This article was downloaded by: [Ryerson University]

On: 25 February 2013, At: 05:03

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Control

Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tcon20

Stability and stabilization of discrete time switched

systems
a b
J. C. Geromel & P. Colaneri
a
DSCE / School of Electrical and Computer Engineering, UNICAMP, CP 6101, 13083 – 970,
Campinas, SP, Brazil
b
Dipartimento di Elettronica e Informazione, Politecnico di Milano, Italy
Version of record first published: 20 Feb 2007.

To cite this article: J. C. Geromel & P. Colaneri (2006): Stability and stabilization of discrete time switched systems,
International Journal of Control, 79:07, 719-728

To link to this article: http://dx.doi.org/10.1080/00207170600645974

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to
anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should
be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims,
proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in
connection with or arising out of the use of this material.
 International Journal of Control
Vol. 79, No. 7, July 2006, 719–728

Stability and stabilization of discrete

time switched systems
J. C. GEROMEL*{ and P. COLANERIz

{DSCE / School of Electrical and Computer Engineering,

UNICAMP, CP 6101, 13083 – 970, Campinas, SP, Brazil
zDipartimento di Elettronica e Informazione,
Politecnico di Milano, Italy

(Received 29 April 2005; revised 8 February 2006; in final form 20 February 2006)
Downloaded by [Ryerson University] at 05:03 25 February 2013

This paper addresses two strategies for stabilization of discrete time linear switched
systems. The first one is of open loop nature (trajectory independent) and is based on the
determination of an upper bound of the minimum dwell time by means of a family of
quadratic Lyapunov functions. The relevant point on dwell time calculation is that the
proposed stability condition does not require the Lyapunov function be uniformly decreasing
at every switching time. The second one is of closed loop nature (trajectory dependent) and
is designed from the solution of what we call Lyapunov–Metzler inequalities from which
the stability condition is expressed. Being non-convex, a more conservative but simpler
to solve version of the Lyapunov–Metzler inequalities is provided. The theoretical results
are illustrated by means of examples.

1. Introduction Stability analysis of continuous time switched linear

systems has been addressed by several authors,
This paper aims at providing new results on stability Branicky (1998), Hockerman-Frommer et al. (1998),
analysis and stabilizing control synthesis for a discrete Johansson and Rantzer (1998), Ye et al. (1998),
time switched linear system of the following general Liberzon (2003) and Hespanha (2004), among others.
form The recent paper (Hespanha 2004), dealing with
extensions of LaSalle’s Invariance Principle, provides
xðk þ 1Þ ¼ A
ðkÞ xðkÞ, xð0Þ ¼ x0 ð1Þ an interesting discussion on a collection of results
on uniform stability of switched systems. For discrete-
defined for all k 2 N where xðkÞ 2 Rn is the state,
(k) time switched systems, see Daafouz and Bernussou
is the switching rule and x0 is the initial condition. (2001), Zhai (2001) and Xie and Wang (2003).
Considering a set of matrices Ai 2 Rn
n , i ¼ 1, . . . , N Liberzon and Morse (1999) and DeCarlo et al. (2000)
be given, the switching rule
(k), for each k 2 N, provide excellent overviews on switched and hybrid
is such that systems.
Generally speaking, when
ðÞ is state independent,
A
ðkÞ 2 fA1 , . . . , AN g: ð2Þ that is, when it is a a priori piecewise constant signal,
if one considers that each matrix of the set
Clearly, (2) imposes that A
ðkÞ is constrained to jump fA1 , . . . , AN g is asymptotically stable then, intuitively,
among the N vertices of the matrix polytope there exists
(k) such that (1) is asymptotically
fA1 , . . . , AN g. stable provided the interval of time between successive
switching (dwell time) is taken large enough. The
main novelty reported in this paper is on the determina-
*Corresponding author. Email: geromel@dsce.fee.unicamp.br tion of an upper bound of the minimum dwell time
International Journal of Control
ISSN 0020–7179 print/ISSN 1366–5820 online
2006 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/00207170600645974
 720 J. C. Geromel and P. Colaneri
preserving global stability. This is accomplished from
the determination of a family of quadratic Lyapunov
functions, associated to the correspondent matrix of
the set fA1 , . . . , AN g such that at every switching time
lq the sequence vðxðlq ÞÞ, for q ¼ 0, . . . , 1, converges
uniformly to zero. From a simple illustrative example,
it is shown that our procedure for the determination
of an upper bound of the minimum dwell time, provides
more accurate results than the ones available in the
literature, in particular the extension to discrete time
systems of Morse (1996) and the average dwell time of
Zhai et al. (2002). In some instances, the determination
of the minimum dwell time, based on a quadratic
guaranteed cost, is related to the results of Xu and
Antsaklis (2004) assuming further that the switching
rule is not a priori given but can be taken arbitrarily,
among the feasible ones.
Downloaded by [Ryerson University] at 05:03 25 February 2013
For switched systems with
ðÞ being state dependent,
the stability condition is expressed in terms of a set of
inequalities that we call Lyapunov–Metzler inequalities
because the variables involved are a set of symmetric
and positive matrices fP1 , . . . , PN g and a Metzler
matrix
. The point to be noted is that our asymptotic
stability condition does not require any stability
property associated to each individual matrix of the
set fA1 , . . . , AN g and it contains as special cases the
quadratic stability condition and the results of
Daafouz and Bernussou (2001) and Zhai (2001). The
price to be payed, however, is the non-convex nature
of the the Lyapunov–Metzler inequalities being thus
difficult to solve numerically. For this reason, a more
conservative but easier to solve asymptotical stability
condition is proposed.
This paper, which is the discrete-time counterpart of
Geromel and Colaneri (2005), is organized as follows.
In x 2 time switching control is considered. The switch-
ing rule
(k) is considered piecewise constant and an
upper bound of the minimum dwell time is determined.
Section 3 is entirely devoted to state switching control
where the goal is to design a proper function uðÞ such
that the system (1) is globally asymptotically stable
with the switching rule
ðkÞ ¼ uðxðkÞÞ. Finally,
x 4 concludes the paper.
The notation used throughout is standard. Capital
letters denote matrices, small letters denote vectors and
small Greek letters denote scalars. For matrices or
vectors ð0 Þ indicates transpose. For symmetric matrices,
X > 0 (0) indicates that X is positive definite (non-
negative definite) and min ðX Þ, max ðX Þ denote
its minimum and maximum eigenvalue respectively.
The sets of real and natural numbers are denoted
by R and N respectively. The L2 squared norm of
n 2
P1 2 R 0 defined for all k 2 N equals kxðkÞk2 ¼
xðkÞ
k¼0 xðkÞ xðkÞ.
2. Time switching control
This section is entirely dedicated to design a time switch-
ing control law for the switched linear system defined
by the model (1) and (2) where it is assumed that each
matrix of the set fA1 , . . . , AN g is asymptotically stable.
The problem under consideration can be stated as
follows: Determine a minimum dwell time  2 N
such that the equilibrium point x ¼ 0 of the system (1)
is globally asymptotically stable with the time switching
control

ðkÞ ¼ i 2 f1, . . . , Ng, k 2 ½lq , lqþ1 Þ, ð3Þ
where lq and lqþ1 are successive switching times satis-
fying lqþ1  lq    1 for all q 2 N. It is interesting
to observe that the index i 2 f1, . . . , Ng selected at
each instant of time k 2 N is arbitrary. Hence, asym-
ptotic stability is preserved whenever it remains
unchanged for a period of time greater or equal to
the minimum dwell time  . The next theorem
provides the theoretical basis towards a possible
solution of the proposed problem by characterizing
an upper bound for  . It uses the concept of
multiple Lyapunov function with the innovation that
the classical assumption on its decreasing at switching
times is no longer needed.
Theorem 1: Assume that, for some   1, there exists
a collection of positive definite matrices fP1 , . . . , PN g
of compatible dimensions such that
A0i Pi Ai  Pi < 0, 8i ¼ 1, . . . , N ð4Þ
and
A0 
i Pj Ai  Pi < 0, 8i 6¼ j ¼ 1, . . . , N: ð5Þ
The time switching control (3) with lqþ1  lq  
makes the equilibrium solution x ¼ 0 of (1) globally
asymptotically stable.
Proof: Consider, in accordance to (3), that
ðkÞ ¼
i 2 f1, . . . , Ng for all k 2 ½lq , lqþ1 Þ where lqþ1 ¼ lq þ q
with q    1 and that at k ¼ lqþ1 the time switching
control jumps to
ðkÞ ¼ j 2 f1, . . . , Ng, otherwise the
result trivially follows. From (4), it is seen that, for all
k 2 ½lq , lqþ1 Þ, the Lyapunov function vðxðkÞÞ ¼
xðkÞ0 P
ðkÞ xðkÞ, along an arbitrary trajectory of (1)
satisfies
vðxðk þ 1ÞÞ ¼ vðxðkÞÞ þ xðkÞ0 ðA0i Pi Ai  Pi ÞxðkÞ
< vðxðkÞÞ ð6Þ
 Stabilization of discrete time switched systems
which enables us to conclude that there exist scalars
 2 ð0, 1Þ and  > 0 such that
kxðkÞk22  ðklq Þ vðxðlq ÞÞ, 8k 2 ½lq , lqþ1 Þ:
On the other hand, using the inequalities (5) we have
0 p¼1
vðxðlqþ1 ÞÞ ¼ xðlqþ1 Þ Pj xðlqþ1 Þ
0 
¼ xðlq Þ0 Ai q Pj Ai q xðlq Þ
0 0ð Þ ð Þ
< xðlq Þ Ai q Pi Ai q xðlq Þ
< xðlq Þ0 Pi xðlq Þ
< vðxðlq ÞÞ,
where the second inequality holds from the fact that
for every  ¼ q   2 N it is true that A0i Pi Ai  Pi .
Downloaded by [Ryerson University] at 05:03 25 February 2013
The consequence is that there exists  2 ð0, 1Þ such that
vðxðlq ÞÞ  q vðx0 Þ, 8q 2 N
which together with (7) implies that the equilib-
rium solution x ¼ 0 of (1) is globally asymptotically
stable. œ 1, P1 >0,..., PN >0
This result deserves some comments. First, it is simple to
determine the scalars ,  and  such that inequalities
(7) and (9) hold. Assuming that positive definite
matrices fP1 , . . . , PN g satisfy the inequalities (4) then
there exists  > 0 such that A0i Pi Ai  Pi  I for all
i ¼ 1, . . . , N yielding  ¼ 1  = maxi max ðPi Þ and
 ¼ 1= mini min ðPi Þ. Furthermore, from (5) there
exists 0   < 1 such that A0  sets of problems expressed in terms of LMIs must be
i Pj Ai  Pi for all
i 6¼ j ¼ 1, . . . , N leading to vðxðlqþ1 ÞÞ  vðxðlq ÞÞ and
consequently (9). Second, since it is assumed that all
matrices of the set fA1 , . . . , AN g are asymptotically
stable then the constraints (4) are always feasible and
the constraints (5) are satisfied whenever  2 N is
taken large enough. Third, assume that matrices
A1 , . . . , AN are quadratically stable, which is the same
to say that they share a positive definite matrix P
such that
A0i PAi  P < 0, 8i ¼ 1, . . . , N:
In this case, the inequality (5) is satisfied for P1 ¼    ¼
PN ¼ P for all   1 meaning that the switching
policy (3) may jump from i to j at each instant of time
k 2 N, preserving once again asymptotical stability.
Hence, Theorem 1 contains, as a particular case, the
quadratic stability condition. It is interesting to mention
that Theorem 1 also contains as a particular case
the stability results of Daafouz and Bernussou (2001)
which are less conservative than quadratic stability.
721
Finally, with   1, it is always possible to define a
time switching control strategy (3) such that A
ðkÞ is
periodic. As a consequence, a necessary condition for
ð7Þ the feasibility of those constraints is

!


YN




ðÞ :¼ max
q Bp
< 1, ð11Þ
q¼1,..., n


where qðÞ denotes a generic eigenvalue of ðÞ
and fB1 , . . . , BN g are matrices corresponding to any
permutation among those of the set fA1 , . . . , AN g.
However, since (3) may produce non-periodic policies
ð8Þ as well, the necessary condition (11) for the existence
of a feasible solution to inequalities (4) and (5), generally
does not meet sufficiency. In the sequel, this aspect
will be illustrated by means of an example.
In this setting, an upper bound for the minimum
dwell time  can be computed by taking the minimum
value of  satisfying the constraints of Theorem 1.
ð9Þ Hence, it can be calculated with no great difficulty
from the optimal solution of the optimization problem
 
min : ð4Þ  ð5Þ ð12Þ
which, for   1 fixed, reduces to a convex program-
ming feasibility problem with linear matrix inequalities
constraints that can be handled by any LMI solver avail-
able in the literature to date, see Boyd et al. (1994) for an
important study on systems and LMIs. (Problem (12)
should be stated with inf instead of min. All feasible
considered closed from the interior within a precision
defined by the user.)
Finally, we generalize the result of Theorem 1 to
define a guaranteed cost to go from any initial condition
to the origin. The next result is related to the time
switching strategy (3) where it is further assumed that
"  lqþ1  lq   holds for all q 2 N for some
"    1. From Theorem 1 asymptotic stability is
preserved whenever    .
Theorem 2: Let Q  0 2 Rn
n and "    1 be given.
ð10Þ Define the set of positive semidefinite matrices
X
"1
Ri :¼ A0k k
i QAi , i ¼ 1, . . . , N: ð13Þ
k¼0
Assume that there exists a collection of positive definite
matrices fP1 , . . . , PN g of compatible dimensions such that
A0i Pi Ai  Pi þ Q < 0, 8i ¼ 1, . . . , N ð14Þ
 722 J. C. Geromel and P. Colaneri

and guaranteed cost is obtained whenever " is chosen as

small as possible. Furthermore, with the possible

A0
i Pj Ai  Pi þ Ri < 0, 8i 6¼ j ¼ 1, . . . , N: ð15Þ choice " ¼ þ1, from (16) it is concluded that the
proposed switching rule (3) with lqþ1  lq  
The time switching control (3) with "  lqþ1  lq  
makes the trajectory of yðkÞ ¼ Q1=2 xðkÞ quadratically
makes the equilibrium solution x ¼ 0 of (1) globally
convergent. Theorem 2 recovers once again the extreme
asymptotically stable and
situations " ¼  ¼ 1 for which no jump occurs and
feasibility is assured with
X
1
xðkÞ0 QxðkÞ < x00 P
ð0Þ x0 ð16Þ
k¼0 X
1
k
Pi ¼ A0k
i ðQ þ "I ÞAi > Ri  0, i ¼ 1, . . . , N ð20Þ
k¼0
Proof: Since matrices Ri defined in (13) are positive
semidefinite then for the same   1, the feasibility of
for all " > 0 and (16) becomes a well known result for
constraints (14) and (15) implies that the conditions
" > 0 taken arbitrarily small. On the other hand, for
(4) and (5) are also satisfied. Hence, from Theorem 1,
 ¼ 1 and any "  1 feasibility holds whenever the
system (1) is globally asymptotically stable. In addition,
set of matrices fA1 , . . . , AN g admits a common or a
from the definition (13) and the inequalities (14) and (15)
linearly dependent parameter Lyapunov function, see
we see that Pi > Ri and
Downloaded by [Ryerson University] at 05:03 25 February 2013

Morse (1996).
For illustration purpose of the theoretical results
A0i ðPi  Ri ÞAi  ðPi  Ri Þ
obtained so far, let us consider the following example
< Q  A0i Ri Ai þ Ri characterized by N ¼ 2, matrices A1 ¼ eB1 T and A2 ¼
X
" X
"1 eB2 T where
k k
< Q  A0k
i QAi þ A0k
i QAi
k¼1 k¼0   
0 1 0 1
< Ai QA"i
0" B1 ¼ , B2 ¼ ð21Þ
10 1 0:1 0:5
<0 ð17Þ
and T ¼ 0.5 which corresponds to get samples from
holds for all i ¼ 1, . . . , N which implies that two continuous-time systems at each T seconds. First,
A0i ðPi ; Ri ÞAi < ðPi  Ri Þ for every  2 N. from (12), we have calculated the upper bound of the
Consequently, taking into account the switching minimum dwell time as being   6. To give an idea
strategy (3) with lqþ1  lq ¼ q   and the inequality of its conservativeness we have calculated from the
(15) one gets plot of figure 1 the value per ¼ 6 corresponding to
the necessary condition for stability (11), arising from
vðxðlqþ1 ÞÞ ¼ xðlqþ1 Þ0 Pj xðlqþ1 Þ linear periodic systems. Both being equal indicates,
0q  q  for this simple example, that the minimum dwell time
< xðlq Þ0 Ai ðPi  Ri ÞAi xðlq Þ
< xðlq Þ0 ðPi  Ri Þxðlq Þ 1.5
< vðxðlq ÞÞ  xðlq Þ0 R
ðlq Þ xðlq Þ ð18Þ

which summing up for all q 2 N and taking into account

that "  lqþ1  lq yields 1

X
1 1 lqþ1
X X 1
q (∆)

0 0ðklq Þ ðklq Þ
xðkÞ QxðkÞ ¼ xðlq Þ0 Ai QAi xðlq Þ
k¼0 q¼0 k¼lq
X
1 0.5
 xðlq Þ0 R
ðlq Þ xðlq Þ
q¼0

< vðx0 Þ, ð19Þ

0
thus proving the proposed theorem. œ 0 2 4 6 8 10 12 14 16 18 20
∆
We observe that the conditions of Theorem 2 are
feasible if and only if "     and a more accurate Figure 1. The function ðÞ.
 Stabilization of discrete time switched systems
is  ¼ 6. On the other hand, for comparison purpose
we have applied an extension to discrete-time systems
of the classical result of Morse (1996) for the determina-
tion of an alternative upper bound for the minimum
dwell time, i.e.  ¼ ½T , where T  maxi¼1,..., N fTi g
and
 
logðÞ
Ti ¼ inf  : kAki k < k ,
>0, 0<<1 logðÞ
For matrices in (21) we have numerically determined
T1 ¼ 13:56 and T2 ¼ 4:66 yielding an estimation for
the minimum dwell time as being   14. Hence,
in this particular example, the result provided by the
solution of problem (12) is much more precise but at
expense of a more expressive computational effort.
Moreover, using the method in Zhai et al. (2002), we
Downloaded by [Ryerson University] at 05:03 25 February 2013
have obtained the upper bound of the so-called average
dwell time yielding again 14 (the average dwell time
is obviously greater than that obtained by our method).
Figure 2 has been constructed by simulation of
system (1) with the time switching control (3), lqþ1 
lq ¼ 6, initial condition x0 ¼ ½1 1 0 ,
ð0Þ ¼ 2 and
Q ¼ I. The family of Lyapunov functions has been
determined from the optimal solution of the convex
programming problem
 
min max x00 Pi x0 : ð14Þ  ð15Þ
P1 >0,..., PN >0 i¼1,..., N
which puts in evidence that, from the result of
Theorem 2 a guaranteed cost can be determined
for the worst case as far as the initial value of
ð0Þ is
concerned. For " ¼  ¼ 6 we have obtained the asso-
ciated minimum guaranteed cost equal to  ¼ 152:17,
valid for any logical initial state. As commented
150
100
v (x(k))
50
ij , such that
0
0 5 10 15 20
k X
ij  0,
ij ¼ 1, 8i, j: ð27Þ
Figure 2. The Lyapunov function.
723
before, the Lyapunov function vðxðkÞÞ ¼ xðkÞ0 P
ðkÞ xðkÞ
goes to zero as k goes to infinity however, it is not
uniformly decreasing with respect to time. In figure 2,
due to the stability conditions of Theorem 2, the
points marked with ‘‘o’’, defines a globally conver-
gent sequence vðxðlq ÞÞ, for all q 2 N. Solving again
problem (23) but for " ¼ þ1 and  ¼ 6 the minimum
cost increases to  ¼ 229:01 as a consequence of allow-
8k  0 : ð22Þ
ing a more flexible switching rule (3) with lqþ1  lq  6.
This assures the global convergence of the sequence
marked in dashed line. This is a clear improvement
on stability conditions, minimum dwell time and
guaranteed cost calculations when compared to the
results available in the literature to date.
3. State switching control
In this section we consider again the system (1) where
the switching rule satisfies (2). The main difference
from the previous section is that, presently, it is assumed
that the state vector x(k) is available for feedback for all
k 2 N and our goal is to determine the function
uðÞ: Rn ! f1, . . . , Ng, such that

ðkÞ ¼ uðxðkÞÞ ð24Þ
ð23Þ
makes the equilibrium point x ¼ 0 of (1) globally asymp-
totically stable. We do not assume that the matrices
of the set fA1 , . . . , AN g are asymptotically stable.
To this end, let us define the simplex
( )
X
N
 :¼  2 R : N
i ¼ 1, i  0 ð25Þ
i¼1
which together with the set of positive definite matrices
fP1 , . . . , PN g enables us to introduce the following
non-quadratic Lyapunov function
!
X
N
0 0
vðxÞ :¼ min x Pi x ¼ min i x Pi x : ð26Þ
i¼1,..., N 2
i¼1
As it will be clear in the sequel, this Lyapunov function
is crucial to our purpose. Before proceeding, let us
recall the class of Metzler matrices denoted by M and
constituted by all matrices
2 RN
N with elements
25 30
N
i¼1
 724 J. C. Geromel and P. Colaneri
It is clear that any
2 M is such that
presents
an unitary eigenvalue since c 0
¼ c 0 where
c 0 ¼ ½1    1. In addition, it is well known that the
eigenvector associated to the unitary eigenvalue of

is non-negative yielding the conclusion that there
always exists 1 2  such that
1 ¼ 1 . The next
theorem summarizes the main result of this section.
Theorem 3: Assume that there exist
2 M and a set
of positive definite matrices fP1 , . . . , PN g satisfying the
Lyapunov–Metzler inequalities
! as a particular case (since the Lyapunov–Metzler
X
N
A0i
ji Pj Ai  Pi < 0, i ¼ 1, . . . , N ð28Þ inequalities do not depend on
anymore) the quadratic
j¼1
The state switching control (24) with
Downloaded by [Ryerson University] at 05:03 25 February 2013
uðxðkÞÞ ¼ arg min xðkÞ0 Pi xðkÞ
i¼1,..., N
makes the equilibrium solution x ¼ 0 of (1) globally
asymptotically stable.
Proof: It follows from the Lyapunov function (26).
Assume, in accordance to (29), that at an arbitrary
k 2 N, the state switching control is given by

ðkÞ ¼ uðxðkÞÞ ¼ i for some i 2 f1, . . . , Ng. Hence,
vðxðkÞÞ ¼ xðkÞ0 Pi xðkÞ and from the system dynamic
equation (1) we have
vðxðk þ 1ÞÞ ¼ min xðkÞ0 A0i Pj Ai xðkÞ
j¼1,..., N
!
X N
¼ min xðkÞ0 A0i j Pj Ai xðkÞ
2
j¼1
! for any initial state x(0) and any initial probability
X
N
 xðkÞ0 A0i
ji Pj Ai xðkÞ, ð30Þ pattern 0 2 , it has been shown (see e.g. Costa and
j¼1
where the inequality holds from the fact that each
column of
belongs to . Finally, using the
Lyapunov–Metzler inequalities (28) one gets
vðxðk þ 1ÞÞ < xðkÞ0 Pi xðtÞ
< vðxðkÞÞ
which proves the proposed theorem since the
Lyapunov function vðxðkÞÞ defined in (26) is radially
unbounded. œ
It is important to observe that Theorem 3 does not
require that the set of matrices fA1 , . . . , AN g be
composed exclusively by asymptotically stable matrices.
Indeed, with
2 M, a necessary condition for
the Lyapunov–Metzler inequalities to be feasible with
respect to fP1 , . . . , PN g is matrices
1=2ii Ai for all then the Lyapunov–Metzler inequalities (28) admit a
i ¼ 1, . . . , N be asymptotically stable. Since 0 
ii  1 solution.
this condition does not imply on the asymptotical
stability of Ai. However, an interesting case occurs
when all matrices fA1 , . . . , AN g are asymptotically
stable for which the choice
¼ I is possible and the
state switching strategy proposed preserves stability.
Furthermore, if the set fA1 , . . . , AN g is quadratically
stable then the Lyapunov–Metzler inequalities admit
a solution P1 ¼    ¼ PN ¼ P. In this classical but
particular case, at any k 2 N, the control law uðxðkÞÞ ¼
i 2 f1, . . . , Ng is arbitrary and asymptotical stability
is once again guaranteed. Hence, Theorem 3, contains
stability condition.
In the literature, the Lyapunov–Metzler inequalities
with
2 M fixed, have been introduced in order to
study the mean-square (MS) stability of Markov Jump
Linear Systems (MJLS). In that context, the Metzler
ð29Þ
matrix
¼
0 2 M is given and
00 represents the
transition matrix of a Markov chain
(k) governing
the dynamical system (1). In this respect, each
component of the vector ðkÞ 2  is the probability of
the Markov chain to be on the ith logical state and
obeys the linear equation
ðk þ 1Þ ¼
0 ðkÞ, ð0Þ ¼ 0 2 , ð32Þ
where the eigenvector 1 2  associated to the unitary
eigenvalue of
0 represents the stationary probability
vector. Hence, using the fact that the stochastic system
under consideration is said to be MS-stable if
lim EðkxðkÞk2 Þ ¼ 0 ð33Þ
k!þ1
Fragoso (1993)) that the system is MS-stable if and
only if there exists a set of positive definite matrices
fP1 , . . . , PN g satisfying the Lyapunov–Metzler inequal-
ities (28) for
¼
0 . Numerically speaking, this
particular case of the Lyapunov–Metzler inequalities
is simple to solve, since (28) reduces to a set of linear
matrix inequalities.
ð31Þ A relevant point presented in the next lemma concerns
the existence of a solution of the Lyapunov–Metzler
inequalities (28) with respect to the variables
2 M
and fP1 , . . . , PN g.
Lemma 1: Suppose that there exist  2  and P > 0
such that
X
N
i ðA0i PAi  PÞ < 0 ð34Þ
i¼1
 Stabilization of discrete time switched systems 725
P
Proof: The proof is constructive. We show that if (34) since P ¼ N i¼1 i Pi is positive definite and multiplying
holds then matrices (28) by i  0, i ¼ 1, . . . , N and summing up we get

Pi ¼ A0i PAi þ "I,
¼    ð35Þ X
N
P > i A0i P Ai > A0 P A ð40Þ
i¼1
with " > 0 sufficiently small solve (28). Indeed, assuming
that (34) holds then there exists " > 0 sufficiently small
The existence of this particular solution to the inequal-
such that its left hand side is less or equal to "I.
ities (28) meets exactly the stability condition provided
From the fact that  2  we get
in Zhai (2001). To see this, in our present context,
let us assume that there exist  ¼ 1 2  and
X
N X
N
P > 0 such that (34) holds. Hence, the switching

ji Pj ¼ j A0j PAj þ "I  P, i ¼ 1, . . . , N ð36Þ
j¼1 j¼1
rule (24) with

which yields uðxðkÞÞ ¼ arg min xðkÞ0 A0i PAi xðkÞ ð41Þ
i¼1,..., N

!
Downloaded by [Ryerson University] at 05:03 25 February 2013

ji Pj Ai  A0i PAi < Pi ,
X
N
makes the equilibrium point x ¼ 0 of the switched
A0i
i ¼ 1, . . . , N ð37Þ
j¼1
system (1) globally asymptotically stable. Indeed,
considering the Lyapunov function vðxðkÞÞ ¼
xðkÞ0 PxðkÞ we have
and thus the lemma is proved from the observation that
the matrices defined in (35) are such that Pi > 0 for all
i ¼ 1, . . . , N and
is a Metzler matrix. œ vðxðk þ 1ÞÞ ¼ xðkÞ0 A0
ðkÞ PA
ðkÞ xðkÞ

Invoking the fact that P > 0 and  2 , the condi- ¼ min xðkÞ0 A0i PAi xðkÞ
i¼1,..., N
tion (34) whenever satisfied, implies that A ¼
P !
N X
N
i¼1 i Ai is asymptotically stable. Indeed, multiplying ¼ min xðkÞ 0
i A0i PAi xðkÞ
the inequality 2
i¼1
!
    X
N
A0i PAi A0i P A0i   xðkÞ 0
1i A0i PAi xðkÞ
¼ P Ai I 0 ð38Þ
PAi P I i¼1

< vðxðkÞÞ: ð42Þ

by i  0, summing up for all i ¼ 1, . . . , N, performing
the Schur Complement of the result with respect to the
It is important to keep in mind that if the set of matrices
second row and column and using (34) we have
fA1 , . . . , AN g does not admit A asymptotically stable
for some  2  then the above stabilizing switching
X
N
rule cannot be determined. Even if it is known that
A0 PA  i A0i PAi < P: ð39Þ
i¼1
there exists  2  such that A is asymptotically stable,
this information is of little use since, as commented
before, the condition (34) of Lemma 1 may still be
Unfortunately, the converse does not hold in general,
infeasible. Furthermore, the numerical determination,
that is, the existence of a stable convex combination
if any, of  2  and P > 0 satisfying (34) is not a
does not suffice for (34) to be true. However, we
simple task due to its non-linear nature. The
conjecture that the existence of a stable combination
Lyapunov–Metzler inequalities introduced in Theorem
suffices to guarantee the existence of a solution to the
3 suffer the same difficulty but fortunately a simple
Lyapunov–Metzler inequalities. Although this point
numerical procedure based on line search can be settled
will not be further addressed our claim is based on the
to determine its solution. This aspect will be considered
proof of Lemma 1 where only a Metzler matrix with a
next. First, let us introduce a guaranteed quadratic
very particular structure has been used to construct a
cost associated to the proposed state switching
feasible solution. Furthermore, notice that if the
control law (29).
Lyapunov–Metzler inequalities (28) admit a
solution with
2PM of the form (35) for some Lemma 2: Let Q  0 be given. Assume that there exist a
 2  then A ¼ N i¼1 i Ai is asymptotically stable set of positive definite matrices fP1 , . . . , PN g and
2 M
 726 J. C. Geromel and P. Colaneri

satisfying the Lyapunov–Metzler inequalities Lyapunov–Metzler inequalities

!
X
N A0i ðPi þ ð1  ÞPj ÞAi  Pi þ Q < 0, j 6¼ i ¼ 1, . . . , N:
A0i
ji Pj Ai  Pi þ Q < 0, i ¼ 1, . . . , N ð43Þ
j¼1
ð47Þ

The state switching control (24) with uðxðkÞÞ given by (29) The state switching control (24) with uðxðkÞÞ given by
makes the equilibrium solution x ¼ 0 of (1) globally (29) makes the equilibrium solution x ¼ 0 of (1) globally
asymptotically stable and asymptotically stable and

X
1 X
1 X
N
0
xðkÞ QxðkÞ < min x00 Pi x0 : ð44Þ xðkÞ0 QxðkÞ < x00 Pi x0 : ð48Þ
i¼1,..., N k¼0 i¼1
k¼0

Proof: The proof follows from the choice of
2 M

Proof: It has the same pattern as the proof of such that
ii ¼  and the remaining elements satisfying
Theorem 3. The Lyapunov function (26) and the
Lyapunov–Metzler inequalities (43) yield
X
N
Downloaded by [Ryerson University] at 05:03 25 February 2013

ð1  Þ1
ji ¼ 1 ð49Þ
vðxðk þ 1ÞÞ  vðxðkÞÞ < xðkÞ0 QxðkÞ ð45Þ j6¼i¼1

which gives for all i ¼ 1, . . . , N. Taking into account that

ji  0
for all j 6¼ i ¼ 1, . . . , N, multiplying (47) by
ji , summing
X
l up for all j 6¼ i ¼ 1, . . . , N and finally multiplying the
vðxðl þ 1ÞÞ  vðxð0ÞÞ ¼ ðvðxðk þ 1ÞÞ  vðxðkÞÞÞ result by ð1  Þ1 > 0 we get
k¼0
X
l !
X
N
< xðkÞ0 QxðkÞ, 8l  0 ð46Þ Pi  Q > A0i Pi þ
ji Pj Ai
k¼0
j6¼i¼1
!
X
N
proving thus the proposed lemma since due to the > A0i
ji Pj Ai ð50Þ
asymptotical stability, vðxðl ÞÞ goes to zero as l goes to j¼1
infinity. œ

The numerical determination, if any, of a solution of
the Lyapunov–Metzler inequalities with respect to the
variables ð
, fP1 , . . . , PN gÞ is not a simple task
and certainly deserves additional attention. The main
source of difficulty stems from its non-convex nature
due to the products of variables and so LMI
solvers do not apply. Perhaps, a point to be
further investigated is that its particular structure
with
ji  0 being scalars may help on the design of an matrices fA1 , . . . , AN g still is not required. In addition,
interactive method based on relaxation.
In this paper we pursue another route. The main idea
is to get a simpler, although certainly more conservative
stability condition that can be expressed by means
of LMIs being thus solvable by the machinery available
in the literature to date. The next theorem shows
that working with a subclass of Metzler matrices,
characterized by having the same diagonal elements,
this goal is accomplished.
Theorem 4: Let Q  0 be given. Assume that there
exist a set of positive definite matrices fP1 , . . . , PN g
and a scalar 0   < 1 satisfying the modified
which being valid for all i ¼ 1, . . . , N are the
Lyapunov-Metzler inequalities (43). From Lemma 2,
the upper bound (44) holds which trivially implies
that (48) is verified. The proposed theorem is thus
proved. œ
The basic theoretical features of Theorem 3 and
Lemma 2 are still present in Theorem 4. The most
important is that the asymptotic stability of the set of
notice that the guaranteed cost (48) is clearly worse
than the one provided by Lemma 2 but the former
being convex makes possible to solve the problem
( )
X
N
min x00 Pi x0: ð47Þ ð51Þ
0<<1, P1 >0,..., PN >0
i¼1
by LMI solvers and line search.
The next example illustrates some aspects of
the theoretical results obtained so far. Consider the
system (1) with N ¼ 2 and matrices A1 ¼ eB1 T ,
 Stabilization of discrete time switched systems 727

× 104 1.5
1.6

1
1.4

0.5
1.2

x (k)
0
d (γ)

−0.5
0.8

−1
0.6

−1.5
0.4 0 10 20 30 40 50 60 70 80 90 100
0 0.005 0.01 0.015 0.02 0.025 k
γ
Figure 4. Time simulation of the state switching control.
Downloaded by [Ryerson University] at 05:03 25 February 2013

Figure 3. Guaranteed cost as a function of .

obtained from the optimal solution of problem (51).

A2 ¼ e B2 T
where As it can be seen, the proposed control strategy is
effective to stabilize the system under consideration.
   
0 1 0 1
B1 ¼ , B2 ¼ ð52Þ
2 9 2 8
4. Conclusion
and T ¼ 0.1. As it can be verified matrices A1 and A2
In this paper we have introduced stability conditions
are both unstable but for  ¼ ½0:69 0:310 the matrix
for switched discrete-time linear systems. They have
A obtained by convex combination is stable. For
been used for control synthesis of state independent
these matrices it has also been verified numerically
(open loop) and state dependent (closed loop) switching
that the condition of Lemma 1 does not hold. Even
rules. In both cases, the determination of a guarant-
though, in this particular case, the modified version of
eed cost associated to each control strategy has been
the Lyapunov–Metzler inequalities (47) have a solution
addressed. Special attention has been devoted
which supports our claim on this point discussed
towards the numerical solvability of the design problems
before. Considering Q ¼ I and the initial condition
by means of methods based on linear matrix
x0 ¼ ½1 10 , problem (51) has been solved by line
inequalities.
search fixing  and minimizing its objective function,
Three issues deserve more attention. The first is on
denoted by ðÞ, with respect to the remaining variables.
the theoretical comparison of dwell time calculation
Figure 3 shows the behavior of the function ðÞ
with the results available in the literature to date. The
which enables us to determine its minimum value
second is related to the development of numerical
 ¼ 4:81
103 , corresponding to   ¼ 6
103 . It is
algorithms for the solution of the introduced
important to stress that, in this particular example, the
Lyapunov–Metzler inequalities, taking into account
function ðÞ has an unique minimum. However, we
their special structure. The third one is the possible gen-
do not have any evidence that this is a generic property
eralization of the stability conditions to cope with linear
valid in all cases. Figure 4 shows the trajectories
control design and guaranteed quadratic cost. Viewed
of the state variable xðkÞ 2 R2 versus time for the
the nonlinear nature of the involved stability condi-
system controlled by the state switching rule
ðkÞ ¼
tions, these points constitute a real theoretical challenge.
uðxðkÞÞ given by (29) with the positive definite matrices
" #
0:3009 0:3312
P1 ¼
103 , Acknowledgements
0:3312 0:4287
" # ð53Þ
0:1341 0:4560 This research was supported by grants from
3
P2 ¼
10 ‘‘Conselho Nacional de Desenvolvimento Cientı́fico e
0:4560 2:3795
Tecnológico – CNPq’’ – Brazil.
 728 J. C. Geromel and P. Colaneri
References
S.P. Boyd, L. El Ghaoui, E. Feron and V. Balakrishnan, Linear
Matrix Inequalities in System and Control Theory, Philadelphia:
SIAM, 1994.
M.S. Branicky, ‘‘Multiple Lyapunov functions and other analysis
tools for switched and hybrid systems’’, IEEE Trans. Automat.
Contr., 43, pp. 475–482, 1998.
O.L.V. Costa and M.D. Fragoso, ‘‘Stability results for discrete-time
linear systems with Markovian jumping parameters’’, Journal of
Mathematical Analysis and Applications, 179, pp. 154–178, 1993.
R.A. DeCarlo, M.S. Branicky, S. Pettersson and B. Lennartson,
‘‘Perspectives and results on the stability and stabilizability of
hybrid systems’’, Proceedings of the IEEE, 88, pp. 1069–1082, 2000.
J. Daafouz and J. Bernussou, ‘‘Parameter dependent Lyapunov
functions for discrete time systems with time varying parametric
uncertainties’’, Systems & Control Letters, 43, pp. 355–359, 2001.
J.C. Geromel and P. Colaneri, ‘‘Stability and stabilization of
continuous time switched systems’’, submitted for publication, 2005.
J.P. Hespanha, ‘‘Uniform stability of switched linear systems:
extensions of LaSalle’s principle’’, IEEE Trans. Automat. Contr.,
49, pp. 470–482, 2004.
Downloaded by [Ryerson University] at 05:03 25 February 2013
J. Hockerman-Frommer, S.R. Kulkarni and P.J. Ramadge,
‘‘Controller switching based on output predictions errors’’,
IEEE Trans. Automat. Contr., 43, pp. 596–607, 1998.
M. Johansson and A. Rantzer, ‘‘Computation of piecewise quadratic
Lyapunov functions for hybrid systems’’, IEEE Trans. Automat.
Contr., 43, pp. 555–559, 1998.
D. Liberzon, Switching in Systems and Control, Boston: Birkhauser,
2003.
D. Liberzon and A.S. Morse, ‘‘Basic problems in stability
and design of switched systems’’, IEEE Control Systems Magazine,
19, pp. 59–70, 1999.
A.S. Morse, ‘‘Supervisory control of families of linear set-point
controllers – Part 1: Exact matching’’, IEEE Trans. Automat.
Contr., 41, pp. 1413–1431, 1996.
G.M. Xie and L. Wang, ‘‘Reachability realization and stabilizability
of switched linear discrete-time systems’’, Journal of Math.
Analysis and Applications, 280, pp. 209–220, 2003.
X. Xu and P.J. Antsaklis, ‘‘Optimal control of switched systems based
on parameterization of the switching instants’’, IEEE Trans.
Automat. Contr., 49, pp. 2–16, 2004.
H. Ye, A.N. Michel and L. Hou, ‘‘Stability theory for hybrid
dynamical systems’’, IEEE Trans. Automat. Contr., 43,
pp. 461–474, 1998.
G.S. Zhai, ‘‘Quadratic stability of discrete-time switched systems
via state and output feedback’’, in Proceedings of the 40th
IEEE CDC, Florida, 2001, pp. 2165–2166.
G.S. Zhai, B. Hu, K. Yasuda and A.N. Michel, ‘‘Quantitative
analysis of discrete-time switched systems’’, in American Control
Conference, Anchorage, pp. 1880–1885, 2002.