Automatica 68 (2016) 266–276

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Brief paper

On asymptotic stability of linear time-varying systems✩

Bin Zhou
Center for Control Theory and Guidance Technology, Harbin Institute of Technology, P.O. Box 416, Harbin, 150001, China

article info abstract

Article history: This paper is concerned with asymptotic stability analysis of linear time-varying (LTV) systems. With
Received 12 June 2014 the help of the notion of stable functions, some differential Lyapunov inequalities (DLIs) based necessary
Received in revised form and sufficient conditions are derived for testing asymptotic stability, exponential stability and uniformly
14 October 2015
exponential stability of general LTV systems. With the help of the concept of (non-uniformly) exponential
Accepted 8 December 2015
Available online 27 February 2016
stability, a class of upper-triangular LTV systems is carefully investigated based on the developed stability
analysis approaches. A couple of numerical examples with some of them borrowed from the literature is
Keywords:
provided to illustrate the effectiveness of the proposed theoretical results.
Linear time-varying systems © 2016 Elsevier Ltd. All rights reserved.
Asymptotic stability
Exponential stability
Stable functions
Differential Lyapunov inequalities

1. Introduction stability of a linear time-invariant system is totally determined

The stability of dynamics systems is the most important
criterion in system design (Harris & Miles, 1980). Hence stability of
linear time-varying (LTV) systems including linear time-invariant
ones as special cases has received considerable attention during
the past several decades (see Anderson & Moore, 1969, Haken &
Naylor, 1966, Malek-zavarei, 1978, Mazenc, Malisoff, & Niculescu,
2014, Mullhaupt, Buccieri, & Bonvin, 2007, Sun, 2007, Zhou,
Cai, & Duan, 2013 and the references therein). There has been
considerable development of the state space approach to stability
theory of linear time-invariant systems in the past several decades,
while the corresponding status of LTV systems is comparatively
retarded (Harris & Miles, 1980). This is because, as pointed out in
Harris and Miles (1980), state transition matrices of LTV systems,
which are impossible to be derived except for very particular cases
(see, for example, Wu, Horowitz, & Dennison, 1975), are generally
needed to ascertain the properties of stability, while for linear
time-invariant systems such properties can be determined directly
(or indirectly) in terms of the system parameters. Particularly, the
✩ This work was supported in part by the National Natural Science Foundation
of China under Grant Numbers 61273028, 61322305 and 61573120, by the
Natural Science Foundation of Heilongjiang Province of China under Grant Number
F2015007, and by the Foundation for the Author of National Excellent Doctoral
Dissertation of China under Grant 201343. The material in this paper was not
presented at any conference. This paper was recommended for publication in
revised form by Associate Editor Akira Kojima under the direction of Editor Ian R.
Petersen.
E-mail addresses: binzhoulee@163.com, binzhou@hit.edu.cn.
http://dx.doi.org/10.1016/j.automatica.2015.12.030
0005-1098/© 2016 Elsevier Ltd. All rights reserved.
by the locations of the eigenvalues of the system matrix, while
the stability of an LTV system cannot be linked with the locations
of the eigenvalues of its system matrix (Rugh, 1996; Wu, 1974).
Consequently, it seems that simple yet necessary and sufficient
conditions guaranteeing stability of general LTV systems are not
available, except for the very special case that the system is
periodic (Zhou & Duan, 2012; Zhou, Hou, & Duan, 2013).
On the other hand, in contrast to linear time-invariant
systems which only have two kinds of stability concepts, namely,
(Lyapunov) stability and asymptotic stability, there are several
different stability concepts for LTV systems by distinguishing
uniform stability from non-uniform stability which are related
with the initial time of the system, making the stability analysis of
LTV systems much more complicated than the time-invariant ones
(see Kalman & Bertram, 1960 for a comprehensive introduction
of different stability concepts for dynamics systems and the
relationship among them). In the literature there are a few
Lyapunov stability theorems for testing uniformly asymptotic
stability (namely, uniformly exponential stability) (see Anderson
& Moore, 1969, Kalman & Bertram, 1960, Ramarajan, 1986, Rugh,
1996 and the references therein) while few results are available for
testing non-uniformly asymptotic stability.
In consideration of the complexity of stability analysis of
LTV systems, considerable research effort has been taken to find
less conservative sufficient conditions. One type of the efficient
methods among them relies on the use of Lyapunov function by
approximating the LTV systems by time-invariant ones (Ilchmann,
Owens, & Pratzel-Wolters, 1987; Mullhaupt et al., 2007). Another
type of methods adopts the ideas found in robust control theory by
 B. Zhou / Automatica 68 (2016) 266–276
modeling the time-varying elements as uncertainties imposed on
the linear time-invariant plants so that the theory built for linear
time-invariant systems can be adopted (see, for example, Cao &
Lam, 2000, Petersen, 1985, Petersen, 1988 and Xu, Lam, & Zou,
2009). Besides, the Bellman–Gronwall inequality approach (which
is also known as perturbation analysis) is also proven to be very
effective in deriving non-conservative conditions for guaranteeing
asymptotic stability for LTV systems (Bellman, 1953; Solo, 1994).
The asymptotic stability test of LTV system has been listed
as the FIRST open problem in mathematical systems and control
theory (see Aeyels & Peuteman, 1999). It is mentioned in Aeyels
and Peuteman (1999) that this problem is non-trivial and some
specific versions of this problem were even examined with the
framework of complexity theory (Aeyels & Peuteman, 1999). It
is further commented in Aeyels and Peuteman (1999) that: ‘‘The
study of the stability of a system ẋ(t ) = A(t )x(t ) with a system
matrix A(t ) which is neither fast time-varying nor slowly time-
varying is an ambitious task’’. Hence, by considering that the
stability test of LTV systems is extremely difficult and has not
yet been fully developed, there is still a need to make an effort
on this issue. In this paper motivated by Rugh (1996), Huang,
Hollot, and Xu (1991) and the comparison principle Khalil (2002),
we make some new observations on the stability analysis of
LTV systems. With the help of the notion of stable functions, a
series of differential Lyapunov inequality (DLI) based necessary
and sufficient conditions is derived to test asymptotic stability,
(non-uniformly) exponential stability and uniformly exponential
stability. The main feature of these criteria is that they share the
same DLI and do not require the right hand side of the DLI be
negative all the time. The (non-uniformly) exponential stability
concept is then utilized to study the asymptotic stability of a class
of triangular LTV systems. A couple of examples with some of them
borrowed from the literature is carried out to demonstrate the
effectiveness of the obtained results.
The remainder of this paper is organized as follows. In Section 2,
we recall the different stability concepts and the corresponding
criteria based on the state transition matrix. The main results
are provided in Section 3 in which the concept of state functions
is introduced and studied in Section 3.1, DLIs based criteria are
introduced in Section 3.2 and stability of a class of upper-triangular
LTV systems is investigated in Section 3.3. Some examples are
provided in Section 4 to illustrate the obtained results. Finally,
Section 5 concludes this paper.
2. Stability concepts of LTV systems
Throughout this paper, if not specified, we let J = [t ∗ , ∞) with
t ∗ being some finite number. We use C1 (J , Ω ) and PC (J , Ω ) to
denote respectively the space of Ω -valued continuously differen-
tiable functions and piecewise continuous functions defined on J.
The acronym WGDR refers to ‘‘with guaranteed decay rate’’ and |·|
refers to the usual Euclidean norm. Consider the following linear
time-varying (LTV) system
ẋ(t ) = A(t )x(t ), t ∈ J, (1)
where A(t ) ∈ PC J , R
 n×n

. Denote the state transition matrix for
this system as Φ (t , t0 ), ∀t , t0 ∈ J , t ≥ t0 .
Definition 1. The LTV system (1) is said to be
(1) stable if for any ε > 0 and for any t0 ∈ J, there exists a
δ = δ(t0 , ε) > 0 such that |x(t0 )| ∈ [0, δ] ⇒ |x(t )| ∈
[0, ε], ∀t , t0 ∈ J , t ≥ t0 ;
(2) uniformly stable if δ in Item 1 is independent of t0 ;
(3) asymptotically stable (AS) if it is stable and, for any t0 ∈ J and
any ε > 0, there exist two positive scalars η = η(t0 ) and
T = T (t0 , ε) such that |x(t0 )| ∈ [0, η] ⇒ |x(t )| ∈ [0, ε], ∀t ≥
T + t0 ;
267
(4) uniformly asymptotically stable (UAS) if it is uniformly stable
and η and T in Item 3 are independent of t0 ;
(5) exponentially stable (ES) [WGDR α with α > 0] if, for any given
t0 ∈ J, there exist two scalars k(t0 ) > 0 such that
|x(t )| ≤ k(t0 ) |x(t0 )| e−α(t −t0 ) , ∀t , t0 ∈ J , t ≥ t0 ; (2)
(6) uniformly exponentially stable (UES) [WGDR α ] if k(t0 ) in the
above item is independent of t0 .
Definition 1 except for Item 5 is recalled from Harris and Miles
(1980). It is well-known that the stability of an LTV system is
completely characterized by its state transition matrix, as recalled
in the following lemma.
Lemma 1. The LTV system (1) is
(1) stable if and only if, for any t0 ∈ J, there exists a k(t0 ) > 0 such
that (Theorem 5.1 in Harris & Miles, 1980)
|Φ (t , t0 )| ≤ k(t0 ), ∀t , t0 ∈ J , t ≥ t0 ; (3)
(2) uniformly stable if and only if k(t0 ) in (3) is independent of t0
(Theorem 6.4 in Rugh, 1996);
(3) AS if and only if (3) is satisfied and (Theorem 5.2 in Harris & Miles,
1980)
lim |Φ (t , t0 )| = 0; (4)
t →∞
(4) ES [WGDR α ] if and only if, for any t0 ∈ J, there exists a scalar
k(t0 ) > 0 such that
|Φ (t , t0 )| ≤ k(t0 )e−α(t −t0 ) , ∀t , t0 ∈ J , t ≥ t0 ; (5)
(5) UES [WGDR α ] if and only if k(t0 ) in (5) is independent of t0
(Theorem 6.7 in Rugh, 1996);
(6) UAS if and only if it is UES (Theorem 6.13 in Rugh, 1996).
Since the notion of the ES (2) is not well recognized in the
literature (we noticed that a slightly different definition of ES was
given in Anderson, Ilchmann, and Wirth (2013)), Item 4 of Lemma 1
seems not available in the literature; yet its proof is simple and can
be carried out by combining the proofs for Items 2 and 5 (see Rugh,
1996).
Remark 1. It follows from Items 5–6 of Lemma 1 that the LTV sys-
tem (1) is UES only if |Φ (t0 + T , t0 )| , t0 ∈ J and |Φ (t0 δ, t0 )| , t0 >
0 are uniformly (with respect to t0 ) bounded for any given T > 0
and δ > 1, namely, if there exists a T > 0 or a δ > 1 such that
limt0 →∞ |Φ (t0 + T , t0 )| = ∞ or limt0 →∞ |Φ (t0 δ, t0 )| = ∞, then
the LTV system (1) is not UES.
It follows from Lemma 1 that ‘‘UAS’’ ⇔ ‘‘UES’’ =⇒ ‘‘ES’’ =⇒
‘‘AS’’ =⇒ ‘‘Stable’’ and ‘‘UAS’’ =⇒ ‘‘Uniformly Stable’’ =⇒ ‘‘Stable’’.
We provide some examples to demonstrate that the converse of
the above relationships is not true. The first system is AS but is not
ES.
Example 1. Consider the following scalar LTV system (see p. 105
in Rugh, 1996)
2t
ẋ(t ) = A(t )x(t ), A(t ) = − , ∀t ∈ J = [0, ∞).
1 + t2
This system is not UES since (see p. 105 in Rugh, 1996)
1 + t02
Φ (t , t0 ) = . (6)
1 + t2
By a similar approach as in Rugh (1996) we can show that it is also
not ES.
The next system is ES but is not UES.
268 B. Zhou / Automatica 68 (2016) 266–276

Example 2. Consider the scalar LTV system (Kalman & Bertram, as shown below. The associated state transition matrix is (see the
1960; Massera, 1956) proof of Theorem 2)

ẋ(t ) = A(t )x(t ), A(t ) = 4t sin t − 2t , ∀t ∈ J = [0, ∞). eα12 t0

 
(7)  e−(t −t0 ) φ12
α12 − 2 , α12 ̸= 2,

It is trivial to show that there exists a scalar k(t0 ) > 0 such that

 

−3(t −t0 )

Φ (t , t0 ) = 0 e
Φ (t , t0 ) = exp 4 sin t − 4t cos t − t 2

e−(t −t0 ) e−(t −3t0 ) (t − t0 )
  

, α12 = 2,

− 4 sin t0 + 4t0 cos t0 + t0 , 2
 

e−3(t −t0 )

0
satisfies (5). Hence this LTV system is ES. Now letting t0 = 2nπ
gives where φ12 = e−(3−α12 )(t −t0 ) − e−(t −t0 ) . Hence this system is AS if
and only if
Φ (t0 + π , t0 ) = Φ ((2n + 1) π , 2nπ ) = eπ (4−π )(4n+1) ,
α12 < 3. (13)
and it follows that limt0 →∞ Φ (t0 + π , t0 ) = ∞. Hence this LTV
system is not UES by Remark 1. Under this condition, this system is moreover ES. However, if α12 >
0, it follows from limt0 →∞ |Φ (t0 + 1, t0 )| = ∞ and Remark 1 that
Since |A(t )| in Example 2 is unbounded, one may argue that UES
this system is not UES. In fact, it is easy to see that this system is
and ES may be equivalent if A(t ) is uniformly bounded. However,
UES if and only if α12 ≤ 0.
this is not true as indicated by the following example.
The LTV system in (12) is also a good example for demonstrating
Example 3. Consider a scalar LTV system (see also Example 9 given that the eigenvalues of A(t ) cannot determine the stability of an
later) LTV system in general (Rugh, 1996) since λ (A(t )) = {−1, −3},
ẋ(t ) = A(t )x(t ), A(t ) = sin (ln t ) + cos (ln t ) − 2α, (8) which is independent of t and is within the open left half plane,
while different α12 determines different stability.
where t ∈ J = (0, ∞) and α is a constant satisfying

1 1 eπ + 1
 3. Stability analysis of LTV systems
α∈ , ≈ (0.5, 0.5452) . (9)
2 2 eπ −1
√ 3.1. Stable functions
It follows that |A(t )| is uniformly bounded by 2α + 2. The state
transition matrix is given by We consider the following scalar LTV system

Φ (t , t0 ) = exp (t sin (ln t ) − t0 sin (ln t0 ) + 2α t0 − 2α t ) . ẏ(t ) = µ(t )y(t ), ∀t ∈ J , (14)

Obviously, we have, for all t , t0 ∈ J , t ≥ t0 , where y(t ) : J → R is the state variable and µ(t ) ∈ PC (J , R).

|Φ (t , t0 )| ≤ exp (2t0 ) exp (− (2α − 1) (t − t0 )) , (10) Definition 2. The function µ(t ) ∈ PC (J , R) is

which implies ES. Now we choose t0 = t0 (n) = exp 2nπ + 23 π
 
(1) AS if the scalar LTV system (14) is AS;
and t = t (n) = eπ t0 to obtain (2) ES [WGDR α ] if the scalar LTV system (14) is ES [WGDR α ];
 3
 (3) UES [WGDR α ] if the scalar LTV system (14) is UES [WGDR α ].
Φ (t , t0 ) = exp e2nπ + 2 π ((2α + 1) − (2α − 1) eπ ) .
By noting that
 the transition matrix for the LTV system (14) is
Hence limt0 →∞ |Φ (eπ t0 , t0 )| = ∞ and the system is not UES by
 t
φ(t , t0 ) = exp t0
µ(s)ds , we obtain immediately the following
Remark 1.
fact.
The next example demonstrates that, even A(t ) is unbounded,
the system can be UES. Lemma 2. The scalar function µ(t ) ∈ PC (J , R) is

Example 4. Consider a scalar LTV system with (1) AS if and only if

 t
ẋ(t ) = A(t )x(t ), A(t ) = t cos t 2 − α, t ∈ J = [0, ∞), (11) lim µ(s)ds = −∞, ∀t0 ∈ J . (15)
t →∞ t0
where α > 0 is a constant. It follows that A(t ) is unbounded.
However, (2) ES [WGDR α ] if and only if there exists a β(t0 ) ∈ [0, ∞) such that
   t
1 1
|Φ (t , t0 )| = exp sin t 2 − α t − sin t02 + α t0 µ(s)ds ≤ −α(t − t0 ) + β(t0 ), ∀t , t0 ∈ J , t ≥ t0 . (16)
2 2 t0

≤ eexp (−α(t − t0 )) , ∀t , t0 ∈ J , t ≥ t0 , (3) UES [WGDR α ] if and only if there exists a β ∈ [0, ∞) such that
which implies that the system is UES.  t
µ(s)ds ≤ −α(t − t0 ) + β, ∀t , t0 ∈ J , t ≥ t0 . (17)
We finally provide a second order LTV system that is ES but is t0
not UES.
The conditions (16) and (17) are not easy to test. We next
Example 5. We consider an LTV system with
provide some criteria that are testable. For a given t0 ∈ J, a

−1 eα12 t
 sequence {ti }∞
i=0 is said to be admissible if limi→∞ ti = ∞ and there
A(t ) = , ∀t ∈ J = [0, ∞), (12) exists a δ > 0 such that δi , ti+1 − ti ∈ (0, δ], ∀i ≥ 0.
0 −3
where α12 is a constant scalar. If α12 = 2, this system appears in Lemma 3. The function µ(t ) ∈ PC (J , R) is ES if, for any given t0 ∈ J,
p. 129 in Rugh (1996) and is claimed to be UES, which is incorrect i=0 and two scalars c > 0 and
there exist an admissible sequence {ti }∞
 B. Zhou / Automatica 68 (2016) 266–276 269

d(t0 ) < ∞ such that, for all i ≥ 0, follows from

 ti+1  ti + 12  t1 i −1 
 tj+1  ti + 12
µ(s)ds ≤ −c . (18) µ(s)ds = µ(s)ds + µ(s)ds + µ(s)ds
ti t0 t0 j =1 tj ti
 ti +θ  t1
1 1
µ(s)ds ≤ d(t0 ), ∀θ ∈ [0, δi ]. (19) = µ(s)ds − (i − 1) + ti
ti t0 4 2
 t1
1 3 1
= µ(s)ds + i + + k0
Lemma 4. The following three statements are equivalent: t0 4 4 2
i→∞
(1) The function µ(t ) ∈ PC (J , R) is UES. −→ ∞,
(2) For any given t0 ∈ J, there exist an admissible sequence {ti }∞
i=0 and that µ(t ) is not ES.
two scalars c > 0 and d < ∞ such that (18) and the following
condition are satisfied for all i ≥ 0 Remark 4. We also mention that (16) does not imply (18) and (19)
 ti +θ if β(t0 ) depends on t0 . To see this, we let J = [0, ∞) and
µ(s)ds ≤ d, ∀θ ∈ [0, δi ]. (20)
−(ρ + 2α), t ∈ [k2 , k2 + k + 1),

ti µ(t ) =
ρ − 2α, t ∈ k2 + k + 1, (k + 1)2 ,

(3) For any given constant c > 0, there exist δ ∗ (c ) > 0 and
d(c ) < ∞ such that, for any δ ≥ δ ∗ (c ) and t ∈ J, the following where α > 0 and ρ > 2α are any constants and 0 ≤ k ∈ Z. We
conditions are satisfied can show that (see Appendix A.4)
 t +δ
 t

µ(s)ds ≤ −c , (21) µ(s)ds ≤ −2α(t − t0 ) + (2k0 + 1) ρ, ∀t , t0 ∈ J , t ≥ t0 , (24)

t t0

t +θ
where k0 is the only non-negative integer satisfying t0 ∈ [k20 ,

µ(s)ds ≤ d(c ), ∀θ ∈ [0, δ]. (22)
t
(k0 + 1)2 ). However, as µ(t ) > 0 in the interval [k2 + k +
1, (k + 1)2 ) which can be made arbitrarily large, (18) and (19)
cannot be satisfied for any admissible sequence {ti }∞
i =0 .
Lemma 5. Assume that µ(t ) ∈ PC (J , R) is periodic with period T .
Then the following statements are equivalent: 3.2. Stability test by differential Lyapunov inequalities
(1) The function µ(t ) is AS.
In this subsection, we first present some DLIs based stability
(2) The function µ(t ) is ES. criteria for LTV systems. If A(t ) is a constant matrix, then system
(3) The function µ(t ) is UES. (1) is UES [WGDR α ] if and only if there exists a P > 0 such that
(4) There exists a constant c > 0 such that A| P + PA ≤ −2α P (see p. 127 in Rugh, 1996). Motivated by this
result, we consider the following DLI
 t +T
ϕ (T ) , µ(s)ds ≤ −c . (23) Ṗ (t ) + A| (t )P (t ) + P (t )A(t ) ≤ 2µ(t )P (t ), t ∈ J, (25)
t
where P (t ) = P (t ) ∈ C J, R and µ(t ) ∈ PC (J , R) are to be
| 1
 n×n


The proofs of Lemmas 3–5 are respectively given in Appen- defined.

dices A.1, A.2, and A.3. Three remarks are given in order for these
Theorem 1. Let t ∗ > −∞. The LTV system (1), where A(t ) ∈
results.
PC J , Rn×n , is


 if there exist an AS function µ(t ), a P (t ) = P (t ) ∈

|
Remark 2. Let ∆ > 0 be a constant and define the moving average (1) AS if
 and only
 t +∆ C1 J , Rn×n , and a positive constant p1 such that (25) and the
of µ(t ) with respect to ∆ as γµ (t , ∆) = ∆ 1
t
µ(s)ds (Khalil,
following inequality are satisfied
2002). Then we can write (21) as γµ (t , δ) ≤ − δc < 0. Hence (21)
is nothing but just requiring that the moving average of µ(t ) with p1 In ≤ P (t ), ∀t ∈ J . (26)
respect to δ is negative for sufficiently large δ . (2) ES [WGDR α ] if and only if there
 exist  an ES function µ(t ) [WGDR
α ], a P (t ) = P | (t ) ∈ C1 J , Rn×n , and a positive constant p1
Remark 3. It is clear that (19) and (20) are satisfied if |µ(t )| is such that (25) and (26) are satisfied.
bounded uniformly (in this case, Item 2 of Lemma 4 coincides with (3) UES [WGDR α ] if (and only if when A (t ) is uniformly bounded)
 UES function µ(t ) [WGDR α ], a P (t ) = P (t ) ∈
|
the conclusion in Theorem 5 of Huang et al., 1991), and they cannot there
 exist an
be removed in general. To see this, we consider J = [0, ∞) and C J, R
1 n×n
, and two constants p2 > p1 > 0 such that (25) and
  the following inequalities are satisfied
1

k, t ∈ k, k + , p1 In ≤ P (t ) ≤ p2 In , ∀t ∈ J .

 (27)
2
µ(t ) =    
1 1 Proof. Choose the Lyapunov function V (x(t )) = x| (t )P (t )x(t ).
, ,k + 1 ,

− k +
 t ∈ k+
2 2 Then, by (25), V̇ (x(t )) ≤ 2µ(t )V (x(t )) , ∀t ∈ J, from which it
follows that
where 0 ≤ k ∈ Z. For any t0 = k0 + θ0 , 0 ≤ k0 ∈ Z, θ0 ∈ [0, 1),   t

we choose ti = k0 + i + 1, 1≤ i ∈ Z. It is easy to show that (18) V (x(t )) ≤ exp 2 µ(s)ds V (x(t0 ))
t
is satisfied with δ = 2 since t i+1 µ(s)ds ≤ − 14 . However, (19) is t0
i

not satisfied since

 ti + 1
2
µ(s)ds = 1 i→∞
t −→ ∞, 1 ≤ i ∈ Z. It then ≤ φ 2 (t , t0 ) |P (t0 )| |x(t0 )|2 , ∀t , t0 ∈ J , t ≥ t0 .
ti 2 i
270 B. Zhou / Automatica 68 (2016) 266–276

This inequality can also be deduced from the so-called comparison = ϕ −1 (t0 ) tr (Φ | (t , t ∗ )Φ (t , t ∗ ))
lemma (Khalil, 2002). Since V (x(t )) ≥ p1 |x(t )|2 and x(t0 ) is √
≤ nϕ −1 (t0 ) Φ (t , t ∗ )
 
arbitrarily chosen, the above condition just implies √
≤ nϕ −1 (t0 ) k t ∗ exp −α t − t ∗
    

|P (t0 )| = κ (t0 ) exp (−α (t − t0 )) ,
|Φ (t , t0 )| ≤ n φ (t , t0 ) , ∀t , t0 ∈ J , t ≥ t0 . (28)
p1 √
where κ (t0 ) = nϕ −1 (t0 ) k (t ∗ ) exp (α (t ∗ − t0 )). Hence µ (t ) is
Proof of Item 1: The ‘‘if’’ part follows from (28) directly. We an ES function [WGDR α ]. √
next prove the ‘‘only if’’ part. Consider the function ϕ (t ) ∈ Proof of
√ Item 3: The ‘‘if’’ part follows from (34) since |P (t0 )| /p1
C1 (J , (0, ∞)) defined as eβ(t0 ) ≤ p2 /p1 eβ which is independent of t0 . We next show the
 ‘‘only if’’ part. For any sufficiently small constant ε > 0, the LTV
ϕ (t ) = tr (Φ | (t , t ∗ )Φ (t , t ∗ )), ∀t ∈ J . (29) system
Since the LTV system (1) is AS, we have limt →∞ ϕ (t ) = 0. Define ẋ(t ) = (A(t ) + (α − ε) In ) x(t ) , Aε (t ) x(t ), (35)
µ (t ) ∈ PC (J , R) as
(α−ε)(t −t0 )
is also UES since its state transition matrix is Φ (t , t0 ) e .
d ϕ̇ (t ) On the other hand, it is well known (see Theorem 7.8 in Rugh, 1996)
µ (t ) = ln (ϕ (t )) = , (30)
dt ϕ (t ) that, if the LTV system (35) is UES and A(t )+(α − ε) In is uniformly
from which it follows that, for all t , t0 ∈ J , t ≥ t0 ,
bounded (which is equivalent to the  uniform boundedness of A(t )),
there exists a P (t ) = P | (t ) ∈ C1 J , Rn×n satisfying (27) and
t
 
ϕ (t ) = ϕ (t0 ) exp µ (s) ds = ϕ (t0 ) φ (t , t0 ) . (31) Ṗ (t ) + A|ε (t )P (t ) + P (t )Aε (t ) ≤ −v In ,
t0
t for some positive constant v . However, the above inequality
Since limt →∞ ϕ (t ) = 0, we have limt →∞ t µ (s) ds = −∞, implies (37) where µ (t ) = − (α − ε). The proof is finished. 
0
which means that µ (t ) is an AS function. We then choose the
symmetric matrix function Remark 5. We make some explanations on Theorem 1.

P (t ) = φ 2
t, t ∗
Φ | (t ∗ , t )Φ (t ∗ , t ) ∈ C 1
J, R
n×n
. (1) In the ‘‘only if’’ part of Item 3 in Theorem 1 we have assumed
   
(32)
that A (t ) is uniformly bounded. However, this assumption
By using the definition of µ (t ) and (31), we have, for all t ∈ J, can be weakened by requiring some uniformly completely
−1 observability property (see Anderson & Moore, 1969). Since
P (t ) = φ 2 t , t ∗ Φ (t , t ∗ )Φ | (t , t ∗ )
 
UES of LTV system has been well recognized in the literature,
−1
≥ φ 2 t , t ∗ tr Φ (t , t ∗ )Φ | (t , t ∗ ) we here do not attempt to explore these subtleties too much.
  
In
(2) Clearly, if we let α = 0 in Items 2–3 of Theorem 1, then
1
= 2 φ 2 t , t ∗ In
 
the corresponding conditions imply respectively that the LTV
ϕ (t ) system (1) is stable and uniformly stable.
(3) The DLI (25) can be written as Ṗ (t ) + A| (t )P (t ) + P (t )A(t ) ≤
−1 2  ∗ 
= ϕ2 t ∗ φ2 t , t ∗ φ t , t In
   

1
−Q (t ), where Q (t ) = −2µ(t )P (t ). Since we do not require
= ϕ − 2 t ∗ In = In , µ(t ) ≤ 0, Q (t ) can be indefinite, which is different from
 
(33)
n the existing results such as Theorem 7.4 in Rugh (1996) and
which indicates that (26) is satisfied. Now by noting that Proposition 2 in Karafyllis and Tsinias (2003). We also do not
∂
∂s
Φ (t , s) = −Φ (t , s)A (s) we can compute require µ(t ) to be bounded.
(4) From the proof of Theorem 1 we can see that Items 1–2 of this
 ∂ Φ | (t ∗ , t )
Ṗ (t ) = 2µ (t ) P (t ) + φ 2 t , t ∗ Φ (t ∗ , t )

theorem are also true if p1 in (26) dependents on t0 , namely,
∂t p1 = p1 (t0 ), ∀t0 ∈ J.
∂ Φ (t ∗ , t ) (5) For UES, if A(t ) and µ(t ) are bounded, then ‘‘Item 3 of
+ φ 2 t , t ∗ Φ | (t ∗ , t )
 
∂t Theorem 1 ’’ + ‘‘Item 2 of Lemma 4 ’’ ⇒ ‘‘Theorem 5 in Huang
= 2µ (t ) P (t ) − (φ t , t A (t ) Φ | (t ∗ , t )Φ (t ∗ , t ) et al. (1991) ’’. Moreover, if A(t ) is periodic, then ‘‘Item 3 of
2
 ∗ |
Theorem 1 ’’ + ‘‘Item 4 of Lemma 5 ’’ ⇒ ‘‘Theorem 6 in Huang
+ φ 2 t , t ∗ Φ | (t ∗ , t )Φ (t ∗ , t )A (t ))
 
et al. (1991) ’’.
= 2µ (t ) P (t ) − (A| (t ) P (t ) + P (t )A (t )) , (6) The construction of P (t ) in (32) was motivated by (22) in
Karafyllis and Tsinias (2003). However, this construction is not
which just shows that P (t ) satisfies the DLI (25).
valid if t ∗ = −∞. So in an early version of this paper, by
Proof of Item 2: By using (16) we can further obtain from (28)
assuming that |A(t )| ≤ aebt , ∀t ∈ J, and b < 2α , the necessity
that, for all t , t0 ∈ J , t ≥ t0 ,
part of Item 2 was proven by constructing
  ∞
|P (t0 )| β(t0 ) −α(t −t0 ) P (t ) = eε(s−t ) |A(s)| Φ | (s, t ) Φ (s, t ) ds, ∀t ∈ J ,
|Φ (t , t0 )| ≤ n e e , (34)
p1 t

which means that the LTV system (1) is ES [WGDR α ]. This proves where ε = 12 (2α − b) > 0. The symmetric matrix P (t ) is well-
defined for all t ∈ J since
the ‘‘if’’ part. The proof for the ‘‘only if’’ part is almost the same as
that for Item 1 and the only difference is that the function µ (t ) in ∞

(30) is an ES function. To show this, it follows from (31), (29) and |P (t )| ≤ eε(s−t ) |A(s)| |Φ (s, t )|2 ds
t
(5) that ∞
ak2 (t )ebt

φ (t , t0 ) = ϕ −1
(t0 ) ϕ (t ) ≤ k2 (t )eε(s−t ) aebs e−2α(s−t ) ds = .
t ε
 B. Zhou / Automatica 68 (2016) 266–276 271

Moreover, P (t ) is bounded below, i.e. (the proof is given in Theorem 2. Consider the following blocked upper-triangular LTV
Appendix A.5), system
1
A11 (t ) A12 (t )
 
P (t ) ≥ In , ∀t ∈ J . (36)
ẋ(t ) = x(t ), ∀t ∈ J , (39)
2 0 A22 (t )
Finally, it is easy to verify that Ṗ (t ) = − |A(t )| In − ε P (t ) −
where Aij (t ) ∈ PC J , Rni ×nj , i, j = 1, 2, j ≥ i, and n1 + n2 = n.
 
A| (t )P (t ) − P (t )A(t ), which is in the form of (25) with µ =
− 21 ε . Assume that the following conditions are met:

Our next result improves Item 2 of Theorem 1. (1) The LTV subsystem ẋ1 (t ) = A11 (t )x1 (t ) is UES WGDR α1 .
(2) The LTV subsystem ẋ2 (t ) = A22 (t )x2 (t ) is ES WGDR α2 .
Corollary 1. There exist a P (t ) = P | (t ) ∈ C J , R (3) There exist a nonnegative constant k12 and a constant α12 such
1
 n×n

and an ES
function [WGDR α ] µ(t ) ∈ PC (J , R) such that the DLI (25) and the that
 some p1 (t0 ) > 0 if and only if there
inequality in (26) are satisfied for
|A12 (t )| ≤ k12 eα12 t , ∀t ∈ J . (40)
exists a P0 (t ) = P0 (t ) ∈ C1 J , Rn×n such that (26) with a new
|

p1 (t0 ) > 0 and the following DLI is satisfied (4) The scalars α2 and α12 satisfy

Ṗ0 (t ) + A (t )P0 (t ) + P0 (t )A(t ) ≤ −2α P0 (t ),

|
∀t , t0 ∈ J , t ≥ t0 . α12 < α2 . (41)
(37)
Then system (39) is ES WGDR min {α2 − α12 , α1 − ε} where ε >
0 is a sufficiently small number. Moreover, if α12 ≤ 0 and the LTV
Proof. The ‘‘if’’ part is obvious. To prove the ‘‘only if’’ part, we subsystem ẋ2 (t ) = A22 (t )x2 (t ) is UES, then system (39) is also UES.
denote µ(t )dt = u(t ) + C where C is some constant, namely,
u(t ) is the indefinite integral of µ(t ). Then we choose Proof. Let Φii (t , t0 ) be the state transition matrix for the system
ẋii (t ) = Aii (t )xii (t ), i = 1, 2. Then the state transition matrix for
P0 (t ) = P (t ) exp (−2α t − 2u(t )) , (38)
system (39) is given by
which, by a direct computation and using (25), satisfies (37). On
Φ11 (t , t0 ) Φ12 (t , t0 )
 
the other hand, by using (16), we have, ∀t , t0 ∈ J , t ≥ t0 , Φ (t , t0 ) = ,
0 Φ22 (t , t0 )
t
  
−2α t0 −2u(t0 )
P0 (t ) = P (t ) exp −2α(t − t0 ) − 2 µ(s)ds e where
t0
 t
≥ p1 (t0 )e−2α(t −t0 )+2α(t −t0 )−2β(t0 ) e−2αt0 −2u(t0 ) In Φ12 (t , t0 ) = Φ11 (t , s)A12 (s)Φ22 (s, t0 )ds. (42)
= p1 (t0 ) exp (−2β(t0 ) − 2α t0 − 2u(t0 )) In , t0

which means that (26) is satisfied with some new p1 (t0 ) (by Item By the assumptions in Items 1 and 2, there exist four positive
4 of Remark 5). The proof is finished.  constants k1 , k2 (t0 ), αi , i = 1, 2, such that

|Φ11 (t , t0 )| ≤ k1 e−α1 (t −t0 ) , ∀t , t0 ∈ J , t ≥ t0 ,


Briefly speaking, Corollary 1 means that, for testing ES, the two
DLIs (25) and (37) are equivalent. |Φ22 (t , t0 )| ≤ k2 (t0 )e−α2 (t −t0 ) , ∀t , t0 ∈ J , t ≥ t0 .
We finally give some explanations on choosing the matrix P (t )
satisfying (25). In principle, the matrix P (t ) satisfying (25) can be Hence we conclude from (40) and (42) that
constructed according to (32). However, this is generally impossi-  t
ble since it involves the state transition matrix Φ (t , t0 ) for the LTV |Φ12 (t , t0 )| ≤ |Φ11 (t , s)| |A12 (s)| |Φ22 (s, t0 )| ds
system (1). If Φ (t , t0 ) is not available, we should mention that, it is t0
t
very difficult, if not impossible, to provide a general procedure for

constructing a Lyapunov matrix P (t ) satisfying (25) even if µ (t ) is ≤ k1 k2 (t0 )k12 e−α1 (t −s) eα12 s e−α2 (s−t0 ) ds
t0
not required to be negative for all t. We can only suggest a simple
t
numerical tractable approach in a particular case: If A(t ) is Hurwitz

= k1 k2 (t0 )k12 eα2 t0 −α1 t e(α1 −α2 +α12 )s ds.
for some ts , then we can choose P (t ) = Ps where Ps > 0 is such that t0
A| (ts )Ps + Ps A (ts ) = −Qs < 0. In this case, we know that (25) is
satisfied with µ (t ) = 12 λmax {(A| (t )Ps + Ps A (t ))Ps−1 } which may be We consider two cases.
Case 1: α1 + α12 ̸= α2 . In this case,
an AS function since µ (ts ) = − λmin {Qs Ps } < 0, namely, µ (t )
1
2
−1

is negative in the neighborhood of ts . This is particularly of the case |Φ12 (t , t0 )| ≤ κ (t0 ) eα12 t0 e−(α2 −α12 )(t −t0 ) − e−α1 (t −t0 )
 
if A(t ) is periodic. This usage of constant Lyapunov matrices will be
illustrated in Section 4.2. Another tractable way is to use piecewise ≤ 2κ (t0 ) eα12 t0 e− min{α1 ,α2 −α12 }(t −t0 ) , (43)
constant Lyapunov function in (25) instead of the continuously dif- k k (t )k
ferentiable Lyapunov function, which deserves a further study. in which κ (t0 ) = |α 1−α
2 0 12
.
1 2 +α12 |
Case 2: α1 + α12 = α2 . In this case, we have
3.3. Stability of a class of triangular LTV systems |Φ12 (t , t0 )| ≤ k1 k2 (t0 )k12 eα2 t0 −α1 t (t − t0 )
Motivated by Example 5 and the concept of ES (2), we inves- = k1 k2 (t0 )k12 eα12 t0 e−α1 (t −t0 ) (t − t0 )
tigate in this subsection a class of LTV systems having triangu- k1 k2 (t0 )k12 eα12 t0 −(α1 −ε)(t −t0 )
lar structures. Since a lower-triangular LTV system is algebraically ≤ e , (44)
equivalent to an upper-triangular one by a time-invariant state ε
transformation x → En x, where En is the anti-identical matrix (see where ε > 0 is any sufficiently small number. Combining (43) and
Example 11), we only consider upper-triangular LTV systems in (44) gives that the LTV system (39) is ES with the guaranteed decay
this paper. We mention that triangular LTV systems are encoun- rate min {α2 − α12 , α1 − ε}.
tered in the observer based output feedback control of LTV systems Finally, if ẋ2 (t ) = A22 (t )x2 (t ) is UES and α12 ≤ 0, it is obvious
(see Chapter 15 in Rugh, 1996). to see that (39) is also UES. 
272 B. Zhou / Automatica 68 (2016) 266–276

Remark 6. The LTV system (12) in Example 5 is in the form of exist). We use Item 1 of Theorem 1 to check the stability of this
(39) and it follows from (13) that the condition (41) is tight for system. Let
guaranteeing the ES of system (39).  
8 t +1
P (t ) = , ∀t ∈ J ,
Applying the above theorem recursively from the bottom up t +1 (t + 1) (10t + 1)
gives immediately the following corollary.
which is unbounded as t → ∞. Direct computation gives

Corollary 2. Consider an LTV system in the form of I2 (1 + t ) (79t + 7)

P (t ) ≥  = I2
tr P −1 (t )

10t 2 + 11t + 9
A11 (t ) A12 (t ) ··· A1n (t )
 
A22 (t ) ··· A2n (t ) (1 + t0 ) (79t0 + 7)
≥ I2 , p1 (t0 )I2 , ∀t , t0 ∈ J , t ≥ t0 ,
ẋ(t ) =   x(t ), ∀t ∈ J ,

.. ..  (45) 10t02 + 11t0 + 9
 . .
Ann (t ) where we have noticed that dtd
p1 (t ) > 0, ∀t ∈ J. Hence P (t )
satisfies (26). Now we compute
where Aij (t ) ∈ PC J , Rni ×nj . If ẋi (t ) = Aii (t )xi (t ), i = 1, 2, . . . ,
 
n − 1, are all UES, ẋn (t ) = Ann (t )xn (t ) is ES, and there exist Ṗ (t ) + A| (t )P (t ) + P (t )A(t ) − 2µ(t )P (t ) = −Q (t ), (47)
nonnegative constants cij , dij and nonnegative integers mij such that
where µ(t ) = − 12(t1+1) , and
Aij (t ) ≤ cij + dij t mij , ∀t ∈ J ,
 
 4 11 
are true for all j > i, then system (45) is ES. If, moreover, ẋn (t ) = 2− + 20t
3 (1 + t ) 6
Ann (t )xn (t ) is also UES and mij = 0, j > i, then system (45) is also Q (t ) =  .
 
11 41 589
UES. + 20t + t + 200t 2
6 6 3
4. Some examples Since 2 − 3(14+t ) > 0, ∀t ∈ J, and

43 + 2443t + 1896t 2
4.1. Some illustrative examples for DLI based stability test det (Q (t )) = > 0, ∀t ∈ J ,
36 (t + 1)
In this subsection, we provide several examples to illustrate the we conclude that Q (t ) > 0, ∀t ∈ J, and it follows from (47) that
obtained results in Section 3.2. the DLI (25) associated with system (46) is guaranteed. Moreover,
it can be verified that, for all t ≥ t0 ∈ J,
Example 6 (Lyapunov Matrix for the Systems in Examples 2 and 3).  t  t
 
For scalar LTV systems that is ES WGDR α , by noting (38), the 1 ds 1 t +1
µ(s)ds = − =− ln ,
associated Lyapunov matrix P (t ) satisfying (26) and (37) can be t0 12 t0 s+1 12 t0 + 1
constructed as P (t ) = exp (−2α t − 2a(t )) where a(t ) is the
which satisfies (15). Hence, by Theorem 1, this LTV system is AS
indefinite integral of A(t ). Hence the matrices Pi (t ), i = 2, 3 for
and, by (28), the state transition matrix for this system satisfies,
the systems in Examples 2 and 3 can be respectively chosen as
for all t ≥ t0 ∈ J,
P2 (t ) = exp −2α2 t − 2 4 sin t − 4t cos t − t 2 ,
  

|P (t0 )|
 
1 t +1
P3 (t ) = exp (−2 (2α − 1) t − 2 (t sin (ln t ) − 2α t )) , |Φ (t , t0 )| ≤ 2 exp − ln
p1 (t0 ) 12 t0 + 1
where α2 > 0 is any constant and α satisfies (9). It is easy to verify 
that P2 (t ) ≥ exp(−8 − 21 (4 + α2 )2 ) and P3 (t ) ≥ 1 for any t ≥ t0 ∈ tr (P (t0 )) 12 t0 + 1


J , [0, ∞). However,

≤2
  both P2 (t ) and P3 (t ) are unbounded as there p1 (t0 ) t +1
exist two series ti,k such that limk→∞ Pi ti,k = ∞, i = 2, 3.  2  
2 10t0 + 11t0 + 9 12 t0 + 1
= √ .
Example 7 (Lyapunov Matrix for the System  in Example 4). For the (1 + t0 ) (79t0 + 7) t +1
system in (11), we construct P (t ) = exp − sin t 2 , ∀t ∈ J, which

Finally, we mention that the following symmetric matrix was
is such that e−1 ≤ P (t ) ≤ e, ∀t ∈ J, and recommended in Zheng (2002) to verify the stability of this system
Ṗ (t ) + A| (t )P (t ) + P (t )A(t ) + 2α P (t ) = 0, ∀t ∈ J .  
1 1 0
P = ,
Hence, by Theorem 1, this system is UES [WGDR α ]. We mention 2 0 t +1
that, for this example, the existence of a solution satisfying (27) to
which is such that P (t ) ≥ , ∀t ∈ J, and
1
I
the DLI (25) is necessary and sufficient for the UES even when A(t ) 2 2

is unbounded. 
0 0

A| (t )P (t ) + P (t )A(t ) + Ṗ (t ) = −10 ≤ 0.
0 t +1
Example 8. We consider an LTV system with (see p. 252 in Zheng,
2002) However, for this matrix P (t ), we can only conclude that the
  system is stable.
0 1
A(t ) =  1 , ∀t ∈ J = [0, ∞). (46)
− −10 4.2. A numerical example for DLI based stability test
t +1
Notice that the state transition matrix for this system is difficult We consider a mass–spring system for which both the damping
to obtain (maybe a closed-form state transition matrix does not coefficient and the elastic constant are time varying (Mullhaupt
 B. Zhou / Automatica 68 (2016) 266–276 273

Table 1
The maximal allowable α for system (48) by different approaches.
Different approaches Rosenbrock (1963) Mullhaupt et al. (2007) Tan and Duan (2009) Theorem 1 with P in (49) Theorem 1 with P in (50)

The maximal value of α 1.2 1.6 1.95 2.125 2.292

Lyapunov function for LTV systems, we attempt to choose some

P (t ) that is intuitively reasonable. Denote A(t ) = A0 + α A1 (t )
where A0 contains all the constant elements of A(t ). We choose
P (t ) = P (t , γ ) = P0 + γ P1 (t ) where γ is some small constant
and P1 (t ) is a time-varying periodic function satisfying Ṗ1 (t ) =
−(A|1 (t )P0 + P0 A1 (t )), which means that P1 (t ) is responsible for
the cross term A1 (t )P0 + P0 A1 (t ) in the time-derivative of Ṗ (t ). It
|

is easy to get
2
 
1+ γ cos(t ) p12 (t )
P (t ) =  5
, (50)
 
17 17
p12 (t ) − γ sin(t )
50 25
which is positive definite ifγ ∈ (−0.5, 0.4480), where p12 (t ) =
1
5
+ γ 1750
cos(t ) − 15 sin(t ) . It is clear that P (t ) is not worse than
P0 since limγ →0 P (t ) = P0 . Again we choose µ(t ) = µ (t , α) =
 2π
1
λ ((A| (t )P (t ) + P (t )A(t ) + Ṗ (t ))P −1 (t )). Then, by applying a
2 max
Fig. 1. The functions µ (t , α) and 0
µ (s, α) ds for both constant and time- linear search procedure for γ , we find that (23) is met for α ≤ αt∗ =
varying Lyapunov matrix P. 2.292 with γ = −0.2 (see Table 1). Hence this result is even better
than that obtained in the above with the constant matrix P0 . For
et al., 2007). This system is in the form of (1) with  2π
the illustration purpose, the functions 0 µ (s, α) ds and µ t , αt∗
 
  with γ = −0.2 associated with this time-varying P (t ) are also
0 1
A(t ) = , recorded in Fig. 1 from which we can again see that µ t , αt∗ is

(48)
− (2 − α sin(t )) − (2 − α cos(t ))
also not negative all the time.
where the scalar α is a constant parameter that accounts for
the variability (Mullhaupt et al., 2007). Clearly, this is a periodic 4.3. Some illustrative examples for triangular systems
system with T = 2π . As noticed in Mullhaupt et al. (2007),
most of the conditions appearing in the literature request that We use some examples to illustrate the results in Section 3.3
the time variation of the system matrix is small. Particularly, the for triangular systems. Since the condition in (41) is independent
time average approach and the Rosenbrock approach (Rosenbrock, of α1 , which is the guaranteed decay rate of the subsystem
1963), which are based on |Ȧ(t )| cannot be improved since |Ȧ(t )| =
ẋ1 (t ) = A11 (t )x1 (t ), (51)
α is a constant. According to the computation in Mullhaupt et al.
(2007), Rosenbrock approach (Rosenbrock, 1963) fails for α = 1.3, we may wander whether the UES imposed on system (51) in
and the computational test established in Mullhaupt et al. (2007) Theorem 2 (namely, Item 1 of this theorem) can be relaxed as ES.
guarantees stability up to α = 1.6. The result in Mullhaupt et al. The following example gives a negative answer.
(2007) was latterly improved in Tan and Duan (2009) where the
test can assess stability when α ≤ 1.95. Example 9. Consider a second order LTV system with
We next apply Item 3 of Theorem 1 and Item 4 of Lemma 5,
sin (ln t ) + cos (ln t ) − 2α
 
e−α t
which is equivalent to Theorem 6 in Huang et al. (1991) (by Item A(t ) =
0 −α
5 of Remark 5), on this system. Since A(t ) is Hurwitz for some t
if α is not too large, as pointed out at the end of Section 3.2, we A11 (t ) A12 (t )
 
can consider a constant matrix P = P0 = [pij ], i, j = 1, 2, where , , (52)
0 A22 (t )
we assume, without loss of generality, p11 = 1. Then by setting
µ(t ) = µ (t , α) = 21 λmax ((A| (t )P + PA(t ))P −1 ) and applying a where α satisfies
linear search procedure for p12 and p22 , we find that (23) is met for 1 1 1 + eπ 1 1 + eπ
α ≤ α0∗ = 2.125 (see Table 1) with <α< < .
2 2 eπ 2 eπ − 1
 1
 This example, due to (Perron, 1930), also appeared in p. 151 in
1 Harris and Miles (1980). By Example 3, the subsystem ẋ1 (t ) =
5 .
P0 =  (49) A11 (t )x1 (t ) is ES WGDR 2α − 1; but it is not UES. Even the A12 (t )

1 17

block is uniformly bounded (and thus Items 2, 3 and 4 of Theorem 2
5 50
are satisfied), this system
 is unstable.
Denote Φ (t , t0 ) = φij (t , t0 ) , i, j = 1, 2. Then, by (42), we

This result is better than that obtained in Mullhaupt et al.
(2007) and Tan and Duan (2009). For the illustration purpose, the have
 2π
functions 0 µ (s, α) ds and µ t , α0∗ are recorded in Fig. 1 from
 
 t

which we can see that µ t , α0∗ is not negative all the time, namely, φ12 (t , t0 ) = eαt0 et sin(ln t )−2αt e−s sin(ln s) ds.
 
t0
µ t , α0∗ ≥ 0, t ∈ [2kπ , 2kπ + 2.1] , k ∈ Z.
 
Let tk = exp 2k − 12 π and β ∈ 0, π2 be a constant to be
    
To further improve the above result, we may use a time-varying
P (t ) instead. By noting that there is no general rule for choosing specified. For any t0 , let k0 be such that tk0 ≥ t0 . Then, for any
274 B. Zhou / Automatica 68 (2016) 266–276

k ≥ k0 , we have 5. Concluding remarks

 tk e π
π
φ12 (tk eπ , t0 ) = eαt0 e−(2α−1)tk e e−s sin(ln s) ds This paper studied asymptotic stability of linear time-varying
t0 (LTV) systems. With the help of the notion of stable functions,
 tk e β unified DLIs based necessary and sufficient conditions were
π
≥ eαt0 e−(2α−1)tk e e−s sin(ln s) ds derived to guarantee asymptotic stability, exponential stability and
tk uniformly exponential stability of LTV systems. By applying the
0
 tk e β obtained DLIs based criteria and the concept of (non-uniformly)
α t0 −(2α−1)tk eπ
≥e e e−s sin(ln s) ds exponential stability, a class of triangular LTV systems was
tk investigated and some sufficient conditions were established to
 tk e β guarantee their exponential stability and uniformly exponential
π
≥ eαt0 e−(2α−1)tk e es cos β ds, stability. The obtained results were illustrated by a couple of
tk
numerical examples with some of them borrowed from the
where we have noticed that − sin (ln s) ≥ cos β, s ∈ tk , tk eβ . literature. The idea of allowing indefinite time derivatives of
 
Hence Lyapunov functions, as indicated by the DLI (25), has been
π
extended to the study of time-varying time-delay systems, time-
φ12 (tk eπ , t0 ) ≥ eαt0 e−(2α−1)tk e etk cos β eβ − 1 tk
 
varying nonlinear systems, and discrete-time LTV systems, which
= eαt0 etk (cos β−(2α−1)e ) eβ − 1 tk .
π   will be reported elsewhere.

Let β be chosen such that cos β > (2α − 1) eπ , which is possible as Appendix
0 < (2α − 1) eπ < 1. Hence limk→∞ φ12 (tk eπ , t0 ) = ∞, namely,
the system is unstable.
A.1. Proof of Lemma 3
Our next example demonstrates that AS of system (39) may be
guaranteed if the subsystem ẋ1 (t ) = A11 (t )x1 (t ) is only AS. But Notice that, for any t , t0 ∈ J , t ≥ t0 , there exists an integer
in this case it seems difficult to build a general result as Theorem 2 k ≥ 0 such that t ∈ [tk , tk+1 ), namely, t = tk + θ , θ ∈ [0, δk ).
since the conditions (3) and (4) for AS provide less information than Hence
condition (5) for UES. k−1 
 t  ti+1  tk +θ
µ(s)ds = µ(s)ds + µ(s)ds
Example 10. We consider a second-order LTV system with t0 i =0 ti tk

2t e(2α−1)t
 ≤ −kc + d(t0 )
−
A(t ) =  1 + t 2 t2 k−1  
1+   ti+1 − ti
0 sin (ln t ) + cos (ln t ) − 2α = −c + d(t0 )
i=0
δi
A11 (t ) A12 (t )
 
, , k−1
c 
0 A22 (t ) ≤− (ti+1 − ti ) + d(t0 )
δ i=0
where α satisfies (9). By Example 1, the  subsystem ẋ1 (t ) = c
A11 (t )x1 (t ) is only AS. Denote Φ (t , t0 ) = φij (t , t0 ) , i, j = 1, 2. = − (tk − t0 ) + d(t0 )

Then we get from (42), (6) and (10) that δ
c c
(2α−1)s = − (tk + θ − t0 ) + d(t0 ) + θ
t 2
δ δ

1+s e
φ12 (t , t0 ) ≤ e2t0 e−(2α−1)(s−t0 ) ds c c
t0 1 + t 2 1 + s2 ≤ − (t − t0 ) + d(t0 ) + δk
t − t0 (2α+1)t0 δ δ
= e , ∀t ≥ t0 ≥ 0, c
1 + t2 ≤ − (t − t0 ) + d(t0 ) + c ,
δ
which implies limt →∞ φ12 (t , t0 ) = 0. Hence this system is AS.
which is just (16) where α = δc and β(t0 ) = d(t0 ) + c.
We finally present an example satisfying all the conditions of
Theorem 2 by modifying system (52) a little. A.2. Proof of Lemma 4

Example 11. Consider the second-order LTV system ẋ(t ) = From the proof of Lemma 3 we can see that β(t0 ) is independent
A| (t )x(t ) where A(t ) is given by (52) and thus A| (t ) is a lower- of t0 if d(t0 ) is. Hence Item 2 implies Item 1. We next show that
triangular matrix. Choosing y(t ) = E2 x(t ) gives Item 1 implies Item 2. Let δ be such that −αδ + β < 0 and choose
  δi = δ, i ≥ 0. Since (17) holds true for all t , t0 ∈ J , t ≥ t0 , we have
−α e−α t
ẏ(t ) = y(t ),
0 sin (ln t ) + cos (ln t ) − 2α ti+1

µ(s)ds ≤ −α (ti+1 − ti ) + β = −αδ + β < 0,
which is exactly in the form of (39) and satisfies all the conditions of ti

Theorem 2 in view of Example 3. Hence, by Theorem 2, this system  ti +θ

and, similarly, µ(s)ds ≤ −αθ + β ≤ β < ∞.
ti
is ES.  t +θ
Item 1 ⇒ Item 3 is obvious since t µ(s)ds ≤ −αθ + β and
We mention that Examples 9 and 11 simply demonstrate that thus δ ∗ (c ) = (β + c )/α and d (c ) = β . We only show Item 3 ⇒
the LTV system ẋ(t ) = A(t )x(t ) and its transposed LTV system Item 1. Notice that for any t ≥ t0 ∈ J, there exists an integer k ≥ 0
ẋ(t ) = A| (t )x(t ) may have totally different stability outcomes. such that t = t0 + kδ + θ , where θ ∈ [0, δ). Then
 B. Zhou / Automatica 68 (2016) 266–276 275

 t k−1  t0 +(i+1)δ  t0 +kδ+θ

 A.5. Proof of inequality (36)
µ(s)ds = µ(s)ds + µ(s)ds
t0 t0 + i δ t0 +kδ
i=0
For any x ∈ Rn , and any s ≥ t ≥ t0 ∈ J, we have
c
≤ −kc + d = − (t − t0 − θ) + d d
δ x| Φ | (s, t ) Φ (s, t ) x
c ds
≤ − (t − t0 ) + d + c , (53)
δ = x| Φ | (s, t ) (A| (s) + A(s)) Φ (s, t ) x
which implies that µ(t ) is UES. ≥ −2x| Φ | (s, t ) |A(s)| Φ (s, t ) x
= −2e−ε(s−t ) x| eε(s−t ) Φ | (s, t ) |A(s)| Φ (s, t ) x
A.3. Proof of Lemma 5 ≥ −2x| eε(s−t ) Φ | (s, t ) |A(s)| Φ (s, t ) x.
Then by (5) we can obtain
‘‘Item 4 ⇒ Item 3’’. This is similar to the proof of Lemma 3 by
−x| x = lim x| Φ | (s, t ) Φ (s, t ) x − x| x
noting that µ (t ) is uniformly bounded. ‘‘Item 3 ⇒ Item 2’’ and s→∞

‘‘Item 2’’ ⇒ ‘‘Item 1’’ are trivial. We only need to show ‘‘Item 1’’ ⇒
 ∞ 
d | |
‘‘Item 4’’. We assume that Item 4 is not true, namely, ϕ (T ) ≥ 0. = x Φ (s, t ) Φ (s, t ) x ds
t ds
Notice that, for any t , t0 ∈ J , t ≥ t0 , there exists an integer k such  ∞ 
that t = t0 + kT + θt , θt ∈ [0, T ). Then we have ε(s−t )
≥ −2x |
e Φ (s, t ) |A(s)| Φ (s, t ) ds x
|

 t  t0 +kT  t0 +kT +θt t

µ (s) ds = µ (s) ds + µ (s) ds = −2x| P (t )x,

t0 t0 t0 +kT
which completes the proof.
k−1 
 t0 +(i+1)T  t0 +kT +θt
= µ (s) ds + µ (s) ds
t0 +iT t0 +kT References
i=0
 t0 +kT +θt Aeyels, D., & Peuteman, J. (1999). Uniform asymptotic stability of linear time-
= kϕ (T ) + µ (s) ds varying systems. In V. Blondel, et al. (Eds.), Open problems in mathematical
systems and control theory (pp. 1–5). London: Springer-Verlag.
t0 +kT
Anderson, B. D. O., Ilchmann, A., & Wirth, F. R. (2013). Stabilizability of linear time-
varying systems. Systems & Control Letters, 62(9), 747–755.
≥ kϕ (T ) − max {|µ (s)|} θt Anderson, B. D. O., & Moore, J. B. (1969). New results in linear system stability. SIAM
s∈[0,T ]
Journal on Control, 7(3), 398–414.
Bellman, R. (1953). Stability theory of differential equations. New York: McGraw-Hill.
≥ −T max {|µ (s)|} , Cao, Y., & Lam, J. (2000). Computation of robust stability bounds for time-delay
s∈[0,T ]
systems with nonlinear time-varying perturbations. International Journal of
Systems Science, 31(3), 359–365.
which contradicts with (15), namely, Item 1. Hence Item 4 is true Haken, R. L., & Naylor, A. W. (1966). A new method of designing time-varying, linear
and the proof is finished. feedback systems. International Journal of Control, 4(2), 137–151.
Harris, C. J., & Miles, J. F. (1980). Stability of linear systems: some aspects of kinematic
similarity. London: Academic Press.
Huang, L., Hollot, O. V., & Xu, Z. (1991). Some problems of Lyapunov method in
A.4. Proof of Eq. (24) linear continuous-time system. Systems Science and Mathematical Sciences, 4(3),
244–253.
Ilchmann, A., Owens, D. H., & Pratzel-Wolters, D. (1987). Sufficient conditions
for stability of linear time-varying systems. Systems & Control Letters, 9,
Clearly, there exist two nonnegative integers k0 and k∗ with 157–163.
k∗ ≥ k0 such that t0 ∈ [k20 , (k0 + 1)2 ) and t ∈ [k2∗ , (k∗ + 1)2 ), or Kalman, R. E., & Bertram, J. E. (1960). Control system analysis and design via the
‘‘second method’’ of Lyapunov-continuous-time systems. Transactions of the
t0 = k20 + θ0 , θ0 ∈ [0, 2k0 + 1) and t = k2∗ + θ∗ , θ∗ ∈ [0, 2k∗ + 1). ASME—Journal of Basic Engineering, 82(2), 371–393.
Then Karafyllis, I., & Tsinias, J. (2003). Non-uniform in time stabilization for linear
systems and tracking control for non-holonomic sysems. International Journal
 t  t  t0 of Control, 26(76), 1536–1546.
Khalil, H. (2002). Nonlinear systems (3rd ed.). New Jersey, NJ, USA: Prentice-Hall Inc.
µ(s)ds = µ(s)ds − µ(s)ds. (54) Malek-zavarei, M. (1978). The stability of linear time-varying systems. International
t0 0 0 Journal of Control, 27(5), 809–815.
Massera, J. L. (1956). Contributions to stability theory. Annals of Mathematics, 64,
Notice that we can compute 182–206.
Mazenc, F., Malisoff, M., & Niculescu, S. (2014). Reduction model approach for linear
time-varying systems with delays. IEEE Transactions on Automatic Control, 59(8),
 t k∗ −1
 (i+1)2  k2∗ +θ∗ 2068–2082.
µ(s)ds = µ(s)ds + µ(s)ds Mullhaupt, Ph., Buccieri, D., & Bonvin, D. (2007). A numerical sufficiently test
for the asymptotic stability of linear time-varying systems. Automatica, 43,
0 i=0 i2 k2∗
632–638.
Petersen, I. R. (1985). A Riccati equation approach to the design of stabilizing
 k2∗ +θ∗ controllers and observers for a class of uncertain linear systems. IEEE
= −2α k2∗ − ρ k∗ + µ(s)ds Transactions on Automatic Control, 30, 904–907.
k2∗ Petersen, I. R. (1988). Quadratic stabilizability of uncertain linear systems
containing both constant and time-varying uncertain parameters. Journal of
Optimization Theory and Applications, 57(3), 439–461.
−2α k∗ + θ∗  − ρ k∗ − θ∗ ρ, if θ∗ ∈ [0, k∗ + 1),
  2 
Perron, O. (1930). Ober eine Matrixtransformation. Mathematische Zeitschrift, 32,

= −2α k2∗ + θ∗ − ρ k∗ − 2ρ (k∗ + 1) + θ∗ ρ, 465–473.
Ramarajan, S. (1986). Time-varying Lyapunov functions for linear time-varying
if θ∗ ∈ [k∗ + 1, 2k∗ + 1),

 systems. International Journal of Control, 44(6), 1699–1702.
Rosenbrock, H. H. (1963). The stability of linear time-dependent control systems.
International Journal of Electronic Control, 15(1), 73–80.
≤ −2α t . (55) Rugh, W. J. (1996). Linear system theory (2nd ed.). Upper Saddle River, New Jersey:
Prentice Hall.
t Solo, V. (1994). On the stability of slowly time-varying linear systems. Mathematics
Similarly, we can get 0 0 µ(s)ds ≥ −2α t0 − ρ (2k0 + 1) substitu- of Control, Signals, and Systems, 7, 331–350.
tion of which and (55) into (54) gives (24). Sun, Z. (2007). A robust stabilizing law for switched linear systems. International
Journal of Control, 77(4), 389–398.
276 B. Zhou / Automatica 68 (2016) 266–276
Tan, F., & Duan, G. (2009). Stability criterion and stabilization of linear time-varying
systems. In: The 48th IEEE conference on decision and control (pp. 3238–3243).
Wu, M. Y. (1974). A note on stability of linear time-varying systems. IEEE
Transactions on Automatic Control, 19, 162–163.
Wu, M. Y., Horowitz, I. M., & Dennison, J. C. (1975). On solution, stability and
transformation of linear time-varying systems. International Journal of Control,
22(2), 169–180.
Xu, S., Lam, J., & Zou, Y. (2009). Robust admissibility of time-varying singular
systems with commensurate time delays. Automatica, 45(11), 2714–2717.
Zheng, D. (2002). Linear system theory (2nd ed.). Beijing, China: Tsinghua University
Press.
Zhou, B., Cai, G., & Duan, G. (2013). Stabilization of time-varying linear systems
via Lyapunov differential equations. International Journal of Control, 86(2),
332–347.
Zhou, B., & Duan, G. (2012). Periodic Lyapunov equation based approaches to the
stabilization of continuous-time periodic linear systems. IEEE Transactions on
Automatic Control, 57(8), 2139–2146.
Zhou, B., Hou, M., & Duan, G. (2013). L∞ and L2 semi-global stabilisation of
continuous-time periodic linear systems with bounded controls. International
Journal of Control, 86(4), 709–720.
Bin Zhou is a professor of the Department of Control
Science and Engineering at the Harbin Institute of Tech-
nology. He received the Bachelor’s degree, the Master’s
Degree and the Ph.D. degree from the Department of
Control Science and Engineering at Harbin Institute of
Technology, Harbin, China in 2004, 2006 and 2010, respec-
tively. His current research interests include constrained
control, time-delay systems, time-varying systems, non-
linear control, multi-agent systems, and control applica-
tions in astronautic engineering. In these areas, he has
published over 80 papers in archival journals. He is the au-
thor of the book Truncated Predictor Feedback for Time-Delay Systems (Springer-
Verlag, Berlin Heidelberg, 2014). He is a reviewer for American Mathematical
Review and is an active reviewer for many journals. He received the ‘‘National Excel-
lent Doctoral Dissertation Award’’ in 2012 from the Academic Degrees Committee
of the State Council and the Ministry of Education of P.R. China. He is currently an
associate editor on the Conference Editorial Board of the IEEE Control Systems So-
ciety and an associate editor of Asian Journal of Control and Journal of System Science
and Mathematical Science.