Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Four-vector wikipedia , lookup
Singular-value decomposition wikipedia , lookup
Matrix (mathematics) wikipedia , lookup
Non-negative matrix factorization wikipedia , lookup
Matrix calculus wikipedia , lookup
Eigenvalues and eigenvectors wikipedia , lookup
Orthogonal matrix wikipedia , lookup
Gaussian elimination wikipedia , lookup
Matrix multiplication wikipedia , lookup
Perron–Frobenius theorem wikipedia , lookup
THE INVERSE MEAN PROBLEM OF GEOMETRIC AND CONTRAHARMONIC MEANS YONGDO LIM Abstract. In this paper we solve the inverse mean problem of contraharmonic and geometric means of positive definite matrices (proposed in [2]) A = X + Y − 2(X −1 + Y −1 )−1 B = X#Y by proving its equivalence to the well-known nonlinear matrix equation X = T − BX −1 B where T = 21 (A+A#(A+8BA−1 B)) is the unique positive definite solution of X = A + 2BX −1 B. The inverse mean problem is solvable if and only if B ≤ A. 1. Introduction In [2] Anderson and coauthors have introduced a natural matrix generalization of the contraharmonic mean of scalars (a2 + b2 )/(a + b) and studied related fixed point and least squares problems on positive semidefinite matrices. The contraharmonic mean C(A, B) of positive definite matrices A and B is defined by C(A, B) = A + B − 2(A−1 + B −1 )−1 . Inverse mean problems involving the contraharmonic mean are considered and answered for the problem of contraharmonic and arithmetric means in [2] where the inverse mean problem of contraharmonic and geometric means is proposed even for the scalar version. The geometric mean A#B of positive definite matrices A and B is defined by A#B = A1/2 (A−1/2 BA−1/2 )1/2 A1/2 and is well-known in theory of matrix inequalities (cf. [4], [5], [6], [15]) and appears in semidefinite programming (scaling point, [20]) and in geometry (geodesic middle, [16], [8], [18]). Although the inverse mean problem occurs on positive semidefinite matrices A and B, we restrict our attention to the positive definite case. Similar questions on geometric, arithmetic and harmonic means are answered in [3]. 1 2 YONGDO LIM The inverse mean problem of contraharmonic and geometric means is to find positive definite matrices X and Y for the system of nonlinear matrix equations: A B = C(X, Y ) = X#Y (1.1) where A and B are given positive definite matrices of same size. The main result of this paper is that the inverse mean problem (1.1) is equivalent to a system of the well-known nonlinear matrix equations (Theorem 2.3): X Y = A + 2BX −1 B = X − BY −1 B. (1.2) The matrix equations X = Q + A∗ X −1 A and X = Q − A∗ X −1 A with Q positive definite have been studied recently by several authors (see [1], [9], [10], [11], [12], [14], [17], [19], [21], [22]). For the application areas where the equations arise, see the references given in [1] and [12]. In section 3, we compute an explicit solution T of the first equation X = A + 2BX −1 B of (1.2) and then assert that the inverse mean problem is equivalent to the matrix equation X = T − BX −1 B. Solving directly the matrix equation X = T − BX −1 B we conclude that the inverse mean problem is solvable if and only if 2B ≤ T, equivalently B ≤ A (Theorem 3.3). Extreme solutions of the inverse mean problem are also discussed. 2. The inverse mean problem The following characterization and properties of the geometric mean are wellknown. See [4], [5], [13], [18], and [7] for multivariable geometric means. Lemma 2.1. (The Riccati Lemma) The geometric mean A#B of positive definite matrices A and B is the unique positive definite solution of the Riccati equation XA−1 X = B. INVERSE MEAN PROBLEM 3 Furthermore, the geometric mean operation has the following properties: (i) A#B = B#A, (ii) (A#B)−1 = A−1 #B −1 , (iii) M (A#B)M ∗ = (M AM ∗ )#(M BM ∗ ) for any nonsingular matrix M, (iv) A0 #B 0 ≤ A#B whenever A0 ≤ A and B 0 ≤ B where A ≤ B means that B − A is positive semidefinite. For given positive definite matrices A and B, the inverse mean problem (1.1) of contraharmonic and geometric means is said to be solvable if there exist positive definite matrices X and Y satisfying the equations of the problem, and in this case the pair (X, Y ) is called a solution of the problem. One immediately sees that solutions of the inverse mean problem do not depend on the order from the commutativity of the contraharmonic and geometric means. That is, if (X, Y ) is a solution of (1.1) then so is (Y, X). By the Riccati Lemma, the second equation of (1.1) is solved in terms of X: B = X#Y implies Y = BX −1 B. Thus the inverse mean problem (1.1) can be alternatively written as A Y The matrix (X −1 +Y −1 −1 ) = C(X, Y ) = BX −1 B. (2.3) in the first equation of (1.1) or in the contraharmonic mean C(X, Y ) is known as the parallel addition of X and Y , and has an alternative expression useful for optimal control theory and algebraic Riccati equations (cf. [2]): (X −1 + Y −1 )−1 = X − X(X + Y )−1 X. But, the following expression looks new and will be crucial for our approach. Lemma 2.2. For positive definite matrices X and Y, (X −1 + Y −1 )−1 = (X#Y )(X + Y )−1 (X#Y ). Proof. The Riccati Lemma implies that X −1 = (Y −1 #X −1 )Y (Y −1 #X −1 ) Y −1 = (X −1 #Y −1 )X(X −1 #Y −1 ). 4 YONGDO LIM It then follows from the commutative and inversion properties of the geometric mean (Lemma 2.1) X −1 = (X#Y )−1 Y (X#Y )−1 Y −1 = (X#Y )−1 X(X#Y )−1 and therefore (X −1 + Y −1 ) = (X#Y )−1 (X + Y )(X#Y )−1 . Taking the inverses of the both sides leads (X −1 + Y −1 )−1 = (X#Y )(X + Y )−1 (X#Y ). ¤ Theorem 2.3. The inverse mean problem (1.1) is equivalent to the system of nonlinear matrix equations X Y = A + 2BX −1 B = X − BY −1 B. (2.4) More precisely, if we denote S and T by the solution sets of the system (1.1) and (2.4), respectively, then the maps Φ : S → T , Φ(X, Y ) = (X + Y, X) Ψ : T → S, Ψ(X, Y ) = (Y, X − Y ) are bijective and S −1 = T . Proof. Suppose that (X, Y ) ∈ S. That is, X and Y are positive definite matrices so that A = X + Y − 2(X −1 + Y −1 )−1 , B = X#Y. Setting S := X and T := X + Y, we have (2.3) S = X = T − Y = T − BX −1 B = T − BS −1 B. INVERSE MEAN PROBLEM 5 The equation A = C(X, Y ) = X + Y − 2(X −1 + Y −1 )−1 of (1.1) leads to T X + Y = A + 2(X −1 + Y −1 )−1 = Lemma(2.2) = A + 2(X#Y )(X + Y )−1 (X#Y ) = A + 2BT −1 B. Therefore, (T, S) = (X + Y, X) is a solution of (2.4). Conversely, suppose that (X, Y ) is a solution of (2.4). That is, X and Y are positive definite matrices so that X = A + 2BX −1 B Y and = X − BY −1 B. (2.5) (2.6) Put T := Y and S := X − Y. Then X = S + Y = S + T. (2.7) It then follows from (2.6) that S = X − Y = BY −1 B > 0 and then by the Riccati Lemma B = S#Y = S#T. (2.8) The equation (2.5) X = A + 2BX −1 B leads to A = (2.7),(2.8) = Lemma(2.2) X − 2BX −1 B (T + S) − 2(S#T )(T + S)−1 (S#T ) = (T + S) − 2(T −1 + S −1 )−1 = C(S, T ). Therefore, (T, S) = (Y, X − Y ) is a solution of the inverse mean problem (1.1). It is obvious to see that the maps Φ and Ψ are bijective and S −1 = T . ¤ Remark 2.4. If (X, Y ) is a solution of (2.4) then (Y, BY −1 B) is a solution of (1.1) and (Y, BY −1 B) = (Y, X − Y ). Indeed, Y = X − BY −1 B implies that X − Y = BY −1 B. 6 YONGDO LIM 3. Solvability In this section we give a complete answer for the inverse mean problem (1.1) by solving the equivalent system of matrix equations (2.4) in Theorem 2.3. First, we solve the first equation X = A + 2BX −1 B in (2.4). The uniqueness and existence of a positive definite solution of the matrix equation X = A + BX −1 B appear in (Proposition 4.1, [12]). However, we compute it explicitly in terms of geometric means. Proposition 3.1. Let A and B be positive definite matrices. Then the nonlinear matrix equation X = A + BX −1 B has the unique positive definite solution X = 12 (A + A#(A + 4BA−1 B)). Proof. By setting Y = A−1/2 XA−1/2 and C = A−1/2 BA−1/2 , we observe that the matrix equation X = A + BX −1 B is equivalent to Y = I + CY −1 C. (3.9) By Lemma 2.1 the equation (3.9) means that C = Y #(Y − I) = (Y 2 − Y )1/2 and hence Y 2 − Y − C 2 = 0. This equation has a unique positive definite solution defined as 1 1 Y = (I + (I + 4C 2 )1/2 ) = (I + I#(I + 4C 2 )). 2 2 Therefore again by Lemma 2.1 the equation (3.9) has a unique positive definite solution 1 X = (A + A#(A + 4BA−1 B)). 2 ¤ Proposition 3.2. The inverse mean problem (1.1) is equivalent to the nonlinear matrix equation X = T − BX −1 B where T = 21 (A+A#(A+8BA−1 B)). In particular, if X is a positive definite solution of X = T −BX −1 B then (X, BX −1 B) = (X, T −X) is a solution of the inverse mean problem (1.1). INVERSE MEAN PROBLEM 7 Proof. By Proposition 3.1, the nonlinear equation X = A + 2BX −1 B of the system (2.4) has the unique positive definite solution T = 21 (A + A#(A + 8BA−1 B)) and therefore (2.4) and hence (1.1) is equivalent to the single matrix equation X = T − BX −1 B. If X is a positive solution of X = T − BX −1 B then (T, X) is a solution of (2.4) and hence Ψ(T, X) = (X, T − X) = (X, BX −1 B) is a solution of the inverse mean problem (1.1) by Theorem 2.3 and Remark 2.4. ¤ The solvability criterion of the matrix equation X = T − BX −1 B (3.10) in Proposition 3.2 appeared in Anderson-Morley-Trapp (Theorem 4.2, [1]) and Engwerda [10]: (3.10) has a positive definite solution if and only if à ! T /2 B ≥ 0, B T /2 and it is well-known that this is equivalent to T /2 ≥ B. Indeed for à requirement ! A B positive definite A and B, ≥ 0 if and only if A ≥ BA−1 B (Lemma 3.3, B A [18]) if and only if B −1/2 AB −1/2 ≥ B 1/2 A−1 B 1/2 if and only if B −1/2 AB −1/2 ≥ I if and only if A ≥ B. We also have a direct proof of solvability of (3.10) as follows. Positive definite solvability of (3.10) is equivalent to solvability of the equation (T −X)#X = B with a side condition T > X > 0. With transforms Z = T −1/2 XT −1/2 and D = T −1/2 BT −1/2 this is equivalent to solvability of the equation, with a side condition I > Z > 0, (I − Z)#Z = (Z − Z 2 )1/2 = D or Z 2 − Z + D2 = 0. (3.11) It is immediate that the solvability in the last form is I − 2D ≥ 0. Therefore we can conclude that positive definite solvability of the original equation is T ≥ 2B. Theorem 3.3. The inverse mean problem (1.1) is solvable if and only if B ≤ A. In this case (X+ , Y+ ), defined as 8 YONGDO LIM 1 (T + T #(T − 4BT −1 B)) = 2 1 ≡ (T − T #(T − 4BT −1 B)) = 2 X+ ≡ Y+ 1 (T + (T + 2B)#(T − 2B)), 2 1 (T − (T + 2B)#(T − 2B)) 2 where 1 T = (A + A#(A + 8BA−1 B)), 2 is the extreme solution in the sense that X ≤ X+ and Y+ ≤ Y for all solutions (X, Y ) of (1.1). Proof. By Theorem 2.3 and Proposition 3.1 we have to consider the solvability of the matrix equation 1 T = (A + A#(A + 8BA−1 B)). 2 But this is the case if and only if 2B ≤ T , equivalently 4B ≤ A + A#(A + 8BA−1 B). Y = T − BY −1 B with Let C = B −1/2 AB −1/2 . Then by Lemma 2.1 the above inequality becomes 4I ≤ C + C#(C + 8C −1 ) = C + (C 2 + 8I)1/2 . (3.12) It is straightforward to see 4≤λ+ √ λ2 + 8 (λ > 0) ⇐⇒ λ ≥ 1. Hence the inequality (3.12) is equivalent to I ≤ C, hence to B ≤ A. Now the equation X = T − BX −1 B (3.13) is equivalent to the equation (3.11) Z 2 − Z + D2 = 0 (3.14) where Z = T −1/2 XT −1/2 and D = T −1/2 BT −1/2 . Since 2B ≤ T is equivalent to 4D2 ≤ I, the maximum and the minimum solution of (3.14) are respectively 1 (I + (I − 4D2 )1/2 ) = 2 1 = (I − (I − 4D2 )1/2 ) = 2 Z+ = Z− 1 (I + (I + 2D)#(I − 2D)) and 2 1 (I − (I + 2D)#(I − 2D)). 2 INVERSE MEAN PROBLEM 9 Therefore again by Lemma 2.1 the maximum and the minimum solutions of (3.13) are respectively 1 1 X+ ≡ (T + T #(T − 4BT −1 B)) = (T + (T + 2B)#(T − 2B)) and 2 2 1 1 X− ≡ (T − T #(T − 4BT −1 B)) = (T − (T + 2B)#(T − 2B)). 2 2 Since X− = T − X+ , the assertion on the extremity follows from Theorem 2.3. ¤ Remark 3.4. The simple condition A ≥ B is very natural in the sense that it corresponds to the fact that twice of the arithmetic mean is larger than the sum of the geometric and harmonic means: If (X, Y ) is a solution of (1.1) then A − B = 2A(X, Y ) − H(X, Y ) − (X#Y ) = (A(X, Y ) − X#Y ) + (A(X, Y ) − H(X, Y )) where A(X, Y ) and H(X, Y ) are the arithmetic and harmonic means of X and Y. The arithmetic-geometric-harmonic mean inequality A(X, Y ) ≥ X#Y ≥ H(X, Y ) implies that A ≥ B. Acknowledgements. The author is very grateful to the referees for many points to improve the paper, particularly on Theorem 3.3 and Remark 3.4. References [1] W. N. Anderson, T. D. Morley and G. E. Trapp, Positive solutions to X = A − BX −1 B ∗ , Linear Algebra Appl. 134 (1990), 53-62. [2] W. N. Anderson, M. E. Mays, T. D. Morley and G. E. Trapp, The contraharmonic mean of HSD matrices, SIAM J. Alg. Disc. Meth. 8 (1987), 674-682. [3] W. N. Anderson and G. E. Trapp, Inverse problems for means of matrices, SIAM J. Alg. Disc. Meth. 7 (1986), 188-192. [4] T. Ando, Topics on Operator Inequalities, Lecture Notes Hokkaido Univ., Sapporo, 1978. [5] —, Concavity of certain maps on positive definite matrices and applications to Hadamard products, Linear Alg. and Appl. 26 (1979) 203-241. [6] —, On the arithmetic-geometric-harmonic-mean inequalities for positive definite matrices, Linear Algebra and Its Applications 52/53 (1983), 31–37. 10 YONGDO LIM [7] T. Ando, C-H. Li, and R. Mathias, Geometric means, Linear Algebra Appl. 385 (2004), 305-334. [8] G. Corach, H. Porta, and L. Recht, Geodesics and operator means in the space of positive operators, International J. of Math. 4 (1993), 193-202. [9] S. M. El-sayed and A. C. M. Ran, On an iteration method for solving a class of nonlinear matrix equations, SIAM J. Matrix Ana. Appl. 23 (2002), 632-645. [10] J. C. Engwerda, On the existence of a positive definite solution of the matrix equation X + AT X −1 A = I, Linear Algebra Appl. 194 (1993), 91-108. [11] J. C. Engwerda, A. C. M. Ran, and A. L. Rijkeboer, Necessary and sufficient conditions for the existence of a positive definite solution of the matrix equation X + A∗ X −1 A = Q, Linear Algebra Appl. 186 (1993), 255-275. [12] A. Ferrante and B. Levy, Hermitian solutions of the equation X = Q + N X −1 N ∗ , Linear Algebra Appl. 247 (1996), 359-373. [13] M. Fiedler and V. Pták, A new positive definite geometric mean of two positive definite matrices, Linear Algebra and Its Applications 251 (1997), 1-20. [14] C.-H. Guo and P. Lancaster, Iterative solution of two matrix equations, Math. Comp. 68 (1999), 1589-1603. [15] F. Kubo and T. Ando, Means of positive linear operators, Math. Ann. 246 (1980), 205-224. [16] S. Lang, Fundamentals of differential geometry, Graduate Texts in Mathematics 191, SpringerVerlag, New York, 1999. [17] X. Liu and H. Gao, On the positive definite solutions of the matrix equations X s ± AT X −t A = In , Linear Algebra Appl. 368 (2003), 83-97. [18] J. D. Lawson and Y. Lim, The geometric mean, matrices, metrics, and more, Amer. Math. Monthly 108 (2001), 797-812. [19] B. Meini, Efficient computation of the extreme solutions of X + A∗ X −1 A = Q and X − A∗ X −1 A = Q, Math. Comp. 71 (2002), 1189-1204. [20] Yu. E. Nesterov and M.J. Todd, Self-scaled barriers and interior-point methods for convex programming, Mathematics of Operations Research 22 (1997), 1-42. [21] A. C. M. Ran and M. C. B. Reurings, On the nonlinear matrix equation X + A∗ F(X)A = Q: solutions and perturbation theory, Linear Algebra Appl. 346 (2002), 15-26. [22] X. Zhan and J. Xie, On the matrix equation X + AT X −1 A = I, Linear Algebra Appl. 247 (1996), 337-345. Department of Mathematics, Kyungpook National University, Taegu 702-701, Korea E-mail address: [email protected]