# Bounds of the logarithmic mean

- Shigeru Furuichi
^{1}Email author and - Kenjiro Yanagi
^{2}

**2013**:535

https://doi.org/10.1186/1029-242X-2013-535

© Furuichi and Yanagi; licensee Springer. 2013

**Received: **4 August 2013

**Accepted: **16 October 2013

**Published: **12 November 2013

## Abstract

We give tight bounds for the logarithmic mean. We also give new Frobenius norm inequalities for two positive semidefinite matrices. In addition, we give some matrix inequalities on the matrix power mean.

**MSC:** 15A39, 15A45.

## Keywords

## 1 Introduction

*a*and

*b*. (We conventionally define $L(a,b)=a$ if $a=b$.) In the paper [1], the following relations were shown:

We now have the following lemma.

**Lemma 1.1**

*For*$a,b>0$,

*we have*

*Proof* The second inequality of (3) can be proven easily. Indeed, we put $f(t)\equiv \frac{2}{3}{t}^{3}+\frac{1}{6}(1+{t}^{6})-\frac{1}{8}{(1+{t}^{2})}^{3}$. Then we have ${f}^{\prime}(t)<0$ for $0<t<1$ and ${f}^{\prime}(t)>0$ for $t>1$. Thus we have $f(t)\ge f(1)=0$ for $t>0$. □

which is known as the classical Pólya inequality [2, 3].

In the paper [2], the following norm inequality was shown.

**Theorem 1.2** ([2])

*For*$A,B\in {M}_{+}(n,\mathbb{C})$, $X\in M(n,\mathbb{C})$

*and the Frobenius norm*${\parallel \cdot \parallel}_{F}$,

*we have*

From Lemma 1.1, we have the following proposition.

**Proposition 1.3**

*For*$A,B\in {M}_{+}(n,\mathbb{C})$, $X\in M(n,\mathbb{C})$

*and the Frobenius norm*${\parallel \cdot \parallel}_{F}$,

*we have*

To the first author’s best knowledge, the first inequality in Proposition 1.3 was suggested in [4].

This proposition can be proven in a similar way to the proof of Theorem 1.2 (or the proof of Theorem 2.2 which will be given in the next section), and this refines inequality (6) shown in [2].

## 2 Lower bound of the logarithmic mean

The following inequalities were given in [5]. Hiai and Kosaki gave the norm inequalities for Hilbert space operators in [5]. See also [6, 7]. Here we give them as a matrix setting to unify the description of this paper.

**Theorem 2.1** ([5])

*For*$A,B\in {M}_{+}(n,\mathbb{C})$, $X\in M(n,\mathbb{C})$, $m\ge 1$

*and every unitarily invariant norm*$\u2980\cdot \u2980$,

*we have*

The Frobenius norm is one of unitarily invariant norms. We give the refinement of the lower bound of the first inequality above for the Frobenius norm. That is, we have the following inequalities.

**Theorem 2.2**

*For*$A,B\in {M}_{+}(n,\mathbb{C})$, $X\in M(n,\mathbb{C})$, $m\ge 1$

*and the Frobenius norm*${\parallel \cdot \parallel}_{F}$,

*we have*

To prove Theorem 2.2, we need a few lemmas.

**Lemma 2.3**

*Let*

*u*,

*v*,

*w*

*be nonnegative integers such that*$w\ge u$,

*and let*

*x*

*be a positive real number*.

*Then we have*

*Proof*It is trivial for the case $x=1$ or $v=0$. We prove for the case $x\ne 1$ and $v\ne 0$. In addition, for the case that $u=w$, the equality holds. Thus we may assume $w>u$ and $v\ge 1$. Then the lemma can be proven in the following way.

□

**Lemma 2.4**

*For a positive real number*

*x*

*and a natural number*

*m*,

*we have*

*Proof*For the case $x=1$, the equality holds. So we prove this lemma for $x\ne 1$. If

*m*is an odd number, then we have $m=2\lfloor m/2\rfloor +1$. Since we then have $m+1=2(\lfloor m/2\rfloor +1)$ and $2(\lfloor m/2\rfloor +1)-1=m$, we have

*m*is an even number, it never happens that the difference of the $\tilde{k}$th term of the both sides in inequality (9) is equal to 0. Therefore we have

where ${a}_{l}=(2l-1)(m+1)$, ${b}_{l}=m-(2l-1)$ and ${c}_{l}=2\{m-(l-1)\}m$ for $l=1,2,\dots ,\lfloor m/2\rfloor $. Here we have ${c}_{l}-{a}_{l}=\{m-(2l-1)\}(2m+1)\ge 0$, whenever ${b}_{l}\ge 0$. By Lemma 2.3, if ${b}_{l}\ge 0$, then we have ${x}^{{a}_{l}}(1-{x}^{{b}_{l}})+{x}^{{c}_{l}}({x}^{{b}_{l}}-1)\ge 0$. Thus the proof of this lemma is completed. □

We then have the following lemma.

**Lemma 2.5**

*For*$a,b>0$

*and*$m\ge 1$,

*we have*

*Proof*The second inequality follows by inequality (11). We use the famous inequality $(x-1)/logx\ge \sqrt{x}$ for $x>0$. We put $x={t}^{1/m}$ in this inequality. Then we have, for $t>0$,

which implies the first inequality. The third inequality can be proven by the use of the arithmetic mean-geometric mean inequality. Thus the proof of this lemma is completed. □

We give some basic properties of the right-hand side of inequality (13) in the Appendix.

*Proof of Theorem 2.2*It is known that the Frobenius norm inequality immediately follows from the corresponding scalar inequality [6]. However, we give here an elementary proof for the convenience of the readers. Let ${U}^{\ast}AU={D}_{1}=diag\{{\alpha}_{1},\dots ,{\alpha}_{n}\}$ and ${V}^{\ast}BV={D}_{2}=diag\{{\beta}_{1},\dots ,{\beta}_{n}\}$. Then ${\alpha}_{1},\dots ,{\alpha}_{n}\ge 0$ and ${\beta}_{1},\dots ,{\beta}_{n}\ge 0$. We put ${U}^{\ast}XV\equiv Y=({y}_{ij})$. By the first inequality of (12), we have

Applying inequality (11), we have the second inequality of (7) in a similar way. The third inequality holds due to Theorem 2.1 (or the third inequality of (12)). □

## 3 Upper bound of the logarithmic mean

In the paper [5], the following norm inequalities were also given for Hilbert space operators. Here we give them for matrices, as we mentioned in the beginning of Section 2.

**Theorem 3.1** ([5])

*For*$A,B\in {M}_{+}(n,\mathbb{C})$, $X\in M(n,\mathbb{C})$, $m\ge 2$

*and every unitarily invariant norm*$\u2980\cdot \u2980$,

*we have*

We also give an improved upper bound of the logarithmic mean on Theorem 3.1 above, only for the Frobenius norm. Namely, we can prove the following inequalities in a similar way to the proof of Theorem 2.2, by the use of scalar inequalities which will be given in Lemma 3.4.

**Theorem 3.2**

*For*$A,B\in {M}_{+}(n,\mathbb{C})$, $X\in M(n,\mathbb{C})$, $m\ge 2$

*and the Frobenius norm*${\parallel \cdot \parallel}_{F}$,

*we have*

To prove Theorem 3.2, we need to prove the following lemmas.

**Lemma 3.3**

*For*$x>0$

*and*$m\ge 2$,

*we have*

*Proof*For $m\ge 2$, we calculate

where ${p}_{l}=(2l-1)m$, ${q}_{l}=2(m-1)(m-l)$ and ${r}_{l}=m-2l$. The last inequality follows from Lemma 2.3, because we have ${q}_{l}-{p}_{l}=(2m-1)(m-2l)\ge $ whenever ${r}_{l}\ge 0$. □

**Lemma 3.4**

*For*$a,b>0$

*and*$m\ge 2$,

*we have*

*Proof*To prove the first inequality, we have only to prove the following inequality:

Inequality (19) can be proven by putting $x\equiv {t}^{1/m}>0$ in the famous inequality $(x-1)/logx\le (x+1)/2$ for $x>0$.

To prove the second inequality of (18), it is sufficient to prove inequality (15) which follows from Lemma 3.3. We obtain actually the second inequality of (18) by putting $t={x}^{m(m-1)}>0$ in inequality (15), and then putting $t=a/b$.

*m*. Indeed, we assume that the inequality (20) holds for some

*m*. Then we add ${t}^{m}>0$ to both sides of inequality (20). Then we have

We put ${f}_{m}(t)\equiv (m-1){t}^{m}-m{t}^{m-1}+1$. Then we can prove ${f}_{m}(t)\ge {f}_{m}(1)=0$ by elementary calculations. Thus inequality (20) holds for $m+1$. □

We give some basic properties of the right-hand side of inequality (19) in the Appendix.

## 4 Matrix inequalities on the geometric mean

Using Lemma 1.1, Lemma 2.5 and Lemma 3.4, we have following Proposition 4.1, Proposition 4.2 and Proposition 4.3, respectively.

**Proposition 4.1**

*For*$A,B\in {M}_{+}(n,\mathbb{C})$,

*we have*

*where* $A{\mathrm{\#}}_{\nu}B\equiv {A}^{1/2}{({A}^{-1/2}B{A}^{-1/2})}^{\nu}{A}^{1/2}$ ($0\le \nu \le 1$) *is* *ν*-*weighted geometric mean introduced in* [8].

**Proposition 4.2**

*For*$A,B\in {M}_{+}(n,\mathbb{C})$

*and*$m\ge 1$,

*we have*

**Proposition 4.3**

*For*$A,B\in {M}_{+}(n,\mathbb{C})$

*and*$m\ge 2$,

*we have*

We give the proof of Proposition 4.3. Proposition 4.1 and Proposition 4.2 are also proven in a similar way, using Lemma 1.1 and Lemma 2.5. In addition, by using the notion of the representing function ${f}_{m}(x)=1mx$ for the operator mean *m*, it is well known [8] that ${f}_{m}(x)\le {f}_{n}(x)$ holds for $x>0$ if and only if $AmB\le AnB$ holds for all positive operators *A* and *B*. However, we give an elementary proof for the convenience of the readers.

*Proof of Proposition 4.3*Since $T\equiv {A}^{-1/2}B{A}^{-1/2}\ge 0$, there exists a unitary matrix

*U*such that ${U}^{\ast}TU=D\equiv diag\{{\lambda}_{1},\dots ,{\lambda}_{n}\}$. Then ${\lambda}_{1},\dots ,{\lambda}_{n}\ge 0$. From Lemma 3.4, for $i=1,\dots ,n$, we have

*U*and ${U}^{\ast}$ to both sides, we have

Inserting $T\equiv {A}^{-1/2}B{A}^{-1/2}$ and then multiplying two ${A}^{1/2}$ to all sides from both sides, we obtain the result. □

Inequalities (22) imply the following result in a similar way to the proof of Proposition 4.3.

**Proposition 4.4**

*For*$A,B\in {M}_{+}(n,\mathbb{C})$,

*we have*

## 5 Comments

Proposition 6.2 given in the Appendix shows that our upper bound is tighter than the standard upper bound for the case $t>1$ and our lower bound is tighter than the standard lower bound for any $t>0$. In addition, our lower bound ${\alpha}_{m}(t)$ of the logarithmic mean $L(t,1)$ is tighter than the lower bound $\sqrt{t}$ given by Lin in [1] for $m\ge 1$. However, it may be a difficult problem to find the minimum $m\in \mathbb{N}$ such that ${\beta}_{m}(t)\le {(({t}^{1/3}+1)/2)}^{3}$ for any $t>0$. The right-hand side of the above inequality is the upper bound given by Lin in [1]. (See inequalities (2).)

## Appendix

for $t>0$.

**Proposition 6.1**

*For any*$t>0$,

*we have following properties*:

- (i)
${\alpha}_{m}(t)\le {\alpha}_{m+1}(t)$.

- (ii)
${\beta}_{m+1}(t)\le {\beta}_{m}(t)$.

- (iii)
${\alpha}_{m}(t)$

*and*${\beta}_{m}(t)$*converges to*$L(t,1)$*as*$m\to \mathrm{\infty}$.*In addition*,*we have*${\alpha}_{m}(t)\le {\beta}_{m}(t)$.

*Proof*We prove (i)-(iii) for $t\ne 1$, since it is trivial for the case $t=1$.

- (i)Since${\alpha}_{m}(t)=\frac{{t}^{1/(2m)}(t-1)}{m({t}^{1/m}-1)},$

- (ii)Since${\beta}_{m}(t)=\frac{({t}^{1/m}+1)(t-1)}{2m({t}^{1/m}-1)},$

- (iii)Since we have$\underset{m\to \mathrm{\infty}}{lim}m({t}^{1/m}-1)=\underset{p\to 0}{lim}\frac{{t}^{p}-1}{p}=logt$

The arithmetic-geometric mean inequality proves ${\alpha}_{m}(t)<{\beta}_{m}(t)$ for $t>0$ and $t\ne 1$. □

Then we have the following relations.

**Proposition 6.2**

- (i)
*For*$0<t<1$,*we have*${\alpha}_{m}(t)>{\gamma}_{m}(t)$*and*${\beta}_{m}(t)>{\delta}_{m}(t)$. - (ii)
*For*$t>1$,*we have*${\alpha}_{m}(t)>{\delta}_{m}(t)$*and*${\beta}_{m}(t)<{\gamma}_{m}(t)$.

*Proof*

- (i)For the case $0<t<1$, the following calculations show assertion:${\alpha}_{m}(t)-{\gamma}_{m}(t)=\frac{(t-1)({t}^{1/(2m)}-{t}^{1/m})}{m({t}^{1/m}-1)}>0,\phantom{\rule{2em}{0ex}}{\beta}_{m}(t)-{\delta}_{m}(t)=\frac{t+1}{2m}>0.$
- (ii)For the case $t>1$, the following calculations show assertion:${\alpha}_{m}(t)-{\delta}_{m}(t)=\frac{t+{t}^{1/(2m)}}{m({t}^{1/(2m)}+1)}>0,\phantom{\rule{2em}{0ex}}{\beta}_{m}(t)-{\gamma}_{m}(t)=\frac{1-t}{2m}<0.$

□

## Declarations

### Acknowledgements

The author (SF) was partially supported by JSPS KAKENHI Grant No. 24540146. The author (KY) was also partially supported by JSPS KAKENHI Grant No. 23540208.

## Authors’ Affiliations

## References

- Lin T-P: The power mean and the logarithmic mean.
*Am. Math. Mon.*1974, 81: 879–883. 10.2307/2319447View ArticleMathSciNetMATHGoogle Scholar - Zou L: Matrix versions of the classical Pólya inequality.
*ScienceAsia*2013, 39: 204–207. 10.2306/scienceasia1513-1874.2013.39.204View ArticleGoogle Scholar - Niculescu CP, Person LE CMS Books in Mathematics 23. In
*Convex Functions and Their Applications: A Contemporary Approach*. Springer, New York; 2006.View ArticleGoogle Scholar - Lin, M: Power-logarithmic mean inequality for Frobenius norm. Personal communicationGoogle Scholar
- Hiai F, Kosaki H: Comparison of various means for operators.
*J. Funct. Anal.*1999, 163: 300–323. 10.1006/jfan.1998.3375MathSciNetView ArticleMATHGoogle Scholar - Hiai F, Kosaki H Lecture Notes in Mathematics 1820. In
*Means of Hilbert Space Operators*. Springer, Berlin; 2003.View ArticleGoogle Scholar - Hiai F: Matrix analysis: matrix monotone functions, matrix means, and majorization.
*Interdiscip. Inf. Sci.*2010, 16: 139–248.MathSciNetMATHGoogle Scholar - Kubo F, Ando T: Means of positive operators.
*Math. Ann.*1980, 264: 205–224.MathSciNetView ArticleMATHGoogle Scholar

## Copyright

This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.