Contents lists available at ScienceDirect
Systems & Control Letters
journal homepage: www.elsevier.com/locate/sysconle
Stability analysis of largescale distributed networked control systems with random communication delays: A switched system approach
Kooktae Lee ^{∗} , Raktim Bhattacharya
Department of Aerospace Engineering, Texas A&M University, College Station, TX 778433141, USA
a r t i c l e
i n f o
Article history:
Received 3 December 2014 Received in revised form 22 April 2015 Accepted 10 August 2015 Available online 18 October 2015
Keywords:
Largescale distributed networked control system Markov jump linear system Switched system Random communication delays
a b s t r a c t
In this paper, we consider the stability analysis of largescale distributed networked control systems with random communication delays. The stability analysis is performed in the switched system framework, particularly as the Markov jump linear system. There have been considerable research on stability analysis of the Markov jump systems. However, these methods are not applicable to largescale systems because large numbers of subsystems result in extremely large number of switching modes. To circumvent this
scalability issue, we propose a new reduced mode model for stability analysis, which is computationally
scalable. We also consider the case in which the transition probabilities for the Markov jump process contain uncertainties. We provide a new method that estimates bounds for uncertain Markov transition probability matrix to guarantee the system stability. Numerical example verifies the computational efficiency of the proposed methods.
© 2015 Elsevier B.V. All rights reserved.
1. Introduction
A networked control system (NCS) is a system that is controlled over a communication network. Recently, NCSs have attracted con siderable research interest due to emerging networked control applications. For example, NCSs are broadly used in applications including traffic monitoring, networked autonomous mobile agents, chemical plants, sensor networks, and distributed software systems in cloud computing architectures. Due to the communica tion network, communication delays or communication losses may occur, resulting in performance degradation or even instability. Therefore, it has led various researchers to analyze NCSs associated with communication delays [1–8]. Particularly in [6], the NCS with communication delays was analyzed by adopting the switched sys tem [7–11], which refers to the dynamical system consists of a family of subsystems and a switching logic governing switching between subsystems. In this paper, we study largescale distributed networked con trol system (DNCS), which denotes NCS with a large number of spatially distributed subsystems (or agents). For such largescale systems, our primary goal is to analyze system stability when ran dom communication delays exist. Typically, the system behavior
^{∗} Corresponding author. Email addresses: animodor@tamu.edu (K. Lee), raktim@tamu.edu (R. Bhattacharya).
01676911/© 2015 Elsevier B.V. All rights reserved.
with random communication delays have been widely modeled as Markov jump linear system (MJLS) [6,12–16] where the switching sequence is governed by a Markovian process. Since stability has been one of the major concerns, considerable effort has been on stability analysis of MJLS [17,10,18–20]. However, these results are only applicable to the systems with a small number of switching modes. Largescale DNCSs, in which we are particularly interested, give rise to an extremely large number of switching modes. Thus, previous conditions developed for the stability analysis of MJLSs cannot be evaluated for largescale DNCSs as they are not compu tationally tractable. Although the literature [21] recently investi gated the switched system that circumvents computation issues associated with a large number of switching modes, it is devel oped for independent and identically distributed (i.i.d.) switching. We consider Markovian switching in this paper. In addition, we are also interested in largescale DNCSs where the transition probabil ities are inaccurately known [20,22,23]. This can happen because in practice it is difficult to accurately estimate the Markov tran sition probability matrix that models the random communication delays. This paper provides two key contributions to analyze the sta bility of largescale DNCSs with random communication delays. Firstly, we guarantee the mean square stability of such systems by introducing a reduced mode model. We prove that the mean square stability for individual switched system implies a neces sary and sufficient stability condition for the entire DNCS. This drastically reduces the number of modes necessary for analysis.
78 K. Lee, R. Bhattacharya / Systems & Control Letters 85 (2015) 77–83
Secondly, we present a new method to estimate the bound for un certain Markov transition probability matrix, in order to guarantee the system stability. These results enable us to analyze largescale systems in a computationally tractable manner. Rest of this paper is organized as follows. We introduce the problem for the largescale DNCS in Section 2. Section 3 presents the switched system framework for the stability analysis with communication delays. In Section 4, we propose the reduced mode model to efficiently analyze stability. Section 5 quantifies the stability region and bounds for uncertain Markov transition probability matrix. This is followed by the application of the proposed method to an example system in Section 6, and we conclude the paper with Section 7.
Notation. The set of real numbers is denoted by R. The symbols ∥·∥ and ∥ · ∥ _{∞} stand for the Euclidean and infinity norm, respectively. The symbol #(·) denotes the cardinality—the total number of elements in the given set. In addition, the symbols tr(·), ρ(·), ⊗, and diag(·) represent trace operator, spectral radius, Kronecker product, and block diagonal matrix operator, respectively.
2. Problem formulation
2.1. Distributed networked control system with no delays
agent i with communication delays can be expressed as:
x _{i} (k + 1) = ^{} A _{i}_{j} x _{j} (k ^{∗} ),
j∈N _{i}
i = 1, 2,
. . .
, N,
(3)
where k ^{∗} k − τ . Note that we have no communication delays when i = j because there is no communication in this case. The random communication delay, represented by the term k ^{∗} , forms a stochastic process. To analyze the stability of the DNCS, we define an augmented state X(k) as X(k) [x(k) ^{⊤} , x(k −
1) ^{⊤} ,
. . .
, x(k − τ _{d} ) ^{⊤} ] ^{⊤} ∈ R ^{N}^{n}^{q}^{×}^{N}^{n}^{q} , where q τ _{d} + 1. Then, the
dynamics for the entire system is given by
X(k + 1) = W(k)X(k),
(4)
where W(k) 
A ˜ _{1} (k) 0 I · · · · · · · · · 0 0 0 0 A ˜ _{2} (k) A ˜ _{q}_{−}_{1} (k) A ˜ _{q} (k) 
_{∈} 
_{R} Nnq×Nnq , 

. . . . . . 0 . . . · · · . . . I . . . 0 

0 
the matrix I denotes an identity matrix with proper dimensions,
and the timevarying matrices
˜
A _{j} (k) ∈ R ^{N}^{n}^{×}^{N}^{n} , j =
1, 2,
. . .
, q,
model the randomness in the communication delays between
neighboring agents.
3. Switched system approach
Consider a discretetime dynamics of each agent in the DNCS, given by: 
Without loss of generality, the dynamics of the largescale DNCS with communication delays in (4) can be transformed into 

x _{i} (k + 1) = ^{} A _{i}_{j} x _{j} (k), 
i = 1, 2, 
, N, 
(1) 
a switched system framework as follows: 

j∈N _{i} 
x(k + 1) = W _{σ}_{(}_{k}_{)} x(k), 
σ(k) ∈ {1, 2, 
, m}, 
(5) 

where k is a discretetime index, N is the total number of agents (subsystems), x _{i} ∈ R ^{n} is a state for the ith agent, N _{i} is a set of neighbors for x _{i} including the agent x _{i} itself, and A _{i}_{j} ∈ R ^{n}^{×}^{n} is a timeinvariant system matrix that represents the linear intercon nections between agents. Note that we have A _{i}_{j} = 0 if there is no interconnection between the agents i and j. To represent the entire systems dynamics, we define the state 
where W _{σ}_{(}_{k}_{)} is the timeinvariant matrix, representing commu nication delays in agents, {σ(k)} is the switching sequence, and m is the total number of switching modes. When the switch ing sequence {σ(k)} is stochastic, (5) is referred to as a stochas tic switched linear system or a stochastic jump linear system, according to the literature [7]. For the stochastic switched lin ear system, the switching sequence {σ(k)} is governed by the 

x(k) ∈ R ^{N}^{n}^{×}^{N}^{n} as x(k) [x _{1} (k) ^{⊤} , x _{2} (k) ^{⊤} , 
, x _{N} (k) ^{⊤} ] ^{⊤} . Then, the 
modeoccupation switching probability π(k) = [π _{1} (k), π _{2} (k), 

system dynamics of the DNCS is given as 
, π _{m} (k)], where π _{i} is a fraction number, satisfying ^{} 
i=1 ^{π} i 

x(k + 1) 
= Ax(k), 
(2) 
and 0 ≤ π _{i} ≤ 1, ∀i. In this case, each π _{i} denotes the modal framework has been widely employed [12–16]. Thus, we make the 

with the following definition for the matrix A ∈ R ^{N}^{n}^{×}^{N}^{n} 
probability corresponding to each mode dynamics W _{i} . In order to properly describe the behavior of random communication delays, it is necessary to adopt a certain switching logic, which is used to 

A A A 
11 
A 
12 
A 
13 
· 
· 
· 
A 
1N 


21 
A A 
22 
A A 
23 
· 
· 
· 
A A 
2N 
update the switching probability π(k). For this purpose, the MJLS 

A 
31 
32 
33 
· 
· 
· 
3N 

, 
following assumption in our analysis. 

. . . 
. . . 
. . . 
. 
. 

. 
. 
. . 

A N1 A ij = 0, A _{i}_{j} , 
A N2 A N3 · · · A NN if no connection between the agents i and j, 
• Assumption: Consider the stochastic jump linear system (5) with the switching probability π(k) = [π _{1} (k), π _{2} (k), π _{m} (k)]. Then, π(k) is updated by the Markovian process given , 

otherwise. 
by π(k+1) = π(k)P, where P ∈ R ^{m}^{×}^{m} is the Markov transition probability matrix. 
For the discretetime system in (2), it is well known that the system is stable if and only if the condition ρ(A) < 1 is satisfied. Throughout the paper we assume that the system without commu nication delays, defined in (2), is stable. Then, we address stability in the presence of random communication delays. We remind the reader that N is very large.
2.2. DNCS with communication delays
Often, network communication between agents encounter time delays or packet losses while sending and receiving data. We
denote the symbol τ as random communication delays and assume
that τ has a discrete value bounded by 0 ≤
τ ≤ τ _{d} < ∞, where
τ _{d} is a finitevalued maximum delay. Then, the dynamics for the
Since the MJLS is a family of the stochastic switched linear system, various stability notions can be defined [10]. In this paper, we will consider the mean square stability condition, defined below.
Definition 3.1 (Definition 1.1 in [11]). The MJLS is said to be mean square stable if for any initial condition x _{0} and arbitrary initial probability distribution π(0), lim _{k}_{→}_{∞} E ^{} ∥x(k, x _{0} )∥ ^{2} ^{} = 0.
The total number of switching modes m depends on the size q and N. Since the communication delays take place independently while receiving and sending the data for each agent, m is calculated by counting all possible scenarios to distribute every matrices A _{i}_{j} ∈ R ^{n}^{×}^{n} for i ̸= j in the block matrix A ∈ R ^{N}^{n}^{×}^{N}^{n} given in (2), into
K. Lee, R. Bhattacharya / Systems & Control Letters 85 (2015) 77–83
79
˜
each A _{j} (k) ∈ R ^{N}^{n}^{×}^{N}^{n} , j = 1, 2,
. . .
, q, given in (4), which results in
m = q ^{N}^{(}^{N}^{−}^{1}^{)} . Note that the largescale DNCS that has considerably
large N results in extremely large m, which makes current analysis
tools useless as they are computationally intractable.
Before we further proceed, we introduce the following propo
sition that was developed for the stability analysis of the MJLS.
Proposition 3.1 (Theorem 1 in [18]). The MJLS with the Markov
transition probability matrix P is mean square stable if and only if
ρ ^{}^{} P ^{⊤} ⊗ I ^{} diag(W _{j} ⊗ W _{j} ) ^{} < 1,
(6)
where I is an identity matrix with a proper dimension,
diag(W _{j} ⊗ W _{j} )
where P ^{i} ∈ R ^{m} ^{i} ^{×}^{m} ^{i} is the transition probability matrix for the
reduced mode MJLS given in (7), I is an identity matrix with a proper
dimension, N is the total number of the agents in the system, m _{i} =
q ^{n}^{ˆ} ^{i} ^{(}^{n}^{ˆ} ^{i} ^{−}^{1}^{)} is the total mode numbers for the reduce mode MJLS, and
diag( W
j
ˆ
i
⊗
ˆ
i
W
j
)
(
ˆ
W
i
1 ^{⊗}
0
.
.
.
0
0
ˆ
W
i
_{1} )
0
(
ˆ
i
W
2 ^{⊗}
ˆ
i
_{2} )
W
0
0
0
0
.
.
.
.
.
.
·
·
·
·
·
·
( W ˆ
i
m _{i} 1 ^{⊗}
0
m _{i} 1 ^{)}
W
ˆ
i
0
0
.
.
.
^{0}
(
m i ^{⊗} m i ^{)}
W
W
ˆ
i
ˆ
i
.
Proof. Let the matrix Q ^{i} (k) be of the form Q ^{i} (k) E[xˆ _{i} (k)xˆ _{i} (k) ^{⊤} ].
Then, Q ^{i} (k) is alternatively obtained by the following equation:
80 K. Lee, R. Bhattacharya / Systems & Control Letters 85 (2015) 77–83
the following recursion equation:
yˆ ^{i} (k)
=
p
^{i}
11
^{(}
ˆ
W
i
1
p
^{i}
12
^{(}
ˆ
W
i
1
^{⊗}
^{⊗}
.
. .
p ^{i}
1m _{i} ^{(}
ˆ
W
i
1 ^{⊗}
ˆ
W
i
_{1}
)
ˆ
W
i
_{1}
)
ˆ
W
i
_{1} )
p
^{i}
21
^{(}
ˆ
i
W
2
p
^{i}
22
^{(}
ˆ
i
W
2
^{⊗}
^{⊗}
ˆ
i
_{2} )
W
ˆ
i
_{2} )
W
. . .
p
^{i}
2m _{i} ^{(}
ˆ
i
W
2 ^{⊗}
ˆ
i
_{2} )
W
.
.
.
.
.
.
.
.
.
.
.
.
p ^{i}
m _{i} 1 ^{(}
m i ^{⊗} m i ^{)}
W
W
ˆ
i
ˆ
i
p ^{i}
m _{i} 2 ^{(}
m i ^{⊗} m i ^{)}
W
W
ˆ
i
ˆ
i
. . .
p ^{i}
m i m i ^{(}
m i ^{⊗} m i ^{)}
W
W
ˆ
i
ˆ
i
y ^{i}
_{1} (k − 1)
y ^{i}
_{2} (k − 1)
y ^{i}
.
. .
_{m} _{i} (k − 1)
.
=(P ^{i} ^{⊤} ⊗I)diag( W
ˆ
i
j
⊗
i
ˆ
W
j
)
=yˆ ^{i} (k−1)
From the above equation, it is clear that ρ (P ^{i} ^{⊤} _{⊗} _{I}_{)}_{d}_{i}_{a}_{g}_{(} W
ˆ
i
j
⊗
W )
ˆ
i
j
<
1 implies lim _{k}_{→}_{∞} yˆ ^{i} (k)
=
0, and hence this leads to
lim _{k}_{→}_{∞} Q ^{i} (k) = 0 ⇐⇒ lim _{k}_{→}_{∞} tr ^{} Q ^{i} (k) ^{} = 0 ⇐⇒ lim _{k}_{→}_{∞} E
^{} ∥xˆ _{i} (k)∥ ^{2} ^{} = 0, which is the sufficient mean square stabil
ity condition for xˆ _{i} (k). On the other hand, if we have ρ (P ^{i} ^{⊤} _{⊗}
I)diag( W
j
ˆ
i
⊗
W ) >
ˆ
i
j
1, then yˆ ^{i} (k) will diverge, resulting in ne
cessity for the mean square stability of xˆ _{i} (k). Hence, the spectral
radius being less than one is the necessary and sufficient mean
square stability condition for the state xˆ _{i} (k). Further, we have
lim _{k}_{→}_{∞} E ^{} ∥xˆ _{i} (k)∥ ^{2} ^{} = 0, ∀i = 1, 2,
. . .
, N ⇐⇒ lim _{k}_{→}_{∞} E
^{} ∥x(k)∥ ^{2} ^{} = 0, where x(k) is the state for the DNCS defined in (5).
This concludes the proof.
Remark 4.1. Theorem 4.1 provides an efficient way to analyze
the stability of largescale DNCSs. The key idea stems from
the hypothesis that the stability of each subsystem obtained
by decomposing the original system will provide the stability
of the entire system. Without any relaxation or conservatism,
Theorem 4.1 provides the necessary and sufficient condition for
stability, which is equivalent to (6) developed for the mean square
stability of the MJLS. Compared to the total number of modes in
the full state model (5), which is q ^{N}^{(}^{N}^{−}^{1}^{)} , the reduced mode model
^{N}
(7) has total ^{} _{i}_{=}_{1}
q ^{n}^{ˆ} ^{i} ^{(}^{n}^{ˆ} ^{i} ^{−}^{1}^{)} modes. Consequently, the growth of
mode numbers in full state model is exponential with respect to
N ^{2} , whereas that in reduced mode model is linear with regard to
N. Therefore, Theorem 4.1 can be checked in a computationally
efficient manner.
5. Stability region and stability bound for uncertain Markov
transition probability matrix
The Markov transition probability matrix can be obtained from
data of communication delays. In general, this Markov transition
probability matrix obtained from statistics can be interpreted as
representation of random communication delays in an average
manner. Thus, one cannot estimate the exact transition probability
in practice, which leads to uncertainty in the Markov transition
probability matrix. In this section, we aim at developing a tool
to measure the stability bound for uncertain Markov transition
probability matrix, in order to guarantee the stability of largescale
DNCSs with uncertain Markovian communication delays.
As explained in Section 3, the dimension of the Markov
transition probability matrix for the full state model is given by
P ∈ R ^{m}^{×}^{m} , where the number of switching modes is m = q ^{N}^{(}^{N}^{−}^{1}^{)} .
For fairly large N, it is practically intractable to handle such a
largescale matrix. For example, even with q = 2 and N =
100, the mode number is given by m = 2 ^{1}^{0}^{0}^{×}^{9}^{9} , resulting in
P ∈ R ^{2} ^{1}^{0}^{0}^{×}^{9}^{9} ^{×}^{2} ^{1}^{0}^{0}^{×}^{9}^{9} . However, the reduced mode model developed
in Section 3 renders computation tractable because it results in
huge decrease in the mode numbers as well as the dimension of
each Markov transition probability matrix, by decomposing the
original system into N numbers of reduced mode systems. As a
consequence, the reduced mode model yields N numbers of new
Markov transition probability matrix P ^{i} , i = 1, 2,
has small size.
. . .
, N, which
Here we assume that the new Markov transition probability
matrix associated with the each reduced mode is given by the
following form: P ^{i} =
¯
P ^{i} + P ^{i} , i
=
1, 2,
. . .
¯
, N, where P ^{i} is the
nominal value and P ^{i} is the uncertainty in the Markov transition
probability matrix for ith subsystem. Due to the variation in P ^{i} ,
the system stability may change and hence we want to estimate the
bound for P ^{i} , to guarantee the system stability. Here we assume
that P ^{i} has the following structure:
P ^{i}
p
^{i}
11
p
^{i}
21
.
.
.
p
^{i}
m _{i} 1
^{}
^{p}
i
_{1}_{2}
^{}
^{p}
i
_{2}_{2}
.
.
.
^{} ^{p}
i
_{m} _{i} _{2}
·
·
·
·
·
·
.
.
.
·
·
·
p
p
^{i}
1m
^{i}
_{i}
2m
_{i}
.
.
.
p ^{i}
m i m i
, ∈ R ^{m} ^{i} ^{×}^{m} ^{i}
s.t.
m
i
s=1
p ^{i}
_{r}_{s} = 0, ∀r = 1, 2,
. . .
,
m _{i} .
(9)
Since we have a constraint such that the row sum has to be zero
for P ^{i} in above equation, we aim to find the feasible maximum
bound for each row, ε _{r} , satisfying the inequality  p ^{i}
i
_{r}_{s}  ≤ ε _{r} , ∀r,
i
i
in order to guarantee the system stability. Then, each ε _{r} for r =
1, 2,
. . .
, m _{i} can be obtained by the following two steps.
Step 1: Solve Linear Programming (LP)
maximize 1 ^{⊤} z 
(for upper bound) 
(10) 

or minimize 1 ^{⊤} z 
(for lower bound) 

subject to Az < b _{s} , ∀s = 1, 2, 
, m _{i} 
(11) 

lb _{s} ≤ z _{s} ≤ ub _{s} , ∀s = 1, 2, 
, m _{i} (12) 

where 

z _{s} [ p ^{i} _{1}_{s} , p ^{i} _{2}_{s} , z [z ⊤ ⊤ i ^{]} ^{⊤} ⊤ 
, p ^{i} m _{i} s ^{]} ⊤ 
, 

1 
, z 
2 
, 
. . . , z m 
, 

A α _{1} , α _{2} , j = 1, 2, 
, α _{m} _{i} , with α _{j} ∥ ˆ W , m _{i} , 
ˆ 

i j ⊗ 
W j 


m 
i 

b _{s} 1 − 

α _{r} p¯ ^{i} rs ^{,} 

r=1 

lb _{s} [−p¯ ^{i} _{1}_{s} , −p¯ _{2}_{s} ^{i} , 
, −p¯ ^{i} m _{i} s ^{]} ⊤ 
, 

ub _{s} [1 − p¯ ^{i} _{1}_{s} , 1 − p¯ _{2}_{s} ^{i} , . . . , 1 − m _{i} s ^{]} ⊤ p¯ ^{i} . The inequality constraint (11) in the LP problem guarantees the 

mean square stability by Lemma 5.1 and Theorem 5.1. The terms lb _{s} 

and ub _{s} in (12) are the lower and upper bounds for z _{s} , respectively, 

according to 0 ≤ p ^{i} _{r}_{s} = (p¯ ^{i} _{r}_{s} + p ^{i} _{r}_{s} ) ≤ 1. Step 2: Obtain Feasible Solution with Hyperplane Constraint 

We can compute the feasible maximum bound for p ^{i} rs 
^{a}^{s} 

follows. 

ε i _{r} = min ^{} min(ε _{r}_{,}_{l}_{b} ), min(ε _{r}_{,}_{u}_{b} ) ^{} , i i r = 1, 2, , m _{i} . (13) where ε _{r}_{,}_{l}_{b} [(∆p ^{i} i r1 ^{)} ⋆ _{l}_{b} ,(∆p ^{i} r2 ^{)} ⋆ _{l}_{b} , ,(∆p ^{i} rm ^{)} lb ^{⋆} ^{]} ^{⊤} , ε _{r}_{,}_{u}_{b} [( p ^{i} i r1 ^{)} ⋆ _{u}_{b} ,(∆p ^{i} r2 ^{)} ⋆ _{u}_{b} , ,(∆p ^{i} rm ^{)} ^{⋆} _{u}_{b} ] ^{⊤} , and (∆p ^{i} _{l}_{b} ,(∆p ^{i} rs ^{)} ^{⋆} rs ^{)} ub ^{⋆} denote optimal lower and upper bounds for ∆p _{r}_{s} ^{i} , obtained from 

the LP, respectively. 
Since upper or lower bounds are solved by maximizing or
minimizing the objective function, ( p _{r}_{s} ^{i} ) ^{⋆} has different values for
upper and lower bounds. Fig. 1 shows the geometry of stability
region analysis for uncertain transition probability matrix. The
K. Lee, R. Bhattacharya / Systems & Control Letters 85 (2015) 77–83
81
Fig. 1. The geometry of the Stability Region Analysis for the uncertain Markov transition probability matrix when m _{i} = 2. Each region is described in the figure.
region S _{1} stands for the bounds that come from −p¯ ^{i}
_{r}_{s}
≤ p ^{i}
rs
^{≤}
1 − p¯ ^{i}
_{r}_{s} . S _{2} can be obtained from inequality constraint (11). The
region S _{3} denotes the solution from the LP, and S is the feasible
maximum bound with a stability guarantee. Note that P ^{i} satisfies
m s=1 ^{} ^{p} ^{i}
i
_{r}_{s} = 0, ∀r and hence, feasible solutions should lie on the
hyperplane, satisfying p ^{i}
_{r}_{1} + p ^{i}
_{r}_{2} +· · ·+ p ^{i}
rm
i
= 0, ∀r. Therefore,
we can compute the feasible maximum bound from (13) for each
row r.
Now we prove that the inequality constraint (11) guarantees
the system stability. For this purpose, Lemma 5.1 will be used.
Lemma 5.1. Consider a block matrix X defined by
X =
X 11
X 21
.
.
.
X m1
X 12
X 22
.
.
.
X m2
·
·
·
·
·
·
.
.
.
X 1m
X 2m
.
.
.
· · · X mm
,
where matrix X _{i}_{j} ∈ R ^{n}^{×}^{n} . Then, we have ρ _{(}_{X}_{)} < 1, if ^{}
< 1, ∀i = 1, 2,
. . .
, m.
m
j=1
^{} X ij ^{}
^{} ∞
Proof. For the block matrix X given above, the following inequality
m
condition ∥X∥ _{∞} ≤ max _{i} ^{} _{j}_{=}_{1} ∥X _{i}_{j} ∥ _{∞} holds. Also, it is well known
that ρ(X) ≤ ∥X∥ _{p} for any choice of p.
Therefore, we conclude that ^{}
m
_{j}_{=}_{1} ∥X _{i}_{j} ∥ _{∞} < 1, ∀i = 1, 2,
. . .
, m ⇒ ρ(X) ≤ ∥X∥ _{∞} < 1.
Theorem 5.1. Consider the MJLS (5) for the largescale DNCS with
communication delays. Then, (5) is mean square stable if
m
i
r=1
α _{r}  p _{r}_{s}
^{i}  < β _{s} ,
∀s = 1, 2,
. . .
, m _{i} ,
∀i = 1, 2,
. . .
, N
where α _{r} = ∥ W
r
ˆ
i
satisfied.
⊗
ˆ
i
W
r
∥ _{∞} and β _{s} = 1 − ^{} ^{m}
i
r=1 ^{p}^{¯} rs
^{i}
^{∥}
ˆ
i
W
r
⊗
W ∥ _{∞} , is
r
ˆ
i
Proof. If the Markov transition probability matrix for the system
¯
in (7) has the uncertainty denoted by P ^{i} = P ^{i} + P ^{i} , then the term
ρ (P ^{i} ^{⊤}
⊗ I)diag(
ˆ
i
W
j
⊗
W ) in (8) can be expressed as
ˆ
i
j
ρ (P ^{i} ^{⊤}
⊗ I)diag(
ˆ
i
W
j
⊗
ˆ
i
W
j
)
= ρ ^{} (
ˆ
P ^{i} + P ^{i} ) ^{⊤} ⊗ I ^{} diag( W
¯
i
j
⊗
ˆ
i
W
j
)
=
ρ (
¯
P
^{i} ^{⊤} ⊗ I) + (∆P ^{i} ^{⊤} ⊗ I) diag(
ˆ
i
W
j
⊗
ˆ
i
W
j
)
= ρ (
¯
P
i
^{⊤}
⊗ I)diag( W
j
ˆ
i
⊗
ˆ
i
W
j
) + (∆P ^{i} ^{⊤} ⊗ I)diag(
ˆ
i
W
j
⊗
ˆ
i
W
j
)
≤ ∥(
¯
P
i
^{⊤}
⊗ I)diag( W
j
ˆ
i
⊗
ˆ
i
W
j
) + (∆P ^{i} ^{⊤} ⊗ I)diag(
ˆ
i
W
j
⊗
ˆ
i
W
j
)∥ _{∞}
≤ ∥(
¯
P
i ^{⊤}
⊗ I)diag( W
j
ˆ
i
⊗
ˆ
i
W
j
)∥ _{∞}
+ ∥(∆P ^{i} ^{⊤} ⊗ I)diag(
ˆ
i
W
j
⊗
W )∥ _{∞} .
j
ˆ
i
(14)
In the first inequality, we used the fact that ρ(·) ≤ ∥ · ∥ _{∞} and
the submultiplicative property was applied in the last inequality.
The block matrix structure for each term of the last inequality
is alternatively expressed as the equation in Box I, where γ _{j} =
(
ˆ
i
W
j
⊗
W ), j = 1, 2,
j
ˆ
i
. . .
, m _{i} , and similarly,
_{}_{(}_{∆}_{P} ^{i} ^{⊤} ⊗ I)diag( W
ˆ
i
j
⊗
ˆ
i
W
j
)
^{} ∞
=
γ _{1} p _{1}_{1} ^{i}
γ _{1} p _{1}_{2} ^{i}
.
.
.
γ _{1} p _{1}_{m} ^{i} _{i}
γ _{2} p ^{i}
_{2}_{1}
γ _{2} p ^{i}
_{2}_{2}
.
.
.
γ _{2} p ^{i}
_{2}_{m} _{i}
γ _{m} _{i} p ^{i}
m _{i} 1
γ _{m} _{i} p ^{i}
m _{i} 2
.
.
.
^{} ∞
·
·
·
·
·
·
.
.
.
· · · γ _{m} _{i} p ^{i}
m i m i
.
By applying the result in Lemma 5.1 into (14), it is guaranteed
that ρ (P ^{i} ^{⊤} ⊗ I)diag( W
ˆ
i
j
⊗
W ) < 1, if the following condition
ˆ
i
j
m
i
r=1
α _{r}  p ^{i}
rs ^{} ^{+}
m
i
r=1
^{i}
α _{r} p¯ _{r}_{s}
_{<} _{1}_{,}
∀s = 1, 2,
. . .
, m _{i} ,
where α _{r} ∥ W
r
ˆ
i
⊗
W ∥ _{∞} , is satisfied.
r
ˆ
i
Therefore, (5) is mean square stable by Theorem 4.1 if it is
guaranteed that
m
i
r=1
α _{r}  p _{r}_{s}
^{i}  < β _{s} ,
∀s = 1, 2,
∀i = 1, 2,
. . .
, m _{i} ,
. . .
, N,
where β _{s} 1 − ^{} ^{m}
i
r=1 ^{α} r ^{p}^{¯} rs
^{i}
^{.}
6. Numerical example We consider the DNCS that has random communication delays,
where the total number of agents is denoted by N = 100. The
state for each agent is updated by the following scalar discrete
time dynamics:
x _{i} (k +
1) =
^{} a _{i}_{j} x _{j} (k ^{∗} ),
j∈N _{i}
i = 1, 2,
. . .
, N,
where a ij ^{} 0.7,
0,
j = i,
−0.14, j
∈ {i − 1, i + 1},
otherwise.
(15)
When there is no communication delay (i.e., k ^{∗} = k), the dynamics
for the entire DNCS is obtained by (2) as follows:
X(k + 1) = AX(k),
A =
0.7
−0.14
0
.
. .
0
0
−0.14
0.7
−0.14
·
·
0
0
−0.14
0.7
.
.
.
·
·
·
·
0
0
−0.14
.
.
.
−0.14
0
·
·
·
·
·
·
·
.
.
.
0.7
−0.14
0
0
0
. . .
−0.14
0.7
_{∈} _{R} N×N
.
In this case, the stability is computed by ρ(A) = 0.9799 < 1,
which assures that the given DNCS without communication delays
is stable. However, communications (e.g., wireless network) may
be delayed in practice and thus, above stability result is no longer
guaranteed. Therefore, it is required to test the system stability in
the existence of random communication delays.
6.1.
Stability analysis with random communication delays
We assume that the communication delay τ is bounded by
0 ≤ τ
≤ τ _{d}
=
1, i.e., k ^{∗}
= {k, k − 1}, ∀i =
1, 2,
. . .
, N, which
82 K. Lee, R. Bhattacharya / Systems & Control Letters 85 (2015) 77–83