Some Properties of Generalized Lyapunov Equations

Weihai Zhang, and Bor-Sen Chen

Abstract— By means of the spectral analysis method, this a unique real symmetric (skew-symmetric) solution? Under
paper studies a class of generalized Lyapunov equations (GLEs) what conditions, does (1) has no or has more than one real
arising from stochastic stability. A necessary and sufficient symmetric (skew-symmetric) solution? The aforementioned
condition is given for the existence and uniqueness of the
symmetric/skew-symmetric solutions of such GLEs. A sufficient problems will be solved in this paper, which will help us to
condition for the GLEs have no and have more than one better understand the solution structure of GLE (1).
symmetric/skew-symmetric solution is also presented, respec- This paper uses a class of generalized Lyapunov operators
tively. to study GLE (1). It is shown that the existence of the
Index Terms— Stochastic systems, generalized Lyapunov
symmetric solutions of GLE (1) is closely associated with
equations, operator spectrum, symmetric solution, skew-
symmetric solution the linear symmetric operator LA,C1 ,C2 ,...,Ck which was first
introduced in [9] and is called “a generalized Lyapunov
I. INTRODUCTION operator”. This paper also introduces a new linear skew-
It is well known that the following generalized Lyapunov symmetric operator L− A,C1 ,C2 ,...,Ck , which is used to present
equation (GLE): the conditions for the existence of skew-symmetric solutions
of GLE (1). What we have obtained in this paper are
k
 new even when GLE (1) reduces to the classical Lyapunov
P A + AT P + CiT P Ci = −Q. (1)
equation (3).
i=1
The organization of this paper is as follows: In Section
plays an important role in studying the mean square stability II, we make some preliminaries, where some useful maps
of the stochastic Itô-type system are defined, and some used lemmas are given. Section III
k
 contains our main results, several theorems about the solution
dx(t) = Ax(t) dt + Ci x(t) dwi (t), x(0) = x0 ∈ Rn structure of GLE (1) are shown. Section IV concludes this
i=1 paper with some remarks.
(2)
For convenience, we adopt the following standard nota-
where A, C1 , · · · , Ck are real constant matrices of suitable
tions:
dimension, x ∈ Rn is the system state, x0 ∈ Rn is a deter-
R (C): the set of all real (complex) numbers. Rm×n : the
ministic initial state, wi (t), i = 1, · · · , k, are independent,
set of all m × n real matrices. Sn (S−n ): the set of all n × n
standard one-dimensional Wiener processes defined on the
symmetric (skew-symmetric) matrices whose entries may be
filtered probability space (Ω, F, P; Ft ); see [4], [5], [6], [9]
complex. AT : the transpose of a matrix A. A > 0(A ≥ 0): A
and the references therein. The GLE (1) is an extension of
is a positive definite (semi-definite) symmetric matrix. σ(L):
the classical Lyapunov equation
the spectrum set of the operator or matrix L.
P A + AT P = −Q (3)
II. S OME P RELIMINARIES
and plays the same role as (3) has done in modern control
theory. In was shown in [4] that systems (2) is asymptotically In order to prove our main theorems, we need to introduce
mean square stable if and only if (1) has a positive solution some definitions and prove a few lemmas that will be used.
−→
P > 0. [6] and [9] further generalized the result of (2) to Definition 2.1. ψ : Xn = (xij )n×n ∈ Rn×n → Xn ∈
the case Q ≥ 0 with the assumptions of exact detectability
2
Rn , ϕ : Xn = (xij )n×n ∈ Sn → X n ∈ Rn(n+1)/2 , where
and observability, respectively. −→
Xn = (x11 , · · · , x1n , x21 , · · · , x2n , · · · , xn1 , · · · , xnn )T ,
Generally speaking, compared with the classical Lyapunov
equation (3), we know little about the properties the solutions
of (1). For example, under what conditions, does (1) has n = (x11 , · · · , x1n , x22 , · · · , x2n , · · · , xn−1,n−1 , xn−1,n , xnn )T .
X
This work was supported by the National Natural Science Foundation of
China (No. 60874032), the Key Project of Natural Science Foundation of
Shandong Province (No. ZR2009GZ001), and Specialized Research Fund Definition 2.2. ϕ− : X−n = (xij )n×n ∈ S−n → X−n ∈
for the Doctoral Program of Higher Education (No. 20103718110006) Rn(n−1)/2 , where
Weihai Zhang is with the College of Information and Electrical Engi-
neering, Shandong University of Science and Technology, Qingdao 266510, X−n = (x12 , · · · , x1n , x23 , · · · , x2n , · · · , xn−1,n )T .
China w [email protected]
Bor-Sen Chen is with the Department of Electrical Engineering,
National Tsing Hua University, Hsinchu 30013, Taiwan That is, X−n consists of all upper triangular elements of
[email protected] X−n .

978-1-4244-8738-7/11/$26.00 2011
c IEEE 3137

Authorized licensed use limited to: National Tsing Hua Univ.. Downloaded on July 16,2024 at 15:40:27 UTC from IEEE Xplore. Restrictions apply.
 Remark 2.1. ψ, ϕ and ϕ− are obviously invertible. It We recall that in order to study the mean square stability
should be noted that ψ, ϕ and ϕ− are defined in different of (2), [9] first introduced a generalized Lyapunov operator
domains. as follows:
−→
Note that Xn is a column vector formed by all elements k
of Xn [3], while X n is a column vector formed by different LA,C1 ,C2 ,...,Ck : X ∈ Sn → AX+XAT + Ci XCiT ∈ Sn .
elements of Xn . It is easy to show that for Xn ∈ Sn , there i=1
is a singular n2 × n(n + 1)/2 transformation matrix Hn It was shown in [9] that (2) is asymptotically mean square
from Xn to − → −→
Xn satisfying Xn = Hn X n . In order for the stable iff σ(LA,C1 ,C2 ,...,Ck ) ⊂ C − , where
uniqueness of such Hn , we confine Hn to possessing of a
special structure that in each row of Hn , only one entry is 1, σ(LA,C1 ,C2 ,...,Ck )
and the other entries are zero. For the reader’s convenience, := {λ ∈ C : LA,C1 ,C2 ,...,Ck X = λX, X ∈ Sn , X = 0}.
we take a 2 × 2 matrix as an example. LA,C1 ,C2 ,...,Ck is a finite-dimensional linear symmetric op-
Example 2.1. Let X2 = (xij )2×2 ∈ S2 , then erator. Similar to LA,C1 ,C2 ,...,Ck , we introduce the following
⎡ ⎤ skew-symmetric generalized Lyapunov operator defined on
x11 ⎡ ⎤ S−n :
x11
−→ ⎢ x ⎥
 ⎣
X2 = ⎢ 12 ⎥
⎣ x12 ⎦ , X2 = x12 .
⎦ 
k

x22 L− T
A,C1 ,C2 ,...,Ck : X ∈ S−n → AX + XA + Ci XCiT ∈ S−n
x22 i=1

Because ⎡ ⎤ and the spectral set of L−

A,C1 ,C2 ,...,Ck can be defined as
1 0 0
−→ ⎢ 0 1 0 ⎥ , σ(L−
A,C1 ,C2 ,...,Ck )
X2 = ⎢
⎣ 0
⎥X
1 0 ⎦ 2 := {λ ∈ C : L−
A,C1 ,C2 ,...,Ck X = λX, X ∈ S−n , X = 0}.
0 0 1
Associated with L− A,C1 ,C2 ,...,Ck and LA,C1 ,C2 ,...,Ck , we
we have ⎡ ⎤ introduce the following two matrices induced respectively
1 0 0
⎢ 0 1 0 ⎥ by L−A,C1 ,C2 ,...,Ck and LA,C1 ,C2 ,...,Ck :
H2 = ⎢
⎣ 0 1
⎥.
0 ⎦ Θ(H−n ) := T
[H−n H−n ]−1 H−n
T
0 0 1 k

Example 2.2. Let X−2 = (xij )2×2 ∈ S−2 , then ·(AT ⊗ I + I ⊗ AT + CiT ⊗ CiT )H−n
i=1
−−→ T
X−2 = 0 x12 −x12 0 , X−2 = x12 . and
So Θ(Hn ) := [HnT Hn ]−1 HnT
−−→ T
X−2 = 0 1 −1 0 x12 , k

T ·(AT ⊗ I + I ⊗ AT + CiT ⊗ CiT )Hn .
H−2 = 0 1 −1 0 . i=1

For Hn and H−n , we have the following property: Lemma 2.4. (i) σ(L−
A,C1 ,C2 ,...,Ck ) = σ(Θ(H−n )). (ii)
Lemma 2.1. Both [HnT Hn ] and [H−n
T
H−n ] are invertible. σ(LA,C1 ,C2 ,...,Ck ) = σ(Θ(Hn )).
Proof. We notice that the columns of Hn and H−n
are linearly independent, i.e., rank(Hn ) = n(n + 1)/2, Proof. To prove (i), we first proceed to show
rank(H−n ) = n(n − 1)/2. According to matrix theory, σ(L− A,C1 ,C2 ,...,Ck ) ⊂ σ(Θ(H−n )). Suppose λ is any given
rank[HnT Hn ]n(n+1)/2×n(n+1)/2 = rank(Hn ) = n(n + 1)/2, eigenvalue of L− A,C1 ,C2 ,...,Ck , and the nonzero symmetric
matrix X−n ∈ S−n is the corresponding eigenvector, then
T
we have
rank[H−n H−n ]n(n−1)/2×n(n−1)/2 = rank(H−n ) = n(n−1)/2.
− T
k
T
This lemma is proved. LA,C1 ,C2 ,...,Ck X−n = AX−n + X−n A + Ci X−n Ci
i=1
The following lemma will be used later.
= λX−n , X−n = 0. (4)
Lemma 2.2 [3]. For the matrices A, P and B of suitable
−−−→ −
→
dimension, AP B = (A ⊗ B T ) P , where A ⊗ B represents Taking ψ operator on (4) and using Lemma 2.2, it yields
the Kronecker product of two matrices A and B T . k
 −−→ −−→ −−→
The following lemma is obvious by Definitions 2.1-2.2. (A ⊗ I + I ⊗ A + Ci ⊗ Ci )X−n = λX−n , X−n = 0
i=1
Lemma 2.3. (i) For Xn ∈ Sn , ϕ(Xn ) = 0 iff Xn = 0. which gives
2
X−n ∈ S−n , ϕ− (X−n ) = 0 iff X−n = 0. (ii) For η ∈ Rn ,
if ψ −1 (η) ∈ Sn , then there exists ξ ∈ Rn(n+1)/2 , such that 
k

2 (A⊗I +I ⊗A+ Ci ⊗Ci )H−n X−n = λH−n X−n , X−n = 0.

η = Hn ξ. (iii) For η ∈ Rn , if ψ −1 (η) ∈ S−n , then there i=1
exists ξ ∈ Rn(n−1)/2 , such that η = H−n ξ. (5)

3138 2011 Chinese Control and Decision Conference (CCDC)

Authorized licensed use limited to: National Tsing Hua Univ.. Downloaded on July 16,2024 at 15:40:27 UTC from IEEE Xplore. Restrictions apply.
 T
Pre-multiplying H−n on both sides of (5), it yields The above leads to
⎧ k
T
⎨ H−n (A ⊗ I + I ⊗ A + i=1 Ci ⊗ Ci )H−n X−n +Q
P = −Θ−1 (Hn )(Q  1 ). (13)
T (6)
= λH−n H−n X−n , In view of our definition for P = ϕ (ζ), we have −1
⎩
X−n = 0. ζ = ϕ(P ) = P. Hence, from (8) and (13), one has
T −Θ−1 (Hn )Q  1 = 0, which gives ϕ(Q1 ) = Q  1 = 0. By
By Lemma 2.1, [H−n H−n ]−1 exists, we have −1
Lemma 2.3, Q1 = 0. In view of (10), P = ϕ (ζ) is actually

Θ(H−n )X −n = λX−n , a symmetric solution to (1).
To prove the uniqueness of the solutions of (1), we proceed
which shows that λ and X−n are respectively the eigen- as follows: If (1) admits two symmetric solutions P1 and P2 ,
value and eigenvector of Θ(H−n ). So σ(L− A,C1 ,C2 ,...,Ck ) ⊂ i.e.,
σ(Θ(H−n )).  k
Secondly, because both the operator L− A,C1 ,C2 ,...,Ck and P1 A + AT P1 + CiT P1 Ci = −Q, (14)
the matrix Θ(H−n ) have and only have n(n − 1)/2 eigen- i=1
values and n(n − 1)/2 linear independent eigenvectors, k

we must have σ(L− A,C1 ,C2 ,...,Ck ) = σ(Θ(H−n ) from P2 A + AT P2 + CiT P2 Ci = −Q. (15)
σ(L−A,C1 ,C2 ,...,Ck ) ⊂ σ(Θ(H −n )). (i) is proved. i=1
Repeating the same procedure as in (i), (ii) can be proved. Subtracting (15) from (14) gives
III. M AIN R ESULTS k

(P1 − P2 )A + AT (P1 − P2 ) + CiT (P1 − P2 )Ci = 0.
In this section, we present several properties of GLE (1),
i=1
our first theorem on GLE (1) is given as follows:
Theorem 3.1. For Q ∈ Sn , the GLE (1) has a unique Repeating the same procedure as in the derivation of (12), it
symmetric solution P iff σ(LA,C1 ,C2 ,...,Ck ) does not contain follows
Θ(Hn )(P1 − P2 ) = 0,
zero eigenvalue.
Proof. Sufficiency: By Lemma 2.4, 1 = P
and P 2 is derived duo to the invertibility of Θ(Hn ).
By Lemma 2.3, we have P1 = P2 , the uniqueness is proved.
σ(LA,C1 ,C2 ,...,Ck ) = σ(Θ(Hn )). (7)
Necessity: We prove it by contradiction. Assume
So we can assume Θ(Hn ) is invertible. Set σ(LA,C1 ,C2 ,...,Ck ) contains zero as an eigenvalue, then
 Θ(Hn ) is singular by Lemma 2.4. So the equation
ζ = −Θ−1 (Hn )ϕ(Q) = −Θ−1 (Hn )Q. (8)

Θ(Hn )x = −Q (16)
By Remark 2.1, ϕ is an invertible map, we now prove P :=
ϕ−1 (ζ) ∈ Sn is a unique symmetric solution to (1). We set admits either more than one solution or no solution. Because
Q1 as ϕ(P ) satisfies (16) with P a unique solution to (1), it follows
that (16) must have more than one solution. We suppose
k
 ζ ∈ Rn(n+1)/2 is another solution which differs from ϕ(P ).
−Q1 = P A + AT P + CiT P Ci + Q. (9)
Noting that (16) is equivalent to
i=1
 k

Obviously, Q1 ∈ Sn . (10) indicates that P is a solution to 
T T T T T
Hn I ⊗ A + A ⊗ I + Ci ⊗ Ci Hn x
k
 i=1
P A + AT P + CiT P Ci = −Q − Q1 . (10) 
= −HnT Hn Q. (17)
i=1

In what follows, we will prove Q1 = 0. By Lemma 2.2, (10) Substituting ζ into (17) yields
 k

yields 
 k
 HnT T
I ⊗A +A ⊗I + T
CiT ⊗ CiT Hn ζ
 −
→ −−−−−−→ i=1
I ⊗ A T + AT ⊗ I + CiT ⊗ CiT P = −(Q + Q1 ).

= −HnT Hn Q. (18)
i=1

and further yields Let

 
k
 −Δ = ϕ−1 (ζ)A + AT ϕ−1 (ζ) + C T ϕ−1 (ζ)C + Q
I ⊗ A T + AT ⊗ I + CiT ⊗ CiT Hn P
i=1
then ϕ−1 (ζ) solves the following GLE
+Q
= −Hn (Q  1 ). (11) ϕ−1 (ζ)A + AT ϕ−1 (ζ) + C T ϕ−1 (ζ)C = −Q − Δ. (19)
Pre-multiplying HnT on both sides of (11), and noting that By (19), we have
HnT Hn is invertible by Lemma 2.1, we have 
k

(I ⊗ AT + AT ⊗ I +  + Δ).
CiT ⊗ CiT )Hn ζ = −Hn (Q  (20)
Θ(Hn )P = −(Q
+Q
 1 ). (12) i=1

2011 Chinese Control and Decision Conference (CCDC) 3139

Authorized licensed use limited to: National Tsing Hua Univ.. Downloaded on July 16,2024 at 15:40:27 UTC from IEEE Xplore. Restrictions apply.
Substituting (20) into (18), we have we will prove that η also solves (22). Let
 = −H T Hn (Q
−HnT Hn Q  + Δ)
 k

n
(I ⊗ AT + AT ⊗ I + CiT ⊗ CiT )Hn η = η1 (26)
which leads to Δ  = 0 due to the invertibility of HnT Hn . i=1
So we obtain two different solutions ϕ−1 (ζ) and P to (1), 2
so we only need to show η1 ∈ Rn is equal to zero. (26)
which contradicts the uniqueness of the solutions of (1). The
can be written as
necessity proof is complete.
k

Remark 3.1. Theorem 3.1 can be viewed as an extension
ϕ−1 (η)A + AT ϕ−1 (η) + CiT ϕ−1 (η)Ci = ψ −1 (η1 ).
and improved version of Lemma 2 of [10], where only a
i=1
sufficient condition for the existence and uniqueness of the (27)
symmetric solutions of GLE (1) with k = 1 was obtained Because the left-hand side of (27) is symmetric, we have
via very complicated method. ψ −1 (η1 ) ∈ Sn . Hence, by Lemma 2.3, there is η2 ∈
Following the same line as in Theorem 3.1, it is easy to Rn(n+1)/2 such that
obtain the following theorem, and the proof is omitted.
Theorem 3.2. For Q ∈ S−n , the GLE (1) has a unique η1 = Hn η2 . (28)
skew-symmetric solution P iff σ(L− A,C1 ,C2 ,...,Ck ) does not Substituting (26) and (28) into (25), it follows
contain zero eigenvalue.
Now, what will happen for the solutions of GLE (1) when HnT Hn η2 = 0
0 ∈ σ(LA,C1 ,C2 ,...,Ck )? To this end, we first investigate the
which deduces η2 = 0 due to the invertibility of HnT Hn , so
homogeneous GLE
η1 = 0, i.e., (24) and (22) have the same solution.
k
 Summarizing the above analysis, we conclude that if
P A + AT P + CiT P Ci = 0. (21) P ∈ Sn is a solution to (21), then P is a solution to (23);
i=1 Conversely, if η ∈ Rn(n+1)/2 is a solution to (23), then
Lemma 3.1. (i) All symmetric (or skew-symmetric) so- ϕ−1 (η) is a solution to (21). By the differential equation
lutions of the homogeneous GLE (21) constitute a vector theory, if rank(Θ(Hn )) = p, then (23) has n(n + 1)/2 −
space Υn (Υ−n ). (ii) When rank(Θ(Hn )) = p for 0 ≤ p ≤ p linear independent solutions e1 , e2 , · · · , en(n+1)/2−p ,
n(n + 1)/2, then dim(Υn ) = n(n + 1)/2 − p. (iii) When while ϕ−1 (e1 ), ϕ−1 (e2 ), · · · , ϕ−1 (en(n+1)/2−p ) constitute
rank(Θ(H−n )) = p, then dim(Υ−n ) = n(n − 1)/2 − p. the bases of Υn , so dim(Θ(Hn )) = n(n + 1)/2 − p. The
Proof. (i) is obvious and the proof is omitted. To prove (ii), proof of (ii) is complete. Repeating the same procedure, (iii)
we note that from the sufficiency proof of Theorem 3.1, P can be proved.
−
→ Using Lemma 3.1 and the the well known result in matrix
is a real symmetric solution of GLE (21) iff P is a solution
to theory about the relationship between the solutions of the
k homogeneous equation Θ(Hn )x = 0 (or Θ(H−n )x = 0)
−
→
(I ⊗ AT + AT ⊗ I + CiT ⊗ CiT t) P = 0 and Θ(Hn )x = −Q  (or Θ(H−n )x = −Q),  it is easy to
i=1 conclude the following result:
which is further equivalent to that P is a solution to GLE Theorem 3.3.
(i) Assume 0 ∈ σ(LA,C1 ,C2 ,...,Ck ), Q ∈ Sn . If
k
 rank(Θ(Hn ), Q)  = rank(Θ(Hn )), then GLE (1) does not
(I ⊗ AT + AT ⊗ I + CiT ⊗ CiT t)Hn P = 0. (22) have any real symmetric solution. If rank(Θ(Hn ), Q)  =
i=1
rank(Θ(Hn ) = p for 0 ≤ p < n(n + 1)/2, then (1) has
In addition, from the proof of Theorem 3.1, we can also see an infinity of real symmetric solutions. If P1 is a special
that P is a real symmetric solution of GLE (21) iff P solves solution of (1), then any solution of (1) can be expressed as
the following homogeneous linear vector-valued equation P = P0 + P1 , where P0 ∈ Υ.
(ii) Assume 0 ∈ σ(L− A,C1 ,C2 ,...,Ck ), Q ∈ S−n . If
Θ(Hn )P = 0. (23)  = rank(Θ(H−n )), then GLE (1)
rank(Θ(H−n ), Q)
By our definition of Θ(Hn ), GLE (23) is equivalent to does not have any real skew-symmetric solution. If
rank(Θ(H−n ), Q)  = rank(Θ(H−n ) = p for 0 ≤ p <
k
 n(n + 1)/2, then (1) has an infinity of real skew-symmetric
HnT (I T
⊗A +A ⊗I + T
CiT ⊗ CiT )Hn P = 0 (24)
solutions. If P1 is a special solution of (1), then any solution
i=1
of (1) can be expressed as P = P0 + P1 , where P0 ∈ Υ−n .
Now, we proceed to show (22) is equivalent to (24). Obvi-
ously, if η := P ∈ Rn(n+1)/2 is a solution to (22), then so IV. CONCLUSIONS
is (24). Conversely, if η solves (24), i.e., This paper has revealed the relation between the spectra of
k a class of generalized Lyapunov operators and the solution

HnT (I ⊗ AT + AT ⊗ I + CiT ⊗ CiT )Hn η = 0 (25) properties of GLE (1). To our best knowledge, Theorems 3.1-
i=1 3.3 are still new even for the classical Lyapunov equation (3).

3140 2011 Chinese Control and Decision Conference (CCDC)

Authorized licensed use limited to: National Tsing Hua Univ.. Downloaded on July 16,2024 at 15:40:27 UTC from IEEE Xplore. Restrictions apply.
In addition, the results of this paper can be easily generalized
to discrete-time Lyapunov equations coming from discrete-
time stochastic systems with multiplicative noise ([1], [2]),
which remains further study.
R EFERENCES
[1] A. El Bouhtouri, D. Hinrichsen, A. J. Pritchard, H∞ -type control
for discrete-time stochastic systems. International Journal of Robust
Nonlinear Control, vol. 9, 1999, pp. 923C948.
[2] C. S. Kubrusly, O. L. V. Costa, Mean square stability conditions for
discrete stochastic bilinear systems, IEEE Trans. Automatic Control,
vol. 30, 1985, pp. 1082-1087.
[3] R. Bellman, Introduction to Matrix Analysis, Philadelphia: SIAM,
1995.
[4] G. Da Prato, J. Zabczyk, Stochastic equations in infinite dimensions,
Cambridge University Press, Cambridge, MA: 1992.
[5] M. D. Fragoso, O. L. V. Costa, and C. E. de Souza, A new approach
to linearly perturbed Riccati equations arising in stochastic control,
Appl. Math. Optim., vol. 37, 1998, pp. 99C126.
[6] W. Zhang, H. Zhang, B. S. Chen, Generalized Lyapunov equation
approach to state-dependent stochastic stabilization/detectability crite-
rion, IEEE Trans. Automatic Control, vol. 53, 2008, pp 1630-1642.
[7] B. S. Chen, W. Zhang, Stochastic H2 /H∞ control with state-
dependent noise, IEEE Trans. Automatic Control, vol. 49, 2004, pp45-
57.
[8] W. Zhang, L. Xie, Interval stability and stabilization of linear stochas-
tic systems, IEEE Trans. Automatic Control, vol. 54, pp. 810-815,
2009.
[9] W. Zhang, B. S. Chen, On stabilization and exact observability of
stochastic systems with their applications, Automatica, vol. 40, 2004,
pp 87-94.
[10] T. Hou, W. Zhang, H. Ma, Essential Instability and Essential Desta-
bilisation of Linear Stochastic Systems, IET Control Theory & Appli-
cations, accepted for publication

2011 Chinese Control and Decision Conference (CCDC) 3141

Authorized licensed use limited to: National Tsing Hua Univ.. Downloaded on July 16,2024 at 15:40:27 UTC from IEEE Xplore. Restrictions apply.