See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/2113953 Generalized Induced Norms Article in Czechoslovak Mathematical Journal · August 2004 DOI: 10.1007/s10587-007-0049-5 · Source: arXiv CITATIONS READS 2 112 3 authors: Shirin Hejazian Madjid Mirzavaziri Ferdowsi University Of Mashhad Ferdowsi University Of Mashhad 41 PUBLICATIONS 200 CITATIONS 66 PUBLICATIONS 874 CITATIONS SEE PROFILE M. S. Moslehian Ferdowsi University Of Mashhad 220 PUBLICATIONS 3,025 CITATIONS SEE PROFILE All content following this page was uploaded by M. S. Moslehian on 15 August 2013. The user has requested enhancement of the downloaded file. SEE PROFILE arXiv:math/0407357v1 [math.FA] 21 Jul 2004 Generalized Induced Norms∗ S. Hejazian, M. Mirzavaziri and M. S. Moslehian Abstract Let k.k be a norm on the algebra Mn of all n × n matrices over C. An interesting problem in matrix theory is that ”are there two norms k.k1 and k.k2 on Cn such that kAk = max{kAxk2 : kxk1 = 1} for all A ∈ Mn . We will investigate this problem and its various aspects and will discuss under which conditions k.k1 = k.k2 . 1 Preliminaries Throughout the paper Mn denotes the complex algebra of all n × n matrices A = [aij ] with entries in C together with the usual matrix operations. Denote by {e1 , e2 , · · · en } the standard basis for Cn , where ei has 1 as its ith entry and 0 elsewhere. We denote by Eij the n × n matrix with 1 in the (i, j) entry and 0 elsewhere. For 1 ≤ p ≤ ∞ the norm ℓp on Cn is defined as follows: ℓp (x) = ℓp ( n X xi ei ) = i=1 ( n X i=1 |xi |p )1/p 1≤p<∞ max{|x |, · · · , |x |} 1 n p=∞ A norm k.k on Cn is said to be unitarily invariant if kxk = kUxk for all unitaries U and all x ∈Cn . ∗ 2000 Mathematics Subject Classification 15A60 (Primary) 47A30, 46B99 (Secondary). Keywords and phrases. induced norm, generalized induced norm, algebra norm, the full matrix algebra, unitarily invariant, generalized induced congruent. 1 By an algebra norm (or a matrix norm) we mean a norm k.k on Mn such that kABk ≤ kAkkBk for all A, B ∈ Mn . An algebra norm k.k on Mn is called unitarily invariant if kUAV k = kAk for all unitaries U and V and all A ∈ Mn . See [2, Chapter IV] for more information. Example 1.1 The norm kAkσ = n X |aij | is an algebra norm, but the norm kAkm = i,j=1 2 1 1 1 1 2 max{|ai,j | : 1 ≤ i, j ≤ n} is not an algebra norm, since k km > k km . 1 1 1 1 kABk : Remark 1.2 It is easy to show that for each norm k.k on Mn , the scaled norm max{ kAkkBk A, B 6= 0}k.k is an algebra norm; cf. [1, p.114] Let k.k1 and k.k2 be two norms on Cn . Then for each A : (Cn , k.k1 ) → (Cn , k.k2 ) we can define kAk = max{kAxk2 : kxk1 = 1}. If k.k1 = k.k2 , then kIk = 1 and there are many examples of k.k1 and k.k2 such that kIk = 6 1. This shows that given k.k on Mn , we cannot deduce in general that there is a norm k.k1 on Cn with kAk = max{kAxk1 : kxk1 = 1}. Let us recall the concept of g-ind norm as follows: Definition 1.3 Let k.k1 and k.k2 be two norms on Cn . Then the norm k.k1,2 on Mn defined by kAk1,2 = max{kAxk2 : kxk1 = 1} is called the generalized induced (or g-ind) norm via k.k1 and k.k2 . If k.k1 = k.k2 , then k.k1,1 is called induced norm. n X n X |ai,j | : 1 ≤ j ≤ n}, kAkR = max{ |ai,j | : 1 ≤ i ≤ n} and j=1 i=1 √ ∗ the spectral norm kAkS = max{ λ : λ is an eigenvalue ofA A} are induced by ℓ1 , ℓ∞ and Example 1.4 kAkC = max{ ℓ2 (or the Eucleadian norm), respectively. It is known that the algebra norm kAk = max{kAkC , kAkR } is not induced [ ] and it is not hard to show that it is not g-ind too; cf. [1, Corollary 3.2.6] We need the following proposition which is a special case of a finite dimensional version of the Hahn-Banach theorem [5, p. 104]: 2 Proposition 1.5 Let k.k be a norm on Cn and y ∈Cn be a given vector. There exists a vector y◦ ∈Cn such that y◦∗ y = kyk and for all x ∈Cn , |y◦∗x| ≤ kxk. (Throughout, ∗ denotes the transpose) [3, Corollary 5.5.15]) In this paper we examine the following nice problems: (i) Given a norm k.k on Mn is there any class A of Mn such that the restriction of the norm k.k on A is g-ind? (ii) When a g-ind norm is unitarily invariant? (iii) If a given norm k.k is g-ind via k.k1 and k.k2 , then is it possible to find k.k1 and k.k2 explicitly in terms of k.k? (iv) When two g-ind norms are the same? (v) Is there any characterization of the g-ind norms which are algebra norms? 2 Main Results We begin with some observations on generalized induced norms. Let k.k1,2 be a generalized induced norm on Mn obtained via k.k1 and k.k2 . Then kEij k1,2 = max{kEij xk2 : kxk1 = 1} = max{kxj ei k2 : k(x1 , · · · , xj , · · · , xn )k1 = 1} = αj kei k2 , where αj = max{|xj | : k(x1 , · · · , xj , · · · , xn )k1 = 1}. In general, for x ∈Cn and 1 ≤ j ≤ n, if Cx,j ∈ Mn is defined by the operator Cx,j (y) = yj x then kCx,j k1,2 = αj kxk2 . Also if for x ∈Cn we define Cx ∈ Mn by Cx = where α = max{| Pn Pn j=1 Cx,j , then clearly kCx k1,2 = αkxk2 , j=1 yj | : k(y1 , · · · , yj , · · · , yn )k1 = 1}. Now we give a partial solution to Problem (i) and useful direction toward solving Problem (iii): Proposition 2.1 Let k.k be an algebra norm on Mn . Then k.k is a g-ind norm on {A ∈ Mn : kAk = kA−1 k = 1}. Proof. Put kxk1 = max{kCAx k : kAk = 1}, λ−1 = max{| λkCx k. 3 n X i=1 xi | : kxk1 = 1} and kxk2 = Then we have kCy k1,2 = max{kCy xk2 : kxk1 = 1} = max{| kyk2λ−1 = kCy k. n X i=1 xi |kyk2 : kxk1 = 1} = It follows that for each y ∈Cn there is some x ∈Cn such that kCy xk2 = kCy kkxk1 = kCy k max{kCDx k : kDk = 1}. Now let A be invertible and kA−1 k = kAk = 1 and z = A−1 Cy x. Then λ−1 kBzk2 = λ−1 kBA−1 Cy xk2 = λ−1 kDxk2 = kCDx k ≤ kC1y k kCy xk2 = kC1y k kAzk2 . Now choose y so that kCy k = 1. Then kCBz k ≤ kCAz k for all B ∈ Mn . This implies that kCAz k is an upper bound for the set {kCBz k : kBk = 1} and indeed kCAz k = max{kCBz k : kBk = 1} = kzk1 . It follows that kAk = 1 = kCA( kzkz ) k = max{kCAu k : kuk1 = 1} = 1 max{kAuk2 : kuk1 = 1} = kAk1,2 .2 Let us now answer Question (ii). Proposition 2.2 An induced norm k.k1,2 is unitarily invariant if and only if so are k.k1 and k.k2 . Proof. Let U, V be unital operators and A be an arbitrary operator on Cn . Suppose that k.k1 and k.k2 are unitarily invariant. Then kUAV xk2 kAV xk2 kAyk2 kAyk2 = max = max −1 = max = kA1,2 . x6=0 x6=0 y6=0 kV y6=0 kyk1 kxk1 kxk1 xk1 kUAV k1,2 = max Conversely, if k.k1,2 is unitarily invariant, then kUxk1 = max{kAUxk2 : kAk1,2 ≤ 1} = max{kBxk2 : kU −1 Bk1,2 ≤ 1} = max{kBxk2 : kBk1,2 ≤ 1} = kxk1 and kUxk2 = α1 kCU x k = 1 kUCx k = α1 kUCx k = α1 kCx k = kxk2 .2 α Modifying the proof of Theorem 5.6.18 of [3], we obtain a similar useful result for g-ind norms: Theorem 2.3 Let k.k1 , k.k2 , k.k3 and k.k4 be four given norms on Cn and Ri,j = max{ kxki : x 6= 0}, 1 ≤ i, j ≤ 4. kxkj Then max{ kAk1,2 : A 6= 0} = R2,4 R3,1 kAk3,4 4 kAk1,1 2,2 : A 6= 0} = max{ kAk : A 6= 0} = R1,2 R2,1 . In particual, max{ kAk kAk1,1 2,2 2 kAxk4 kxk3 2 = kAxk . . . Hence kAk1,2 ≤ Proof. Let A be a matrix and x 6= 0. Then kAxk kxk1 kAxk4 kxk3 kxk1 kAk1,2 R2,4 kAk3,4 R3,1 . Thus kAk ≤ R2,4 R3,1 . 3,4 There are vectors y, z in Cn such that kyk2 = kzk2 = 1, kyk2 = R2,4 kyk4 and kzk3 = R3,1 kzk1 . By Proposition 1.15, there exists a vector z◦ ∈ Cn such that |z◦∗ x| ≤ kxk3 and z◦∗ z = kzk3 . ∗ kyk2 ◦ zk2 2 kzk3 ◦ zk2 Put A◦ = yz◦. Then kAkzk = kyz = kykkzk = kyk2R3,1 . Hence kA◦ k1,2 ≥ kyk R3,1 kyk4 = kzk1 1 1 4 ∗ ∗ ◦ xk4 4 |z◦ x| ◦ xk4 R2,4 .R3,1 kyk4. On the other hand, kAkxk = kyzkxk = kykkxk ≤ kyk4. Thus kA◦ k3,4 ≤ kyk4. 3 3 3 R3,1 kyk4 ◦ k1,2 Hence kA ≥ R2,4 kyk = R2,4 R3,1 .2 kA◦ k3,4 4 Corollary 2.4 (i) k.k1,2 ≤ k.k3,2 if and only if k.k1 ≥ k.k3 , (ii) k.k1,2 ≤ k.k1,4 if and only if k.k2 ≤ k.k4 . 1,2 : A 6= 0} = R2,2 R3,1 ≤ 1 and this if and Proof. (i) k.k1,2 ≤ k.k3,2 if and only if max{ kAk kAk3,2 only if R3,1 ≤ 1 or equivalently k.k3 ≤ k.k1 . The proof of (ii) is similar.2 The following corollary completely answers to Question (iv): Corollary 2.5 k.k1,2 = k.k3,4 if and only if there exists γ > 0 such that k.k1 = γk.k3 and k.k2 = γk.k4 . kAk1,2 3,4 : A 6= 0} = 1 = max{ kAk : A 6= Proof. If kAk1,2 = kAk3,4 , then R4,2 R1,3 = max{ kAk kAk1,2 3,4 kxk2 kxk1 1 0} = R2,4 R3,1 . Hence max{ kxk : x 6= 0} = R2,4 = R13,1 = min{ kxk : x 6= 0} ≤ max{ kxk : kxk3 4 3 kxk2 x 6= 0} = R1,3 = R14,2 = min{ kxk : x 6= 0}. Thus there exists a number γ such that 4 kxk1 kxk2 = γ = kxk .2 kxk4 3 Remark 2.6 It is known that each induced norm k.k is minimal in the sense that for any matrix norm k.k, the inequality k.k ≤ k.k1,1 implies that k.k = k.k1,1 . But this is not true for g-ind norms in general. For instance, put k.kα = ℓ∞ (.), k.kβ = 2ℓ2 (.) and k.kγ = ℓ2 (.). Then k.kγ,β ≤ k.kα,β but k.kγ,β 6= k.kα,β . 5 The following theorem is one of our main theorems and provide a complete solution for Problem (v): Theorem 2.7 Let k.k1 and k.k2 be two norms on Cn . Then k.k1,2 is an algebra norm on Mn if and only if k.k1 ≤ k.k2 . Proof. For each A and B in Mn we have kABxk2 ≤ kAk1,2 kBxk1 ≤ kAk1,2 kBxk2 ≤ kAk1,2 kBk1,2 kxk1 . Hence kABk1,2 ≤ kAk1,2 kBk1,2 . Conversely, let k.k1,2 be an algebra norm. Then for each A, B ∈ Mn we have kABk2 ≤ kAk1,2 kBk1,2 kxk1 . Let B be an arbitrary member of Mn . For Bx 6= 0, take M to be the 1 linear span of {Bx} and define f : (M, k.k1 ) →C by f (cBx) = ckBxk . By the Hahn-Banach kBxk2 Theorem, there is an F : (Cn , k.k1 ) →C with F |M = f and kF k = kf k = max{|f (cBx)| : 1 1 : |c|kBxk1 = 1} = kBxk . Define A : (Cn , k.k1 ) → (Cn , k.k2) by kcBxk1 = 1} = max{ |c|kBxk kBxk2 2 Ay = F (y)Bx. Then kAk1,2 = max{kAyk2 : kyk1 = 1} = max{|F (y)|kBxk2 : kyk1 = 1} = kBxk1 1, and kABxk2 = |F (Bx)|kBxk2 = |f (Bx)|kBxk2 = kBxk kBxk2 = kBxk1 . Thus for all B, 2 kBxk1 = kABxk2 ≤ kAk1,2 kBk1,2 kxk1 = kBk1,2 kxk1 , or kBxk1 ≤ kBk1,2 kxk1 . 1 Now take N to be the linear span of {x} and define g : (N, k.k1 ) →C by g(cx) = ckxk . By kxk2 the Hahn-Banach Theorem, there is a G : (Cn , k.k1 ) →C with G|N = g and kGk = kgk = 1 1 max{|g(cx)| : kcxk1 } = max{ |c|kxk : |c|kxk1 = 1} = kxk . Define B : (Cn , k.k1) → (Cn , k.k2 ) kxk2 2 by By = G(y)x. Then kBk1,2 = max{kByk2 : kyk1 = 1} = max{|G(y)|kxk2 : kyk1 = 1} = kxk2 kxk1 kxk2 kGk = 1, and kBxk1 = |G(x)|kxk1 = |g(x)|kxk1 = kxk kxk1 = kxk21 . 2 So Thus k.k1 ≤ k.k2 .2 kxk21 = kBxk1 ≤ kBk1,2 kxk1 = kxk1 . kxk2 6 Proposition 2.8 Suppose that k.k1,2 is a g-ind norm and λ > 0. Then the scaled norm λk.k1,2 is a g-ind algebra norm if and only if λ ≥ R1,2 . Proof. Evidently, λk.k1,2 = k.kk.k1 ,λk.k2 . If k.k3,4 = λk.k1,2 = k.kk.k1 ,λk.k2 then Corollary 2.5 implies that there exists α > 0 such that k.k3 = αk.k1 and k.k4 = αλk.k2 . Now Theorem 2.7 follows that λk.k1,2 = k.k3,4 is an algebra norm if and only if αk.k1 ≤ αλk.k2 or equivalently R1,2 ≤ λ.2 Proposition 2.9 Let k.k1 and k.k2 be two norms on Cn and 0 6= α, β ∈C. Define k.kα and k.kβ on Cn by kxkα = kαxk1 and kxkβ = kβxk2 , respectively. Then k.kα and k.kβ are two norms on Cn and k.kα,β = | αβ |k.k1,2. Proof. We have kAkα,β = max{kAxkβ : kxkα = 1} = max{kβAxk2 : kαxk1 = 1} = | αβ | max{kAyk2 : kyk1 = 1} = | αβ |kAk1,2 .2 The preceding proposition leads us ti give the following definition: Definition 2.10 Let (k.k1 , k.k2) and (k.k3 , k.k4 ) be two pairs of norms on Cn . We say that (k.k1 , k.k2 ) is generalized induced congruent (gi-congeruent) to (k.k3 , k.k4) and we write (k.k1 , k.k2 ) ≡gi (k.k3 , k.k4 ) if k.k1,2 = γk.k3,4 for some 0 < γ ∈R. Clearly ≡gi is an equivalence relation. We denote by [(k.k1 , k.k2 )]gi the equivalence class of (k.k1 , k.k2 ). Proposition 2.9 shows that for each 0 < α, β ∈R we have (αk.k1 , βk.k2) ≡gi (k.k1 , k.k2 ). Indeed, we have the following result: Theorem 2.11 For each pair (k.k1 , k.k2 ) of norms on Cn we have [(k.k1 , k.k2)]gi = {(αk.k1, βk.k2 ) : 0 < α, β ∈R}. We can extend the above method to find some other norms on Mn which are not necessarily gi-congruent to a given pair (k.k1 , k.k2 ): Proposition 2.12 Let (k.k1 , k.k2 ) be a pair of norms on Cn and K, L ∈Mn be two invertible matrices. Define kkK and kkL and Cn by kxkK = kKxk1 and kxkL = kLxk2 . Then kkK and kkL are norms on Cn and kAkK,L = kLAK −1 k1,2 . 7 Proof. Clear and see also Lemma 3.1 of [4].2 Remark 2.13 Note that the case K = αI and L = βI gives Proposition 2.9. References [1] Belitskiı̆, G. R.; Lyubich, Yu. I. Matrix norms and their applications. Translated from the Russian by A. Iacob. Operator Theory: Advances and Applications, 36. Birkhuser Verlag, Basel, 1988. [2] Bhatia, Rajendra. Matrix analysis. Graduate Texts in Mathematics, 169. SpringerVerlag, New York, 1997. [3] Horn, Roger A.; Johnson, Charles R. Matrix analysis. Cambridge University Press, Cambridge, 1985. [4] Li, Chi-Kwong; Tsing, Nam-Kiu; Zhang, Fuzhen. Norm hull of vectors and matrices. Linear Algebra Appl. 257 (1997), 1-27. [5] Rudin, Walter. Real and Complex Analysis, McGraw-Hill, Singapore, 1987. 8 View publication stats