1

Robustness Analysis of Asynchronous Sampled-Data

Multi-Agent Networks With Time-Varying Delays
Feng Xiao, Member, IEEE, Yang Shi, Fellow, IEEE, Wei Ren, Fellow, IEEE
arXiv:1710.02014v1 [math.OC] 4 Oct 2017

Abstract

In this paper, we study the simultaneous stability problem of a finite number of locally inter-connected linear subsystems under
practical constraints, including asynchronous and aperiodic sampling, time-varying delays, and measurement errors. We establish
a new Lyapunov-based stability result for such a decentralized system. This system has a particular simple structure of inter-
connections, but it captures some key characteristics of a large class of intermediate models derived from the consensus analysis
of multi-agent systems. The stability result is applicable to the estimation of the maximum allowable inter-sampling periods and
time delays based on individual dynamics and coupling structures in the scenarios of consensus control via asynchronous sampling
of relative states and asynchronous broadcasting of self-sampled states respectively. The asynchrony of aperiodic sampling and the
existence of measurement errors allow the utilization of some kinds of quantizing devices, such as Logarithmic quantizers, in the
process of data sampling, and allow the introduction of a period of dwell time after each update of state measurement to eliminate the
Zeno behavior of events in event-based control. The extension in the case with input saturations and input delays is also discussed.

Keywords

Asynchronous multi-agent systems, simultaneous stability, consensus, aperiodic sampling, time-varying delays, event-triggered
control.

I. I NTRODUCTION

Due to the attractive advantages in signal processing and transmission, digital devices have found their wide applications in
modern control systems [1]. An analog-to-digital converter is responsible for converting a continuous-time signal into a digital
signal by sampling and quantization. A digital controller gathers all input signals and compute an output to achieve the desired
control purpose. Sampled-data control deals with such hybrid continuous-time and discrete-time systems and usually assumes
that these digital devices share the same clocks and process their data periodically and synchronously [2]. This ideal assumption
can formulate the considered digital systems as standard LTI discrete-time systems by plant discretization and servers as the
basis of many fundamental results, for example, in stability and optimal control [3], [4]. When these digital components, such
as A/D samplers and D/A zero-order holds, work at different frequencies or they are spatially scattered in a large area without a
central clock, the sampled-data systems become inherently asynchronous [5]. Although synchronous sampled-data models may

This work was supported by the National Natural Science Foundation of China under Grant 61422302, the Program for New Century Excellent Talents in
University under Grant NCET-13-0178, and the 111 Project under Grant B17048.
F. Xiao is with the School of Control and Computer Engineering, North China Electric Power University, Beijing 102206, China (e-mail: [email protected]).
Yang Shi is with the Department of Mechanical Engineering, University of Victoria, Victoria, BC, V8N 3P6, Canada (e-mail: [email protected]).
Wei Ren is with the Department of Electrical and Computer Engineering, University of California, Riverside, CA 92521 USA (email: [email protected]).
2

be used to approximate the asynchronous ones in some of such cases [6], this treatment could sacrifice some dynamic details in
plants.
Literature Review: Asynchronous sampling has been brought to the attention of researchers in the control community for more
than three decades [7]. It has obtained enormous results, and is now developing rapidly with the growth of multi-agent theory
as one of the main research areas. Generally, asynchronous sampling can be found in the following three types of systems:
(1) Multi-rate systems. Multi-rate sampling is one of the earliest motivating examples for studying asynchronous systems,
where sampling and hold elements work periodically at different rates with irrational ratios [5], [6], [8]–[10]. Examples include
dual-rate linear systems with a single sampler and a single hold [5], [8] and multi-rate linear systems with multiple asynchronous
samplers and zero-order holds [6], [9], [10]. The corresponding control problems include the optimal LQG control [5], [8] and
stability problems [6], [9], [10].
(2) Networked control systems (NCSs). Theoretically, some NCSs can be modeled as multi-rate systems. Due to the long-
distance transmission of information, communication networks in NCSs may suffer from packet loss and time delays [11]–[15]
and often have multiple independent information channels [12]–[14], [16]. To ensure effective information transmission with
reduced costs, asynchronous event-triggered sampling also has been considered in NCSs [14], [15]
(3) Multi-agent systems. The asynchronous property in signal processing becomes more prominent in multi-agent networks, in
which, finite numbers of subsystems, equipped with independent signal-sensing and data-processing devices, are inter-connected
to perform some cooperative tasks [17], [18]. For example, in the formation control of multi-robot systems, each robot should
detect the positions and velocities of other adjacent robots for route planning. However, the installed sensors, such as ultrasonic
sensors or laser sensors, usually cannot monitor all objects in 360-degree coverage at a time. Furthermore, they could also
be affected by environmental interference. So in such cases, it is practically preferable that each robot collects the data of its
neighbors in some order (not necessarily periodically) according to its own schedule and within its sensing/processing capacities.
These kinds of realistic scenarios raise the problem of asynchronous and aperiodic sampling. The sensor scheduling can be
time-driven [19]–[21], event-driven [22], or mixture (of time and event)-driven [23], [24].
With asynchronous sampling, even if the involved digital devices operate periodically, sampled-data systems become non-
periodic and time-varying [8]. Asynchronous sampling can destroy the stability of systems which are stable in synchronous
environments [25], and induce additional time-varying delays in system analysis [26]. So the analysis of asynchronous systems
is more difficult and challenging than their synchronous counterparts, and the asynchronous mechanism also sets a strict
requirement on the robustness of designed controllers or algorithms with respect to aperiodic sensor scheduling, sensing errors
and processing/communcation delays. In asynchronous multi-agent systems, most results have been developed based on the
simple individual models of single-integrators [24] and double-integrators [19], [27], [28], but relatively few studies have been
done on the general linear models, with the exception of the event-triggered control [29]–[31].
Contributions: This paper will establish a Lyapunov-based stability result for an asynchronous multi-agent system with the
consideration of the above-mentioned practical issues and then apply it to solving several representative asynchronous consensus
problems in one framework. The contributions are summarized as follows:
First, we set up a basic stability model for asynchronous coordination of multi-agent systems and solve the problem of how
to estimate the maximum length of sampling intervals by subsystem matrices and inter-connection structures. In the model,
all subsystems, represented by the general linear state-space models, share the same system matrix and feedback matrix. They
communicate with each other via discrete-time signals produced by samplers and zero-order holds. Examples of such a model
without sampling can be found in the coordination analysis of a large range of multi-agent systems [32], [33]. We impose
 3

relaxed assumptions on the A/D and D/A devices and information channels, including asynchronous and aperiodic sampling,
measurement errors, and time-varying delays. The proposed result well describes the robustness of the asynchronous sampled-
data system based on the stability of the continuous-time system without sampling. In [26], by a Lyapunov-Krasovskii method
and LMIs, the authors studied the protocol design problem of a similar model. But it is different from ours in the following two
aspects: (1) in our model, given any inter-connection structure, the subsystems are inherently coupled by discrete-time signals; in
[26], the subsystems are coupled by continuous-time signals and the feedback matrices and scheduling protocols for discrete-time
signals need to be designed accordingly; (2) In our model, the information is aperiodically sampled with measurement errors
and transmitted with time delays; in [26], the information is sampled and transmitted according to the proposed Round-Robin
(RR) or Try-Once-Discard (TOD) protocol. Note that the sampling with time-varying periods is also referred to as asynchronous
sampling in [34]–[36] but it is a different definition from the “asynchronous sampling” in this paper.
Second, we solve the sampled-data consensus problems of the following asynchronous multi-agent systems in one framework:
(1) networks of general linear agents with asynchronous sampling of relative states; (2) networks of single-integrators and
marginally stable systems with asynchronous broadcasting of self-sampled states. Feedbacks with relative states and broadcast
communication have been widely used in the consensus and formation control of multi-agent systems; particularly, broadcasting
is an important way of keeping the state average unchanged; see [28], [37]–[39], [41]–[45] and references therein. Our work on
asynchronous broadcasting is partly inspired by this observation and the protocol design method presented for the event-based
control of single-integrator and double-integrator agents in [43]. Due to the challenging difficulties of asynchronous consensus
analysis, the existing results on sampled-data consensus mainly focus on synchronous sampling; that is, all data should be sampled
at the same time with constant or variant frequencies. The involved individual models are usually single-integrators [46], double-
integrators [47]–[49], or second-order oscillators [50]. There are relatively rare reports on asynchronous sampled-data consensus
of general linear agents and they mostly deal with event-triggered consensus [29]–[31], which is intrinsically different from the
time-driven style of asynchrony in this paper. Our work is also different from the previous results based on dynamic outputs of
controllers [19], [27].
Finally, taking advantages of asynchronous sampling and allowable measurement errors, we can extend the obtained results
further to deal with some of the general cases with quantization in discretizing continuous-time signals, and we rigorously prove
the effectiveness of the method for eliminating the Zeno behavior by introducing a dwell time after each measurement in event-
based control. To the best of our knowledge, the related problems on general linear multi-agent systems have not been studied
in the existing literature. This extension is illustrated by the implementation of Logarithmic quantizers [51] and event-triggering
conditions based on state errors [23], [39]. We also show the application of the results in the case with input saturations and
input delays.
This paper is organized as follows. In Section II, we set up the asynchronous sampled-data model and give sufficient conditions
for stability by a Lyapunov-based approach. In Section III, the stability result is applied to solving some asynchronous consensus
problems. In Section IV, further extensions of the stability result are discussed. The paper is concluded in Section V.
Notations: In denotes the identity matrix in Rn×n ; ⊗ denotes the Kronecker product; 1n = [1, 1, . . . , 1]T ∈ Rn .

II. LYAPUNOV- BASED ASYNCHRONOUS STABILITY

In this paper, we are interested in the protocol design and robustness/effectiveness analysis in coordinating agents’ states in a
general scenario of aperiodic asynchronous sampling and time-varying transmission delays.
4

Assume that there are m subsystems in a multi-agent system with zero-order holds and their states are represented by zi (t),
i = 1, 2, . . . , m, respectively, in a common state space RN . The dynamics of subsystem i is given as follows:
m
X
żi (t) = Azi (t) − gij K ẑj (t), (1)
j=1

where A ∈ RN ×N is the common system matrix of subsystems, gij represents the coupling weight, K ∈ RN ×N can be viewed
as the common state feedback gain, and ẑj (t) is the sampled state of subsystem j.
Let tik , k = 0, 1, 2, . . . , be the sampling instants of zi (t), τki , k = 0, 1, 2, . . . , be the corresponding time delays in sampling,
transmission, or computation, and eik , k = 0, 1, 2, . . . , be the measurement errors. The sampled state ẑi (t), incorporating time
delays, is given by
i
ẑi (t) = zi (tik ) − eik , t ∈ [tik + τki , tik+1 + τk+1 ), k = 0, 1, 2, . . . ,

which is a piece-wise constant function of time t. Here, the sequences of {tik }, {τki }, and {eik }1 are completely independent
with respect to different i, and satisfy the following assumption:
Assumption 1: For k = 0, 1, 2, . . . ,
(1) tik+1 − tik ≤ h;
(2) τki < tik+1 − tik ;
(3) τki ≤ τ ;
T T
(4) eik eik ≤ ω ẑi (tik + τki ) ẑi (tik + τki ),
where h is the maximum sampling period, τ is the maximum time delay, and ω ≥ 0 .
Remark 1: (1) The simultaneous stability model, with each individual represented by system (1), originates from the
consensus coordination of multi-agent systems. Several examples will be given in Section III. Furthermore, this model is
also of interest in the networked situation where some individual systems cannot stabilize themselves by their sampled
self-states especially when they cannot access their own states due to device failure. Particularly, when m = 1, system
(1) can be viewed a networked control system with aperiodic discrete-time signals and time-varying delays, which is also
of its own significance [40].
(2) Assumption 1 (4) says that the measurement error is of multiplicative type and proportional to the measurement. The
additive errors can be modeled as uncertainties, which will be considered in Section III.
(3) It can be shown that Assumption 1 (4) can be ensured by
T ω
eik eik ≤ √ 2 zi (tik )T zi (tik ).
(1 + ω)
(4) Note that the sequences of {tik }, {τki }, and {eik } are only indexed by the ith subsystem. So it can be understood that the
information is transmitted via broadcasting. Another interpretation will be given in Section III.
Denote z(t) = [z1 (t)T z2 (t)T . . . zm (t)T ]T and ẑ(t) = [ẑ1 (t)T ẑ2 (t)T . . . ẑm (t)T ]T . The overall system is given by

ż(t) = (Im ⊗ A)z(t) − (G ⊗ K)ẑ(t), (2)

where G = [gij ] ∈ Rm×m .

Let λAs denote the largest eigenvalue of (1/2)(A + AT ) and let σA , σG , and σK denote the largest singular values of A, G,
and K, respectively. We have the following result:
1 Notations {tik }, {τki }, and {eik } stand for the sequences ti0 , ti1 , . . . , τ0i , τ1i , . . . , and ei0 , ei1 , . . . , respectively.
 5

Theorem 1: In system (1), assume that there exists a lower bounded function V (t) with the following property:
dV (t)
≤ −µẑ(t)T ẑ(t) + ε(z(t) − ẑ(t))T(z(t) − ẑ(t)), (3)
dt
where µ > 0 and ε > 0. If Assumption 1 holds for all i and there exist some positive numbers α and β, such that
1 1
µ − εω(1 + )(1 + )e2λAs (h+τ )
α β
 1
− ε (1 + α)(1 + )σA 2 (4)
β
7 2 2
+ (1 + β) σG σK (h + τ )2 e2λAs (h+τ ) > 0,
3
then
lim z(t) = 0.
t→∞

Proof: The proof is based on the method of “analytic synchronization” [18], [19] and given in the Appendix.
Remark 2: (1) In (3), if ε ≤ 0, the stability analysis becomes trivial. So we only consider the case with positive ε.
(2) The left side of (4) is a continuous function with respect to variables h and τ . If h = τ = 0, then the left side converges
to µ − εω as parameters α and β approach to ∞. Therefore, when εω < µ, from (4), it can be seen that we can always
find maximum allowable sampling period h and the corresponding maximum time delay τ to ensure the stability of the
system. However, if εω ≥ µ, then the decrease of V (t) cannot be decided by (3) and thus the stability of system (2)
cannot be decided either.
(3) By Theorem 1, the maximum allowable h and τ can be calculated by (4), where there exists a trade-off between τ and
h. The larger the h is, the smaller the upper bound of allowable time delays is.

III. A SYNCHRONOUS CONSENSUS

Consider a group of n linear autonomous agents interacting with each other through local information flow. Label these agents
with 1 to n and suppose that the i-th agent takes the following dynamics:

ẋi (t) = Axi (t) + Bui (t), (5)

where xi (t) ∈ RN is the state, ui (t) ∈ RM is an input signal, called protocol in multi-agent coordination and designed based on
the information received from locally linked agents, and system matrix A and input matrix B are with compatible dimensions.

A. Asynchronous sampling of relative states

Assume that the interaction topology is modeled by an undirected simple graph G with vertices v1 , v2 , . . . , vn , which represent
the n agents respectively. Assume that there are total m edges, labeled with 1, 2, · · · , m. The existence of an edge (vi , vj )
represents that agents i and j have the capacity of knowing the relative state xi − xj at the same time, and in such a case,
suppose that agents i and j sample the relative state xi − xj at discrete times tij ij ij
k , and they receive the data xi (tk ) − xj (tk )
with measurement error eij ij ij ij ji ij ji ij ji
k at time tk + τk , k = 0, 1, 2, . . . . Here, tk = tk , τk = τk and ek = −ek . For different pair of
adjacent agents, the initial sampling of relative states may start at different times, and their subsequent sampling instants are also
independent. This is referred to as asynchrony, which is an intrinsic characteristic in distributed networks without any global
clocks to synchronize agents’ actions. Our assumption about time delay τkij is also quite general. Time delay τkij can be changing
with respect to k and independent of those on other channels. Let Ni denote the set of all js, such that (vi , vj ) exists, and let
6

tij (t) = tij ij ij ij ij ij ij

k , e (t) = ek , t ∈ [tk + τk , tk+1 + τk+1 ), k = 0, 1, 2, . . . . In what follows, if (vi , vj ) is indexed by p (the p-th
edge), tpk , τkp , and tp (t) are also used instead of tij ij ij
k , τk , and t (t).

With the utilization of zero-order holds, the protocol ui (t) takes the following form:
X
ui (t) = K (xj (tij (t)) − xi (tij (t)) + eij (t)). (6)
j∈Ni

We will design the feedback matrix K and give sufficient conditions in terms of sampling period and time delays to ensure that
Pn
limt→∞ (xi (t) − (1/n)eAt j=1 xj (0)) = 0, which implies that the system solves the average consensus problem [37].
Remark 3: In the protocol, we require that tij ji ij ji ij ji
k = tk , τk = τk and ek = −ek , which ensure the symmetry of information
sharing and are the least requirement for distributed average consensus. To get this symmetric information, each pair of adjacent
agents should communicate beforehand to synchronize each mutual sampling and ensure the sampled data, like xi (tij ij
k ) − xj (tk ),
used in both controllers, is updated at the same time. The synchronization of data sampling and controller update for any pair
of adjacent agents is technically possible because it is only based on local communication. In many engineering applications,
such as formation control and attitude alignment, relative states (position, attitude, etc), as a whole, are easily obtainable, which
is one of the reasons why we assume the same time delays on both parts of each relative state [41].
1) Edge dynamics [52]: Assign each edge an arbitrary direction in interaction topology G and define the n × m incidence
matrix D = [dij ] by (see [54])

 1, if vi is the head of the oriented edge j,


dij = −1, if vi is the tail of the oriented edge j,


 0, otherwise.

Denote x(t) = [x1 (t)T x2 (t)T . . . xn (t)T ]T and z(t) = (DT ⊗IN )x(t), which is the vector obtained by stacking the relative states
corresponding to edges 1, 2, . . . , m in sequence. Let z(t) = [z1 (t)T z2 (t)T . . . zm (t)T ]T with zp (t) ∈ RN , p = 1, 2, . . . , m. For
the pth edge (vj , vi ), if it is oriented from vj to vi in the definition of D, denote zp (t) = xi (t)− xj (t), ẑp (t) = zp (tp (t))− eij (t),
and epk = eij T
k , k = 0, 1, . . . . Denote ẑ(t) = [ẑ1 (t) ẑ2 (t)
T
. . . ẑm (t)T ]T . The system (5) with protocol (6) has the following
equivalent representation:

ẋ(t) = (In ⊗ A)x(t) − (D ⊗ BK)ẑ(t),

and

ż(t) = (Im ⊗ A)z(t) − (DT D ⊗ BK)ẑ(t). (7)

Matrix DDT is called the graph Laplacian of G, which is independent of the selection of D [53]. Let the eigenvalues of
DDT be λ1 , λ2 , · · · , λn in the increasing order. Then λ1 = 0. λ2 is called the algebraic connectivity of G, which is positive
when G is connected [54]. It can be observed that matrices DDT and DT D share the same non-zero eigenvalues with the same
algebraic multiplicities. So DT D also serves the purpose of algebraic characterization of G and is called edge Laplacian [52].
2) Average consensus: The feedback matrix K is designed with the requirement that there exist a positive definite matrix P
and a positive number µ satisfying the following Lyapunov inequalities:

(A + λi BK)T P + P (A + λi BK) + 2µIN ≤ 0, i = 2, 3, . . . , n. (8)

 7

Let λP BKs denote the largest eigenvalue of (1/2)(P BK + K T B T P ), σ denote the largest singular value of matrix (DT D ⊗
P BK) − 2µInN , σBK denote the largest singular value of matrix BK, and

S = {γ :γ > 0, µ − σ/(2γ) > 0,

γσ/2 − µ + λn λP BKs > 0}.

Theorem 2: In system (5), assume that (A, B) is stabilizable, the interaction topology G is connected, and Assumption 1 holds
for any i. If there exist positive numbers α and β such that
1 1
ω(1+ )(1 + )e2λAs (h+τ )
α β
 1
+ (1 + α)(1 + )σA 2
β
7 2  (9)
+ (1 + β) λn σBK 2 (h + τ )2 e2λAs (h+τ )
3
σ
µ − 2γ
< sup γσ ,
γ∈S 2 − µ + λn λP BKs

then protocol (6) solves the average consensus problem; mathematically,

 n 
1 At X
lim xi (t) − e xi (0) = 0.
t→∞ n i=1

Remark 4: (1) In the proof of Theorem 2, we will see that if there exists γ > 0 such that µ − σ/(2γ) > 0 and γσ/2 −
µ + λn λP BKs ≤ 0, then the derivative of the employed Lyapunov candidate is always negative (in the case with nonzero
ẑ(t); see (13)), and thus the system solves the average consensus problem for any sampling periods, which is usually not
possible for the systems with zero-order holds. So Theorem 2 only considers the case that γ ∈ S.
(2) Since (A, B) is stabilizable, we can always find proper matrices K, P , and positive number µ such that inequalities (8)
holds. One feasible solution K = B T P can be obtained by solving the following Riccati equation in the LQR problem
with 0 < λ ≤ λ2 :

P A + AT P − 2λP BB T P = −2µIN . (10)

Note that the above Riccati equation has been used in the design of consensus protocols in the literature [29], [33].
Furthermore, λ can be estimated only by the number of agents without the knowledge of the algebraic connectivity of G
[55].
Proof of Theorem 2: Consider the following Lyapunov candidate:

1
V (t) = z(t)T (Im ⊗ P )z(t).
2
Then by (7),

dV (t)
=z(t)T (Im ⊗ P )ż(t)
dt  1 
=z(t)T Im ⊗ (P A + AT P ) − DT D ⊗ P BK z(t)
2
+ z(t)T (DT D ⊗ P BK)(z(t) − ẑ(t)). (11)
8

Let DDT = C −1 ΛC, where C −1 = C T and Λ is the diagonal matrix with eigenvalues of DT D in the diagonal positions. In
equation (11), we have that
 1 
z(t)T Im ⊗ (P A + AT P ) − DT D ⊗ P BK z(t)
2
1
=x(t) (C −1 ⊗ IN )(Λ ⊗ (P A + AT P )
T
2
2
+ Λ ⊗ P BK)(C ⊗ IN )x(t)
≤ − x(t)T (C −1 ⊗ IN )(Λ ⊗ µIN )(C ⊗ IN )x(t)
≤ − µz(t)T z(t). (12)

Substituting inequality (12) into equation (11) gives that

dV (t)
≤ − µz(t)T z(t) + z(t)T (DT D ⊗ P BK)(z(t) − ẑ(t))
dt
= − µẑ(t)T ẑ(t)
+ ẑ(t)T ((DT D ⊗ P BK) − 2µInN )(z(t) − ẑ(t))
+ (z(t)− ẑ(t))T(DTD⊗P BK −µInN )(z(t)− ẑ(t)).

Thus, for any γ > 0,

dV (t) σ
≤ − (µ − )ẑ(t)T ẑ(t)
dt 2γ (13)
γσ
+( −µ+λnλP BKs )(z(t)− ẑ(t))T (z(t)− ẑ(t)).
2
For any γ ∈ S, by Theorem 1 and inequality (9), limt→∞ z(t) = 0, and thus limt→∞ V (t) = 0. Let
n
1 At X
δ(t) = x(t) − 1n ⊗ e xi (0). (14)
n i=1

Noticing that z(t) = (DT ⊗ IN )δ(t) and δ(t)⊥1nN , we obtain that

1 λ2
V (t) = δ(t)T DDT δ(t) ≥ δ(t)T δ(t).
2 2
Therefore, limt→∞ δ(t) = 0.

B. Asynchronous broadcasting of self-sampled data

Consider the multi-agent system (5) with an undirected interaction topology G. Different from the sampling of relative states in
the previous subsection, we assume that each agent samples its own state with measurement errors and then broadcasts them to
its neighbors with time delays. When all neighbors get the information, the agent and the neighbors all update their controllers.
Denote the sampling instants of agent i, the associated measurement errors and time delays by tik , eik , and τki , k = 0, 1, 2, . . . ,
i
respectively. Denote x̂i (t) = xi (tik ) − eik for t ∈ [tik + τki , tik+1 + τk+1 ), k = 0, 1, 2, . . . . Under Assumption 1 (1-3), we give the
following protocol:
X
ui (t) = K (x̂j (t) − x̂i (t)). (15)
j∈Ni

The matrix K will be designed later.

 9

Denote x(t) = [x1 (t)T x2 (t)T . . . xn (t)T ]T and x̂(t) = [x̂1 (t)T x̂2 (t)T . . . x̂n (t)T ]T . Substituting equation (15) into
equation (5) gives that
ẋ(t) = (In ⊗ A)x(t) − (DDT ⊗ BK)x̂(t). (16)
Pn
Let δ(t) = [δ1 (t)T δ2 (t)T . . . δn (t)T ]T be defined by (14) with δi (t) ∈ RN , and let κ(t) = (1/n)e At
i=1 xi (0). For
t ∈ [tik + τki , tik+1 + i
τk+1 ), let δ̂i (t) = δi (tik ) − eik and κ̂i (t) = κ(tik ), k = 0, 1, . . . , i = 1, 2, . . . , n. Denote δ̂(t) =
T T T T
[δ̂1 (t) δ̂2 (t) . . . δ̂n (t) ] and κ̂(t) = [κ̂1 (t)T κ̂2 (t)T T T
. . . κ̂n (t) ] . Then we have

δ̇(t) = (In ⊗ A)δ(t) − (DDT ⊗ BK)(δ̂(t) + κ̂(t)), (17)

and δ(t) ⊥ 1nN .

Remark 5: (1) By the definition of incidence matrix D, if the system solves a consensus problem, then the final consensus
trajectory should be κ(t). However, it does not satisfy the differential equation (16) with asynchronous sampling and
unstable system matrix A; that is, (DDT ⊗ BK)κ̂(t) 6= 0 in most cases. Thus in such cases, the asynchronous consensus
cannot be achieved.
(2) In the case of multiplicative type of measurement errors, even if the consensus is reached, the final consensus state is
usually not zero and thus eik may never converge to zero as k → ∞. Therefore, the multiplicative type of measurement
errors can destroy the consensus of the considered system (16).
1) Asynchronous consensus of single-integrators: In the case of multiple single-integrators with N = 1, A = 0, and B = 1,
κ(t) is time-invariant and so is κ̂(t). Therefore, choose K = 1 and by (17),

δ̇(t) = −DDT δ̂(t).

Consider the following Lyapunov candidate

1
V (t) = δ(t)T δ(t). (18)
2
With the same arguments as in proving (13), we have
dV (t)
≤ − λ2 δ(t)T δ(t) + δ(t)T DDT (δ(t) − δ̂(t))
dt
σ
≤ − (λ2 − )δ̂(t)T δ̂(t)
2γ
γσ
+( + λn − λ2 )(δ̂(t) − δ(t))T (δ̂(t) − δ(t)),
2
where σ = max{2λ2 , λn − 2λ2 } is the largest singular value of matrix DDT − 2λ2 In , and γ > 0 with λ2 − σ/(2γ) > 0.
By Theorem 1, we have the following result:
Theorem 3: In system (5), suppose that each agent takes the single-integrator dynamics without measurement errors, and
suppose that the sampling instants and time delays satisfy Assumption 1 (1-3), i = 1, 2, . . . , n. If the interaction topology G is
connected and
σ
3 λ2 − 2γ
(h + τ )2 < sup γσ , (19)
7λn 2 γ>0 2 + λn − λ2
then protocol (15) with K = 1 solves the average consensus problem.
Remark 6: In [46], it has been shown that the necessary and sufficient condition for the synchronous sampled-data consensus
of single-integrators with sampling period h is 0 < h ≤ 2/λn . Compared with this condition, the condition (19) may be a bit
conservative, which is understandable since this condition is valid in the general asynchronous sampling case and obtained based
on the general model of linear subsystems.
10

2) Asynchronous consensus of marginally stable systems: Suppose that matrix A is marginally stable. Then maxs keAs k2 is
bounded and by Lemma 1, we have2
n
√ X
kκ̂(t) − 1n ⊗ κ(t)k2 <(1/ n)k xi (0)k2 maxs keAs k2
i=1
σA (eλAs h − 1)
× .
λAs
Denote the right side of the above equation by ∆κ (h). Clearly, limh→0 ∆κ (h) = 0.
For any i, denote ei (t) = eik and δ̃i (t) = δi (tik ), t ∈ [tik + τki , tik+1 + τk+1
i
), k = 0, 1, . . . . Denote e(t) = [e1 (t)T e2 (t)T . . .
en (t)T ]T and δ̃(t) = [δ̃1 (t)T δ̃2 (t)T . . . δ̃n (t)T ]T . Then δ̂(t) = δ̃(t) − e(t) and

δ̇(t) =(In ⊗ A)δ(t) − (DDT ⊗ BK)δ̃(t)

(20)
+ (DDT ⊗ BK)(e(t) − κ̂(t)).

If ke(t)k2 ≤ ∆e , then

k(DDT ⊗ BK)(e(t) − κ̂(t))k2

≤ kDDT ⊗ BKk2 (∆κ (h) + ∆e ) , ∆(h).

Choose the feedback matrix K = B T P with a positive definite matrix P and a positive number µ by (10) and consider the
following Lyapunov function candidate
1
V (t) = δ(t)T (In ⊗ P )δ(t).
2
Then we have
dV (t)
≤δ(t)T (In ⊗ P A)δ(t)−δ(t)T (DDT ⊗ P BB T P )δ̃(t)
dt
λP λP η
+ δ(t)T δ(t) + ∆(h)2
2η 2
λP σ
≤ − (µ − − )δ̃(t)T δ̃(t)
2η 2γ
γσ λP
+( −µ+ +λn σP B 2 )(δ(t)− δ̃(t))T(δ(t)− δ̃(t))
2 2η
λP η
+ ∆(h)2 , (21)
2
where λP is the largest eigenvalue of P , σP B is the largest singular value of P B, σ is the largest singular value of (DDT ⊗
P BB T P ) − 2(µ − λP /(2η))InN , and γ and η are any positive numbers with µ − λP /(2η) − σ/(2γ) > 0.
It can be seen that equations (17) and (21) do not exactly match the model (2), (3). However, when the largest sampling
period h and measurement errors are small, the system represented by (20) and (21) can be seen as the model (2), (3) with
uncertainties. Let σBB T P denote the largest singular value of BB T P ,

¯ γσ λP 1
∆(h, τ, α, γ, η) =( −µ+ + λn σP B 2 )(1 + )
2 2η α
2 2
× maxs keAs k∞ n(h + τ )2 ∆(h)
1 2
+ λP η∆(h) ,
2
2 In case of a 0-valued λAs serving as a denominator, the value of the expression is set to the limit when λAs → 0.
 11

and
λP σ
Γ(h, τ, α, β, γ, η) =µ − −
2η 2γ
γσ λP
−( −µ+ + λn σP B 2 )(1 + α)
2 2η
 1 7 
× (1 + )σA 2 + (1 + β) λn 2 σBB T P 2
β 3
× (h + τ )2 e2λAs (h+τ ) ,
where k · k∞ denotes the maximum row sum matrix norm. By the same arguments as in the proof of Theorem 1, we can obtain
the following result:
Theorem 4: In system (5), assume that (A, B) is stabilizable, A is marginally stable, Assumption 1 (1-3) holds, ke(t)k2 ≤ ∆e ,
tik+1 − tik is lower bounded by a positive number independent of k, i = 1, 2, . . . , n, and the interaction topology is connected.
Under the proposed protocol (15) with K given by (10), if S 6= ∅, then for any time t′ and any θ > 1, there exists some t,
t ≥ t′ , such that
¯
θ∆(h, τ, α, γ, η)
δ̃(t)T δ̃(t) ≤ inf , (22)
[α,β,γ,η]∈S Γ(h, τ, α, β, γ, η)
where
n
S = [α, β, γ, η] :α, β, γ, η > 0,

µ − λP /(2η) − σ/(2γ) > 0,

γσ λP
−µ+ + λn σP B 2 ≥ 0,
2 2η
o
Γ(h, τ, α, β, γ, η) > 0 .

Proof: The proof is omitted here.

IV. E XTENSIONS

This section only considers the model (2), (3). So the obtained results are also valid for the consensus models in Section III.

A. Quantized sampling
This subsection gives a simple example to show the effectiveness of Theorem 1 in the scenario of quantization. Further
discussions will continue in the next subsection.
In system (2), if apply the following Logarithmic quantizer to each entry of zi (tik ) [51]:
(
0, if ξ = 0,
Qlog ε (ξ) = (23)
⌊log ǫ |ξ|⌋
sign(ξ)ǫ , otherwise,
then
(zi (tik )−Qlog ǫ (zi (tik )))T (zi (tik ) − Qlog ǫ (zi (tik )))
(24)
≤ (ǫ − 1)2 Qlog ε (zi (tik ))T Qlog ǫ (zi (tik )),
where ǫ > 1 is the quantizing level, ξ is the scalar quantizer input, ⌊·⌋ denotes the largest integer not greater than the considered
variable, and Qlog ǫ (zi (tik ))) is the vector obtained by applying the quantizer (23) to zi (tik ) entrywise. So Qlog ǫ (zi (tik )) can be
viewed as the sampled value zi (tik ) with measurement error eik = zi (tik ) − Qlog ǫ (zi (tik )).
12

Redefine ẑi (t) = Qlog ǫ (zi (tik ))), t ∈ [tik + τki , tik+1 + τk+1
i
), k = 0, 1, . . . , i = 1, 2, . . . , m, and define the same ẑ(t) as in
Section II. By Theorem 1, we have the following corollary:
Corollary 1: In system (2) with the utilization of Logarithmic quantizer (23) in sampling, if Assumption 1 (1-3), i =
1, 2, . . . , m, and inequality (3) hold, and there exist some positive numbers α and β, such that
1 1
µ − ε(ǫ − 1)2 (1 + )(1 + )e2λAs (h+τ )
α β
 1
− ε (1 + α)(1 + )σA 2
β
7 
+ (1 + β) σG 2 σK 2 (h + τ )2 e2λAs (h+τ ) > 0,
3
then
lim z(t) = 0.
t→∞

Note that the same property as (24) of the above Logarithmic quantizer can be preserved by the output of an event-based
sampler with event-triggering condition that ei (t)T ei (t) > (ǫ − 1)2 ẑi (t)T ẑi (t), where ẑi (t) is the latest measurement of zi (·) up
to t and ei (t) is the measurement error. So similar discussions in the following subsection can be given to quantized systems.

B. Event-triggered sampling
This subsection shows an interesting application of asynchronous aperiodic sampling in removing the Zeno behavior in event-
based control.
Consider the multi-agent system with m subsystem in the absence of time delays and introduce the same notations zi (t) and
z(t) as in Section II to represent the state of agents. Let ẑi (t) be the measurement of zi (t) and it is updated in an event-triggered
way. In detail, for agent i, let ẑi (t) first updated at ti0 and set ẑi (ti0 ) = zi (ti0 ). For k = 0, 1, . . . , as time increases from tik + h,
agent i examine the condition that

(zi (t) − zi (tik ))T (zi (t) − zi (tik )) ≥ ωzi (tik )T zi (tik ), (25)

where h is a predetermined positive dwell time and called rest time in [24]. For the first time t when the above inequality become
correct, set tik+1 = t and ẑi (tik+1 ) = zi (tik+1 ). Note that the sequence ti0 , ti1 , . . . may terminate at some finite k. Obviously,
ẑi (t) is piece-wise constant; that is, in the case with tik given through the above procedure, ẑi (t) = zi (tik ) for t ∈ [tik , tik+1 ) if
tik+1 exists; and ẑi (t) = zi (tik ) for t ∈ [tik , ∞) if tik+1 doesn’t exist.
Denote ẑ(t) = [ẑi (t)T ẑ2 (t)T . . . ẑm (t)T ]T . The overall system, represented by (2), has the following stability result:
Corollary 2: In system (2), assume that ẑ(t) is updated according to the event-triggering condition (25) with dwell time h
and without time delays, and there exists a lower bounded function V (t) with property described by (3). If there exist some
positive numbers α and β, such that
1 1
µ − εω(1 + )(1 + )e2λAs h
α β
 1
− ε (1 + α)(1 + )σA 2
β
7 
+ (1 + β) σG 2 σK 2 h2 e2λAs h > 0,
3
then
lim z(t) = 0.
t→∞
 13

Proof: Note that the definition of tik is different from that in Section II and Assumption 1 (1) doesn’t hold. However, by
event-triggering condition (25), for any t, t 6∈ ∪k (tik , tik + h),

(zi (t) − ẑi (t))T (zi (t) − ẑi (t)) < ω ẑi (t)T ẑi (t). (26)

So we can view ẑi (t) as the sampled state of zi (t) at time t with measurement error zi (t) − ẑi (t), which is exactly Assumption
1 (4).
(Case 1) If time sequence ti0 , ti1 , . . . , is of finite length, let k̄ = arg maxk tik . For any k, k < k̄, there exists a positive integer
nik such that tik + nik h < tik+1 ≤ tik + (nik + 1)h. Let sik , k = 0, 1, 2, . . . , denote all the time instants tik + ph, k = 0, 2, . . . ,
k̄ − 1, p = 0, 1, . . . , nik , and tik̄ + ph, p = 0, 1, . . . , in the increasing order. (Case 2) If time sequence ti0 , ti1 , . . . , is of infinite
length, for any k, there exists a positive integer nik such that tik + nik h < tik+1 ≤ tik + (nik + 1)h. Let sik , k = 0, 1, 2, . . . , denote
all the time instants tik + ph, k = 0, 2, . . . , p = 0, 1, . . . , nik , in the increasing order.
Then we have that sik+1 − sik ≤ h and inequality (26) holds at any t = sik . By Theorem 1, we have the stability of the system.

Remark 7: (1) The idea of removing the Zeno behavior in event-based control by adding a dwell time was previously
presented for networks of single-integrators in [24].
(2) Event-triggering condition (25) can be replaced by
√
ω
kzi (t) − zi (tik )k2 ≥ √ kzi (t)k2 ,
1+ ω
which is classified as a quadratic event-triggering condition in [23].

C. Saturated control with input delay

This subsection discusses the extension of Theorem 1 in the case with input saturations and input delays [56]–[61].
Pm
In system (1), j=1 gij K ẑj (t) can be viewed as the input. Instead of applying a saturation function to the input or to each
sampled state ẑj (t), we use a scaler function ρ(·), defined below, to scale down the sampled states to meet the restriction of
input saturation:

 1, if ξ = 0,
ρ(ξ) = 1
, otherwise,
⌈ kξk
 ∞
ρs ⌉

where ξ ∈ RN , ⌈·⌉ denotes the ceiling function (which returns the smallest integer greater than or equal the considered variable)
and ρs is some positive number decided by the maximum allowable magnitude of input. Clearly, for any ξ, 0 < ρ(ξ) ≤ 1 and
kρ(ξ)ξk∞ ≤ ρs . Function ρ(ξ) can be also defined in terms of other norms. Then the subsystem i takes the following dynamics:
m
X
żi (t) = Azi (t) − gij ρ(ẑj (t − τin ))K ẑj (t − τin ), (27)
j=1

where τin is the input delay.

Theorem 5: Assume that Assumption 1 holds for all i and there exists a lower bounded function V (t) such that
dV (t)
≤ − µẑ(t − τin )T ẑ(t − τin )
dt (28)
+ ε(z(t) − ẑ(t − τin ))T(z(t) − ẑ(t − τin )),
14

where µ > 0 and ε > 0. If there exist some positive numbers α and β, such that
1 1
µ − εω(1 + )(1 + )e2λAs (h+τ +τin)
α β
 1
− ε (1 + α)(1 + )σA 2 (29)
β
7 
+ (1 + β) σG 2 σK 2 (h + τ + τin )2 e2λAs (h+τ +τin) > 0,
3
then
lim z(t) = 0.
t→∞

Proof: The proof follows the same procedure as in the proof of Theorem 1. Here, τki + τin is viewed as the time delay in
sampling zi (t). We also use the property that the largest singular value of G is not less than that of matrix [gij ρ(ẑj (t − τin ))].
Note that neither ρ(·) nor ρs appears in Theorem 5. However, the saturated input with function ρ(·) as well as initial states
has a significant impact on the existence of a valid feedback gain K and a function V (t) satisfying inequality (28). For example,
when we know that the states of all subsystem (27) are bounded by Lyapunov methods, then ρ(ẑj (t − τin )) has a finite set
of all possible values. The considered system is a switched system with state-dependent switching between a finite number of
subsystems. V (t) can be chosen to be a common Lyapunov function [62]. Initial states are closely related to the upper bound
of system states and thus affect the number of subsystems in the resulting switched system.
Example 1: Consider the network of single-integrators described by equation (5), where N = 1, A = 0, and B = K = 1.
Revise protocol (6) by multiplying each sampled state by the function ρ(·). Denote Λ(t) = diag([ρ(ẑ1 (t − τin )) ρ(ẑ2 (t −
τin )) . . . ρ(ẑm (t − τin ))]), and suppose that ρm = min{ρ(ẑi (t − τin )) : i = 1, 2, . . . , m} is larger than 0. We obtain that

ẋ(t) = −DΛ(t)ẑ(t − τin ).

Consider the Lyapunov candidate given by (18). We have

dV (t)
= − z(t)Λ(t)ẑ(t − τin )
dt
= − ẑ(t − τin )T Λ(t)ẑ(t − τin )
− ẑ(t − τin )T Λ(t)(z(t) − ẑ(t − τin ))
1
≤ − (ρm − )ẑ(t − τin )T Λ(t)ẑ(t − τin )
2γ
γ
+ (z(t) − ẑ(t − τin ))T (z(t) − ẑ(t − τin )),
2
where parameter γ > 1/(2ρm ).

V. N UMERICAL E XAMPLES

Fig. 1. Interaction Topology

 15

This section presents three numerical examples to demonstrate the effectiveness of the theoretical results.
Example 2 (Quantized sampling): We first consider the network of harmonic oscillators under the topology depicted in Fig. 1.
These oscillators take the dynamics given by (5). The system matrix A and input B are given by
" # " #
0 1 0
A= , B= .
−1 0 1

Let λ = λ2 and µ = 1. Solving equation (10), we obtain K = [0.5626 1.0633]. Apply the Logorithmic quantizer (23) with
ǫ = 1.1 to the relative states in protocol (6). By Theorem 2, we obtain that if h + τ < 0.017, then the system solves the average
consensus problem. Let the sampling periods randomly generated in [0.005, 0.012] and the delays are random and not larger
than 0.005. The states of oscillators and quantized relative states with time delays are given in Fig. 2 and Fig. 3
10

States 5

−5

−10
0 5 10 15
Time
10

5
States

−5

−10
0 5 10 15
Time

Fig. 2. State trajectories of oscillators. The two figures show the first and second components of states, respectively.
Quantized relative states

−5

−10
0 5 10 15
time
Quantized relative states

10

−5
0 5 10 15
time

Fig. 3. Quantized relative states with time delays. The two figures show the first and second components of quantized relative states, respectively.

Example 3: Consider the network of single integrators under the topology depicted in Fig. 1. If protocol (15) is used, then
by Theorem 3, h + τ < 0.0691 is sufficient for state consensus in the absence of measurement errors. By the result in [46] (see
Remark 6), h ≤ 0.5528 is a necessary and sufficient condition for state consensus in the synchronous periodic sampling case
without time delays and measurement errors. This result is not applicable to our system because of different model setups and
assumptions. Moreover, no necessary and sufficient results have been derived for the asynchronous time-delayed system studied
in this paper in the literature.
Example 4 (Event-triggered broadcasting): Consider the system in Example 3 with h = 0.025 and τ = 0.02. Denote x̂i ((tik +
τki )− ) = limt↑tik +τki x̂i (t), which denotes the most recent broadcasted data of xi (t) before time tik . If the following event-triggering
condition holds:
|xi (tik ) − x̂i ((tik + τki )− )| > min{0.3|x̂i ((tik + τki )− )|, 0.08},
16

then xi (tik ) is broadcasted to the neighbors of agent i. Then we have that ∆(h) = 0.2894. By Theorem 4, the system solves
the average consensus problem with the error 0.4535 given by (22), where α = 0.5, γ = 3.188, and η = 1.6. Let the sampling
periods randomly generated in [0.02, 0.025] and the delays are random and not larger than 0.02. The state trajectories and event
numbers are given in Fig. 4 and Fig. 5, respectively.

18

17

states
16

15

14

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

time
14
12
10
8
6
4
0.4535
2
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time

Fig. 4. The two figures show the state trajectories and the trajectory of δ(t)T δ(t), respectively.

14

12

10
event number

0
0 1 2 3 4 5
time

Fig. 5. Event number every 1/10 unit of time.

VI. C ONCLUSIONS

This paper studied the asynchronous stability of a sampled-data multi-agent system and showed its application in solving
asynchronous consensus problems and its potential extensions to quantized and event-triggered systems. Although the presented
results have addressed several control problems in one framework, the studied system indeed has a quite special structure. So
we are also expecting further detailed work on more general asynchronous systems, such as with directed information links and
heterogeneous individual dynamics.

A PPENDIX

Preliminary lemmas:

Lemma 1: For positive number t and any real square matrix A with AT + A 6= 0, the largest singular values of eAt , eAt − I,
Rt
and 0 eA(t−s) ds are bounded above by eλAs t , σA (eλAs t − 1)/λAs , (eλAs t − 1)/λAs , respectively, where I is the identity matrix
 17

with compatible dimensions, and λAs and σA are the largest eigenvalue of (1/2)(A + AT ) and the largest singular value of A,
respectively.
Proof: (1) To show that eλAs t is an upper bound of the singular values of eAt , it is sufficient to show that xT (eAt )T eAt x ≤
e2λAs t xT x holds for any vector x. Let y(t) = eAt x and let V (t) = y(t)T y(t). Then dV (t)/dt = 2y(t)T AeAt x = y(t)T (A +
AT )y(t) ≤ 2λAs y(t)T y(t) = 2λAs V (t). By the Comparison Principle of differential equations, V (t) ≤ e2λAs t V (0); that is,
T
xT (eAt ) eAt x ≤ e2λAs t xT x.
(2) Let y(t) = (eAt − I)x and let V (t) = y(t)T y(t). We have that dy(t)/dt = AeAt x = Ay(t) + Ax and
dV (t)
= 2y(t)T (Ay(t) + Ax)
dt
≤ 2λAs y(t)T y(t) + 2xT (eAt − I)T Ax,

where, by the Comparison Principle of differential equations, xT (eAt − I)T Ax ≤ (σA 2 /λAs )(eλAs t − 1)xT x (In detail,
d/dt(xT (eAt − I)T Ax) = xT AT (eAt )T Ax ≤ eλAs t xT AT Ax ≤ σA 2 eλAs t xT x = d/dt((σA 2 /λAs )(eλAs t − 1)xT x).) Employing
the Comparison Principle of differential equations again,
2σA 2 t 2λAs (t−s) λAs s
Z
V (t) ≤ e (e − 1)dsxT x
λAs 0
σA 2 λAs t
= (e − 1)2 xT x.
λAs 2
Thus the singular values of eAt − I are bounded above by σA (eλAs t − 1)/λAs .
Rt
(3) Let y(t) = 0 eA(t−s) dsx and let V (t) = y(t)T y(t). Then dy(t)/dt = Ay(t) + x and
Z t
dV (t)
= 2y(t)T Ay(t) + 2xT eA(t−s) dsx.
dt 0
Rt
With the same arguments as before, we have that xT 0 eA(t−s) dsx ≤ ((eλAs t − 1)/λAs )xT x, and thus the singular values of
R t A(t−s)
0 e ds are bounded above by (eλAs t − 1)/λAs .
Lemma 2: The following inequalities hold for any nonnegative number t:
(1) e2t − 4et + 3 + 2t ≤ (2t3 /3)e2t ;
(2) t ≤ et − 1 ≤ tet .

Proof of Theorem 1:

Only the asymptotical stability of system (2) is studied, so the system dynamics in the first few seconds is ignored and assume
that t ≥ maxi (ti0 + τ0i ). Denote all the time instants tik , tik + τki , i = 1, 2, . . . , m, k = 0, 1, . . . , at or after time maxi (ti0 + τ0i ),
by a single sequence t0 , t1 , t2 , . . . , in the increasing order tk < tk+1 , k = 0, 1, . . . . Clearly, ẑ(t) is a constant in each time
interval [tk , tk+1 ), k = 0, 1, . . . .

Rt
(1) Evaluation of s=t0 (z(s) − ẑ(s))T (z(s) − ẑ(s))ds.
Consider the time interval t ∈ [tik + τki , tik+1 + τk+1
i
]. By (2),
i i
zi (t) − ẑi (t) =(eA(t−tk ) − IN )ẑi (tik + τki ) + eA(t−tk ) eik
Z t
− eA(t−s) (Gi ⊗ K)ẑ(s)ds
tik
18

and thus

(zi (t)−ẑi (t))T (zi (t) − ẑi (t))

1 i
≤(1 + α)(1 + )ẑi (tik + τki )T (eA(t−tk ) − IN )T
β
i
× (eA(t−tk ) − IN )ẑi (tik + τki )
1 1 T i i
+ (1 + )(1 + )eik (eA(t−tk ) )T eA(t−tk ) eik
α β
Z t T
+ (1 + β) eA(t−s) (Gi ⊗ K)ẑ(s)ds
tik
Z t
× eA(t−s) (Gi ⊗ K)ẑ(s)ds, (30)
tik

where Gi is the i-th row of matrix G.

Consider the first term on the right side of inequality (30). By Lemma 1 and Lemma 2 in the appendix,
i i
ẑi (tik +τki )T (eA(t−tk ) − IN )T (eA(t−tk ) − IN )ẑi (tik + τki )
σA 2 λAs (t−tik )
≤ (e − 1)2 ẑi (tik + τki )T ẑi (tik + τki )
λAs 2
≤σA 2 (h+τ )2 e2λAs (h+τ ) ẑi (tik +τki )T ẑi (tik +τki ). (31)

Consider the third term on the right side of inequality (30). Suppose that tp1 = tik , tp2 = tik + τki , and tp3 < t ≤ tp3 +1 ≤
i
tik+1 + τk+1 , and denote hik = tik+1 − tik .
Z t
eA(t−s) (Gi ⊗ K)ẑ(s)ds
tik
3 −1
pX
tl+1 − tl hik + τk+1
i
=
hik + τk+1
i tl+1 − tl
l=p1
Z tl+1
× eA(tl+1 −s) ds eA(t−tl+1 ) (Gi ⊗ K)ẑ(tl )
tl
tp3 +1 − tp3 hik + τk+1
i
+
hik + τk+1
i tp3 +1 − tp3
Z t
× eA(t−s) ds(Gi ⊗ K)ẑ(tp3 ).
tp3

By the convexity of the square function and Lemma 1,

Z t TZ t
A(t−s)
e (Gi ⊗ K)ẑ(s)ds eA(t−s) (Gi ⊗ K)ẑ(s)ds
tik tik
3 −1
pX
hik i  Z tl+1
+ τk+1
≤ eA(tl+1 −s) ds
tl+1 −tl tl
l=p1
T
× eA(t−tl+1 ) (Gi ⊗ K)ẑ(tl )
Z tl+1
× eA(tl+1 −s) ds eA(t−tl+1 ) (Gi ⊗ K)ẑ(tl )
tl
T
hi + τk+1
i t
Z
+ k e A(t−s)
ds(Gi ⊗ K)ẑ(tp3 )
tp3 +1 − tp3 tp3
 19

Z t
× eA(t−s) ds(Gi ⊗ K)ẑ(tp3 )
tp3
3 −1
pX
hik + τk+1
i
≤ (eλAs (tl+1 −tl ) − 1)2 e2λAs (t−tl+1 )
l=p1
λAs 2 (tl+1 − tl )

× ẑ(tl )T (Gi T Gi ⊗ K T K)ẑ(tl )

hik + τk+1
i
+ (eλAs (t−tp3 ) − 1)2
λAs 2 (tp3 +1 − tp3 )
× ẑ(tp3 )T (Gi T Gi ⊗ K T K)ẑ(tp3 ). (32)

Notice that inequality (32) holds for any t and the corresponding properly defined indexes tp1 , tp2 , and tp3 . Combining
inequalities (30), (31) and (32) gives that
Z t
(zi (s) − ẑi (s))T (zi (s) − ẑi (s))ds
tik +τki
 1  1
≤ (1 + α)σA 2 (h + τ )2 + (1 + )ω (1 + )
α β
× e2λAs (h+τ ) (t − tp2 )ẑi (tik + τki )T ẑi (tik + τki )
2 −1
pX
hik + τk+1
i
+ (1 + β) (eλAs (tl+1 −tl ) − 1)2
l=p1
λAs 2 (tl+1 − tl )
e 2λAs (t−tl+1 )
− e2λAs (tp2 −tl+1 )
×
2λAs
× ẑ(tl )T (Gi T Gi ⊗ K T K)ẑ(tl )
pX3 −1 
hik + τk+1 i 
+ (1 + β) 3 e2λAs (tl+1 −tl )
l=p2
2λ A s (t l+1 − t l )

λAs (tl+1 −tl )
− 4e + 3 + 2λAs (tl+1 − tl )
hik + τk+1
i
+ (eλAs (tl+1 −tl ) − 1)2
λAs 2 (tl+1 − tl )
e2λAs (t−tl+1 ) − 1

× ẑ(tl )T (Gi T Gi ⊗ K T K)ẑ(tl )
2λAs
hik + τk+1
i 
+ (1 + β) 3 e2λAs (t−tp3 )
(t
2λAs p3 +1 − t p3 )

λAs (t−tp3 )
− 4e + 3 + 2λAs (t − tp3 )

× ẑ(tp3 )T (Gi T Gi ⊗ K T K)ẑ(tp3 )

 1  1
≤ (1 + α)σA 2 (h + τ )2 + (1 + )ω (1 + )
α β
× e2λAs (h+τ ) (t − tp2 )ẑi (tik + τki )T ẑi (tik + τki )
2 −1
pX
+ (1 + β)(hik + τk+1
i
) (tl+1 − tl )(t − tp2 )
l=p1

× e2λAs (t−tl ) ẑ(tl )T (Gi T Gi ⊗ K T K)ẑ(tl )

3 −1 
pX
1
+ (1 + β)(hik + τk+1 i
) (tl+1 − tl )2
3
l=p2
20


× e2λAs (tl+1 −tl ) + (tl+1 − tl )(t − tl+1 )e2λAs (t−tl )

× ẑ(tl )T (Gi T Gi ⊗ K T K)ẑ(tl )

1 (t − tp3 )3
+ (1 + β)(hik + τk+1
i
)
3 tp3 +1 − tp3
× e2λAs (t−tp3 ) ẑ(tp3 )T (Gi T Gi ⊗ K T K)ẑ(tp3 )
 1  1
≤ (1+α)σA 2 (h+τ )2 +(1+ )ω (1 + )e2λAs (h+τ )
α β
X p3 
× (tl+1 −tl )+(t−tp3 ) ẑi (tik +τki )T ẑi (tik +τki )
l=p2

+ (1 + β)(h + τ )2 e2λAs (h+τ )

2 −1
pX
× (tl+1 − tl )ẑ(tl )T (Gi T Gi ⊗ K T K)ẑ(tl )
l=p1
4
+ (1 + β)(h + τ )2 e2λAs (h+τ )
3
pX3 −1

× (tl+1 − tl )ẑ(tl )T (Gi T Gi ⊗ K T K)ẑ(tl )

l=p2
1
+ (1 + β)(h + τ )2 e2λAs (h+τ )
3
× (t − tp3 )ẑ(tp3 )T (Gi T Gi ⊗ K T K)ẑ(tp3 ). (33)

(2) Evaluation of V (t).

Let p− be any index such that tip− ≥ maxj (tj0 + τ0j ) for any i. Let tp0 = maxi (tip− + τpi− ), tpi1 = tip− , and tpi2 = tip− + τpi− .
Note that inequality (33) is correct for any t with the associated tp1 , tp2 , and tp3 . By inequalities (3) and (33), we have the
estimation of V (t) in the following form:
p−1 
X  1
V (t) ≤V0 + (tl+1 − tl ) − µ + ε (1 + )ω
α
l=p0

 1
+(1+α)σA 2 (h+τ )2 (1+ )e2λAs(h+τ ) ẑ(tl )Tẑ(tl )
β
7 2 2λAs (h+τ )
+ (1 + β)ε (h + τ ) e
3
p−1
X
× (tl+1 − tl )ẑ(tl )(GT G ⊗ K T K)ẑ(tl )T ,
l=p0

where
Z tp0
V0 =V (t0 ) − µ ẑ(t)T ẑ(t)
t=t0
m Z
X tpi
2
+ε (z(t) − ẑ(t))T (z(t) − ẑ(t))dt
i=1 t0
m
X  1 
+ε (1 + α)σA 2 (h + τ )2 + (1 + )ω
i=1
α
 21

p0 −1
1 2λAs (h+τ ) X
× (1 + )e (tl+1 − tl )zi (tik )T zi (tik )
β i l=p2
2 2λAs (h+τ )
+ (1 + β)(h + τ ) e
0 −1
pX
× (tl+1 − tl )ẑ(tl )T (Gi T Gi ⊗ K T K)ẑ(tl )
l=pi1
4
+ (1 + β) (h + τ )2 e2λAs (h+τ )
3
0 −1
pX
!
T T T
× (tl+1 − tl )ẑ(tl ) (Gi Gi ⊗ K K)ẑ(tl ) .
l=pi2

Let
1 1
Γ =µ − ε(1 + )((1 + )ω
β α
+ (1 + α)σA 2 (h + τ )2 )e2λAs (h+τ )
7
− ε(1 + β) σG 2 σK 2 (h + τ )2 e2λAs (h+τ ) .
3
Then by (4), Γ > 0 and

p−1
X
V (tp ) ≤V0 − Γ (tl+1 − tl )ẑ(tl )T ẑ(tl ).
l=p0

Pp−1
By the lower boundedness of V (t), limp→∞ l=p0 (tl+1 − tl )ẑ(tl )T ẑ(tl ) exists and

∞
X
lim (tl+1 − tl )ẑ(tl )T ẑ(tl ) = 0. (34)
p→∞
l=p

Denote ei (t) = eik and z̃i (t) = zi (tik ), t ∈ [tik +τki , tik+1 +τk+1
i
), i = 1, 2, . . . , m, k = 0, 1, . . . . Denote e(t) = [e1 (t)T e2 (t)T . . . em (t)T ]T
and z̃(t) = [z̃1 (t)T z̃2 (t)T . . . z̃m (t)T ]T . By (34) and Assumption 1 (4),
∞
X
lim (tl+1 − tl )e(tl )T e(tl ) = 0,
p→∞
l=p

and
∞
X
lim (tl+1 − tl )z̃(tl )T z̃(tl ) = 0. (35)
p→∞
l=p

(3) Asymptotically Stability.

For any given k, the same indexes p1 , p2 , and p3 as in the previous part are defined. With the same arguments as in proving
inequalities (31) and (32) and by Lemma 1, for t ∈ [tik + τki , tik+1 + τk+1
i
],

i i
zi (t) − zi (tik + τki ) =(eA(t−tk −τk ) − I)zi (tik + τki )
Z t
− eA(t−s) (Gi ⊗ K)ẑ(s)ds,
tik +τki
22

and

(zi (t)−zi (tik + τki ))T (zi (t) − zi (tik + τki ))

≤(tik+1 + τk+1
i
− tik − τki )2σA 2 e2λAs (h+τ )
T
× (t − tik − τki )zi (tik + τki ) zi (tik + τki )
+ (tik+1 + τk+1
i
− tik − τki )2λGi K e2λAs (h+τ )
3 −1
pX
× (tl+1 − tl )ẑ(tl )T ẑ(tl )
l=p2

≤(tik+1 i
+ τk+1 − tik − τki )4σA 2 e2λAs (h+τ )
T
× (t − tik − τki )z̃i (tik + τki ) z̃i (tik + τki )
+ (tik+1 + τk+1
i
− tik − τki )
× 4σA 2 e2λAs (h+τ ) (t − tik − τki )
× (zi (tik +τki )−zi (tik ))T (zi (tik +τki )−zi (tik ))
+ (tik+1 + τk+1
i
− tik − τki )2λGi K e2λAs (h+τ )
3 −1
pX
× (tl+1 − tl )ẑ(tl )T ẑ(tl ), (36)
l=p2

where λGi K is the largest eigenvalue of Gi T Gi ⊗ K T K. Furthermore, since

i
zi (tik + τki ) − zi (tik ) =(eAτk − I)zi (tik )
Z tik +τki
− eA(t−s) (Gi ⊗ K)ẑ(s)ds,
tik

we have that

(t−tik − τki )(zi (tik + τki ) − zi (tik ))T (zi (tik + τki ) − zi (tik ))
T
≤2σAs 2 τ 2 e2λAs τ (t − tik − τki )z̃i (tik + τki ) z̃i (tik + τki )
(37)
2 −1
pX
+ 2λGi K τ e2λAs τ (t−tik −τki ) (tl+1 −tl )ẑ(tl )T ẑ(tl ).
l=p1

T Pp3 −1
By (34) and (35), in (36) and (37), all the quantities (t − tik − τki )z̃i (tik + τki ) z̃i (tik + τki ), T
l=p2 (tl+1 − tl )ẑ(tl ) ẑ(tl ),
Pp2 −1
and l=p1 (tl+1 − tl )ẑ(tl )T ẑ(tl ) converge to 0 as k goes to infinity. Therefore, we have the following inequality with positive
f (i, t, k) converging to 0 as k goes to infinity:

(zi (t)−zi (tik + τki ))T (zi (t) − zi (tik + τki ))

≤ (tik+1 + τk+1
i
− tik − τki )f (i, t, k),

which implies that the change of zi (t) over any time intervals with a fixed length will be infinitesimally small when t is sufficiently
large. By (35),
lim zi (t) = 0.
t→∞
 23

R EFERENCES

[1] R. C. Dorf and R. H. Bishop, Modern Control Systems, Pearson Education, 2016.

[2] T. Chen and B. Francis, Optimal Sampled-Data Control Systems, Springer, 1995.

[3] B. A. Francis and T. T. Georgiou, Stability theory for linear time-invariant plants with periodic digital controllers, IEEE Transactions on Automatic Control,
vol. AC-33, pp. 820-832, 1988.

[4] B. A. Bamieh and J. B. Pearson, A general framework for linear periodic systems with applications to H ∞ sampled-data control, IEEE Transactions on
Automatic Control, vol. 37, no. 2, pp. 418-435, 1992.

[5] P. Voulgaris and B. Bamieh, Control of asynchronous sampled-data systems, Proceedings of the 32nd Conference on Decision and Control, pp. 785-786,
1993.

[6] V. S. Ritchey and G. F. Franklin, A stability criterion for asynchronous multirate linear systems, IEEE Transactions on Automatic Control, vol. 34, no. 5,
pp. 529-535, 1989.

[7] G. Sridharan, M. C. Srisailam, and V. S. Rao, A note on the effect of asynchronous sampling on estimation accuracy, Automatica, vol. 21, no. 4,
pp. 491-493, 1985.

[8] M. F. Sågfors and H. T. Toivonen, H∞ and LQG control of asynchronous sampled-data systems, Automatica, vol. 33, no. 3. pp. 1663-1668, 1997.

[9] Y. Fang and X. Chu, Stability analysis of multirate sampled-datla systems, Proceedings of the American Control Conference, pp. 3241-3242, 1994.

[10] M. Moarref and L. Rodrigues, Stability and stabilization of linear sampled-data systems with multi-rate samplers and time driven zero order holds,
Automatica, vol. 50, pp. 2685-2691, 2014.

[11] L.-S. Hu, T. Bai, P. Shi, and Z. Wu, Sampled-data control of networked linear control systems, Automatica, vol. 43, pp. 903-911, 2007.

[12] X. Ge and Q.-L. Han, Distributed sampled-data asynchronous H∞ filtering of Markovian jump linear systems over sensor networks, Signal Processing,
vol. 127, pp. 86-99, 2016.

[13] B. Tavassoli, Stability of nonlinear networked control systems over multiple communication links with asynchronous sampling, IEEE Transactions on
Automatic Control, vol. 59, no. 2, pp. 511-515, 2014.

[14] X. Wang, Y. Sun, and N. Hovakimyan, Asynchronous task execution in networked control systems using decentralized event-triggering, Systems & Control
Letters, vol. 61, 936-944, 2012.

[15] M. Mazo Jr. and M. Cao, Asynchronous decentralized event-triggered control, Automatica, vol. 50, pp. 3197-3203, 2014.

[16] D. Antunes, J. Hespanha, and C. Silvestre, Stability of networked control systems with asynchronous renewal links: An impulsive systems approach,
Automatica, vol. 49, pp. 402-413, 2013.

[17] J. N. Tsitsiklis, D. P. Bertsekas, and M. Athans, Distributed asynchronous deterministic and stochastic gradient optimization algorithms, IEEE Transactions
on Automatic Control, vol. AC-31, no. 9, pp. 803-812, 1986.

[18] J. Lin, A. S. Morse, and B. D. O. Anderson, The multiagent rendezvous problem–the asynchronous case, Proceedings of the 43rd IEEE Conference on
Decision and Control, pp. 1926-1931, 2004.

[19] M. Cao, A. S. Morse, and B. D. O. Anderson, Agreeing asynchronously, IEEE Transactions on Automatic Control, vol. 53, no. 8, pp. 1826-1838, 2008.

[20] Y. Gao and L. Wang, Asynchronous consensus of continuous-time multi-agent systems with intermittent measurements, International Journal of Control,
vol. 83, no. 3, pp. 552-562, 2010.

[21] Y. Gao and L. Wang, Sampled-data based consensus of continuous-time multiagent systems with time-varying topology, IEEE Transactions on Automatic
Control, vol. 56, no. 5, pp. 1226-1231, 2011.

[22] D. V. Dimarogonas, E. Frazzoli, and K. H. Johansson, Distributed event-triggered control for multi-agent systems, IEEE Transactions on Automatic
Control, vol. 57, no. 5, pp. 1291-1297, 2012.

[23] W. P. M. H. (Maurice) Heemels, M. C. F. (Tijs) Donkers, and A. R. Teel, Periodic event-triggered control for linear systems, IEEE Transactions on
Automatic Control, vol. 58, no. 4, pp. 847-861, 2013.

[24] F. Xiao and T. Chen, Sampled-data consensus in multi-agent systems with asynchronous hybrid event-time driven interactions, Systems & Control Letters,
vol. 89, pp. 24-34, 2016.
24

[25] L. Fang and P. J. Antsaklis, Information consensus of asynchronous discrete-time multi-agent systems. Proceedings of the American control conference,
pp. 1883-1888, 2005.

[26] D. Freirich and E. Fridman, Decentralized networked control of systems with local networks: A time-delay approach, Automatica, vol. 69, pp. 201-209,
2016.

[27] J. Zhan and X. Li, Asynchronous consensus of multiple double-integrator agents with arbitrary sampling intervals and communication delays, IEEE
Transactions on Circuits and Systems-I: Regular Papers, vol. 62, no. 9, pp. 2301-2311, 2015.

[28] X. Meng and T. Chen, Event based agreement protocols for multi-agent networks, Automatica, vol. 49, no. 7, pp. 2123-2132, 2013.

[29] W. Hu, L. Liu, and G. Feng, Consensus of linear multi-agent systems by distributed event-triggered strategy, IEEE Transactions on Cybernetics, vol. 46,
no. 1, pp. 148-157, 2016.

[30] D. Yang, W. Ren, X. Liu, and W. Chen, Decentralized event-triggered consensus for linear multi-agent systems under general directed graphs, Automatica,
vol. 69, pp. 242-249, 2016.

[31] W. Zhu, Z.-P. Jiang, and G. Feng, Event-based consensus of multi-agent systems with general linear models, Automatica, vol. 50, pp. 552-558, 2014.

[32] F. Xiao and L. Wang, Consensus problems for high-dimensional multi-agent systems. IET Control Theory and Applications, vol. 1, no. 3, pp. 830-837,
2007.

[33] J. Qin, C. Yu, and H. Gao, Collective behavior for group of generic linear agents interacting under arbitrary network topology, IEEE Transactions on
Control of Network Systems, vol. 2, no. 3, pp. 288-297, 2015.

[34] A. Seuret, A novel stability analysis of linear systems under asynchronous samplings, Automatica, vol. 48, pp. 177-182, 2012.

[35] W. Jiang and A. Seuret, Improved stability analysis of networked control systems under asynchronous sampling and input delay, 2nd IFAC Workshop on
Distributed Estimation and Control in Networked Systems, pp. 79-84, 2010.

[36] H. Omran, L. Hetel, and J.-P. Richard, Local stability of bilinear systems with asynchronous sampling, Proceedings of the 4th IFAC Conference on Analysis
and Design of Hybrid Systems (ADHS 12), pp. 19-24, 2012.

[37] R. Olfati-Saber and R. M. Murray, Consensus problems in networks of agents with switching topology and time-delays, IEEE Transactions on Automatic
Control, vol. 49, no. 9, pp. 1520-1533, 2004.

[38] L. Tassiulas and C. J. Su, Information delivery through broadcasting in satellite communication networks, Automatica, vol. 35, no. 12, pp. 2013-2030,
1999.

[39] X. Wang and M. D. Lemmon, Event-triggered broadcasting across distributed networked control systems. Proceedings of American control conference,
pp. 3139-3144, 2008.

[40] W. Zhang, M. S. Branicky, and S. M. Phillips, Stability of networked control systems. IEEE Control Systems, vol. 21, no. 1, pp. 84-99, 2001.

[41] J. Cortés, Global and robust formation-shape stabilization of relative sensing networks, Automatica, vol. 45, pp. 2754-2762, 2009.

[42] M. Guo and D. V. Dimarogonas, Consensus with quantized relative state measurements, Automatica, vol. 49, pp. 2531-2537, 2013.

[43] G. S. Seyboth, D. V. Dimarogonas, and K. H. Johansson, Event-based broadcasting for multi-agent average consensus, Automatica, vol. 49, pp. 245-252,
2013.

[44] A. Nedic, Asynchronous broadcast-based convex optimization over a network, IEEE Transactions on Automatic Control, vol. 56, no. 6, pp. 1337-1351,
2011.

[45] A. Wang, Event-based consensus control for single-integrator networks with communication time delays, Neurocomputing, vol. 173, part 3, pp. 1715-1719,
2016.

[46] G. Xie, H. Liu, L. Wang, and Y. Jia, Consensus in networked multi-agent systems via sampled control: Fixed topology case. Proceedings of the 2009
American Control Conference, pp. 3902-3907, 2009.

[47] Y. Cao and W. Ren, Sampled-data discrete-time coordination algorithms for double-integrator dynamics under dynamic directed interaction, International
Journal of Control, vol. 83, no. 3, pp. 506-515, 2010.

[48] Y. Zhang and Y. P. Tian, Consensus of data-sampled multi-agent systems with random communication delay and packet loss, IEEE Transactions on
Automatic Control, vol. 55, vol. 4, pp. 939-943, 2010.
 25

[49] J. Qin and H. Gao, A sufficient condition for convergence of sampled-data consensus for double-integrator dynamics with nonuniform and time-varying
communication delays, IEEE Transactions on Automatic Control, vol. 57, no. 9, pp. 2417-2422, 2012.

[50] H. Zhang and J. Zhou, Synchronization of sampled-data coupled harmonic oscillators with control inputs missing, Systems & Control Letters, vol. 61,
pp. 1277-1285, 2012.

[51] M. A. Aldajani, Logarithmic quantization in the least mean squares algorithm, Digital Signal Processing, vol. 18, pp. 321-333, 2008.

[52] D. Zelazo, A. Rahmani, and M. Mesbahi, Agreement via the edge Laplacian, Proceedings of the 46th IEEE Conference on Decision and Control,
pp. 2309-2314. 2007.

[53] R. A. Horn and C. R. Johnson, Matrix Analysis, New York: Cambridge University Press, 1985.

[54] C. Godsil and G. Royal, Algebraic Graph Theory, New York: Springer-Verlag, 2001.
[55] B. Mohar, The Laplacian spectrum of graphs, Graph Theory, Combinatorics, and Applications, vol. 2, pp. 871-898, 1991.

[56] J. Huang, C. Wen, W Wang, and Z.-P. Jiang, Adaptive stabilization and tracking control of a nonholonomic mobile robot with input saturation and
disturbance, Systems & Control Letters, vol. 62, pp. 234-241, 2013.

[57] T. Yang, A. A. Stoorvogel, H. F. Grip, and A. Saberi, International Journal of Robust and Nonlinear Control, vol. 24, pp. 548-566, 2014.

[58] H. Su, M. Z. Q. Chen, and G. Chen, Robust semi-global coordinated tracking of linear multi-agent systems with input saturation, International Journal
of Robust and Nonlinear Control, vol. 25, pp. 2375-2390, 2015.

[59] Y.-H. Lim and H.-S. Ahn, Consensus under saturation constraints in interconnection states, IEEE Transactions on Automatic Control, vol. 60, no. 11,
pp. 3053-3058, 2015.

[60] Z. Zhao, Y. Hong, and Z. Lin, Semi-global output consensus of a group of linear systems in the presence of external disturbances and actuator saturation:
An output regulation approach, International Journal of Robust and Nonlinear Control, vol. 26, pp. 1353-1375, 2016.

[61] X. Yin, D. Yue, and S. Hu, Adaptive periodic event-triggered consensus formulti-agent systems subject to input saturation, International Journal of Control,
vol. 89, no. 4, pp. 653-667, 2016.

[62] Z. Sun, and S. S. Ge. Stability theory of switched dynamical systems, Springer Science & Business Media, 2011.