Systems & Control Letters 61 (2012) 1023–1031

Contents lists available at SciVerse ScienceDirect

Systems & Control Letters

journal homepage: www.elsevier.com/locate/sysconle

Stabilization of non-minimum phase switched nonlinear systems with

application to multi-agent systems✩
Hao Yang a,∗ , Bin Jiang a , Huaguang Zhang b
a
College of Automation Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing, 210016, China
b
School of Information Science and Engineering, Northeastern University, Shenyang, 110004, China

article info abstract

Article history: This paper addresses the stabilization issue of a class of switched nonlinear systems where each mode
Received 29 July 2011 may be non-minimum phase, and the states of linearized dynamics of all modes compose the whole state
Received in revised form space. Time dependent and state dependent stabilization switching laws are provided by considering both
30 March 2012
common and multiple Lyapunov functions. The new results are applied to the aggregation problem of
Accepted 6 July 2012
Available online 5 September 2012
nonlinear multi-agent systems with designable switching connection topology. Finally, an aircraft team
example illustrates the efficiency of the proposed approaches.
Keywords:
© 2012 Elsevier B.V. All rights reserved.
Switched systems
Non-minimum phase
Stabilization
Multi-agent systems

1. Introduction where an approximated minimum phase model is utilized. The

same problem is also investigated in [12] by means of an inversion-
Many engineering applications can be modeled by switched based control strategy.
systems due to the existence of various jumping parameters [1]. The main idea of the above results consists of two steps:
Fruitful results have been reported on stability and stabilization of
switched nonlinear systems, e.g., [2–5] to name a few. 1. Design the individual controller respectively in each mode to
Stabilization of non-minimum phase nonlinear systems is a compensate for its own unstable internal dynamics such that
quite challenging problem. Several fundamental methods have all modes become stable individually. This can be done by using
been proposed including state feedback control [6], and output the existing techniques for non-minimum phase nonlinear
feedback control [7–9]. The main idea behind these methods is to systems.
compensate for the unstable zero dynamics by means of output 2. Apply the standard stability condition of switched systems,
synthesis or auxiliary systems such that the system becomes sta- e.g. common/multiple Lyapunov functions methods to achieve
ble. The controllability of the zero dynamics is the basic require- the stability of the whole switched system.
ment, otherwise, the stabilization cannot be achieved.
This idea is natural and extends the approaches of non-
Some contributions have also been devoted to non-minimum
switching systems to the switched one. However, it is well known
phase switched nonlinear systems where each nonlinear mode
that the stabilization for non-minimum phase nonlinear systems is
may be non-minimum phase. In [10], H∞ control goal is achieved
quite difficult, and is even impossible to be achieved if the unstable
for a class of non-minimum phase cascade switched nonlinear
zero dynamics is uncontrollable.
systems where the internal dynamics of each mode is assumed to
In this work, we focus on the stabilization of non-minimum
be asymptotically stabilizable. Output tracking of non-minimum
phase switched nonlinear systems has been considered in [11], phase switched nonlinear systems where the internal dynamics
of each mode may be unstable and uncontrollable. Different
from [10–12], we do not try to compensate for the unstable internal
dynamics respectively in each mode. Instead, we achieve the
✩ This work is supported by the National Natural Science Foundation of China
stabilization from the overall system point of view. It will show that
(61034005, 61010121, and 61104116), the Doctoral Fund of Ministry of Education
under some conditions, the negative effects of internal dynamics in
of China (20113218110011) and the NUAA Research Funding (NS2011016).
∗ Corresponding author. Tel.: +86 25 84892301x6060; fax: +86 25 84892300. some modes may be compensated by other modes, and the overall
E-mail addresses: [email protected] (H. Yang), [email protected] switching process can still be stable. The similar idea can be seen
(B. Jiang), [email protected] (H. Zhang). in [13] for switched linear systems and [14] for switched nonlinear
0167-6911/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.sysconle.2012.07.001
1024 H. Yang et al. / Systems & Control Letters 61 (2012) 1023–1031

systems. This special property of a switched system provides a new

clue for stabilizing the non-minimum phase systems. Two benefits
follow:
1. It relaxes significantly the condition on the internal dynamics (it
is allowed to be unstable and uncontrollable simultaneously).
2. It makes easier the design of each mode’s individual controller.
We consider a class of switched nonlinear systems where each
mode may be non-minimum phase, and the states of linearized dy-
namics of all modes compose the whole state space. Consequently,
we provide novel time dependent and state dependent stabiliza-
tion switching laws with considering both common and multiple
Lyapunov functions. The proposed results are more general and Fig. 1. The structure of switched systems.
flexible than that in [14] where only a time-dependent switching
law is provided.
A basic assumption throughout the paper is given:
As an important application of the new results, we consider
the target aggregation problem of nonlinear multi-agent systems
which can be divided into several groups of subsystems with Assumption 2.1. For a switched system (1) with all modes satis-
1 , . . . , xm ] .
fying (2)–(4), it holds that x = [x⊤ ⊤ ⊤
each group being in the leader-following structure, and only one 
group is allowed to be interconnected at one time. Inspired by the Assumption 2.1 implies that the states of unstable internal
linearized feedback control idea [15], we design a novel ‘‘feedback dynamics in each mode can be controlled and linearized in other
control topology’’ for each group, under which the multi-agent modes. The states of all linearized dynamics compose the whole
system with switching topologies can be regarded as a non-
state space as illustrated in Fig. 1. This allows us to achieve
minimum phase switched system and the proposed switched
the stabilization by fully utilizing the tradeoff among different
system results are applied.
modes of the switched system. Assumption 2.1 can be relaxed to a
In the rest of the paper, some preliminaries are given in
more moderate case where more than one mode may share some
Section 2. Sections 3 and 4 propose time dependent and state
states in their linearized dynamics. The proposed methods can be
dependent stabilization switching laws respectively, which are
straightforwardly extended to this case.
applied to multi-agent systems in Section 5, followed by some
concluding remarks in Section 6.
Remark 2.1. We would like to show how to fully utilize the
switching properties to achieve the stabilization goal. Therefore,
2. Preliminaries
an idea of ‘‘overall system point of view’’ will be followed
throughout the paper. For more general switched systems where
Consider the following switched nonlinear control systems
Assumption 2.1 is not satisfied, the combination between the
ẋ = fσ (x) + gσ (x)uσ (1) existing control techniques for non-switching systems and the
where x ∈ ℜ are available states. Define M = {1, 2, . . . , m},
n switching properties can be developed. 
where m is the number of modes. σ (t ) : [0, ∞) → M denotes
the switching signal, which is assumed to be a piecewise constant
function continuous from the right. ∀i ∈ M , ui ∈ ℜ is the input, fi 3. Time-dependent switching law
and gi are smooth functions with fi (0) = gi (0) = 0. It is required
that the activating period of each mode is not less than τ where In this section, we provide time-dependent stabilizing switch-
τ > 0 is called ‘‘dwell-time’’ [1]. We also assume that the states do ing laws for the switched system satisfying Assumption 2.1.
not jump at the switching instants.
Suppose that for each mode i ∈ M , we can find a function yi 3.1. Common Lyapunov function
i , xi ] where xi ∈ ℜ , x̄i ∈ ℜ
and a partition x = [x̄⊤ ⊤ ⊤ ri n−ri
to rewrite
the system (1) into the normal form [15]: Define a special function that will be used in the following
sections:
x̄˙ i = ψi (x̄i , xi ) (2)
ẋi = Axi + bi (x̄i , xi ) + ai (x̄i , xi )ui (3) Definition 3.1. A class GKL function γ : [0, ∞) × [0, ∞) →
yi = Cxi i = 1, 2, . . . , m (4) [0, ∞) if γ (·, t ) is of class K 1 for each fixed t ≥ 0 and γ (s, t )
increases to infinity as t → ∞ for each fixed s ≥ 0. 
where
···
0 1 0
 
Assumption 3.1. For the switched system (1) satisfying Assump-
.. .. . . ..  tion 2.1, we can design ui for each mode i under which there exists
. . . . ,

Ai ∈ ℜri ×ri =  a continuous non-negative function
0 0 ··· 1
 
0 0 ··· 0 V : ℜn → ℜ≥0 = V1 (x1 ) + V2 (x2 ) + · · · + Vm (xm ) (5)
1×ri
= 1, 0, . . . , 0
 
Ci ∈ ℜ where Vk (xk ) ∈ C 1 : ℜrk → ℜ≥0 , k ∈ M , and there exist
and bi (x̄i , xi ) = [0, 0, . . . , 0, b̄i (x̄i , xi )] , ai (x̄i , xi ) = [0, 0, . . . , 0,
⊤ λi > 0, α1 , α2 ∈ K∞ , γab ∈ K∞ , for a, b ∈ M, and φj ∈ GKL ∀j ∈
āi (x̄i , xi )]⊤ with b̄i and āi being scalar functions, āi ̸= 0. M − {i} such that
Mode i is a non-minimum phase if its zero dynamics x̄˙ i =
ψi (x̄i , 0) is unstable, otherwise it is a minimum phase. The problem
to be solved in this paper is whether the switched system (1) with 1 Class K is a class of strictly increasing and continuous functions [0, ∞) →
all non-minimum phase modes satisfying (2)–(4) can be stabilized by [0, ∞) which are zero at zero. Class K∞ is the subset of K consisting of all those
using σ and uσ ? functions that are unbounded.
 H. Yang et al. / Systems & Control Letters 61 (2012) 1023–1031 1025

Theorem 3.1. Consider a switched system (1) satisfying Assump-

tions 2.1, 3.1 and 3.2. Under switching law S1 , there exist ϵ > 0 such
that for any given δ0 > 0, limt →∞ sup V (t ) = mϵ, ∀Vi (0) ≤ δ0 , i ∈
M.
Proof. We first suppose that Vj (t ) > ϵ, ∀t ≥ 0. Consider xi
dynamics in the first period [0, T ). Since the switching sequence
is mode 1 → mode 2 → · · · mode i · · · → mode m, the solution
of Vi (i ̸= 1) at t = 1t can be represented as

φj (Vi (0), 1t ) + max γik (∥Vk ∥[0,1t ) )

 
k∈M −{i}
Vi (1t ) ≤ Vi (0). (10)
Vi (0)
Assumption 3.2 ensures that Vi (1t ) ≤ ηVi (0) if ∥Vk ∥[0,1t ) ≤
ηm−1 δ0 for k ∈ M − {i}. Note that the initial Vj (0) ≤ δ0 , ∀j ∈ M.
Therefore at [0, 1t ), V1 (t ) would decrease, other Vi (t ) increase
until t = 1t at which Vi (1t ) ≤ ηVi (0) ≤ ηδ0 .
By induction, it is not hard to find that under S1 , Vi would
Fig. 2. Behaviors of Vi under S1 . decrease when mode i is activated, and increase when other modes
are working. It follows from Assumption 3.2 that ∥Vk ∥[0,T ) ≤
α1 (|xs |) ≤ Vs (xs ) ≤ α2 (|xs |), ∀s ∈ M (6) ηm−1 δ0 for k ∈ M − {i}.
Vi (xi (t )) ≤ e −λi (t −tik )
Vi (xi (tik )), t ≥ tik (7) Since λi can be assigned arbitrarily by ui , we choose λi such that
Vj (xj (t )) ≤ φj (Vj (xj (tik )), t − tik ) ⋆
e−λi 1t ηm−1 ≤ e−λ m1t , for λ⋆ > 0. (11)
+ max γjk (∥Vk (xk )∥[tik ,t ) ) ,
 
∀j ∈ M − {i}, t ≥ tik (8) One further has that
k∈M −{j}
⋆
where tik , i ∈ M , k = 1, 2, . . . denotes the kth time when mode Vi (T ) ≤ ηm−i · e−λi 1t · ηi−1 · Vi (0) ≤ e−λ T Vi (0). (12)
i is switched on. | · | is the Euclidean norm, ∥ · ∥[a,b] denotes the
supremum norm on [a, b].  Thus Vi is bounded in [0, T ) and its value decreases once a period
of switching is completed as shown in Fig. 2.
Inequality (7) implies that when mode i is activated, xi dynamics Combining all Vi yields
is fully linearizable and is exponentially stable with decay rate
⋆
λi under appropriate design of ui . Inequality (8) is a general V (T ) ≤ e−λ T V (0). (13)
formulation to describe the instability of other states of internal
dynamics in mode i, γjk denotes the gain from Vk to Vj and evaluates In the second interval [T , 2T ), one has ∥Vk ∥[T ,2T ) ≤ δ , for δ < δ0 .
⋆
the interconnection between xk and xj dynamics. Note that for Assumption 3.2 still holds, It follows that V (2T ) ≤ e−λ T V (T ).
xj , γjk may differ among different modes. Following a similar way, we can obtain that V ((k + 1)T ) ≤
⋆
Choose 1t ≥ τ , and define T , m1t. A time-dependent e−λ T V (kT ), k = 0, 1, 2, . . . . Once at some time instants t =
periodic switching law is designed: t ⋆ , Vj (t ⋆ ) ≤ ϵ, Vi may not decrease. Therefore limt →∞ sup Vi (t ) =
Switching law S1 ϵ , and limt →∞ sup V (t ) = mϵ . This completes the proof. 
1. Let k = 0, i = 1.
2. Activate mode i at t = kT + (i − 1)1t, until t = kT + i1t, Remark 3.1. The number ϵ is designed according to system struc-
go to 3. ture. A small gain γjk allows a small ϵ as shown in Assumption 3.2.
3. Let i = i + 1, if i = m + 1, then let i = 1, k = k + 1, go This is consistent with the small gain idea in [16] for large scale
to 2; else, go to 2.  interconnected systems. If there is no interconnection among all
We focus on the decreasing behavior of each xi dynamics (each xi dynamics (as in Section 5), γjk disappears. In this case, we could
Vi ) respectively throughout the system process. The expected let ϵ = 0, the origin is rendered asymptotically stable. An alter-
behavior under S1 is illustrated in Fig. 2. Vi decreases in each native stabilization approach follows [14] that makes xi input-to-
activating period of mode i and increases when other modes are state stable in each [kT , (k + 1)T ], i.e.
activated. The following assumption is needed to guarantee that |xi ((k + 1)T )| ≤ β(|xi (kT )|, (k + 1)T )
Vi (xi (kT )), k = 0, 1, . . . , converges to a small region. We will
+ max γ̄ik (∥xk ∥[kT ,(k+1)T ) ) , β ∈ KL, γ̄ik ∈ K∞
 
denote Vi (t ) instead of Vi (xi (t )), ∀i ∈ M for the sake of simplicity. (14)
k∈M −{i}

Assumption 3.2. For a given δ0 > 0, there exist positive numbers and then applies the small gain conditions among γ̄ik , ∀i ∈ M .
η > 1, ϵ > 0, 1t ≥ τ such that when mode i is activated and However, such input-to-state stability is not easy to verify for gen-
Vi (tik ) > ϵ, xj dynamics (j ̸= i) satisfies eral switched nonlinear systems. 
φj (Vj (tik ), 1t ) + max γjk (ηm−1 δ0 )
 
k∈M −{j} 3.2. Multiple Lyapunov functions
≤ η.  (9)
Vj (tik )
This section extends the result in Section 3.1 to the case of
It follows from (8) that the left side of (9) is an upper bound multiple Lyapunov functions.
of the gain of Vj in the period 1t, if ∥Vk ∥[tik ,tik +1t ) ≤ ηm−1 δ0 .
Inequality (9) implies that the diverging speed of each mode’s Assumption 3.3. For each mode i ∈ M of system (1) satisfying
internal dynamics is bounded during a interval that is not less than Assumption 2.1, we can design ui under which there exists a
the dwell time τ . Since the switching sequence and instants are continuous non-negative function
designed a priori, on-line information cannot be used. Such a fixed (i) (i)
upper bound has to be imposed. η is not hard to find if γjk is small. V (i) : ℜn → ℜ≥0 = V1 (x1 ) + V2 (x2 ) + · · · + Vm(i) (xm ) (15)
1026 H. Yang et al. / Systems & Control Letters 61 (2012) 1023–1031

(i)
where Vk (xk ) ∈ C 1 : ℜrk → ℜ≥0 , k ∈ M satisfies (6)–(8).
Moreover, for p, q ∈ M , p ̸= q,
(p) (q)
Vi (xi ) ≤ ζ (Vi (xi )) (16)
where ζ ∈ K∞ and ζ (ab) ≤ aζ (b) for 0 ≤ a ≤ 1, b ≥ 0. 

Remark 3.2. Inequality (16) is a condition imposed on multiple

(j)
functions Vi in different modes. This formulation covers two
(p)
general forms: (1) The µ condition in [1] under which Vi (xi ) ≤
(q)
µVi (xi ) for µ ≥ 1 and (2) the difference condition in [17] under
(p) (q)
which Vi (xi ) ≤ Vi (xi ) + α(|xi (0)|) for α ∈ K∞ . 
For a positive number ρ > 0, define ζ 2 (ρ) , ζ (η · ζ (ηρ)),
ζ (ρ) , ζ (η · ζ (η · ζ (ηρ))) · · ·.
3

Fig. 3. Behaviors of Vi and V under S2 .

Assumption 3.4. For a given δ0 > 0, there exists positive numbers
η > 0, 1t ≥ τ such that when mode i is activated and Vj (tik ) > Assumption 4.1. For each mode i ∈ M of system (1) satisfying
ϵ, xj dynamics (j ̸= i) satisfies Assumption 2.1, we can design ui under which there exists a con-
tinuous non-negative function satisfying (5)–(7), and 1V̄i (t )tik ≤
φj (Vj(i) (tik ), 1t ) + max γjk (ζ m−1 (δ0 ))
 
k∈M −{j} β(V̄i (tik ), t − tik ) for β ∈ GKL. Moreover, there exist positive
(i)
≤ η.  (17)
ϵ, υ > 0 such that if we let tik = t for i = min arg maxi⋆ ∈M Vi⋆ (t ),
Vj (tik )
then

Theorem 3.2. Consider a switched system (1) satisfying Assump- 1V̄i (tik + τ )tik ≤ −1Vi (tik + τ )tik − υ (22)
tions 2.1, 3.3 and 3.4. Under switching law S1 , there exists ϵ > 0 such for Vi (tik ) > ϵ . 
(1)
that for any given δ0 > 0, limt →∞ sup V (1) (t ) = mϵ, ∀Vi (0) ≤
δ0 , i ∈ M . Assumption 4.1 means that the diverging speed of each mode’s
internal dynamics is bounded. Moreover, if Vi at each tik is maximal
Proof. Consider xi dynamics in the first period [0, T ). Suppose that among all Vi⋆ , i⋆ ∈ M , then during each activating period τ of mode
(j)
Vi (t ) > ϵ, ∀j ∈ M , t ≥ 0. Similar to the proof of Theorem 3.1, i, the increasing amount of V̄i is less than the decreasing amount of
we have Vi . The decay rate of Vi can be designed arbitrarily. Inequality (22)
(m) is satisfied if the increasing rate of V̄i is small enough, and can be
Vi (T − ) ≤ ηζ m−1−i (e−λi 1t ζ i−1 (δ0 )). (18)
checked conveniently when V̄i follows the exponential diverging
Note that ηζ m−1−i (e−λi 1t ζ i−1 (δ0 )) ≤ ηe−λi 1t ζ m−2 (δ0 ), and form as shown in Section 5.
(1) (m) A state-dependent switching law is designed:
Vi (T ) ≤ ζ (Vi (T − )). Assign λi such that
⋆ (1)
Switching law S2
e−λi 1t ζ m−1 (δ0 ) ≤ e−λ T Vi (0), for λ⋆ > 0 (19) 1. Let k = 0, t0 = 0, choose ϵ > 0, κ > 0.
we further obtain that
2. At t = tk , choose i = min arg maxi⋆ ∈M Vi⋆ (tk ).
3. If Vi (tk ) > ϵ , go to 4; Else, go to 5.
(1) ⋆
Vi (T ) ≤ e−λ T Vi(1) (0). (20) 4. Choose σ (tk ) = i. Until t = t ⋆ such that 1V̄i (t ⋆ )tk =
(1)
−1Vi (t ⋆ )tik − υ , let k = k + 1, go to 2.
Combining all Vi yields 5. If k = 0, choose σ (t0 ) = i, let k = k + 1, go to 6; Else,
⋆ go to 6.
V (1) (T ) ≤ e−λ T V (1) (0). (21) 6. Let mode σ (tk−1 ) work until t = t ⋆⋆ such that ∃j ∈ M ,
(i)
Thus all V , i ∈ M are bounded in [0, T ], and V decreases (1) Vj (t ⋆⋆ ) > ϵ and V (t ⋆⋆ ) < mϵ + κ . let tk = t ⋆⋆ , goto 2. 
once a period of switching is completed. The main idea behind S2 is that once Vi is maximal among
Following a similar way, we can obtain that V (1) ((k + 1)T ) ≤ all functions (Step 2), mode i is activated, and both Vi and V are
⋆
e−λ T V (1) (kT ), k = 0, 1, 2, . . .. Once at some time instants t = decreased (Step 4). The decreasing performance of V is expected to
(j) (j) (1)
t , Vi (t ⋆ ) ≤ ϵ, Vi may not decrease, consequently, Vi may also
⋆
be maintained at each switching instant until V ≤ mϵ as illustrated
(1)
not decrease. Therefore limt →∞ sup V (t ) = mϵ . This completes in Fig. 3.
the proof. 
Theorem 4.1. Consider a switched system (1) satisfying Assump-
4. State-dependent switching law tions 2.1 and 4.1. Under switching law S2 , there exists ϵ > 0 such
that for any given δ0 > 0, limt →∞ sup V (t ) = mϵ + κ , where κ > 0
In this section, we propose a more flexible state-dependent can be arbitrarily small, ∀Vi (0) ≤ δ0 , i ∈ M .
switching law where the dwell period of each mode is not designed
Proof. Consider the interval [0, t1 ) during which mode i is sup-
and fixed a priori but relies on real-time state values. We will
posed to be activated, t1 ≥ τ . We consider two cases:
focus on the decreasing behavior of the whole system in each mode
Case 1: Vi (0) > ϵ . We have
respectively. This is quite different from the idea in Section 3.
V (t1 ) = Vi (t1 ) + V̄i (t1 )
4.1. Common Lyapunov function
= Vi (0) + 1Vi (t1 )0 + V̄i (0) + 1V̄i (t1 )0 . (23)
Similar to Section 3.1, consider a series of functions V1 (x1 ), Assumption 4.1 guarantees that 1V̄i (τ )0 ≤ −1Vi (τ )0 − υ . There
V2 (x2 ), . . . , Vm (xm ) with Vk (xk ) ∈ C 1 : ℜrk → ℜ≥0 . Define always exists t1 ≥ τ such that
V̄i , ( s=1 Vs ) − Vi , and 1Vi (ta )tb , Vi (ta ) − Vi (tb ) for two instants
m
ta , tb . 1V̄i (t1 )0 = −1Vi (t1 )0 − υ. (24)
 H. Yang et al. / Systems & Control Letters 61 (2012) 1023–1031 1027

Substituting (24) into (23) yields 4. Choose σ (tk ) = i. Until t = t ⋆ such that 1V̄i(i) (t ⋆ )tk =
(i)
V (t1 ) < V (0). (25) 1Vi (t ⋆ )tk − υ − ϵ , let k = k + 1, go to 3.
5. If k = 0, choose σ (t0 ) = i, let k = k + 1, go to 6; Else,
⋆⋆
Case 2: Vi (0) ≤ ϵ . It follows that V (0) ≤ mϵ . Once Vi (t ) > ϵ go to 6.
and V (t ⋆⋆ ) < mϵ + κ , at t = t ⋆⋆ , for i ∈ M , the switching would 6. Let mode σ (tk−1 ) work until t = t ⋆⋆ such that
occur at t = t ⋆⋆ , thus t1 = t ⋆⋆ . We have V (σ (tk−1 )) (t ⋆⋆ ) > 2mϵ and ζ (V (σ (tk−1 )) (t ⋆⋆ )) < 2mϵ + κ , for
i ∈ M . let tk = t ⋆⋆ , go to 2. 
V (t1 ) < mϵ + κ. (26)

Combining (25) and (26) leads to Theorem 4.2. Consider a switched system (1) satisfying Assump-
tions 2.1, 4.2 and 4.3. Under S3 , there exists ϵ > 0 such that for any
V (t1 ) < max[V (0), mϵ + κ]. (27) given δ0 > 0, limt →∞ sup V (i) (t ) = 2mϵ + κ, ∀i ∈ M where κ can
(i)
By induction, we further have be arbitrarily small, ∀Vi (0) ≤ δ0 , i ∈ M .

V (t2 ) < max[V (t1 ), mϵ + κ] Proof. Consider the interval [0, t1 ) during which mode i is sup-
.. posed to be activated. We still consider two cases:
(i)
. Case 1: Vi (0) > 2ϵ . It follows from Assumption 4.2 that
V (tk ) < max[V (tk−1 ), mϵ + κ], k = 1, 2, . . . . (28)
V (i) (t1 ) < V (i) (0) − ϵ. (31)
Finally we obtain V (t ) < max[V (tk ), mϵ + κ], for t > tk , k =
0, 1, . . . . One can find that V (t ) would decrease if it is larger than We also have ∀j ∈ M − {i}
mϵ + κ . Once V (t ) ≤ mϵ , it may decrease or increase and is always
V (j) (t1 ) ≤ ζ̄ V (i) (t1 )
 
not larger than mϵ + κ . Therefore limt →∞ sup V (t ) = mϵ + κ . This (32)
completes the proof.  ζ̄ −1
V (i) (0) ≤ V (j) (0).
 
(33)

Remark 4.1. The decreasing speed of V depends on υ , the smaller Assumption 4.3 ensures that
is υ , the slower V decreases. ϵ could also be small if −1Vi (tik +τ )tik
V (i) (0) − ζ̄ −1 ◦ ζ̄ −1 (V (i) (0)) < ϵ. (34)
is large enough compared with 1V̄i (tik + τ )tik . If (22) holds for all
Vi (tik ), we can choose an arbitrarily small ϵ . In this case, the states Combining (31) and (34) yields
converge to an arbitrarily small region. 
V (i) (t1 ) < ζ̄ −1 ◦ ζ̄ −1 (V (i) (0)). (35)
4.2. Multiple Lyapunov functions One further has

ζ̄ V (i) (t1 ) < ζ̄ −1 V (i) (0) .

   
Now we extend the result in Section 4.1 to the case of multiple (36)
Lyapunov functions.
This together with (32)–(33) leads to
Assumption 4.2. For each mode i ∈ M of system (1) satisfying V (j) (t1 ) < V (j) (0). (37)
Assumption 2.1, we can design ui under which there exists a
(j)
(i)
continuous non-negative function that satisfies (15) with Vk (xk ) Inequality (37) implies that at t1 , V , ∀j ∈ M − {i} also decreases.
(i)
satisfying (6)–(7), (16) and Case 2: Vi (0) ≤ 2ϵ . It follows from condition (29) that
(j)
V (0) ≤ 2mϵ, ∀j ∈ M .
Vi
(i)
≥ Vi(j) , ∀j ∈ M − {i}. (29) The switching would occur at t = t1 if V (i) (t1 ) > 2mϵ . More-
(i) (i) over, whatever mode j is activated at t = t1 , it holds that V (j) (t1 ) <
Moreover, 1V̄i (t )tik < β(V̄i (tik ), t − tik ) for β ∈ GKL. There 2mϵ + κ .
exists positive ϵ, υ > 0 such that To this end we have
(i) (i)
1V̄i (tik + τ )tik ≤ −1Vi (tik + τ )tik − υ − ϵ V (i) (t1 ) < max[V (i) (0) − ϵ, 2mϵ + κ] (38)
(i) (j) (j)
for Vi (tik ) > 2ϵ .  V (t1 ) < max[V (0), 2mϵ + κ], ∀j ∈ M − {i}. (39)
(i) By induction, we further have
We can design ui and choose appropriate Vi to satisfy (29).
Under condition (16), we can also find a K∞ function denoted as
V (σ (tk )) (tk+1 ) < max[V (σ (tk )) (tk ) − ϵ, 2mϵ + κ] (40)
ζ̄ such that
(j) (j)
V (tk+1 ) < max[V (tk ), 2mϵ + κ],
V (p) (x) ≤ ζ̄ (V (q) (x)). (30)
∀j ∈ M − {σ (tk )} k = 1, 2, . . . . (41)

Similar to the proof of Theorem 4.1, we finally obtain that

Assumption 4.3. When mode i ∈ M is activated at t = tik , there
limt →∞ sup V (i) (t ) ≤ 2mϵ + κ . This completes the proof. 
exists a positive ϵ > 0 such that V (i) (tik ) − ζ̄ −1 ◦ ζ̄ −1 (V (i) (tik )) < ϵ ,
∀i ∈ M, for Vi(i) (tik ) > 2ϵ .  Remark 4.2. Each V (i) decreases at each switching instant of the
A state-dependent switching law is designed: whole process until V (i) ≤ 2mϵ whatever mode is activated. A key
condition to achieve this goal is Assumption 4.3, which implies that
Switching law S3 the effect of ζ̄ (·) among V (i) can be compensated if the decreasing
1. Let k = 0, t0 = 0. Choose ϵ > 0, κ > 0, 0 < υ < ϵ . amount of V (i) is large enough when mode i is activated. This
(i⋆ ) (i)
2. At t = tk , choose i = min arg maxi⋆ ∈M Vi⋆ (tk ). requires that both the percent rate of Vi in V (i) and the decreasing
(i) (i)
3. If Vi (tk ) > 2ϵ , go to 4; Else, go to 5. amount of Vi are large enough. 
1028 H. Yang et al. / Systems & Control Letters 61 (2012) 1023–1031

5. Application to multi-agent systems 5.2. Feedback control topology

This section applies the results in Sections 3–4 to the target Motivated by the linearized feedback control idea [15], we
aggregation problem of a class of multi-agent systems. Previous propose a ‘‘feedback control topology’’ with its corresponding
assumptions will be further verified. cooperative controllers.
Since there are m groups of subsystems, we build m connection
topologies respectively by using the following rule.
5.1. Problem formulation
Rule R of building topology i (i ∈ M)
1. Pick agents in Ni .
The system consists of n agents modeled by a directed graph
2. Set arcs (is+1 , is ), for s = 1, 2, . . . , ri − 1.
G = (N , E ), where N = {1, 2, . . . , n} is the set of agents and E is
3. Set arcs (iι , iri ), for ι = 1, . . . , ri − 1. 
the set of arcs, (j, i) ∈ E denotes an arc from agent j to agent i, such
that agent i can receive information directly from agent j. Agent j In the topology resulting from R, there is a chain from
is the neighbor of agent i if (j, i) exists. the leader to followers, and each follower also feeds back its
The dynamics of agents are given as: information to the leader.
The cooperative law of agent is is designed as
ẋi = gi (xi ) + uci (x1 , . . . , xn ), i∈N (42)   ⊤
∂ Vi
 ∂ xiss

where xi ∈ ℜq , ∀i ∈ N , gi is a smooth function, uci represents the

 2 (−Ki Φi ) s = ri

cooperative law of agent i. Under topology i H⇒ ucis =  ∂ Vis  (45)
Given a constant target point, denoted as a constant vector
  ∂ xi 
s

a(xis+1 − xis ) s = 1, . . . , ri − 1

χ ∈ ℜq , define Vi , (xi − χ )⊤ P (xi − χ ), i ∈ N with P ∈ ℜq×q

a symmetric positive definite matrix. Vi is a potential function that
where Φi , [Vi1 , Vi2 , . . . , Viri ]⊤ . Ki = [ki1 ki2 · · · kiri ] is the
evaluates the distance between the target point and agent i. Define
feedback gain vector with each element kis > 0. a ≥ λ1 is a positive
N l ⊆ N such that each agent i ∈ N l is the leader who knows the  2
 ∂V 
target information. number. Due to the structure of Vis , the term  ∂ x is  is impossible
is
to become zero unless xis − χ = 0.
Assumption 5.1. There exist λ0 , λ1 > 0, such that
Theorem 5.1. Consider a group Ni of multi-agent system (42) satis-
uci = 0 ⇒ V̇i ≤ −λ0 Vi , ∀i ∈ N l (43) fying Assumption 5.1. The cooperative law (45) under R guarantees
uci = 0 ⇒ V̇i ≤ λ1 Vi , ∀i ∈ N − N l .  (44) that |xi (t ) − χ| exponentially tends to zero with the arbitrary decay
rate.

Inequality (43) implies that agent i ∈ N l can reach the target by Proof. According to Assumption 5.1, the time derivative of Vis , s =
itself. All other agents are the followers. Without cooperation, these 1, . . . , ri − 1 along (42) satisfies
agents may neither reach χ nor keep a constant distance with χ , V̇is ≤ λ1 Vis + 2a(xis − χ )⊤ P (xis+1 − xis )
and may run far away from the target as in (44). Assumption 5.1 is
more general than the setting in most literature, e.g. [18,19] where ≤ λ1 Vis + 2a(−(xis − χ )⊤ P (xis − χ )
the follower keeps a constant distance from the target without + (xis − χ )⊤ P (xis+1 − χ )). (46)
cooperation.
The connection topology of a multi-agent system is often There exists a constant matrix η ∈ ℜ q×q
such that P = η η. We
⊤

changeable due to some intentional reasons, e.g. communication further have

cost limitation or protection action, such that any allowable indi-
2(xis − χ )⊤ P (xis+1 − χ ) ≤ (xis − χ )⊤ η⊤ η(xis − χ )
vidual topology cannot achieve the goal of the whole group [20].
In this case, the topology can be designable and switched among a + (xis+1 − χ )⊤ η⊤ η(xis+1 − χ )
family of candidate topologies.
≤ Vis + Vis+1 . (47)
In this work, the original multi-agent system is divided into
several groups of subsystems as in [21]. Only one group is allowed Substituting (47) into (46) yields
to be interconnected at one time. There exists a supervisor that
determines which group would be connected. V̇is ≤ (λ1 − a)Vis + aVis+1 ≤ aVis+1 . (48)
m
More precisely, divide N = i=1 Ni , where each group Ni One can also get from Assumption 5.1 that the time derivative of
includes ri agents denoted as Viri along (42) satisfies
 
i−1 i−1 i
V̇iri ≤ −λ0 Viri − Ki Φi ≤ −Ki Φi .
  
Ni = rs + 1, rs + 2, . . . , rs (49)
s=0 s=0 s=0
Combining (48)–(49) leads to the compact form

where r0 = 0, and r1 + · · · + rm = n. It is clear that Ni Nj =
∅, ∀i, j ∈ M and i ̸= j. Φ̇i ≤ Āi Φi (50)
i−1
For the sake of simplicity, define ij = s=0 rs + j. Therefore, where
the states of agents in Ni are [xi1 , . . . , xiri ]⊤ . Suppose that iri ∈ N l ,  0 a ··· 0

i.e., in each group, there is a unique leader who knows the target .. .. .. .. 
. . . . .

information. The problem to be solved is to let the states of all Āi = 
agents (42) satisfying Assumption 5.1 reach χ in the case that only 0 0 ··· a
 
one group Ni , i ∈ M is allowed to be interconnected at one time. −λ0 − ki1 −ki2 ··· −kiri
 H. Yang et al. / Systems & Control Letters 61 (2012) 1023–1031 1029

Since the eigenvalues of Āi can be assigned arbitrarily according Proof. Condition (1) guarantees that under each topology,
to Ki , there exists a nonnegative function Assumption 3.1 is satisfied. Since there is no interconnection
among groups, Assumption 3.2 is also satisfied under condition (2),
Ωi (Φi ) , Φi⊤ P̄ Φi (51) and we can choose ϵ = 0. Therefore, it follows from Theorem 3.1
that limt →∞ Ω (t ) = 0 under S1 . 
with P̄ ∈ ℜri ×ri a symmetric positive definite matrix, whose time
derivative along (50) satisfies Theorem 5.3 (State-Dependent Switching Topology). Consider m a
⋆ multi-agent system (42) satisfying Assumption 5.1. limt →∞ sup i=1
Ω̇i (Φi ) ≤ −λ Ωi (Φi ) (52) Ω (t ) = ς , with ς > 0 an arbitrarily small number, ∀Ωi (Φi (0)) ≤
where λ⋆ > 0. This completes the proof.  δ0 , i ∈ M with any given δ0 > 0 if
(1) Each group has the cooperative law (45) and (53) under R.
⋆⋆ ⋆
Remark 5.1. The similar target aggregation problem is also ad- (2) (m − 1)(eλ τ − 1) < 1 − eλ τ .
dressed in [22] with uci = (xj − xi ), where aij is the weight (3) m candidate topologies are switched following S2 with Ωi

j̸=i a ij
of the arc between agent i and agent j. aij > 0 if (j, i) exists, oth- replacing Vi and Ω replacing V .
erwise aij = 0. The strategy provided in [22] requires each agent Proof. The proof is based on Theorem 4.1. It follows from (54) that
to calculate weights of its own arcs that depend on the weights of
⋆⋆ τ
other arcs. Comparatively, a benefit of the cooperative law (45) un- 1Ωj (Φj (tik + τ ))tik ≤ (eλ − 1)Ωj (Φj (tik )). (56)
der the feedback control topology is that all calculations are done
only by the leader, which significantly reduces the real-time com- Since mode i is activated at t = tik if Ω (Φi (tik )) is maximal among
putational burden of the following agents, and is simpler to be all functions, we further have
implemented.  ⋆⋆ τ
1Ωj (Φj (tik + τ ))tik ≤ (m − 1)(eλ

− 1)Ωj (Φi (tik )). (57)
Now let us turn to other groups in N − Ni under topology i. j∈M −{i}

Since there is no interconnection in these groups, all cooperative On the other hand, one has
laws are set zero, i.e.
⋆
1Ωi (Φi (tik + τ ))tik ≥ (1 − eλ τ )Ωi (Φi (tik )). (58)
Under topology i H⇒ ucj = 0, ∀j ∈ N − Ni . (53)
Combining (57), (58) and Condition (2) yields
According to Assumption 5.1, we further obtain that

1Ωj (Φj (tik + τ ))tik ≤ Ωi (Φi (tik )) − υ (59)
Ω̇k (Φk ) ≤ λ⋆⋆ Ωk (Φk ) ∀k ∈ M − {i} (54)
j∈M −{i}

for λ⋆⋆ > 0, where Ωk is defined in (58). for υ > 0 that can be arbitrarily small. Therefore, under conditions
We can find that under topology i, the whole multi-agent (1)–(2), Assumption 4.1 is satisfied. It follows from Theorem 4.1
system can be regarded as a non-minimum phase system with that limt →∞ sup Ω (t ) = ς under S2 . 
states

Φ , [Φ1 Φ2 · · · Φm ]⊤ (55) 5.4. Unmanned aerial vehicles team example

where the group Ni is controllable and linearizable, while other We are now taking an unmanned aerial vehicles (UAVs) team
groups of agents compose the ‘‘uncontrollable internal dynamics’’ example to illustrate the above results. Consider a team of 4
which may be unstable. UAVs divided into two groups. UAV 1 and UAV 3 are leaders that
determine the flying behavior of the group, while UAVs 2 and 4
have no behavior information by themselves. Such a system can be
5.3. Target aggregation via switching topology
naturally modeled by a multi-agent system with each UAV being
an agent [23].
As the topology switch among m candidate ones, the system
The UAV’s longitudinal differential equations under small at-
process naturally become a switched system with m modes where
tack angle can be expressed as [24]:
mode i corresponds to the topology i. The dynamics of group Ni can
ϑ̇i = ωi

be regarded as xi dynamics in Section 2. Assumption 2.1 is naturally
satisfied. ω̇i = Miω ωi + Miϑ cos(ϑi − αi ) + Mi ui i = 1, 2, . . . , 4.
Moreover, each Φi (i ∈ M ) is continuous, all modes share a
For UAV i, the two states ϑi and ωi denote respectively the pitch an-
common Lyapunov function
gle and the pitch rate, the input ui is the elevator deflection angle,
m
 αi denotes the attack angle. Miω , Miϑ and Mi are the longitudinal
Ω, Ωi . dynamics parameters.
i=1 Consider the ‘‘climbing’’ process which requires that all UAVs in
Next we provide two kinds of switching topologies. To avoid the team have the same pitch rates ω⋆ = 5 (°/s). The initial values
arbitrary fast switching, the ‘‘dwell time’’ τ is also involved. ωi (0) = 2 (°/s), for i = 1, 2, 3, 4. The dynamic equations of pitch
rates under appropriate controllers are:
Theorem 5.2 (Time-Dependent Switching Topology). Consider a UAV i : ω̇1 = −(ωi − ω⋆ ) + uci , i = 1, 3 (60)
multi-agent system (42) satisfying Assumption 5.1. limt →∞ Ω (t ) = ⋆
0, ∀Ωi (Φi (0)) ≤ δ0 , i ∈ M with any given δ0 > 0 if UAV i : ω̇i = 0.5(ωi − ω ) + uci , i = 2, 4. (61)
It can be seen that UAVs 1 and 3 know the prescribed pitch rate.
(1) Each group takes the cooperative law (45) and (53) under R.
⋆⋆ However, without cooperation, UAVs 2 and 4 may track their own
(2) There exist positive numbers η > 0, 1t ≥ τ such that eλ 1t ≤
⋆ pitch rates, ω2 and ω4 run far away from ω⋆ . Define Vi = (ωi −
η, and e−λ 1t ηm−1 ≤ e−λ̄m1t for λ̄ > 0. ω⋆ )2 , i = 1, 2, 3, 4, Assumption 5.1 is satisfied with λ0 = 2,
(3) m candidate topologies are switched following S1 . λ1 = 1.
1030 H. Yang et al. / Systems & Control Letters 61 (2012) 1023–1031

(a) Under S1 . (b) Under S2 .

Fig. 4. Trajectories of pitch rates.

The topology 1 for group 1 is built based on R : (1, 2) and and construction of these functions for general switched sys-
(2, 1). Let tems deserve further investigation.
ω1 − ω ⋆ 2. In this work, all states are available. Output-feedback control
uc1 = (−3V1 − 5V2 ), for k1 , k2 > 0 together with observer design would be considered in the ab-
2(ω1 − ω⋆ )2
sence of full state measurements.
uc2 = ω1 − ω2 . 3. The dynamics of each mode is supposed to be known and fixed.
Similarly, topology 2 for group 2 is arcs (3, 4) and (4, 3). Let Robust issue would be addressed in the presence of disturbance
or unknown parameters, etc.
ω3 − ω ⋆
uc3 = (−3V3 − 5V4 )
2(ω3 − ω⋆ )2
References
uc4 = (ω3 − ω4 ).

Theorem 5.1 guarantees that under topologies 1 and 2, groups 1 [1] D. Liberzon, Switching in Systems and Control, Birkhäuser, Boston, MA, 2003.

and 2 approach ω⋆ respectively. [2] J.P. Hespanha, D. Liberzon, D. Angeli, E.D. Sontag, Nonlinear norm-
observability notions and stability of switched systems, IEEE Transactions on
Suppose that only one group can be connected at one time due Automatic Control 50 (2) (2005) 154–168.
to the communication cost limitation. In this case, the switching [3] J. Zhao, D.J. Hill, On stability, L2 gain and H∞ control for switched systems,
among two topologies has to be applied. Automatica 44 (5) (2008) 1220–1232.
[4] P. Colaneri, J.C. Geromel, A. Astolfi, Stabilization of continuous-time switched
Define Ξ1 , [V1 V2 ]P [V1 V2 ]⊤ and Ξ2 , [V3 V4 ]P [V3 V4 ]⊤ with nonlinear systems, Systems & Control Letters 57 (1) (2008) 95–103.
7.5

1
P = 1 0.5
. Simple calculation yields [5] T.-T. Han, S.S. Ge, T.H. Lee, Persistent dwell-time switched nonlinear systems:
variation paradigm and gauge design, IEEE Transactions on Automatic Control
55 (2) (2010) 321–337.
Ξ̇1 ≤ −1.2Ξ1

[6] M. Cannon, M. Bacic, B. Kouvaritakis, Dynamic non-minimum phase compen-
Under topology 1 :
Ξ̇2 ≤ 0.4Ξ2 sation for SISO nonlinear, affine in the input systems, Automatica 42 (11)
(2006) 1969–1975.
Ξ̇2 ≤ −1.2Ξ2

[7] A. Isidori, A tool for semiglobal stabilization of uncertain non-minimum-
Under topology 2 :
Ξ̇1 ≤ 0.4Ξ1 . phase nonlinear systems via output feedback, IEEE Transactions on Automatic
Control 45 (10) (2000) 1817–1827.
One can find that group 1 (resp. 2) is uncontrollable under topology [8] S. Nazrulla, H.K. Khalil, Robust stabilization of non-minimum phase nonlinear
systems using extended high-gain observers, IEEE Transactions on Automatic
2 (resp. 1), which can be regarded as an unstable zero dynamics. Control 56 (4) (2011) 802–813.
Now we illustrate switching laws S1 and S2 . Choose dwell time [9] S.M. Hoseini, M. Farrokhi, A.J. Koshkouei, Robust adaptive control of nonlinear
τ = 0.5 (s), η = 1.5, 1t = 1 (s), all conditions of Theorem 5.2 hold. non-minimum phase systems with uncertainties, Automatica 47 (2) (2011)
348–357.
Fig. 4(a) shows the switching function and trajectories of 4 pitch
[10] M. Wang, G.M. Dimirovski, J. Zhao, H∞ control for a class of non-minimum-
rates under S1 , from which we can see that all pitch rates reach phase cascade switched nonlinear systems, in: Proc. of 2008 American Control
ω⋆ . All conditions of Theorem 5.3 also hold. Choose ϵ = υ = 0.2, Conference, Seattle, USA, 2008, pp. 5080–5085.
Fig. 4(b) shows the switching law and trajectories of 4 pitch rates [11] M. Oishi, C. Tomlin, Switching in nonminimum phase systems: applications to
a VSTOL aircraft, in: Proc. of 2000 American Control Conference, Chicago, USA,
under S2 ; all pitch rates still reach ω⋆ . The performance under S1 2000, pp. 487–491.
is better than that under S2 , this is because υ is chosen very small, [12] M. Benosman, K.Y. Lum, Output trajectory tracking for a switched nonlinear
which makes the decay speed of states under S2 slower. non-minimum phase system, in: Proc. of 16th IEEE International Conference
on Control Applications, Singapore, 2007, pp. 262–269.
[13] G.M. Xie, L. Wang, Periodic stabilizability of switched linear control systems,
6. Conclusion Automatica 45 (9) (2009) 2141–2148.
[14] H. Yang, B. Jiang, V. Cocquempot, H. Zhang, Stabilization of switched nonlinear
systems with all unstable modes: applications to multi-agent systems, IEEE
This paper discusses the stabilization problem for a class of Transactions on Automatic Control 56 (9) (2011) 2230–2235.
non-minimum phase switched nonlinear systems as well as its [15] A. Isidori, Nonlinear Control Systems, third ed., Springer Verlag, London, 1995.
application. Some further remarks are as follows: [16] Z.P. Jiang, Y. Wang, A generalization of the nonlinear small-gain theorem for
large-scale complex systems, in: Proc. of the 7th World Congress on Intelligent
Control and Automation, Chongqing, China, 2008, pp. 1188–1193.
1. The developed results rely on a series of Lyapunov-like func-
[17] H. Yang, V. Cocquempot, B. Jiang, On stabilization of switched nonlinear
tions that are established systematically in the target aggrega- systems with unstable modes, Systems & Control Letters 58 (10–11) (2009)
tion problem of multi-agent systems in Section 5. The existence 703–708.
 H. Yang et al. / Systems & Control Letters 61 (2012) 1023–1031 1031

[18] G. Shi, Y. Hong, Global target aggregation and state agreement of nonlinear topologies and communication delays, Systems & Control Letters 59 (6) (2010)
multi-agent systems with switching topologies, Automatica 45 (5) (2009) 340–348.
1165–1175. [22] H. Yang, M. Staroswiecki, B. Jiang, J. Liu, Fault tolerant cooperative control for a
[19] R.O. Saber, R.M. Murray, Consensus problems in networks of agents with class of nonlinear multi-agent systems, Systems & Control Letters 60 (4) (2011)
switching topology and time-delays, IEEE Transactions on Automatic Control 271–277.
49 (9) (2004) 1520–1533. [23] F. Giulietti, L. Pollini, M. Innocenti, Autonomous formation flight, IEEE Control
[20] J. Zhao, D. Hill, T. Liu, Synchronization of complex dynamical networks with Systems Magazine 20 (6) (2000) 34–44.
switching topology: a switched system point of view, Automatica 45 (11) [24] X. Mu, W. Zhang, W. Zhang, An adaptive backstepping design for longitudinal
(2009) 2502–2511. flight path control, in: Proc. of 7th World Congress on Intelligent Control and
[21] J. Yu, L. Wang, Group consensus in multi-agent systems with switching Automation, Chongqing, China, 2008, pp. 5249–5251.