Abstract
Some Choi–Davis–Jensen’s type trace inequalities for convex functions are proved. Also, we generalize these inequalities for any arbitrary operator mean via operator monotone decreasing functions. In particular, we present some new order among \(\mathrm{tr}(\Phi (C)A)\) and \(\mathrm{tr}(\Phi (C)A^{-1})\). New refinements of some power type trace inequalities via reverse and refinement of Young’s inequality are established. Among our results, we obtain new versions of the Hölder type trace inequality for any arbitrary operator mean.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
For this purpose, let \({\mathcal {B}}(\mathcal {H})\) stand for the \(C^*\)-algebra of all bounded linear operators on a complex Hilbert space \((\mathcal {H},\langle \cdot ,\cdot \rangle )\). An operator \(A\in {\mathcal {B}}(\mathcal {H})\) is called positive if \(\langle Ax, x\rangle \ge 0\) for all \(x\in \mathcal {H}\), and we then write \(A\ge 0\). We write \(A>0\) if it is a positive invertible operator. Let \(A,B \in {\mathcal {B}}(\mathcal {H})\) be two self-adjoint operators. The partial order \(A\le B\) is defined as \(B-A\ge 0\). The absolute value of A is denoted by |A|, that is \(|A|=(A^*A)^{\frac{1}{2}}\). A continuous real-valued function f defined on interval J is said to be operator monotone increasing (decreasing) if for every two positive operators A and B with spectral in J, the inequality \(A \le B\) implies \(f(A) \le f(B)\) (\(f(A) \ge f(B)\)), respectively. We recall that every operator monotone decreasing function is operator convex. An operator mean \(\sigma _{f}\) in the sense of Kubo–Ando [2] is defined by a positive operator monotone increasing function f on the half interval \((0, \infty )\) with \(f(1)=1\) as
for positive invertible operators \(A,B\in {\mathcal {B}}(\mathcal {H})\). For \(A, B>0\) the weighted operator arithmetic, geometric and harmonic means, respectively, by
and \(A !_\nu B = \big ( (1-\nu ) A^{-1} +\nu B ^{-1} \big )^{-1} \), where \(\nu \in [0,1]\). A linear map \(\Psi :{\mathcal {B}}(\mathcal {H}) \rightarrow {\mathcal {B}}(\mathcal {K})\) is called positive if \(A\ge 0\) implies \(\Psi (A)\ge 0\). It is said to be unital if \(\Psi (I)=I\). Davis [9] and Choi [7] showed that if \(\Psi \) is a unital positive linear map on \({\mathcal {B}}(\mathcal {H})\) and if f is an operator convex function on an interval J, then so-called the Choi–Davis–Jensen inequality \(f\big (\Psi (A)\big )\le \Psi \big (f(A)\big )\) holds for every self-adjoint operator A on \(\mathcal {H}\) whose spectrum is contained in J.
Let \(\{e_i\}_{i\in I}\) be an orthonormal basis of \(\mathcal {H}\). An operator \(A\in {\mathcal {B}}(\mathcal {H})\) is said to be trace class (denoted by \({\mathcal {B}}_1(\mathcal {H})\)) if \(\Vert A\Vert _1:=\sum _{i\in I} \langle |A| e_i, e_i\rangle < \infty \). We define the trace of a trace class operator \(A\in {\mathcal {B}}_1(\mathcal {H})\) to be
If \(\Vert A\Vert _2:= \sum _{i\in I} \Vert Ae_i\Vert ^2 <\infty \), A is said to be Hilbert–Schmidt operator (denotes \(A\in {\mathcal {B}}_2(\mathcal {H})\)). The definitions of \(\Vert A\Vert _1\), \(\Vert A\Vert _2\) and trace do not depend on the choice of the orthonormal basis \(\{e_i\}_{i\in I}\).
The Hölder’s type inequality [21] is given as follows:
where \(\alpha \in (0,1)\) and \(A,B\in {\mathcal {B}}(\mathcal {H})\) with \(|A|^{1/\alpha },|B|^{{1}/{(1-\alpha )}}\in {\mathcal {B}}_1(\mathcal {H})\).
In particular, for \(\alpha =\frac{1}{2}\), we get the Schwarz inequality
where \(|A|^2, |B|^2 \in {\mathcal {B}}_1(\mathcal {H})\). We have the following Hölder type trace inequality for the weighted geometric mean [10]: if A, B are positive invertible operators, \(p, q > 1\) with \(\frac{1}{p} + \frac{1}{q} = 1\) and \(A^{p}, B^{q} \in {\mathcal {B}}_{1}(\mathcal {H})\), then \(B^{q} \sharp _{\frac{1}{q}} A^{p} \in {\mathcal {B}}_{1}(\mathcal {H})\) and
Also, if \(C \in {\mathcal {B}}_{1}(\mathcal {H})\) and \(C \ge 0\), then \(CA^{p}, CB^{q}, C(B^{q} \sharp _{\frac{1}{p}} A^{p}) \in {\mathcal {B}}_{1}(\mathcal {H})\) and
According to [13], if A and B are self-adjoint operators with \(A\le B\) and \(P\in {\mathcal {B}}_1(\mathcal {H})\) with \(P\ge 0\), then
Moreover, if \(A \ge 0\), then \(0\le \mathrm{tr} (PA) \le \Vert A\Vert \mathrm{tr} (P)\), and
for a self-adjoint operator A and \(P\in {\mathcal {B}}_1(\mathcal {H})\) with \(P\ge 0\).
For the theory of trace functional and their applications, the reader is referred to [22]. Some classical trace inequalities investigated in [8, 20, 24], which are continuations of the work of Bellman [4].
We express the following results which is derived from [6, Sections 31.2.1 and 31.2.2].
Lemma 1.1
Let \(\Phi :{\mathcal {B}}_1(\mathcal {K})\rightarrow {\mathcal {B}}_1(\mathcal {H})\)be a positive linear mapping such that
for all \(W\in {\mathcal {B}}_1(\mathcal {K})\)with \(W\ge 0\)and \(\mathrm{tr}(W)=1\), i.e., for all density operators W on Hilbert space \(\mathcal {K}\). Then its dual map \(\Phi ^{*}\)is a linear map \({\mathcal {B}}(\mathcal {H})\rightarrow {\mathcal {B}}(\mathcal {K})\)which is well defined by
for all \(B\in {\mathcal {B}}(\mathcal {H})\)and \(A\in {\mathcal {B}}_1(\mathcal {K})\).
\(\Phi \)is continuous and \(\Phi ^{*}(I)\le I\)is equivalent to \(\mathrm{tr}\big (\Phi (W)\big )\le 1\)for all density operators W on \(\mathcal {K}\). A linear map \(\Phi :{\mathcal {B}}_1(\mathcal {K}) \rightarrow {\mathcal {B}}_1(\mathcal {H})\)is positive if and only if its dual map \(\Phi ^*\)is positive.
Corollary 1.2
For a positive linear mapping \(\Phi : {\mathcal {B}}_{1}(\mathcal {K}) \rightarrow {\mathcal {B}}_{1}(\mathcal {H})\)the following statements are equivalent:
-
(i)
\(\mathrm{tr}\big (\Phi (W)\big )\le 1\)for all density operators W on \(\mathcal {K}\);
-
(ii)
\(\Phi \)is continuous and \(\Phi ^{*}(I) \le I\).
Theorem 1.3
A linear mapping \(\Phi :{\mathcal {B}}_1(\mathcal {K})\rightarrow {\mathcal {B}}_1(\mathcal {H})\)is positive if and only if \(\Phi ^{*}:{\mathcal {B}}(\mathcal {H})\rightarrow {\mathcal {B}}(\mathcal {K})\)is positive.
For a comprehensive account on positive linear maps see [5, 6, 23].
Dragomir in [11,12,13,14] proved Jensen’s type trace inequalities for convex functions.
In Sect. 2, we first extend the Choi–Davis–Jensen inequality for trace. In particular, with the help of this we obtain some power type trace inequalities for positive linear maps. Also we generalize Choi–Davis–Jensen’s type trace inequalities for any arbitrary operator mean via operator monotone decreasing functions. As an application of these, we prove some new order among \(\mathrm{tr}(\Phi (C)A)\) and \(\mathrm{tr}(\Phi (C)A^{-1})\), where \(0<mI\le A\le MI\), \(C \in {\mathcal {B}}_{1}(\mathcal {K})\), \(C>0\) and \(\Phi \) is a positive linear mapping satisfying (1.6). Next, we present reverses of the Choi–Davis–Jensen inequality for trace.
In Sect. 3 by applying a recent reverse and refinement of Young’s inequality, we establish some power type trace inequalities. Also we show new versions of the Hölder type trace inequality for any arbitrary operator mean.
Some examples for the power function and logarithm are presented in Sect. 4.
2 Choi–Davis–Jensen’s type trace inequalities
We recall the gradient inequality for the convex function \(f:[m,M]\rightarrow \mathbb {R}\), namely
for any \(s,t \in (m, M)\), where \(\delta _f(t)\in [{f'}_{-}(t),{f'}_{+}(t)]\).
Now, we are ready to present our first result.
Theorem 2.1
Let \(\Phi :{\mathcal {B}}_1(\mathcal {K})\rightarrow {\mathcal {B}}_1(\mathcal {H})\)be a positive linear mapping satisfying (1.6), whose adjoint is \(\Phi ^{*}\), A be a self-adjoint operator on the Hilbert space \(\mathcal {H}\)and assume that \(\mathrm{Sp}(A)\subseteq [m,M]\)for some scalars m, M with \(m<M\). If f is a continuous convex function on [m, M] and \(B \in {\mathcal {B}}_1(\mathcal {K})\setminus \{0\}\)is a strictly positive operator, then we have \(\frac{\mathrm{tr}\big (B(\Phi ^*(A)\big )}{\mathrm{tr}\big (\Phi (B)\big )}\in [m,M]\)and the Choi–Davis–Jensen inequality
Proof
According to the hypothesis, we have
for any \(i\in I\), which by summation (2.3), we get
If follows from the properties of trace and equality (1.7) that
By inequality (2.4) and equality (2.5), we conclude that \(\frac{\mathrm{tr}\big (B\Phi ^*(A)\big )}{\mathrm{tr}\big (\Phi (B)\big )}\in [m,M]\). Utilising the gradient inequality (2.1), we have
for any \(s\in [m,M]\), where
Inequality (2.6) implies in the operator order of \({\mathcal {B}}(\mathcal {H})\) that
which can be written as
for any \(y\in \mathcal {H}\). If we take in (2.7), \(y=\big (\Phi (B)\big )^{\frac{1}{2}}e_i\) and summing over \(i\in I\), we obtain
Thus, using the properties of trace and equality (1.7) to the above inequality, we have that
and inequality (2.2) is thus proved. \(\square \)
As an application of Theorem 2.1, we have the following inequality.
Corollary 2.2
If A is a self-adjoint operator on the Hilbert space \(\mathcal {H}\)satisfying \(Sp(A)\subseteq [m, M]\)for some scalars m, M with \(0\le m<M\)and \(B\in {\mathcal {B}}_{1}(\mathcal {H}) \setminus \{0\}\), \(B>0\). Then for every positive linear map \(\Phi ^{*}:{\mathcal {B}}(\mathcal {H})\rightarrow {\mathcal {B}}(\mathcal {H})\)which is adjoint of a continuous linear map with \(\Phi ^{*}(I) \le I\)and every \(p \ge 1\)
Proof
Let \(\Phi ^*\) be the adjoint of the continuous linear map \(\Phi :{\mathcal {B}}_1(\mathcal {H})\rightarrow {\mathcal {B}}_1(\mathcal {H})\), then \(\Phi \) is positive and \(\mathrm{tr}\big (\Phi (W)\big )\le 1\) for any density operators W on Hilbert space \(\mathcal {H}\) by Theorem 1.3 and Corollary 1.2. Since \(\mathrm{tr}\big (\Phi (B)\big )=\mathrm{tr}\big ( B\Phi ^*(I)\big )\) and \(\Phi ^*(I) \le I\); hence, using inequality (1.4), we have \(\mathrm{tr}\big (\Phi (B)\big )\le \mathrm{tr}(B)\). Therefore, by convexity of the function \(f(x)=x^p\) for \(p\ge 1\) and applying inequality (2.2), the proof is complete. \(\square \)
In [17], the authors proved that for two positive operators \(0<mI\le A, B \le MI\) and \(!_{\nu } \le \sigma _{1}, \sigma _{2} \le \bigtriangledown _\nu \),
where \(g:(0, \infty ) \rightarrow (0, \infty )\) is an operator monotone increasing, \(\Psi \) is a positive unital linear map and k stands for the known Kantorovich constant \(k=\dfrac{(M+m)^{2}}{4mM}\).
It is well known that for positive invertible operators \(A, B \in {\mathcal {B}}(\mathcal {H})\), if \(\sigma _{\nu }\) is a symmetric operator mean, then
Furuichi et al. [16] showed the following new reverse inequalities of (2.9): if \(0<mI\le A, B \le MI\). Then
where \(\lambda = \min \lbrace \nu , 1-\nu \rbrace \), \(\mu = \max \lbrace \nu , 1-\nu \rbrace \) and \(\nu \in [0, 1]\).
In the following result by Theorem 2.1, we extend the reverse of inequality (2.8) for a positive operator monotone decreasing function on \((0, \infty )\) involving trace.
Theorem 2.3
Let \(0<mI\le A, B \le MI\). Then for every positive linear map \(\Psi \) on \({\mathcal {B}}(\mathcal {H})\)and \(\sigma _1\), \(\sigma _2\)between \(\bigtriangledown _\nu \)and \(!_\nu \)
where \(f: (0, \infty ) \rightarrow (o, \infty )\)is an operator monotone decreasing function, \(\lambda =\min \{ \nu , 1-\nu \}\), \(\mu =\max \{\nu , 1-\nu \}\) and \(\nu \in [0,1]\).
Moreover, if \(\nu =\frac{1}{2}\), then for any \(! \le \sigma _{1}, \sigma _{2} \le \bigtriangledown \),
Proof
Since \(0<mI \le A, B \le MI\), we can write
By virtue of [3, Remark 2.6], it follows that
Using the left-hand side of inequality (2.10) and the fact that for \(\alpha \ge 1\), \(f(\alpha t)\ge \frac{1}{\alpha } f(t)\) when f is an operator monotone decreasing we have
and similarly from the right-hand side of inequality (2.10) we obtain
It follows from (2.12), (2.13) and (2.14) that
Applying positive linear map \(\Psi \) and [1, Theorem 3] and inequality (2.12), we get the required inequality (2.11). \(\square \)
Since the power function \(x^{p}\) on \((0, \infty )\) is operator monotone decreasing for \(p \in [-1, 0]\), we get the following result.
Corollary 2.4
Let \(0<mI\le A,B \le MI\), \(\Psi \)be a positive linear map and operator means \(\sigma _1,\sigma _2\)between \(\bigtriangledown _\nu \)and \(!_\nu \)and \(-1 \le p \le 0\). Then
where \(\lambda =\min \{\nu , 1-\nu \}\), \(\mu =\{\nu , 1-\nu \}\)and \(\nu \in [0,1]\).
In the next result, we present new versions of Choi–Davis–Jensen’s type trace inequalities.
Proposition 2.5
Let \(\Phi : {\mathcal {B}}_1(\mathcal {K}) \rightarrow {\mathcal {B}}_1(\mathcal {H})\)be a positive linear mapping satisfying (1.6), whose adjoint is \(\Phi ^*\), \(0 < mI \le A\), \(B \le MI\)and \(C\in {\mathcal {B}}_1(\mathcal {K})\setminus \{0\}\), \(C>0\). If f is a positive operator monotone decreasing on [m, M] and \(!_\nu \le \sigma _1, \sigma _2 \le \bigtriangledown _\nu \), then
where \(\lambda =\min \{\nu , 1-\nu \}\), \(\mu =\max \{\nu -1, \nu \}\)and \(\nu \in [0,1]\).
If \(\nu =\frac{1}{2}\), then for any \(! \le \sigma _{1}, \sigma _{2} \le \bigtriangledown \),
Proof
Due to Theorems 2.1, 2.3 and inequality (1.4), we obtain (2.15). \(\square \)
In the following remark, we can obtain the relation between \(\mathrm{tr}(\Phi (C)A)\) and \(\mathrm{tr}(\Phi (C)A^{-1})\).
Remark 2.6
Let \(\Phi :{\mathcal {B}}_1(\mathcal {K})\rightarrow {\mathcal {B}}_1(\mathcal {H})\) be a positive linear mapping satisfying (1.6), \(A, B \in {\mathcal {B}}(\mathcal {H})\) such that \(0<mI \le A, B \le MI\) and \(C\in {\mathcal {B}}_1(\mathcal {K})\setminus \{0\}\), \(C>0\). If we put \(\sigma _{1}=\sigma _{2}=\bigtriangledown \) and \(f(t)=t^{-1}\) in (2.16), we have
If in (2.17) we take \(A = B\), then we get
To present our next result, we will need the following lemma.
Lemma 2.7
Let T be a self-adjoint operator such that \(\alpha 1_\mathcal {H}\le T\le \beta 1_\mathcal {H}\)for some real constant \(\beta \ge \alpha \)and assume that \(\Phi :{\mathcal {B}}_1(\mathcal {K})\rightarrow {\mathcal {B}}_1(\mathcal {H})\)be a positive linear mapping satisfying (1.6), whose adjoint is \(\Phi ^*\). Then for any strictly positive operator \(S\in {\mathcal {B}}_1(\mathcal {K})\setminus \{0\}\)we have
Proof
The first inequality follows from Choi–Davis Jensen’s inequality (2.2) for the convex function \(f(t)=t^2\). Now observe that
Since \(\mathrm{tr}\big (S\Phi ^*(1_\mathcal {H})\big )=\mathrm{tr}\big (\Phi ^*(1_\mathcal {H})S\big )=\mathrm{tr}\big (\Phi (S)\big ),\) we have
Now, since \(\alpha 1_\mathcal {H}\le T\le \beta 1_\mathcal {H}\), we infer
which implies that
Taking the modulus in (2.19) and using the property (1.5), we obtain (2.21)
Put \(v=T-\frac{\beta +\alpha }{2}\cdot 1_\mathcal {H}\) and \(u=T-\frac{\mathrm{tr}\big (S\Phi ^*(T)\big )}{\mathrm{tr}\big (\Phi (S)\big )}\cdot 1_\mathcal {H}\). Let \(vu=w'|vu|\) be the polar decomposition of vu, where \(w'\) is a unique partial isometry on \(\mathcal {H}\), and \(w''={w'}^*vw'\). Then \(|vu|={w'}^*vu=w''|u|\). Hence,
so
Now by (2.22), (2.20) and using the property (1.4), we get
this proves the first part of (2.18).
Using Schwarz’s inequality (1.1), one can obtain
Observe that
By (2.24) and (2.25), we can write
and by (2.21) and (2.26), we have
which implies that
Therefore, by (2.26), we get
which proves the last part of (2.18). \(\square \)
The following result provides reverses for the inequality (2.2).
Theorem 2.8
Let \(\Phi :{\mathcal {B}}_1(\mathcal {K})\rightarrow {\mathcal {B}}_1(\mathcal {H})\)be a positive linear mapping satisfying (1.6), whose adjoint is \(\Phi ^*\), A be a self-adjoint operator on the Hilbert space \(\mathcal {H}\)and assume that \(\mathrm{Sp}(A)\subseteq [m,M]\)for some scalars m, M with \(m<M\).
If f is a continuously differentiable convex function on [m, M] and \(B\in {\mathcal {B}}_1(\mathcal {K})\setminus \{0\}\)is a strictly positive operator, then we have
and
Proof
By the gradient inequality we have
for any \(s,t\in [m,M]\). This inequality implies in the operator order
which is equivalent to
for any \(y\in \mathcal {H}\), which is of interest in itself as well. If we take \(y=\big (\Phi (B)\big )^{\frac{1}{2}}e_i\) in (2.29), we obtain (2.27).
Since f is continuously convex on [m, M], then \(f'\) is monotonic non-decreasing on [m, M] and \(f'(m)\le f'(t)\le f'(M)\) for any \(t\in [m,M]\). We also observe that
Since
we have
Then by taking the modulus in (2.30) and using the property (1.5), we get the following inequality
By (2.22) and (2.31), we have the part (i) of (2.28).
If follows from Lemma 2.7 that
By applying (2.32) and (2.33) we get the part (ii) of (2.28). We observe that \(L(\Phi ,\Phi ^*,f',B,A)\) can be represented as
Applying a similar argument as above for this representation, we get parts (iii) and (iv) of (2.28). \(\square \)
Remark 2.9
For the convex function \(f(t)=t^p\), \(p\ge 1\), we get the following inequalities of interest:
for some constants m, M with \(M>m>0\) and \(mI \le A \le MI\).
3 The Hölder type trace inequality for an operator mean
The main purpose of this section is to find new refinements of some power type trace inequalities via reverse and refinement of Young’s inequality. Also we show new versions of the Hölder type trace inequality for an operator mean \(\sigma _{f}\).
Let a and b be positive numbers. The famous Young inequality states that \(a^{1-\nu }b^\nu \le (1-\nu ) a+\nu b \) for every \(0\le \nu \le 1\). Liao et al. [18] obtained reverse of the Young inequality as
where \(h=\frac{b}{a}\), \(h>0\), \(k(h)=\frac{(h+1)^2}{4h}\) is the Kantorovich constant and \(R=\max \{1-\nu , \nu \}\). Note that the function k is decreasing on (0, 1) and increasing on \([1,\infty )\), \(k(h)\ge 1\) and \(k(h)=k\big (\frac{1}{h}\big )\) for any \(h>0\). Recently, Moradi et al. [19] obtained the following refinement of inequality (3.1):
where \(R=\max \{\nu , 1-\nu \}\) and \(D=\max \{a,b\}\). In [15], it has been shown that if \(0 < mI \le A \le MI\) and \(B \in {\mathcal {B}}_{1}(\mathcal {H})\), \(B>0\), then
where \(p,q>1\) with \(\frac{1}{p} + \frac{1}{q} = 1\).
In the following theorem, we present a refinement of inequality (3.3) for positive linear map.
Theorem 3.1
Let \(\Phi :{\mathcal {B}}_1(\mathcal {K})\rightarrow {\mathcal {B}}_1(\mathcal {H})\)be a positive linear mapping satisfying (1.6), whose adjoint is \(\Phi ^*\), \(mI\le A\le MI\)for some constants \(m,\ M\)with \(M>m >0\)and \(B\in {\mathcal {B}}_1(\mathcal {K})\), \(B>0\). Then for any \(p, q >1\)with \(\frac{1}{p}+\frac{1}{q}=1\)we have
Proof
Assume that \(\nu \in (0,1)\), \(a,b\in [m,M]\), \(D=\max \{a,b\}\) and \(R=\max \{\nu ,1-\nu \}\). Since
for each \(0\le \nu \le 1\), we can write
Let \(0<m<M\), then \(\frac{m}{M}<1 <\frac{M}{m}\) and \(\frac{m}{M}\le \frac{a}{b} \le \frac{M}{m}\). Indeed, for \(\frac{m}{M}\le \frac{a}{b}<1\) and \(1<\frac{a}{b}\le \frac{M}{m}\) we have \(k^R\big (\frac{b}{a}\big )=k^R\big (\frac{a}{b}\big )\le k^R\big (\frac{M}{m}\big )\). Thus, for any \(a,b\in [m, M]\), it follows from (3.2) and (3.5) that
According to the hypothesis for \(p>1\) we have \(m^pI\le A^p\le M^pI\). Now, applying functional calculus for \(\nu =\frac{1}{p}\) in (3.6), we obtain
for any \(y\in \mathcal {H}\) and \(t\in [m^p, M^p]\). If we take in (3.7) \(y=\big (\Phi (B)\big )^{\frac{1}{2}}e_i\) and summing over \(i\in I\), then by the properties of trace and equality (1.7), we get
that is, putting \(t=\frac{\mathrm{tr}\big (B\Phi ^*(A^p)\big )}{\mathrm{tr}\big (\Phi (B)\big )}\in [m^p, M^p]\), so we have inequality (3.4). \(\square \)
Remark 3.2
To see that inequality (3.4) refines the inequality (3.3), let \(p=q=2\), we have
On the other hand, since \(\left( \frac{1}{2pq}-\frac{\max \{\frac{1}{p},\frac{1}{q}\}}{4}\right) \le 0\) for \(p,q>2\) we have
In the next theorem, we extend inequalities (1.2) and (1.3) for any arbitrary operator mean \(\sigma _{f}\).
Theorem 3.3
Let \(A, B, C \in {\mathcal {B}}(\mathcal {H})\)such that \(A, B >0\).
-
(i)
If \(A, C \in {\mathcal {B}}_{1}(\mathcal {H})\), then \(C(A \sigma _{f} B), |C^{*}|^{2}A^{\frac{1}{2}}f^{2}(A^{\frac{-1}{2}}BA^{\frac{-1}{2}})A^{\frac{1}{2}} \in {\mathcal {B}}_{1}(\mathcal {H})\)and
$$\begin{aligned} |\mathrm{tr}(C(A \sigma _{f} B))|^{2} \le \mathrm{tr}\big (|C^{*}|^{2}A^{\frac{1}{2}}f^{2}(A^{\frac{-1}{2}}BA^{\frac{-1}{2}})A^{\frac{1}{2}}\big ) \mathrm{tr}(A). \end{aligned}$$(3.8) -
(ii)
If \(A, f(A^{\frac{-1}{2}}BA^{\frac{-1}{2}}) \in {\mathcal {B}}_{1}(\mathcal {H})\), then \(A \sigma _{f} B, A^{\frac{1}{2}}f^{2}(A^{\frac{-1}{2}}BA^{\frac{-1}{2}})A^{\frac{1}{2}} \in {\mathcal {B}}_{1}(\mathcal {H})\)and
$$\begin{aligned} |\mathrm{tr}(A \sigma _{f} B)|^{2} \le \mathrm{tr}(A^{\frac{1}{2}}f^{2}(A^{\frac{-1}{2}}BA^{\frac{-1}{2}})A^{\frac{1}{2}}) \mathrm{tr}(A). \end{aligned}$$
Proof
(i) Let \(\{e_i\}_{i\in I}\) be an orthonormal basis of \(\mathcal {H}\) and F a finite part of I, then
for any \(i\in I\). Using the generalized triangle inequality for the modulus and the Cauchy–Bunyakovsky–Schwarz inequality for finite sums, we have from (3.9) that
for any F a finite part of I.
By the properties of trace and (3.10), we obtain (3.8). \(\square \)
Letting \(f(t) = t^{\frac{1}{2}}\), in Theorem 3.3 we get
Corollary 3.4
Let \(A, B, C \in {\mathcal {B}}(\mathcal {H})\)such that \(A, B >0\). If \(A, C \in {\mathcal {B}}_{1}(\mathcal {H})\), then
In particular, for \(A, B \in {\mathcal {B}}_{1}(\mathcal {H})\)
4 Some examples
We start this section with a well-known theorem.
Theorem 4.1
[6, First Representation Theorem of Kraus] Given an operator \(\Phi :{\mathcal {B}}_1(\mathcal {H})\rightarrow {\mathcal {B}}_1(\mathcal {H})\), there exists a finite or countable family \(\{A_j:j\in J\}\)of bounded linear operators on \(\mathcal {H}\), satisfying
for all finite \(J_0\subset J\), such that for every \(A\in {\mathcal {B}}_1(\mathcal {H})\)and every \(B\in {\mathcal {B}}(\mathcal {H})\)one has
respectively,
and
Conversely, if a countable family \(\{A_j:j\in J\}\)of bounded linear operators on \(\mathcal {H}\)is given which satisfies (4.1) then Eq. (4.2) defines an operation \(\Phi \)whose adjoint \(\Phi ^*\)is given by (4.3) and \(\Phi ^*(I)\)defines by (4.4).
Employing the above theorem and Theorem 2.8 for some examples of convex functions, we can present new versions of trace inequalities.
Example 4.2
Consider the power function \(f:(0,\infty )\rightarrow (0,\infty )\), \(f(t)=t^r\) with \(t\in \mathbb {R}\setminus \{0\}\). For \(r\in (-\infty ,0)\cup [1,\infty )\), f is convex and for \(r\in (0,1)\), f is concave. Let \(r\ge 1\) and A be a self-adjoint operator on the Hilbert space \(\mathcal {H}\) and assume that \(\mathrm{Sp}(A) \subseteq [m,M]\) for some scalars m, M with \(0 \le m < M\). If \(B \in {\mathcal {B}}_1(\mathcal {H})\setminus \{0\}\) is a strictly positive operator and \(\{A_j:j\in J\}\) is a countable family of bounded linear operators on \(\mathcal {H}\) which satisfies (4.1), then by Theorem 2.8 and equality (1.7) respectively, we get
Example 4.3
Consider the convex function \(f:(0,\infty )\rightarrow (0,\infty )\), \(f(t)=-\ln t\) and let A be a self-adjoint operator on the Hilbert space \(\mathcal {H}\) and assume that \(\mathrm{Sp}(A)\subseteq [m,M]\) for some scalars m, M with \(0<m<M\). If \(B\in {\mathcal {B}}_1(\mathcal {H})\setminus \{0\}\) is a strictly positive operator and \(\{A_j:j\in J\}\) is a countable family of bounded linear operators on \(\mathcal {H}\) which satisfies (4.1), then by Theorem 2.8 and equality (1.7), respectively, we get
Example 4.4
Consider the convex function \(f(t)=t\ln t\) and let A be a self-adjoint operator on the Hilbert space \(\mathcal {H}\) and assume that \(\mathrm{Sp}(A)\subseteq [m,M]\) for some m, M with \(0<m<M\). If \(B\in {\mathcal {B}}_1(\mathcal {H})\setminus \{0\}\) is a strictly positive operator and \(\{A_j:j\in J\}\) is a countable family of bounded linear operators on \(\mathcal {H}\) which satisfies (4.1), then by Theorem 2.8 and equality (1.7), respectively, we have
References
Ando, T.: Concavity of certain maps on positive definite matrices and applications to Hadamard products. linear Algebra Appl. 26, 203–241 (1979)
Ando, T., Zhan, X.: Norm inequalities related to operator monoton functions. Math. Ann. 315, 771–780 (1999)
Ando, T., Hiai, F.: Operator log-convex functions and operator means. Math. Ann 350(3), 611–630 (2011)
Bellman, R.: Some inequalities for positive definite matrices. In: Beckenbach, E.F. (ed.) General Inequalities 2, Proceedings of the 2nd International Conference on General Inequalities, pp. 89–90. Birkhauser, Basel (1980)
Bhatia, R.: Positive Definite Matrices. Princeton University Press, Princeton (2007)
Blanchard, P., Brüning, E.: Mathematical Methods in Physics. Springer, Switzerland (2015)
Choi, M.D.: A Schwarz inequality for positive linear maps on \(C^*\)-algebras. Ill. J. Math. 18, 565–574 (1974)
Coop, I.D.: On matrix trace inequalities and related topics for products of Hermitian matrix. J. Math. Anal. Appl. 188, 999–1001 (1994)
Davis, C.: A Schwarz inequality for convex operator functions. Proc. Am. Math. Soc. 8, 42–44 (1957)
Dragomir, S.S.: On some Hölder type trace inequalities for operator weighted geometric mean. RGMIA Res. Rep. Coll. 18, Art. 152 (2015)
Dragomir, S.S.: Jensen’s type trace inequalities for convex functions of selfadjoint operators in Hilbert spaces. Facta Univ. Ser. Math. Inform. 31(5), 981–998 (2016)
Dragomir, S.S.: Reverse Jensen’s type trace inequalities for convex functions of selfadjoint operators in Hilbert spaces. Ann. Math. Sil. 30(1), 39–62 (2016)
Dragomir, S.S.: Some Slater’s type trace inequalities for convex functions of selfadjoint operators in Hilbert spaces. Toyama Math. J. 38, 75–99 (2016)
Dragomir, S.S.: Some trace inequalities for convex functions of selfadjoint operators in Hilbert spaces. Korean J. Math. 24(2), 273–296 (2016)
Dragomir, S.S.: Power and Hölder type trace inequalities for positive operators in Hilbert spaces. Stud. Sci. Math. Hungar. 55(3), 383–406 (2018)
Furuichi, Sh, Moradi, H.R., Sababheh, M.: New sharp inequalities for operator means. Linear Multilinear Algebra 66(1), 1–12 (2018)
Khosravi, M., Moslehian, M.S., Sheikhhosseini, A.: Some operator inequalities involving operator means and positive linear maps. Linear Multilinear Algebra 66(6), 1186–1198 (2018)
Liao, W., Wu, J., Zhao, J.: New versions of reverse Young and Heinz mean inequalities with the Kantorovich constant. Taiwan. J. Math. 19(2), 467–479 (2015)
Moradi, H.R., Furuichi, S., Mitroi-Symeonidis, F.-C., Naseri, R.: An extension of Jensen’s operator inequality and its application to Young inequality. Rev. Real Acad. Cie. Exactas Fís. Nat. Ser. A. Mat. 113, 605–614 (2019)
Neudecker, H.: A matrix trace inequality. J. Math. Anal. Appl. 166, 302–303 (1992)
Ruskai, M.B.R.: Inequalities for traces on Von Neumann algebras. Commun. Math. Phys. 26, 280–289 (1972)
Simon, B.: Trace Ideals and Their Applications. Cambridge University Press, Cambridge (1979)
Stormer, E.: Positive Linear Maps of Operator Algebras. Springer, Berlin (2013)
Yang, Y.: A matrix trace inequality. J. Math. Anal. Appl. 133, 573–574 (1988)
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
About this article
Cite this article
Jafarmanesh, H., Shateri, T.L. & Dragomir, S.S. Choi–Davis–Jensen’s type trace inequalities for convex functions of self-adjoint operators in Hilbert spaces. Bol. Soc. Mat. Mex. 26, 1195–1215 (2020). https://doi.org/10.1007/s40590-020-00300-4
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s40590-020-00300-4