6746 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 69, NO.

10, OCTOBER 2024

A Framework on Fully Distributed State

Estimation and Cooperative Stabilization
of LTI Plants
Peihu Duan , Yuezu Lv , Senior Member, IEEE, Guanghui Wen , Senior Member, IEEE,
and Maciej Ogorzałek , Fellow, IEEE

Abstract—How to realize high-level autonomy of individ- I. INTRODUCTION

uals is one of the key technical issues to promote swarm
OOPERATION is one of the most significant character-
intelligence of multiagent (node) systems with collective
tasks, while the fully distributed design is a potential way to
achieve this goal. This article works on the fully distributed
C istics of multiagent systems, which can greatly facilitate
their implementation, ranging from collaborative manipulation
state estimation and cooperative stabilization problem of of multiple robots in smart factories [1] to cooperative fire
linear time-invariant (LTI) plants with multiple nodes com-
municating over general directed graphs, and is aimed to monitoring of multiple unmanned aerial vehicles in forests [2].
provide a fully distributed framework for each node to per- Among these applications, one crucial task for agents is to
form cooperative stabilization tasks. First, by incorporating synthesize effective cooperative controllers with limited com-
a novel adaptive law, a consensus-based estimator is de- munication and computation resources. For this issue, plenty of
signed for each node to obtain the plant state based on research efforts have been devoted into distributed control of
its local measurement and local interaction with neighbors,
without using any global information of the communication multiagent systems over the last two decades [3], [4], [5], [6],
topology. Subsequently, a local controller is developed for [7], [8]. In this control scheme, agents are arranged to complete
each node to stabilize the plant collaboratively with per- a collective task only using local information and local commu-
formance guaranteed under mild conditions. Specifically, nication with neighbors. In comparison to the centralized and
the proposed method only requires that the communication decentralized structures, the distributed framework can enhance
graph be strongly connected, and the plant be collectively
controllable and observable. Further, the proposed method the system robustness and scalability [9]. Hence, the distributed
can be applied to pure fully distributed state estimation cooperation deserves in-depth and extensive investigation to
scenarios and modified for noise-bounded LTI plants. Fi- promote the application of multiagent systems.
nally, two numerical examples are provided to show the
effectiveness of the theoretical results.
Index Terms—Cooperative stabilization, fully distributed A. Literature Review and Motivations
state estimation, linear time-invariant (LTI) plants, multia-
gent systems.
Inspired by the well-developed graph theory and control
theory, a large number of existing works formulate the dis-
tributed cooperation behavior design of multiagent systems into
various consensus-based control problems, such as consensus
control [10], containment control [11], formation control [12],
Manuscript received 12 January 2024; accepted 9 March 2024. Date and synchronization [13]. In these studies, the states of all agents
of publication 13 March 2024; date of current version 27 Septem- are required to converge to a common value or maintain a
ber 2024. This work was supported in part by the National Natural
Science Foundation of China through under Grant 62088101, Grant constant error with neighbors. However, agents in a group may
62325304, Grant 62273045, Grant 62073079, Grant U2341213, and need to behave differently for collective tasks, and subsequently
Grant U22B2046; and in part by Beijing Nova Program under Grant the states are completely inconsistent. For example, in the co-
20230484481. Recommended by Associate Editor F. Pasqualetti. (Cor-
responding author: Yuezu Lv.) operative planar carrying task of a huge object by a network
Peihu Duan is with the School of Electrical Engineering and Computer of robots, robots at different positions need to exert different
Science, KTH Royal Institute of Technology, 114 28 Stockholm, Sweden operation forces on the object [14]. In such scenarios, the above
(e-mail: [email protected]).
Yuezu Lv is with the MIIT Key Laboratory of Complex-field Intelligent consensus-based formulation is not applicable. An alternative
Sensing, Beijing Institute of Technology, Beijing 100081, China (e-mail: way is to interpret the collective task as a cooperative control
[email protected]). problem of a multichannel plant, where each channel is managed
Guanghui Wen is with the Department of Systems Science, School
of Mathematics, Southeast University, Nanjing 211189, China (e-mail: by an agent [15]. Moreover, considering the heterogeneous
[email protected]). system dynamics and channel parameters, each agent often has
Maciej Ogorzałek is with the Department of Information Technolo- access to different partial information of the plant state and exerts
gies, Jagiellonian University, 30-348 Krakow, Poland (e-mail: maciej.
[email protected]). its particular input on the plant. This article focuses on this kind
Digital Object Identifier 10.1109/TAC.2024.3376791 of distributed cooperative control problems.
1558-2523 © 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
DUAN et al.: FRAMEWORK ON FULLY DISTRIBUTED STATE ESTIMATION AND COOPERATIVE STABILIZATION OF LTI PLANTS 6747

In recent years, some researches on distributed cooperative B. Contributions

control of a multichannel plant supervised by a multiagent
Motivated by the above observations, this article aims at
system have been reported, generally adopting a “distributed
addressing the fully distributed state estimation and cooperative
state estimation + feedback control” structure [16], [17], [18],
stabilization problem for LTI plants with a multiagent system
[19], [20]. In this structure, each agent first attempts to esti-
under a directed communication graph. Specifically, the plant
mate the plant state via distributed estimators. Further, each
consists of multiple channels and each channel is managed by
agent generates a local feedback control signal based on the
an agent (node). By virtue of the channel, every node equips
estimate to cooperatively fulfill the global control task. As a
with a sensor as well as an actuator. To address this problem,
preliminary attempt, Liu et al. [16], [17] addressed the cooper-
this article focuses on a novel “fully distributed output estimation
ative stabilization and cooperative output regulation problems
+ local state feedback control” framework. First, a distributed
for linear time-invariant (LTI) plants by embedding a class of
state estimator using output estimation is developed such that all
distributed Luenberger observers, respectively. However, the
nodes can have a complete knowledge of the plant state. Subse-
complete communication topology of agents had to be utilized
quently, a local state feedback controller is designed to stabilize
to design the coupling gains in observers. A similar problem
the plant. Compared with literature, this article possesses four
formulated in a discrete-time form was considered in [18].
advantageous features as follows.
Lately, a simultaneous distributed state estimation and control
1) The proposed state estimation and control strategy is fully
framework for LTI plants with dynamics uncertainties and com-
distributed, even without using any global information of
munication constraints was established in [20], but its feasibility
the communication topology graph (Theorems 1 and 2).
condition should be checked by a control center in advance.
Compared with the recent works [16], [17], [18], [19], the
Altogether, albeit the above existing cooperative control algo-
proposed strategy entirely avoids the need for a central
rithms for LTI plants can be performed in a distributed way,
authority.
the algorithm parameters need to be carefully selected by lever-
2) The proposed strategy achieves the fully distributed state
aging global information. In this sense, a central supervisory
estimation and cooperative stabilization of LTI plants un-
is still needed. Besides, the coupling gains designed in [16],
der general directed communication graphs (Theorems 3
[17], [18], [19], and [20] are fixed and large, which usually
and 4). Subsequently, it exhibits wider applications com-
requires a small integration step, leading to a high commu-
pared to its undirected counterparts.
nication frequency among nodes in the practical discrete-time
3) The proposed method can be directly applied to the pure
implementations.
fully distributed state estimation (Theorem 5). Different
To avoid using the central authority and the global graph
from the fixed coupling gains designed in the traditional
information, some fully distributed state estimation approaches
estimators [26], [27], an adaptive coupling gain is adopted
have been proposed for LTI plants, such as distributed Kalman
in this article, which has a smaller value. According
filtering [21], [22], [23] and cooperative observation [24], [25],
to [26], a smaller coupling gain allows for a larger integra-
[26], [27]. In particular, a unified Kalman filter-based distributed
tion step in the discrete-time applications, benefiting the
state estimation framework was proposed in [28]. Moreover, a
reduction in the communication frequency among nodes.
creative parameter estimation-based distributed state estimation
4) By utilizing the σ modification technique [32], a robust
approach was designed for both discrete-time and continuous-
fully distributed state estimation and cooperative stabi-
time LTI plants [29], [30]. In these works, however, only the
lization strategy is proposed for noisy LTI plants, which
state estimation problem is tackled while the control issue is
can simultaneously guarantee the uniform boundedness
missing. Note that the simultaneous consideration of distributed
of the adaptive coupling gain and the plant state, in the
state estimation and control for LTI plants is a very challenging
presence of both the process and measurement distur-
and complicated issue. The difficulty arises from the strong
bances (Theorem 6).
coupling between the distributed state estimation and control
processes of LTI plants. Subsequently, the distributed state esti-
mator and the controller cannot be designed separately. This fact
results in the inadequacy of conventional pure state estimation C. Organization and Notations
techniques when it comes to addressing the joint distributed The rest of this article is organized as follows. In Section II,
state estimation and control problem. Recently, Kim et al. [31] some preliminary results and the problem formulation are pre-
proposed a novel joint state estimation and control framework, sented. In Section III, a fully distributed state estimation and
where each node could self-organize its own controller using cooperative stabilization control framework is proposed for LTI
the local measurement and local interaction with its neighbors, plants. In Section IV, the stability of the proposed framework is
which builds on undirected graphs. Noting that the Laplacian analyzed. In Section V, a discussion on the choice of estimator
matrices of directed graphs are generally asymmetric, which and controller gains is provided. In Section VI, the proposed
renders the traditional method utilizing the symmetry of Lapla- framework is applied to pure distributed state estimation. In
cian matrices ineffective for cases over such graphs. Until now, Section VII, a robust strategy based on the designed framework
research on fully distributed state estimation and cooperative is proposed for noisy plants. In Section VIII, the effectiveness
stabilization of LTI plants under general directed graphs is still of the proposed method is illustrated by numerical examples.
demanding. Finally, Section IX concludes this article.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
6748 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 69, NO. 10, OCTOBER 2024

Let Ip represent the p-dimensional identity matrix and 1p not necessarily guaranteed. Let Ni denote the neighbor set of
denote the p-dimensional column vector with all elements being node i. A directed path from node i1 to node il is a sequence
1. Let 0p denote the p-dimensional column vector with all of edges (i1 , i2 ), . . ., (il−1 , il ) ∈ E. The directed graph G is
elements being 0. Let L2 denote the Hilbert space of square strongly connected if there is a directed path from node i to node
integrable and Lebesgue measurable functions. Let the symbol j ∀i, j ∈ V. For the graph G, its adjacency matrix A = [aij ]N ×N
diag(x1 , . . . , xn ) represent a diagonal matrix with diagonal is defined as aij = 1 if (j, i) ∈ E and aij = 0 otherwise. We
elements being xi , i = 1, . . ., n. Let σmax (X) represent the assume that there is no self-loop in the graph, i.e., aii = 0. The
maximum singular value of the matrix X. For a symmetric ma- Laplacian matrix L = [lij ]N ×N is defined as lii = N j=1 aij and
trix Z, let λmax (Z), λmin (Z), and Tr(Z) denote the maximum lij = −aij if i = j. In this article, the following assumption on
and minimal eigenvalues, and the trace of Z, respectively. Let the graph G is needed.
A ⊗ B denote the Kronecker product of the matrices A and B. Assumption 2: The directed graph G is strongly connected.
Definition 1 ([25], [33]): A square matrix is called a nonsin-
II. PROBLEM FORMULATION gular M-matrix if all its off-diagonal elements are nonpositive
and all its eigenvalues have positive real parts.
A. Model Description Lemma 1: Suppose that Assumption 2 holds. Then, the matrix
This article considers the distributed state estimation and
Lj = L + Aj ∀j ∈ V
cooperative stabilization problem for a class of continuous-time
LTI plants containing N nodes, where each node equips with is a nonsingular M -matrix, where L is the Laplacian matrix of
a sensor as well as an actuator. The dynamics of the plant are G and Aj = diag(a1j , . . ., aN j ). Moreover, the matrix
described by
L̂ = L ⊗ Im + Â

N 
ẋ = Ax + Bi ui is a nonsingular M -matrix, where m = N i=1 mi and  =
i=1 diag(A1 , . . ., AN ) with Ai = diag(ai1 Im1 , . . ., aiN ImN ).
The proof of Lemma 1 is presented in Appendix A.
yi = Ci x, i ∈ V  {1, . . . , N } (1) Lemma 2 ([33, Th. 4.25]): For a nonsingular M -matrix M,
where x ∈ Rn , yi ∈ Rmi , and ui ∈ Rpi represent the state of the there is a diagonal matrix G > 0 such that the matrix GM +
plant, the measurement of node i, and the control input of node i, MT G is positive definite.
respectively. Besides, A ∈ Rn×n , Ci ∈ Rmi ×n , and Bi ∈ Rn×pi
are the state matrix of the plant, the output matrix of node i, and C. Estimator and Controller Structure
the input matrix of node i, respectively. Let y = [y1T , . . . , yN ]
T T
Many existing researches on distributed state estimation adopt
and u = [u1 , . . . , uN ] denote the augmented measurement and
T T T
the following estimator for plants [16], [17], [18], [19], [25],
input, respectively. [26], [27]:
In this article, node i is regarded as an independent agent

N
such that it only obtains the measurement yi via its own sensor, x̂˙ i = Ax̂i + Fi (yi − Ci x̂i ) + γPi aij (x̂j − x̂i ) + Bu (2)
and implements the control input ui via its own actuator. Before j=1
moving on, a useful assumption is made as follows.
Assumption 1: The triple (A, B, C) is controllable and ob- where x̂i is the state estimate of node i, Pi , and Fi are the
servable, where B = [B1 , . . . , BN ] and C = [C1T , . . . , CN ] .
T T feedback gain matrices, and γ is the constant coupling gain.
Remark 1: Assumption 1 requires that the plant be collec- To ensure the stability of the estimator (2), the closed-loop state
tively controllable and observable, which allows (A, Bi ) to be matrix
uncontrollable and (A, Ci ) to be unobservable ∀i ∈ V. Hence, IN ⊗ A − Fdiag Cdiag − γPdiag (L ⊗ In )
this assumption is very mild and even a necessary and suffi-
cient condition in centralized settings. Particularly, if there is a should be Hurwitz stable, where Fdiag = diag(F1 , . . . , FN ),
fusion center collecting all the measurements y and designing Cdiag = diag(C1 , . . . , CN ), and Pdiag = diag(P1 , . . . , PN ). For
all the control inputs u, the cooperative state estimation and this goal, the existing works usually require a sufficiently large
stabilization problem under Assumption 1 naturally turns into γ, of which the lower bound depends on the global connectivity
the trivial output-feedback control problem. In contrast, this information of the communication topology [26]. To avoid the
article focuses on the cooperative problem under Assumption 1 requirement on global information, we will design a novel fully
in a fully distributed manner, i.e., each node self-organizes its distributed state estimation and cooperative stabilization control
behavior without a fusion center. framework for the plant (1), particularly under general directed
graphs. Specifically, we will focus on a joint state estimation and
control problem formulated as follows.
B. Graph Theory
Problem 1: Design a fully distributed state estimator (FDSE)
The communication topology among the N nodes is repre- and a local controller for node i ∀i ∈ V, in the form of
sented by a directed graph G = {V, E}, where V defined in (1) is
ŷ˙ i = fi (x̂i , yj , ŷj , j∈Ni ), x̂˙ i = gi (x̂i , ŷi ), ui = hi (x̂i ) (3)
the node set and E ⊂ V × V is the edge set. In this directed graph,
let (i, j) ∈ E denote that node i is a neighbor of node j such that where ŷi ∈ Rm is the estimate of the augmented measurement y
node j can receive information from node i, but the converse is defined following (1) for node i; x̂i ∈ Rn is the estimate of the
Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
DUAN et al.: FRAMEWORK ON FULLY DISTRIBUTED STATE ESTIMATION AND COOPERATIVE STABILIZATION OF LTI PLANTS 6749

plant state x defined in (1) for node i; and ui is the local input manner in Section V. Moreover, x̂i in (4) is the estimate of the
defined in (1) and exerted by node i. In addition, fi (·), gi (·), and plant state x in (1), designed as
hi (·) are certain linear or nonlinear functions to be designed for
two objectives. First, ŷi and x̂i are expected to converge to y and 
N
x̂˙ i = (A + BK)x̂i + Fj (Cj x̂i − ŷij ) (6)
x, respectively, i.e., j=1
lim ŷi (t) − y(t) = 0 with Fj ∈ Rn×mi , j ∈ V, being the estimator gain matrix. Sub-
t→∞

and sequently, a local controller is proposed for node i to stabilize

(1) as
lim x̂i (t) − x(t) = 0
t→∞ ui = Ki x̂i . (7)
for all i ∈ V. Further, the cooperative stabilization task of the
plant (1) is expected to be realized in the sense that Remark 2: The relation between (3) and (4)–(7) is demon-
strated as follows. Before moving on, let ŷi = [ŷi1 T
, . . ., ŷiN
T T
]
lim x(t) = 0. with ŷij defined in (4). It follows from (4) that:
t→∞
⎡ ⎤
The salient feature of the proposed structure (3) in Problem 1 
is that each node has only access to its own and neighbors’ ŷ˙ i = − (γi Im + Ψi ) ⎣ (ŷi − ŷj ) + Ai (ŷi − y)⎦
information. By utilizing this structure, this article is aimed at j∈Ni
developing a state estimator and a local controller for each node
+ C(A + BK)x̂i (8)
in a fully distributed way, particularly without using the global
connectivity information of the directed communication graph. where Ψi = diag(ψi1 Im1 , . . ., ψiN ImN ), and the matrices Ai ,
In contrast to centralized methods that rely on a central authority B, C, and K are defined in Lemma 1, Assumption 1, and (4),
to gather global information, and decentralized methods which respectively. It can be found from (8) that the dynamics of ŷi
are constrained by structural limitations during the design of are described by a differential equation about x̂i , yj , ŷj , j ∈ Ni ,
estimator and controller gains [34], the proposed method with a which corresponds to the function fi (·) in (3). That is, fi (·) in (3)
distributed structure offers enhanced system robustness, scala- is explicitly designed as the right-hand side of (8), equivalently,
bility, and greater design flexibility. the augmented form of the right-hand side of (4). In addition,
gi (·) and hi (·) in (3) are specified as (6) and (7), respectively.
III. FULLY DISTRIBUTED STATE ESTIMATION AND Remark 3: The underlying idea behind the proposed strategy
COOPERATIVE STABILIZATION DESIGN (4)–(7) is that each node first employs an output estimator
In this section, a novel fully distributed state estimation and (4) to obtain the value of y based on its own and neighbors’
stabilization control framework based on (3) is designed for each information, and then performs a state observer (6) to estimate
node to stabilize the plant (1) cooperatively. the state of the plant (1) based on the output estimate. Further,
First, an output estimate ŷij ∈ Rmi is designed for node i to each node uses a local controller (7) based on the state estimate
estimate the measurement yj as to cooperatively stabilize the plant (1).
In many relevant works, such as [25], [26], and [27], the
ŷ˙ ij = −(γi + ψij )ζij + Cj (A + BK)x̂i , i, j ∈ V (4) distributed state estimation structure (2) involves a common
where constant γ relying on the global connectivity information of the
communication topology. For example, γ introduced in [25],

N
[26], and [27] needs to satisfy
ζij = aim (ŷij − ŷmj ) + aij (ŷij − yj )
m=1 λmax (fL (L))
γ>

N λmin (gL (L))
ψij = μζij
T
ζij , γ̇i = ψij (5)
where fL (·) and gL (·) are two complex matrix functions that
j=1
depend on the Laplacian matrix of the communication topology.
K = [K1T , . . ., KN ] is the augmented control gain matrix with
T T
On the contrary, in this article, this global information is not
pi ×n
Ki ∈ R , i ∈ V, being the gain of node i to be designed in needed, since we introduce an adaptive coupling gain γi in (4)
Section V, B is defined in Assumption 1, and μ is any positive to effectively address its absence. Moreover, the adaptive gain
scalar. Besides, ζij , ψij , and γi are the consensus error of the is much smaller compared with the fixed gains designed in [25],
output estimate, the quadratic form of the consensus error, and [26], and [27], which will be illustrated in the simulation part.
the adaptive estimator gain for node i, respectively. Accord- Note that the continuous-time state estimation laws are usually
ing to the definition of the adjacency matrix A = [aij ]N ×N in implemented in a discrete-time digital form, and a large coupling
Section II-B, ζij is a local fusion law of node i based only on the gain in the continuous-time state estimation requires a small
interaction with its neighbors. Further, ψij and γi can be locally integration step, resulting in a high communication frequency
computed by each node. It is worth mentioning that although the among nodes [26], [35]. Hence, in the practical applications, the
augmented control gain K is needed in (4), we will demonstrate adaptive gain designed in this article benefits the reduction in
that this gain can be obtained by each node in a fully distributed the communication frequency among nodes.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
6750 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 69, NO. 10, OCTOBER 2024

In addition, the information transmitted among nodes in the It follows from (9) that the dynamics of ỹ j can be computed as
proposed estimators (4) and (6) is quite different from that in
the traditional distributed estimator (2). Specifically, node i ex- ỹ˙ j = −[(Γ + Ψj )Lj ⊗Imj ]ỹ j + [IN ⊗(Cj Ā) − 1N ⊗(Cj B̄)]x̃
changes the state estimate x̂i with its neighbors when using (2), (11)
while it shares the output estimate ŷij with its neighbors when where Γ = diag(γ1 , . . ., γN ), Ψj = diag(ψ1j , . . ., ψN j ), Ā
using (4). Moreover, to realize the fully distributed design of the and B̄ are defined following (9), and Lj = L + Aj is a non-
state estimator (6), two extra dynamical internal states, namely, singular M -matrix according to Lemma 1. Next, by letting
the output estimate ŷij and the adaptive gain γi , are developed ζ j = [ζ1j , . . . , ζN j ] , we have
T T T
for node i ∀i ∈ V. By doing so, an additional computational
cost is introduced, which is the price that the proposed structure ζ j = (Lj ⊗ Imj )ỹ j
(3) takes. In comparison to the reduced communication costs and
the elimination of the requirement for the global communication with the dynamics being derived as
topology, a slight increase in computational burden is generally ζ̇ j = −[Lj (Γ + Ψj )⊗Imj ]ζ j + [Lj ⊗(Cj Ā) − αj ⊗(Cj B̄)]x̃
tolerable for most autonomous intelligent agents equipped with (12)
advanced computers.
where αj = [a1j , . . . , aN j ]T is the augmented adjacency gain
IV. STATE ESTIMATION AND STABILIZATION CONTROL with elements defined in Section II-B. Now, the performance
PERFORMANCE ANALYSIS of the proposed distributed state estimator (6) and the feedback
controller (7) is guaranteed as shown in the following theorems.
In this section, the state estimation and control performance
Theorem 1: Suppose that Assumption 2 holds. By choosing
of the proposed cooperative strategy (4)–(7) is analyzed.
the feedback gain matrices Ki and Fi such that Ā and Acl
First of all, let ỹij denote the output estimation error of yj for
defined following (9) and (10), respectively, are Hurwitz stable,
node i, i.e.,
the output estimate ŷij in (4) and the state estimate x̂i in (6)
ỹij = ŷij − yj , i, j ∈ V. asymptotically converge to the measurement yj and the state x,
respectively.
Further, let x̃i denote the state estimation error of x for node i, The proof of Theorem 1 is presented in Appendix B, where
i.e., the dynamics of ỹ j in (11) are used for the stability analysis. In
x̃i = x̂i − x particular, the derivation using ỹ j is much more straightforward
compared to the one using ỹj ( ŷj − yj ). Note that ỹ j denotes
and x̃ = [x̃T1 , . . . , x̃TN ]T be the augmented state estimation error. the augmented estimates of yj by all nodes, which decouples
By combining (1) with (4), the dynamics of ỹij can be directly the estimation of y into N subsystems. As a result, we can
derived as adopt the N Lyapunov function candidates V1j , j ∈ V, in (25)
to analyze the N subsystems, respectively. On the other hand,
ỹ˙ ij = −(γi + ψij )ζij + Cj Āx̃i − Cj B̄x̃ (9)
ỹj is the estimate of y by node j and coupled with other
where Ā = A + BK, B̄ = [B1 K1 , . . . , BN KN ], and the dy- estimates ỹm ∀m ∈ V/{j}. This indicates that the stability
namics of ζij can be rewritten as analysis using ỹj needs to be established on a high-dimensional
system. Hence, we choose the derivation route using ỹ j in this

N
article.
ζij = aim (ỹij − ỹmj ) + aij ỹij
Moreover, as shown in Appendix B, the fully distributed
m=1
design of the proposed estimators (4) and (6) benefits from the
since ŷij − ŷmj = ỹij − ỹmj . Similarly, the dynamics of the quadratic terms of the consensus error, i.e., ψij = μζij T
ζij , to
augmented state estimation error x̃ can be derived as derive the adaptive gain. Precisely, this quadratic form proves
effective in addressing the asymmetric Laplacian matrix associ-
x̃˙ = Acl x̃ − (IN ⊗F)ỹ (10)
ated with directed communication graphs. It is worth mentioning
where that the derivative of the adaptive gain γi in (5) is the sum of the
feedback gains ψij , j ∈ V, making the closed-loop dynamics of
Acl = IN ⊗ (A + FC + BK) − 1N ⊗ B̄ the consensus errors ζ j , j ∈ V, coupled with each other, which
F = [F1 , . . . , FN ], K = [K1T , . . . , KN ]
T T inevitably introduces significant challenges in the convergence
analysis. We have developed a novel Lyapunov function (25) to
ỹ = [ỹ1T , . . . , ỹN ] , ỹi = [ỹi1
T T T
, . . . , ỹiN
T T
] . handle this challenge.
In addition, the stabilization control performance of the pro-
Let ỹ j = [ỹ1j
T
, . . . , ỹN j ] denote the augmented form of the jth
T T
posed cooperative strategy (4)–(7) is summarized as follows.
output estimation error by all nodes, which is different from ỹi
Theorem 2: Suppose that Assumption 2 holds. Problem 1 is
defined above. Then, we have
solved using the FDSE (4)–(6) and the feedback controller (7)

N 
N by choosing the feedback gain matrices Ki and Fi such that Ā
ỹ ỹ =
T
(ỹi ) ỹi =
T
(ỹ j )T ỹ j . and Acl defined following (9) and (10), respectively, are Hurwitz
i=1 j=1 stable.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
DUAN et al.: FRAMEWORK ON FULLY DISTRIBUTED STATE ESTIMATION AND COOPERATIVE STABILIZATION OF LTI PLANTS 6751

Based on Theorems 1 and 2 can be directly proved as follows. LMIs (13)–(15) and the system controllability and observability
First, the closed-loop dynamics of x in (1) can be rewritten as in Assumption 1.
Theorem 4: Suppose that Assumption 1 holds. Then, the
ẋ = Āx + B̄x̃ LMIs (13)–(15) are solvable, i.e., there must exist positive-
definite matrices Q and P enabling that the LMIs (13)–(15)
where Ā and B̄ are defined following (9). According to The- simultaneously hold.
orem 1, we have limt→∞ x̃(t) = 0 and x̃(t) ∈ L2 . Moreover, The proof of Theorem 4 is presented in Appendix D. The
since Ā is Hurwitz stable, we have [36, Ch. 3] idea of the proof is designing appropriate gains K and F by 1)
 decoupling the closed-loop estimation and stabilization system
t
into a slow system and a fast system; 2) guaranteeing the stability
lim x(t) = lim eĀt x(0) + eĀ(t−τ ) B̄x̃(τ )Dτ = 0.
t→∞ t→∞ 0 of two decoupled systems. The feasibility of this idea is ensured
by the fact that the eigenvalues of the closed-loop state matrices
A + BK and A + FC can be arbitrarily assigned if and only if
Thus, the proof of Theorem 2 is complete.
Assumption 1 holds [36, Ch. 3.4]. It is worth mentioning that if
Theorem 2 reveals that the cooperative stabilization of the
the state matrix A is Hurwitz stable, the solutions to the LMIs
plant (1) can be achieved by the proposed FDSE and controller
(13)–(15) must exist, even when Assumption 1 does not hold. It
(4)–(7). It is worth noting that the feedback gain matrices Ki
is straightforward according to the properties of the Lyapunov
and Fi should be carefully chosen to make Ā and Acl Hurwitz
equations [36, Lemma 3.18]. Hence, the controllability and
stable. We can simply choose an appropriate Ki to make Ā
observability condition stated in Assumption 1 is a sufficient
stable. However, it is difficult to design Fi to ensure that Acl is
but not necessary condition to ensure the existence of solutions
Hurwitz due to its N n × N n dimension. In the next section, a
to the LMIs (13)–(15).
novel method for choosing the gains will be introduced.
On the other hand, the LMIs (13)–(15) are centralized since
they are dependent on the global input and output matrices, i.e.,
V. ESTIMATOR AND CONTROLLER GAINS DESIGN B and C, which indicates that the estimator and controller gains
In this section, a thorough analysis on the choice of the have to be designed using a central authority. To address this
estimator and controller gains, i.e., Fi and Ki ∀i ∈ V, with issue, a novel approach based on the derivation of Theorem 4 is
lower dimensions is provided. In addition, a fully distributed developed as follows, which enables each node to self-organize
method for designing the gains is proposed for each node. its own estimator and controller gains in a fully distributed way.
Theorem 3: Suppose that Assumption 2 holds. Problem 1 is First, let B̂ij ∈ Rn×pj and Ĉij ∈ Rmj ×n be the estimates of
solved using the FDSE (4)–(6) and the feedback controller (7) Bj and Cj for node i ∀i, j ∈ V, respectively. The dynamics of
by choosing the feedback gain matrices Ki and Fi such that B̂ij and Ĉij are designed, respectively, as
A + BK, A + FC, and A + BK + FC are Hurwitz stable. 
N
˙
The proof of Theorem 3 is presented in Appendix C. In B̂ij = − aim (B̂ij − B̂mj ) − aij (B̂ij − Bj )
comparison to the N n × N n-dimensional condition in Theo- m=1
rem 2 that Acl should be Hurwitz stable, three low-dimensional
˙ 
N
conditions are derived in Theorem 3, which renders the design Ĉij = − aim (Ĉij − Ĉmj ) − aij (Ĉij − Cj ). (16)
of the estimator and controller gains more straightforward. To m=1
simultaneously meet the three requirements, node i ∀i ∈ V,
Since the matrix Lj defined in Lemma 1 is a nonsingular M -
can first set Ki = −BiT P −1 with P being any positive-definite
matrix, by utilizing the augmented form of the above equation,
matrix satisfying the following linear matrix inequality (LMI):
we can directly prove that B̂ij and Ĉij asymptotically converge
AP + P AT − BB T < 0. (13) to Bj and Cj , respectively. In particular, since Bj and Cj are
constants, it is feasible for each node to obtain their values by
utilizing (16) prior to implementing the cooperative strategy
Then, the estimator gain of node i is chosen as Fi = −Q−1 CiT
(4)–(7). Next, according to the proof of Theorem 4, for any
with Q being any positive-definite matrix satisfying the follow-
given positive-definite matrix T2 , each node can select an suffi-
ing two LMIs:
ciently small positive-definite matrix T1 satisfying (33) and (36),
denoted by T1,i . Then, node i ∀i ∈ V, performs the following
QA + AT Q − C T C < 0 (14)
fusion law:
QA + AT Q − C T C + QBK + (BK)T Q < 0. (15)  N
˙
T̂1,i = − aij (T̂1,i − T̂1,j ) (17)
The LMI (13) ensures that A + BK is Hurwitz stable, since j=1
it shows that there exists a positive-definite matrix P −1 mak-
where T̂1,i (0) = T1,i . Similarly, we can prove that T̂1,i asymp-
ing P −1 (A + BK) + (A + BK)T P −1 < 0 hold. Similarly, the 
LMIs (14) and (15) are a guarantee of A + FC and A + BK + totically converges to N1 N T1,i . Therefore, each node can
1
Ni=1
FC being Hurwitz stable, respectively. In the following, we es- obtain the value of N i=1 T1,i using (17). Moreover, the

tablish a direct relation between the existence of solutions to the solution P1 to (33) with T1 = N1 N i=1 T1,i also satisfies (36),

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
6752 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 69, NO. 10, OCTOBER 2024

Algorithm 1: A Fully Distributed Method for Designing the 

N

Estimator and Controller Gains in (4)–(7) for Node i. + Fj (Cj x̂i − υij ) (20)
j=1
1: initialize T2 > 0;
2: obtain the values of B and C by utilizing (16); with
3: select any T1 satisfying (33) and (36), denoted by T1,i ; 
N

4: obtain the value of N1 N i=1 T1,i by utilizing (17);
χij = aim (υij − υmj ) + aij (υij − yj )
5: design K and F by utilizing (18). m=1


N
φij = μχTij χij , θ̇i = φij
since (33) is a linear equation with respect to T1 . Based on this j=1
finding, the control and estimator gains K and F are designed where υij is the estimate of yj by node i, i, j ∈ V; x̂i is the
for node i as estimate of the plant state in (19) by node i; μ is any positive
scalar; θi is the adaptive coupling gain with a positive initial
K = −B T P1 , F = −Q1 C T (18)
value θi (0) > 0; φij is the feedback gain; and Fj is any feedback
where P1 and Q1 being the unique solutions, respectively, to gain matrix satisfying that A + FC is Hurwitz stable.
Theorem 5: Suppose that Assumption 2 holds and (A, C)
1 
N is observable. The FDSE (20) ensures that the estimate xi
AT P1 + P1 A − P1 BB T P1 + T1,i = 0. converges to the system state x in (19), i.e.,
N i=1
lim x̂i (t) − x(t) = 0 ∀i ∈ V.
Q1 AT + AQ1 − Q1 C T CQ1 + T2 = 0 t→∞

The proof of Theorem 5 can be conducted using the same

Altogether, the proposed method is summarized in Algorithm 1. derivation process as that of Theorem 3, except setting B = 0
It is worth mentioning that Algorithm 1 is fully distributed and and K = 0. Compared with the joint state estimation and sta-
applicable to cases under directed communication graphs. bilization strategy (4)–(7), only the estimator gain matrix F is
Remark 4: Note that a fully distributed cooperative control needed in the FDSE (20), which should ensure that A + FC is
framework has been proposed in [31]. This result builds on Hurwitz stable. The proposed estimator (20) has many potential
undirected graphs, while the strategy proposed in this article is applications. For example, it can be directly integrated into
applicable to general directed graphs, making it more versatile the well-established distributed localization framework [37] for
in the applications. It is worth mentioning that establishing a position estimation, and the engineer-friendly prescribed control
fully distributed state estimation and cooperative stabilization architecture [38] for state estimation.
strategy under directed graphs is much more challenging. The Remark 5: In the relevant work [25], a fully distributed
reason lies in that the Laplacian matrices of directed graphs are state estimation scheme has been proposed. When utilizing this
generally asymmetric, which renders the conventional method scheme to address the joint state estimation and stabilization
of designing the coupling gains using the symmetry of the problem formulated in Section II-C, a potential challenge may
Laplacian matrix ineffective in this article. arise due to the strong coupling among the state estimate, the
adaptive coupling gain and the control input. On the contrary, this
VI. APPLICATION TO PURE STATE ESTIMATION article bypasses this strong coupling by introducing the output
estimation framework (4) or (20), where the coupling gain relies
In this section, the proposed structure (3) is further applied to
on the output estimate instead of the state estimate. Hence, the
the pure distributed state estimation problem. The dynamics of
new framework renders the design and analysis of the FDSE and
the system (1) turn into
cooperative stabilization controller more straightforward.
ẋ = Ax, yi = Ci x, i ∈ V (19)
VII. ROBUST ESTIMATOR AND CONTROLLER DESIGN
where the variables are defined the same as those in (1). The In the previous section, a joint fully distributed state estima-
objective for the N nodes is to design appropriate distributed tion and cooperative stabilization control framework (4)–(7) is
state estimates x̂i , i ∈ V, to observe the plant state in the sense deigned for the noise-free plant (1), where γi may diverge if there
that limt→∞ x̂i (t) − x(t) = 0. Compared with the cooper- exist disturbances in the system process and measurement. To
ative stabilization problem of the system (1), the distributed address this issue, a robust strategy modified from (4)–(7) is
state estimation problem is not concerned with the stability of proposed for a noisy plant as follows.
the plant state, but only focuses on the state observation. As a First of all, the dynamics of the noisy plant are described by
result, the control input is omitted in the system dynamics for
simplification here. 
N
ẋ = Ax + Bi ui + ω, yi = Ci x + νi i ∈ V (21)
Based on the structure (3), a FDSE is proposed for node
i=1
i ∀i ∈ V, as
where ω ∈ Rn and νi ∈ Rmi are the process and measurement
υ̇ij = − (θi + φij )χij + Cj Ax̂i , x̂˙ i = Ax̂i noise, respectively. In this section, we assume that ω and νi are

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
DUAN et al.: FRAMEWORK ON FULLY DISTRIBUTED STATE ESTIMATION AND COOPERATIVE STABILIZATION OF LTI PLANTS 6753

bounded in the sense that there exist three positive constants ωb ,

νb , and νd such that
ω(t) 2 ≤ ωb , ν(t) 2 ≤ νb , ν̇(t) 2 ≤ νd (22)
for all the time t ≥ 0, where ν(t) = [ν1T (t), . . . , νN
T
(t)]T . It is
worth noting that the values of ωb , νb , and νd are not needed for
the robust strategy design as follows.
By utilizing the σ modification technique [32], the original
strategy (4)–(7) remains unchanged except that γi in (5) is
redesigned as

N
γ̇i = −(γi − 1)2 + ψij (23) Fig. 1. Six mobile robots carry a huge plant cooperatively.
j=1

where  is any positive constant scalar and the initial value of γi

3 → robot 4 → robot 5 → robot 6 → robot 1. In this task, let
is greater than 1, i.e., γi (0) > 1. In the following, we provide a
pini ∈ R2 , pdes ∈ R2 , p(t) ∈ R2 , and v(t) ∈ R2 denote the initial,
significant result about the modified cooperative strategy.
desired, real-time positions, and the real-time velocity of the
Theorem 6: Suppose that Assumptions 1 and 2 hold. By uti-
plant, respectively. According to [31], the dynamics of the plant
lizing Algorithm 1 to design the feedback gain matrices Ki and
can be modeled as the noisy system (21), where the state x
Fi such that the matrices A + BK, A + FC, and A + BK + FC
denotes the position and velocity error [pT (t) − pTdes , v T (t)]T
are Hurwitz stable, the cooperative control strategy (4)–(7) with
and the system matrices are
γi redesigned in (23) ensures that the state of the noisy plant ⎡ ⎤ ⎡ ⎤
(21) exponentially converges to the residual set {x : x 22 ≤ Π}, 0 0 1 0 0
⎢ 0 0 0 1⎥ 1 ⎢ ⎥
where Π is a constant defined in (44). Moreover, γi (t) ∀i ∈ V, A= ⎢ ⎥ ⎢ 0 ⎥ , i = 1, . . . , 6
is uniformly bounded. ⎣ 0 0 0 0⎦ , Bi =
M cos αi ⎦
⎣
The proof of Theorem 6 is presented in Appendix E. 0 0 0 0 sin αi
Theorem 6 indicates that the proposed cooperative strategy
with M = 5 being the inertial of the plant. Particularly, pini 
(4)–(7) with γi redesigned in (23) guarantees the boundedness
[pini,1 , pini,2 ] and pdes  [pdes,1 , pdes,2 ] are set as [30, 40]
of both the plant state and the adaptive gain. This result benefits
and [230, 200], respectively. The force directions are set as
from the introduction of the negative term −(γi − 1)2 into the
α1 = π/3, α2 = 3π/4, α3 = 4π/3, α4 = 3π/2, α5 = 7π/4,
dynamics of the adaptive gain γi in (23). Moreover, for pure
and α6 = 2π. Moreover, the vision-aided measuring strategy
distributed state estimation of the noisy plant (21) without the
in [39] is adopted in this example. Specifically, we assume that
inputs, we can adopt the strategy (20) except that θi is redesigned
Robots 1 and 3 are equipped with cameras capable of measuring
as
the positions of the plant, while Robots 2, 4, 5, and 6 are not

N
equipped with any sensors to reduce the costs. Subsequently,
θ̇i = −(θi − 1)2 + φij (24) the output matrices in (21) are set as
j=1 
1 0 0 0
where  is defined in (23) and the initial value of θi is greater Ci = , i = 1, 3
0 1 0 0
than 1, i.e., θi (0) > 1. Similarly, the performance of the re-
designed strategy can be ensured using the techniques in the and Ci = 0, i = 2, 4, 5, and 6. In addition, the unknown process
proof of Theorem 6. In comparison to the innovative parameter and measurement disturbances in (21) are chosen as
estimation-based distributed state estimation algorithm [30] for ω(t) = 0.02 ∗ [sin(t), sin(2t), sin(3t), sin(4t)]T
noise-free systems, the above new strategies show robustness
against system disturbances. νi (t) = 0.02 ∗ [cos(t), cos(2t)]T , i = 1, 3.
The proposed fully distributed cooperative strategy (FDCS)
VIII. SIMULATION
(4)–(7) with γi redesigned in (23) is used to complete the above
In this section, two numerical examples are presented to illus- task. The estimator and controller gains Ki and Fi , i = 1, . . ., 6,
trate the effectiveness of the proposed fully distributed methods: are designed by applying Algorithm 1. The parameters μ and  in
1) planar carrying of a plant by a group of mobile robots; 2) (5) and (23) are set as 0.003 and 0.01, respectively. The position
distributed state estimation of a large-scale sensor network. and velocity errors of the object under the proposed FDCS are
Example 1: A cooperative planar transportation task of a plant illustrated in Fig. 2. It can be found that all errors converge to
by six mobile robots is considered, as illustrated in Fig. 1 . In this small residual sets, which indicates that the transportation task
task, each robot exerts a partial force on the plant and intends can be completed by the proposed strategy.
to carry it to a desired position on the c1 − c2 coordinate in a In addition, we compare the proposed FDCS with three rel-
distributed manner. The communication topology graph of the evant methods reported in literature, namely, the centralized
six robots is a directed circle, i.e., robot 1 → robot 2 → robot estimation and control method (CECM) [36], the distributed

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
6754 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 69, NO. 10, OCTOBER 2024

Fig. 2. Position and velocity errors of the plant under the pro-
posed FDCS (4)–(7), where x(t) = [p1 (t) − pdes,1 , p2 (t) − pdes,2 , v1 (t),
Fig. 5. Communication topology of 100 sensors, where two nodes are
v2 (t)]T . neighbors if the distance between them is less than 60 in the coordinate-
frame scale.

will not diverge in the presence of both the system process

and measurement noise, which benefits from the negative term
−(γi − 1)2 in (23). Furthermore, the coupling gains adopted
in this article are smaller than the ones in the DCSM and PPDC.
Since a greater coupling gain usually requires a higher commu-
nication frequency among nodes [26], the proposed FDCS needs
a lower communication frequency than the DCSM and PPDC.
Altogether, the fully distributed state estimation and cooperative
stabilization strategy designed in this article have advantages in
guaranteeing the control performance at a lower communication
Fig. 3. Trajectories of the noisy plant under different cooperative con- frequency and without using any global information simultane-
trol methods, where the filled square and circle represent the initial and
terminal positions, respectively. ously.
Example 2: In this example, a network of 100 sensors is used
to cooperatively observe the state of a plant of the 100th order.
The communication topology graph of the 100 sensors is shown
in Fig. 5. The model of the plant is described by (19) with
parameters being chosen as

099 0.01 × I99
A= , Ci = [0i−1 , 1, 0100−i ],
0 0T99

where i = 1, . . . , 100. The initial values of the system state

and the associated estimates are set as x(0) = [x1 (0), . . .,
xh (0), . . ., x100 (0)]T , and x̂i (0) = [x̂i,1 (0), . . ., x̂i,h (0), . . .,
x̂i,100 (0)]T , where xh (0) = 1 + 0.01i and x̂i,h (0) = i ∀h =
1, . . ., 100 ∀i = 1, . . ., 100. The proposed FDSE in (20) with
θi redesigned in (24) is adopted to estimate the plant state for
Fig. 4. Coupling gains of different cooperative control methods, where
6 all nodes, where both the parameters μ and  are set as 0.01.
γ̄ = 16 γ denotes the average coupling gain in the FDCS.
i=1 i It can be found from Fig. 6 that the state estimate using
the FDSE tends to the real system state, which indicates that
the proposed FDSE can guarantee the estimation performance
cooperative stabilization method with the global connectivity for large-scale systems. Moreover, we compare the proposed
information of the communication topology (DCSM) [16], and FDSE with four relevant distributed state estimation methods
the plug-and-play distributed control method (PPDC) [31]. It (DSEM) in literature, namely the completely decentralized state
can be seen from Fig. 3 that the proposed FDCS can guarantee a estimation methods with a fixed coupling gain (CDFC) or an
carrying trajectory very close to the centralized method. More- adaptive coupling gain (CDAC) in [25], the asymptotically
over, the coupling gains of different methods are illustrated in optimal consensus-based filtering method (AOCF) in [26], and
Fig. 4. Since the PPDC is designed based on undirected graphs, the DSEM in [27]. Note that the unified estimation framework
the coupling gain γ is computed using a bidirected circle graph. proposed in [28] at the steady state has a similar structure to [35]
Fig. 4 shows that the adaptive coupling gains designed in (23) (or the continuous-time counterpart [26]). Since this article

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
DUAN et al.: FRAMEWORK ON FULLY DISTRIBUTED STATE ESTIMATION AND COOPERATIVE STABILIZATION OF LTI PLANTS 6755

Fig. 6. Estimation performance of the proposed FDSE (20). Without

loss of generality, take the 10th, 30th, and 70th elements in x(t) and the
Fig. 8. estimation errors by different DSEMs, where ea (t) =
50th sensor for example, where x̂10 30 70
50 , x̂50 , and x̂50 are the estimates of 1
Average
100 100
x10 (t), x30 (t), and x70 (t) obtained by node 50, respectively. 10000 i=1 h=1
|x̂i,h (t) − xh (t)| denotes the average estimation
error.

counterpart of the unified algorithm designed in [28]. The av-

erage estimation errors by different DSEMs are illustrated in
Fig. 8, which indicates that the proposed FDSE possesses better
estimation performance in the presence of system disturbances.
Until now, the effectiveness of the theoretical results obtained
in this article has been illustrated.

IX. CONCLUSION
This article investigated a fully distributed state estimation
and cooperative stabilization problem for LTI plants with multi-
ple nodes under directed graphs. To achieve the estimate of the
plant state for each node, a FDSE has been introduced with a
Fig. 7. Coupling gains designed in different state estimation methods,
1
100 1
100 novel adaptive law such that the global connectivity information
where θ̄ = 100 θ and γ̄ = 100
i=1 i
γ denote the average cou-
i=1 i of the communication topology could be avoided. Further, a
pling gains in the FDSE, CDFC, and CDAC, respectively.
local controller has been developed for nodes to stabilize the
plant collaboratively. Particularly, the stability of the joint state
estimation and control framework has been ensured and a spe-
considers the continuous-time LTIs, we utilize the continuous- cific method for designing the estimator and controller gains
time algorithm AOCF [26] as a representative for numerical has been introduced. Altogether, the proposed method enables
comparisons. The coupling gains designed in these methods each node to self-organize its behavior (estimator and controller)
are illustrated in Fig. 7, where the adaptive gain designed in for a collective task (cooperative stabilization) only using local
this article is much smaller than the ones in [25], [26], and information and local interaction with its neighbors, which has
[27]. According to [26], this means that the estimator (20) the potential in promoting the warm intelligence of multiagent
allows a larger integration step, benefiting the reduction in the systems. Our future works will focus on the relevant security
communication frequency among nodes for distributed state issue.
estimation.
In addition, we consider the case where the above plant and
sensors are affected by noise, whose dynamics are redescribed APPENDIX A
by (21) without the inputs. Particularly, the unknown process PROOF OF LEMMA 1
and measurement disturbances in (21) are set as ω(t) = 0.2 ∗ For any strongly connected graph G, we have that the as-
sin(t/100) ∗ 1100 , νi (t) = 0.2 ∗ cos(t/100) ∀i = 1, . . ., 100. sociated Laplacian matrix L is a singular but irreducible M -
The proposed FDSE in (20) with θi redesigned in (24) is adopted matrix [33, Ch. 4.3.4] and there exists at least a positive principal
to estimate the plant state for all nodes, where  is set as 0.01 and element in Aj ∀j ∈ V. Further, according to [33, Corollary
other parameters remains uncharged. We compare the proposed 4.33], the matrix Lj is a nonsingular M -matrix. Similarly, we
FDSE with the DSEM based on parameter estimation [30], can prove that the matrix L̂ = L ⊗ Im + Â is a nonsingular
and AOCF [26], which can be regarded as the continuous-time M -matrix.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
6756 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 69, NO. 10, OCTOBER 2024

APPENDIX B It following from (12) and (5) that the first term on the right-hand
PROOF OF THEOREM 1 side of the above equation satisfies
Based on the notations defined above Theorem 1, it suffices to 
N 
N
demonstrate the convergence of x̃ and ζj with respect to time. To gij (γi + ψij )ψ̇ij
achieve this target, we consider a Lyapunov function candidate j=1 i=1
as follows:

N


N =− μ(ζ j )T [(Γ+Ψj )(Gj Lj +(Lj )T Gj )(Γ+Ψj )⊗Imj ]ζ j
V = β1 x̃T Qx̃ + V1j (25) j=1
j=1

N
with +2 μ(ζ j )T [(Γ + Ψj )Gj ⊗Imj ](Lj ⊗Cj Ā − αj ⊗Cj B̄)x̃.
j=1

N
gij 2
V1j = [(2γi + ψij )ψij + (γi − β) ]
i=1
2 Since

where β1 and β are positive constants to be determined later. 

N
− μ(ζ j )T [(Γ + Ψj )(Gj Lj + (Lj )T Gj )(Γ + Ψj )⊗Imj ]ζ j
Besides, gij > 0, i ∈ V, are the diagonal elements of any diag-
j=1
onal matrix Gj satisfying Gj Lj + (Lj )T Gj > 0, as shown in
Lemma 2. In addition, since Acl is assumed to be Hurwitz stable, 
N

there must exist a positive-definite matrix Q satisfying ≤− λ0j μ(ζ j )T [(Γ + Ψj )2 ⊗Imj ]ζ j
j=1
QAcl + ATcl Q  −W < 0.

N 
N 
N 
N

It is worth mentioning that β1 is introduced just for stability =− λ0j (γi + ψij )2 ψij ≤ − λ0 (γi + ψij )2 ψij
analysis, rather than the estimator or controller design. Then, j=1 i=1 j=1 i=1
the derivative of V1 with respect to time is derived by where λ0j denotes the minimal eigenvalue of the positive definite

N matrix Gj Lj + (Lj )T Gj and λ0 = minj∈V {λ0j }, and
V̇1 = −β1 x̃T W x̃ − 2β1 x̃T Q(IN ⊗F)ỹ + V̇1j . (26)

N
j=1
2 μ(ζ j )T [(Γ + Ψj )Gj ⊗Imj ](Lj ⊗Cj Ā − αj ⊗Cj B̄)x̃
By using the Young’s inequality, the second term on the right- j=1
hand side of (26) satisfies
 
N
λ0 μ j T j j 2
2σj2 μ
− 2β1 x̃T Q(IN ⊗F)ỹ ≤ (ζ ) [(Γ + Ψ ) ⊗Imj ]ζ +
j
x̃ W x̃
T

j=1
2 λ0 λmin (W)
(Q(IN ⊗F))  j T j 
2 N
β1 T 4β1 σmax N  N
≤ x̃ W x̃ + (ỹ ) ỹ λ0 2 N σ2 μ T
4 λmin (W) ≤ (γi + ψij )2 ψij + x̃ W x̃
j=1
j=1 i=1
2 λ0 λmin (W)
β1 T N 2
4β1 σmax (Q(IN ⊗F)) j T j
≤ x̃ W x̃ + (ζ ) ζ where σj = σmax (Gj Lj ⊗Cj Ā − Gj αj ⊗Cj B̄) and σ=
4 j=1
2 (L̂)
λmin (W)σmin maxj∈V {σj }, we have

β1 T 2
4β1 σmax (Q(IN ⊗F))  
N N 
N 
N
= x̃ W x̃ + ψij (27) gij (γi + ψij )ψ̇ij
4 2 (L̂)
μλmin (W)σmin j=1 i=1 j=1 i=1

where the second “≤” holds since L̂ is a nonsingular 2 N σ2 μ T N  N

λ0
M -matrix ≤ x̃ W x̃ − (γi + ψij )2 ψij .
according to Lemma 1, and the last “=” holds due to N ψ = λ0 λmin (W) 2
N i=1 ij j=1 i=1
μ(ζ ) ζ from (5). Further, the derivative of j=1 V1j with
j T j

respect to time is derived as Since γ̇i = N j=1 ψij , the second term on the right-hand side of
(28) satisfies

N 
N 
N
V̇1j = gij [(γi + ψij )ψ̇ij + (ψij + γi − β)γ̇i ] 
N 
N 
N 
N
j=1 j=1 i=1 − gij β γ̇i ≤ − g0 βN ψij
j=1 i=1 j=1 i=1

N 
N 
N 
N
= gij (γi + ψij )ψ̇ij − gij β γ̇i where g0 = mini,j∈V {gij }. Similarly, the last term on the right-
j=1 i=1 j=1 i=1 hand side of (28) satisfies

N 
N
N N 
N N  N 2 
+ gij (γi + ψij )γ̇i . (28) gij (γi +ψij )γ̇i ≤ ḡ0 N γi ψij + ψij
j=1 i=1
j=1 i=1 i=1 j=1 j=1

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
DUAN et al.: FRAMEWORK ON FULLY DISTRIBUTED STATE ESTIMATION AND COOPERATIVE STABILIZATION OF LTI PLANTS 6757


N 
N
1 
N
≤ ḡ0 N (γi + ψij )ψij ėx̂i = (A + BK + FC)ex̂i + F ỹi − F ỹj . (30)
j=1 i=1
N j=1

where ḡ0 = maxi,j∈V {gij } and the second “≤” is derived using When the matrices A + BK, A + FC and A + BK + FC are
⎛ ⎞2 Hurwitz stable, there must exist three positive definite matrices
N 
N P̄ , Q, and Q̄, respectively, satisfying
⎝ ψij ⎠ ≤ N 2
ψij .
j=1 j=1
X = −[P̄ (A + BK) + (A + BK)T P̄ ] > 0

Now, by choosing W = −[Q(A + FC) + (A + FC)T Q] > 0

2
4β1 σmax (Q(IN ⊗F)) 2ḡ02 N W̄ = −[Q̄(A + BK + FC) + (A + BK + FC)T Q̄] > 0
β= +
2 (L̂)
μg0 N λmin (W)σmin λ0 g 0 Now, we introduce a Lyapunov function candidate as
2
with β1 = λ0 λmin (W) ,
8 Nσ μ
and noticing that 
N
V2 = x̄T P̄ x̄ + β21 eTx̄ Qex̄ + β22 eTx̂ (IN ⊗Q̄)ex̂ + V1j

N 
N
λ0 2ḡ02 N 2 
N 
N j=1
− (γi + ψij )2 ψij − ψij (31)
j=1 i=1
2 λ0 j=1 i=1
where ex̂ = [eTx̂1 , . . . , eTx̂N ]T is the augmented form of ex̂i , i ∈

N 
N
V; the parameters β21 and β22 are positive constant scalars to be
≤ −2ḡ0 N (γi + ψij )ψij determined later; and V1j , j ∈ V, is defined in (25). It follows
j=1 i=1
from (29) and (30) that the derivative of V2 with respect to time
we have can be derived as

β1  N N 
N
V̇1 ≤ − x̃T W x̃ − ḡ0 N (γi + ψij )ψij V̇2 = V̇1j − x̄T X x̄ − β21 eTx̄ W ex̄ − β22 eTx̂ (IN ⊗W̄ )ex̂
2 j=1 i=1 j=1

 N 2 T T
β1 T + 2x̄T P̄ FCex̄ + 2β21 eTx̄ QB̄ex̂ − x̄ (1N ⊗P̄ F)ỹ
=− x̃ W x̃−μḡ0 N (ζ j )T [(Γ+Ψj )Gj ⊗Imj ]ζ j ≤ 0. N
2 j=1 2β21 T T
− e (1 ⊗QF)ỹ + 2β22 eTx̂ (H⊗Q̄F)ỹ (32)
N x̄ N
Therefore, V1 (t) is uniformly bounded, and so are x̃, ζj , and
1 1T
γi , which further implies the boundedness of x̃˙ and ζ̇ j . Since where H = IN − NN N . In the following, the last five terms
V1 ≥ 0, it has a finite limit V1∞ as t → ∞. That is, on the right-hand side of (32) are discussed. First, by utilizing
 the Young’s inequality, the terms 2x̄T P̄ FCex̄ and 2β21 eTx̄ QB̄ex̂
∞
β1 T N  N
satisfy, respectively,
x̃ W x̃+ḡ0 N (γi + ψij )ψij dt≤V1 (0)−V1∞ .
0 2 2
j=1 i=1 1 T 4σmax (P̄ FC)
2x̄T P̄ FCex̄ ≤ x̄ X x̄ + eT W ex̄
In light of the Barbalat’s Lemma [40], it can be concluded 4 λmin (X)λmin (W ) x̄
that limt→∞ x̃(t) = 0, limt→∞ ψij (t) = 0 and limt→∞ ζ j (t) = and
0 ∀i, j ∈ V. Thus, the proof of Theorem 1 is complete.  β21 T 2
4β21 σmax (QB̄) T
2β21 eTx̄ QB̄ex̂ ≤ ex̄ W ex̄ + e W̄ ex̂ .
4 λmin (W )λmin (W̄ ) x̂
APPENDIX C
PROOF OF THEOREM 3 Similarly, the last three terms satisfy
 2 T T 2β21 T T
First of all, let x̄ = N1 Nj=1 x̂j denote the average state 2β22 eTx̂ (H⊗Q̄F)ỹ− x̄ (1N ⊗P̄ F)ỹ− e (1 ⊗QF)ỹ
estimate of all nodes, whose dynamics can be derived as N N x̄ N
1 β21 T β22 T
1  ≤ x̄T X x̄ + e W ex̄ + e (IN ⊗W̄ )ex̂ + 4β0 ỹ T ỹ
N
x̄˙ = (A + BK)x̄ + FCex̄ − F ỹj (29) 4 4 x̄ 4 x̂
N j=1 where
2 2 2
where ex̄ = x̄ − x is the error between the average state estimate σmax (P̄ F) β21 σmax (QF) β22 σmax (Q̄F)
β0 = + + .
and the plant state. Further, let ex̂i = x̄ − x̂i denote the error N λmin (X) N λmin (W ) λmin (W̄ )
between the average state estimate and the state estimate of node
Particularly, it follows from (11) and Lemma 1 that the term ỹ T ỹ
i, i ∈ V. Then, the dynamics of the errors ex̄ and ex̂i can be
in the above inequality can be derived as
written as
 N
(ζ j )T ζ j 
N
(ζ j )T ζ j
1 
N N
ėx̄ = (A + FC)ex̄ + Bj Kj ex̂j − F ỹj ỹ ỹ ≤
T
2 ≤ .
N j=1 σ (Lj ) j=1 σmin
j=1 min
2 (L̂)
j=1

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
6758 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 69, NO. 10, OCTOBER 2024

Hence, we have unique positive-definite matrix Q1 such that


N
4β0 4β0 
N 
N
Q1 AT + AQ1 − Q1 C T CQ1 + T2 = 0. (34)
4β0 ỹ ỹ ≤
T
2 (L̂)
(ζ ) ζ =
j T j
2 (L̂)
ψij
σmin μσmin By choosing Q = Q−1
j=1 j=1 i=1 1 , we have
 N
since N i=1 ψij = μ(ζ ) ζ . Further, the term
j T j
j=1 V̇1j in (32)
AT Q + QA − C T C = −QT2 Q < 0 (35)
is derived
 as follows. Noting that x̃ = 1N ⊗ e x̄ − ex̂ , the term which indicates that the LMI (14) is solvable. In the following,
in N V̇
j=1 1j following (28) satisfies we will prove that there always exists a solution Q to (35)

N guaranteeing that the LMI (15) holds. Note that the LMI (15)
2 μ(ζ j )T [(Γ + Ψj )Gj ⊗Imj ](Lj ⊗Cj Ā − αj ⊗Cj B̄)x̃ can be rewritten as
j=1
 AT Q + QA − C T C − QBB T P1 − P1 BB T Q < 0
2 2
2σ 2N μ T 2 Nμ T
≤ ex̄ W ex̄ + e (IN ⊗W̄ )ex̂ where K = −B T P1 is substituted. To make the above LMI hold,
λ0 λmin (W ) λmin (W̄ ) x̂
it follows from (35) that we only need to guarantee that:

N 
N
λ0
+ (γi + ψij )2 ψij . QT2 Q > −QBB T P1 − P1 BB T Q
j=1 i=1
2
equivalently,
Altogether, by combining the derivation in Appendix B and
T2 > −BB T P1 Q1 − Q1 P1 BB T .
choosing
2 It suffices to prove that there exist positive-definite matrices Q1 ,
16σmax (P̄ FC) 16 N 2 σ 2 μ
β21 = + T2 , and P1 in (33) and (34) satisfying
λmin (X)λmin (W ) λ0 λmin (W )
2
2
λmin (T2 ) BB T 2 2
16β21 σmax (QB̄) 16 N σ 2 μ > κ1 + P1 2 (36)
β22 = + λmax (Q1 ) κ1
λmin (W )λmin (W̄ ) λ0 λmin (W̄ )
since
4β0 2ḡ02 N
β= + 1
2 (L̂)
μg0 N σmin λ0 g 0 −BB T P1 Q1 − Q1 P1 BB T ≤ κ1 Q1 + BB T P1 Q1 P1 BB T
κ1
after some complex but straightforward calculation, we have always holds for any positive scalar κ1 . By referring to the
1 β21 T β22 T argument following (33), for any positive scalar κ, we can find a
V̇2 ≤ − x̄T X x̄ − ex̄ W ex̄ − e (IN ⊗W̄ )ex̂ pair of solution T1 and P1 to (33) ensuring P1 2 ≤ κ. Hence,
2 4 2 x̂
for any T2 and Q1 satisfying (34), by setting κ1 = κ and

N 
N
− ḡ0 N (γi + ψij )ψij . λmin (T2 )
κ< 2)
j=1 i=1 λmax (Q1 )(1 + BB T 2
Following the same steps in the proof of Theorem 1, it can be the inequality (36) always hold. Thus, the proof of Theorem 4
concluded that x̄, ex̄ , ex̂ and ζ j asymptotically converge to zero. is complete.
Thus, the proof of Theorem 3 is complete. 
APPENDIX E
APPENDIX D PROOF OF THEOREM 6
PROOF OF THEOREM 4
To proceed, a useful lemma is given as follows.
To prove Theorem 4, it suffices to provide a feasible solution Lemma 3 ([41]): For any nonnegative scalars a and b, and
of P and Q to the LMIs (13)–(15) for any controllable and positive scalars p and q with 1/p + 1/q = 1, the inequality ab ≤
observable triple (A, B, C). First of all, when (A, B) is control- ap /p + bq /q always holds.
lable, for any positive definite matrix T1 , there always exists For the noisy plant (21), the dynamics of x̄, ex̄ , and ex̂i defined
a unique positive-definite matrix P1 such that the following in (29) and (30), respectively, can be rederived as
algebra Riccati equation holds [36, Ch. 14.2]:
1 
N
AT P1 + P1 A − P1 BB T P1 + T1 = 0. (33) x̄˙ = (A + BK)x̄ + FCex̄ − F ỹj −Fν
N j=1
Meanwhile, there always exist a pair of positive-definite matrices
T1 and P1 satisfying the above equation with P1 2 ≤ κ, where 
N
1 
N

κ is any positive scalar. By choosing P = P1−1 , we have ėx̄ = (A + FC)ex̄ + Bj Kj ex̂j − F ỹj −Fν − ω
j=1
N j=1
P AT + AP − BB T = −P T1 P < 0.
1 
N

Hence, the LMI (13) is solvable. Similarly, when (A, C) is ėx̂i = (A + BK + FC)ex̂i + F ỹi − F ỹj . (37)
N j=1
observable, for any positive-definite matrix T2 , there exists a

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
DUAN et al.: FRAMEWORK ON FULLY DISTRIBUTED STATE ESTIMATION AND COOPERATIVE STABILIZATION OF LTI PLANTS 6759

When A + BK, A + FC, and A + BK + FC are Hurwitz sta- where zj = Cj ω + ν̇j and other variables are defined the same
ble, there always exist positive-definite matrices P̄ , Q, and Q̄ as those in Appendix C. Further, note that
satisfying [42, Proposition 1], respectively,
− 2(ζ j )T [(Γ + Ψj )Gj αj ⊗zj ]
X = −[P̄ (A + BK) + (A + BK)T P̄ ] − δ P̄ > 0 √ 
= − 2(ζ j )T [(Γ − I) Gj ⊗Imj ][ Gj /αj ⊗zj ]
W = −[Q(A + FC) + (A + FC)T Q] − δQ > 0 √ √
− 2(ζ j )T [ Gj ⊗Imj ][ Gj αj ⊗zj ]
W̄ = −[Q̄(A + BK + FC) + (A + BK + FC)T Q̄] − δ Q̄ > 0  √
− 2(ζ j )T [ Ψj Gj /2αj ⊗Imj ][ 2Ψj Gj αj ⊗zj ]
where δ is any positive scalar. Now, we adopt the Lyapunov   
≤ (ζ j )T (Γ − I)2 Gj + (Ψj + IN )Gj ⊗Imj ζ j + f1j
function candidate (31), whose time derivative can be similarly
(40)
derived for the noisy plant (21) as
where “≤” is derived by utilizing Lemma 3 several times, and

N
f1j is expressed by
V̇2 = Z1 − 2x̄T P̄ Fν − 2β21 eTx̄ Q(Fν + ω) + V̇1j
j=1 f1j = (−1 + 1)Gj αj2 ⊗Imj 2 ( Cj 2 ωb + νd )2
√

N
+ 2 Gj αj2 ⊗Imj 22 ( Cj 2 ωb + νd )4
− δ(V2 − V1j ) (38)
j=1 since zj 2 ≤ Cj ωb + νd . Now, substituting (40) and (39)
and the dynamics of Z1 in Appendix C into (38) yields
where Z 1 denotes all the terms on the right-hand side of (32)
expect N 
N 
N 
N 
N
j=1 V̇1j . The second and third terms on the right-hand
V̇2 ≤ − (1 + ḡ0 δ) (γi + ψij )ψij + (1 + ψij )ψij
side of (38) satisfy, respectively,
j=1 i=1 j=1 i=1

1 T 4σ 2 (P̄ F)νb2 ⎛ ⎞
2x̄T P̄ Fν ≤ x̄ X x̄ + max 
N 
N 
N
4 λmin (X) − gij (γi −β)(γi − 1)2 +f2 − δ ⎝V2 − V1j⎠
j=1 i=1 j=1
and
1 β21 T β22 T
β21 T − x̄T X x̄ − e W ex̄ − e (IN ⊗W̄ )ex̂ (41)
2β21 eTx̄ Q(Fν + ω) ≤ e W ex̄ 4 8 x̄ 2 x̂
8 x̄
where parameters β21 and β22 are defined the same as those in
16β21 Q 22 ( F 22 νb2 + ωb2 ) Appendix C while β is rechosen as
+ .
λmin (W )
4β0 (ḡ0 N + 1 + ḡ0 δ)2
β= +
Besides, noticing that ζ j in (12) can be rederived as
2 (L̂)
μg0 N σmin 2λ0 g0 N
and f2 is a constant as
ζ̇ j = − [Lj (Γ + Ψj )⊗Imj ]ζ j + [Lj ⊗(Cj Ā) − αj ⊗(Cj B̄)]x̃

N 2
4σmax (P̄ F)νb2 16β21 Q 22 ( F 22 νb2 +ωb2 )
− αj ⊗(Cj ω + ν̇j ) f2 = μf1j + + .
j=1
λmin (X) λmin (W )
for the noisy plant (21), by following the similar derivation
process of Theorem 3, we have Further, since ri ≥ 1, we have

 
N 
N 
N

N
2σ 2 2
2N μ T 2 Nμ T V̇2 ≤ − δV2 − ḡ0 δ (γi + ψij )ψij + δ V1j
V̇1j ≤ ex̄ W ex̄ + e (IN ⊗W̄ )ex̂
j=1
λ0 λmin (W ) λmin (W̄ ) x̂ j=1 i=1 j=1


N 
N
 
N 
N − gij (γi − β)(γi − 1)2 + f2
−( βg0 N λ0 /2 − ḡ0 N ) (γi + ψij )ψij j=1 i=1
j=1 i=1
1 β21 T β22 T

N − x̄T X x̄ − ex̄ W ex̄ − e (IN ⊗W̄ )ex̂ . (42)
4 8 2 x̂
− 2μ(ζ j )T [(Γ + Ψj )Gj αj ⊗zj ]
j=1
Before moving on, it follows from Lemma 3 that the following
inequalities always hold:

N
1 16
− μ(ζ j )T [(Γ − I)2 Gj αj ⊗Imj ]ζ j −(γi − β)(γi − 1)2 ≤ − (γi − 1)3 + (β − 1)3
j=1 2 27
and

N 
N
− gij (γi − β)(γi − 1) 2
(39) δ 1 2δ 3 δ
(γi − β)2 ≤ (γi − 1)3 + + (β − 1)2 .
j=1 i=1 2 2 272 2

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
6760 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 69, NO. 10, OCTOBER 2024

Combining
N N the above two inequalities with the expression of [12] K.-K. Oh, M.-C. Park, and H.-S. Ahn, “A survey of multi-agent formation
i=1 in Appendix B, V̇2 in (42) satisfies
control,” Automatica, vol. 53, pp. 424–440, 2015.
j=1 [13] Y. Tang, F. Qian, H. Gao, and J. Kurths, “Synchronization in complex net-
works and its application—A survey of recent advances and challenges,”
1 β21 T
V̇2 ≤ − δV2 + fb − x̄T X x̄ − e W ex̄ (43) Annu. Rev. Control, vol. 38, no. 2, pp. 184–198, 2014.
4 8 x̄ [14] P. Duan, L. He, Z. Duan, and L. Shi, “Distributed cooperative LQR design
for multi-input linear systems,” IEEE Trans. Control Netw. Syst., vol. 10,
where fb is a constant about the noise bounds ωb , νb , and νd , no. 2, pp. 680–692, Jun. 2023.
explicitly expressed by as [15] J. Lavaei and A. G. Aghdam, “Overlapping control design for multi-
 channel systems,” Automatica, vol. 45, no. 5, pp. 1326–1331, 2009.
2 16 3 2δ 3 δ 2 [16] K. Liu, H. Zhu, and J. Lü, “Cooperative stabilization of a class of LTI
fb = f2 + N ḡ0 (β − 1) + + (β − 1) plants with distributed observers,” IEEE Trans. Circuits Syst. I: Regular
27 272 2 Papers, vol. 64, no. 7, pp. 1891–1902, Jul. 2017.
[17] K. Liu, Y. Chen, Z. Duan, and J. Lü, “Cooperative output regulation of
where f2 is defined following (41). According to [43, Lemma LTI plant via distributed observers with local measurement,” IEEE Trans.
2.5], V2 is uniformly bounded, which indicates that the adaptive Cybern., vol. 48, no. 7, pp. 2181–2191, Jul. 2018.
gain γ1 is uniformly bounded. Further, it follows from (43) that [18] L. Wang, D. Fullmer, F. Liu, and A. S. Morse, “Distributed control of linear
multi-channel systems: Summary of results,” in Proc. Amer. Control Conf.,
V2 ≤ − δV2 when 2020, pp. 4576–4581.
1 T β21 T [19] X. Zhang, K. Hengster-Movrić, M. Šebek, W. Desmet, and C. Faria,
x̄ X x̄ + e W ex̄ ≤ fb . “Distributed observer and controller design for spatially interconnected
4 8 x̄ systems,” IEEE Trans. Control Syst. Technol., vol. 27, no. 1, pp. 1–13,
Hence, x̄ and ex̄ converge, respectively, to the sets Jan. 2017.
[20] F. C. Rego, Y. Pu, A. Alessandretti, A. P. Aguiar, A. M. Pascoal, and C.
 
4fb 8fb N. Jones, “A distributed Luenberger observer for linear state feedback
x̄ : x̄ 22 ≤ , ex̄ : ex̄ 22 ≤ systems with quantized and rate-limited communications,” IEEE Trans.
λmin (X) β21 λmin (W ) Autom. Control, vol. 66, no. 9, pp. 3922–3937, Sep. 2021.
[21] S. Das and J. M. Moura, “Consensus+innovations distributed Kalman
with a convergence rate faster than eδt . Further, x in (21) are filter with optimized gains,” IEEE Trans. Signal Process., vol. 65, no. 2,
ultimately bounded by pp. 467–481, Jan. 2017.
[22] G. Battistelli, L. Chisci, and D. Selvi, “A distributed Kalman filter with
2 2 2 8fb 16fb event-triggered communication and guaranteed stability,” Automatica,
x 2 ≤ 2 x̄ 2 + 2 ex̄ 2 ≤Π + . vol. 93, pp. 75–82, 2018.
λmin (X) β21 λmin (W ) [23] X. He, C. Hu, Y. Hong, L. Shi, and H.-T. Fang, “Distributed Kalman
(44) filters with state equality constraints: Time-based and event-triggered
communications,” IEEE Trans. Autom. Control, vol. 65, no. 1, pp. 28–43,
Thus, the proof of Theorem 6 is complete. Jan. 2020.
[24] S. Park and N. C. Martins, “Design of distributed LTI observers for state
omniscience,” IEEE Trans. Autom. Control, vol. 62, no. 2, pp. 561–576,
REFERENCES Feb. 2017.
[1] Y.-H. Liu, Y. Xu, and M. Bergerman, “Cooperation control of multiple [25] T. Kim, C. Lee, and H. Shim, “Completely decentralized design of dis-
manipulators with passive joints,” IEEE Trans. Robot. Automat., vol. 15, tributed observer for linear systems,” IEEE Trans. Autom. Control, vol. 65,
no. 2, pp. 258–267, Apr. 1999. no. 11, pp. 4664–4678, Nov. 2020.
[2] D. Casbeer, R. Beard, T. McLain, S.-M. Li, and R. Mehra, “Forest fire [26] S. Battilotti, F. Cacace, M. D’Angelo, and A. Germani, “Asymptotically
monitoring with multiple small UAVs,” in Proc. Amer. Control Conf., 2005, optimal consensus-based distributed filtering of continuous-time linear
vol. 5, pp. 3530–3535. systems,” Automatica, vol. 122, 2020, Art. no. 109189.
[3] R. D’Andrea and G. E. Dullerud, “Distributed control design for spatially [27] P. Duan, J. Qian, Q. Wang, Z. Duan, and L. Shi, “Distributed state
interconnected systems,” IEEE Trans. Autom. Control, vol. 48, no. 9, estimation for continuous-time linear systems with correlated measure-
pp. 1478–1495, Sep. 2003. ment noise,” IEEE Trans. Autom. Control, vol. 67, no. 9, pp. 4614–4628,
[4] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and coop- Sep. 2022.
eration in networked multi-agent systems,” Proc. IEEE, vol. 95, no. 1, [28] S. Wang and W. Ren, “On the convergence conditions of distributed dy-
pp. 215–233, Jan. 2007. namic state estimation using sensor networks: A unified framework,” IEEE
[5] G. Wen, Z. Duan, W. Yu, and G. Chen, “Consensus of multi-agent systems Trans. Control Syst. Technol., vol. 26, no. 4, pp. 1300–1316, Jul. 2018.
with nonlinear dynamics and sampled-data information: A delayed-input [29] R. Ortega, E. Nuno, and A. Bobtsov, “Distributed observers for LTI
approach,” Int. J. Robust Nonlinear Control, vol. 23, no. 6, pp. 602–619, systems with finite convergence time: A parameter-estimation-based ap-
2013. proach,” IEEE Trans. Autom. Control, vol. 66, no. 10, pp. 4967–4974,
[6] F. Dörfler, J. W. Simpson-Porco, and F. Bullo, “Breaking the hierarchy: Oct. 2021.
Distributed control and economic optimality in microgrids,” IEEE Trans. [30] R. Ortega, E. Nuno, and A. Bobtsov, “An algebraic, distributed state
Control Netw. Syst., vol. 3, no. 3, pp. 241–253, Sep. 2016. observer for continuous-and discrete-time linear time-invariant systems
[7] Y. Lv, Z. Li, Z. Duan, and J. Chen, “Distributed adaptive output feedback with time-varying communication graphs,” Int. J. Adaptive Control Signal
consensus protocols for linear systems on directed graphs with a leader of Process., vol. 36, no. 6, pp. 1340–1352, 2022.
bounded input,” Automatica, vol. 74, pp. 308–314, 2016. [31] T. Kim, D. Lee, and H. Shim, “Decentralized design and plug-and-play
[8] P. Duan, K. Liu, N. Huang, and Z. Duan, “Event-based distributed tracking distributed control for linear multi-channel systems,” IEEE Trans. Autom.
control for second-order multiagent systems with switching networks,” Control, to be published, doi: 10.1109/TAC.2023.3293036.
IEEE Trans. Syst., Man, Cybern. Syst., vol. 50, no. 9, pp. 3220–3230, [32] P. A. Ioannou and J. Sun, Robust Adaptive Control. Upper Saddle River,
Sep. 2020. NJ, USA: Courier Corporation, 1996.
[9] J. Ferber and G. Weiss, Multi-Agent Systems: An Introduction to Dis- [33] Z. Qu, Cooperative Control of Dynamical Systems: Applications to Au-
tributed Artificial Intelligence. Reading, MA, USA: Addison-Wesley, tonomous Vehicles. London, U.K.: Springer-Verlag, 2009.
1999. [34] L. Bakule, “Decentralized control: An overview,” Annu. Rev. Control,
[10] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of vol. 32, no. 1, pp. 87–98, 2008.
agents with switching topology and time-delays,” IEEE Trans. Autom. [35] S. Battilotti, F. Cacace, and M. D’Angelo, “A stability with optimality
Control, vol. 49, no. 9, pp. 1520–1533, Sep. 2004. analysis of consensus-based distributed filters for discrete-time linear
[11] Y. Cao, W. Ren, and M. Egerstedt, “Distributed containment control with systems,” Automatica, vol. 129, 2021, Art. no. 109589.
multiple stationary or dynamic leaders in fixed and switching directed [36] K. Zhou et al., Robust and Optimal Control. Englewood Cliffs, NJ, USA:
networks,” Automatica, vol. 48, no. 8, pp. 1586–1597, 2012. Prentice Hall, 1996.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.
DUAN et al.: FRAMEWORK ON FULLY DISTRIBUTED STATE ESTIMATION AND COOPERATIVE STABILIZATION OF LTI PLANTS 6761

[37] X. Fang, L. Xie, and X. Li, “Distributed localization in dynamic networks Guanghui Wen (Senior Member, IEEE) re-
via complex Laplacian,” Automatica, vol. 151, 2023, Art. no. 110915. ceived the Ph.D. degree in mechanical sys-
[38] R. Ji, S. S. Ge, and D. Li, “Saturation-tolerant prescribed control for non- tems and control from Peking University, Beijing,
linear systems with unknown control directions and external disturbances,” China, in 2012.
IEEE Trans. Cybern., vol. 54, no. 2, pp. 877–889, Feb. 2024. He is currently a Young Endowed Chair Pro-
[39] P. Zhu and W. Ren, “Fully distributed joint localization and target tracking fessor with the Department of Systems Sci-
with mobile robot networks,” IEEE Trans. Control Syst. Technol., vol. 29, ence, Southeast University, Nanjing, China. His
no. 4, pp. 1519–1532, Jul. 2021. research interests include autonomous intelli-
[40] H. Khalil, Nonlinear Systems. Englewood Cliffs, NJ, USA: Prentice-Hall, gent systems, complex networked systems, dis-
2002. tributed control and optimization, resilient con-
[41] D. S. Bernstein, Matrix Mathematics: Theory, Facts, and Formulas. Prince- trol, and distributed reinforcement learning.
ton, NJ, USA: Princeton Univ. Press, 2009. Dr. Wen is an IET Fellow. He was the recipient of the National Sci-
[42] Z. Li, Z. Duan, and G. Chen, “Dynamic consensus of linear multi-agent ence Fund for Distinguished Young Scholars, and Australian Research
systems,” IET Control Theory Appl., vol. 5, no. 1, pp. 19–28, 2011. Council Discovery Early Career Researcher Award. He is a Reviewer
[43] Y. Lv, Z. Li, Z. Duan, and G. Feng, “Novel distributed robust adaptive for American Mathematical Review and is an Active Reviewer for many
consensus protocols for linear multi-agent systems with directed graphs journals. He currently serves as an Associate Editor for IEEE TRANS-
and external disturbances,” Int. J. Control, vol. 90, no. 2, pp. 137–147, ACTIONS ON INDUSTRIAL INFORMATICS, IEEE TRANSACTIONS ON NEURAL
2017. NETWORKS AND LEARNING SYSTEMS, IEEE TRANSACTIONS ON INTELLIGENT
VEHICLES, IEEE JOURNAL OF EMERGING AND SELECTED TOPICS IN INDUS-
TRIAL ELECTRONICS, IEEE TRANSACTIONS ON SYSTEMS, MAN AND CYBER-
NETICS: SYSTEMS, IEEE OPEN JOURNAL OF THE INDUSTRIAL ELECTRONICS
SOCIETY, and Asian Journal of Control. He has been named a Highly
Cited Researcher by Clarivate Analytics since 2018.

Peihu Duan received the B.Eng. degree in

mechanical engineering from the Huazhong Maciej Ogorzałek (Fellow, IEEE) received the
University of Science and Technology, Wuhan, Ph.D. degree in electrical engineering from the
China, in 2015, and the Ph.D. degree in me- AGH University of Science and Technology,
chanical systems and control from Peking Uni- Krakow, in 1987. He held several visiting posi-
versity, Beijing, China, in 2020. tions in Denmark, Switzerland, Germany, Spain,
From 2020 to 2022, he was a Postdoc with Japan, and Hong Kong. He worked for one year
the Hong Kong University of Science and Tech- as a Guest Professor with the National Micro-
nology and the University of Hong Kong, Hong electronic Center, Sevilla, Spain. He is currently
Kong, respectively. He currently works as a a Professor of Electrical Engineering and Com-
Postdoc with the School of Electrical Engineer- puter Science and Head of the Department of
ing and Computer Science, KTH Royal Institute of Technology, Stock- Information Technologies, Jagiellonian Univer-
holm, Sweden. His research interests include cooperative control and sity, Krakow, Poland. He has authored or coauthored over 350 technical
state estimation of networked systems. papers published in journals and conference proceedings, an authored
the book Chaos and Complexity in Nonlinear Electronic Circuits (World
Scientific, 1997) .
Dr. Ogorzałek is a Member of the Polish Academy of Sciences and
Member of the European Academy of Sciences (Academia Europaea).
He was the recipient of the Research Award from the Ministry of Ed-
ucation of Spain in 2000. He was the recipient of the Senior Award
from the Japan Society for Promotion of Science as visiting Professor
at Kyoto University and in 2005 Hertie Foundation Fellowship at The
Goethe Universty Frankfurt-am-Main. During 2006–2009, he held the
Chair of Biosignals and Systems, Hong Kong Polytechnic University
under the Distinguished Scholars Scheme. He was also the recipient
Yuezu Lv (Senior Member, IEEE) received the of the 2002 Guillemin-Cauer Award, IEEE-CAS Golden Jubilee Award
B.S. degree in engineering mechanism and and IEEE CAS Meritorious Service Award. He gave over 40 plenary
Ph.D. degree in mechanical systems and control and keynote lectures at major conferences world-wide. He served as
from the College of Engineering, Peking Univer- Editor-in-Chief for Circuits and Systems Magazine from 2004 to 2007,
sity, Beijing, China, in 2013 and 2018, respec- Associate Editor for IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS
tively. PART I, from 1993 to 1995 and 1999–2001; he was elected Member
From 2018 to 2021, he worked as a Lec- of the Editorial Board Proceedings of the IEEE from 2004 to 2009. He
turer with the Department of Systems Science, serves also as an Associate Editor for Journal of the Franklin Institute
School of Mathematics, Southeast University, from 1997 to present, Member of the Editorial Board of the International
Nanjing, China. He is currently an Associate Journal of Bifurcation and Chaos, Secretary of the Editorial Board for
Researcher with Advanced Research Institute the Quarterly of Electrical Engineering from 1993–2000, Member of
of Multidisciplinary Sciences, Beijing Institute of Technology, Beijing, the Editorial Board of Automatics (both in Polish), and Member of the
China. His research interests include cooperative control of multiagent Editorial board of the International Journal of Circuit Theory and Ap-
systems, adaptive control, robust control of uncertain systems, and plications from 2000 to present and Associate Editor of the Nonlinear
distributed resilient control. Theory and Its Applications Journal IEICE Japan. He served the IEEE
Dr. Lv was a finalist for Zhang Si-Ying (CCDC) Outstanding Youth Circuits and Systems Society (CAS) in various capacities including VP
Paper Award in 2015. He was the recipient of the 2021 APNNS Young for Region 8, Administrative Vice-president and finally 2008 Society
Researcher Award by Asia Pacific Neural Network Society, and the President. He was CAS Society Distinguished Lecturer from 2004 to
Lotfi A. Zadeh Best Conference Paper Award at IEEE ICCSS 2022. He 2005 and received the 2002 Guillemin-Cauer Award, IEEE-CAS Golden
was selected for the fifth Young Elite Scientists Sponsorship Program Jubilee Award and IEEE CAS Meritorious Service Award. He currently
by CAST in 2020. He is an Associate Editor for IEEE Systems, Man, serves as IEEE Division 1 Director (2016-2017), Member of the IEEE
and Cybernetics Magazine, and a young editorial board member of Board of Directors. Until 2006, he was an Executive Vice-President of
International Journal of Dynamics and Control. the Sniadecki Science Foundation in Poland.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:23:52 UTC from IEEE Xplore. Restrictions apply.