Hermitian Nonnegative-Definite and Positive-Definite Solutions of The Matrix Equation AXB C
Hermitian Nonnegative-Definite and Positive-Definite Solutions of The Matrix Equation AXB C
Hermitian Nonnegative-Definite and Positive-Definite Solutions of The Matrix Equation AXB C
Abstract
We propose necessary and sufficient conditions for the existence of Hermitian
nonnegative-definite or positive-definite solutions to the matrix equation AXB =
C. A representation of these solutions is also given.
1 Introduction
Many authors have studied various solutions to the matrix equation
AXB = C (1)
with unknown matrix X. For instance, Kuo [7] has studied the structure of the solutions
to (1) by using the tensor product representation of (1). Dai [3] and Khatri and
Mitra [6] have obtained different sufficient and necessary conditions for the existence
of Hermitian (symmetric) solutions and given the general expressions of all Hermitian
(symmetric) solutions to (1). The Hermitian nonnegative-definite solutions to the
special case B = A∗ of (1) have been studied by Baksalary [2], Dai and Lancaster [4],
Groß [5], Khatri and Mitra [6] and also Zhang and Cheng [8]. Furthermore, Khatri
and Mitra [6] investigated a necessary and sufficient condition for the existence of
Hermitian nonnegative definite solutions to (1), and deduced a representation of all
Hermitian nonnegative-definite solutions to (1) if such solutions exist.
The purpose of this paper is to provide a new approach which gives a representation
of all Hermitian nonnegative-definite (respectively, positive-definite) solution to (1).
The proposed approach is different from that of Khatri and Mitra [6] which lies on the
generalized inverses of matrices.
Let Cm×n be the set of m × n complex matrices. We denote by H> ≥
n , Hn , Un and
n×n
GLn the subsets of C consisting of Hermitian positive-definite matrices, Hermitian
nonnegative-definite matrices, unitary matrices and nonsingular matrices, respectively.
For X ∈ Cm×n , let X ∗ , X + and R(X) be, respectively, the conjugate transpose, the
Moore-Penrose inverse and the column space of X. The notation ⊕ denotes the direct
∗ Mathematics Subject Classifications: 15A06, 15A24.
† Department of Mathematics, Heilongjiang University, Harbin, 150080, P. R. China
‡ Presently at School of Mechanical and Manufacturing Engineering, The Queen’s University of
40
X. Zhnag 41
sum. We denote by Ik and Om×n the k × k identity matrix and the m × n zero matrix,
respectively; We also write them as I and O respectively when their dimensions are
clear.
Let A ∈ Cm×n , B ∈ Cn×p and C ∈ Cm×p . Since
are necessary conditions for the existence of solutions to (1), in the following we will
always assume that (2) is satisfied. Furthermore, we have the following two observa-
tions.
NOTE 1. If A = O or B = O, then there exists a Hermitian nonnegative-definite
(respectively, positive-definite) solution to (1) if and only if C = O. In this case,
any matrix in H≥ >
n (respectively, Hn ) is a Hermitian nonnegative-definite (respectively,
positive-definite) solution to (1).
T
NOTE 2. If rankA < m, then A = PA AT1 O , where PA ∈ Um and A1 is of
full-row rank. Let
T
PA∗ C = C1T C2T , (3)
where C1 and C2 have appropriate sizes. Combining (2) and (3) yields C2 = O. This
implies that the pair of matrix equations (1) and A1 XB = C1 have the same solutions.
Therefore, we may assume that rankA = m. Similarly, we may also assume that
rankB = p.
Based on the above two notes, in the rest of this paper we assume
Obviously, (2) holds if (4) is satisfied. Therefore, we are now concerned with the
following problem.
PROBLEM 1. Given matrices A ∈ Cm×n , B ∈ Cn×p and C ∈ Cm×p satisfying
(4). Determine a necessary and sufficient condition for the existence of Hermitian
nonnegative-definite (respectively, positive-definite) solutions to (1). Furthermore, give
a representation of all Hermitian nonnegative-definite (respectively, positive-definite)
solutions to (1) when it has such solutions.
Now we introduce two special equivalent decompositions of matrices. For an arbi-
trary but fixed E ∈ Cm×n with rankE = r, let
E = U1 (Σ ⊕ O)V1 (5)
ensure good numerical reliability. Furthermore, (5) can be easily obtained by using the
function svd in MATLAB.
In order to solve problem 1, some preliminary results are given in the next section.
We solve Problem 1 in Section 3. An example is presented in Section 4 to illustrate
our approach.
2 Preliminary results
We first introduce the following lemma.
M Y
LEMMA 1. ([1]) Given matrix Ψ = with M ∈ Cn1 ×n1 , N ∈ Cn2 ×n2
Y∗ N
and Y ∈ Cn1 ×n2 . Then
(i) Ψ ∈ H≥ ≥ ∗ + ≥
n1 +n2 if and only if M ∈ Hn1 , N − Y M Y ∈ Hn2 and R(Y ) ⊆ R(M ).
> ∗ −1
(ii) Ψ ∈ Hn1 +n2 if and only if M ∈ Hn1 and N − Y M Y ∈ H>
>
n2 .
To solve Problem 1, we need the following algorithm which produces an integer
s and three matrices P ∈ GLm , T ∈ GLp and Q ∈ GLn based on two matrices
A ∈ Cm×n and B ∈ Cn×p satisfying (4).
ALGORITHM 1.
B1
= Q1 B. (7)
B2
Step 6 Calculate the matrices B11 ∈ Cm×s and B12 ∈ Cm×(p−s) according to
Ip−s
P3 Q3 . (Note that: by (4), (7), (8) and (9), rankB12 = p − s can be
O
deduced as follows.)
B11 B12 B11 B12 Q2
rankB12 = rank Is O = Is O −s
O O P2 Q 2
O O
B1
= rank − s = rankB − s = p − s.
B2
O Q3
P = P1 P3 , T = Q2
Is O
and
P3∗ O O
∗ Im O
Q = B11 Is O Q1 .
O P2∗
O O In−m−s
To some extent the following lemma reveals the sense of the above algorithm.
LEMMA 2. Given matrices A ∈ Cm×n and B ∈ Cn×p satisfying (4). Let the
integer s and the matrices P , Q and T be obtained by Algorithm 1. Then
T
Ip−s O O O
A=P Im O Q, B = Q∗ T. (10)
O Os×(m−p+s) Is O
where X14 , X22 , X24 , X33 , X34 and X44 are parameter matrices with appropriate
sizes which satisfy
R (X14 ) ⊆ R (C1 )
≥
Y1 ∈ Hm−p+s
R C0 Y2 ⊆ R (Y1 )
∗ + (14)
Y
3
− C Y
0 1 C0 ∈ H≥ s
R (Z) ⊆ R Y3 − C0∗ Y1+ C0
+
Y5 − Y2∗ Y1+ Y2 − Z ∗ Y3 − C0∗ Y1+ C0 Z ∈ H≥
n−m−s
with
C0 = C4 − C3 C1+ C2
Y1 = X22 − C3 C1+ C3∗
Y2 = X24 − C3 C1+ X14
Y3 = X33 − C2∗ C1+ C2 . (15)
∗ +
Y4 = X34 − C C
2 1 X14
∗
Y5 = X44 − X14 C1+ X14
∗ +
Z = Y4 − C0 Y1 Y2
where X11 ∈ H≥ ≥ ≥
p−s , X22 ∈ Hm−p+s and X33 ∈ Hs . Using Lemma 2, (17) and (18)
yields
X11 X13
C=P ∗ T.
X12 X23
This, together with (11), implies
C1 C3∗ C2 X14
C3 X22 C4 X24
QX̃Q =
∗
C2∗
. (19)
C4∗ X33 X34
∗ ∗ ∗
X14 X24 X34 X44
Combining (19), (i) of Lemma 1 and X̃ ∈ H≥ n deduces (12).
(ii) It follows from (i) and (12) that the matrix equation (1) has at least a Her-
mitian nonnegative-definite solution. Suppose X is an arbitrary but fixed Hermitian
nonnegative-definite solution to (1). Then, by a similar argument to the proof of (19),
the matrix X possesses the form (13). Now, to complete the proof, it suffices to show
that X ∈ H≥ n is equivalent to (14). Indeed, by (i) of Lemma 1, (12) and (13), we have
that X ∈ H≥ n if and only if
Y1 C0 Y2
R (X14 ) ⊆ R (C1 ) , C0∗ Y3 Y4 ∈ H≥ n−p+s , (20)
Y2∗ Y4∗ Y5
where C0 and Yi , i = 1, · · · , 5, are defined in (15). Again applying (i) of Lemma 1 to
the second relation in (20), we see that (20) is equivalent to
R (X14 ) ⊆ R (C1 )
≥
Y1 ∈ Hm−p+s
R C0 Y2 ⊆ R (Y1 ) . (21)
∗ + ∗ +
Y − C Y C Y − C Y Y
3 0 1 0 4 0 1 2
∈ H≥
Y4∗ − Y2∗ Y1+ C0 Y5 − Y2∗ Y1+ Y2 n−m
with (15).
46 Hermitian Solutions of AXB = C
4 An Example
Consider a matrix equation in the form of (1) with the parameter matrices:
1 2 0
0 5 9 3 −1 2 1 1 −1 2
A= , B=
6
, C = .
2 2 1 3 2 −2 1 0 2
3 1 −1
b1 c1
where a, b, c, d ∈ C are parameters satisfying a > 2.4448 and ∈ H≥
2 with
c̄1 d1
−3 0.0000
b1 = b − 10 × 0.0106 − a−2.4448
0.0004
c1 = c + 10−3 × 0.0546 + a−2.4448 .
0.0004
−3
d1 = d − 10 × 0.2824 − a−2.4448
References
[1] A. Albert, Condition for positive and nonnegative definite in terms of pseudoinverse,
SIAM J. Appl. Math., 17(1969), 434—440.
[2] J. K. Baksalary, Nonnegative definite and positive definite solutions to the matrix
equation AXA∗ = B, Linear and Multilinear Algebra, 16(1984), 133—139.
[3] H. Dai, On the symmetric solutions of linear matrix equations, Linear Algebra
Appl., 131(1990), 1—7.
[4] H. Dai and P. Lancaster, Linear matrix equations from an inverse problem of vi-
bration theory, Linear Algebra Appl., 246(1996), 31—47.
[5] J. Groß, Nonnegative-define and positive-definite solutions to the matrix equation
AXA∗ = B– revisited, Linear Algebra Appl., 321(2000), 123—129.
[6] C. G. Khatri and S. K. Mitra, Hermitian and nonnegative definite solutions of linear
matrix equations, SIAM J. Appl. Math., 31(1976), 579—585.
[7] Y. E. Kuo, The use of tensor products for solving matrix equations, Matrix Tensor
Quart., 26(4)(1975/76), 148—151.
[8] X. Zhang and M. Y. Cheng, The rank-constrained Hermitian nonnegative-definite
and positive-definite solutions to the matrix equation AXA∗ = B, Linear Algebra
Appl., 370(2003), 163—174.