Logarithmic trace inequalities (Role of Operator Inequalities in

Title Operator Theory)

Author(s) Furuta, Takayuki

Citation 数理解析研究所講究録 (2005), 1427: 114-127

Issue Date 2005-04

URL http://hdl.handle.net/2433/47304

Right

Type Departmental Bulletin Paper

Textversion publisher

Kyoto University
数理解析研究所講究録 1427 巻 2005 年 114-127

114

Logarithmic trace inequalities

Takayuki Furuta (Tokyo University of Science)

東京理科大学理学部 古田 孝之

Abstract We shall extend logarithmic trace inequalities shown by Bebiano, Lemos and
Providencia and also by Hiai and Petz, by applying log majorization equivalent to an order
preserving operator inequality. We shall consider the convergence of certain logarithmic
trace inequalities, as some extensions of Bebiano, Lemos and Providencia and Hiai-Petz.
As an appendix, we state the following result. Let and be strictly positive definite $A$ $B$

matrices such that $M_{1}I\geq A\geq m_{1}I>0$ and $M_{2}I\geq B\geq m_{2}I>0$ . Put . $h= \frac{\Lambda I_{1}M_{\mathrm{Z}}}{m1m2}>1$

Then the following inequalities hold’.

$\log S(1)’\mathrm{b}[A]$ $+S(A, B)\geq-\mathrm{T}\mathrm{r}[\hat{S}(A|B)]$ $\geq S(A, B)$ .

where $S(A, B)=\mathrm{T}\mathrm{r}$ [ $A(\log$ A-log $B)$ ], $S\wedge(A|B)=A^{\frac{1}{2}}(\log A^{\frac{-1}{2}}BA^{\frac{-1}{2}})A^{\frac{1}{2}}$

and $S$
(1) $= \frac{h^{\frac{1}{h-1}}}{e\log h^{\frac{1}{h-1}}}$

$(h>1)$ . The first inequality is the reverse one of the well known second one.
\S 1. Introduction
In this paper a capital letter means $n><n$ matrix. Following Ando and Hiai [1], let us
define the $log$ majorization for positive semideflnite matrices , $B\geq 0$ , denoted by $A$
$A\succ B(\log)$

if
$\prod_{i=1}^{k}\lambda_{i}(A)$ $\geq\prod_{i=1}^{k}\lambda_{i}(B)$
, $k=1,2$ , $\ldots$ , $n-$ , $1$

and
$\mathrm{I}\mathrm{I}^{\lambda_{i}(A)=\prod_{i=1}^{n}\lambda_{i}(B)}$ , $\mathrm{i}.\mathrm{e}.$

, $\det A=\det B$ ,

where $\lambda_{1}(A)$ $\geq\lambda_{2}(A)$ $\geq\ldots\geq\lambda_{n}(A)$ and $\lambda_{1}(B)\geq\lambda_{2}(B)\geq\ldots\geq\lambda_{n}(B)$ are the eigenvalues

of and , respectively, arranged in decreasing order. When
$A$ $B$ $0\leq$ a , the a-power
$\leq 1$

mean of positive invertible matrices , $B>0$ is defined by $A$

$A\neq_{\alpha}B=A^{\frac{1}{2}}(A^{\frac{-1}{2}}BA^{\frac{-1}{2}})^{\alpha}A^{\frac{1}{2}}$
.
Further, $\mathrm{A}\#\mathrm{a}\mathrm{B}$
for $A,$ $B\geq 0$ is defined by $\mathrm{A}\#\mathrm{a}\mathrm{B}=\lim_{\epsilon\downarrow 0}(A+\epsilon I)\neq_{\alpha}(B+\epsilon I)$.

For the sake of convenience for symbolic expression, we define $A\# sB$ , for any real number
$s\geq 0$ and for $A>0$ and $B\geq 0$ , by the following

$A\#_{S}B=A^{\frac{1}{2}}(A^{\frac{-1}{2}}BA^{\frac{-1}{2}})^{s}A^{\frac{1}{2}}$
.
$A\mathfrak{h}_{\alpha}B$
in the case $0\leq\alpha\leq 1$ just coincides with the usual a-power mean
 115

The following excellent and useful $\log$ majorization is shown in Ando and Hiai [1, The-
orem 2.1].
Theorem A. For every A, B $\geq 0$
and $0\leq\alpha\leq 1$ ,
(1.1) $(A\#\alpha B)^{r}\succ A^{r}\neq_{\alpha}B^{r}$
for $r\geq 1$ .
(log)

Also, (1.1) can be transformed into the following matrix inequality (1.2) of Theorem
$\mathrm{B}$

in
Ando and Hiai [1, Theorem 3.5]:
Theorem B. if A $\geq B\geq 0$ with A $>0_{J}$ then

(1.2) $A^{r}\geq\{A^{\frac{r}{2}}(A^{\frac{-1}{2}}B^{p}A^{\frac{-1}{2}})^{r}A^{\frac{r}{2}}\}^{\frac{1}{\mathrm{p}}}$

for $r,p\geq 1$ .
We obtained the following extension of Theorem A in Furuta [11, Therorem 2.1] applying
the method in Ando and Hiai [1] to Theorem (see Q3). $\mathrm{G}$

Theorem C. For every A $>0$ , B , and for each t $\in[0,$1], $\geq 0$ $0\leq\alpha\leq 1$

$(A\neq_{\alpha}B)^{h}\succ A^{1-t+r}\neq_{\beta}(A^{1-t}\mathfrak{h}_{s}B)$

(log)

holds for $s\geq 1_{f}$ and $r\geq t\geq 0$ , where $\beta=\frac{\alpha(1-t+r)}{(1-\alpha t)s+\alpha r}$ and $h= \frac{(1-t+r)s}{(1-\alpha t)s+\alpha r}$ .

Next, we state the following result which is shown in Hiai and Petz [13, Theorem 3.5]
and, recently, a new proof is given in Bebiano, Lemos and Providencia [2, Theorem 2.2].
Theorem D. If A, B $\geq 0$ , then for every p $\geq 0$

(1.3) $\frac{1}{p}$

Tr[A $\log(A^{\epsilon B}\mathrm{z}B^{p}A2)$

] $\geq$
Tr $[A(\log A+\log B)]$

left hand side of (1.3) converges the right hand side

hola and the
$s$
as $p\downarrow \mathrm{O}$
.

Theorem E. if A , B $>0$ , and p $>0$ , then

$\geq 0$ $0\leq\alpha\leq 1$

(1.4) $\frac{\mathit{1}}{p}\mathrm{T}\mathrm{r}[A\log(A^{p}\#\alpha B^{\mathrm{p}})]+\frac{\alpha}{p}\mathrm{T}\mathrm{r}[A\log(A\%_{B^{-p}A^{2})]\geq}tt$ Tr $[A\log A]$

holds and the left hand side of (1.4) converges the the right hand side as $p\downarrow \mathrm{O}$
.
The inequality (1.4) is shown in Ando and Hiai [1, Theorem 5.3], and the convergence
of (1.4) is shown in Bebiano, Lemos and Providencia [2, Corollary 2.2].
We shall extend Theorem and Theorem by applying the inequality derived
$\mathrm{D}$
$\mathrm{E}$
$\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{e}$

from majorization equivalent to an order preserving inequality, and also by applying

$\log$

the generalized Lie-Trotter formulae of Lemma 6.1 in \S 6 and Lemm a 7.1 in \S 7.

52, majorization equivalent to an order preserving operator inequality
${\rm Log}$

We shall show a $\log$ majorization equivalent to an order preserving operator inequality.

Theorem 2.1. The following (i) and (ii) hold and are equivalent
118

(i) if $A$
, $B\geq 0$ , then for each $t\in[0,1]$ and $r\geq t$

$A^{\frac{1}{2}}(A^{\frac{r-t}{2}}B^{p}A^{\frac{r-t}{2}})^{\mathrm{A}}pA^{\frac{1}{2}}\succ A^{\frac{(p-tq)s+rq}{2\mathrm{p}s}}\{B^{E}2(B^{2}2A^{r}B^{E}2)^{s-1}B\mathrm{z}\}\epsilon p\overline{s}s_{A^{\frac{(p-tq)s+\tau q}{2ps}}}$

(log)

holds for any $s\geq 1$ and $p\geq q>0$ .

(ii) if $A\geq B\geq 0$ with $A>0$ , then for each $t\in[0,1]$ and $r\geq t$
$A^{\frac{\{p-tq\}s+rq}{\mathrm{p}\epsilon}}\geq\{A^{\frac{r}{2}(q}A^{\frac{-t}{2}}B^{I\mathrm{i}}A^{\frac{-t}{2}})^{s}A^{\frac{r}{2}}\}^{ps}\neq$
.

holds for any $s\geq 1$ and $p\geq q>0$ .

Corollary 2.2. The following (i) and (ii) holl and are equivalent:
(i) if $A$
, $B\geq 0$ , then for each $r\geq 0$

$A^{\frac{1}{2}}(A^{\frac{r}{2}}B^{p}A^{\frac{r}{2}})^{\mathrm{z}}\mathrm{p}A^{\frac{1}{2}}\succ A^{\frac{1}{2}(1+\frac{r}{p}q)}B^{q}A^{\frac{1}{2}(1+\frac{r}{p}q)}$

(log)

holds for any $p\geq q>0$ .

(ii) ij $A\geq B\geq 0$ , then for each $r\geq 0$

$A^{1+\frac{\Gamma}{p}q}\geq(A^{\frac{r}{2}}B^{R}qA^{\frac{r}{2}})^{\mathrm{a}}p$

holds for any $p\geq q>0$ .

Corollary 2.3. The following (i) and (ii) hold and are equivalent:
(i) If $A$
, $B\geq 0$ , then for each $r\geq 1$

$A^{\frac{1}{2}}(A^{\frac{r-1}{2}}BA^{\frac{r-1}{2}})^{q}A^{\frac{1}{2}}\succ A^{\frac{1}{2}}\{B^{\frac{1}{2}}(B^{\frac{1}{2}}A^{r}B^{\frac{1}{2}})^{r-1}B^{\frac{1}{2}}\}^{\mathrm{a}}rA^{\frac{1}{2}}$

(log)

holds for any $1\geq q>0$ .

(ii) If $A\geq B\geq 0$ with $A>0$ , then for each $r\geq 1$

$A\geq\{A^{\frac{r}{2}}(A^{\frac{-1}{2}}B^{\frac{1}{q}}A^{\frac{-1}{2}})^{r}A^{\frac{r}{2}}\}^{q}r$
.
holds for any $1\geq q>0$ .

Q3. Results needed to give proofs of the results in \S 2

Throughout this section, a capital letter means a bounded linear operator on a complex
Hilbert space $H$ . An operator is said to be positive (denoted by $T\geq 0$ ) if $(Tx, x)$ for $T$

all $x\in H$ . Also, an operator is strictly positive (denoted by $T>0$ ) if is positive and $T$ $T$

invertible. We state the following celebrated L\"owner-Heinz inequality in operator theory.

Theorem L-H (L\"owner-Heinz inequality).

If $A\geq B\geq 0$ , then for any $ce\in[0, 1]$ . $A^{\alpha}\geq B^{\alpha}$

Lemma A. [11, Lemma 1], Let A $>0$ and also let B be an invertible operator. Then
$(BAB^{*})^{\lambda}=BA^{\frac{1}{2}}(A^{\frac{1}{2}}B^{*}BA^{\frac{1}{2}})^{\lambda-1}A^{\frac{1}{2}}B^{*}$
 117

holds for any real number A.

Theorem $\mathrm{F}$

(Furuta inequality).

If $A\geq B\geq 0$ , then for each $r\geq 0$ ,

(i) $(B^{\frac{\Gamma}{2}}A^{p}B^{\frac{r}{2}})^{\frac{1}{q}}\geq(B^{\frac{f}{2}}B^{p}B^{\frac{f}{2}})^{\frac{1}{q}}$

and

(ii) $(A^{\frac{f}{2}}A^{p}A^{\frac{f}{2}})^{\frac{1}{q}}\geq(A^{\frac{f}{2}}B^{p}A^{\frac{r}{2}})^{\frac{1}{q}}$

hold for $p\geq 0$ and $q\geq 1$ with $(1+r)q\geq p+r$ .

It is shown in Tanahashi [16], that the domain drawn for $p,g$ and in Figure 1 is the $r$

best possible one for Theorem F. Theorem yields L\"owner-Heinz inequality asserting that $\mathrm{F}$

$A\geq B\geq 0$ ensures for any when we put $r=0$ in (i) or (ii) of Theorem
$A^{\alpha}\geq B^{\alpha}$ $\alpha$ $\in[0, 1]_{;}$

F. The original proof is in Furuta [9], alternative proofs can be found in Fujii [4], Kamei
[14] and one page proof in Furuta [10].

As an extension of Theorem , we obtain the following Theorem which interpolates

$\mathrm{G}$
$\mathrm{F}$

Theorem and Theorem B. Theorem is used to prove Theorem C.

$\mathrm{F}$
$\mathrm{G}$

Theorem G. If A $\geq B\geq 0$ with A $>0$ , then for each t $\in[0,$ 1] and p $\geq 1$

$A^{1-t+r}=\{A^{\frac{r}{2}}(A^{\frac{-t}{2}}A^{p}A^{\frac{-t}{2}})^{s}A^{\frac{r}{2}}\}^{\frac{1-t+r}{(\mathrm{p}-t)s+r}}$

$\geq\{A^{\frac{r}{2}}(A^{\frac{-t}{2}}B^{p}A^{\frac{-t}{2}})^{s}A^{\frac{f}{2}}\}^{\frac{1-l+\underline{r}}{\{p-t\rangle s+r}}$

for any $s\geq 1$ and $r\geq t$ .

The original proof of Theorem is in Furuta [11, Theorem 1.1], alternative proofs can $\mathrm{G}$

be found in Fujii and Kamei [5] and one page proof in Furuta [12]. The original proof
$1-t$ $+r$
of the best possible exponent $\overline{(p-t)s+r}$ in Theorem is obtained in Tanahashi [17], $\mathrm{G}$

and alternative proofs can be found in M.Fujii, Matsumoto and Nakamoto [6], and also in
Yamazaki [19].

Q4 Proofs of the results in \S 2

Applying Theorem and [Theorem 2.1, Ando-Hiai [1]], we can give a proof of Theorem
$\mathrm{G}$

2.1 and we omit it. Corollary 2.2 and Corollary 2.3 are immediate consequence of Theorem
2.1,
118

\S 5. Logarithmic trace inequalities as an application of Theorem 2.1

For $A$
, $B>0$ , the relative operator entropy $\hat{S}(A|B)$ is defined by
$\hat{S}(A|B)=A^{\frac{1}{2}}\log(A^{\frac{-1}{2}}BA^{\frac{-1}{2}})A^{\frac{1}{2}}$

in i.Fujii and Kamei [3], and $\hat{S}(A|I)=$ -Alog is the usual operator entropy, (see [15]).
$\mathrm{J}$
$A$

The Umegaki operator entropy $S(A, B)$ is defined by

$S(A, B)=$ Tr [ $A(\log$ A-log $B)$ ]

(see Umegaki [18]). For $A$

, $B>0$ , let $\triangle(A|B)$ are defined by
$\Delta(A|B)=-\mathrm{h}[\hat{S}(A|B)]-S(A, B)$ .

We shall discuss the low er bound of $\Delta(A|B)$ in terms of the of and and a $\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{e}$ $A$ $B$

parameter, and this result implies the well known inequality $\triangle(A|B)\geq 0$ (for example,
[13],[2] . $)$

Theorem 5.1. if A, B $\geq 0$ , then, for every t $\in[0,$ 1] and p $\geq 0$ ,

(5.1) Tr[A $\log(A^{\frac{r-t}{2}}B^{p}A^{\frac{r-t}{2}})^{s}$

]
$\geq(r-ts)\mathrm{T}\mathrm{r}[A \log A]$ $+\mathrm{T}\mathrm{r}[A\log\{B^{R}2(B^{\mathrm{E}}2A^{r}B^{\epsilon}2)^{s-1}B^{\epsilon}2\}]$

holds for any $r\geq f$ and $s\geq 1$ .

Corollary 5.2. If A, B $\geq 0$ , then, for every p $\geq 0$
and r $\geq 0$ ,

(5.2) Tr $[A\log(A^{\frac{r}{2}}B^{p}A^{\frac{r}{2}})^{s}]\geq$
Tr[Alog $A^{r}$
] $+\mathrm{T}\mathrm{r}[A\log\{B^{E}2(B^{\epsilon}2A^{r}B^{E}2)^{s-1}B^{R}2\}]$

holds for any $s\geq 1$ . In particular,

(5.3) Tr $[A \log(A^{\frac{r}{2}}B^{p}A^{\frac{r}{2}})]$ $\geq$
Tr $[A \log A^{r}+A\log B^{p}]$

and
(5.4) Tr $[A \log(A^{\frac{r}{2}}B^{p}A^{\frac{r}{2}})^{2}]$ $\geq$
Tr $[A \log A^{r}]$ $+\mathrm{T}\mathrm{r}[A \log(B^{p}A^{r}B^{p})]$ .

The inequality (5.3) of Corollary 5.2 may be considered as the two variable version of
(1.3) in Theorem D. In fact, (5.3) of Corollary 5.2 is equivalent to (1.3) in Theorem (see $\mathrm{D}$

Remark 5,1).

Corollary 5.3. If A, B $\geq 0$ , then

(5.5) $\mathrm{b}[A\log(A^{\frac{r-1}{2}}BA^{\frac{r-1}{2}})^{r}]\geq \mathrm{B}[A\log\{B^{\frac{1}{2}}(B^{\frac{1}{2}}A^{r}B^{\frac{1}{2}})^{r-1}B^{\frac{1}{2}}\}]$

 118

holds for every real number $r\geq 1$ . In particular,

(5.6) Tr[Alog $(A^{\frac{1}{2}}BA^{\frac{1}{2}})^{2}$
] $\geq \mathrm{R}[A \log(BA^{2}B)]$ .

Corollary 5.4. if A, B $>0$ , then

(5.7) $\Delta(A|B)\geq\prime \mathrm{b}[A\log B]-\frac{2(s-1)}{s}\mathrm{T}\mathrm{r}[A\log A]+\frac{1}{s}$

Tr $[A \log\{A^{-1}(AB^{-1}A)^{s}A^{-1}\}]$

holds for every real number $s\geq 1$ . In particular, $\triangle(A|B)\geq 0$ holds.

We remark that the right hand side of (5.7) is zero when $s=1$ , or when $A$
commutes with
.
$B$

Proof of Theorem 5.1 is obtained by Theorem 2.1 and we omit it and Corollaries in this
section are shown by Theorem 5.1.

\S 6 Generalized Lie-Trotter formulae, I

We adopt the usual convention $X^{0}=I$ for $X>0$ . We obtain a convenient generalization
of the Lie-Totter formulae to prove the results in 98. The famous Lie-Trotter formula states

$e^{A+B}= \lim_{p\downarrow 0}(e^{\mathrm{E}_{\frac{A}{2}}}e^{pB}e^{l\mathrm{i}^{\underline{A}}}2)^{\frac{1}{p}}$

for any Hermitian $A$
and $B$ .

The following lemma is an -mean variant of the Lie-Trotter formula. $\alpha$

Lemma H[13, Lemma 3.3]. ij A and B crre Hemitian and $\alpha\in[0,$ 1], then
$e^{(1-\alpha)A+\alpha B}= \lim_{p\downarrow 0}(e^{pA}\beta_{\alpha}e^{pB})^{\frac{1}{p}}$ .

We remark that the Lie-Trotter formula and the a-mean variant of the Lie-Trotter for-
mula are both quite useful in operator theory.

By retracing the proof of Lemma $\mathrm{H}$

, we shall obtain the following lemma.

, $B$ , $C$ and are Hermitian, then, for any positive numbers and
Lemma 6.1. if
$\beta$
$D$
$\alpha$

$A$

(6.1)
$e^{A+\alpha B+\alpha\beta(C+D)}= \lim_{p\downarrow 0}\{e^{L_{-}^{A}}2(e^{L_{-}^{B}}2(e^{L_{-}^{C}}2e^{pD}e^{L^{C}}2)^{\beta}e^{\mathrm{a}_{2}^{\underline{B}}})^{\alpha}e^{\mathrm{L}_{-}^{A}}2\}^{\frac{1}{\mathrm{p}}}.$
,

in particular,

(6.2) $e^{A+\alpha(B+C)}= \lim_{p\downarrow 0}\{e^{L_{-}^{A}}2(e^{\mathrm{a}_{2}^{\underline{B}}}e^{pC}e^{L^{\underline{B}}}2)^{\alpha}e^{E^{\underline{A}}}2\}^{\frac{1}{p}}$ .

To prove the results in \S 8, we rewrite Lemma 6.1 in the following convenient form.
Lemma 6.1’. if A, B, C and D are positive, then, for any positive numbers $\alpha$
and , $\beta$
120

(6.1’) $\log A+\alpha\log B+\alpha\beta(\log C+\log D)=\lim_{p\downarrow 0}\log\{A^{2}2(B2$

$(C2\epsilon\epsilon D^{p}C^{\epsilon}2)^{\beta}B2)^{\alpha}RA^{\epsilon}2\}^{\frac{1}{\mathrm{p}}}$

In particular,

(6.2’) $\log A+a(\log B+\log C)$ $=1\mathrm{i}\mathrm{r}\mathrm{n}\log\{A^{2}2(B^{q}p\downarrow 02C^{p}B^{\mathrm{E}}2)^{\alpha}A^{B}2\}^{\frac{1}{p}}$ .

Following analogous steps to those in the proof of Lemma 6.1, we can easily prove (6.7).

Q7 Generalized Lie-Trotter formula, II

In this section, we present generalizations of the Lie-botter formulae different from those
in Lemma 6.1 in Q6 in order to prove the results in Q8.

Lemma 7.1. if A, B and C are Hermitian, then, for any $\alpha$

$\in[0,$ 1] and r $\geq 0$ ,

(7.1) $e^{r\langle 1-\alpha)A+r\alpha B+C}= \lim_{p\downarrow 0}(e^{\epsilon_{\frac{c}{2}}}((1-\alpha)e^{pA}+\alpha e^{pB})^{r}e^{L^{\underline{C}}}2)^{\frac{1}{p}}$

In particular,

(7.2) $e^{(1-\alpha)A+\alpha B}= \lim_{p\downarrow 0}((1-\alpha)e^{pA}+ae^{pB})^{\frac{1}{p}}$ .

Lemma 7.1 can be rewritten as follows.

Lemma 7.1. If $A$

, $B$
and $C$
are positive definite, then, for any a $\in[0,1]$ and $r\geq 0$ ,

(7. 1 ,) $e^{r\langle 1-\alpha)\log A+r\alpha\log B+\log C}= \lim_{p\downarrow 0}(C^{R}\mathit{2}$

$((1-\alpha)A^{p}+\alpha B^{p})^{r}C^{\epsilon}2)^{\frac{1}{p}}$

In particular,

(72’) $e^{(1-\alpha\}\log A+\alpha\log B}= \lim_{p\downarrow 0}((1-a)A^{p}+aB^{p})^{\frac{1}{p}}$ .

Next, we shall state an application of Lemma 7.1’. M.Fujii and R.Nakamoto [7] defined
the chaotically -geometric mean which is different from the usual a-geometric
$\alpha$ $A\phi_{\alpha}B$

mean $A\#\alpha B$
:

A $\phi_{\alpha}B=e^{(1-a)\log A+\alpha\log B}$

, for $A$
, $B>0$ and $\alpha\in[0,1]$ .

Among others, M.Fujii and R.Nakamoto [7] proved the following result.

Theorem I. If A and B are strictly positive operators on a Hilbert space and a $\in[0,$ 1],
then and strongly converge to the chaotically a-geometric
$\frac{1}{p}$

$(A^{p}\nabla_{\alpha}B^{p})^{\frac{1}{p}},(A^{p}\#\alpha B^{p})^{\frac{1}{p}}$
$(A^{p}!_{\alpha}B^{p})$

mean as , where SVaT $=(1-\alpha)S+aT$ cvnd $S!_{\alpha}T=((1-\alpha)S^{-1}+\alpha T^{-1})^{-1}$

$A\phi_{\alpha}B$ $p\downarrow \mathrm{O}$

for strictly positive operators and . $S$ $T$

 121

Two proofs to Theorem I are given in (Theorem 4, [7]) and (\S 4, [8]). We shall extend
Theorem I as an application of Lemma 7.1, that is, we shall show that:
The chaotically -geometric mean $a$ $A\phi_{\alpha}B$
is the uniform limit of $(A^{p}\nabla_{\alpha}B^{p})^{\frac{1}{p}}$
.
;,
Proposition 7.2. If A and B are Hermitian and a $\in[0,$ 1], then $(e^{pA}\#\alpha e^{pB})^{\frac{1}{p}}$
$(e^{pA}\nabla_{\alpha}e^{pB})$

and uniformly converge to as

$(e^{pA}!_{\alpha}e^{pB})^{\frac{1}{\mathrm{p}}}$ $e^{A}\phi_{\alpha}e^{B}$ $p\downarrow \mathrm{O}$
.
Proposition 7.2 can be rewritten as follow . $\mathrm{s}$

Proposition 7.2’. ij A and B are positive definite and a $\in[0,$ 1], then
$(A^{p} \nabla_{\alpha}B^{p})^{\frac{1}{p}},(A^{p}\oint_{\alpha}B^{p})^{\frac{1}{p}}$

and ; uniformly converge to the chaotically -geometric mean

$(A^{p}!_{\alpha}B^{p})$
$\alpha$
$A\phi_{\alpha}B$
as p $\downarrow \mathrm{O}$

We remark that Proposition 7.2’ remains valid for Hilbert space operators because
Lemma 6.1 still remains valid for operators, so that Proposition 7.2’ may be considered to
be a strong version of Theorem I.
Q8. Convergence of logarithmic trace inequalities via generalized Lie- rotter
formulae
In this section, We shall discuss the convergence of the logarithmic inequalities $\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{e}$

obtained in \S 5 by applying generalized Lie-Trotter formulae of Lemma 6.1’ in \S 6 and the

Lemma 7.1’ in \S 7.
Theorem 8.1. if A, B $\geq 0$ , then, for every p $\geq 0$ ,

(8.1) $\frac{s}{p}$
Tr[A tog $(A^{\mathrm{E}}2B^{p}A^{\mathrm{E}}2)$
] $- \frac{1}{p}\mathrm{T}\tau$
[ $A\log\{B^{R}2$ (B2 $A^{p}B^{\epsilon\epsilon}\dot{\mathrm{Z}})^{s-1}B2\}$
]

$\geq \mathrm{h}[A\log A]$

holds for any $p\geq 0$ and $s\geq 1$ , and the left hand side converges to the right hand side as
$p\downarrow \mathrm{O}$
.
Theorem 8.1 yields the following Corollary 8.2.
Corollary 8.2.

(i) if $A$
, $B\geq 0$ , then, for every $p\geq 0$ ,

(8.2) $\frac{1}{p}$

Tr[A $\log(A^{2}2B^{p}A^{R}2)$ ] $\geq$

Tr[A $\log A+A\log B$ ]

holds and the left hand side converges to the right hand side as ,
$p\downarrow \mathrm{O}$

(ii) If $A$
, $B\geq 0$ , then, for every $p\geq 0$ ,

(8.3) $\frac{2}{p}\mathrm{h}[A \log(A^{\mathrm{E}}2B^{p}A^{\mathrm{E}}2)]$ $- \frac{1}{p}\mathrm{b}[A \log(B^{p}A^{p}B^{p})]$

122

$\geq \mathrm{R}$
[ $A\log$ A3]

holds and the left hand side converges to the right hand side as $p\downarrow \mathrm{O}$
.

We remark that (i) of Corollary 8.2 is Theorem D.

Theorem 8.3. ij A $>0$ and B $\geq 0$ , then, for every positive number $\beta$
,

(8.4) $\frac{s}{p}\mathrm{T}\mathrm{r}[A \log(A^{p}\mathfrak{h}_{\beta}B^{p})]$ $- \frac{1}{p}\mathrm{b}[A \log\{A^{-B^{-}}A^{-}\mathrm{z}(A^{p}\mathfrak{h}_{\beta}B^{p})^{s}A^{-_{2}I}\}]$

$\geq \mathrm{b}[A\log A]$

holds for any $p\geq 0$ , $s\geq 1$ , and the left hand side converges to the right hand side as $p\downarrow \mathrm{O}$
.
Theorem 8.3 implies the following Corollary 8.4.

Corollary 8.4.

(i) If $A_{7}B>0$ , then, for every positive number , $\beta$

(S.5) $p\underline{1}$

Tr[A $\log(A^{p}\mathfrak{h}_{\beta}B^{p})$
] $+ \frac{\beta}{p}$
Tr[A $\log(A^{R}2B^{-p}A^{R}2)$ ]

$\geq$
Tr $[A \log A]$

holds for any $p\geq 0_{f}$ and the left hand side converges to the right hand side as $p\downarrow 0$
.

(ii) if $A$
, $B>0$ , then, for every positive rvurnber $\beta_{\lambda}$

(8.6) $\frac{2}{p}\mathrm{R}[A \log(A^{p}\#\beta B^{p})]$ $- \frac{1}{p}\mathrm{T}\mathrm{r}[A \log(A^{-R^{-B}}-_{2}B^{p}A^{-}2)^{\beta}A^{p}(A^{-}2B^{p}A^{-}2)^{\beta}]-B^{-B}$

$\geq \mathrm{T}\mathrm{r}[A\log A]$

holds for any $p\geq 0$ and the left hand side converges to the right hand side as $p\downarrow \mathrm{O}$
.

We remark that, when $A\geq 0_{?}B>0$ and $\beta\in[0,1]$ , (i) of Corollary 8.4 becomes Theorem
E.

Theorem 8.5. If A $>0$ and B $\geq 0$ , then for every $\alpha\in[0,$ 1]

(8.7) $\frac{s}{p}$
Tr[A $\log(A^{p}\nabla_{\alpha}B^{p})$
] $- \frac{1}{p}i\mathrm{R}[A \log\{A^{\cdot}2(A^{p}\nabla_{\alpha}B^{\mathrm{p}})^{s}\lrcorner \mathrm{i}^{-B}A^{-}-2 \}]$

$\geq \mathrm{b}[A\log A]$

holds for any $p\geq 0$ , $s\geq 1$ , and the left hand side converges to the right hand side as $p\downarrow \mathrm{O}$
.

Theorem 8.5 implies the following Corollary 8.6,

Corollary 8.6. If A $>0$ and B $\geq 0$ , then

 123

(S.8) $p\underline{1}$

Tr[A $\log((1-a)A^{p}+\alpha B^{p})$ ] $- \frac{1}{p}\mathrm{b}[A \log\{(1-a)I+A^{-}2B^{p}A^{-\lrcorner \mathrm{i}}-B^{-}2 \}]$

$\geq$
Tr [A $\log$ $A$ ]

holds for any $p\geq 0$ , a $\in[0,1]$ , and the left hand side converges to the right hand side as
. Moreover,
$p\downarrow \mathrm{O}$

(8.9) $\frac{s}{p}\mathrm{T}\mathrm{r}[A \log\frac{A^{p}+B^{p}}{2}]$ $- \frac{1}{p}\mathrm{T}\mathrm{r}[A \log\{A^{-}2(-B\frac{A^{p}+B^{p}}{2})^{s-s}A^{-}2\}]$

$\geq \mathrm{T}\mathrm{r}[A\log A]$

holds for any $p\geq 0_{f}s\geq 1$ , and the left hand sile converges to the right hand side as $p\downarrow \mathrm{O}$
.

\S 9 Proofs of the results in \S 8

The results in \S 8 are show $\mathrm{n}$

by Theorem 5.1 and Lemma 6.1’ and Lemma 7.1’ and we

omit them.

References

[1] T.Ando and F.Hiai, Log majorization and complementary Golden-Thompson type in-
equealities, Linear Alg. and Its AppL,197, 198 (1994),113-131.

[2] N.Bebiano, R.Lemos and J. da Providencia, Inequalities for quantum relative entropy,
preprint.

[3] J.I.Fujii and E.Kamei, Relative operator entropy in noncommutative information theory,
Math. Japon, 34(1989), 341-348.
[4] M. Fujii, Furuta’s inequality and its mean theoretic approach, J. Operator Theory,
23 (1990), 67-72.
[5] M. Fujii and E.Kamei, Mean theoretic approach to the grand Furuta inequality, Proa
Amer. Math. Soc, 124(1996), 2751-2756.

[6] M. Fujii, A.Matsumoto and R.Nakamoto, A short proof of the best possibility for the
grand Furuta inequality, J. of InequaL and AppL, $4(1999)$ , 339-344.
[7] M. Fujii and R.Nakamoto, A geometric mean in the Furuta inequality, Scientiae Math-
ematicae Japonicae Online, $5(2001)$ , 435-441.
[8] M. Fujii, Y. Seo and M. Tominaga, Golden-Thompson type inequalities related to a
geometric mean via Specht’s ratio, Math. Ineq. Appl., $5(2002)$ , 573-582
124

[9] T.Furuta, A $\geq B\geq 0$ assures for r $\geq 0,p\geq 0$ , q

$(B^{r}A^{p}B^{r})^{1/q}\geq B^{(p+2r)/q}$ $\geq 1$ with
$(1+2r)q\geq p+2r$ , Proc. Amer. Math. Soc.,101 (1987),85-88.
.
.Furuta, Elementary proof of an order preserving inequality, Proc. Japan
$\mathrm{A}\mathrm{c}\mathrm{a}\mathrm{d}.,65(1989),126$
$[10]\mathrm{T}$

[11] T,Furuta, An extension of the Furuta inequality and Ando-Hiai

$\log$ majorization,
Linear Alg. and Its App1.,219 (1995), 139-155.
[12] T.Furuta, Simplified proof of an order preserving operator inequality, Proc, Japan
Acad., 74(1998), 114.

[13] Hiai and Petz, The Golden-Thompson trace inequality is complemented, Linear Alg.
and Its App1.,181 (1993), 153-185.
[14] E.Kamei, A sattelite to Furuta’s inequality, Math. Japon., 33(1988), 883-886.
[15] M.Nakamura and H.Umegaki, A note on entropy for operator algebras, Proc. Japan
Acad., 37(1961), 149-154.

[16] K.Tanahashi, Best possibility of the Furuta inequality, Proc. Amer. Math. Soc,
124(1996) )141-146.

[17] K.Tanahashi, The best possibility of the grand Furuta inequality,Proc. Amer. Math.
Soc, 128(2000) )511-519.

[18] H.Umegaki, Conditional expectation in an operator algebra IV, Kodai Math. Sem.
Rep., 14(1962), 59-85.
[19] T.Yamazaki, Simplified proof of Tanahashi ’s result on the best possibility of generalized
Furuta inequality, Math. Inequal. Math., $2(1999),$ , 473-477.

This paper except Appendix is early announcement of the following paper:

T.Furuta, Convergence of logarithmic inequalities via generalized Lie-Trotter for
$\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{e}$

mulae, to appear in LAA

 125

\S 10 Appendix.
Inequalities associated with Umegaki relative entropy $S(A, B)=\mathrm{b}$ $A\log$ A–Alog ] [ $B$

and the relative operator entropy $\hat{S}(A|B)=A^{\frac{1}{2}}(\log A^{\frac{-1}{2}}BA^{\frac{-1}{2}})A^{\frac{1}{2}}$

by J.LFujii and
E.Kamei

On December 2, 2004 I have spoken this appendix in my talk at the Mathematics Re-
search Insitute of Kyoto University.

A capital letter means complex matrix and Tr[X] means the

$n\mathrm{x}$ $n$ on the matrix $\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{e}$

$X$ . A matrix $X$ is said to be strictly positive definite if $X$ is positive definite and invertible

(denoted by $X>0$ ). Let and be strictly positive definite matrices. Umegaki relative
$A$ $B$

entropy $S(A, B)$ in [8] is defined by

(1.1) $S(A, B)=\mathrm{T}\mathrm{r}[A(\log A-\log B)]$

and the relative operator entropy $\hat{S}(A|B)$ in [3] is defined by

(1.2) $\hat{S}(A|B)=A^{\frac{1}{2}}(\log A^{\frac{-1}{2}}BA^{\frac{-1}{2}})A^{\frac{1}{2}}$

as an extension of [7]. Let $h>1$ . $S(p)$ is defined by

(1.3) $S(p)= \frac{h^{\frac{p}{h\mathrm{p}-1}}}{e\log h^{B}\overline{h}\mathrm{P}-\overline{1}}$

for any real number . In particular

$p$
$S(1)= \frac{h^{\frac{1}{h-1}}}{e\log h^{\frac{1}{h-1}}}$
is said to be the Specht ratio and

$S(1)>1$ is well known. We shall show the following inequalities associated with $S(A, B)$

and -Tr $[\hat{S}(A|B)]$ .

Theorem 1. Let and $A$ $B$ be strictly positive definite matrices such that $M_{1}I\geq A$ $\geq$

$m_{1}I>0$ and $M_{2}I\geq B\geq m_{2}I>0$ . Put $h= \frac{M_{1}M_{2}}{m_{1}m_{2}}>1$ . Then the following inequalities
hold:
(1.4) $\log S(1)^{r}\mathrm{b}[A]+S(A, B)$

$\geq\log S(1)\mathrm{T}\mathrm{r}[A]+$ Tr [ $A(\log$ Tr $[A]-\log \mathrm{i}\mathrm{b}[B])$ ]

$\geq-\mathrm{R}[\hat{S}(A|B)]$

$\geq S(A, B)$

in particular,
(1.5) $\log S(1)\mathrm{h}[A]+S(A, B)\geq-\mathrm{R}[\hat{S}(A|B)]\geq S(A, B)$ .

The first inequality of (1.5) is the reverse one of the second inequality which is well known
in $[5],[6]$ and [1]. We prepare the following results to prove Theorem 1.
128

Proposition 2. Let and be strictly positive $A$ $B$

definite matrices such that $M_{1}I\geq A\geq$

$m_{1}I>0$ and $M_{2}I\geq B\geq m_{2}I>0$ . Put $h= \frac{M_{1}M_{2}}{m_{1}m_{2}}>1$ . Let $\varphi$
be a normalized positive
linear functional on $M_{n}(C)$ . Then
(1.6) $\log S(1)\varphi(A)+\varphi(\hat{S}(A|B))\geq\varphi(A)(\log\varphi(B)-\log\varphi(A))\geq\varphi(\hat{S}(A|B))$ .
Proof. Let and be two matrices satisfying the hypotheses in Proposition 2. By (iii)
$A$ $B$

of of [Theorem 2.1, [4]], if I is a normalized positive linear map from $M_{n}(C)$ into itself,
then

(1.7) $\log S(1)\Phi(A)+\Phi(\hat{S}(A|B))\geq\hat{S}(\Phi(A)|\Phi(B))\geq\Phi$ (S $(A|B)$ )

and (1.6) follows from (1.7) since $\varphi$
be a normalized positive linear functional on $M_{n}(C)$

and
$\hat{S}(\varphi(A)|\varphi(B))=\varphi(A)(\log\varphi(B)-\log\varphi(A)).\square$

Proposition 3 (Peierls-Bogoliubov inequality). The following (i) and (ii) hold and are
equivalent:

(i) $\mathrm{T}\mathrm{r}e^{A+B}\geq \mathrm{b}e^{A}\exp(\frac{\mathrm{T}\mathrm{r}e^{A}B}{\mathrm{h}e^{A}})$

for Hermitian $A$

and $B$

and
(ii) $S(A, B)\geq \mathrm{H}$ $A(\log$ [ Tr[A] - $\log$ Th[B])] for $A>0$ and $B>0$ .
Peierls-Bogoliubov inequality is well known in statistic dynamics and the equivalence
relation between (i) and (ii) is stated in ([6] and [2]).
Proposition 4 ( $[5],[6]$ and [1]) The following inequality holds:
(1.8) $-\mathrm{T}\mathrm{r}[\hat{S}(A|B)]\geq S(A, B)$
for $A>0$ and $B>0$ .
Proof of Theorem 1. Let and be two matrices satisfying the hypotheses in $A$ $B$

Theorem 1 and recall that these hypotheses are the same as ones in Proposition 2. Put
in Proposition 2. Then the first inequality of (1.6) implies
$\varphi(X)=\frac{1}{n}\mathrm{b}[X]$

(1.9) Jog $S(1)\mathrm{T}\mathrm{r}[A]+\mathrm{T}\mathrm{r}$

[ $A(\log$ Tr $[A]$ $-\log$ Tr $[B])$ ] $\geq-\mathrm{T}\mathrm{r}[\hat{S}(A|B)]$
.
Therefore we have
(1.4) $\log S(1)\mathrm{T}\mathrm{r}[A]+S(A, B)$

$\geq\log S(1)\mathrm{R}[A]+\mathrm{T}\mathrm{r}$
[ $A(\log$ Tr $[A]$ $-\log$ Tr $[B])$ ] by (ii) of Proposition 3
$\geq-\mathrm{b}[\hat{S}(A|B)]$
by (1.9)
$\geq S(A, B)$ by (1.8) of Proposition 4
so the proof is complete since (1.5) is an immediate consequence of (1.4). $[]$
 127

References
[1] N.Bebiano, R.Lemos and J. da Providencia, inequalities for quantum relative entropy,
to appear in Linear Alg. and Its AppL
[2] N. Bebiano, R. Lemos and J. da Providencia, Matrix inequalities in statistical mechanics,
Linear Alg. and Its App1.,376 (2004), 265-273.
[3] J.Pujii and E.Kamei, Relative operator entropy in noncomm utative information theory,
Math. Japon, 34(1989), 341-348.
[4] T.Furuta, Operator inequalities associated with A $\log A$ via Specht ratio, Linear Alg.
and Its AppL, 375(2003), 251-273.
[5] Hiai and Petz, The proper formula for relative entropy in asymptotics in quantum
probability, Comm. math.Phys., 143(1991),99-114.
[6] Hiai and Petz, The Golden-Thompson trace inequality is complemented, Linear Alg.
and Its AppL,181 (1993), 153-185.
[7] M.Nakamura and H.Umegaki, A note on entropy for operator algebras, Proc. Japan
Acad., 37(1961) $149- 154$ .
)

[8] H.Umegaki, Conditional expectation in an operator algebra IV, Kodai Math. Sem.
Rep., 14(19621259-85.

Further extension of this appendix will appear in the following:

T.Furuta, Reverse inequalities involving two relative operator entropies and two relative
entropies, to appear in LAA