1506

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 51, NO. 9, SEPTEMBER 2006

Eigenvalue Inequalities for Matrix Product

where jAj2 = max (A) is the spectral norm, i.e., the largest singular ^ = (1)=(2)(A + AT ), it is stated value of A. In the same setting, with A in [3, Th. 2] that

AbstractWe present a family of eigenvalue inequalities for the product of a Hermitian matrix and a positivesemidenite matrix. Our theorem contains or extends some existing results on trace and eigenvalues. Index TermsEigenvalue, inequality, singular value, trace.

^)tr(B ): ^)tr(B )  tr(AB )  max (A 0max (0A

(3)

I. INTRODUCTION Because of deep connections with control theory, dynamical systems, and other related areas, Lyapunov and Riccati equations have been extensively studied, and the problem of estimating the solutions to the equations is of central importance. In [1], some lower and upper bounds for the trace of the solutions to both equations are explicitly presented, while in [2][9], much work is done particularly on the bounds of traces and the extreme eigenvalues of matrix product. We take, for instance, the Lyapunov equation in the matrix form

The inequalities in (3) are extended in [5, Lemma II.1] to complex matrices with the eigenvalues of A and B involved in a symmetric form: For n 2 n Hermitian matrices A and B , assuming that all eigenvalues are arranged in decreasing order 1 (X )  1 1 1  n (X ) for n 2 n Hermitian matrix X , then
n i=1

n0i+1 (A)i (B )  tr(AB ) 

n i=1

 i (A ) i (B ):

(4)

More generally, [5, Lemma II.2] for any complex A and Hermitian B , both n 2 n
n

AT X + XA = 0Q

(LE)

where Q is a real positivedenite matrix. The equation has a positivedenite solution X if and only if matrix A is stable, i.e., all of its eigenvalues have negative real parts. Given a stable matrix A and a matrix Q of the same size, by taking trace, a necessary condition for X to be a solution to the equation is

i=1

~)i (B )  <(tr(AB ))  n0i+1 (A

n i=1

~)i (B ) i (A

(5)

)T is the conjugate trans~ = (1)=(2)(A + A3 ) and A3 = (A where A pose matrix of the matrix A. In particular, if A and B are n 2 n real and if, in addition, B is symmetric, then
n i=1

1 tr(XA) = 0 tr(Q): 2
This is one of the reasons leading to the study of the trace of the product of two matrices, one of which is assumed to be positive semidenite. It is shown in [1, Lemma 1] that for real symmetric matrix A and real positivesemidenite (or nonnegative denite) matrix B of the same size

~)i (B )  tr(AB )  n0i+1 (A

n i=1

~)i (B ): i (A

(6)

Relaxing the symmetry of B and using singular values (arranged in decreasing order), it is shown in [8, Th. 1] that for any n 2 n real matrices A and B

min (A)tr(B )  tr(AB )  max (A)tr(B )

min (AS )

n i=1

i (B )  tr(AB )  max (AS )

n i=1

i (B ) (7)

(1)

where min (A) and max (A) are the smallest and largest eigenvalues of A, respectively. We note that (1) holds when A and B are switched in (1). With these inequalities, some trace bounds on the solutions of the algebraic matrix Riccati and Lyapunov equations are obtained in [1]. Under the condition that A be arbitrary and B be positive semidenite, both n 2 n and real, it is proven in [2] that

tr(AB )  jtr(AB )j  jAj2 tr(B )

(2)

Manuscript received October 17, 2005; revised February 2, 2006. Recommended by Associate Editor A. Hansson. This work was supported in part by the NSU-FAR Faculty Development Funds and the China National Nature Science Foundation under Grant 60574011. F. Zhang is with the Division of Math, Science and Technology, Nova Southeastern University, Fort Lauderdale, FL 33314 USA, and also with Shenyang Normal University, Shenyang, Liaoning Province, China (e-mail: zhang@nova. edu). Q. Zhang is with the College of Sciences, Northeastern University, Shenyang, Liaoning Province 110004, China (e-mail: [email protected]). Digital Object Identier 10.1109/TAC.2006.880787

where S = U V T if B = U DV T is a singular value decomposition of B . It seems that much of the effort has been made on removing the symmetry or the positive deniteness of A and B , though the proofs of most of the results boil down to those of Hermitian or symmetric matrices. We note that for any square complex matrix A, if we de^ = (1)=(2)(A + AT ) and A ~ = (1)=(2)(A + A3 ); then note by A ^ ~ ) if B tr(AB ) = tr(AB ) if B is symmetric, and <(tr(AB )) = tr(AB ^ ) = tr(A ^B ^) is Hermitian. Thus for any n 2 n matrices A and B; tr(AB ~ )) = tr(A ^ = A ~B ~ when A is a real ~ ). Obviously, A and <(tr(AB matrix. We remark that (1) is immediate from (4) which is the same as our later (15) and that (7) [or our (20)] implies (3) which is stronger than ~)  kAk2 (see, for instance, [13, p. 236]). (2) because max (A In view that the trace tr(AB ) is the sum of all eigenvalues of the matrix product AB , we will extend the sum to partial sum and arrive at a family of eigenvalue inequalities which yield the aforementioned inequalities. We explain why our results may be regarded as the most general ones in certain sense. At end we derive a variety of inequalities from our theorem and present an application.

0018-9286/$20.00 2006 IEEE

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 51, NO. 9, SEPTEMBER 2006

1507

II. MAIN RESULTS Throughout this note, In is the n 2 n identity matrix and the eigenvalues (i ), if all real, and the singular values (i ) are always arranged in decreasing order. Moreover we assume that all the matrices are complex and of size n 2 n unless otherwise stated. Our results are of course valid for real matrices. We proceed with two basic known results in matrix theory as our lemmas. These results can be found in [10, pp. 511 and 513], respectively. Lemma 1: If H is an n 2 n Hermitian matrix, then for every k; 1 

To show the second inequality in (8), use Lemma 2 to (1)=(2) tr(UP HP (1)=(2) U 3 ) to get
tr

UP HP U 3

k i=1

i (H )i (UP U 3 ):

n

k i=1

i (H ) = UU max =I

tr(UHU )

where the extrema are over all k 2 n complex matrices U satisfying UU 3 = Ik . Lemma 2: If H is an n 2 n Hermitian matrix and V is a k 2 n matrix, k  n, then

The desired inequality then follows immediately. This completes the proof. Note that the proof in [5] by substituting H with a positive semidenite H + I (for sufciently large ) does not work here, since a partial sum is no longer a trace in general. To compare with existing results, it would be appealing if the second inequality in (8) held without the appearance of U , as is well known [13, p. 232] (see also [12] for more general inequalities), if H and P are both positive semidenite (see [10, p. 242] for an analog for H + P ), then for k = 1; 2; . . . ; n

k i=1

i (HP ) 

k i=1

 i (H ) i (P ):

(9)

k i=1

i (H )i (V V 3 )  tr(V HV 3 ) 

k i=1

n0i+1 (H )i (V V 3 ):

However, this is not true when H is Hermitian in general. A counterexample comes handy: Set H = 0I3 , take P to be the diagonal matrix diag(3; 2; 1) and k = 1. Moreover, recall [11] that when H and P are both positive semidenite,

Now, we present our main result. Theorem 3: Let H be an n 2 n Hermitian matrix and P be an n 2 n positivesemidenite matrix. Then, for every integer k; 1  k  n,
max 111<t n i=1 n0i+1 (H )t (P ) t < k  i (HP ) i=1 k i (H )i (UP U 3 ):  UU max =I i=1

k i=1

i (HP ) 

k i=1

n0t +1 (H )t (P ):

(10)

(8)

With this in mind we would hope that the summand in the rst term in (8) could more generally be n0t +1 (H )t (P ) in place of n0i+1 (H )t (P ). This is not true in general. Take P = diag(3; 2; 1) and H = diag(01; 02; 03) with k = 2 and t1 = 2; t2 = 3. Then, the right-hand side of (10) is 05, while the left-hand side is 06. So (9) and (10) no longer hold if one matrix is Hermitian, but not positive semidenite. On the other hand, it is also known ([5] or [10, p. 262]) that when H and P are both Hermitian

Proof: First notice (see, for instance, [13, p. 51]) that HP (= HP (1)=(2) P (1)=(2) ) and P (1)=(2) HP (1)=(2) have the same eigenvalues, where P (1)=(2) is the unique positive semidenite square root matrix of P . By Lemma 1, for each k , where 1  k  n, we have

k i=1

i (HP ) =
=

k i=1

i P HP
tr

max UU =I

UP HP U 3 :

This obviously follows from our theorem, since i (UP U 3 ) = i (P ) when k = n. Note that the upper index n in (11) cannot be replaced by k in general. In light of inequalities (9)(11) and in the sense of going from nonnegative deniteness to Hermity, that is, one matrix is Hermitian and the other one is positive semidenite, our theorem may be regarded as the most general one. III. COROLLARIES We now present variances of (8) and derive more inequalities from it, some of which may have existed possibly in different forms. Replacing H with 0H in (8) and noting that i (0H ) = 0n0i+1 (H ), we have Corollary 4: Let H be Hermitian and P be positive semidenite. Then

i=1

i (HP ) 

i=1

 i (H ) i (P ):

(11)

Let P (1)=(2) = R3 DR, where R is a unitary matrix and D is a di(1)=(2) (1)=(2) (1)=(2) (P );  t (P ); . . . ;  t (P ) and the agonal matrix with t square roots of the rest eigenvalues of P on the main diagonal. Set U0 = (Ik ; 0)R. Then
max UU =I

UP HP U 3  tr U0 P HP U03 3 3 = tr[(Ik ; 0)DRHR D (Ik ; 0) ] k  n0i+1 (RHR3 )t (P ) i=1 k n0i+1 (H )t (P ): = i=1
tr

k i=1

 i (H ) t (P )
k i=1

n0i+1 (HP )
k i=1
=

 UU min I

n0i+1 (H )i (UP U 3 ):

(12)

1508

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 51, NO. 9, SEPTEMBER 2006

Setting ti = i in (8) and (12), we have Corollary 5: Let H be Hermitian and Then
k i=1

In particular, for arbitrary matrix C be positive semidenite.

n i=1

~ ) i (C ): i (C

n0i+1 (H )i (P )
k i=1

(18) Corollary 9: Let A and B be complex matrices and B = UDV 3 be a singular value decomposition of B , where U and V are unitary, and D is diagonal. Denote S = UV 3 . Then (13)
n i=1

i (HP )
k i=1
=

 UU max I
and
k i=1

i (H )i (UP U 3 )

n0i+1 (AS )i (B )  <(tr(AB ))

 i (H ) i (P )
k i=1


n i=1

n i=1

i (AS )i (B ):

(19)

n0i+1 (HP )
k i=1
=

Consequently, for real A and B (see [8]) (14)

 UU min I
n i=1

n0i+1 (H )i (UP U 3 ):

min (AS )

i (B )  tr(AB )  max (AS )

n i=1

 i (B ):
(20)

Corollary 6: Let H and P be both Hermitian. Then

n0i+1 (H )i (P )  tr(HP ) 

n i=1

Proof: We make use the idea in the proof of [8]. Notice that

 i (H ) i (P ):
(15)

Proof: First, note that U is unitary when k = n, and thus = i (P ). If H is Hermitian and P is positive semidenite, then (15) follows from (13) or (14). If P is Hermitian, we choose  so large that P + I is positive semidenite. A simple computation for H and P + I will yield (15). The well known HardyLittlewoodPlya rearrangement theorem (see, for instance, [10, p. 141]) follows from (15) at once by taking H and P to be n 2 n real diagonal matrices. Corollary 7: Let x1  x2  1 1 1  xn be a set of real numbers and  be a permutation of f1; 2; . . . ; ng such that jx(1) j  jx(2) j  1 1 1  jx(n)j. Let p be a permutation of f1; 2; . . . ; ng. Then, for k =

<(tr(AB)) = <(tr(AUDV 3 )) 3 3 = <(tr(AUV V DV )) 3 = <(tr(ASV DV ))

i (UP U 3 )

Letting H = AS and P = V DV 3 in (15), we obtain (19) and thus (20). Corollary 10: Let l denote the number of eigenvalues of the n 2 n Hermitian matrix H (including multiplicities) that lie in the open left half-plane. (If l = n then H is a stable matrix.) Let P be an n 2 n positive semidenite matrix. Then for every integer k , if 1  k  l, then

1; 2; . . . ; n

k i=1

xn0i+1 jx(i) j  max p H

k i=1

xp(i) jxp(i) j:
and

t <111<t n i=1 n0i+1 (H )t (P )

(16)

Proof:
diag(
1

of the theorem immediately. We remark that Corollary 4 does not seem to follow from the Hardy LittlewoodPlya rearrangement theorem in which the upper limit for summation needs to be n. (Note: (2a) of A.3.a in [10, p. 141] is false. Take (03; 02) and (2; 1)). ~ ) for any matrix A and Hermitian matrix Since <(tr(AB )) = tr(AB B , (15) yields (5). For any complex matrix X , denote jX j = (X 3 X )(1)=(2) . Observing that jX j is positive semidenite and its eigenvalues are the singular values of X , we have the following. Corollary 8: Let A be Hermitian and B be arbitrary. Then
n i=1

jx j; jx j; . . . ; jxn j). Then, (16) follows from the rst inequality

Let

diag(x1 ; x2 ; . . . ; xn )

i=1 k i=1

n0i+1 (H )n0k+i (P ) i (HP ):

Proof: Notice that if 1  k  l, all n0i+1 (H ) < 0 for i = 1; 2; . . . ; k . Thus, the maximum value is attained by taking the k smallest eigenvalues of P , that is, n0k+1 (P )  1 1 1  n01 (P )  n (P ); i.e., n0k+i (P ) for i = 1; 2; . . . ; k . Corollary 11: Let H be Hermitian and P be positive semidenite. Then
max 1 (H )t (P )

3   (HP )  uu max  (H ) (uP u )

and
uu
min
=1

n0i+1 (A)i (B )  tr(AjB j) 

n i=1

 i (A ) i (B ):
(17)

n (H )1 (uPu3 )  n (HP )  min  1 (H ) t (P ): t

Proof: Putting k = 1 in Theorem 3 and Corollary 4 yields the previous inequalities, respectively.

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 51, NO. 9, SEPTEMBER 2006

1509

As an immediate application to the Lyapunov equation (LE), if A is stable and Hermitian and X is positive semidenite, then we have from Corollary 11
 n (A ) n (X )

 1 (AX )  1 (A)n(X )  n (AX )  1(A)1 (X ):

[12] B.-Y. Wang and F. Zhang, Trace and eigenvalue inequalities for ordinary and Hadamard products of positive semidenite Hermitian matrices, SIAM J. Matrix Anal. Appl., vol. 16, pp. 11731183, Oct. 1995. [13] F. Zhang, Matrix Theory: Basic Results and Techniques. New York: Springer-Verlag, 1999. [14] B. Cain, R. Horn, and L. Li, Inequalities for monotonic arrangements of eigenvalues, Linear Alg. Appl., vol. 222, pp. 18, 1995.

and
 n (A ) 1 (X )

We remark that some singular value inequalities may be obtained for an arbitrary (rectangle) matrix A by applying our theorem and corollaries to the partitioned Hermitian matrix M = 0 A3 ( ) and the partitioned positive semidenite matrix A 0 1 (A)In A3 N = ( ). Note that the eigenvalues of M A  1 (A )I n are 1 (A); . . . ; n (A); 0n (A); . . . ; 01 (A), where 1 (A)  1 1 1  n (A) are the singular values of the matrix A. IV. CONCLUSION We have seen that our main result (Theorem 3) presents a set of general inequalities concerning the product of a positive semidenite matrix and a Hermitian matrix. These inequalities give derivation of the inequalities in Section I, including the upper and lower bounds on the trace, by which the solutions to the matrix Riccati and Lyapunov equations have been estimated. ACKNOWLEDGMENT The rst author would like to thank R. Horn for communication of [14] after completion of this note. Some related results are discussed and obtained in [14]. Both authors would like to thank the referees for carefully reading the manuscript.

Exact Stability Analysis of 2-D Systems Using LMIs

AbstractIn this note, we propose necessary and sufcient conditions for the asymptotic stability analysis of two-dimensional (2-D) systems in terms linear matrix inequalities (LMIs). By introducing a guardian map for the set of Schur stable complex matrices, we rst reduce the stability analysis problems into nonsingularity analysis problems of parameter-dependent complex matrices. Then, by means of the discrete-time positive real lemma and the generalized -procedure, we derive LMI-based conditions that enable us to analyze the asymptotic stability in an exact (i.e., nonconservative) fashion. It turns out that, by employing the generalized -procedure, we can derive smaller size of LMIs so that the computational burden can be reduced. Index TermsLinear matrix inequalities (LMIs), stability analysis, twodimensional (2-D) systems.

I. INTRODUCTION In this note, we address asymptotic stability analysis problems of two-dimensional (2-D) systems described by the FornasiniMarchesini second model [6]
x(i + 1; j + 1) = A1 x(i; j + 1) + A2 x(i + 1; j ) A1 ; A2

REFERENCES
[1] S.-D. Wang, T.-S. Kuo, and C.-F. Hsu, Trace bounds on the solution of the algebraic matrix Riccati and Lyapunov equation, IEEE Trans. Autom. Control, vol. AC-31, no. 7, pp. 654656, Jul. 1986. [2] J. M. Saniuk and I. B. Rhodes, A matrix inequality associated with bounds on solutions of algebraic Riccati and Lyaponov equations, IEEE Trans. Autom. Control, vol. AC-32, no. 8, pp. 739740, Aug. 1987. [3] T. Mori, Comments on A matrix inequality associated with bounds on solutions of algebraic Riccati and Lyaponov equations, IEEE Trans. Autom. Control, vol. 33, no. 11, p. 1088, Nov. 1988. [4] Y. Fang, K. Loparo, and X. Feng, Inequalities for the trace of matrix product, IEEE Trans. Autom. Control, vol. 39, no. 12, pp. 24892490, Dec. 1994. [5] J. B. Lasserre, A trace inequality for matrix product, IEEE Trans. Autom. Control, vol. 40, no. 8, pp. 15001501, Aug. 1995. [6] P. Park, On the trace bound of a matrix product, IEEE Trans. Autom. Control, vol. 41, no. 12, pp. 17991802, Dec. 1996. [7] J. B. Lasserre, Tight bounds for the trace of a matrix product, IEEE Trans. Autom. Control, vol. 42, no. 4, pp. 578581, Apr. 1997. [8] W. Xing, Q. Zhang, and Q. Wang, A trace bound for a general square matrix product, IEEE Trans. Autom. Control, vol. 45, no. 8, pp. 15631565, Aug. 2000. [9] V. R. Karanam, Eigenvalue bounds for algebraic Riccati and Lyapunov equation, IEEE Trans. Autom. Control, vol. AC-27, no. 4, pp. 461463, Apr. 1982. [10] A. W. Marshall and I. Olkin, Inequalities: Theory of Majorization and Its Applications. New York: Academic, 1979. [11] B.-Y. Wang and F. Zhang, Some inequalities for the eigenvalues of the product of positive semidenite Hermitian matrices, Linear Alg. Appl., vol. 160, pp. 113118, 1992.

2 Cn2n :

(1)

Precise denition of the asymptotic stability of the 2-D system was rst made in [5]. Since then, various types of necessary and sufcient conditions have been proposed for the analysis of the asymptotic stability. We summarize some of them in the following proposition. Proposition 1: [6] The following conditions are equivalent. i) The 2-D system (1) is asymptotically stable. ii) det(In 0 z1 A1 0 z2 A2 ) 6= 0 for all (z1 ; z2 ) 2  2  where  denotes the closure of the open unit disc on the complex plane. iii) (A()) < 1 for all  2 [0; 2] where A() := A1 + ej A2 and ( 1 ) denotes the spectral radius. Unfortunately, the conditions in ii) and iii) are not numerically tractable since they should be checked at innitely many points over

D D

Manuscript received April 3, 2004; revised March 22, 2005 and March 23, 2006. Recommended by Associate Editor D. Nesic. This work was supported in part by the Ministry of Education, Culture, Sports, Science, and Technology of Japan under Grant-in-Aid for Young Scientists (B) 15760314. Y. Ebihara and T. Hagiwara are with the Department of Electrical Engineering, Kyoto University, Kyoto 615-8510, Japan (e-mail: [email protected]). Y. Ito is with the Graduate School of Engineering, Osaka University, Osaka 565-0871, Japan. Digital Object Identier 10.1109/TAC.2006.880789

0018-9286/$20.00 2006 IEEE