Schr¨odinger uncertainty relation, Wigner-Yanase-Dyson skew information and metric adjusted correlation measure Shigeru Furuichi1∗and Kenjiro Yanagi2† 1Department of Computer Science and System Analysis, 2 College of Humanities and Sciences, Nihon University, 1 3-25-40,Sakurajyousui,Setagaya-ku,Tokyo, 156-8550,Japan 0 2 2Division of Applied Mathematical Science, Graduate School of Science and Engineering, Yamaguchi University, n a 2-16-1,Tokiwadai, Ube City, 755-0811,Japan J 5 1 Abstract. In this paper, we give Schro¨dinger-type uncertainty relation using the Wigner- ] Yanase-Dyson skew information. In addition, we give Schro¨dinger-type uncertainty relation by h p useof atwo-parameter extended correlation measure. We finallyshow thefurthergeneralization - ofSchro¨dinger-typeuncertaintyrelationbyuseofthemetricadjustedcorrelationmeasure. These t n results generalize our previous result in [Phys. Rev. A, Vol.82(2010), 034101]. a u q Keywords : Trace inequality, Wigner-Yanase-Dyson skew information, Schro¨dinger un- [ certainty relation and metric adjusted correlation measure 4 2000 Mathematics Subject Classification : 15A45, 47A63 and 94A17 v 2 9 3 1 Introduction 0 . 0 1 In quantum information theory, one of the most important results is the strong subadditivity 0 of von Neumann entropy [22]. This important property of von Neumann entropy can be proven 1 by the use of Lieb’s theorem [16] which gave a complete solution for the conjecture of the : v convexity of Wigner-Yanase-Dyson skew information. In addition, the uncertainty relation has i X been widely studied in quantum information theory [21, 31, 29]. In particular, the relations r betweenskewinformationanduncertaintyrelationhavebeenstudiedin[17,4,8,9,7]. Quantum a Fisher information is also called monotone metric which was introduced by Petz [23] and the Wigner-Yanase-Dyson metric is connected to quantum Fisher information (monotone metric) as a special case. Recently, Hansen gave a further development of the notion of monotone metric, so-called metric adjusted skew information [12]. The Wigner-Yanase-Dyson skew information is also connected to the metric adjusted skew information as a special case. That is, the metric adjusted skew information gave a class including the Wigner-Yanase-Dyson skew information, whilethemonotonemetricgave aclassincludingtheWigner-Yanase-Dyson metric. Inthepaper [12], the metric adjusted correlation measure was also introduced as a generalization of the quantum covariance and correlation measure defined in [17]. Therefore there is a significance to givetherelation amongtheWigner-Yanase-Dyson skewinformation, metricadjustedcorrelation measure and uncertainty relation for the fundamental studies on quantum information theory. ∗E-mail:[email protected] †E-mail:[email protected] 1 We start from the Heisenberg uncertainty relation [13]: 1 V (A)V (B) Tr[ρ[A,B]]2 (1) ρ ρ ≥ 4| | for a quantum state (density operator) ρ and two observables (self-adjoint operators) A and B. The further stronger result was given by Schro¨dinger in [27, 28]: 1 V (A)V (B) Re Cov (A,B) 2 Tr[ρ[A,B]]2, (2) ρ ρ ρ −| { }| ≥ 4| | where the covariance is defined by Cov (A,B) Tr[ρ(A Tr[ρA]I)(B Tr[ρB]I)]. ρ ≡ − − The Wigner-Yanase skew information represents a measure for non-commutativity between a quantum state ρ and an observable H. Luo introduced the quantity U (H) representing a ρ quantum uncertainty excluding the classical mixture [18]: U (H) V (H)2 (V (H) I (H))2, (3) ρ ρ ρ ρ ≡ − − q with the Wigner-Yanase skew information [32]: 1 I (H) Tr (i[ρ1/2,H ])2 = Tr[ρH2] Tr[ρ1/2H ρ1/2H ], H H Tr[ρH]I ρ ≡ 2 0 0 − 0 0 0 ≡ − h i and then he successfully showed a new Heisenberg-type uncertainty relation on U (H) in [18]: ρ 1 U (A)U (B) Tr[ρ[A,B]]2. (4) ρ ρ ≥ 4| | As stated in [18], the physical meaning of the quantity U (H) can be interpreted as follows. ρ For a mixed state ρ, the variance V (H) has both classical mixture and quantum uncertainty. ρ Also, the Wigner-Yanase skew information I (H) represents a kind of quantum uncertainty ρ [19, 20]. Thus, the difference V (H) I (H) has a classical mixture so that we can regard that ρ ρ − the quantity U (H) has a quantum uncertainty excluding a classical mixture. Therefore it is ρ meaningful and suitable to study an uncertainty relation for a mixed state by the use of the quantity U (H). ρ Recently, a one-parameter extension of the inequality (4) was given in [33]: U (A)U (B) α(1 α)Tr[ρ[A,B]]2, (5) ρ,α ρ,α ≥ − | | where U (H) V (H)2 (V (H) I (H))2, ρ,α ρ ρ ρ,α ≡ − − q with the Wigner-Yanase-Dyson skew information I (H) is defined by ρ,α 1 I (H) Tr (i[ρα,H ])(i[ρ1−α,H ]) = Tr[ρH2] Tr[ραH ρ1−αH ], ρ,α ≡ 2 0 0 0 − 0 0 (cid:2) (cid:3) It is notable that the convexity of I (H) with respect to ρ was successfully proven by Lieb ρ,α in [16]. The further generalization of the Heisenberg-type uncertainty relation on U (H) has ρ been given in [34] using the generalized Wigner-Yanase-Dyson skew information introduced in [3]. See also [1, 5, 7, 8] for the recent studies on skew informations and uncertainty relations. Motivated by the fact that the Schro¨dinger uncertainty relation is a stronger result than the Heisenberg uncertainty relation, a new Schro¨dinger-type uncertainty relation for mixed states using Wigner-Yanase skew information was shown in [4]. That is, for a quantum state ρ and two observables A and B, we have 1 U (A)U (B) Re Corr (A,B) 2 Tr[ρ[A,B]]2, (6) ρ ρ ρ −| { }| ≥ 4| | 2 where the correlation measure [17] is defined by Corr (X,Y) Tr[ρX∗Y] Tr[ρ1/2X∗ρ1/2Y] ρ ≡ − for any operators X and Y. This result refined the Heisenberg-type uncertainty relation (4) shownin[18]formixedstates(generalstates). Weeasilyfindthattheinequality(6)isequivalent to the following inequality: U (A)U (B) Corr (A,B)2. (7) ρ ρ ρ ≥ | | The main purpose of this paper is to give some extensions of the inequality (7) by using the Wigner-Yanase-Dyson skew information I (H) and the metric adjusted correlation measure ρ,α introduced in [12]. 2 Schr¨odinger uncertainty relation with Wigner-Yanase-Dyson skew information In this section, we give a generalization of the Schro¨dinger type uncertainty relation (7) by the use of the quantity U (H) defined by the Wigner-Yanase-Dyson skew information I (H). ρ,α ρ,α Theorem 2.1 For α [1/2,1], a quantum state ρ and two observables A and B, we have ∈ U (A)U (B) 4α(1 α)Corr (A,B)2. (8) ρ,α ρ,α ρ,α ≥ − | | where the generalized correlation measure [14, 36] is defined by Corr (X,Y) Tr[ρX∗Y] Tr[ραX∗ρ1−αY] ρ,α ≡ − for any operators X and Y. To prove Theorem 2.1, we need the following lemmas. ∞ Lemma 2.2 ([33])Foraspectral decomposition ofρ= λ φ φ , puttingh φ H φ , j=1 j| jih j| ij ≡ h i| 0| ji we have the following relations. P (i) For the Wigner-Yanase-Dyson skew information, we have I (H) = λα λα λ1−α λ1−α h 2. ρ,α i − j i − j | ij| Xi<j (cid:0) (cid:1)(cid:16) (cid:17) (ii) For the quantity associated to the Wigner-Yanase-Dyson skew information: 1 J (H) Tr ( ρα,H )( ρ1−α,H ) = Tr[ρH2]+Tr[ραH ρ1−αH ], ρ,α ≡ 2 { 0} 0 0 0 0 (cid:2) (cid:8) (cid:9) (cid:3) where X,Y XY +YX is an anti-commutator, we have { } ≡ J (H) λα+λα λ1−α+λ1−α h 2. ρ,α ≥ i j i j | ij| Xi<j (cid:0) (cid:1)(cid:16) (cid:17) Lemma 2.3 ([2, 33]) For any t > 0 and α [0,1], we have ∈ (1 2α)2(t 1)2 (tα t1−α)2. − − ≥ − 3 ∞ Proof of Theorem 2.1: We take a spectral decomposition ρ = λ φ φ . If we put j=1 j| jih j| a = φ A φ and b = φ B φ , where A = A Tr[ρA]I and B = B Tr[ρB]I, then ij i 0 j ji j 0 i 0 0 h | | i h | | i − P − we have Corr (A,B) = Tr[ρAB] Tr[ραAρ1−αB] ρ,α − = Tr[ρA B ] Tr[ραA ρ1−αB ] 0 0 0 0 − ∞ = (λ λαλ1−α)a b i − i j ij ji i,j=1 X = (λ λαλ1−α)a b i − i j ij ji i6=j X = (λ λαλ1−α)a b +(λ λαλ1−α)a b . (9) i − i j ij ji j − j i ji ij Xi<j n o Thus we have Corr (A,B) λ λαλ1−α a b + λ λαλ1−α a b . | ρ,α | ≤ | i− i j || ij|| ji| | j − j i || ji|| ij| Xi<j n o Since a = a and b = b ,takingasquareofbothsidesandthenusingSchwarzinequality ij ji ij ji | | | | | | | | and Lemma 2.2, we have 4α(1 α)Corr (A,B)2 ρ,α − | | 2 4α(1 α) λ λαλ1−α + λ λαλ1−α a b ≤ − | i − i j | | j − j i | | ij|| ji| Xi<j n o 2 = 2 α(1 α) λα+λα λ1−α λ1−α a b − i j | i − j || ij|| ji| Xi<j p (cid:0) (cid:1) 2 2 α(1 α) λ λ a b i j ij ji ≤ − | − || || | Xi<j p 2 1/2 λα λα λ1−α λ1−α λα+λα λ1−α+λ1−α a b ≤ i − j i − j i j i j | ij|| ji| Xi<j n(cid:0) (cid:1)(cid:16) (cid:17)(cid:0) (cid:1)(cid:16) (cid:17)o λα λα λ1−α λ1−α a 2 λα+λα λ1−α+λ1−α b 2 ≤ i − j i − j | ij| i j i j | ij| Xi<j (cid:0) (cid:1)(cid:16) (cid:17) Xi<j (cid:0) (cid:1)(cid:16) (cid:17) I (A)J (B) ρ,α ρ,α ≤ In the above process, the inequality (xα+yα)x1−α y1−α x y for x,y 0 and α [1,1] | − | ≤ | − | ≥ ∈ 2 and the inequality 4α(1 α)(x y)2 (xα yα) x1−α y1−α (xα+yα) x1−α+y1−α for − − ≤ − − x,y 0 and α [0,1], which can be proven by Lemma 2.3, were used. By the similar way, we ≥ ∈ (cid:0) (cid:1) (cid:0) (cid:1) also have 4α(1 α)Corr (A,B)2 I (B)J (A). ρ,α ρ,α ρ,α − | | ≤ Thus for α 1 we have ≥ 2 4α(1 α)Corr (A,B)2 U (A)U (B). (10) ρ,α ρ,α ρ,α − | | ≤ Note that Theorem 2.1 recovers the inequality (7), if we take α = 1. 2 4 Remark 2.4 We take α = 0.1 and 1 1 0 2 2 i 2 i ρ= ,A = − ,B = , 3 0 2 2+i 1 i 1 (cid:18) (cid:19) (cid:18) (cid:19) (cid:18) − (cid:19) then we have U (A)U (B) 4α(1 α)Corr (A,B)2 0.28332. ρ,α ρ,α ρ,α − − | | ≃ − Therefore the inequality (8) does not hold for α [0,1/2) in general. ∈ Corollary 2.5 Underthe same assumptions withTheorem 2.1, we have the following inequality: U (A)U (B) 4α(1 α) Re Corr (A,B) 2 Im Tr[ραAρ1−αB] 2 ρ,α ρ,α ρ,α − − | { }| −| | α(1 α)Tr[ρ[A,B]]2. (11) (cid:0) (cid:8) (cid:9) (cid:1) ≥ − | | Proof: From 1 Im Corr (A,B) = Tr[ρ[A,B]] Im Tr[ραAρ1−αB] , ρ,α { } 2i − (cid:8) (cid:9) we have 1 Tr[ρ[A,B]] 2 Im Corr (A,B) 2+ Im Tr[ραAρ1−αB] 2. ρ,α 4| | ≤ | { }| | | Thus we have (cid:8) (cid:9) Corr (A,B)2 = Re Corr (A,B) 2+ Im Corr (A,B) 2 ρ,α ρ,α ρ,α | | | { }| | { }| 1 Re Corr (A,B) 2+ Tr[ρ[A,B]] 2 Im Tr[ραAρ1−αB] 2, ρ,α ≥ | { }| 4| | −| | (cid:8) (cid:9) which proves the corollary. Remark 2.6 The following inequality does not hold in general for α [1,1]: ∈ 2 Re Corr (A,B) 2 Im Tr[ραAρ1−αB] 2. (12) ρ,α | { }| ≥| | Because we have a counter-example as follows. We t(cid:8)ake α = 2 and (cid:9) 3 1 2 3 2 2 i 2 i ρ= ,A = − ,B = , 7 3 5 2+i 1 i 1 (cid:18) (cid:19) (cid:18) (cid:19) (cid:18) − (cid:19) then we have Re Corr (A,B) 2 Im Tr[ραAρ1−αB] 2 0.0548142. ρ,α | { }| −| | ≃ − This shows Theorem 2.1 does not refine the i(cid:8)nequality (5) in(cid:9)general. 3 Two-parameter extensions Inthissection,weintroducetheparametricextendedcorrelationmeasureCorr (X,Y)bythe ρ,α,γ convex combination between Corr (X,Y) and Corr (X,Y). Then we establish the para- ρ,α ρ,1−α metric extended Schro¨dinger-type uncertainty relation applying the parametric extended corre- lation measure Corr (X,Y). In addition, introducing the symmetric extended correlation ρ,α,γ (sym) measureCorr (X,Y)bytheconvexcombination betweenCorr (X,Y)andCorr (Y,X), ρ,α,γ ρ,α ρ,α we show its Schro¨dinger-type uncertainty relation. 5 Definition 3.1 We define the parametric extended correlation measure Corr (X,Y) for two ρ,α,γ parameters α,γ [0,1] by ∈ Corr (X,Y) γCorr (X,Y)+(1 γ)Corr (X,Y) (13) ρ,α,γ ρ,α ρ,1−α ≡ − for any operators X and Y. Note that we have Corr (H,H) = I (H) for any observable H. Then we can prove the ρ,α,γ ρ,α following inequality. Theorem 3.2 If 0 α,γ 1 or 1 α,γ 1, then we have ≤ ≤ 2 2 ≤ ≤ U (A)U (B) 4α(1 α)Corr (A,B)2 ρ,α ρ,α ρ,α,γ ≥ − | | for two observables A, B and a quantum state ρ. Proof: By the similar way of the proof of Theorem 2.1, we have Eq.(9) and we also have Corr (A,B) = Tr[ρAB] Tr[ρ1−αAραB] ρ,1−α − = (λ λ1−αλα)a b +(λ λ1−αλα)a b . (14) i − i j ij ji j − j i ji ij Xi<j n o Thus we have Corr (A,B) = γCorr (A,B)+(1 γ)Corr (A,B) ρ,α,γ ρ,α ρ,α − = γλα(λ1−α λ1−α)+(1 γ)λ1−α(λα λα) a b i i − j − i i − j ij ji Xi<j n o + γλα(λ1−α λ1−α)+(1 γ)λ1−α(λα λα) a b . j j − i − j j − i ji ij Xi<j n o Since we have a = a and b = b , we then have ij ji ij ji | | | | | | | | Corr (A,B) γ(λα+λα)λ1−α λ1−α +(1 γ)(λ1−α +λ1−α)λα λα a b | ρ,α,γ | ≤ i j | i − j | − i j | i − j| | ij|| ji| Xi<j n o λ λ a b , i j ij ji ≤ | − || || | i<j X thanks to the inequality γ(xα+yα)x1−α y1−α +(1 γ)(x1−α +y1−α)xα yα x y (15) | − | − | − |≤ | − | for 0 α,γ 1 or 1 α,γ 1, and x,y 0. The rest of the proof goes similar way to that of ≤ ≤ 2 2 ≤ ≤ ≥ Theorem 2.1. Corollary 3.3 For any α [0,1], two observables A, B and a quantum state ρ, we have ∈ Uρ,α(A)Uρ,α(B) ≥ 4α(1−α)|Corrρ,α,21(A,B)|2. Proof: If γ = 1, then the equality of the inequality (15) holds for any α [0,1] and x,y 0. 2 ∈ ≥ Therefore we have the present corollary from Theorem 3.2. We may define the following correlation measure instead of Definition 3.1. 6 (sym) Definition 3.4 We define a symmetric extended correlation measure Corr (X,Y) for two ρ,α,γ parameters α,γ [0,1] by ∈ Corr(sym)(X,Y) γCorr (X,Y)+(1 γ)Corr (Y,X) (16) ρ,α,γ ≡ ρ,α − ρ,α for any operators X and Y. (sym) (sym) Note that we have Corr (A,B) = Corr (B,A) for self-adjoint operators A and B. ρ,α,γ ρ,α,γ Then we have the following therem by the similar proof of the above using the inequality (xα+yα)x1−α y1−α x y | − |≤ | − | for x,y 0 and α 1. ≥ ≥ 2 Theorem 3.5 For α [1,1] and γ [0,1], we have ∈ 2 ∈ U (A)U (B) 4α(1 α)Corr(sym)(A,B)2 ρ,α ρ,α ≥ − | ρ,α,γ | for two observables A, B and a quantum state ρ. 4 A further generalization by metric adjusted correlation mea- sure Inspired by the recent results in [10] and the concept of metric adjusted skew information intro- duced by Hansen in [12], we here give a further generalization for Schro¨dinger-type uncertainty relation applying metric adjusted correlation measure introduced in [12]. We firstly give some notations according to those in [10]. Let M (C) and M (C) be the set of all n n complex n n,sa × matrices and all n n self-adjoint matrices, equipped with the Hilbert-Schmidt scalar product × A,B = Tr[A∗B], respectively. Let M (C) be the set of all positive definite matrices of n,+ h i M (C) and M (C) be the set of all density matrices, that is n,sa n,+,1 M (C) ρ M (C)Trρ= 1,ρ > 0 M (C). n,+,1 n,sa n,+ ≡ { ∈ | } ⊂ Here X M (C) means we have φX φ 0 for any vector φ Cn. In the study of n,+ ∈ h | | i ≥ | i ∈ quantum physics, we usually use a positive semidefinite matrix with a unit trace as a density operator ρ. In this section, we assume the invertibility of ρ. A function f : (0,+ ) R is said operator monotone if the inequalities 0 f(A) ∞ → ≤ ≤ f(B) hold for any A,B M (C) such that 0 A B. An operator monotone function n,sa ∈ ≤ ≤ f : (0,+ ) (0,+ ) is said symmetric if f(x) = xf(x−1) and normalized if f(1) = 1. We ∞ → ∞ represents the set of all symmetric normalized operator monotone functions by . We have op F the following examples as elements of : op F Example 4.1 ([12, 10, 6, 25]) 2x x+1 x 1 f (x) = , f (x) = , f (x) = − , RLD SLD BKM x+1 2 logx √x+1 2 (x 1)2 f (x) = , f (x) = α(1 α) − , α (0,1). WY 2 WYD − (xα 1)(x1−α 1) ∈ (cid:18) (cid:19) − − 7 The functions f (x) and f (x) are normalized in the sense that lim f (x) = 1 BKM WYD x→1 BKM and lim f (x) = 1. Note that a simple proof of the operator monotonicity of f (x) x→1 WYD WYD was given in [6]. See also [30] for the proof of the operator monotonicity of f (x) by use of WYD majorization. Remark 4.2 ([10, 15, 24, 25]) For any f , we have the following inequalities: op ∈F 2x x+1 f(x) , x > 0. x+1 ≤ ≤ 2 That is, all f lies in between the harmonic mean and the arithmetic mean. op ∈ F Forf wedefinef(0) = lim f(x). Wealsodenotethesetsofregularandnon-regular op x→0 ∈ F functions by r = f f(0) = 0 and n = f f(0) = 0 . Fop { ∈ Fop| 6 } Fop { ∈ Fop| } Definition 4.3 ([8, 10]) For f r , we define the function f˜by ∈ Fop 1 f(0) f˜(x) = (x+1) (x 1)2 , (x >0). 2 − − f(x) (cid:26) (cid:27) Then we have the following theorem. Theorem 4.4 ([8, 6, 26]) The correspondence f f˜is a bijection between r and n. → Fop Fop We can use matrix mean theory introduced by Kubo-Ando in [15]. Then a mean m corre- f sponds to each operator monotone function f by the following formula op ∈F m (A,B) = A1/2f(A−1/2BA−1/2)A1/2, f for A,B M (C). By the notion of matrix mean, we may define the set of the monotone n,+ ∈ metrics [23] by the following formula A,B = Tr[Am (L ,R )−1(B)], ρ,f f ρ ρ h i where L (A) = ρA and R (A) = Aρ. ρ ρ Definition 4.5 ([12, 8]) For A,B M (C), ρ M (C) and f r , we define the ∈ n,sa ∈ n,+,1 ∈ Fop following quantities: f(0) Corrf(A,B) i[ρ,A],i[ρ,B] , If(A) Corrf(A,A), ρ ≡ 2 h iρ,f ρ ≡ ρ Cf(A,B) Tr[m (L ,R )(A)B], Cf(A) Cf(A,A), ρ ≡ f ρ ρ ρ ≡ ρ Uf(A) V (A)2 (V (A) If(A))2. ρ ≡ ρ − ρ − ρ q f The quantity I (A) is known as metric adjusted skew information [12]. It is notable that the ρ metric adjusted correlation measure Corrc(A,B) was firstly introduced in [12] for a regular ρ Morozova-Chentsov function c. Recently the notation Ic(A,B) in [1] and the notation If(A,B) ρ ρ in [11] were used. In addition, it is useful for the readers to be noted that the correlation f I (A,B) can be expressed as a difference of covariances [11]. Throughout the present paper, we ρ f usethe notation Corr (A,B) as the metric adjusted correlation measure, to avoid the confusion ρ of the readers. (In the previous sections, we have already used Corr (A,B), Corr (A,B) and ρ ρ,α Corr (A,B) as correlation measures and done I (H) and I (H) as skew informations.) ρ,α,γ ρ ρ,α Then we have the following proposition. 8 Proposition 4.6 ([8, 10]) For A,B M (C), ρ M (C) and f r , we have the ∈ n,sa ∈ n,+,1 ∈ Fop following relations, where we put A A Tr[ρA]I and B B Tr[ρB]I. 0 0 ≡ − ≡ − (1) If(A) = Tr[ρA2] Tr[m (L ,R )(A )A ]= V (A) Cf˜(A ). ρ 0 − f˜ ρ ρ 0 0 ρ − ρ 0 (2) Jf(A) = Tr[ρA2]+Tr[m (L ,R )(A )A ]= V (A)+Cf˜(A ). ρ 0 f˜ ρ ρ 0 0 ρ ρ 0 f f (3) 0 I (A) U (A) V (A). ρ ρ ρ ≤ ≤ ≤ f f f (4) U (A) = I (A)J (A). ρ ρ ρ q (5) Corrf(A,B) = 1Tr[ρA B ] + 1Tr[ρB A ] Tr[m (L ,R )(A )B ] = 1Tr[ρA B ] + ρ 2 0 0 2 0 0 − f˜ ρ ρ 0 0 2 0 0 1Tr[ρB A ] Cf˜(A ,B ). 2 0 0 − ρ 0 0 Thefollowinginequality isthefurthergeneralization ofCorollary 3.3bytheuseofthemetric adjusted correlation measure. Theorem 4.7 For f r , if we have ∈ Fop x+1 +f˜(x) 2f(x), (17) 2 ≥ then we have Uf(A)Uf(B) 4f(0)Corrf(A,B)2, (18) ρ ρ ≥ | ρ | for A,B M (C) and ρ M (C). n,sa n,+,1 ∈ ∈ In order to prove Theorem 4.7, we use the following two lemmas. Lemma 4.8 ([35]) If Eq.(17) is satisfied, then we have the following inequality: 2 x+y m (x,y)2 f(0)(x y)2. 2 − f˜ ≥ − (cid:18) (cid:19) Proof: By Eq.(17), we have x+y +m (x,y) 2m (x,y). 2 f˜ ≥ f We also have x m (x,y) = yf˜ f˜ y (cid:18) (cid:19) 2 y x x f(0) = +1 1 2 y − y − f(x/y) ( ) (cid:18) (cid:19) x+y f(0)(x y)2 = − . 2 − 2m (x,y) f Therefore 2 x+y x+y x+y m (x,y)2 = m (x,y) +m (x,y) 2 − f˜ 2 − f˜ 2 f˜ (cid:18) (cid:19) (cid:26) (cid:27)(cid:26) (cid:27) f(0)(x y)2 − 2m (x,y) f ≥ 2m (x,y) f = f(0)(x y)2. − 9 f f f f We have the following expressions for the quantities I (A), J (A), U (A) and Corr (A,B) ρ ρ ρ ρ by using Proposition 4.6 and a mean m . f˜ Lemma 4.9 ([10]) Let φ , φ , , φ be a basis of eigenvectors of ρ, corresponding to 1 2 n {| i | i ··· | i} the eigenvalues λ ,λ , ,λ . We put a = φ A φ ,b = φ B φ , where A A 1 2 n jk j 0 k jk j 0 k 0 { ··· } h | | i h | | i ≡ − Tr[ρA]I and B B Tr[ρB]I for A,B M (C) and ρ M (C). Then we have 0 n,sa n,+,1 ≡ − ∈ ∈ 1 If(A) = (λ +λ )a a m (λ ,λ )a a ρ 2 j k jk kj − f˜ j k jk kj j,k j,k X X λ +λ = 2 j k m (λ ,λ ) a 2, 2 − f˜ j k | jk| j<k(cid:26) (cid:27) X 1 Jf(A) = (λ +λ )a a + m (λ ,λ )a a ρ 2 j k jk kj f˜ j k jk kj j,k j,k X X λ +λ 2 j k +m (λ ,λ ) a 2, ≥ 2 f˜ j k | jk| j<k(cid:26) (cid:27) X 2 2 1 Uf(A)2 = (λ +λ )a 2 m (λ ,λ )a 2 ρ 4 j k | jk| − f˜ j k | jk| j,k j,k X X and 1 1 Corrf(A,B) = λ a b + λ a b m (λ ,λ )a b ρ 2 j jk kj 2 k jk kj − f˜ j k jk kj j,k j,k j,k X X X λ +λ λ +λ j k k j = m (λ ,λ ) a b + m (λ ,λ ) a b . 2 − f˜ j k jk kj 2 − f˜ k j kj jk j<k(cid:18) (cid:19) j<k(cid:18) (cid:19) X X (19) We are now in a position to prove Theorem 4.7. Proof of Theorem 4.7: From Eq.(19), we have λ +λ λ +λ Corrf(A,B) j k m (λ ,λ ) a b + j k m (λ ,λ ) a b | ρ | ≤ 2 − f˜ j k jk kj 2 − f˜ k j kj jk j<k(cid:12)(cid:18) (cid:19) (cid:12) j<k(cid:12)(cid:18) (cid:19) (cid:12) X(cid:12) (cid:12) X(cid:12) (cid:12) (cid:12)λ +λ (cid:12) (cid:12)λ +λ (cid:12) (cid:12) j k m (λ ,λ ) a b (cid:12)+ (cid:12) j k m (λ ,λ ) a b (cid:12) ≤ 2 − f˜ j k | jk|| kj| 2 − f˜ k j | kj|| jk| j<k(cid:12) (cid:12) j<k(cid:12) (cid:12) X(cid:12) (cid:12) X(cid:12) (cid:12) (cid:12) λ +λ (cid:12) (cid:12) (cid:12) = 2 (cid:12) j k m (λ ,λ(cid:12)) a b (cid:12) (cid:12) 2 − f˜ j k | jk|| kj| j<k(cid:12) (cid:12) X(cid:12) (cid:12) (cid:12) (cid:12) λ(cid:12)j λk ajk bkj . (cid:12) ≤ | − || || | j<k X Then we have 2 f(0)Corrf(A,B)2 f(0)1/2 λ λ a b | ρ | ≤ | j − k|| jk|| kj| j<k X 10