This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2017.2671029, IEEE
Transactions on Automatic Control

Synchronization of multi-agent systems using

event-triggered and self-triggered broadcasts
João Almeida, Carlos Silvestre, and António M. Pascoal

Abstract—This paper addresses the problem of synchronizing received attention due to a flurry of theoretical developments.
a group of identical linear time-invariant agents that exchange Known as event-triggered control, in this new approach such
information through a communication network. The agents may tasks as sampling a signal or broadcasting information are
only broadcast information at discrete time instants and the
decision to execute a broadcast is based on an event-triggered only executed when deemed necessary according to some
communication protocol. We prove that with the proposed control triggering conditions, often dependent on the state of each
architecture the state of each agent converges to and remains in agent. For more details on this approach, the interested reader
a neighborhood of a desired reference signal and the closed- is referred to, e.g., [6]–[9] for the single plant case and to [10]–
loop system does not exhibit Zeno solutions. A self-triggered [12] for the case of multiple plants. It is important to point
implementation of the proposed event-triggered communication
protocol is also derived. out that in event-triggered control, triggering conditions must
be constantly monitored which may be infeasible for some
applications. To circumvent this issue, self-triggered control
I. I NTRODUCTION strategies were developed where instead of continuously test-
In this paper, we define a multi-agent system as a dynamical ing a triggering condition, an event scheduler computes when
system formed by a set of agents, each with dynamics mod- the next event should occur by using information available at
eled by a linear time-invariant (LTI) system, connected by a the current time instant (see, e.g., [13]–[16]).
communication network that provides them with the means to In a multi-agent scenario where agents have to communicate
exchange information. A survey of applications of multi-agent with each other, the event-triggered strategy is even more
systems presented in [1] illustrates how local decentralized relevant since the communication medium is often shared by
coordination strategies can be employed so that a desired all agents, meaning that if each agent tried to transmit too
global behavior is observed. A special class of applications often, successful communications would become impossible.
requires the agents to align their states in a well-defined sense, Hence, by resorting to event-triggered control techniques, a
with the most representative examples being the consensus and communication protocol that avoids redundant broadcasts of
synchronization problems (see, e.g., [1]–[5]). information is sought. These techniques have been applied
We address the synchronization problem for groups of to the consensus problem in [17]–[20]. We note that the
identical agents. Although the authors of [5] solve this problem consensus problem is a particular type of synchronization
for groups of heterogeneous agents, our goal is to drop the problem where the reference signal is constant.
assumption of continuous communication links present in [5] The contribution of this paper is twofold: i) we extend
by employing sampled-data control techniques. The objective the event-triggered consensus results reported in [18] for 1st
is to derive decentralized control laws and communication and 2nd order integrators with an undirected communica-
protocols capable of making the state of each agent converge tion network; this is done by deriving an event-triggered
to the same reference signal. communication protocol capable of achieving synchronization
Due to the digital nature of the communication network, an for a class of agents with LTI dynamics that are connected
additional constraint on the protocol design arises from the fact by a directed communication network and ii) we offer a
that communications can only occur at discrete time instants. self-triggered implementation of the proposed event-triggered
The standard approach would be to broadcast information communication protocol.
periodically. However, in recent years a different strategy has Notation: If {ak }k≥0 and {bk }k≥0 are two strictly in-
creasing sequences with elements in R, then their union is a
João Almeida and António M. Pascoal are with the Institute for Systems sequence {ck }k≥0 defined as the set of unique elements in
and Robotics (ISR/IST), LARSyS, Instituto Superior Técnico, Universidade de
Lisboa, Portugal (e-mail: {jalmeida,antonio}@isr.ist.utl.pt). Carlos Silvestre is
{ak }k≥0 and {bk }k≥0 reordered to satisfy ck < ck+1 for all
with the Department of Electrical and Computer Engineering of the Faculty k ≥ 0. We denote this by writing {ck }k≥0 = {ak }k≥0 ∪
of Science and Technology, University of Macau, Macao, China, on leave {bk }k≥0 . For a complex number z, <{z} denotes its real part.
from Instituto Superior Técnico, Universidade de Lisboa, Portugal (e-mail:
csilvestre@umac.mo).
For a signal x : [0, +∞) → Rn , if the limit from below at time
This work was supported by Fundação para a Ciência e a Tecnologia t ∈ [0, +∞) exists, then it is defined as x− (t) = lims↑t x(s).
(FCT), through ISR, under the LARSyS FCT (UID/EEA/50009/2013) funding If t is understood from context, we simply write x and x− to
program, projects MYRG2016-00097-FST and MYRG2015-00127-FST of the
University of Macau, the Macao Science and Technology Development Fund,
stand for x(t) and x− (t), respectively. A vector of dimension
under Grant FDCT/048/2014/A1, the European Union’s H2020 research and n whose entries are all equal to one is denoted by 1n . Given
innovation programme under the Marie Sklodovska-Curie grant agreement a collection of vectors {x1 , . . . , xN } where xi ∈ Rni , the
No. 642153, and the EU H2020 Framework programme under the WiMUST
project (H2020-ICT-645141). João Almeida benefited from Ph.D. Student
vector obtained by stacking all xi column-wise is represented
> >
by z = (x1 , . . . , xN ) = x>

Scholarship SFRH/BD/30605/2006 of FCT. 1 . . . x N . The symbol In

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2017.2671029, IEEE
Transactions on Automatic Control

denotes the identity matrix of dimension n. For a square initial conditions, the agents are not guaranteed to converge
matrix X, eX , kXk, and σ(X) denote its matrix exponential, to the same trajectory. In order to correct this misalignment,
its spectral norm (defined as its largest singular value), and the agents must exchange information among them by using
its spectrum (the set of eigenvalues of X), respectively. The a communication network. This network is modeled by a
symbol ⊗ denotes the Kronecker product. communication graph G that is assumed to be fixed over time
and where an edge (j, i) ∈ E means that agent i receives
II. G RAPH THEORY REVIEW information from agent j. Convergence to the same trajectory
must be distributed in the sense that vi may only depend on
In this section we introduce some necessary concepts and
information obtained from the in-neighbors of agent i. When
results from graph theory (adapted from [1], [21]) required for
continuous communication links among agents are allowed,
the presentation and analysis of our proposed solution for the
one possible solution is to let vi be given by
problem of event-triggered synchronization.
A (directed) graph G = G(V, E) consists of a finite set N
X
V = {1, 2, . . . , N } of N vertices and a finite set E ⊆ V × V vi = aij (ζj − ζi ), (3)
of m ordered pairs of vertices (i, j) named edges (in this j=1
paper, self-edges (i, i) are not allowed). An undirected graph where aij denotes the entries of the adjacency matrix associ-
is defined as a graph where (i, j) ∈ E if and only if ated with G. In [3] it is shown that if the following assumption
(j, i) ∈ E. If (i, j) ∈ E, then we say that vertex i is an in- is satisfied, then all agents synchronize asymptotically, that
neighbor of vertex j and that j is an out-neighbor of vertex is, for all initial conditions ζi (0) ∈ Rm , limt→+∞ kζi (t) −
i. The set of in-neighbors and the set of out-neighbors of ζj (t)k = 0, for all i, j ∈ {1, . . . , N }.
vertex i are defined as Ni− = {j ∈ V : (j, i) ∈ E} and
Ni+ = {j ∈ V : (i, j) ∈ E}, respectively. A path in G from Assumption 1. The communication graph G is a rooted graph
vertex i to vertex j is a sequence of distinct edges of the form and, for all i ∈ {1, . . . , m} and j ∈ {2, . . . , N },
{(i, i1 ), (i1 , i2 ), . . . , (ik , j)}. A vertex i is a root of a graph G
<{λi (Ar ) − λj (L)} < 0. (4)
if there exists a path in G from vertex i to every other vertex in
G. If G has at least one root, we say that it is a rooted graph. If a Assumption 1 requires the connectivity of the graph to be
graph G is undirected and rooted, then it is said to be connected strong enough to dominate the unstable dynamics in Ar so that
(in this case, all vertices are roots). A weighted graph is a (4) holds. One way to accomplish this is to design the graph
graph where a real number (weight) is associated with every topology or edge weights such that the non-zero eigenvalues of
edge in the graph (in this paper, all graphs are weighted). The L meet condition (4). This is always possible, as shown next.
adjacency matrix of a graph, denoted by A = [aij ] ∈ RN ×N , Suppose Gρ is a rooted graph with all edge weights equal to
is a square matrix with rows and columns indexed by the ρ > 0 and let Lρ denote its corresponding Laplacian matrix.
vertices and whose entries satisfy aij > 0 if (j, i) ∈ E and are Notice that Lρ = ρL1 . In this case, the stability condition (4)
zero otherwise (aij denotes the weight for edge (j, i) ∈ E). becomes
The in-degree matrix D of a graph is a diagonal matrix where
<{λi (Ar )} maxi <{λi (Ar )}
the (i, i)-entry is equal to the in-degree of vertex i defined as ρ > max = . (5)
P N i,j <{λj (L1 )} minj6=1 <{λj (L1 )}
j=1 aij . The Laplacian matrix of a graph G is defined as j6=1
L = D − A and has the following properties: i) L1N = 0
and there exists β ∈ RN , β > 1N = 1 such that β > L = 0; Thus, by selecting ρ sufficiently large, the unstable dynamics
ii) σ(L) = {0, λ2 , . . . , λN } with <{λi } > 0 for all non- of Ar (eigenvalues of Ar with positive real part) can be
zero eigenvalues; iii) G is a rooted graph if and only if 0 dominated. Note that if all eigenvalues of Ar are imaginary,
is a simple eigenvalue of L; iv) if G is a rooted graph, then that is, σ(Ar ) ⊂ iR, then (4) is satisfied for all rooted graphs.
there exist matrices L̂ ∈ R(N −1)×(N −1) , V ∈ RN×(N −1) , and
W ∈ R(N −1)×N such that σ(L̂) = σ(L)\{0}, 1N V is A. Event-triggered synchronization
>  −1
nonsingular, β W > = 1N V

, and To avoid the need for continuous communication links in
> (3) and inspired by the work reported in [18] for event-
L = 1N V diag(0, L̂) β W > .
  
(1)
triggered consensus, in this section we propose an event-
triggered solution for the multi-agent synchronization problem.
III. S YNCHRONIZATION OF MULTI - AGENT SYSTEMS
The proposed control architecture is represented in Fig. 1,
The multi-agent system that we consider consists of N from the point of view of agent i. The agent has been
agents with identical LTI dynamics. Each agent has a state augmented with additional state variables and is responsible
denoted by ζi ∈ Rm such that ζi (0) ∈ Rm and, for all t ≥ 0, for deciding when its current state should be broadcast to
the network, as represented by the broadcast event detector.
ζ̇i = Ar ζi + vi (2)
This event detector triggers the broadcast of the current value
where vi ∈ Rm is the control input and Ar ∈ Rm×m (Ar of ζi to the out-neighbors of agent i whenever a given state
may have unstable eigenvalues). To achieve synchronization, dependent condition is violated. The sequence of time instants
the state ζi must evolve in such a way that its trajectory is where these violations occur is referred to as the sequence of
eventually the same across all agents. Note that due to different broadcast times of agent i and is denoted by {bik }k≥0 .

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2017.2671029, IEEE
Transactions on Automatic Control

Augmented Agent ⇣i t = bik

The setup described in [18] for event-triggered consensus
Agent is recovered by taking m = 1 and Ar = 0. In this case, the
⇣i ⇣ˆi authors proved the following result.
Broadcast
⇣ˆi , {⇣ˆji }j2N event detector Theorem 1 ([18, Theorem 3.2]). If G is an undirected
i

connected graph and c0 > 0, then the closed-loop system does

⇣ˆli not exhibit Zeno solutions and each agent’s trajectory satisfies
.. ⇣j .. √
j
. t = bp .
N kLk
lim kζi (t) − ak ≤ c0 , (11)
t→+∞ λ2 (L)
Ni Ni+
PN
Network
for all ζi (0) ∈ R, where a = N1 i=1 ζi (0) and λ2 (L) is the
smallest nonzero eigenvalue of L.
Fig. 1. Proposed control architecture from the point of view of agent i.
In the next section, we extend Theorem 1 by allowing
directed graphs and an arbitrary Ar matrix as long as As-
The state of the augmented agent is described by three sumption 1 is satisfied.
variables: ζi , ζ̂i , and ζ̂ji with j ∈ Ni− . The dynamics of ζi
are as in (2). The state variable ζ̂i evolves according to a
B. Stability analysis
unperturbed model between broadcast times of agent i and is
reset to the current value of ζi when a broadcast occurs. The For analysis purposes, it is more convenient to work with
dynamics of ζ̂i may be written in the form of an impulsive the errors ei = ζ̂i − ζi that originate from the fact that ζ̂i is
system as used for feedback rather than ζi . The dynamics of ei are given
by
˙
(
ζ̂i = Ar ζ̂i , t ∈ [bik , bik+1 ), (6a)
ėi = Ar ei − vi , t ∈ [bik , bik+1 ),

(12a)
ζ̂i = ζi− , t = bik (6b) ei = 0, t = bik . (12b)
(for an introduction to impulsive systems see, e.g., [22]). The Using the error ei , (10) is equivalent to
additional states ζ̂ji represent local replicas of the state ζj
of agent i’s in-neighbors and are used to store information bik+1 = inf{t > bik : kei (t)k = c(t)}. (13)
received from them. When an in-neighbor of agent i, say
Let ζ = (ζ1 , . . . , ζN ) and e = (e1 , . . . , eN ) denote new
j ∈ Ni− , broadcasts the current value of ζj , this value is used
state vectors. Their dynamics are derived from (2), (9), and
to reset the value of ζ̂ji , as modeled by the impulsive system
(12), and may be written as
˙
(
ζ̂ji = Ar ζ̂ji , t ∈ [bjk , bjk+1 ),
  
(7a)
 
ζ̇ Z −L ⊗ Im ζ
= , t ∈ [bk , bk+1 ),


ζ̂ji = ζj− , t = bjk . ė L ⊗ I I ⊗ A + L ⊗ I e

(7b) m N r m


(14a)

  −
To remove the need for continuous communication links
  
 ζ I N ⊗ Im 0 ζ
 = , t = bk ,
among agents, (3) is replaced by (IN − Rk ) ⊗ Im e−

 e 0


N
(14b)
X
aij (ζ̂ji − ζ̂i ).
SN
vi = (8) where {bk }k≥0 = i=1 {bik }k≥0 ,
j=1
Z = IN ⊗ Ar − L ⊗ Im , (15)
Without loss of generality, suppose that, for all i, j ∈
{1, . . . , N }, ζ̂ji is initialized with the value ζj (0). Since ζ̂j and Rk = diag(r1,k , r2,k , . . . , rN,k ) is a diagonal matrix
and ζ̂ji have the same dynamics (compare (6) with i = j and whose entries satisfy ri,k = 1 if bip = bk for some p ≥ 0
(7)), we have that ζ̂ji (t) = ζ̂j (t) for all t ≥ 0. Therefore, for and are zero otherwise.
analysis purposes, only the state variables ζ̂j are required and We will show that each ζi converges to a neighborhood of
we may write (8) as the reference signal a(t) = (β > ⊗ Im )ζ(t). Note that if the
graph is undirected, then L is symmetric, β = 1N /N , and
N
X a(t) becomes the average of all ζi (t). The signal a satisfies
vi = aij (ζ̂j − ζ̂i ). (9)
j=1 ȧ = (β > ⊗ Im ) (Zζ − (L ⊗ Im )e)
If the broadcasted information were to arrive at each out- = (β > ⊗ Ar )ζ − ((β > L) ⊗ Im )(ζ + e)
neighbor of agent i at different times due to, e.g., transmission = (1 ⊗ Ar )(β > ⊗ Im )ζ
delays, then the previous simplification would not be possible.
= Ar a, (16)
Finally, the sequence of broadcast times satisfies
for all t ∈ [bk , bk+1 ), with initial condition a(0) = (β > ⊗
bik+1 = inf{t > bik : kζ̂i (t) − ζi (t)k = c(t)}, (10)
Im )ζ(0). When t = bk , we have that a = a− .
for all k ≥ 0, where bi0 = 0 and c(t) represents a time-varying Let δ(t) = ζ(t) − 1N ⊗ a(t). The norm of δ(t) is a measure
threshold defined as c(t) = c0 + c1 e−αt with c0 , c1 , α ≥ 0. of the mismatch among the state variables ζi of each agent,

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2017.2671029, IEEE
Transactions on Automatic Control

at time t. From (14a) and (16), it follows that, for all t ∈ Moreover, if c0 > 0, then, for all k ≥ 0 and all i ∈
[bk , bk+1 ), {1, . . . , N },
δ̇ = (IN ⊗ Ar )ζ − (L ⊗ Im )(ζ + e) − 1N ⊗ (Ar a) 1  ωc0 
bik+1 − bik ≥ θmin = log 1 + >0 (26)
= (IN ⊗ Ar )(ζ − 1N ⊗ a) − (L ⊗ Im )(δ + 1N ⊗ a + e) ω v̄
√
= (IN ⊗ Ar )δ − (L ⊗ Im )(δ + e). (17) where ω = λmax (Ar +A> r )/2 and v̄ = kLk(δ̄+ N (c0 +c1 )).

When t = bk , we have that δ = ζ −1N ⊗a = ζ − −1N ⊗a− = Proof. From (18), it follows that, for all t ≥ 0,
δ − . Note also that, using the properties of β, we obtain that, Z t
for all t ≥ 0, (β > ⊗ Im )δ = (β > ⊗ Im )ζ − (β > 1N ⊗ a) = δ(t) = eZt δ(0) − eZ(t−s) (L ⊗ Im )e(s)ds. (27)
0
a − 1 ⊗ a = 0. In summary, δ satisfies
 The triggering condition in (13) implies that, for all t ≥ 0,
δ̇ = Zδ − (L ⊗ Im )e, t ∈ [bk , bk+1 ), (18a) v
−
uN
δ=δ , t = bk , (18b) uX √
ke(t)k = t kei (t)k2 ≤ N c(t). (28)
and (β > ⊗ Im )δ(t) = 0 for all t ≥ 0. To derive a bound on i=1
the asymptotic behavior of δ, we need the following lemma.
Taking the norm in (27) and using Lemma 1, yields
Lemma 1. Let v ∈ RN m be such that (β > ⊗ Im )v = 0. If Z t
Assumption 1 holds, then there exist κ ≥ 1 and λ > 0 such kδ(t)k ≤ κe−λt kδ(0)k + κe−λ(t−s) k(L ⊗ Im )e(s)kds
that, for all t ≥ 0, 0
(28)
Z t √
keZt vk ≤ κe−λt kvk. (19) ≤ κe−λt kδ(0)k + κe−λ(t−s) kLk N c(s)ds
0
√ c0
Proof. Let L be decomposed as in (1). Then, the matrix Z −λt
1 − e−λt


≤ κe kδ(0)k + κ N kLk
defined in (15) may be written as λ
c1
e−αt − e−λt .




> + (29)
Z = 1N V ⊗ Im diag(Ar , Ẑ) β W > ⊗ Im , λ−α
 
 

(20) If Zeno solutions are avoided, then the limit in (25) exists and
> −1 δ̄∞ is obtained from (29) by letting t → +∞. The bound in
where β W > ⊗ Im = 1N V ⊗ Im
 
 

and (24) is obtained by rewriting (29) as
Ẑ = IN −1 ⊗ Ar − L̂ ⊗ Im . (21)
n √ c
0 c1 −αt o
kδ(t)k ≤ κ e−λt (kδ(0)k − c̄) + N kLk + e
λ λ−α
It then follows that 
≤ κ{max kδ(0)k − c̄, 0 + c̄}, (30)

>
eZt v = 1N V ⊗ Im ediag(Ar ,Ẑ)t β W > ⊗ Im v
 
 
and using the fact that max{a − b, 0} + b = max{a, b}.
= (V ⊗ Im ) eẐt (W ⊗ Im ) v, (22) To prove that the closed-loop system does not exhibit Zeno
solutions, we show that the time interval between consecutive
where we used the fact that (β > ⊗ Im )v = 0. Notice that
broadcasts of any agent is lower bounded by a positive number
σ(Ẑ) = σ(Ar ) − σ(L̂) (see, e.g., [23, Theorem 4.4.5]), hence
(this implies that the sequence {bk }k≥0 cannot have any
(4) implies that Ẑ is Hurwitz. Therefore, there exist κ1 ≥ 1
accumulation points). Let k ≥ 0 and i ∈ {1, . . . , N } be fixed.
and λ > 0 such that, for all t ≥ 0,
Using the fact that ei (bik ) = 0, (12a) implies that
keẐt k ≤ κ1 e−λt . (23) Z t
ei (t) = − eAr (t−s) vi (s)ds, (31)
Using (22), (23), and the fact that kX ⊗ Ik = kXk for bik
any matrix X, we conclude that (19) is satisfied for κ =
κ1 kV kkW k. for all t ∈ [bik , bik+1 ). Applying norms on both sides, we obtain
Z t Z t
Lemma 1 is an extension of Lemma 2.1 in [18] that is kei (t)k ≤ keAr (t−s) kkvi (s)kds ≤ eω(t−s) kvi (s)kds,
recovered by considering only undirected connected graphs bik bik (32)
and taking Ar = 0 (in this case, we may set κ = 1 and λ =
λ2 (L)). Using (18) and Lemma 1, we conclude the following. where we used the fact that ω is such that keAr t k ≤ eωt for all
t ≥ 0 (see, e.g., [24, Section 2]). Letting v = (v1 , . . . , vN ) =
Theorem 2 (Theorem 1 for Ar 6= 0 and directed graphs).
(L ⊗ Im )(ζ + e), we have that
If Assumption 1 holds, then, for all initial conditions ζ(0) ∈
RN m and all α < λ, the vector δ satisfies kvi k ≤ kvk = k(L ⊗ Im )(ζ + e)k = k(L ⊗ Im )(δ + e)k
kδ(t)k ≤ δ̄ = κ max {kδ(0)k, c̄} , (24) ≤ kLk(kδk + kek) ≤ v̄. (33)
√
for all t ≥ 0, where c̄ = N kLk(c0 /λ + c1 /(λ − α)) and Replacing (33) in (32) yields
√ Z t
v̄  ω(t−bik ) 
κ N kLk kei (t)k ≤ eω(t−s) v̄ds = e −1 . (34)
lim kδ(t)k ≤ δ̄∞ = c0 . (25) ω
t→+∞ λ bik

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2017.2671029, IEEE
Transactions on Automatic Control

Hence, a lower bound on the minimum time interval between from which we obtain
any two consecutive broadcast times of agent i is given by the
kei (t)k ≤ eω(t−bk ) (kei (bk )k + kv̄i (bk )k(t − bk )) . (38)
solution of v̄ eωθ − 1 = c0 ω, whose closed form is given in

(26). Since θmin is independent of both k and i, the lower The next broadcast time is then defined as bipi +1 = bk + θi∗
bound holds for all k ≥ 0 and i ∈ {1, . . . , N }. where θi∗ is the positive solution of
Notice that the asymptotic bound in (25) can be made eωθ (kei (bk )k + kv̄i (bk )kθ) = c(bk + θ). (39)
arbitrarily small by decreasing c0 , albeit at the expense of
making θmin smaller as well. Also, both κ and λ depend on Note that ei (bk ) and v̄i (bk ) are known to agent i at time
the weights assigned to each edge. Further study is required t = bk , thus they may be used to compute the next broadcast
to analyze how to exploit this degree of freedom (weight time. Taking c1 = α = kei (bk )k = 0 and using the fact
assignment) to achieve some desired closed-loop properties. that kv̄i (bk )k ≤ v̄, a lower bound on the minimum broadcast
interval of each agent is defined as the positive solution of
v̄θeωθ = c0 and denoted by θmin self
.
C. Self-triggered communication protocol
Remark 1. Solving (39) using a generic root finder may be
To avoid spending computational resources by constantly time consuming. As an alternative, we propose a method that
testing if the broadcast condition has been violated, in this computes an approximation that is strictly smaller. Note that
section we propose a self-triggered implementation of the (39) may be written as
event-triggered communication protocol defined in (13).
Suppose agent ` executes a broadcast at time t = bk . Let kv̄i (bk )kθ + kei (bk )k = c0 e−ωθ + c1 e−αbk e−(ω+α)θ , (40)
pj = max{p ≥ 0 : bjp ≤ bk } with j ∈ {1, . . . , N } denote the which is an equation of the form
index of the last broadcast of agent j (notice that b`p` = bk ). At
this point, instead of continuously testing the event condition ax + b = ce−αx + de−βx (41)
defined in (13) to determine the next broadcast time, agent `
where a, b, c, d, α, β ≥ 0. Let x∗ denote the unique positive
computes b`p` +1 using the information available at the current
solution of (41) (that exists if b < c + d). An approximation
time instant bk . At the same time, all its out-neighbors have
x1 < x∗ is obtained by exploiting the convexity of the
to recompute their next broadcast times as well to guarantee
exponential terms in (41). For fixed x0 ≥ 0 and γ ≥ 0, we
that their corresponding event conditions are satisfied. This is
have that e−γx ≥ e−γx0 (1 − γ(x − x0 )) for all x ≥ 0. Using
necessary because when ζ̂` is updated, vj (bk ) changes for all
this fact in (41), x1 is defined as
j ∈ N`+ thereby altering the trajectory of ζj and ej for t ≥ bk .
In what follows, let i ∈ {`} ∪ N`+ . To derive an expression ˜ − β(x1 − x0 ))
ax1 + b = c̃(1 − α(x1 − x0 )) + d(1 (42)
for the computation of bipi +1 at time t = bk , we start by ˜ + βx0 ) − b
c̃(1 + αx0 ) + d(1
solving (12a) in t, yielding ⇔ x1 = , (43)
˜ +a
c̃α + dβ
Z t
ei (t) = eAr (t−bk ) ei (bk ) − eAr (t−s) vi (s)ds, (35) where c̃ = ce−αx0 and d˜ = de−βx0 . A better approximation
bk x1 < x2 < x∗ is obtained by repeating this process, taking
for all t ∈ [bk , bk+1 ). We have that e` (bk ) = 0 but, in general, this time x0 = x1 . Starting with x0 = 0, this iterative process
ej (bk ) 6= 0 for j ∈ N`+ . Given (35), finding a closed-form generates a strictly increasing sequence {xk }k≥0 that tends to
solution for the triggering condition kei (bi,∗ i,∗
pi +1 )k = c(bpi +1 )
x∗ from below, that is, for all k ≥ 0, xk < xk+1 < x∗ and
is, in general, impossible. Instead of the exact solution, we will limk→+∞ xk = x∗ .
compute bipi +1 such that bipi +1 ≤ bi,∗ pi +1 thereby guaranteeing
that kei (bipi +1 )k ≤ c(bipi +1 ) is satisfied. The goal is to keep IV. E XAMPLE
the gap between bipi +1 and bi,∗ pi +1 as small as possible. The self- In this section, we compare the proposed event-triggered
triggered implementation is therefore expected to generate a and self-triggered communication protocols. We consider N =
sequence of broadcast times with an higher average broadcast 6 agents with the dynamics of a fourth-order oscillator where
rate than the one obtained in the event-triggered case.  
To compute bipi +1 , we note that the dynamics of ζ̂i given in 0 −1 1 0
1 0 0 1 
(6) imply that, for all t ∈ [bk , bk+1 ), ζ̂i (t) = eAr (t−bk ) ζi (bk ). Ar = 
0 0 0 − 1  .
 (44)
Thus, (9) may be written as vi (t) = eAr (t−bk ) v̄i (bk ) where 1
3
0 0 3 0
N
X The agents exchange information according to the communica-
v̄i (bk ) = aij (ζ̂j (bk ) − ζ̂i (bk )). (36)
j=1
tion graph shown in Fig. 2a. Since σ(Ar ) ⊂ iR, Assumption 1
is satisfied. Selecting all edge weights equal to ρ = 1, we
Using this fact in (35) yields have that σ(L) = {0, 0.5344, 1.5±i0.8660, 2.2328±i0.7926},
Z t kLk = 2.9364, and β = (1/3, 1/3, 1/3, 0, 0, 0). The value
Ar (t−bk )
ei (t) = e ei (bk ) − eAr (t−s) eAr (s−bk ) v̄i (bk )ds of λ is obtained by finding P  Im(N −1) and λ > 0
bk
such that Ẑ > P + P Ẑ + 2λP  0, where Ẑ is defined in
= eAr (t−bk ) ei (bk ) − (t − bk )eAr (t−bk ) v̄i (bk ), (37) (21). Then, κ is as defined in the proof of Lemma 1 with

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2017.2671029, IEEE
Transactions on Automatic Control

2 4 protocol extends the work reported in [18] for event-triggered

1 6
consensus, by allowing directed communication graphs and
more general agent dynamics. We showed that the proposed
3 5
control architecture achieves bounded synchronization errors
(a) and that the closed-system does not exhibit Zeno solutions.
1

R EFERENCES
ζ i, 1

[1] W. Ren and R. W. Beard, Distributed Consensus in Multi-vehicle

−1
6 Cooperative Control, ser. Communications and Control Engineering.
London: Springer-Verlag, 2008.
b ik

1 [2] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks

0 5 10 15 20 25 30 35 40 45 50
t (s) of agents with switching topology and time-delays,” IEEE Transactions
on Automatic Control, vol. 49, no. 9, pp. 1520–1533, Sep. 2004.
(b) [3] J. Fax and R. Murray, “Information flow and cooperative control of
vehicle formations,” IEEE Transactions on Automatic Control, vol. 49,
1
no. 9, pp. 1465–1476, Sep. 2004.
ζ i, 1

0 [4] L. Scardovi and R. Sepulchre, “Synchronization in networks of identical

linear systems,” Automatica, vol. 45, no. 11, pp. 2557–2562, Nov. 2009.
−1 [5] P. Wieland, R. Sepulchre, and F. Allgöwer, “An internal model principle
6
is necessary and sufficient for linear output synchronization,” Automat-
b ik

1 ica, vol. 47, no. 5, pp. 1068–1074, May 2011.

0 5 10 15 20 25 30 35 40 45 50 [6] P. Tabuada, “Event-triggered real-time scheduling of stabilizing control
t (s) tasks,” IEEE Transactions on Automatic Control, vol. 52, no. 9, pp.
(c) 1680–1685, Sep. 2007.
[7] K. J. Åström, “Event based control,” in Analysis and Design of Nonlin-
ear Control Systems, A. Astolfi and L. Marconi, Eds. Springer Berlin
0
10 Heidelberg, 2008, pp. 127–147.
[8] J. Lunze and D. Lehmann, “A state-feedback approach to event-based
!δ !

−2
10
control,” Automatica, vol. 46, no. 1, pp. 211–215, Jan. 2010.
event
sel f [9] M. C. F. Donkers and W. P. M. H. Heemels, “Output-based event-
−4
δ̄ ∞ triggered control with guaranteed L∞ -gain and improved and decentral-
10
0 5 10 15 20 25 30 35 40 45 50 ized event-triggering,” IEEE Transactions on Automatic Control, vol. 57,
t (s) no. 6, pp. 1362–1376, Jun. 2012.
(d) [10] X. Wang and M. D. Lemmon, “Event-triggering in distributed networked
control systems,” IEEE Transactions on Automatic Control, vol. 56,
Fig. 2. Example. (a) Communication graph. (b)-(c) Trajectories of ζi,1 and no. 3, pp. 586–601, Mar. 2011.
sequences of broadcast times for different communication protocols: (b) event- [11] C. De Persis, R. Sailer, and F. Wirth, “On a small-gain approach to
triggered; (c) self-triggered. (d) Trajectories of kδk and its asymptotic bound. distributed event-triggered control,” in Preprints of the 18th IFAC World
Congress, Milano, Italy, 28 Aug. – 2 Sept. 2011, pp. 2401–2406.
p [12] X. Wang, Y. Sun, and N. Hovakimyan, “Asynchronous task execution
κ1 = λmax (P )/λmin (P ). This yields λ = 0.5344 and in networked control systems using decentralized event-triggering,”
κ = 10.66. The triggering parameters in (10) are c0 = 0.001, Systems & Control Letters, vol. 61, no. 9, pp. 936–944, Sep. 2012.
[13] M. Velasco, P. Martı́, and J. M. Fuertes, “The self triggered task model
c1 = 0.499, and α = 0.25. The agents initial conditions for real-time control systems,” in Proc. of the 24th IEEE Real-Time
are ζ(0) = ζ̂(0) = g/kgk where g ∈ RmN has entries Systems Symp., Cancun, Mexico, 3–5 Dec. 2003, pp. 67–70.
gj = (2j−mN −1)/(mN −1) for j ∈ {1, . . . , mN [14] X. Wang and M. D. Lemmon, “Self-triggered feedback control systems
√ }. Note that with finite-gain L2 stability,” IEEE Transactions on Automatic Control,
kζ(0)k = 1 implies kδ(0)k ≤ kIN −1N β > k = 2. According vol. 54, no. 3, pp. 452–467, Mar. 2009.
to Theorem 2, we have δ̄ = 1.347 × 102 , v̄ = 3.990 × 102 , [15] A. Anta and P. Tabuada, “To sample or not to sample: Self-triggered
δ̄∞ = 1.435 × 10−1 , and θmin = θmin self
= 2.506 × 10−6 s. control for nonlinear systems,” IEEE Transactions on Automatic Control,
vol. 55, no. 9, pp. 2030–2042, Sep. 2010.
The simulation results are presented in Fig. 2b-d. The tra- [16] M. Mazo Jr., A. Anta, and P. Tabuada, “An ISS self-triggered implemen-
jectories of ζi,1 when using event-triggered and self-triggered tation of linear controllers,” Automatica, vol. 46, no. 8, pp. 1310–1314,
communication protocols are shown in Fig. 2b and Fig. 2c, Aug. 2010.
[17] D. V. Dimarogonas, E. Frazzoli, and K. H. Johansson, “Distributed
respectively (in the latter case, we solved (39) using two event-triggered control for multi-agent systems,” IEEE Transactions on
iterations of the method presented in Remark 1). In both cases, Automatic Control, vol. 57, no. 5, pp. 1291–1297, May 2012.
the difference between the trajectories of any two agents is [18] G. S. Seyboth, D. V. Dimarogonas, and K. H. Johansson, “Event-based
broadcasting for multi-agent average consensus,” Automatica, vol. 49,
within a certain error tolerance, a fact that is corroborated by no. 1, pp. 245–252, Jan. 2013.
the trajectory of kδk shown in Fig. 2d. The average sampling [19] G. Guo, L. Ding, and Q.-L. Han, “A distributed event-triggered trans-
intervals observed were between 1.4416 s and 2.8512 s in the mission strategy for sampled-data consensus of multi-agent systems,”
Automatica, vol. 50, no. 5, pp. 1489–1496, May 2014.
event-triggered case and between 1.2654 s and 1.6235 s in the [20] E. Garcia, Y. Cao, and D. W. Casbeer, “Decentralized event-triggered
self-triggered case, illustrating the conservativeness introduced consensus with general linear dynamics,” Automatica, vol. 50, no. 10,
in the derivation of the latter communication protocol. pp. 2633–2640, Oct. 2014.
[21] C. Godsil and G. Royle, Algebraic Graph Theory, ser. Graduate Texts
in Mathematics. New York, USA: Springer-Verlag, 2001.
V. C ONCLUSIONS [22] R. Goebel, R. G. Sanfelice, and A. R. Teel, “Hybrid dynamical systems,”
In this paper, we proposed and analyzed a control archi- IEEE Control Systems Magazine, vol. 29, no. 2, pp. 28–93, Apr. 2009.
[23] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis. New York,
tecture designed to achieve synchronization of a multi-agent NY, USA: Cambridge University Press, 1991.
system using event-triggered and self-triggered communica- [24] C. Van Loan, “The sensitivity of the matrix exponential,” SIAM J. on
tion protocols. The proposed event-triggered communication Numer. Analysis, vol. 14, no. 6, pp. 971–981, Dec. 1977.

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.