊗ (A − GC1 ) +Λ ⊗ BF ŵ
ŵ
ẇ = (IN −1 ⊗ (A − GC1 ))w + (L1 ⊗ BF )w + (L1 ⊗ G)ξ,
U⊤ ⊗ E
 
u = (IN −1 ⊗ F )w. (6) + d,
U ⊤ L1 ⊗ GD1
By interconnecting the error system (5) with the dynamic  
protocol (6), we obtain the controlled error system ê
ϵ = [U ⊗ C2 U ⊗ D2 F ] . (11)
     ŵ
ė I ⊗A IN −1 ⊗ BF e
= N −1 Note that after the transformation (10), the impulse re-
ẇ L1 ⊗ GC1 IN −1 ⊗ (A − GC1 ) + L1 ⊗ BF w
  sponse matrix from the disturbance input d to the output
IN −1 ⊗ E ϵ still equals the impulse response matrix (8). To proceed,
+ d,
L1 ⊗ GD1 the following N −1 auxiliary linear systems are introduced:
 
ϵ = [IN −1 ⊗ C2 IN −1 ⊗ D2 F ]
e
. (7) ee˙ i (t) = Ae
ei (t) + Be
ui (t) + E dei (t) ,
w i = 1, . . . , N − 1,
  ξei (t) = C1 eei (t) + D1 dei (t) ,
IN −1 ⊗ A IN −1 ⊗ BF ϵi (t) = C2 eei (t) + D2 uei (t) ,
Denote Ao = , e
L1 ⊗ GC1 IN −1 ⊗ (A − GC1 ) + L1 ⊗ BF

IN −1 ⊗ E
 where eei ∈ Rn , u ei ∈ Rm , dei ∈ Rq , ξei ∈ Rr and
Co = [IN −1 ⊗ C2 IN −1 ⊗ D2 F ] , Eo = p
L1 ⊗ GD1
. The ϵi ∈ R are, respectively, the state, the coupling input, the
e
external disturbance, the measured output and the output
impulse response matrix for the controlled error system to be controlled of the ith auxiliary system. By using the
(7) from the external disturbance d to the performance associated dynamic feedback controllers
output ϵ is then equal to
w˙ i = Awei + Be ui + G(ξei − C1 w ei ),
TF,G (t) = Co eAo t Eo . (8) e
(12)
u ei , i = 1, . . . , N − 1.
ei = λi F w
Thus, the associated cost functional H2 is given by
Z ∞
 ⊤  where λi > 0, i = 1, . . . , N − 1 are the eigenvalues of L1 ,
J(F, G) := tr TF,G (t)TF,G (t) dt. (9) the closed-loop systems can be written as
0
ee˙ i
      
A λi BF eei E
which measures the performance of the system (7) as the ˙ i = GC1 A − GC1 + λi BF w + de ,
w ei GD1 i
square of the L2 - norm of its impulse response. The H2 e
 
optimal leader-follower consensus control problem is a non- ee
convex optimization problem due to the communication ϵi = [C2 λi D2 F ] i , i = 1, . . . , N − 1. (13)
w
e
ei
constraints among the agents, and it is not yet known  
whether a closed-form solution exists in the literature. Denote A ei = A λi BF
,C ei = [C2 λi D2 F ],
Alternatively, we solve a version of the problem that GC1 A − GC1 + λi BF
 
requires only suboptimality. Eei = E . The impulse response matrix for each system
Definition 3. The protocol (3) is said to solve the dis- GD1
tributed H2 suboptimal leader-follower consensus problem (13) from the disturbance dei to the output e ϵi is Tei,F,G (t) =
for the multi-agent system (1) and (2) if, Cei eAei t E ei . The associated H2 cost functional is given by
R∞ h ⊤ i
• whenever the external disturbances of all followers are Ji (F, G) := 0 tr Tei,F,G (t)Tei,F,G (t) dt, i = 1, . . . , N − 1.
equal to zero, i.e., d = 0, we have xi → xN and Consequently, the following theorem holds.
wi → 0 for all i = 1, . . . , N − 1.
• J(F, G) < γ, where γ is a given upper bound. Theorem 4. Let G ∈ Rn×r and F ∈ Rm×n . Assume that
D1 E ⊤ = 0, D2⊤ C2 = 0, D1 D1⊤ = Ir and D2⊤ D2 = Im .
The problem that we want to address is the following: The dynamic protocol (3) with gain matrices F, G achieves
Problem 1. Let γ > 0. Design local feedback gain matrices leader-follower consensus for the agents (1) and (2) if and
G ∈ Rn×r and F ∈ Rm×n such that the dynamic protocol only if the controllers (12) with the same F, G internally
(3) achieves leader-follower consensus and J(F, G) < γ. stabilize all N − 1 systems in (13). Moreover, we have
PN −1
J(F, G) := i=1 Ji (F, G).
4. PROTOCOL DESIGN
Proof. It can be derived from (10) that ê = 0 and ŵ = 0
if and only if ei = 0, wi = 0, i.e., xi → xN and wi → 0 for
In this section, we deal with Problem 1 and establish a
all i = 1, . . . , N − 1. Hence, the leader-follower consensus
design method for obtaining gain matrices F and G.
problem is solved if and only if limt→∞ ê(t) = 0 and
According to Assumption 1 and Lemma 2, L1 is pos- limt→∞ ŵ(t) = 0. Recall that U ⊤ L1 U = Λ and by using
itive definite, which implies that L1 is diagonalizable. two transformations
 dˆ = (U ⊤ ⊗ In )d, ϵ̂ = (U ⊤ ⊗ In )ϵ, (14) be easily relaxed to the regularity condition D1 D1⊤ > 0
the controlled error system (11) can be transferred into as and D2⊤ D2 > 0. By applying Theorem 4, the distributed
H2 suboptimal leader-follower consensus problem for the
ê˙
    
IN −1 ⊗ A IN −1 ⊗ BF ê multi-agent system (1) and (2) can be recast into H2
˙ = Λ ⊗ GC1 IN −1 ⊗ (A − GC1 ) + Λ ⊗ BF ŵ
ŵ suboptimal control problems of N −1 independent systems
(13) by using dynamic output feedback controllers (12).
 
IN −1 ⊗ E ˆ
+ d,
Λ ⊗ GD1 Next, we show that, for given gain matrices G ∈ Rn×r
and F ∈ Rm×n , the following lemma is presented to solve
 
ê
ϵ̂ = [IN −1 ⊗ C2 IN −1 ⊗ D2 F ] . (15) the problem of H2 suboptimal control for N − 1 systems
ŵ
  (13), i.e., all N − 1 systems are internally stable, while
I ⊗A IN −1 ⊗ BF PN −1
Denote Âo = N −1 , i=1 Ji (F, G) < γ.
Λ ⊗ GC1 IN −1 ⊗ (A − GC1 ) + Λ ⊗ BF
  Lemma 5. The dynamic controllers (12) internally stabi-
I ⊗E
Ĉo = [IN −1 ⊗ C2 IN −1 ⊗ D2 F ], Êo = N −1
PN −1
. It is lize all N − 1 systems (13) and i=1 Ji (F, G) < γ if and
Λ ⊗ GD1
easily seen that for i = 1, . . . , N − 1 the decomposed only if there exist Pi > 0, i = 1, . . . , N − 1 and Q > 0
satisfying
subsystems (Âoi , Eˆoi , Cˆoi ) in (15) and auxiliary systems
(A + λi BF )⊤ Pi + Pi (A + λi BF )
(Aei , Eei , C
ei ) in (13) are isomorphic. So limt→∞ ê(t) = 0
and limt→∞ ŵ(t) = 0 if and only if ee1 = · · · = eeN −1 = 0, + (C2 + λi D2 F )⊤ (C2 + λi D2 F ) < 0, (16)
we1 = · · · = w eN −1 = 0. AQ + QA − QC1⊤ C1 Q + EE ⊤ < 0,
⊤
(17)
 
w N −1
Let ρi = eei − w fi and by using the transformation i =
f X
tr C1 QPi QC1⊤ + tr C2 QC2⊤ < γ.



ρei (18)
     i=1
0 In eei A λ i BF
, then Ai = e in
In −In w fi GC1 A − GC1 + λi BF Proof. The proof follows from the proof of (Jiao et al.,
 
A + λi BF −GC1 2020, Lemma 2) by taking Ā = A, B̄ = λi B, C̄1 = C1 ,
(13) will be transformed into Aρ i = f .
0 A − GC1 D̄1 = D1 , C̄2 = C2 , D̄2 = λi D2 , and Ē = E and thus it is
In this regard, it is obvious that the state of eei and w ei , for omitted here. □
i = 1, . . . , N −1 converge asymptotically to zero if and only
if the matrices A+λi BK and A−GC1 of the N −1 systems Note that, however, Lemma 5 does not yet provide a
are stable. Subsequently, the leader-follower consensus is method for computing matrices F, G. The following the-
achieved. orem then provides a design method for finding suitable
PN −1 matrices F, G.
Next, we prove J(F, G) := i=1 Ji (F, G). Let F ,G be Theorem 6. Assume that D1 E ⊤ = 0, D2⊤ C2 = 0, D1 D1⊤ =
such that matrices A+λi BF , matrix A−GC1 are Hurwitz. Ir and D2⊤ D2 = Im . Let γ > 0. Consider the controlled
Note that U ⊤Co U = Ĉo , U ⊤ Eo  U = Êo , U ⊤ Ao U = Âo , error system (7) with associated H2 cost functional (9).
⊤
U ⊗ In 0 Let Q > 0 satisfies
where U ⊤ = . Then, by substituting
0 U ⊤ ⊗ In AQ + QA⊤ − QC1⊤ C1 Q + EE ⊤ < 0. (19)
(8) in (9), we have
Z ∞ Furthermore, consider the following two cases:
 ⊤ 
J(F, G) := tr TF,G (t)TF,G (t) dt (i) if 0 < c < λ1 +λ2N −1 , where λ1 is the smallest
Z0 ∞ eigenvalue and λN −1 is the largest eigenvalue of L1 .
tr (Co eAo t Eo )⊤ (Co eAo t Eo ) dt
 
= Then there exists P > 0 satisfying
Z0 ∞ h i A⊤ P + P A + (c2 λ21 − 2cλ1 )P BB ⊤ P + C2 ⊤ C2 < 0.
= tr U (Ĉo eÂo t Êo )⊤ (Ĉo eÂo t Êo )U ⊤ dt. (20)
0
(ii) if λ1 +λ2N −1 ≤ c < λN2−1 , then there exists P > 0
Recall that U ⊤ L1 U = Λ = diag(λ1 , . . . , λN −1 ), D1 E ⊤ =
0, D2⊤ C2 = 0, D1 D1⊤ = Ir , D2⊤ D2 = Im , the decom- satisfying
posed subsystems (Âoi , Eˆoi , Cˆoi ) in (15) and auxiliary A⊤ P +P A+(c2 λ2N −1 −2cλN −1 )P BB ⊤ P +C2 ⊤ C2 < 0.
systems (A ei , Eei , C
ei ) in (13) are isomorphic. Consequently, (21)
⊤ Aˆ ⊤ t ⊤ ˆ
tr(Eˆo i e o i Cˆo i Cˆoi eAo i t Eˆoi ) = tr(E e ⊤ eAe⊤
i tC e⊤ C ei eAei t E
ei ), In both cases, if P and Q also satisfy
i i
for i = 1, . . . , N − 1. Therefore, γ
tr C1 QP QC1⊤ + tr C2 QC2⊤ <



, (22)
Z ∞ NX −1 N −1
J(F, G) := ei⊤ eAe⊤
tr(U E ei⊤ C
i tC ei U ⊤ )dt
ei eAei t E then the protocol (3) with F := −cB ⊤ P and G :=
0 i=1 QC1 ⊤ achieves leader-follower consensus for the agents (1)
Z ∞N
X −1 N
X −1 and (2) and the protocol is suboptimal, i.e., J(F, G) < γ.
= tr(Tei,F,G (t)⊤ Tei,F,G (t))dt = Ji (F, G).
0 i=1 i=1 Proof. First, note that (19) is exactly (17). For case (ii)
The proof is now complete. □ above, using the upper and lower bound on c, c2 λ2N −1 −
2cλN −1 < 0 can be verified. Since the Riccati inequality
Note that, in Theorem 4, the assumptions D1 D1⊤ = Ir (21) has a positive definite solution P . For i = 1, . . . , N −1,
and D2⊤ D2 = Im are made to simplify notation and can taking Pi = P and F = −cB ⊤ P in (16) immediately yields
 5. SIMULATION EXAMPLE

This section provides a simulation example to validate the

performance of our proposed protocol by dynamic output
feedback. Consider a leader-follower multi-agent system
consisting of one leader with the 
form (1)
 and six followers

−2 2 0
with the form (2), where A = , B = , E =
−1 1 1
     
0.6 0 1 1.2 0
, C2 = , D2 = , C1 = [1 0] , D1 = [0 1].
1 0 0 0 1
Fig. 1. The communication topology between the leader
The pair (A, B) is stabilizable and the pair (C1 , A) is
and the followers
detectable. We also have D1 E ⊤ = [0 0], D2⊤ C2 = [0 0]
(A − cλi BB ⊤ P )⊤ P + P (A − cλi BB ⊤ P ) and D1 D1⊤ = 1, D2⊤ D2 = 1.
+ (C2 − λi D2 B ⊤ P )⊤ (C2 − λi D2 B ⊤ P ) < 0. For illustration, let the communication graph G be given
by Figure 1, where node 7 is the leader and the others are
Recall the conditions D2 ⊤ C2 = 0 and D2 ⊤ D2 = Im followers. Correspondingly, due to the specific partition
this yields (A − cλi BB ⊤ P )⊤ P + P (A − cλi BB ⊤ P ) + form of the Laplacian matrix L (4) associated with graph
c2 λ2i P BB ⊤ P + C2 ⊤ C2 < 0. G, the smallest and largest eigenvalue of the matrix L1 are
Since c2 λ21 − 2cλ1 ≤ c2 λ2i − 2cλi ≤ c2 λ2N −1 − 2cλN −1 < λ1 = 0.1088 and λ6 = 4.2784. Now we use the method
0 and λi ≤ λN −1 for i = 1, . . . , N − 1, the positive proposed in Theorem 6 to compute the gain matrices F, G
definite solution P of (21) also satisfies the N − 1 Riccati of the dynamic output feedback protocol (3) to solve the
inequalities H2 suboptimal leader-follower consensus control problem.
Let the desired upper bound for the H2 cost (9) be γ = 92.
A⊤ P + P A + (c2 λ2i − 2cλi )P BB ⊤ P + C2 ⊤ C2 < 0. Following Theorem 6, we first compute a solution P > 0
Next, it follows from (22) that also (18) holds. By Lemma in case (ii) by solving
5 then, all N − 1 systems (13) are internally stabilized A⊤ P +P A+(c2 λ26 −2cλ6 )P BB ⊤ P +C2 ⊤ C2 +δI2 = 0 (23)
PN −1
and i=1 Ji (F, G) < γ. Subsequently, it follows from 2
with δ = 0.001. In addition, we choose c = λ1 +λ = 0.4559,
Theorem 4 that the protocol (3) achieves leader-follower 6
which is the ’best’ choice to find a small upper bound γ in
consensus for the agents (1) and (2) while J(F, G) < γ.
the sense as explained in Remark 7. Then, by solving (23)
For case (i) above, the proof is similar and is omitted. □
in Matlab with the command icare, we compute the gain
Remark 7. Theorem 6 states that by choosing suitable c, matrix F = (1.3414, −4.5669).
P and Q, the distributed dynamic output protocol with
gain matrices F = −cB ⊤ P and G = QC1 ⊤ is suboptimal. Next, we compute a solution Q > 0 in (19) by solving
Thus, for this suboptimal problem the question arises: how AQ + QA⊤ − QC1⊤ C1 Q + EE ⊤ + ηI2 = 0
to select the upper bound γ as
small as possible such that with η = 0.001 in Matlab using command icare,the
tr C1 QP QC1⊤ + tr C2 QC2⊤ < Nγ−1 ? The point can be


gain matrix G = (1.0407, 1.2213)⊤ . Moreover, we compute
easily made that, in general, smaller
P and Q lead to 6(tr(C1 QP QC1 ⊤ ) + tr(C2 QC2 ⊤ )) = 91.0974, which is
smaller tr C1 QP QC1⊤ + tr C2 QC2⊤ , and consequently,


indeed smaller than the upper bound γ = 92. Then by
the smaller feasible given γ. To find a small feasible γ, we using the command norm(sys,2) in Matlab, the actual
could try to find P and Q as small as possible. With η > 0, H2 norm of the controlled error system (7) is computed
we can establish a equality from (19) as ||TF,G ||H2 = 8.7388, which is indeed smaller than
to be √
√
AQ + QA⊤ − QC1⊤ C1 Q + EE ⊤ + ηIN = 0. γ = 92 = 9.5917.
By using the standard argument, it can be shown that Q In the following, we compare the performance of our
decreases as η decreases. Consequently, if we chose η > 0 protocol with that of the proposed protocol in Li and
very close to 0, we can find a small solution for Q(η) > 0. Duan (2017). The corresponding feedback gains of the
Similarly, a small solution P (c, δ) > 0 with δ > 0 for the protocol in Li and Duan (2017)) are computed as K̄ =
two cases (20) and (21) can be founded by establishing two (0.8250, −4.2000) and F̄ = (−29.2078, −16.8628)⊤ . The
equalities as follows associated actual H2 norm of the controlled system (7) is
A⊤ P + P A − r1 P BB ⊤ P + C2 ⊤ C2 + δIn = 0, computed to be ||TK̄,F̄ ||H2 = 23.4079. Since ||TK̄,F̄ ||H2 =
23.4079 > ||TF,G ||H2 = 8.7388, it is shown that our
A⊤ P + P A − r2 P BB ⊤ P + C2 ⊤ C2 + δIn = 0,
protocol outperforms the protocol in Li and Duan (2017).
where r1 = (−c2 λ21 +2cλ1 ) and r2 = (−c2 λ2N −1 +2cλN −1 ).
As an illustrative example, we take the initial states
Obviously, the larger r1 (or r2 ) and the smaller δ, the ⊤ ⊤
smaller P is. It can be computed that the maximum of of the agents to be x10 = [−13 10] , x20 = [5 12] ,
⊤ ⊤ ⊤
r1 is obtained when c∗ = λN1−1 and the maximum of r2 is x30 = [−9 −15] , x40 = [18 11] , x50 = [−2 −4] ,
⊤ ⊤
obtained when c∗ = λ11 . Therefore, for both two cases, if x60 = [12 12] and x70 = [2.5 7.5] . To further compare
we choose δ > 0 very close to 0 and c = λ1 +λ2N −1 , we find the performance of our proposed protocol with that in
Li and Duan (2017), the same white noise d with an
the ‘best’ solution to the Riccati inequalities (20) and (21) amplitude ranging between -4 and 4 is applied. In Figure
as explained above. 2, the trajectories of the performance output ϵ using our
Fig. 2. Trajectories of the performance outputs ϵi1 , ϵi2 for i = 1, . . . , 6 using our proposed protocol (plot on the left)
and those of using the protocol in Li and Duan (2017) (plot on the right)

designed protocol and the protocol in Li and Duan (2017) Li, Z., Ren, W., Liu, X., and Fu, M. (2013). Distributed
are plotted. It is shown that our protocol has a better containment control of multi-agent systems with general
performance than that of the protocol proposed in Li and linear dynamics in the presence of multiple leaders.
Duan (2017), in the sense that our protocol has a better International Journal of Robust and Nonlinear Control,
tolerance for external disturbances. 23(5), 534–547.
Liu, W., Liu, A.L., and Zhou, S.L. (2015). Distributed H∞
6. CONCLUSIONS control of multi-agent systems with directed networks.
Chinese Physics B, 24(9), 090208.
In this paper, we have investigated the distributed H2 sub- Meng, Z., Ren, W., and You, Z. (2010). Distributed finite-
optimal leader-follower consensus control problem using time attitude containment control for multiple rigid
dynamic output feedback. Consider a multi-agent system bodies. Automatica, 46(12), 2092–2099.
with N agents consisting of an autonomous leader and Ni, W. and Cheng, D. (2010). Leader-following consensus
N − 1 followers, and an associated H2 cost functional of multi-agent systems under fixed and switching topolo-
with a desired upper bound, we have developed a design gies. Systems & Control Letters, 59(3-4), 209–217.
method for computing a distributed protocol that achieves Oh, K.K., Park, M.C., and Ahn, H.S. (2015). A survey of
H2 suboptimal leader-follower consensus, i.e., the states of multi-agent formation control. Automatica, 53, 424–440.
the followers converge to the state of the leader and the Olfati-Saber, R. and Murray, R.M. (2004). Consensus
associated H2 cost is smaller than this given upper bound. problems in networks of agents with switching topology
REFERENCES and time-delays. IEEE Transactions on Automatic
Control, 49(9), 1520–1533.
Hong, Y., Chen, G., and Bushnell, L. (2008). Distributed Pirani, M., Sandberg, H., and Johansson, K.H. (2019).
observers design for leader-following control of multi- A graph-theoretic approach to the H∞ performance of
agent networks. Automatica, 44(3), 846–850. leader–follower consensus on directed networks. IEEE
Jadbabaie, A., Lin, J., and Morse, A.S. (2003). Coordina- Control Systems Letters, 3(4), 954–959.
tion of groups of mobile autonomous agents using near- Ren, W. and Beard, R.W. (2005). Consensus seeking in
est neighbor rules. IEEE Transactions on Automatic multiagent systems under dynamically changing inter-
Control, 48(6), 988–1001. action topologies. IEEE Transactions on Automatic
Jiao, J., Trentelman, H.L., and Camlibel, M.K. (2018). Control, 50(5), 655–661.
A suboptimality approach to distributed H2 optimal Trentelman, H.L., Takaba, K., and Monshizadeh, N.
control. IFAC-Papers OnLine, 51(23), 154–159. (2013). Robust synchronization of uncertain linear
Jiao, J., Trentelman, H.L., and Camlibel, M.K. (2020). multi-agent systems. IEEE Transactions on Automatic
A suboptimality approach to distributed H2 control by Control, 58(6), 1511–1523.
dynamic output feedback. Automatica, 121, 109164. Wang, J., Duan, Z., Li, Z., and Wen, G. (2014). Dis-
Li, Z. and Duan, Z. (2017). Cooperative control of multi- tributed H∞ and H2 consensus control in directed net-
agent systems: a consensus region approach. CRC press. works. IET Control Theory & Applications, 8(3), 193–
Li, Z., Duan, Z., and Chen, G. (2011). On H∞ and H2 201.
performance regions of multi-agent systems. Automat- Zhang, D., Xu, Z., Wang, Q.G., and Zhao, Y.B. (2017).
ica, 47(4), 797–803. Leader–follower H∞ consensus of linear multi-agent sys-
Li, Z., Duan, Z., Chen, G., and Huang, L. (2009). Consen- tems with aperiodic sampling and switching connected
sus of multiagent systems and synchronization of com- topologies. ISA Transactions, 68, 150–159.
plex networks: A unified viewpoint. IEEE Transactions
on Circuits and Systems I: Regular Papers, 57(1), 213–
224.