Improvements on Geometric Means Related to the Tracy-Singh Products of Positive Matrices

Transcription

1 MATEMATIKA, 2005, Volume 21, Number 2, pp c Department of Mathematics, UTM. Improvements on Geometric Means Related to the Tracy-Singh Products of Positive Matrices 1 Adem Kilicman & 2 Zeyad Al Zhour Department of Mathematics and Institute for Mathematical Research, University Putra Malaysia UPM, 43400, Serdang, Selangor, Malaysia 1 ailic@fsas.upm.edu.my / 2 zeyad1968@yahoo.com Abstract In this paper, a family of geometric means for positive matrices is studied; we discuss possible definitions of the geometric means of positive matrices, and some counter examples are given. It is still an open problem to find a completely satisfactory definition. Other problems are how to define the geometric, arithmetic, harmonic, power and operator means of finitely many positive matrices. We generalize these means of two positive matrices to arrive the definitions of the weighted means of positive matrices. We recover and develop the relationship between the Ando s geometric mean and the Kronecer product to the Tracy-Singh product and other means. Some new attractive inequalities for the Tracy-Singh product, Khatri-Rao product and geometric means of several positive matrices are established. The results lead to the case of Kronecer and Hadamard products of any finite numbers of matrices. Keywords Tracy-Singh product, Khatri-Rao product, Kronecer product, Hadamard product, Positive definite matrix, Geometric means, weighted geometric means. 1 Introduction and Preliminary Results Consider matrices A [a ij ] and C [c ij ] of order m n and B [b l ] of order p q. Let A and B be partitioned as A [A ij ] and B [B l ] 1 i t, 1 j c, where A ij is an t c t c m i n j matrix and B l is a p q l matrix m m i, n n j, p p i, q q j. Let A B, A B, AΘB and A B be the Kronecer, Hadamard, Tracy-Singh and Khatri- Rao products, respectively. The definitions of the mentioned four matrix products are given by Liu in [8, 9] as A B a ij B ij, A C a ij c ij ij C A, A B A ij B ij ij and AΘB A ij ΘB ij A ij B l l ij.additionally, Liu [8] shows that the Khatri-Rao product can be viewed as a generalized Hadamard product and the Tracy-Singh product as a generalized Kronecer product, i.e., for non-partitioned matrices A and B, their AΘB is A B and their A B is A B. The Khatri-Rao and Tracy-Singh products of matrices A i 1 i, 2 will be denoted by respectively. A i A 1 A 2... A and ΘA i A 1 ΘA 2 Θ...ΘA,

2 50 Adem Kilicman & Zeyad Al Zhour For Hermitian matrices A and B, the relation A > B means that A B > 0 is a positive definite and the relation A B means A B 0 is a positive semi-definite. Given a positive definite matrix A, its positive definite square root is denoted by. Notice that for positive definite matrices A and B, the relation A B implies B 1/2, A 2 B 2 and B 1 A 1. Let us introduce some notations. The notation M m,n M m,n + is the set of all m n positive definite matrices over M and when m n, we write M m M m + instead of M m,n M m,n +. The notations A T, A, A 1 are the transpose, conjugate transpose and inverse of matrix A, respectively. The Khatri - Rao and Tracy - Singh products are related by the following relation [5,14]: A i Z1 T ΘA i Z Here, A i M mi,ni 1 i, 2 are compatibly partitioned matrices, t c t m mi, n ni, r m j i, s n j i, mi m j i, ni t n j i. Z 1 and Z 2 are real matrices of order m r and n s, respectively such that Z T 1 Z I 1, Z T 2 Z I 2, where I 1 and I 2 are identity matrices of order r r and s s, respectively. In particular, if mi ni, we then have A i Z T ΘA i Z. 1-2 We shall mae frequent use the following properties of the Tracy-Singh and Khatri- Rao products and their proofs are straightforward by using induction on. Let A i and B i 1 i, 2 be compatible partitioned matrices, we have a ΘA i ΘB i ΘA i B i. 1-3 b ΘA i r c ΘA i ΘA i and A i A i. 1-4 ΘA r i if A i M m + i 1 i, 2 and r is any real number. 1-5 d A i ΘB i A i Θ B i. 1-6 In this paper, the results are established in three ways. First, we discuss possible definitions of the geometric means of two matrices and weighted means of positive matrices, provided some counter examples are given. It is still an open problem to find a completely

3 Improvements on Geometric Means Related to the Tracy-Singh Products 51 satisfactory definition. Second, we discover some interesting nenequalities involving Tracy-Singh product, Khatri-Rao product and geometric means of several positive matrices. Finally, the results lead to the case of Kronecer and Hadamard products of any finite numbers of positive matrices. 2 Geometric Means of Two Positive Matrices If A and B are arbitrary n n matrices, then the arithmetic mean is defined by as A B 1 A + B Similarly, when A and B are positive n n matrices, their harmonic mean can be defined A!B } A 1 + B However, it is not at all obvious how to define the geometric mean of positive matrices. In what follows, if A M n + and is any real number, then A will denote its unique positive semidefinite th power. Ideally, the geometric mean AB of two positive matrices and should satisfy the following properties: i AB AB 1/2 when A and B commute 2-3 ii AB 0 Positive property 2-4 iii AB BA Symmetry property iv A!B AB A B Arithmetic-Geometric-Harmonic inequality 2-5 The obvious candidates for AB are not satisfactory. We now explain the problems which arise with some of these candidates in the case that A and B are positive n n matrices. Candidate 1 : A 1 B AB 1/2 2-6 This has the drawbac need not a positive semi definite square root unless AB BA. To see this, observe that if A > 0, B > 0 and AB 0, then AB AB B A BA So, if AB BA, we must conclude that AB 0 and so it is impossible for any square root of AB to be positive semi definite. Candidate 2 : A 2 B B 1/2 2-7 Again, there are problems with the positivity property. Notice that if A > 0, B > 0 and B 1/2 B 1/2, then AB BA. Hence if AB BA, we must conclude that B 1/2 B 1/2, and so, as above, A 2 B B 1/2. Candidate 3 : A 3 B A 1/4 B 1/2 A 1/4 2-8

4 52 Adem Kilicman & Zeyad Al Zhour Certainly A 3 B satisfies property i and by design, also satisfies the positivity property ii. However, A 3 B need not be symmetric. To see this, consider [ ] [ ] A, B For these matrices, mathematica shows that: [ A B ], B 3 A [ and these are clearly different. In 1979, Ando [2] gave an apparently complicated variant of candidate3, which turns out to be very useful and satisfy all properties i-iv. When A and B are positive n n matrices, Ando s geometric mean is defined by AB D 1/2 with D B. 2-9 It is clear to show that Ando s geometric mean enjoy the following properties: aaa A. ba p A q A p+q/2, for all < p, q < caba 1 AB B dab 1 AB A e ABB 1/2 is a unitary matrix Now we use the fact that if X and Y are positive n n matrices, then if and only if X 2 Y 2. This follows from the observation that a positive n n matrix has a unique positive square root. Thus AB BA is equivalent to B 1/2 B 1/2 B 1/2 2 B 1/2 AB 1/ To see that this is indeed true, expand out the square B 1/2 B 1/2 B 1/2 2 ] B 1/2 B 1/2 B 1/2 B 1/2 B 1/2 B 1/2 B 1/2 B 1/2 B 1/2 B 1/2 B 1/2 B 1/2 B 1/2 B 1/2 B 1 B 1/2 B 1/2 B 1/2 B 1/2 B 1/2 AB 1/2. Ando s geometric mean also satisfies the arithmetic-geometric-harmonic mean inequality. The geometric-arithmetic means inequality can be established by taing inverses. Let A > 0 and B > 0 be n n matrices and write D B. It is clear that D > 0. Now AB A B B} 1/2 1 A + B 2 B} 1/2 1 2 A1/2 I + B B} 1/2 1 2 I + A 1/2 B D 1/2 1 I + D

5 Improvements on Geometric Means Related to the Tracy-Singh Products 53 This last inequality is the arithmetic-geometric mean inequality for the positive definite matrix. Since A > 0, there are a unitary matrix U and a diagonal matrix S diag[s 1,..., s n ] with s i > 0 1 i n satisfying A U SU. Thus D 1/2 1 2 I + D is equivalent to s 1/2 i s i, for all 1 i n and this is obviously true. Ando [2] established further nice properties of his geometric mean of two positive n n matrices vc ABC C ACC BC, for all C M n. Distributive property 2-15 via 1 B 1 A 2 B 2 A 1 A 2 B 1 B 2. Mixed property 2-16 viiab 1 A 1 B 1. Inverse property 2-17 We examine these in the next section, where we will give some extensions and generalizations. Now, we have A BA!B AB 2-18 To see that this is indeed true: A BA!B I DI!D} 1/2I + D 1/2 1/2I + D 1 1/2 } D 1/2 AB. Ando used his definition to study monotone functions of matrices and obtained many nice results [2]. In spite of all this, Ando s definition is not completely satisfactory. It fails to satisfy one very desirable property: viii AB 2 is similar to AB. Eigenvalue property 2-19 [ ] [ ] To see why this is a problem, consider A and B. Then, computing with Mathematica, we find AB [ ] [ 4 2, and AB 1 1 Here, the eigenvalues of AB are and , while the eigenvalues of AB 2 are and Hence AB 2 is not similar to AB, since similar matrices must have the same eigenvalues. In 1997, Fiedler and Pta [7] succeeded in finding a definition of geometric mean that does satisfy the eigenvalue property. They used A B A 1 B 1/2 AA 1 B 1/2, 2-20 for positive matrices A and B. Unfortunately, this definition fails the arithmetic-geometric -harmonic inequality. It satisfy properties from i to iii and from v to viii. We give a new direct proof of the eigenvalue property, assuming that A B B A, for positive matrices A and B. Observe that A B A 1 B 1/2 AA 1 B 1/2 B 1 BB ]

6 54 Adem Kilicman & Zeyad Al Zhour Thus, by property iii of Ando s geometric mean, we have A B 2 A BB A A 1 B 1/2 AA 1 B 1/2 B 1 BB 1 A 1 B 1/2 A B 1 1 B 1 BB 1 A 1 B 1/2 ABB 1 B 1 1 ABB 1 Q 1 ABQ Q B 1. Hence, A B 2 is not similar to AB. In 1998, Ando and Hiai [4] succeeded in generalizing Ando s geometric mean to the power mean that satisfy properties from i to vii. They used A B D with D B, 2-22 for any real number and positive matrices A and B. In particular if 1 2, we get Ando s geometric mean, i.e., A B AB. 1/2 It is clear to show that power mean enjoy the following new properties: aa A A ba p A q A 1 p+q, for all < p, q < In 2000, Micic, Pecaric and Seo [10] succeeded in generalizing the power mean to the operator mean. They used AσB fd with D B, 2-24 for any non-negative operator monotone function ft on [0, and positive matrices A and B. As a matter of fact, the power means are determined by the operator monotone function ft t when 0 < 1 or by the operator monotone function ft t 1/ when 1 <. It remains an open question whether there is a definition of the geometric mean of two positive n n matrices that satisfies all of properties from i - viii. We extend the mixed property to include the Tracy-Singh product as in next result. Lemma 2.1 Let A i and B i M + n i 1, 2 be compatible partitioned matrices. Then a A 1 ΘB 1 A 2 ΘB 2 A 1 A 2 ΘB 1 B b A 1 B 1 ΘA 2 B 2 A 1 ΘA 2 B 1 ΘB

7 Improvements on Geometric Means Related to the Tracy-Singh Products 55 Proof a By using 1-4 and 1-5, we have A 1 ΘB 1 A 2 ΘB 2 A 1 ΘB 1 1/2 A 1 ΘB 1 1/2 A 2 ΘB 2 A 1 ΘB 1 1/2} 1/2 A1 ΘB 1 1/ } 1/2 1 ΘB 1/2 1 A 2 ΘB 2 1 ΘB 1/2 1/2 1 A } 1 A 2 1 ΘB 1/2 1 B 2 B 1/ A 2 1 1/2 ΘB 1/2 1 A 2 1 1/2} 1 A 1 A 2 ΘB 1 B 2. We can prove b in a similar manner B 2 B 1/2 1 1/2} B 1/2 1 B 2 B 1/2 1 1/2} B 1/2 1 We extend also Lemma 2.1 to include the Fiedler and Pta definition and power mean definition related to the Tracy -Singh product as in next result. Lemma 2.2 For any real number and positive compatible partitioned matrices A i and B i i 1, 2, we have aa 1 ΘB 1 A 2 ΘB 2 A 1 A 2 ΘB 1 B ba 1 ΘB 1 A 2 ΘB 2 A 1 B 1 ΘA 2 B ca 1 B 1 ΘA 2 B 2 A 1 ΘA 2 B 1 ΘB da 1 B 1 ΘA 2 B 2 A 1 ΘA 2 B 1 ΘB Proof a To see that this is indeed true, let Then A 1 ΘB 1 A 2 ΘB 2 D 1 1 A 2 1 and D 2 B 1/2 1 B 2 B 1/2 1. A 1 ΘB 1 1/2 A 1 ΘB 1 1/2 A 2 ΘB 2 A 1 ΘB 1 1/2 A1 ΘB 1 1/2 1 1 ΘB 1/2 1 A 2 ΘB 2 1 ΘB 1/2 1/2 A 1 1 A 2 1 ΘB 1/2 1 B 2 B 1/ } 1 1 D 1 ΘD D 1 1 D 1 B 1/2 A 1 A 2 ΘB 1 B 2. 1

8 56 Adem Kilicman & Zeyad Al Zhour ba 1 ΘB 1 A 2 ΘB 2 A 1 ΘB 1 1 A 2 ΘB 2 } 1/2 A1 ΘB 1 A 1 ΘB 1 1 A 2 ΘB 2 } 1/2 A 1 1 ΘB 1 1 A 2ΘB 2 } 1/2 A1 ΘB 1 A 1 1 ΘB 1 1 A 2ΘB 2 } 1/2 A 1 1 A 2ΘB1 1 B 2 } 1/2 A1 ΘB 1 A 1 1 A 2 B1 1 B 2 } 1/2 A 1 1 A 2 1/2 ΘB1 1 B 2 1/2} A 1 ΘB 1 A 1 1 A 2 1/2 ΘB1 1 B 2 1/2} A 1 1 A 2 1/2 A 1 A 1 1 A 2 1/2} Θ A 1 B 1 ΘA 2 B 2 B 1 1 B 2 1/2 B 1 B 1 1 B 2 1/2} We can prove c and d in a similar manner. 3 Geometric Means of Several Positive Matrices Other problems are how to define the geometric, arithmetic, harmonic and operator means of finitely many positive matrices. We generalize these means of two positive matrices to arrive the definitions of the weighted means of positive matrices. Definition 3.1 Let w 1, w 2,..., w be positive numbers such that 1 and let A i M n + 1 i, 2. The weighted geometric, weighted arithmetic and weighted harmonic means of A i 1 i, 2 are defined by a Weighted geometric mean: A i w A 1 2 s+1 where u s 1 w s+1 / w j b Weighted arithmetic mean: u1 1/2 A 2 u2 } u 1 1/ A 1 A 1/2 1/2 A, 3-1 for s 1, 2,..., 1. A i w 1 A w A w c Weighted harmonic mean:! w A i w 1 A w 1 A 1 1 A i. 3-2 A 1 i

9 Improvements on Geometric Means Related to the Tracy-Singh Products 57 Definition 3.2 Let A i M n + 1 i, 2 and f j 1 j 1 be non-negative operator monotone functions on [0,. The weighted operator mean of A i 1 i is defined by σ A i w f f A 1/2 3 2 f 1 2 A }. 3-4 In fact, the weighted geometric means are determined by the operator monotone functions f 1 t t u 1, f 2 t t u 2,..., f 2 t t u2, f 1 t t u1 where u j 1 w j+1 / j+1 w r, 0 < u j 1, for j 1, 2,, 1 and w 1, w 2,, w r1 are positive numbers such that 1. Note that also the weighted geometric mean becomes Ando s geometric mean, weighted arithmetic mean becomes arithmetic mean and weighted harmonic mean becomes harmonic mean when 2 and w 1 w These general definitions have many good properties analogous. The fundamental relationship is no surprise. Theorem 3.3 Let w 1, w 2,..., w be positive numbers such that 1 and let A i M n 1 i, 2 be positive matrices. Then see [12].! w A i A i w A i 3-5 w All the inequalities are strict unless A 1 A 2... A. Surprisingly, it has not previously been observed that the geometric mean for more than two matrices fails the symmetry property. For example, if 3 and w 1 w 2 w 3 1 3, corresponding geometric means of three positive matrices A 1, A 2 and A 3 are A 3 A 2 A 1 1 and A 1 A 2 A 3 3 [ ] 3 1 But if A 1, A mathematica, we find A 3 A 2 A A 3 2 1/ [ [ A 1 2 1/2 ] [ 1 1 and A ] and A 1 A 2 A } 2/3 A 1/2 1, } 2/3 A 1/2 3. ]. Then, after computing with [ In other words, A 3 A 2 A 1 A 1 A 2 A 1. Even though property iii fails, properties vi and vii are still applicable. ].

10 58 Adem Kilicman & Zeyad Al Zhour 4 Further Developments and Applications In this section, we prove several theorems related to geometric means and Tracy-Singh products for compatible partitioned matrices. Theorem 4.1 Let A i and B i M n +, 1 i be compatible partitioned matrices. Then a A i ΘB i A i Θ B i b ΘA i B i ΘA i ΘB i. Proof a The proof is by induction on ; we now by 2-28 that A 1 ΘB 1 A 2 ΘB 2 A 1 B 1 ΘA 2 B 2. This gives the claimed when 2. Suppose that 1 holds. Then A i ΘB i 1 1 A i ΘB i A i A ΘB A Θ We can prove b in a similar manner. 1 A i ΘB i 1 A i Θ A i Θ B i B i A ΘB B i B A i Θ B i. In general case, for any real number and positive matrices A i and B i i 1, 2. The power mixed property related to the Tracy-Singh product can be extended to any finite number of positive matrices as in next results Theorem 4.2 and Theorem 4.3. Theorem 4.2 Let A i and B i M n + 1 i, 2 be compatible partitioned matrices. Then a A i ΘB i A i Θ B i. 4-2 b ΘA i B i ΘA i ΘB i. Proof Follows immediately by on induction on. When A i M + n 1 i, 2 and i are real scalars, we have defined Now continue recurrently, setting A 1 A 2 A A 2 2 A 1 A A A i A 1 A , 2,..., 1 1, 2,..., 1 1, 2,..., 1 1, 2,..., 1 A

11 Improvements on Geometric Means Related to the Tracy-Singh Products , 2,..., 2 A i }. 4-4 Theorem 4.3 Let A i and B i M n + 1 i and be compatible partitioned matrices and let i 1 i be real scalars. Then for any m 2, m m m A i ΘB i Θ , 2,..., m 1 A i 1, 2,..., m 1 B i 1, 2,..., m 1 Proof We use induction on m. In particular, when m 2 we obtain Theorem 4.2. Suppose that A i ΘB i A i Θ B i holds. 1, 2,..., 1 1, 2,..., 1 1, 2,..., 1 Then +1 A i ΘB i A i ΘB i A +1 ΘB +1 1, 2,..., 1, 2,..., 1, 2,..., } A +1 ΘB +1 1/2 A +1 ΘB +1 1/2 A i ΘB i A +1 ΘB +1 1/2 1, 2,..., 1 +1 ΘB1/ ΘB 1/ ΘB1/ A i 1, 2,..., 1, 2,..., 1 A i 1, 2,..., 1 Θ B 1/2 +1 A i 1, 2,..., 1 Θ Θ +1 B 1/2 +1 B i 1, 2,..., Θ +1 ΘB 1/ B i 1, 2,..., 1 +1 B 1/2 +1 } +1 B i 1, 2,..., 1 B 1/2 +1 B i 1, 2,..., 1 A +1 ΘB +1 1/2 } +1 ΘB1/ ΘB1/2 +1 B 1/2 +1 } B 1/ Theorem 4.3 can be extended by using induction as given in the following theorem.

12 60 Adem Kilicman & Zeyad Al Zhour Theorem 4.4 Let A ij M n + 1 i m, 1 j be compatible partitioned matrices and let i 1 i m be real scalars. Then ΘA 1j... ΘA mj 1, 2,..., m 1 1, 2,..., m 1 Θ A 1j... A mj , 2,..., m 1 1, 2,..., m 1 Property vii for two positive matrices namely A B 1 A 1 B 1, can be extended to power mean of two positive matrices: To see that this is indeed true: A 1 B 1 A A 1 B A 1 B A B 1 A A A A 1 2 BA 1 2 } 1 A A 1 2 A 1 2 BA } 1 1 A 2 A B 1. A quic induction is sufficient to extend this to means of more than two matrices. Lemma 4.5 Let A i M + n 1 i, 2 and let i 1 i be real scalars. Then A 1 i 1, 2,..., 1 A i 1, 2,..., We now turn to a generalization of the arithmetic-geometric-harmonic mean inequality related to Tracy-Singh Product. Theorem 4.6 Let A ij M n + 1 i m, 1 j be compatible partitioned matrices and let 1 i m be positive scalars such that m 1. Let u i 1 i 1, 2,..., m 1. Then a b ΘA 1 ij 1 1 i m ΘA 1 i jj i 1,,i m1 +1 / i+1 s1 w s Θ A 1j A mj ΘA ij. 4-9 w w 1 ΘA ij ΘA mj w w i 11 Equalities in 4-9 and 4-10 hold if and only if 1 m i m1 ΘA 1j... ΘA mj., ΘA ijj 4-10

13 Improvements on Geometric Means Related to the Tracy-Singh Products 61 Proof By Theorem 3.3 and Theorem 4.4, we have Θ A 1j... A mj ΘA 1j... w w w w Substituting A 1 ij for A ij and using Lemma 4.5, we obtain ΘA mj 1 Θ A 1j... A mj Θ A 1 1j... A 1 mj w w w w ΘA ij. ΘA 1 ij. Now, we have and hence Θ A 1j... A mj w w ΘA 1 ij 1 1 ΘA 1 ij, Θ A 1j... A mj. w w Similarly, we have the second inequalities b. Since ΘA 1j... ΘA mj Θ A 1j... A mj w w w w m Θ A ij m i 11 i m1 ΘA ijj. We shall use Theorem 4.6 to extend a result of Ando [2, pp. 229, Theorem 12] in the following theorem. Theorem 4.7 Let A i M + n 1 i m be commutative matrices. Then see [2] m m 1 m A i m A i We need some further ingredients before we address Ando s result. In 1997, Alic and others [1] gave a generalization of Theorem 3.3 in the negative weight case. The various means are defined just as before, even if there are negative weights. Theorem 4.8 Let A i 1 i be positive matrices and let 1 i be real numbers such that w 1 > 0, < 02 i and 1. Then see [1] A i w A i w

14 62 Adem Kilicman & Zeyad Al Zhour If A 1 i > 0, then A i w! w Equalities hold if and only if A 1... A. Developing Theorem 4.8, we have the following. A i Theorem 4.9 Let A ij M n + 1 i m, 1 j and let 1 i m be real numbers such that w 1 > 0, < 02 i m, and 1. Let u i 1 +1 / i+1 s1 w s, for i 1, 2,, m 1. Then a b... i 11 If w 1 ΘA ij Θ A 1j... A mj w w 1...m ΘA ijj i m1 ΘA 1j w m ΘA 1j ΘA mj > 0, then c Θ A 1j... A mj w w d ΘA ij... ΘA mj w w Equalities hold if and only if ΘA 1 ij... w w 1 ΘA 1j... ΘA mj. ΘA mj 1...i m ΘA 1 i jj i 1,...,i m Using Theorem 4.6 and the connection between the Khatri-Rao and Tracy-Singh products in 1-1, we have the following two theorems: Theorem 4.10 Let A ij M n + 1 i m, 1 j be compatible partitioned matrices and let 1 i m be positive real numbers such that m 1. Let u i 1 +1 / i+1 s1 w s, for i 1, 2,, m 1. Then a b A 1j... A mj w w A 1j... A mj w w... i 11 A ij m i m1 A ijj 4-19 Equalities hold if and only if A 1j... A mj. Proof It follows immediately by using Theorem 4.6 and Eq.1-1.

15 Improvements on Geometric Means Related to the Tracy-Singh Products 63 Finally, we arrive at our extension and generalization of Ando s result quoted in Theorem 4.7. Theorem 4.11 Let A i M n + 1 i m be compatible partitioned matrices. Then A 1... A m A 2... A m A 1... A m A 1... A m 1 w w w w w w w w m A i Proof In Theorem 4.10, let Then A 1... A m w w A 11,..., A m1 A 1,..., A m A 12,..., A m2 A 2,..., A m, A 1 A 1m, A 2m,..., A mm A m, A 1,..., A m 1. A 2... w w A m A 1... w A 1j... A mj w w A m A 1... A m 1 w w w m A ij A j m A i To see how Ando s result is a special case of Theorem 4.11, note that if A i 1 i m commute and non-partitioned matrices, then A 1... A m A 2... A m A 1 w w w w w A m A 1... A m w w w Theorem 4.12 Let A i M + n i 1, 2 be compatible partitioned matrices. Then for any real number and for all < p, q < A 1 ΘA 2 p A 1 ΘA 2 q A 1 ΘA 2 1 p+q 4-22

16 64 Adem Kilicman & Zeyad Al Zhour In particular if 1 2, we have A 1 ΘA 2 p A 1 ΘA 2 q A 1 ΘA 2 p+q/ Proof By using 1-5, 2-23 and Lemma 2.1, we have A 1 ΘA 2 p A 1 ΘA 2 q A p 1 ΘAp 2 A q 1 ΘAq 2 Ap 1 A q 1 ΘAp 2 A q 2 A 1 p+q 1 ΘA 1 p+q 2 A 1 ΘA 2 1+p+q. Theorem 4.13 Let A i M n + 1 i be compatible partitioned matrices. Then for any real number and for all < p, q < 4-24 ΘA p i In particular if 1 2, we have ΘA p i ΘA q i ΘA q i ΘA 1 p+q i ΘA p+q/2 i 4-25 Proof The proof is a consequence of Theorem This gives the claimed when 2. We can now proceed by induction on. Assume that the corollary holds for products of 1 matrices. Then ΘA p i ΘA q i [ 1 1 ΘA p i ΘA 1 p+q i 1 ΘA q i ] Θ[A p A q ] ΘA 1 p+q ΘA 1 p+q i Remar: The results obtained in Section 2, Section 3 and Section 4 are quite general. As a special case, consider the matrices in above sections are non-partitioned, we then have equalities and inequalities involving Kronecer and Hadamard products by replacing Θ by and by.. 5 Conclusion In the present wor we study a family of geometric means for positive matrices. It is still an open problem to find a completely satisfactory definition that satisfies all properties i-viii. To find a completely satisfactory definition is subject for the future study. In this study we have succeeded to find new connections between geometric means and Tracy- Singh products and establish several equalities and inequalities for geometric means and Tracy-Singh products of several positive matrices. Acnowledgment The authors gratefully acnowledge that this research has been partially supported by Universiti Putra Malaysia UPM under the Grant IRPA EA001.

17 Improvements on Geometric Means Related to the Tracy-Singh Products 65 References [1] M. Alic, B. Mond, J. Pecaric & V. Volence, The Arithmetic-Geometric-Harmonic Mean and Related Matrix Inequalities Lin. Alg. and its Appl , [2] T. Ando, Concavity of Certain Maps on Positive Definite Matrices and Applications to Hadamard Products, Lin. Alg. and its Appl , [3] T. Ando, On Arithmetic-Geometric-Harmonic Mean Inequalities for Positive Definite Matrices, Lin. Alg. and its Appl. 52/ , [4] T. Ando & F. Hiai, Holder Type Inequalities for Matrices, Math. Ineq. Appl ,1-30. [5] C.-G. Cao, X. Zhang & Z.-P. Yang Some Inequalities for the Khatri-Rao Products of Matrices, Elec. J. Lin. Alg. ISSN , , [6] B. Q. Feng, Matrix inequalities, Ph.D. Thesis, Kent State University [7] M. Fiedler & V. Pta, A New Positive Definite Geometric Mean of Two Positive definite Matrices, Lin. Alg. and its Appl., , [8] S. Liu, Matrix Results on the Khatri-Rao and Tracy-Singh Products, Lin. Alg. and its Appl , [9] S. Liu, Several Inequalities Involving Khatri-Rao Products of Positive Semi-definite Matrices, Lin. Alg. and its Appl , [10] J. Micic, J. Pecaric & Y. Seo, Complementary Inequalities to Inequalities of Jensen and Ando based on the Mond-Pecaric Method, Lin. Alg. and its Appl , [11] B. Mond, J. Pecaric & Zagreb, A mixed Arithmetic-Mean-Harmonic-Mean Matrix Inequality, Lin. Alg. and its Appl. 237/ , [12] M. Sagae & K. Tanabe, Upper and Lower Bounds for the Arithmetic-Geometric- Harmonic Means of Positive Definite Matrices, Lin. Multilin. Alg , [13] Zeyad Al Zhour & Adem Kilicman, Extension and Generalization Inequalities Involving the Hadamard Product of Several Matrices, ICREM 2, 2005, [14] Zeyad Al Zhour & Adem Kilicman, New Holder-Type Inequalities for the Tracy-Singh and Khatri-Rao Products of Positive Matrices, IRCMSA 2005, Proceeding ISBN: , 1-7.