INVITED

PAPER

Consensus and Cooperation in

Networked Multi-Agent Systems
Algorithms that provide rapid agreement and teamwork between all participants
allow effective task performance by self-organizing networked systems.
By Reza Olfati-Saber, Member IEEE, J. Alex Fax, and Richard M. Murray, Fellow IEEE

ABSTRACT | This paper provides a theoretical framework for I. INTRODUCTION

analysis of consensus algorithms for multi-agent networked Consensus problems have a long history in computer
systems with an emphasis on the role of directed information science and form the foundation of the field of distributed
flow, robustness to changes in network topology due to computing [1]. Formal study of consensus problems in
link/node failures, time-delays, and performance guarantees.
groups of experts originated in management science and
An overview of basic concepts of information consensus in
statistics in 1960s (see DeGroot [2] and references therein).
networks and methods of convergence and performance The ideas of statistical consensus theory by DeGroot re-
analysis for the algorithms are provided. Our analysis frame- appeared two decades later in aggregation of information
work is based on tools from matrix theory, algebraic graph with uncertainty obtained from multiple sensors1 [3] and
theory, and control theory. We discuss the connections medical experts [4].
between consensus problems in networked dynamic systems Distributed computation over networks has a tradition
and diverse applications including synchronization of coupled in systems and control theory starting with the pioneering
oscillators, flocking, formation control, fast consensus in small-
work of Borkar and Varaiya [5] and Tsitsiklis [6] and
world networks, Markov processes and gossip-based algo-
Tsitsiklis, Bertsekas, and Athans [7] on asynchronous
rithms, load balancing in networks, rendezvous in space, asymptotic agreement problem for distributed decision-
distributed sensor fusion in sensor networks, and belief making systems and parallel computing [8].
propagation. We establish direct connections between spectral In networks of agents (or dynamic systems), Bcon-
and structural properties of complex networks and the speed sensus[ means to reach an agreement regarding a certain
of information diffusion of consensus algorithms. A brief quantity of interest that depends on the state of all agents.
introduction is provided on networked systems with nonlocal A Bconsensus algorithm[ (or protocol) is an interaction
information flow that are considerably faster than distributed
rule that specifies the information exchange between an
systems with lattice-type nearest neighbor interactions. Simu-
agent and all of its neighbors on the network.2
lation results are presented that demonstrate the role of small- The theoretical framework for posing and solving
world effects on the speed of consensus algorithms and consensus problems for networked dynamic systems was
cooperative control of multivehicle formations. introduced by Olfati-Saber and Murray in [9] and [10]
building on the earlier work of Fax and Murray [11], [12].
KEYWORDS | Consensus algorithms; cooperative control; The study of the alignment problem involving reaching an
flocking; graph Laplacians; information fusion; multi-agent agreementVwithout computing any objective functionsV
systems; networked control systems; synchronization of cou-
appeared in the work of Jadbabaie et al. [13]. Further
pled oscillators
theoretical extensions of this work were presented in [14]
and [15] with a look toward treatment of directed infor-
Manuscript received August 8, 2005; revised September 7, 2006. This work was
mation flow in networks as shown in Fig. 1(a).
supported in part by the Army Research Office (ARO) under Grant W911NF-04-1-0316.
R. Olfati-Saber is with Dartmouth College, Thayer School of Engineering, Hanover,
NH 03755 USA (e-mail: [email protected]).
1
J. A. Fax is with Northrop Grumman Corp., Woodland Hills, CA 91367 USA This is known as sensor fusion and is an important application of
(e-mail: [email protected]). modern consensus algorithms that will be discussed later.
R. M. Murray is with the California Institute of Technology, Control and Dynamical 2
The term Bnearest neighbors[ is more commonly used in physics
Systems, Pasadena, CA 91125 USA (e-mail: [email protected]).
than Bneighbors[ when applied to particle/spin interactions over a lattice
Digital Object Identifier: 10.1109/JPROC.2006.887293 (e.g., Ising model).

Authorized licensed use

0018-9219/$25.00 limited IEEE
Ó 2007 Vol.July
to: Universite De Sherbrooke. Downloaded on 95,16,2023
No. 1, January 2007UTC
at 14:01:05 | Proceedings
from IEEE Xplore.
ofRestrictions
the IEEE apply.215
Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

More recently, there has been a tremendous surge of

interestVamong researchers from various disciplines of
engineering and scienceVin problems related to multia-
gent networked systems with close ties to consensus prob-
lems. This includes subjects such as consensus [26]–[32],
collective behavior of flocks and swarms [19], [33]–[37],
sensor fusion [38]–[40], random networks [41], [42], syn-
chronization of coupled oscillators [42]–[46], algebraic
connectivity3 of complex networks [47]–[49], asynchro-
nous distributed algorithms [30], [50], formation control
for multirobot systems [51]–[59], optimization-based co-
operative control [60]–[63], dynamic graphs [64]–[67],
complexity of coordinated tasks [68]–[71], and consensus-
based belief propagation in Bayesian networks [72], [73].
A detailed discussion of selected applications will be pre-
sented shortly.
In this paper, we focus on the work described in five key
papersVnamely, Jadbabaie, Lin, and Morse [13], Olfati-
Saber and Murray [10], Fax and Murray [12], Moreau [14],
and Ren and Beard [15]Vthat have been instrumental in
paving the way for more recent advances in study of self-
organizing networked systems, or swarms. These networked
systems are comprised of locally interacting mobile/static
agents equipped with dedicated sensing, computing, and
Fig. 1. Two equivalent forms of consensus algorithms: (a) a network communication devices. As a result, we now have a better
of integrator agents in which agent i receives the state xj of its understanding of complex phenomena such as flocking
neighbor, agent j, if there is a link ði; jÞ connecting the two nodes;
and (b) the block diagram for a network of interconnected
[19], or design of novel information fusion algorithms for
dynamic systems all with identical transfer functions PðsÞ ¼ 1=s. sensor networks that are robust to node and link failures
The collective networked system has a diagonal transfer function [38], [72]–[76].
and is a multiple-input multiple-output (MIMO) linear system. Gossip-based algorithms such as the push-sum protocol
[77] are important alternatives in computer science to
Laplacian-based consensus algorithms in this paper.
The common motivation behind the work in [5], [6], Markov processes establish an interesting connection
and [10] is the rich history of consensus protocols in com- between the information propagation speed in these two
puter science [1], whereas Jadbabaie et al. [13] attempted categories of algorithms proposed by computer scientists
to provide a formal analysis of emergence of alignment in and control theorists [78].
the simplified model of flocking by Vicsek et al. [16]. The The contribution of this paper is to present a cohesive
setup in [10] was originally created with the vision of de- overview of the key results on theory and applications of
signing agent-based amorphous computers [17], [18] for consensus problems in networked systems in a unified
collaborative information processing in networks. Later, framework. This includes basic notions in information
[10] was used in development of flocking algorithms with consensus and control theoretic methods for convergence
guaranteed convergence and the capability to deal with and performance analysis of consensus protocols that
obstacles and adversarial agents [19]. heavily rely on matrix theory and spectral graph theory. A
Graph Laplacians and their spectral properties [20]–[23] byproduct of this framework is to demonstrate that seem-
are important graph-related matrices that play a crucial role ingly different consensus algorithms in the literature [10],
in convergence analysis of consensus and alignment algo- [12]–[15] are closely related. Applications of consensus
rithms. Graph Laplacians are an important point of focus problems in areas of interest to researchers in computer
of this paper. It is worth mentioning that the second smallest science, physics, biology, mathematics, robotics, and con-
eigenvalue of graph Laplacians called algebraic connectivity trol theory are discussed in this introduction.
quantifies the speed of convergence of consensus algo-
rithms. The notion of algebraic connectivity of graphs has A. Consensus in Networks
appeared in a variety of other areas including low-density The interaction topology of a network of agents is rep-
parity-check codes (LDPC) in information theory and com- resented using a directed graph G ¼ ðV; EÞ with the set of
munications [24], Ramanujan graphs [25] in number theory nodes V ¼ f1; 2; . . . ; ng and edges E  V  V. The
and quantum chaos, and combinatorial optimization prob-
3
lems such as the max-cut problem [21]. To be defined in Section II-A.

216Authorized licensed useof

Proceedings limited
theto:IEEE | Vol.De
Universite 95,Sherbrooke.
No. 1, January 2007 on July 16,2023 at 14:01:05 UTC from IEEE Xplore.
Downloaded Restrictions apply.
 Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

neighbors of agent i are denoted by Ni ¼ fj 2 V : ði; jÞ 2 Eg. network to compute the average. There are a variety of
According to [10], a simple consensus algorithm to reach an functions that can be computed in a similar fashion using
agreement regarding the state of n integrator agents with synchronous or asynchronous distributed algorithms (see
dynamics x_ i ¼ ui can be expressed as an nth-order linear [10], [28], [30], [73], and [76]).
system on a graph
B. The f -Consensus Problem and Meaning
X 
of Cooperation
x_ i ðtÞ ¼ xj ðtÞ xi ðtÞ þ bi ðtÞ; xi ð0Þ ¼ zi To understand the role of cooperation in performing
j2Ni coordinated tasks, we need to distinguish between un-
2 R; bi ðtÞ ¼ 0: (1) constrained and constrained consensus problems. An
unconstrained consensus problem is simply the alignment
problem in which it suffices that the state of all agents
The collective dynamics of the group of agents following asymptotically be the same. In contrast, in distributed
protocol (1) can be written as computation of a function f ðzÞ, the state of all agents has to
asymptotically become equal to f ðzÞ, meaning that the
consensus problem is constrained. We refer to this con-
x_ ¼ Lx (2) strained consensus problem as the f -consensus problem.
Solving the f -consensus problem is a cooperative task
and requires willing participation of all the agents. To
where L ¼ ½lij is the graph Laplacian of the network and its demonstrate this fact, suppose a single agent decides not to
elements are defined as follows: cooperate with the rest of the agents and keep its state
unchanged. Then, the overall task cannot be performed
 despite the fact that the rest of the agents reach an agree-
1; j 2 Ni ment. Furthermore, there could be scenarios in which
lij ¼ (3)
jNi j; j ¼ i: multiple agents that form a coalition do not cooperate with
the rest and removal of this coalition of agents and their
links might render the network disconnected. In a dis-
Here, jNi j denotes the number of neighbors of node i (or connected network, it is impossible for all nodes to reach
out-degree of node i). Fig. 1 shows two equivalent forms of an agreement (unless all nodes initially agree which is a
the consensus algorithm in (1) and (2) for agents with a trivial case).
scalar state. The role of the input bias b in Fig. 1(b) is From the above discussion, cooperation can be infor-
defined later. mally interpreted as Bgiving consent to providing one’s
According to the definition of graph
P Laplacian in (3), all state and following a common protocol that serves the
row-sums of L are zero because of j lij ¼ 0. Therefore, L group objective.[
always has a zero eigenvalue 1 ¼ 0. This zero eigenvalues One might think that solving the alignment problem is
corresponds to the eigenvector 1 ¼ ð1; . . . ; 1ÞT because 1 not a cooperative task. The justification is that if a single
belongs to the null-space of LðL1 ¼ 0Þ. In other words, an agent (called a leader) leaves its value unchanged, all
equilibrium of system (2) is a state in the form x ¼ others will asymptotically agree with the leader according
ð; . . . ; ÞT ¼ 1 where all nodes agree. Based on ana- to the consensus protocol and an alignment is reached.
lytical tools from algebraic graph theory [23], we later show However, if there are multiple leaders where two of whom
that x is a unique equilibrium of (2) (up to a constant are in disagreement, then no consensus can be asymptot-
multiplicative factor) for connected graphs. ically reached. Therefore, alignment is in general a coop-
One can show that for a connected network, the erative task as well.
equilibrium x ¼ ð; . . . ; ÞT is globally exponentially
P Formal analysis of the behavior of systems that involve
stable. Moreover, the consensus value is  ¼ 1=n i zi more than one type of agent is more complicated, partic-
that is equal to the average of the initial values. This im- ularly, in presence of adversarial agents in noncooperative
plies that irrespective of the initial value of the state of games [79], [80]. The focus of this paper is on cooperative
each agent, all agents reach an asymptotic P consensus multi-agent systems.
regarding the value of the function f ðzÞ ¼ 1=n i zi .
While the calculation of f ðzÞ is simple for small net- C. Iterative Consensus and Markov Chains
works, its implications for very large networks is more In Section II, we show how an iterative consensus
interesting. For example, if a network has n ¼ 106 nodes algorithm that corresponds to the discrete-time version of
and each node can only talk to log10 ðnÞ ¼ 6 neighbors, system (1) is a Markov chain
finding the average value of the initial conditions of the
nodes is more complicated. The role of protocol (1) is to
provide a systematic consensus mechanism in such a large ðk þ 1Þ ¼ ðkÞP (4)

Vol.July
Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on 95,16,2023
No. 1, January 2007UTC
at 14:01:05 | Proceedings
from IEEE Xplore.ofRestrictions
the IEEE apply.217
Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

with P ¼ I L and a small  9 0. Here, the ith element shed light on performance and convergence analysis of
of the row vector ðkÞ denotes the probability of being in self-synchrony in oscillator networks [42].
state i at iteration k. It turns out that for any arbitrary
graph G with Laplacian L andPa sufficiently small , the 2) Flocking Theory: Flocks of mobile agents equipped
matrix P satisfies the property j pij ¼ 1 with pij  0; 8i; j. with sensing and communication devices can serve as
Hence, P is a valid transition probability matrix for the mobile sensor networks for massive distributed sensing in an
Markov chain in (4). The reason matrix theory [81] is so environment [87]. A theoretical framework for design and
widely used in analysis of consensus algorithms [10], analysis of flocking algorithms for mobile agents with
[12]–[15], [64] is primarily due to the structure of P in (4) obstacle-avoidance capabilities is developed by Olfati-
and its connection to graphs.4 Saber [19]. The role of consensus algorithms in particle-
There are interesting connections between this Markov based flocking is for an agent to achieve velocity matching
chain and the speed of information diffusion in gossip- with respect to its neighbors. In [19], it is demonstrated
based averaging algorithms [77], [78]. that flocks are networks of dynamic systems with a
One of the early applications of consensus problems dynamic topology. This topology is a proximity graph that
was dynamic load balancing [82] for parallel processors depends on the state of all agents and is determined locally
with the same structure as system (4). To this date, load for each agent, i.e., the topology of flocks is a state-
balancing in networks proves to be an active area of dependent graph. The notion of state-dependent graphs
research in computer science. was introduced by Mesbahi [64] in a context that is
independent of flocking.
D. Applications
Many seemingly different problems that involve inter- 3) Fast Consensus in Small-Worlds: In recent years,
connection of dynamic systems in various areas of science network design problems for achieving faster consensus
and engineering happen to be closely related to consensus algorithms has attracted considerable attention from a
problems for multi-agent systems. In this section, we pro- number of researchers. In Xiao and Boyd [88], design of
vide an account of the existing connections. the weights of a network is considered and solved using
semi-definite convex programming. This leads to a slight
1) Synchronization of Coupled Oscillators: The problem of increase in algebraic connectivity of a network that is a
synchronization of coupled oscillators has attracted numer- measure of speed of convergence of consensus algorithms.
ous scientists from diverse fields including physics, An alternative approach is to keep the weights fixed and
biology, neuroscience, and mathematics [83]–[86]. This design the topology of the network to achieve a relatively
is partly due to the emergence of synchronous oscillations high algebraic connectivity. A randomized algorithm for
in coupled neural oscillators. Let us consider the network design is proposed by Olfati-Saber [47] based on
generalized Kuramoto model of coupled oscillators on a random rewiring idea of Watts and Strogatz [89] that led to
graph with dynamics creation of their celebrated small-world model. The random
rewiring of existing links of a network gives rise to
X considerably faster consensus algorithms. This is due to
_i ¼ sinðj i Þ þ !i (5) multiple orders of magnitude increase in algebraic
j2Ni connectivity of the network in comparison to a lattice-
type nearest-neighbort graph.
where i and !i are the phase and frequency of the ith
4) Rendezvous in Space: Another common form of
oscillator. This model is the natural nonlinear extension of
consensus problems is rendezvous in space [90], [91]. This
the consensus algorithm in (1) and its linearization around
is equivalent to reaching a consensus in position by a num-
the aligned state 1 ¼ . . . ¼ n is identical to system (2)
ber of agents with an interaction topology that is position
P a nonzero input bias bi ¼ ð!i !Þ=
plus  with ! ¼ 1=
induced (i.e., a proximity graph). We refer the reader to
n i !i after a change of variables xi ¼ ði !tÞ=  .
[92] and references therein for a detailed discussion. This
In [43], Sepulchre et al. show that if is sufficiently
type of rendezvous is an unconstrained consensus problem
large, then for a network with all-to-all links, synchroni-
that becomes challenging under variations in the network
zation to the aligned state is globally achieved for all ini-
topology. Flocking is somewhat more challenging than
tial states. Recently, synchronization of networked
rendezvous in space because it requires both interagent
oscillators under variable time-delays was studied in [45].
and agent-to-obstacle collision avoidance.
We believe that the use of convergence analysis methods
that utilize the spectral properties of graph Laplacians will
5) Distributed Sensor Fusion in Sensor Networks: The
4 most recent application of consensus problems is distrib-
In honor of the pioneering contributions of Oscar Perron (1907) to
the theory of nonnegative matrices, were refer to P as the Perron Matrix of uted sensor fusion in sensor networks. This is done by
graph G (See Section II-C for details). posing various distributed averaging problems require to

218Authorized licensed useof

Proceedings limited
theto:IEEE | Vol.De
Universite 95,Sherbrooke.
No. 1, January 2007 on July 16,2023 at 14:01:05 UTC from IEEE Xplore.
Downloaded Restrictions apply.
 Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

implement a Kalman filter [38], [39], approximate vehicle systems is provided in Section IV. Some simulation
Kalman filter [74], or linear least-squares estimator [75] results related to consensus in complex networks including
as average-consensus problems. Novel low-pass and high- small-worlds are presented in Section V. Finally, some
pass consensus filters are also developed that dynamically concluding remarks are stated in Section VI.
calculate the average of their inputs in sensor networks
[39], [93].
II. INFORMATION CONSENSUS
6) Distributed Formation Control: Multivehicle systems Consider a network of decision-making agents with
are an important category of networked systems due to dynamics x_ i ¼ ui interested in reaching a consensus via
their commercial and military applications. There are two local communication with their neighbors on a graph
broad approaches to distributed formation control: i) rep- G ¼ ðV; EÞ. By reaching a consensus, we mean asymptot-
resentation of formations as rigid structures [53], [94] and ically converging to a one-dimensional agreement space
the use of gradient-based controls obtained from their characterized by the following equation:
structural potentials [52] and ii) representation of form-
ations using the vectors of relative positions of neighboring x1 ¼ x2 ¼ . . . ¼ xn :
vehicles and the use of consensus-based controllers with
input bias. We discuss the later approach here.
A theoretical framework for design and analysis of This agreement space can be expressed as x ¼ 1 where
distributed controllers for multivehicle formations of type 1 ¼ ð1; . . . ; 1ÞT and  2 R is the collective decision of the
ii) was developed by Fax and Murray [12]. Moving in group of agents. Let A ¼ ½aij be the adjacency matrix of
formation is a cooperative task and requires consent and graph G. The set of neighbors of a agent i is Ni and
collaboration of every agent in the formation. In [12], defined by
graph Laplacians and matrix theory were extensively used
which makes one wonder whether relative-position-based
Ni ¼ fj 2 V : aij 6¼ 0g; V ¼ f1; . . . ; ng:
formation control is a consensus problem. The answer is
yes. To see this, consider a network of self-interested agents
whose individual
P desire is to minimize their local cost Agent i communicates with agent j if j is a neighbor of i (or
Ui ðxÞ ¼ j2Ni kxj xi rij k2 via a distributed algorithm aij 6¼ 0). The set of all nodes and their neighbors defines
(xi is the position of vehicle i with dynamics x_ i ¼ ui and rij the edge set of the graph as E ¼ fði; jÞ 2 V  V : aij 6¼ 0g.
is a desired intervehicle relative-position vector). Instead, A dynamic graph GðtÞ ¼ ðV; EðtÞÞ is a graph in which
if the agents P use gradient-descent algorithm on the the set of edges EðtÞ and the adjacency matrix AðtÞ are
collective cost ni¼1 Ui ðxÞ using the following protocol: time-varying. Clearly, the set of neighbors Ni ðtÞ of every
agent in a dynamic graph is a time-varying set as well.
X X
x_ i ¼ ðxj xi rij Þ ¼ ðxj xi Þ þ bi (6) Dynamic graphs are useful for describing the network
j2Ni j2Ni topology of mobile sensor networks and flocks [19].
It is shown in [10] that the linear system
P
with input bias bi ¼ j2Ni rji [see Fig. 1(b)], the objective
X  
of every agent will be achieved. This is the same as the x_ i ðtÞ ¼ aij xj ðtÞ xi ðtÞ (7)
consensus algorithm in (1) up to the nonzero bias terms bi . j2Ni
This nonzero bias plays no role in stability analysis of sys-
tem (6). Thus, distributed formation control for integrator
agents is a consensus problem. The main contribution of is a distributed consensus algorithm, i.e., guarantees con-
the work by Fax and Murray is to extend this scenario to vergence to a collective decision via local interagent
the case where all agents are multiinput multioutput linear interactions. Assuming that the graph is undirected
systems x_ i ¼ Axi þ Bui . Stability analysis of relative- (aij ¼ aji for all i; j), it follows that the sumPof the state
position-based formation control for multivehicle systems of all nodes is an invariant quantity, or i x_ i ¼ 0. In
is extensively covered in Section IV. particular, applying this condition twice at times t ¼ 0 and
t ¼ 1 gives the following result
E. Outline
The outline of the paper is as follows. Basic concepts 1X
and theoretical results in information consensus are ¼ xi ð0Þ:
n i
presented in Section II. Convergence and performance
analysis of consensus on networks with switching topology
are given in Section III. A theoretical framework for In other words, if a consensus is asymptotically reached,
cooperative control of formations of networked multi- then necessarily the collective decision is equal to the

Vol.July
Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on 95,16,2023
No. 1, January 2007UTC
at 14:01:05 | Proceedings
from IEEE Xplore.
ofRestrictions
the IEEE 219
apply.
Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

average of the initial state of all nodes. A consensus algo- A. Algebraic Connectivity and Spectral Properties
rithm with this specific invariance property is called an of Graphs
average-consensus algorithm [9] and has broad applications Spectral properties of Laplacian matrix are instrumen-
in distributed computing on networks (e.g., sensor fusion tal in analysis of convergence of the class of linear
in sensor networks). consensus algorithms in (7). According to Gershgorin
The dynamics of system (7) can be expressed in a theorem [81], all eigenvalues of L in the complex plane are
compact form as located in a closed disk centered at  þ 0j with a radius of
 ¼ maxi di , i.e., the maximum degree of a graph. For
undirected graphs, L is a symmetric matrix with real
x_ ¼ Lx (8)
eigenvalues and, therefore, the set of eigenvalues of L can
be ordered sequentially in an ascending order as
where L is known as the graph Laplacian of G. The graph
Laplacian is defined as
0 ¼ 1  2      n  2: (12)
L¼D A (9)

The zero eigenvalue is known as the trivial eigenvalue of L.

where D ¼ diagðdP 1 ; . . . ; dn Þ is the degree matrix of G with For a connected graph G, 2 9 0 (i.e., the zero eigenvalue
elements di ¼ j6¼i aij and zero off-diagonal elements. By is isolated). The second smallest eigenvalue of Laplacian
definition, L has a right eigenvector of 1 associated with 2 is called algebraic connectivity of a graph [20]. Algebraic
the zero eigenvalue5 because of the identity L1 ¼ 0. connectivity of the network topology is a measure of
For the case of undirected graphs, graph Laplacian performance/speed of consensus algorithms [10].
satisfies the following sum-of-squares (SOS) property:
Example 1: Fig. 2 shows two examples of networks of
1 X integrator agents with different topologies. Both graphs
xT Lx ¼ aij ðxj xi Þ2 : (10) are undirected and have 0–1 weights. Every node of the
2 ði;jÞ2E
graph in Fig. 2(a) is connected to its 4 nearest neighbors on
a ring. The other graph is a proximity graph of points that
By defining a quadratic disagreement function as are distributed uniformly at random in a square. Every
node is connected to all of its spatial neighbors within a
closed ball of radius r 9 0. Here are the important degree
1
’ðxÞ ¼ xT Lx (11) information and Laplacian eigenvalues of these graphs
2

it becomes apparent that algorithm (7) is the same as aÞ1 ¼ 0; 2 ¼ 0:48; n ¼ 6:24;  ¼ 4
bÞ1 ¼ 0; 2 ¼ 0:25; n ¼ 9:37;  ¼ 8: (13)
x_ ¼ r’ðxÞ

or the gradient-descent algorithm. This algorithm globally In both cases, i G 2 for all i.
asymptotically converges to the agreement space provided
that two conditions hold: 1) L is a positive semidefinite
B. Convergence Analysis for Directed Networks
The convergence analysis of the consensus algorithm in
matrix; 2) the only equilibrium of (7) is 1 for some .
(7) is equivalent to proving that the agreement space
Both of these conditions hold for a connected graph and
follow from the SOS property of graph Laplacian in (10). characterized by x ¼ 1;  2 R is an asymptotically stable
Therefore, an average-consensus is asymptotically reached equilibrium of system (7). The stability properties of
for all initial states. This fact is summarized in the system (7) is completely determined by the location of the
following lemma. Laplacian eigenvalues of the network. The eigenvalues of
the adjacency matrix are irrelevant to the stability analysis
of system (7), unless the network is k-regular (all of its
Lemma 1: Let G be a connected undirected graph. Then,
nodes have the same degree k).
the algorithm in (7) asymptotically solves an average-
The following lemma combines a well-known rank
consensus problem for all initial states.
property of graph Laplacians with Gershgorin theorem to
provide spectral characterization of Laplacian of a fixed
5
These properties were discussed earlier in the introduction for directed network G. Before stating the lemma, we need to
graphs with 0–1 weights. define the notion of strong connectivity of graphs. A graph

220Authorized licensed useof

Proceedings limited
theto:IEEE | Vol.De
Universite 95,Sherbrooke.
No. 1, January 2007 on July 16,2023 at 14:01:05 UTC from IEEE Xplore.
Downloaded Restrictions apply.
 Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

spanning tree if there exists a node r (a root) such that all

other nodes can be linked to r via a directed path. This type
of condition on existence of directed spanning trees have
appeared in [13]–[15]. The root node is commonly known
as a leader [13].
The essential results regarding convergence and deci-
sion value of Laplacian-based consensus algorithms for
directed networks with a fixed topology are summarized in
the following theorem. Before stating this theorem, we
need to define an important class of digraphs that appear
frequently throughout this section.

Definition 1: (balanced
P digraphs
P [10]) A digraph G is
called balanced if j6¼i aij ¼ j6¼i aji for all i 2 V.
In a balanced digraph, the total weight of edges
entering a node and leaving the same node are equal for all
nodes. The most important property of balanced digraphs
is that w ¼ 1 is also a left eigenvector of their Laplacian
(or 1T L ¼ 0).

Theorem 1: Consider a network of n agents with topol-

ogy G applying the following consensus algorithm:

X  
x_ i ðtÞ ¼ aij xj ðtÞ xi ðtÞ ; xð0Þ ¼ z: (14)
j2Ni

Fig. 2. Examples of networks with n ¼ 20 nodes: (a) a regular network

Suppose G is a strongly connected digraph. Let L be the
with 80 links and (b) a random network with 45 links. Laplacian of G with a left eigenvector ¼ ð 1 ; . . . ; n Þ
satisfying T L ¼ 0. Then
i) a consensus is asymptotically reached for all
initial states;
is strongly connected (SC) if there is a directed path ii) the algorithm solves the f -consensus problem with
connecting any two arbitrary nodes s; t of the graph.6 the linear function f ðzÞ T T
P¼ ð zÞ=ð P1Þ, i.e., the
group decision is  ¼ i wi zi with i wi ¼ 1;
Lemma 2: (spectral localization) Let G be a strongly iii) if the digraph is balanced, an average-consensus
P is
connected digraph on n nodes. Then rankðLÞ ¼ n 1 and asymptotically reached and  ¼ ð i xi ð0ÞÞ=n.
all nontrivial eigenvalues of L have positive real parts. Proof: The convergence of the consensus algorithm
Furthermore, suppose G has c  1 strongly connected follows from Lemma 2. To show part ii), note that the
components, then rankðLÞ ¼ n c. collective dynamics of the network is x_ ¼ Lx. This means
Proof: The proof of the rank property for digraphs is that y ¼ T x is an invariant quantity due to y_ ¼ T
Lx ¼
T T
given in [10]. The proof for undirected graphs is available 0; 8 x. Thus, limt!1 yðtÞ ¼ yð0Þ, or ð1Þ P ¼ xð0Þ that
T
in the algebraic graph theory literature [23]. The positivity implies theP group decision is  ¼ ð zÞ= i i . Setting
of the real parts of the eigenvalues follow from the fact that wi ¼ i = i i , we get  ¼ wT z. Part iii) follows as a special
all eigenvalues are located in a Gershgorin disk in the case of the statement in part ii) because for a balanced
closed right-hand plane that touches the imaginary axis at digraph ¼ 1 and wi ¼ 1=n; 8i. h
zero. The second part follows from the first part after
relabeling the nodes of the digraph so that its Laplacian Remark 2: In [10], it is shown that a necessary and suf-
becomes a block diagonal matrix. h ficient condition for L to have a left eigenvector of ¼ 1 is
that G must be a balanced digraph.
Remark 1: Lemma 2 holds under a weaker condition of A challenging problem is to analyze convergence of a
existence of a directed spanning tree for G. G has a directed consensus algorithm for a dynamic network with a switching
topology GðtÞ that is time-varying. Various aspects of this
6
The notion of strong connectivity applies to directed graphs (or problem has been addressed by several groups during the
digraphs). For undirected graphs SC is the same as connectivity. recent years [10], [13]–[15] and will be discussed in detail.

Vol.July
Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on 95,16,2023
No. 1, January 2007UTC
at 14:01:05 | Proceedings
from IEEE Xplore.ofRestrictions
the IEEE apply.221
Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

C. Consensus in Discrete-Time and Matrix Theory Based on Gershgorin theorem, all eigenvalues of L are in the
An iterative form of the consensus algorithm can be disk js j  . Defining z ¼ 1 s=, we have jzj  1
stated as follows in discrete-time: which proves part ii). If G is a balanced digraph, then 1 is
the left eigenvector of L, or 1T L ¼ 0. This means that
X   1T P ¼ 1T 1T L ¼ 1T which implies the column sums of
xi ðk þ 1Þ ¼ xi ðkÞ þ  aij xj ðkÞ xi ðkÞ : (15) P are 1. This combined with the result in part i) gives
j2Ni
part iii). To prove part iv), note that if G is strongly con-
nected, then P is an irreducible matrix [81]. To prove that P
The discrete-time collective dynamics of the network is primitive, we need to establish that it has a single eigen-
under this algorithm can be written as value with maximum modulus of 1. For all 0 G  G 1=, the
transformation ¼ 1 s maps the circle js j ¼ 
into a circle that is located strictly inside a unit disk passing
xðk þ 1Þ ¼ PxðkÞ (16) through the point ¼ 1. This means that only a single
eigenvalue at 1 ¼ 1 can have a modulus of 1. h
with P ¼ I L (I is the identity matrix) and  9 0 is the
step-size. In general, P ¼ expð LÞ and the algorithm in Remark 3: The condition  G 1= in part iv) is nec-
(15) is a special case that only uses communication with essary. If an incorrect step-size of  ¼ 1= is used. Then, P
first-order neighbors.7 We refer to P as Perron matrix of a would no longer be a primitive matrix because it could
graph G with parameter . have multiple eigenvalues of modulus 1. The counter-
Three important types of nonnegative matrices are example is a directed cycle of length n with a Laplacian
irreducible, stochastic, and primitive (or ergodic) matrices that has n roots on the boundary of the Gershgorin disk
[81]. A matrix A is irreducible if its associated graph is js j  . With the choice of  ¼ 1= ¼ 1, one gets a
strongly connected. A nonnegative matrix is called row (or Perron matrix that is irreducible but has n eigenvalues on
column) stochastic if all of its row-sums (or column-sums) the boundary of the unit circle. This is a common mistake
are 1. An irreducible stochastic matrix P is primitive if it has that is repeated by some of the researchers in the past.
only one eigenvalue with maximum modulus. The convergence analysis of the discrete-time consen-
sus algorithm relies on the following well-known lemma in
Lemma 3: Let G be aP digraph with n nodes and maxi- matrix theory.
mum degree  ¼ maxi ð j6¼i aij Þ. Then, the Perron matrix
P with parameter  2 ð0; 1= satisfies the following Lemma 4: (Perron-Frobenius, [81]) Let P be a primitive
properties. nonnegative matrix with left and right eigenvectors w and
i) P is a row stochastic nonnegative matrix with a v, respectively, satisfying Pv ¼ v, wT P ¼ wT , and vT w ¼ 1.
trivial eigenvalue of 1; Then limk!1 Pk ¼ vwT .
ii) All eigenvalues of P are in a unit circle; The convergence and group decision properties of
iii) If G is a balanced graph, then P is a doubly iterative consensus algorithms x Px with row stochastic
stochastic matrix; Perron matrices is stated in the following result. It turns
iv) If G is strongly connected and 0 G  G 1=, then out that this discrete-time convergence result is almost
P is a primitive matrix. identical to its continuous-time counterpart.
Proof: Since P ¼ I L, we get P1 ¼ 1 L1 ¼ 1
which means the row sums of P is 1. Moreover, 1 is a trivial Theorem 2: Consider a network of agents xi ðk þ 1Þ ¼
eigenvalue of P for all graphs. To show that P is non- xi ðkÞ þ ui ðkÞ with topology G applying the distributed
negative, notice that P ¼ I D þ A due to definition of consensus algorithm
Laplacian L ¼ D A. A is a nonnegative matrix. The di-
agonal elements of I D are 1 di  1 di =  0 X  
xi ðk þ 1Þ ¼ xi ðkÞ þ  aij xj ðkÞ xi ðkÞ (18)
which implies I D is nonnegative. Since the sum of two j2Ni
nonnegative matrices is a nonnegative matrix, P is a non-
negative row stochastic matrix. To prove part ii), one no-
tices that all eigenvectors of P and L are the same. Let j be where 0 G  G 1= and  is the maximum degree of the
the jth eigenvalue of L. Then, the jth eigenvalue of P is network. Let G be a strongly connected digraph. Then
i) A consensus is asymptotically reached for all
initial states; P
j ¼1 j : (17) ii) PThe group decision value is  ¼ i wi xi ð0Þ with
i wi ¼ 1;
iii) If the digraph is balanced (or P is doubly-
7
The set of mth-order neighbors is the set of neighbors of node i on a stochastic), an average-consensus
P is asymptoti-
graph with adjacency matrix Am . cally reached and  ¼ ð i xi ð0ÞÞ=n.

222Authorized licensed useof

Proceedings limited
theto:IEEE | Vol.De
Universite 95,Sherbrooke.
No. 1, January 2007 on July 16,2023 at 14:01:05 UTC from IEEE Xplore.
Downloaded Restrictions apply.
 Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

Table 1 Continuous-Time versus Discrete-Time Consensus Based on the following lemma, one can readily show that
ðÞ ¼ T  is a valid Lyapunov function for the CT system
that quantifies the collective disagreement in the network.

Theorem 3: (algebraic connectivity of digraphs) Let G be

a balanced digraph (or undirected graph) with Laplacian L
with a symmetric part Ls ¼ ðL þ LT Þ=2 and Perron matrix
P with Ps ¼ ðP þ PT Þ=2. Then,
i) 2 ¼ min1T ¼0 ðT L=T Þ with 2 ¼ 2 ðLs Þ, i.e.
Proof: Considering that xðkÞ ¼ Pk xð0Þ, a consensus is
reached in discrete-time, if the limit limk!1 Pk exists.
According to Lemma 4, this limit exists for primitive T L  2 kk2
matrices. Based on part iv) of Lemma 3, P is a primitive
matrix. Thus, limk!1 xðkÞ ¼ vðwT xð0ÞÞ with v ¼ 1, or for all disagreement vectors ;
xi !  ¼ wT xð0Þ for all
P i as k ! 1. P Hence, the group T T
ii) 2 ¼ max1T ¼0 ð P= Þ with 2 ¼1 2 , i.e.
decision value is  ¼ i wi xi ð0Þ with i wi ¼ 1 (due to
vT w ¼ 1). If the graph is balanced, based on part iii) of
Lemma 3, P is a column stochastic matrix with a left eigen- T P  2 kk
2
vector of w ¼ ð1=nÞ1. The group decision becomes equal to
 ¼ ð1=nÞ1T xi ð0Þ and average-consensus is asymptotically
reached. h for all disagreement vectors .
So far, we have presented a unified framework for Proof: Since G is a balanced digraph, 1T L ¼ 0 and
analysis of convergence of consensus algorithms for di- L1 ¼ 0. This implies that Ls is a valid Laplacian matrix
rected networks with fixed topology in both discrete-time because of Ls 1 ¼ ðL1 þ LT 1Þ=2 ¼ 0. Similarly, Ps is a valid
and continuous-time. A comparison between the two cases Perron matrix which is a nonnegative doubly stochastic
of continuous-time and discrete-time consensus are listed matrix. Part i) follows from a special case of Courant-
in Table 1. Fisher theorem [81] for a symmetric matrix Ls due to

D. Performance of Consensus Algorithms

The speed of reaching a consensus is the key in design T L T Ls 
min ¼ min ¼ 2 ðLs Þ:
of the network topology as well as analysis of performance T
1 ¼0 T  1T ¼0  T 
of a consensus algorithm for a given network. Let us first
focus on balanced directed networks that include
undirected networks as a special case. This is primarily To show part ii), note that for a disagreement vector 
due to the fact that the collective dynamics of the network satisfying 1T  ¼ 0, we have
of agent applying a continuous- or discrete-time consen-
sus algorithm
P in this case has an invariant quantity
T P T P T  T L
 ¼ ð i xi Þ=n. To demonstrate this in discrete-time, note max ¼ max ¼ max
that 1T P ¼ 1T and  T   T   T 
T L
¼ 1  min T ¼ 1 2 ðLs Þ
  
1 1
ðk þ 1Þ ¼ 1T xðk þ 1Þ ¼ ð1T PÞxðkÞ ¼ ðkÞ ¼ 2 ðPs Þ (21)
n n

which implies  is invariant in at iteration k. Let us define Corollary 1: A continuous-time consensus is globally
the disagreement vector [10] exponentially reached with a speed that is faster or equal to
2 ¼ 2 ðLs Þ with Ls ¼ ðL þ LT Þ=2 for a strongly connected
¼x 1 (19) and balanced directed network.
Proof: For CT consensus, we have
P T
and note that i i ¼ 0, or 1  ¼ 0. The consensus
algorithms result in the following disagreement dynamics: _ ¼ 2T L  22 T  ¼ 22 :

CT : _ ¼ LðtÞ
ðtÞ
Therefore, ðÞ ¼ kk2 exponentially vanishes with a
DT : ðk þ 1Þ ¼ PðkÞ: (20) speed that is at least 22 . Since kk ¼ 1=2 , the norm of

Vol.July
Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on 95, 16,2023
No. 1, January 2007UTC
at 14:01:05 | Proceedings
from IEEE Xplore.
ofRestrictions
the IEEE 223
apply.
Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

the disagreement vector exponentially vanishes with a speed dynamics of (22), i.e., x_ ¼ Qx. Then, an alternative form
of at least 2 . h of graph Laplacian is
Recently, Olfati-Saber [47] has shown that quasi-
random small-world networks have extremely large 2
values compared to regular networks with nearest neighbor Q¼I D 1 A: (23)
communication such as the one in Fig. 2(a). For example
for a network, with n ¼ 1000 nodes and uniform degree This is identical to the standard Laplacian of the weighted
di ¼ 10; 8i, the algebraic connectivity of a small-world net- graph G due to L ¼ D A ¼ I D 1 A. The conver-
work can become more than 1500 times of the 2 of a gence analysis of this algorithm is identical to the con-
regular network [47]. sensus algorithm presented earlier. The Perron matrix
According to Theorem 3, 2 is the second largest eigen- associated with Q is in the form P ¼ I L with 0 G  G 1.
value of Ps Vthe symmetric part of the Perron matrix P. In explicit form, this gives the following iterative consensus
The speed of convergence of the iterative consensus algo- algorithm:
rithm is provided in the following result.
 
Corollary 2: A discrete-time consensus is globally xðk þ 1Þ ¼ ð1 ÞI þ D 1 A xðkÞ:
exponentially reached with a speed that is faster or equal
to 2 ¼ 1 2 ðLÞ for a connected undirected network. The aforementioned algorithm for  ¼ 1 takes a rather
Proof: Let ðkÞ ¼ ðkÞT ðkÞ be a candidate Lyapunov simple form xðk þ 1Þ ¼ D 1 AxðkÞ that does not converge
function for the discrete-time disagreement dynamics of for digraphs such as cycles of length n. Therefore, this
ðk þ 1Þ ¼ PðkÞ. For an undirected graph P ¼ PT and all discretization with  ¼ 1 is invalid. Interestingly, the
eigenvalues of P are real. Calculating ðk þ 1Þ, one gets Markov process

ðk þ 1Þ ¼ ðk þ 1ÞT ðk þ 1Þ ðk þ 1Þ ¼ ðkÞP (24)

2 2 2
¼ kPðkÞk  2 kðkÞk
¼ 2 with transition probability matrix P ¼ D 1 A is known as
2 k
the process of random walks on a graph [95] in graph theory
and computer science literature with close connections to
with 0 G 2 G 1 due to the fact that P is primitive. Clearly, gossip-based consensus algorithms [78].
kðkÞk exponentially vanishes with a speed faster or equal Keep in mind that based on algorithm (22), if graph G is
to 2 . h undirected (or balanced), the quantity

Remark 4: The proof of Corollary 2 for balanced di- ! !

X X
graphs is rather detailed and beyond the scope of this ¼ di xi di
paper. i i

E. Alternative Forms of Consensus Algorithms is invariant in time and a weighted-average consensus is

In the context of formation control for a network of P
asymptotically reached. The weighting wi ¼ di =ð i di Þ is
multiple vehicles, Fax and Murray [12] introduced the specified by node degree di ¼ jNi j. Only for regular net-
following version of a Laplacian-based system on a graph G works (i.e., d1 ¼ d2 ¼    ¼ dn ), (22) solves an average-
with 0–1 weights: consensus problem. This is a rather restrictive condition
because most networks are not regular.
1 X Another popular algorithm proposed in [13] (also used
x_ i ¼ ðxj xi Þ: (22) in [14], [15]) is the following discrete-time consensus
jNi j j2Ni
algorithm for undirected networks:

!
This is a special case of a consensus algorithm on a graph 1 X
G with adjacency elements aij ¼ 1=jNi j ¼ 1=di for j 2 Ni xi ðk þ 1Þ ¼ xi ðkÞ þ xj ðkÞ (25)
1 þ jNi j
P zero for j 62 Ni . According to this form, di ¼
and j2Ni

j6¼i aij ¼ 1 for all i that means the degree matrix of G

is D ¼ I and its adjacency matrix is A ¼ D 1 A provided which can be expressed as
that all nodes have nonzero degrees (e.g., for connected
graphs/digraphs). In graph theory literature, A is called
normalized adjacency matrix. Let Q be the key matrix in the xðk þ 1Þ ¼ ðI þ DÞ 1 ðI þ AÞxðkÞ:

224Authorized licensed useof

Proceedings limited
theto:IEEE | Vol.De
Universite 95,Sherbrooke.
No. 1, January 2007 on July 16,2023 at 14:01:05 UTC from IEEE Xplore.
Downloaded Restrictions apply.
 Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

Note that the stochastic Perron matrix P ¼ ðI þ Table 2 Forms of Laplacians

DÞ 1 ðI þ AÞ is obtained from the following normalized
Laplacian matrix with  ¼ 1:

Ql ¼ I ðI þ DÞ 1 ðI þ AÞ: (26)

This Laplacian is a modification of (23) and has the In contrast, Pl does not suffer from this problem because
drawback that it does not solve average-consensus problem of the n nonzero diagonal elements. Again, v is an eigen-
for general undirected networks. vector of Pl associated with the eigenvalue 1, but Pl w ¼
Now, we demonstrate that algorithm (25) is equivalent ðm 1Þ=ðm þ 1Þw and due to ðm 1Þ=ðm þ 1Þ G 1 for
to (23) (and, thus, a special case of (7)). Let G be a graph all m  1, 1 is no longer an eigenvalue of Pl .
with adjacency matrix A and no self-loops, i.e., aii ¼ 0; 8i. Table 2 summarizes three types of graph Laplacians
Then, the new adjacency matrix Al ¼ I þ A corresponds to used in systems and control theory. The alternative
a graph Gl that is obtained from G by adding n self-loops forms of Laplacians in the second and third rows of
with unit weights ðaii ¼ 1; 8iÞ. As a result, the corres- Table 2 are both special cases of L ¼ D A that is widely
ponding degree matrix of Gl is Dl ¼ I þ D. Thus, the nor- used as the standard definition of Laplacian in algebraic
malized Laplacian of Gl in (26) is Ql ¼ I Dl 1 Al . In other graph theory [23].
words, the algorithm proposed by Jadbabaie et al. is The algorithms in all three cases are in two forms
identical to the algorithm of Fax and Murray for a graph
with n self-loops. In both cases  ¼ 1 is used to obtain the
x_ ¼ Lx (28)
stochastic nonnegative matrix P.
xðk þ 1Þ ¼ PxðkÞ: (29)
Remark 5: A undirected cycle is not a counterexample
for discretization of x_ ¼ Ql x with  ¼ 1. Since the Perron
Based on Example 2, the choice of the discrete-time
matrix Pl ¼ ðI þ DÞ 1 ðI þ AÞ is symmetric and primitive.
consensus algorithm is not arbitrary. Only the first and the
third row of Table 2 guarantee stability of a discrete-time
Example 2: In this example, we clarify that why P ¼
linear system for all possible connected networks. The
D 1 A can be an unstable matrix for a connected graph G,
second type requires a further analysis to verify whether P
whereas Pl ¼ ðI þ DÞ 1 ðI þ AÞ remains stable for the same
is stable, or not.
exact graph. for doing so, let us consider a bipartite graph G
with n ¼ 2m nodes and adjacency matrix
F. Weighted-Average Consensus
The choice of the Laplacian for the continuous-time
consensus depends on the specific application of interest.
0m Jm In cases that reaching an average-consensus is desired,
A¼ (27)
Jm 0m only L ¼ D A can be used. In case of weighted-average
consensus with a desired weighting vector ¼ ð 1 ; . . . ; n Þ,
the following algorithms can be used:
where 0m and Jm denote the m  m matrices of zeros and
ones, respectively. Note that D ¼ mIn and P ¼ D 1 A ¼
ð1=mÞA. On the other hand, the Perron matrix of G with n K x_ ¼ Lx (30)
self-loops is
with K ¼ diagð 1 ; . . . ; n Þ and L ¼ D A. This is equiva-
lent to a nodes with a variable rate of integration based on
1 Im Jm
Pl ¼ ðIn þ DÞ 1 ðIn þ AÞ ¼ : the protocol
m þ 1 Jm Im
X
i x_ i ¼ aij ðxj xi Þ:
Let v ¼ 12m be the vector of ones with 2m elements and j2Ni
w ¼ colð1m ; 1m Þ. Both v and w are eigenvectors of P
associated with eigenvalues 1 and 1, respectively, due to
Pv ¼ v and Pw ¼ w. This proves that P is not a primitive In the special case the weighting is proportional to the
matrix and the limit limk!1 Pk does not exist (since P has node degrees, or K ¼ D, one obtains the second type of
two eigenvalues with modulus 1). Laplacian in Table 2, or x_ ¼ D 1 Lx ¼ ðI D 1 AÞx.

Vol.July
Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on 95, 16,2023
No. 1, January 2007UTC
at 14:01:05 | Proceedings
from IEEE Xplore.
ofRestrictions
the IEEE 225
apply.
Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

G. Consensus Under Communication Time-Delays delays. Networks with hubs (having very large degrees)
Suppose that agent i receives a message sent by its that are commonly known as scale-free networks [96] are
neighbor j after a time-delay of . This is equivalent to a fragile to time-delays. In contrast, random graphs [97] and
network with a uniform one-hop communication time- small-world networks [47], [89] are fairly robust to time-
delay. The following consensus algorithm: delays since they do not have hubs. In conclusion, con-
struction of engineering networks with nodes that have
X   high degrees is not a good idea for reaching a consensus.
x_ i ðtÞ ¼ aij xj ðt Þ xi ðt Þ (31)
j2Ni
I II . CONSENSUS I N S WITCHING
NETWORKS
was proposed in [10] to reach an average-consensus for In many scenarios, networked systems can possess a
undirected graphs G. dynamic topology that is time-varying due to node and link
failures/creations, packet-loss [40], [98], asynchronous
Remark 6: Keep in mind that the algorithm consensus [41], state-dependence [64], formation recon-
figuration [53], evolution [96], and flocking [19], [99].
X   Networked systems with a dynamic topology are
x_ i ðtÞ ¼ aij xj ðt Þ xi ðtÞ (32)
j2Ni
commonly known as switching networks. A switching
network can be modeled using a dynamic graph GsðtÞ
parameterized with a switching signal sðtÞ : R ! J that
P
does not preserve the average xðtÞ ¼ ð1=nÞ i xi ðtÞ in time takes its values in an index set J ¼ f1; . . . ; mg. The
for a general graph. The same is true when the graph in consensus mechanism on a network with a variable
(31) is a general digraph. It turns out that for balanced topology becomes a linear switching system
digraphs with 0–1 weights, xðtÞ is an invariant quantity
along the solutions of (31).
The collective dynamics of the network can be x_ ¼ LðGk Þx; (34)
expressed as

with the topology index k ¼ sðtÞ 2 J and a Laplacian of the

_ ¼
xðtÞ Lxðt Þ: type D A. The set of topologies of the network is  ¼
fG1 ; G2 ; . . . ; Gm g. First, we assume at any time instance,
the network topology is a balanced digraph (or undirected
Rewriting this equation after taking Laplace transform of graph) that is strongly connected. Let us denote 2 ððL þ
both sides, we get LT Þ=2Þ by 2 ðGk Þ for a topology dependent Laplacian L ¼
LðGk Þ. The following result provides the analysis of
HðsÞ average-consensus for dynamic networks with a perfor-
XðsÞ ¼ xð0Þ (33) mance guarantee.
s
Theorem 5: (Olfati-Saber and Murray, 2004) Consider a
with a proper MIMO transfer function HðsÞ ¼ ðIn þ network of agents applying the consensus algorithm in
ð1=sÞ expð sÞLÞ 1 . One can use Nyquist criterion to (34) with topologies Gk 2 . Suppose every graph in  is a
verify the stability of HðsÞ. A similar criterion for stability balanced digraph that is strongly connected and let
of formations was introduced by Fax and Murray [12]. The 2 ¼ mink2J 2 ðGk Þ. Then, for any arbitrary switching sig-
following theorem provides an upper bound on the time- nal, the agents asymptotically reach an average-consensus
delay such that stability of the network dynamics is main- for all initial states with a speed faster or equal to 2 .
tained in presence of time-delays. Moreover, ðÞ ¼ T  is a common Lyapunov function for
the collective dynamics of the network.
Theorem 4: (Olfati-Saber and Murray, 2004) The Proof: See the proof of Theorem 9 in [10]. h
algorithm in (31) asymptotically solves the average- Note that  is a finite set with at most nðn 1Þ
consensus problem with a uniform one-hop time-delay  elements and this allows the definition of 2 . Moreover,
for all initial states if and only if 0   G =2n . the use of Pnormal Laplacians does not render the average
Proof: See the proof of Theorem 10 in [10]. h x ¼ ð1=nÞ i xi invariant in time, unless all graphs in  are
Since n G 2, a sufficient condition for conver- d-regular (all of their nodes have degree d). This is hardly
gence of the average-consensus algorithm in (31) is that the case for various applications.
 G =4. In other words, there is a trade-off between The following result on consensus for switching
having a large maximum degree and robustness to time- networks does not require the necessity for connectivity

226Authorized licensed useof

Proceedings limited
theto:IEEE | Vol.De
Universite 95,Sherbrooke.
No. 1, January 2007 on July 16,2023 at 14:01:05 UTC from IEEE Xplore.
Downloaded Restrictions apply.
 Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

in all time instances and is due to Jadbabaie et al. [13]. This Thus, an alignment is asymptotically reached and the
weaker form of network connectivity is crucial in analysis group decision is an undetermined quantity in the convex
of asynchronous consensus with performance guarantees hull of all initial states.
(which is currently an open problem). We need to re-
phrase the next result for the purpose of compatibility with Remark 7: Since normal Perron matrices in the form
the notation used in this paper. ðI þ DÞ 1 ðI þ AÞ are employed in [13], the agents (in
Consider the following discrete-time consensus general) do not reach an average-consensus. The use of
algorithm: Perron matrices in the form I L with 0 G  G 1=
ð1 þ maxk2J ðGk ÞÞ resolves this problem.
Recently, an extension of Theorem 6 with connectivity
xkþ1 ¼ Psk xk ; t ¼ 0; 1; 2; . . . (35) of the union of graphs over an infinite interval has been
introduced by Moreau [14] (also, an extension is presented
in [15] for weighted graphs). Here, we rephrase a theorem
with sk 2 J. Let P ¼ fP1 ; . . . ; Pm g denote the set of Perron due to Moreau and present it based on notation. First, let
matrices associated with a finite set of undirected graphs  us define a less restrictive notion of connectivity of
with n self-loops. We say a switching network with the set switching networks compared to periodic connectivity. Let
of topologies  is periodically connected with a period N 9 1  be a finite set of undirected graphs with n self-loops. We
if the unions of all graphs over a sequence of intervals say a switching networks with topologies in  is ultimately
½j; jNÞ for j ¼ 0; 1; 2; . . . are connected graphs, i.e., Gj ¼ connected if there exists an initial time k0 such that over
jN 1
[k¼j Gsk is connected for j ¼ 0; 1; 2; . . .. the infinite interval ½k0 ; 1Þ the graph G ¼ [1 k¼k0 Gsk with
sk 2 J is connected.
Theorem 6: (Jadbabaie, Lin, and Morse, 2003) Consider
the system in (35) with Psk 2 P for k ¼ 0; 1; 2; . . .. Assume Theorem 7: (Moreau, 2005) Consider an ultimately
the switching network is periodically connected. Then, connected switching network with undirected topologies
limk!1 xk ¼ 1, or an alignment is asymptotically in  and dynamics (35). Assume Psk 2 P where P is the set
reached. of normal Perron matrices associated with . Then, a
Proof: See the proof of Theorem 2 in [13]. h consensus is globally asymptotically reached.
The solution of (35) can be explicitly expressed as Proof: See the proof of Proposition 2 in [14]. h
Similarly, the algorithm analyzed in Proposition 2 of
! [14] does not solve the f -consensus problem. This can be
Y
t
resolved by using the first form of Perron matrices in
xt ¼ Psk x0 ¼ t x0 Table 2. The proof in [14] uses a nonquadratic Lyapunov
k¼0
function and no performance measures for reaching a
consensus is presented.
with t ¼ Pst    Ps2 Ps1 . the convergence of the consensus
algorithm in (35) depends on whether the infinite product IV. COOPERATION I N NETWORKED
of nonnegative stochastic matrices Pst    Ps2 Ps1 has a limit. CONTROL SYSTEMS
The problem of convergence of infinite product of
stochastic matrices has a long history and has been studied This section provides a system-theoretic framework for
by several mathematicians including Wolfowitz [100]. The addressing the problem of cooperative control of net-
proof in [13] relies on Wolfowitz’s lemma: worked multivehicle systems using distributed controllers.
On one hand, a multivehicle system represents a col-
Lemma 5: (Wolfowitz, 1963) Let P ¼ fP1 ; P2 ; . . . ; Pm g lection of decision-making agents that each have limited
be a finite set of primitive stochastic matrices such that for knowledge of both the environment and the state of the
any sequence of matrices Ps1 ; Ps2 ; . . . ; Psk 2 P with k  1, other agents. On the other hand, the vehicles can influ-
the product Psk    Ps2 Ps1 is a primitive matrix. Then, there ence their own state and interact with their environment
exists a row vector w such that according to their dynamics which determines their
behavior.
The design goal is to execute tasks cooperatively
exercising both the decision-making and control capabilities
lim Psk    Ps2 Ps1 ¼ 1w: (36)
k!1 of the vehicles. In real-life networked multivehicle sys-
tems, there are a number of limitations including limited
sensing capabilities of the vehicles, network bandwidth
According to Wolfowitz’s lemma, we get limk!1 xk ¼ limitations, as well as interruptions in communications
1ðwx0 Þ ¼ 1 with  ¼ wx0 . The vector w depends on the due to packet-loss [40], [98] and physical disruptions to
switching sequence and cannot be determined a priori. the communication devices of the vehicle.

Vol.July
Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on 95,16,2023
No. 1, January 2007UTC
at 14:01:05 | Proceedings
from IEEE Xplore.
ofRestrictions
the IEEE 227
apply.
Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

other vehicles. We assume that Ni 6¼ ;, meaning that each

vehicle can sense at least one other vehicle. Note that a
single vehicle cannot drive all the zij terms to zero simul-
taneously; the errors must be fused into a single signal
error measurement

1 X
zi ¼ zij (41)
jNi j j2Ni

Fig. 3. The block diagram of cooperative and distributed where jNi j is the cardinality of the set Ni . We also
formation control of networked multivehicle systems. define a distributed controller K which maps yi , zi to ui
The Kronecker product  is defined in (37).
and has internal states vi 2 Rs , represented in state-space
form by
The system framework we analyze is presented in a
schematic form in Fig. 3. The Kronecker product  between v_i ¼ Fvi þ G1 yi þ G2 zi
two matrices P ¼ ½pij and Q ¼ ½qij is defined as ui ¼ Hvi þ D1 yi þ D2 zi : (42)

P  Q ¼ ½pij Q : (37)
Now, we consider the collective system of all n vehicles.
For dimensional compatibility, we use the Kronecker
This is a block matrix with the ijth block of pij Q. product to assemble the matrices governing the formation
The dynamics of each vehicle, represented by PðsÞ, is behavior. The collective dynamics of n vehicles can be
decoupled from the dynamics of other vehicles in the represented as follows:
networkVthus, the system transfer function In  PðsÞ. The
output of PðsÞ represents observable elements of the state of     
x_ M11 M12 x
each vehicle. Similarly, the controller of each vehicle, ¼ : (43)
v_ M21 M22 v
represented by KðsÞ, is decoupled from the controller of
othersVthus, the controller transfer function In  KðsÞ.
The coupling occurs through cooperation via the consensus where the Mij ’s are block matrices defined as a function of
feedback. Since all vehicles apply the same controller, they the normalized graph Laplacian L (i.e., the second type in
form a cooperative team of vehicles with consensus feed- Table 2) and other matrices as follows:
back gain matrix L  Im . This cooperation requires sharing
of information among vehicles, either through interagent
sensing, or explicit communication of information. M11 ¼ In  ðA þ BD1 C1 Þ þ ðIn  BD2 C2 ÞðL  Im Þ
M12 ¼ In  BH
A. Collective Dynamics of Multivehicle Formations M21 ¼ In  G1 C1 þ ðIn  G2 C2 ÞðL  Im Þ;
Let us consider a group of n vehicles, whose (identical)
linear dynamics are denoted by M22 ¼ In  F:

x_ i ¼ Axi þ Bui (38) B. Stability of Relative Dynamics of Formations

The main stability result on relative-position-based
formations of networked vehicles is due to Fax and Murray
where xi 2 Rm , ui 2 Rp are the vehicle states and controls, [12] and can be stated as follows:
and i 2 V ¼ f1; . . . ; ng is the index for the vehicles in the
group. Each vehicle receives the following measurements: Theorem 8: (Fax and Murray, 2004) A local controller K
stabilizes the formation dynamics in (43) if and only if it
stabilizes all the n systems
yi ¼ C1 xi (39)
zij ¼ C2 ðxi xj Þ; j 2 Ni (40)
x_ i ¼ Axi þ Bui
yi ¼ C1 xi
Thus, yi 2 Rk represents internal state measurements, and
zij 2 Rl represents external state measurements relative to zi ¼ i C2 xi (44)

228Authorized licensed useof

Proceedings limited
theto:IEEE | Vol.De
Universite 95,Sherbrooke.
No. 1, January 2007 on July 16,2023 at 14:01:05 UTC from IEEE Xplore.
Downloaded Restrictions apply.
 Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

where fi gni¼1 is the set of eigenvalues of the normalized is scaled by the eigenvalues of the (normalized) Laplacian
graph Laplacian L. of the network. Note that i may be complex, leading to a
Theorem 8 reveals that the stability of a formation of n complex-valued LTI system in the above formulation. This
identical vehicles can be verified by stability analysis of a formalism lends itself to applications of tools from robust
single vehicle with the same dynamics and an output that control theory [101].

Fig. 4. (a) A small-world with 300 links, (b) a regular lattice with interconnections to k ¼ 3 nearest neighbors and 300 links,
(c) a regular lattice with interconnections to k ¼ 10 nearest neighbors and 1000 links; (d), (e), (f) the state evolution corresponding to
networks in (a), (b), and (c), respectively. [Note: only the links of a single node are depicted in parts (b) and (c).]

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded Vol. 95, 16,2023
on July No. 1, January 2007UTC
at 14:01:05 | Proceedings
from IEEE Xplore. the IEEEapply.229
ofRestrictions
Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

Fig. 5. (a) Interconnection graph of a multivehicle formation and (b) the Nyquist plot.

The zero eigenvalue of L can be interpreted as the A. Consensus in Complex Networks

unobservability of absolute motion of the formation in the In this experiment, we demonstrate the speed of
measurements zi . A prudent design strategy is to close an convergence of consensus algorithm (7) for three different
inner loop around yi such that the internal vehicle dyna- networks with n ¼ 100 nodes in Fig. 4. The initial state is
mics are stable, and then to close an outer loop around zi set to xi ð0Þ ¼ i for i ¼ 1; . . . ; 100. In Fig. 4(a) and (c), the
which achieves desired formation performance. For the network has 300 links and on average each node
remainder of this section, we concern ourselves solely with communicates with d ¼ 6 neighbors. Apparently, the
the outer loop. Hence, we assume from now on that C1 is group with a small-world network topology reaches an
empty and that A has no eigenvalues in the open right half average-consensus more than 2 ðGa Þ=2 ðGc Þ  22 times
plane. We do not wish to exclude eigenvalues along the j! faster. To create a regular lattice with comparable alge-
axis because they are characteristic of vehicle systems, braic connectivity, every node has to communicate with 20
representing the directions in which motion is possible. other nodes on average to gain an algebraic connectivity
The controller K is also presumed to be stable. If K stabi- 2 ðGe Þ=2 ðGa Þ  1:2 that is close to that of the small-
lizes the system in (44) for all i other than the zero world network. Of course, the regular network in Fig. 4(e)
eigenvalue, we say that it stabilizes the relative dynamics of has 3.33 times as many links as the small-world network.
a formation. For further information on small-world networks, we refer
Let us refer to the system from ui to yi as P, its transfer the reader to [47], [89], and [102].
function as PðsÞ, and that of the controller from yi to ui as
KðsÞ. For single-input single-output (SISO) systems, we B. Multivehicle Formation Control
can state a second version of Theorem 8 which is useful for Consider a system of the form PðsÞ ¼ e sT =s2 , model-
stability and robustness analysis. ing a second-order system with time-delay and suppose
this system has been stabilized with a proportional-
Theorem 9: (Fax and Murray, 2004) Suppose P is a derivative (PD) controller. Fig. 5 shows a formation graph
SISO system. Then K stabilizes the relative dynamics of and the Nyquist plot of KðsÞPðsÞ with the location of
a formation if and only if the net encirclement of 1=i Laplacian eigenvalues. The Bo[ locations correspond to the
by the Nyquist plot of KðsÞPðsÞ is zero for all non- eigenvalues of the graph defined by the solid arcs in Fig. 5,
zero i . and the F_ locations are for eigenvalues of the graph
The application of the above theorem is demonstrated when the dashed arc is included as well. This example
in Section V-B. clearly shows the effect the formation has on stability
margins. The standard Nyquist plot reveals a system with
reasonable stability marginsVabout 8 dB and 45 . When
V. SIMULATIONS one accounts for the effects of the formation, however, one
In this section, we present the simulation results for sees that for the Bo[ formation, the stability margins are
three applications of consensus problems in networked substantially degraded, and for the B[ formation, the
systems. system is in fact unstable. Interestingly, the formation is

230Authorized licensed useof

Proceedings limited
theto:IEEE | Vol.De
Universite 95,Sherbrooke.
No. 1, January 2007 on July 16,2023 at 14:01:05 UTC from IEEE Xplore.
Downloaded Restrictions apply.
 Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

rendered unstable when additional information (its nonnegative stochastic matrices, and algebraic connectivity
position relative to vehicle 6) is used by vehicle 1. This is of graphs and digraphs play an instrumental role in analysis
primarily due to the fact that changing the topology of a of consensus algorithms. We proved that algorithms
network directly effects the location of eigenvalues of the introduced by Jadbabaie et al. and Fax and Murray are
Laplacian matrix. This example clarifies that the stability identical for graphs with n self-loops and are both special
analysis of formations of networked vehicles with directed cases of the consensus algorithm of Olfati-Saber and
switching topology in presence of time-delays is by no Murray. The notion of Perron matrices was introduced as
means trivial. the discrete-time counterpart of graph Laplacians in con-
sensus protocols. A number of fundamental spectral pro-
perties of Perron matrices were proved. This led to a unified
VI. CONCL US ION framework for expression and analysis of consensus
A theoretical framework was provided for analysis of algorithms in both continuous-time and discrete-time.
consensus algorithms for networked multi-agent systems Simulation results for reaching a consensus in small-worlds
with fixed or dynamic topology and directed information versus lattice-type nearest-neighbor graphs and cooperative
flow. The connections between consensus problems and control of multivehicle formations were presented. h
several applications were discussed that include synchro-
nization of coupled oscillators, flocking, formation control,
fast consensus in small-world networks, Markov processes Acknowledgment
and gossip-based algorithms, load balancing in networks, R. Olfati-Saber would like to thank Richard Murray for
rendezvous in space, distributed sensor fusion in sensor giving him the opportunity to teach a significant portion of
networks, and belief propagation. The role of Bcoop- consensus theory in this paper at California Institute of
eration[ in distributed coordination of networked auton- Technology in the Fall of 2002. The authors would also like
omous systems was clarified and the effects of lack of to thank the anonymous reviewers and the associate
cooperation was demonstrated by an example. It was editors for their helpful remarks that improved the
demonstrated that notions such as graph Laplacians, presentation of the paper.

REFERENCES IEEE Trans. Autom. Control, vol. 49, no. 9, IEEE Trans. Autom. Control, vol. 51, no. 3,
pp. 1520–1533, Sep. 2004. pp. 401–420, Mar. 2006.
[1] N. A. Lynch, Distributed Algorithms.
San Francisco, CA: Morgan Kaufmann, [11] J. A. Fax, BOptimal and cooperative control [20] M. Fiedler, BAlgebraic connectivity of
1997. of vehicle formations,[ Ph.D. dissertation, graphs,[ Czechoslovak Math. J., vol. 23,
Control Dynamical Syst., California Inst. no. 98, pp. 298–305, 1973.
[2] M. H. DeGroot, BReaching a consensus,[ Technol., Pasadena, CA, 2001.
J. Am. Statist. Assoc., vol. 69, no. 345, [21] B. Mohar, The Laplacian spectrum of
pp. 118–121, 1974. [12] J. A. Fax and R. M. Murray, BInformation graphs,[ in Graph Theory, Combinatorics,
flow and cooperative control of vehicle and Applications, Y. Alavi, G. Chartrand,
[3] J. A. Benediktsson and P. H. Swain, formations,[ IEEE Trans. Autom. Control, O. Ollermann, and A. Schwenk, Eds.
BConsensus theoretic classification vol. 49, no. 9, pp. 1465–1476, Sep. 2004. New York: Wiley, 1991, pp. 871–898.
methods,[ IEEE Trans. Sys., Man, Cybern.,
vol. 22, no. 4, pp. 688–704, Apr. 1992. [13] A. Jadbabaie, J. Lin, and A. S. Morse, [22] R. Merris, BLaplacian matrices of a graph:
BCoordination of groups of mobile A survey,[ Linear Algebra its Appl., vol. 197,
[4] S. C. Weller and N. C. Mann, BAssessing autonomous agents using nearest neighbor pp. 143–176, 1994.
rater performance without a Fgold standard_ rules,[ IEEE Trans. Autom. Control, vol. 48,
using consensus theory,[ Med. Decision [23] C. Godsil and G. Royle, Algebraic Graph
no. 6, pp. 988–1001, Jun. 2003. Theory, ser. Graduate Texts in Mathematics.
Making, vol. 17, no. 1, pp. 71–79, 1997.
[14] L. Moreau, BStability of multi-agent systems Berlin, Germany: Springer-Verlag, 2001,
[5] V. Borkar and P. Varaiya, BAsymptotic with time-dependent communication links,[ vol. 207.
agreement in distributed estimation,[ IEEE IEEE Trans. Autom. Control, vol. 50, no. 2,
Trans. Autom. Control, vol. AC-27, no. 3, [24] M. Sipser and D. A. Spielman, BExpander
pp. 169–182, Feb. 2005. codes,[ IEEE Trans. Inf. Theory, vol. 42, no. 6,
pp. 650–655, Jun. 1982.
[15] W. Ren and R. W. Beard, BConsensus pp. 1710–1772, Nov. 1996.
[6] J. N. Tsitsiklis, BProblems in decentralized seeking in multi-agent systems under
decision making and computation,[ [25] A. Lubotzky, R. Phillips, and P. Sarnak,
dynamically changing interaction BRamanujan graphs,[ Combinatorica, vol. 8,
Ph.D. dissertation, Dept. Electr. Eng. topologies,[ IEEE Trans. Autom.
Comput. Sci., Lab. Inf. Decision Syst., no. 3, pp. 261–277, 1988.
Control, vol. 50, no. 5, pp. 655–661,
Massachusetts Inst. Technol., Cambridge, [26] Z. Lin, M. Brouke, and B. Francis, BLocal
May 2005.
MA, Nov. 1984. control strategies for groups of mobile
[16] T. Vicsek, A. Cziroók, E. Ben-Jacob, autonomous agents,[ IEEE Trans. Autom.
[7] J. N. Tsitsiklis, D. P. Bertsekas, and I. Cohen, and O. Shochet, BNovel type
M. Athans, BDistributed asynchronous Control, vol. 49, no. 4, pp. 622–629,
of phase transition in a system of self-deriven Apr. 2004.
deterministic and stochastic gradient particles,[ Phys. Rev. Lett., vol. 75, no. 6,
optimization algorithms,[ IEEE [27] V. Blondel, J. M. Hendrickx, A. Olshevsky,
pp. 1226–1229, 1995.
Trans. Automatic Control, vol. 31, no. 9, and J. N. Tsitsiklis, BConvergence in
pp. 803–812, Sep. 1986. [17] H. Abelson, D. Allen, D. Coore, C. Hanson, multi-agent coordination, consensus, and
G. Homsy, J. Knight, T. F. R. Nagpal, flocking,[ in Proc. 44th IEEE Conf. Decision
[8] D. P. Bertsekas and J. Tsitsiklis, Parallell and E. Rauch, G. J. Sussman, and R. Weiss,
Distributed Computation. Upper Saddle and Control, 2005 and 2005 Eur. Control Conf.
BAmorphous computing,[ Commun. ACM, (CDC-ECC ’05), Dec. 2005, pp. 2996–3000.
River, NJ: Prentice-Hall, 1989. vol. 43, no. 5, pp. 74–82, 2000.
[9] R. O. Saber and R. M. Murray, BConsensus [28] D. Bauso, L. Giarré, and R. Pesenti,
[18] R. Nagpal, BProgrammable self-assembly BNonlinear protocols for optimal distributed
protocols for networks of dynamic agents,[ using biologically-inpired multi-agent
in Proc. 2003 Am. Control Conf., 2003, consensus in networks of dynamic agents,[
control,[ in Proc. 1st Int. Conf. Autonomous Systems and Control Letters, vol. 55, no. 11,
pp. 951–956. Agents and Multi-Agent Systems, 2002, pp. 918–928, 2006.
[10] R. Olfati-Saber and R. M. Murray, pp. 418–425.
BConsensus problems in networks of agents [29] J. Cortés, BAchieving coordination tasks
[19] R. Olfati-Saber, BFlocking for multi-agent in finite time via nonsmooth gradient
with switching topology and time-delays,[ dynamic systems: Algorithms and theory,[

Vol.July
Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on 95,16,2023
No. 1, January 2007UTC
at 14:01:05 | Proceedings
from IEEE Xplore.
ofRestrictions
the IEEE apply.231
Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

flows,[ in Proc. 44th IEEE Conf. Decision and [44] A. Jadbabaie, N. Motee, and M. Barahona, behavior,[ in 44th IEEE Conf. Decision and
Control, 2005 and 2005 Eur. Control Conf. BOn the stability of the Kuramoto model of Control, 2005 and 2005 Eur. Control Conf.
(CDC-ECC ’05), Dec. 2005, pp. 6376–6381. coupled nonlinear oscillators,[ in Proc. 2004 (CDC-ECC ’05), Dec. 2005, pp. 84–89.
[30] M. Mehyar, D. Spanos, J. Pongsjapan, Am. Control Conf., vol. 5, pp. 4296–4301, Jun. [60] R. O. Saber, W. B. Dunbar, and
S. H. Low, and R. M. Murray, BDistributed 2004. R. M. Murray, BCooperative control of
averaging on asynchronous communication [45] A. Papachristodoulou and A. Jadbabaie, multivehicle systems using cost graphs and
networks,[ in Proc. 44th IEEE Conf. Decision BSynchronization in oscillator networks: optimization,[ in Proc. 2003 Am. Control
and Control, 2005 and 2005 Eur. Control Conf. Switching topologies and non-homogeneous Conf., Jun. 2003, pp. 2217–2222.
(CDC-ECC ’05), Dec. 2005, pp. 7446–7451. delays,[ in 44th IEEE Conf. Decision and [61] T. Keviczky, G. J. Borelli, and F. Balas,
[31] P.-A. Bliman and G. Ferrari-Trecate, Control, 2005 and 2005 Eur. Control Conf. BA study on decentralized receding
BAverage consensus problems in networks (CDC-ECC ’05), Dec. 2005, pp. 5692–5697. horizon control for decoupled systems,[ in
of agents with delayed communications,[ in [46] N. Chopra and M. W. Spong, Proc. 2004 American Control Conf., 2004,
Proc. 44th IEEE Conf. Decision and Control, BOn synchronization of Kuramoto pp. 4921–4926.
2005 and 2005 Eur. Control Conf. (CDC-ECC oscillators,[ in 44th IEEE Conf. Decision [62] J. Hu, M. Prandini, and C. Tomlin,
’05), Dec. 2005, pp. 7066–7071. and Control, 2005 and 2005 Eur. Control Conf. BInteresting conjugate points in formation
[32] J. Sandhu, M. Mesbahi, and T. Tsukamaki, (CDC-ECC ’05), Dec. 2005, pp. 3916–3922. constrained multi-agent coordination,[ in
BRelative sensing networks: Observability, [47] R. Olfati-Saber, BUltrafast consensus in Proc. 2005 Am. Control Conf., Jun. 2005,
estimation, and the control structure,[ in small-world networks,[ in Proc. 2005 Am. pp. 1871–1876.
Proc. 44th IEEE Conf. Decision and Control Conf., Jun. 2005, pp. 2371–2378. [63] M. Alighanbari and J. P. How,
Control, 2005 and 2005 Eur. Control Conf. [48] M. Cao, D. A. Spielman, and A. S. Morse, BDecentralized task assignment for
(CDC-ECC ’05), Dec. 2005, pp. 6400–6405. BA lower bound on convergence of a unmanned aerial vehicles,[ in 44th IEEE
[33] A. V. Savkin, BCoordinated collective motion distributed network consensus algorithm,[ Conf. Decision and Control, 2005 and 2005 Eur.
of groups of autonomous mobile robots: in 44th IEEE Conf. Decision and Control, 2005 Control Conf. (CDC-ECC ’05), Dec. 2005,
Analysis of Vicsek’s model,[ IEEE Trans. and 2005 Eur. Control Conf. (CDC-ECC ’05), pp. 5668–5673.
Autom. Control, vol. 49, no. 6, pp. 981–982, Dec. 2005, pp. 2356–2361. [64] M. Mesbahi, BOn state-dependent dynamic
Jun. 2004. [49] Y. Kim and M. Mesbahi, BOn maximizing graphs and their controllability properties,[
[34] N. Moshtagh, A. Jadbabaie, and K. Daniilidis, the second smallest eigenvalue of IEEE Trans. Autom. Control, vol. 50, no. 3,
BDistributed geodesic control laws for state-dependent graph Laplacian,[ pp. 387–392, Mar. 2005.
flocking of nonholonomic agents,[ in Proc. IEEE Trans. Autom. Control, vol. 51, no. 1, [65] A. Muhammad and M. Egerstedt,
44th IEEE Conf. Decision and Control, 2005 pp. 116–120, Jan. 2006. BConnectivity graphs as models of local
and 2005 Eur. Control Conf. (CDC-ECC ’05), [50] L. Fang, P. J. Antsaklis, and A. Tzimas, interactions,[ J. Appl. Math. Computation,
Dec. 2005, pp. 2835–2841. BAsynchronous consensus protocols: 2004, (preprint, submitted).
[35] W. Xi, X. Tan, and J. S. Baras, BA stochastic Preliminary results, simulations and open [66] M. Ji and M. Egerstedt, BConnectedness
algorithm for self-organization of questions,[ in 44th IEEE Conf. Decision and preserving distributed coordination control
autonomous swarms,[ in Proc. 44th IEEE Control, 2005 and 2005 Eur. Control Conf. over dynamic graphs,[ in Proc. 2005 Am.
Conf. Decision and Control, 2005 and 2005 Eur. (CDC-ECC ’05), Dec. 2005, pp. 2194–2199. Control Conf., Jun. 2005, pp. 93–98.
Control Conf. (CDC-ECC ’05), Dec. 2005, [51] M. Egerstedt and X. Hu, BFormation [67] M. M. Zavlanos and G. J. Pappas,
pp. 765–770. control with virtual leaders and reduced BControlling connectivity of dynamic
[36] R. A. Freeman, P. Yang, and K. M. Lynch, communications,[ IEEE Trans. Robot. graphs,[ in 44th IEEE Conf. Decision and
BDistributed estimation and control of Autom., vol. 17, no. 6, pp. 947–951, 2001. Control, 2005 and 2005 Eur. Control Conf.
swarm formation statistics,[ in Proc. 2006 [52] R. Olfati-Saber and R. M. Murray, (CDC-ECC ’05), Dec. 2005, pp. 6388–6393.
Am. Control Conf., pp. 749–755, Minneapolis, BDistributed cooperative control of multiple [68] D. Hristu and K. Morgansen, BLimited
MN, Jun. 2006. vehicle formations using structural potential communication control,[ Syst. Control Lett.,
[37] R. Olfati-Saber, BSwarms on sphere: functions,[ presented at the 15th IFAC vol. 37, pp. 193–205, Jul. 1999.
A programmable swarm with synchronous World Congr., Barcelona, Spain, Jul. 2002. [69] E. Klavins, Communication complexity
behaviors like oscillator networks,[ in [53] VV, BGraph rigidity and distributed of multirobot systems,[ in Algorithmic
Proc. 45th IEEE Conf. Decision and Control, formation stabilization of multivehicle Foundations of Robotics V, ser. Springer Tracts
pp. 5060–5066, San Diego, CA, Dec. 2006. systems,[ in Proc. 41st IEEE Conf. Decision in Advanced Robotics, J.-D. Boissonnats,
[38] VV, BDistributed Kalman filter with and Control, 2002, pp. 2965–2971. J. Burdik, K. Goldberg, and
embedded consensus filter,[ in 44th IEEE [54] T. Eren, W. Whiteley, A. S. Morse, S. Huchinson, Eds. Budapest: Springer,
Conf. Decision and Control, 2005 and 2005 Eur. P. N. Belhumeur, and B. D. O. Anderson, 2003, vol. 7, pp. 275–292.
Control Conf. (CDC-ECC ’05), Dec. 2005, BSensor and network topologies of [70] Martı́nez, F. Bullo, J. Cortés, and E. Frazzoli,
pp. 8179–8184. formations with direction, bearing BOn synchronous robotic networks Part I:
[39] R. Olfati-Saber and J. S. Shamma, and angle information between Models, tasks, and complexity notions,[ in
BConsensus filters for sensor networks agents,[ in Proc. 42nd IEEE Conf. Decision and Proc. 44th IEEE Conf. Decision and Control,
and distributed sensor fusion,[ in 44th Control, Dec. 2003, pp. 3064–3069. 2005 and 2005 Eur. Control Conf.
IEEE Conf. Decision and Control, 2005 [55] R. Teo, D. Stipanović, and C. J. Tomlin, (CDC-ECC ’05), Dec. 2005, pp. 2847–2852.
and 2005 Eur. Control Conf. (CDC-ECC ’05), BDecentralize spacing control of a string of [71] VV, BOn synchronous robotic networks
Dec. 2005, pp. 6698–6703. multiple vehicles over lossy datalinks,[ in Part II: Time complexity of the rendezvous
[40] V. Gupta, V. Hassibi, and R. M. Murray, Proc. 42nd IEEE Conf. Decision and Control, and deployment algorithms,[ in Proc. 44th
BOn sensor fusion in the presence of Dec. 2003, pp. 682–687. IEEE Conf. Decision and Control, 2005 and
packet-dropping communication channels,[ [56] H. G. Tanner, G. J. Pappas, and V. Kumar, 2005 Eur. Control Conf. (CDC-ECC ’05),
in 44th IEEE Conf. Decision and Control, 2005 BLeader-to-formation stability,[ IEEE Trans. Dec. 2005, pp. 8313–8318.
and 2005 Eur. Control Conf. (CDC-ECC ’05), Robot. Autom., vol. 20, no. 3, pp. 443–455, [72] V. Saligrama, M. Alanyali, and O. Savas,
Dec. 2005, pp. 3547–3552. Jun. 2004. Asynchronous distributed detection in sensor
[41] Y. Hatano and M. Mesbahi, BAgreement over [57] Z. Lin, B. Francis, and M. Maggiore, networks, 2005, Preprint.
random networks,[ IEEE Trans. Autom. BNecessary and sufficient graphical [73] R. Olfati-Saber, E. Franco, E. Frazzoli, and
Control, vol. 50, no. 11, pp. 1867–1872, conditions for formation control of J. S. Shamma, BBelief consensus and
Nov. 2005. unicyles,[ IEEE Trans. Autom. Control, distributed hypothesis testing in sensor
[42] V. M. Preciado and G. C. Verghese, vol. 50, no. 1, pp. 121–127, Jan. 2005. networks,[ presented at the Workshop on
BSynchronization in generalized Erd [58] X. Xi and E. H. Abed, BFormation Network Embedded Sensing and Control,
os-Rénye networks of nonlinear oscillators,[ control with virtual leaders and reduced Notre Dame, IN, Oct. 2005.
in 44th IEEE Conf. Decision and Control, 2005 communications,[ in 44th IEEE Conf. [74] D. P. Spanos, R. Olfati-Saber, and
and 2005 Eur. Control Conf. (CDC-ECC ’05), Decision and Control, 2005 and 2005 Eur. R. M. Murray, BApproximate distributed
Dec. 2005, pp. 4628–4633. Control Conf. (CDC-ECC ’05), Dec. 2005, Kalman filtering in sensor networks with
[43] R. Sepulchre, D. Paley, and N. Leonard, pp. 1854–1860. quantifiable performance,[ in Proc. 4th
BCollective motion and oscillator [59] D. V. Dimarogonas and K. J. Kyriakopoulos, Int. Symp. Information Processing in Sensor
synchronization,[ in Proc. Block Island BFormation control and collision avoidance Networks, Apr. 2005, pp. 133–139.
Workshop Cooperative Control, Block Island, for multi-agent systems and a connection [75] L. Xiao, S. Boyd, and S. Lall, BA scheme
RI, Jun. 2003. between formation infeasibility and flocking for asynchronuous distributed sensor fusion

232Authorized licensed useof

Proceedings limited
theto:IEEE | Vol.De
Universite 95,Sherbrooke.
No. 1, January 2007 on July 16,2023 at 14:01:05 UTC from IEEE Xplore.
Downloaded Restrictions apply.
 Olfati-Saber et al.: Consensus and Cooperation in Networked Multi-Agent Systems

based on average consensus,[ in Proc. 4th oscillators,[ SIAM J. Appl. Math., vol. 50, 16th IFAC World Congr., Prague, Czech,
Int. Symp. Information Processing in Sensor pp. 1645–1662, 1990. 2005.
Networks, pp. 63–70, Apr. 2005. [85] S. H. Strogatz, BFrom Kuramoto to [94] T. Eren, P. N. Belhumeur, and A. S. Morse,
[76] J.-Y. Chen, G. Pandurangan, and D. Xu, Crawford: Exploring the onset of BClosing ranks in vehicle formations
BRobust computation of aggregates in synchronization in populations of based on rigidity,[ in Proc. 41st IEEE
wireless sensor networks: Distributed coupled oscillators,[ Physica D, Conf. Decision and Control, Dec. 2002,
randomized algorithms and analysis,[ in vol. 143, pp. 1–20, 2000. pp. 2959–2964.
Proc. 4th Int. Symp. Information Processing in [86] G. B. Ermentrout and N. Kopell, BFrequency [95] L. Lovász, Random walks on graphs:
Sensor Networks, Apr. 2005, pp. 348–355. plateaus in a chain of weakly coupled A survey,[ in Combinatorics, Paul Erdos
[77] D. Kempe, A. Dobra, and J. Gehrke, oscillator,[ SIAM J. Math. Anal., vol. 15, is Eighty, D. Milos, V. T. Sos, and
BGossip-based computation of aggregate pp. 215–237, 1984. T. Szony, Eds. Budapest, Hungary: János
information,[ in Proc. 44th Annu. IEEE [87] J. Cortés, S. Martı́nez, T. Karatas, and Bolyai Math. Soc., 1996, pp. 353–398.
Symp. Foundations of Computer Science F. Bullo, BCoverage control for mobile [96] A.-L. Barabási and R. Albert, BEmergence
(FOCS ’03), 2003, vol. 3, no. 3, pp. 482–491. sensing networks,[ IEEE Trans. Robot. of scaling in random networks,[ Science,
[78] S. Boyd, A. Ghosh, D. Prabhakar, and Autom., vol. 20, no. 2, pp. 243–255, vol. 286, pp. 509–512, 1999.
B. Shah, BGossip algorithms: Design, analysis Apr. 2004. [97] P. Erdös and A. Rényi, BOn the evolution
and applications,[ in Proc. 24th Annu. Joint [88] L. Xiao and S. Boyd, BFast linear iterations of random graphs,[ Pub. Math. Inst. Hungarian
Conf. IEEE Computer and Communications for distributed averaging,[ Systems & Control Acad. Science, vol. 5, pp. 17–61, 1960.
Societies (INFOCOM ’05), Mar. 2005, Letters, vol. 52, pp. 65–78, 2004. [98] B. Sinopoli, L. Schenato, M. Franceschetti,
pp. 1653–1664. [89] D. J. Watts and S. H. Strogatz, BCollective K. Poola, M. I. Jordan, and S. S. Sastry,
[79] D. Fudenberg and D. K. Levine, The Theory of dynamics of Fsmall-world_ networks,[ BKalman filtering with intermittent
Learning in Games. Cambridge, MA: MIT Nature, vol. 393, pp. 440–442, Jun. 1998. observations,[ IEEE Trans. Autom.
Press, 1998. [90] H. Ando, Y. Oasa, I. Suzuki, and Control, vol. 49, no. 9, pp. 1453–1464,
[80] J. S. Shamma and G. Arslan, BDynamic M. Yamashita, BDistributed memoryless Sep. 2004.
fictitious play, dynamic gradient play, and point convergence algorithm for mobile [99] R. O. Saber and R. M. Murray, BFlocking
distributed convergence to Nash equilibria,[ robots with limited visibility,[ IEEE Trans. with obstacle avoidance: Cooperation
IEEE Trans. Autom. Control, vol. 50, no. 3, Robot. Autom., vol. 15, no. 5, pp. 818–828, with limited communication in mobile
pp. 312–327, Mar. 2005. Oct. 1999. networks,[ in Proc. 42nd IEEE Conf. Decision
[81] R. A. Horn and C. R. Johnson, Matrix [91] J. Lin, A. S. Morse, and B. D. O. Anderson, and Control, 2003, vol. 2, pp. 2022–2028.
Analysis. Cambridge, U.K.: Cambridge BThe multi-agent rendezvous problem,[ in [100] J. Wolfowitz, BProducts of indecomposable,
Univ. Press, 1987. Proc. 42nd IEEE Conf. Decision and Control, aperiodic, stochastic matrices,[ in Proc.
[82] G. Cybenko, BDynamic load balancing Dec. 2003, pp. 1508–1513. Am. Math. Soc., 1963, vol. 15, pp. 733–736.
for distributed memory multiprocessors,[ [92] J. Cortés, S. Martı́nez, and F. Bullo, [101] K. Zhou and J. C. Doyle, Essentials of Robust
J. Parallel and Distributed Computing, vol. 7, BRobust rendezvous for mobile autonomous Control. Upper Saddle River, NJ: Prentice
no. 2, pp. 279–301, Oct. 1989. agents via proximity graphs in arbitrary Hall, 1997.
[83] Y. Kuramoto, Chemical Oscillators, Waves, dimensions,[ IEEE Trans. Autom.
Control, vol. 51, no. 8, pp. 1289–1298, [102] M. E. J. Newman, BThe structure and
and Turbulance. Berlin, Germany: function of complex networks,[ SIAM Rev.,
Springer-Verlag, 1984. Aug. 2004.
vol. 45, pp. 167–256, 2003.
[84] R. E. Mirollo and S. H. Strogatz, [93] D. Spanos, R. Olfati-Saber, and
BSynchronization of pulse-coupled biological R. M. Murray, BDynamic consensus
on mobile networks,[ presented at the

ABOUT THE AUTHORS

Reza Olfati-Saber (Member, IEEE) received the Richard M. Murray (Fellow, IEEE) received the
S.M. and Ph.D. degrees in electrical engineering B.S. degree in electrical engineering from the
and Computer Science from the Massachusetts California Institute of Technology, Pasadena, in
Institute of Technology, Cambridge, in 1997 and 1985 and the M.S. and Ph.D. degrees in electrical
2001, respectively, and the B.S. degree in electrical engineering and computer sciences from the
engineering from Sharif University of Technology, University of California, Berkeley, in 1988 and
Tehran, Iran, in 1994. 1991, respectively.
He is an Assistant Professor at Thayer School of He is currently the Thomas E. and Doris
Engineering at Dartmouth College. Prior to joining Everhart Professor of Control and Dynamical
Dartmouth, he was a Postdoctoral Scholar at the Systems and the Director for Information Science
California Institute of Technology, Pasadena, from 2001 to 2004 and a and Technology at the California Institute of Technology. His research is
Visiting Scientist at the University of California, Los Angeles, from 2004 in the application of feedback and control to mechanical, information,
to 2005. His research interests include distributed control theory, multi- and biological systems. His current projects include integration of
agent systems, swarms, sensor networks, evolutionary dynamics of control, communications, and computer science in multi-agent systems,
behavior, sensor fusion and Kalman filtering, machine learning and information dynamics in networked feedback systems, analysis of insect
inference, dynamic social networks, and consensus theory. flight control systems, and synthetic biology using genetically encoded
Dr. Olfati-Saber is a member of Sigma Xi. finite state machines.

J. Alex Fax received the undergraduate degree in

mechanical and aerospace engineering from
Princeton University, Princeton, NJ, in 1993 and
the Ph.D. degree in control and dynamical
systems from the California Institute of Technol-
ogy (Caltech), Pasadena, in 2002.
He manages the Networked Navigation Sys-
tems group at Northrop Grumman’s Navigation
Systems Division, which focuses on novel navi-
gation applications and architectures to meet
evolving military needs. While at Caltech, he researched cooperative
control methods for groups of unmanned vehicles.

Vol.July
Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on 95, 16,2023
No. 1, January 2007UTC
at 14:01:05 | Proceedings
from IEEE Xplore.
ofRestrictions
the IEEE 233
apply.