数学物理学报  2017, Vol. 37 Issue (3): 450-456   PDF    
扩展功能
加入收藏夹
复制引文信息
加入引用管理器
Email Alert
RSS
本文作者相关文章
邹黎敏
吴艳秋
矩阵酉不变范数Hölder不等式及其应用
邹黎敏, 吴艳秋     
重庆三峡学院数学与统计学院 重庆 404100
摘要:讨论了现有的两个矩阵酉不变范数Hölder不等式之间的关系.同时,利用矩阵酉不变范数Hölder不等式以及一些现有的矩阵酉不变范数不等式,得到了几个新的矩阵酉不变范数不等式.所得结果是Alakhrass和Lee等所得相关不等式的推广或改进.
关键词酉不变范数        lder不等式    算子凸函数    算子凹函数    
Hölder Inequality of Unitarily Invariant Norms for Matrices and Its Applications
Zou Limin, Wu Yanqiu     
School of Mathematics and Statistics, Chongqing Three Gorges University, Chongqing 404100
Abstract: The relationship between two existing Hölder inequalities of unitary invariant norms for matrices is discussed. Meanwhile, we present some inequalities of unitarily invariant norms for matrices by using Hölder inequality of unitary invariant norms for matrices and some existing inequalities of unitary invariant norms. Our results are generalizations or refinements of ones shown Alakhrass and Lee.
Key words: Unitarily invariant norms          lder inequality     Operator convex functions     Operator concave functions    
1 前言

文中, $M_{n} $表示$n\times n$复矩阵的全体, $\left| A \right| = \left( {A^* A} \right)^{1/2} $表示矩阵$A$的绝对值算子.我们分别用$\left| {\lambda _1 \left( A \right)} \right| \ge \cdots \ge \left| {\lambda _n \left( A \right)} \right|$$s_1 \left( A \right) \ge \cdots \ge s_n \left( A \right)$来表示矩阵$A$的按照降序排列的特征值和奇异值.用$\left\| \cdot \right\|$表示任意的酉不变范数, 两类常见的酉不变范数是Schatten $p$ -范数

$\left\| A \right\|_p =\bigg( {\sum\limits_{j=1}^n {s_j^p \left( A \right)} } \bigg)^{1 /p}=\left( {{\rm tr}\left| A \right|^p} \right)^{1/p}, 1\le p<\infty $

以及Ky Fan $k$-范数

$ \left\| A \right\|_{\left( k \right)} =\sum\limits_{j=1}^k {s_j \left( A \right)}, k=1, \cdots, n. $

显然, $\left\| A \right\|_1 = \left\| A \right\|_{\left( n \right)} $为矩阵$A$的迹范数, $\left\| A \right\|_2 = \Big( {\sum\limits_{j = 1}^n {s_j^2 \left( A \right)} } \Big)^{1/2}$为矩阵$A$的Frobenius范数, $\left\| A \right\|_\infty = \mathop {\lim }\limits_{p \to \infty } \left\| A \right\|_p = s_1 \left( A \right) = \left\| A \right\|_{\left( 1 \right)} $为矩阵$A$的谱范数, 即是由向量的欧氏范数诱导的$ M_n$上的算子范数.

$A, X, B \in M_{n}$, $\displaystyle \frac{1}{p} + \frac{1}{q}=1, p, q > 1, r \ge 0$, Horn和Zhan在文献[1]中证明了

$\begin{equation} \label{eq1-1} \left\| {\left| {A^* XB} \right|^r } \right\| \le \left\| {\left| {\left| {A^* } \right|^p X} \right|^r } \right\|^{1/p} \left\| {\left| {X\left| {B^* } \right|^q } \right|^r } \right\|^{1/q}, \end{equation}$ (1.1)

这是一个矩阵酉不变范数Hölder不等式. Albadawi在文献[2]中也得到了一个矩阵酉不变范数Hölder不等式

$\begin{equation} \label{eq1-2} \left\| {\left| {A^* XB} \right|^r } \right\| \le \left\| {\left| {AA^* X} \right|^{rp/2} } \right\|^{1/p} \left\| {\left| {XBB^* } \right|^{rq/2} } \right\|^{1/q} . \end{equation}$ (1.2)

比较不等式(1.1) 和(1.2), 自然的, 我们希望知道

$ \left\| {\left| {\left| {A^* } \right|^p X} \right|^r } \right\|^{1/p} \left\| {\left| {X\left| {B^* } \right|^q } \right|^r } \right\|^{1/q} $

$ \left\| {\left| {AA^* X} \right|^{rp/2} } \right\|^{1/p} \left\| {\left| {XBB^* } \right|^{rq/2} } \right\|^{1/q} $

之间的大小关系, 这是本文的动机之一.

Bourin在文献[3]中提出了如下问题:设$A, B\in M_n$为半正定矩阵, $t \in \left[{0, 1} \right]$, 则不等式

$\begin{equation} \label{eq1-3} \left\| {A^t B^{1 - t} + B^t A^{1 - t} } \right\| \le \left\| {A + B} \right\| \end{equation}$ (1.3)

是否成立?利用矩阵酉不变范数三角不等式和Young不等式可知, 对于迹范数, 不等式(1.3) 是成立的.最近, 不等式(1.3) 的一些特殊情形被证明是成立的, 可参见文献[4-6]. Alakhrass在文献[7]中证明了

$\begin{equation} \label{eq1-4} \left\| {A^t B^{1 - t} + B^t A^{1 - t} } \right\| \le 2^{\left| {1/2 - t} \right|} \left\| {A + B} \right\|. \end{equation}$ (1.4)

虽然这个不等式比Bourin的问题弱, 但是比文献[4-6]中的相关结果都更一般化.

$A, B \in M_{n}$, Lee在文献[8]中得到了一个关于Schatten $p$ -范数的三角不等式

$\begin{equation} \label{eq1-5} \left\| {A + B} \right\|_p \le 2^{1/2 - 1/\left( {2p} \right)} \left\| {\left| A \right| + \left| B \right|} \right\|_p . \end{equation}$ (1.5)

在这篇短文中, 我们首先讨论了不等式(1-1) 和(1-2) 之间的关系, 接着, 利用矩阵酉不变范数Hölder不等式, 我们得到了不等式(1-4) 和(1-5) 的一个推广以及$\left\| {A^t B^{1 - t} + B^t A^{1 - t} } \right\|$的一个上界.

2 主要结果

首先讨论

$ \left\| {\left| {\left| {A^* } \right|^p X} \right|^r } \right\|^{1/p} \left\| {\left| {X\left| {B^* } \right|^q } \right|^r } \right\|^{1/q} \mbox{ 和} \left\| {\left| {AA^* X} \right|^{rp/2} } \right\|^{1/p} \left\| {\left| {XBB^* } \right|^{rq/2} } \right\|^{1/q} $

之间的大小关系, 我们有如下的结果.

定理2.1 设$A, X, B \in M_n, \displaystyle \frac{1}{p} + \frac{1}{q} = 1, p, q > 1, r \ge 0$, 则

$ \left\| {\left| {A^* XB} \right|^r } \right\| \le \min \left\{ {\left\| {\left| {\left| {A^* } \right|^p X} \right|^r } \right\|^{1/p} \left\| {\left| {X\left| {B^* } \right|^q } \right|^r } \right\|^{1/q}, \left\| {\left| {AA^* X} \right|^{rp/2} } \right\|^{1/p} \left\| {\left| {XBB^* } \right|^{rq/2} } \right\|^{1/q} } \right\}. $

 只需证明

$ \left\| {\left| {\left| {A^* } \right|^p X} \right|^r } \right\|^{1/p} \left\| {\left| {X\left| {B^* } \right|^q } \right|^r } \right\|^{1/q} \mbox{ 和} \left\| {\left| {AA^* X} \right|^{rp/2} } \right\|^{1/p} \left\| {\left| {XBB^* } \right|^{rq/2} } \right\|^{1/q} $

之间没有谁一定比谁大或是小的关系.因为$\frac{1}{p} + \frac{1}{q}=1, p, q > 1$, 所以$p, q$中一个属于$\left[{2, \infty } \right)$, 一个属于$\left( {1, 2} \right]$, 于是可设$p \in \left( {1, 2} \right), q \in \left( {2, 4} \right)$, 同时, 设$X$为压缩矩阵, $r=2$.因为$f\left( t \right) = t^{p/2}$是算子凹函数, 所以

$ s_j \left( {\left| {AA^* X} \right|^p } \right) = \lambda _j \left( {X^* \left| {A^* } \right|^4 X} \right)^{p/2} \ge \lambda _j \left( {X^* \left| {A^* } \right|^{2p} X} \right) = s_j \left( {\left| {\left| {A^* } \right|^p X} \right|^2 } \right), $

这蕴含着

$ \left\| {\left| {AA^* X} \right|^p } \right\|^{1/p} \ge \left\| {\left| {\left| {A^* } \right|^p X} \right|^2 } \right\|^{1/p}. $

另一方面, 因为$f\left( t \right) = t^{q/2} $是算子凸函数, 所以

$ \begin{array}{lll} s_j \left( {\left| {XBB^* } \right|^q } \right) &=& \lambda _j \left( {BB^* X^* XBB^* } \right)^{q/2} = \lambda _j \left( {X\left| {B^* } \right|^4 X^* } \right)^{q/2} \\ &\le& \lambda _j \left( {X\left| {B^* } \right|^{2q} X^* } \right) = s_j \left( {\left| {X\left| {B^* } \right|^q } \right|^2 } \right), \\ \end{array} $

这蕴含着

$ \left\| {\left| {XBB^* } \right|^q } \right\|^{1/q} \le \left\| {\left| {X\left| {B^* } \right|^q } \right|^2 } \right\|^{1/q}. $

于是可知

$ \left\| {\left| {\left| {A^* } \right|^p X} \right|^r } \right\|^{1/p} \left\| {\left| {X\left| {B^* } \right|^q } \right|^r } \right\|^{1/q} \mbox{ 和} \left\| {\left| {AA^* X} \right|^{rp/2} } \right\|^{1/p} \left\| {\left| {XBB^* } \right|^{rq/2} } \right\|^{1/q} $

之间没有谁一定比谁大或是小的关系.

接下来, 我们将给出不等式(1-4) 的推广, 为此, 我们需要如下的引理.

引理2.1[9] 设$A_1, \cdots, A_m \in M_n$为半正定矩阵, $f\left( t \right)$是非负凹函数, 则

$ \left\| {f\left( {\sum\limits_{i = 1}^m {A_i } } \right)} \right\| \le \left\| {\sum\limits_{i = 1}^m {f\left( {A_i } \right)} } \right\|. $

引理2.2[10] 设$A, B \in M_{n}$, $f\left( t \right)$是递增的凸函数.若

$ \left\| A \right\| \le \left\| B \right\|, $

$ \left\| {f\left( A \right)} \right\| \le \left\| {f\left( B \right)} \right\|. $

定理2.2 设$A_i, X_i, B_i \in M_n, i = 1, \cdots, m, \displaystyle \frac{1}{p} + \frac{1}{q} = 1, p, q > 1$, 则

$\begin{equation} \label{eq2-1} \left\| {\sum\limits_{i = 1}^m {A_i^* X_i B_i } } \right\| \le m^{\left| {1/2 - 1/p} \right|} \max \left\{ {\left\| {X_1 } \right\|_\infty, \cdots, \left\| {X_m } \right\|_\infty } \right\}\left\| {\sum\limits_{i = 1}^m {\left| {A_i } \right|^p } } \right\|^{1/p} \left\| {\sum\limits_{i = 1}^m {\left| {B_i } \right|^q } } \right\|^{1/q}. \end{equation}$ (2.1)

 由不等式(1-1) 以及$\left\| {\left| {A^* } \right|^r } \right\| = \left\| {\left| A \right|^r } \right\|, r \ge 0$可得

$\begin{eqnarray} \label{eq2-2} \left\| {A^* XB} \right\| &\le& \left\| {\left| {A^* } \right|^p X} \right\|^{1/p} \left\| {X\left| {B^* } \right|^q } \right\|^{1/q} \\ &\le& \left\| X \right\|_\infty \left\| {\left| {A^* } \right|^p } \right\|^{1/p} \left\| {\left| {B^* } \right|^q } \right\|^{1/q} \\ &=& \left\| X \right\|_\infty \left\| {\left| A \right|^p } \right\|^{1/p} \left\| {\left| B \right|^q } \right\|^{1/q}. \end{eqnarray}$ (2.2)

$ A = \left[{\begin{array}{*{20}c} {A_1 } & 0 & \cdots & 0 \\ {A_2 } & 0 & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ {A_m } & 0 & \cdots & 0 \\ \end{array}} \right],  X = \left[{\begin{array}{*{20}c} {X_1 } & 0 & \cdots & 0 \\ 0 & {X_2 } & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & {X_m } \\ \end{array}} \right],  B = \left[{\begin{array}{*{20}c} {B_1 } & 0 & \cdots & 0 \\ {B_2 } & 0 & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ {B_m } & 0 & \cdots & 0 \\ \end{array}} \right], $

于是, 由不等式(2-2) 可得

$\begin{eqnarray} \label{eq2-3} \left\| {A^* XB} \right\| &=& \left\| {\sum\limits_{i = 1}^m {A_i^* X_i B_i } } \right\| \\ &\le& \max \left\{ {\left\| {X_1 } \right\|_\infty, \cdots, \left\| {X_m } \right\|_\infty } \right\}\left\| {\left( {\sum\limits_{i = 1}^m {\left| {A_i } \right|^2 } } \right)^{p/2} } \right\|^{1/p} \left\| {\left( {\sum\limits_{i = 1}^m {\left| {B_i } \right|^2 } } \right)^{q/2} } \right\|^{1/q} . \end{eqnarray}$ (2.3)

因为$\displaystyle \frac{1}{p} + \frac{1}{q}=1, p, q > 1$, 所以$p, q$中一个属于$\left[{2, \infty } \right)$, 一个属于$\left( {1, 2} \right]$, 由于$p, q$的对称性, 于是可设$1 < p \le 2 \le q$.因此, 由引理2.1可知

$\begin{equation} \label{eq2-4} \left\| {\left( {\sum\limits_{i = 1}^m {\left| {A_i } \right|^2 } } \right)^{p/2} } \right\|^{1/p} \le \left\| {\sum\limits_{i = 1}^m {\left| {A_i } \right|^p } } \right\|^{1/p}. \end{equation}$ (2.4)

因为$q \ge 2$, 所以函数$g\left( t \right) = t^{2/q} $是算子凹函数, 于是可得

$ \left\| {\sum\limits_{i = 1}^m {\left| {B_i } \right|^2 } /m} \right\| \le \left\| {\left( {\sum\limits_{i = 1}^m {\left| {B_i } \right|^q /m} } \right)^{2/q} } \right\|. $

结合引理2.2有

$\begin{equation} \label{eq2-5} \left\| {\left( {\sum\limits_{i = 1}^m {\left| {B_i } \right|^2 } } \right)^{q/2} } \right\|^{1/q} \le m^{1/2 - 1/q} \left\| {\sum\limits_{i = 1}^m {\left| {B_i } \right|^q } } \right\|^{1/q}. \end{equation}$ (2.5)

由不等式(2-3), (2-4) 以及(2-5)

$ \left\| {\sum\limits_{i = 1}^m {A_i^* X_i B_i } } \right\| \le m^{1/2 - 1/q} \max \left\{ {\left\| {X_1 } \right\|_\infty, \cdots, \left\| {X_m } \right\|_\infty } \right\}\left\| {\sum\limits_{i = 1}^m {\left| {A_i } \right|^p } } \right\|^{1/p} \left\| {\sum\limits_{i = 1}^m {\left| {B_i } \right|^q } } \right\|^{1/q} . $

同样的, 当$1 < q \le 2 \le p$时, 我们有

$ \left\| {\sum\limits_{i = 1}^m {A_i^* X_i B_i } } \right\| \le m^{1/2 - 1/p} \max \left\{ {\left\| {X_1 } \right\|_\infty, \cdots, \left\| {X_m } \right\|_\infty } \right\}\left\| {\sum\limits_{i = 1}^m {\left| {A_i } \right|^p } } \right\|^{1/p} \left\| {\sum\limits_{i = 1}^m {\left| {B_i } \right|^q } } \right\|^{1/q} . $

结合上面两个不等式以及$\left| {1/2 - 1/p} \right| = \left| {1/2 - 1/q} \right|$可得到不等式(2-1).

注2.1 在不等式(2-1) 中令$X_1 = \cdots = X_m = I$可得文献[2]中的定理17.

注2.2 设$A_1, A_2, X_1, X_2, B_1, B_2\in M_n, \frac{1}{p} + \frac{1}{q}=1, p, q > 1$, 则由不等式(2-1) 可得

$\begin{eqnarray} \label{eq2-6} && \left\| {A_1^* X_1 B_1 + A_2^* X_2 B_2 } \right\| \\ &\le& 2^{\left| {1/2 - 1/p} \right|} \max \left\{ {\left\| {X_1 } \right\|_\infty, \left\| {X_2 } \right\|_\infty } \right\}\left\| {\left| {A_1 } \right|^p + \left| {A_2 } \right|^p } \right\|^{1/p} \left\| {\left| {B_1 } \right|^q + \left| {B_2 } \right|^q } \right\|^{1/q} . \end{eqnarray}$ (2.6)

在不等式(2-6) 中, 令$X_1 =X_2 = I$可得文献[11]中的定理11.另一方面, 在不等式(2-6) 中, 令

$ A_1^* = A^t, B_1 = B^{1 - t}, A_2^* = B^t, B_2 = A^{1 - t}, p = \frac{1}{t}, q = \frac{1}{{1 - t}}, $

可得

$ \left\| {A^t X_1 B^{1 - t} + B^t X_2 A^{1 - t} } \right\| \le 2^{\left| {1/2 - t} \right|} \max \left\{ {\left\| {X_1 } \right\|_\infty, \left\| {X_2 } \right\|_\infty } \right\}\left\| {A + B} \right\|. $

这是不等式(1-4) 的一个推广.

接下来, 同样的, 利用矩阵酉不变范数Hölder不等式, 我们将不等式(1.5) 推广到了多个矩阵的情形.

定理2.3 设$A_i\in M_n, i = 1, \cdots, m$, 则

$\begin{equation} \label{eq2-7} \left\| {\sum\limits_{i = 1}^m {A_i } } \right\|_p \le m^{1/2 - 1/\left( {2p} \right)} \left\| {\sum\limits_{i = 1}^m {\left| {A_i } \right|} } \right\|_p . \end{equation}$ (2.7)

 设$A_i = U_i \left| {A_i } \right|$为矩阵$A_i\in M_n$的极分解, 于是可知

$ A_i = \left| {A_i^* } \right|^{1/2} U_i \left| {A_i } \right|^{1/2}, i = 1, \cdots, m, $

所以我们有

$ \sum\limits_{i = 1}^m {A_i } = \left[{\left| {A_1^* } \right|^{1/2} \cdots \left| {A_m^* } \right|^{1/2} } \right]{\rm diag}\left( {U_1, \cdots, U_m } \right)\left[\begin{array}{c} \left| {A_1 } \right|^{1/2} \\ \vdots \\ \left| {A_m } \right|^{1/2} \\ \end{array} \right]. $

由不等式(2-2) 可得

$\begin{equation} \label{eq2-8} \left\| {\sum\limits_{i = 1}^m {A_i } } \right\| \le \left\| {\sum\limits_{i = 1}^m {\left| {A_i^* } \right|} } \right\|^{1/2} \left\| {\sum\limits_{i = 1}^m {\left| {A_i } \right|} } \right\|^{1/2}. \end{equation}$ (2.8)

对于Schatten $p$ -范数, 由函数$f\left( t \right) = t^{1/p}, 1 \le p < \infty $的凹性和引理2.1可知

$\begin{eqnarray} \label{eq2-9} \left\| {A_1 } \right\|_p + \cdots + \left\| {A_m } \right\|_p &=& \left( {{\rm tr}\left| {A_1 } \right|^p } \right)^{1/p} + \cdots + \left( {{\rm tr}\left| {A_m } \right|^p } \right)^{1/p} \\ &\le& m^{1 - 1/p} \left( {{\rm tr}\left| {A_1 } \right|^p + \cdots + {\rm tr}\left| {A_m } \right|^p } \right)^{1/p} \\ &=& m^{1 - 1/p} \left\| {\left( {\left| {A_1 } \right|^p + \cdots + \left| {A_m } \right|^p } \right)^{1/p} } \right\|_p \\ &\le& m^{1 - 1/p} \left\| {\left| {A_1 } \right| + \cdots + \left| {A_m } \right|} \right\|_p. \end{eqnarray}$ (2.9)

由酉不变范数的三角不等式和不等式(2-9) 可知

$\begin{equation} \label{eq2-10} \left\| {\sum\limits_{i = 1}^m {\left| {A_i^* } \right|} } \right\|_p \le \sum\limits_{i = 1}^m {\left\| {\left| {A_i^* } \right|} \right\|_p } = \sum\limits_{i = 1}^m {\left\| {A_i } \right\|_p } \le m^{1 - 1/p} \left\| {\sum\limits_{i = 1}^m {\left| {A_i } \right|} } \right\|_p . \end{equation}$ (2.10)

结合不等式(2-8) 和(2-10) 可知结论成立.

注2.3 显然, 不等式(2-7) 是不等式(1-5) 的一个推广.

注2.4 设$A_i\in M_n, i = 1, \cdots, m$, Lee在文献[12]中证明了

$ \left\| {\sum\limits_{i = 1}^m {A_i } } \right\| \le m^{1/2} \left\| {\sum\limits_{i = 1}^m {\left| {A_i } \right|} } \right\|. $

显然, 对于Schatten $p$ -范数, 不等式(2-7) 是Lee所得结果的一个改进.

最后, 我们给出$\left\| {A^t B^{1 - t} + B^t A^{1 - t} } \right\|$的一个上界.

定理2.4 设$A, B\in M_n$为半正定矩阵, $t \in \left[{0, 1} \right]$, 则

$\begin{equation} \label{eq2-11} \left\| {A^t B^{1 - t} + B^t A^{1 - t} } \right\| \le \left\| A \right\| + \left\| B \right\| - 2r_0 \left( {\sqrt {\left\| A \right\|} - \sqrt {\left\| B \right\|} } \right)^2, \end{equation}$ (2.11)

其中$r_0 =\min \left\{ {t, \;1-t} \right\}$.

 由矩阵酉不变范数三角不等式和不等式(2-2) 可得

$ \begin{array}{lll} \left\| {A^t B^{1 - t} + B^t A^{1 - t} } \right\| &\le& \left\| {A^t B^{1 - t} } \right\| + \left\| {B^t A^{1 - t} } \right\| \\ &\le& \left\| {A^{pt} } \right\|^{1/p} \left\| {B^{q\left( {1 - t} \right)} } \right\|^{1/q} + \left\| {B^{pt} } \right\|^{1/p} \left\| {A^{q\left( {1 - t} \right)} } \right\|^{1/q}. \end{array} $

在上面这个不等式中, 令$ p = \frac{1}{t}, $ $q = \frac{1}{{1 - t}}, $可得

$\begin{equation} \label{eq2-12} \left\| {A^t B^{1 - t} + B^t A^{1 - t} } \right\| \le \left\| A \right\|^t \left\| B \right\|^{1 - t} + \left\| A \right\|^{1 - t} \left\| B \right\|^t. \end{equation}$ (2.12)

Kittaneh和Manasrah在文献[13]中证明了:若$a, b \ge 0, t\in [0, 1]$, 则

$\begin{equation} \label{eq2-13} a^t b^{1 - t} + a^{1 - t} b^t \le a + b - 2r_0 \left( {\sqrt a - \sqrt b } \right)^2, \end{equation}$ (2.13)

其中$r_0 =\min \left\{ {t, \;1-t} \right\}$, 结合不等式(2-12) 和(2-13) 有

$ \left\| {A^t B^{1 - t} + B^t A^{1 - t} } \right\| \le \left\| A \right\| + \left\| B \right\| - 2r_0 \left( {\sqrt {\left\| A \right\|} - \sqrt {\left\| B \right\|} } \right)^2. $

证毕.

注2.5 对于迹范数, 当$A, B\in M_n$为半正定矩阵时, 我们有$\left\| A \right\|_1 + \left\| B \right\|_1 = \left\| {A + B} \right\|_1$, 所以, 由不等式(2-11) 可得

$\begin{equation} \label{eq2-14} \left\| {A^t B^{1 - t} + B^t A^{1 - t} } \right\|_1 \le \left\| {A + B} \right\|_1 - 2r_0 \left( {\sqrt {\left\| A \right\|_1 } - \sqrt {\left\| B \right\|_1 } } \right)^2, \end{equation}$ (2.14)

其中$r_0 =\min \left\{ {t, \;1-t} \right\}$, 这是不等式(1-3) 的一个改进.由不等式(2-9) 和(2-11)

$ \left\| {A^t B^{1 - t} + B^t A^{1 - t} } \right\|_p \le 2^{1 - 1/p} \left\| {A + B} \right\|_p - 2r_0 \left( {\sqrt {\left\| A \right\|_p } - \sqrt {\left\| B \right\|_p } } \right)^2, $

其中$r_0 =\min \left\{ {t, \;1-t} \right\}$, 这是不等式(2-14) 的一个推广.

注2.6 对比不等式(1-3) 和(2-11), 自然的, 我们希望知道$\left\| A \right\| + \left\| B \right\| - 2r_0 \big( \sqrt {\left\| A \right\|} -$ $ \sqrt {\left\| B \right\|} \big)^2$$\left\| {A + B} \right\|$之间的大小关系.取$\left\| \cdot \right\| = \left\| \cdot \right\|_2, t =1/4$, 同时令

$ A = \left[{\begin{array}{*{20}c} 5 & 1 \\ 1 & 5 \\ \end{array}} \right],  B = \left[{\begin{array}{*{20}c} 1 & 0 \\ 0 & 1 \\ \end{array}} \right]. $

简单计算可知

$ \left\| {A + B} \right\|_2 = {\rm{8}}{\rm{.6023}} \ge {\rm{7}}{\rm{.5061}} = \left\| A \right\|_2 + \left\| B \right\|_2 - 2r_0 \left( {\sqrt {\left\| A \right\|_2 } - \sqrt {\left\| B \right\|_2 } } \right)^2. $

另一方面, 若取

$ A = \left[{\begin{array}{*{20}c} 1 & 1 \\ 1 & 1 \\ \end{array}} \right],  B = \left[{\begin{array}{*{20}c} 1 & 0 \\ 0 & 1 \\ \end{array}} \right]. $

简单计算可知

$ \left\| {A + B} \right\|_2 = {\rm{3}}{\rm{.1623}} \le {\rm{3}}{\rm{.3889}} = \left\| A \right\|_2 + \left\| B \right\|_2 - 2r_0 \left( {\sqrt {\left\| A \right\|_2 } - \sqrt {\left\| B \right\|_2 } } \right)^2. $

所以, 它们这两者之间没有谁一定比谁大或是小的关系.

参考文献
[1] Horn R A, Zhan X. Inequalities for C-S seminorms and Lieb functions. Linear Algebra Appl, 1999, 291(1-3): 103–113. DOI:10.1016/S0024-3795(98)10245-8
[2] Albadawi H. Hölder-type inequalities involving unitarily invariant norms. Positivity, 2012, 16(2): 255–270. DOI:10.1007/s11117-011-0122-z
[3] Bourin J C. Matrix subadditivity inequalities and block-matrices. Int J Math, 2009, 20(6): 679–691. DOI:10.1142/S0129167X09005509
[4] Hayajneh S, Kittaneh F. Lieb-thirring trace inequalities and a question of Bourin. J Math Phys, 2013, 54: 033504. DOI:10.1063/1.4793993
[5] Bhatia R. Trace inequalities for product of positive definite matrices. J Math Phys, 2014, 55: 013509. DOI:10.1063/1.4863899
[6] Yousef A, Sababheh M, Khalil R. Operator-norm inequalities and interpolated trace inequalities. Linear Multilinear Algebra, 2015, 63(3): 455–465. DOI:10.1080/03081087.2013.872639
[7] Alakhrass M. Inequalities related to Heinz mean. Linear Multilinear Algebra, 2016, 64(8): 1562–1569. DOI:10.1080/03081087.2015.1106435
[8] Lee E Y. Rotfel'd type inequalities for norms. Linear Algebra Appl, 2010, 433(3): 580–584. DOI:10.1016/j.laa.2010.03.029
[9] Bourin J C, Uchiyama M. A matrix subadditivity inequality for f (A + B) and f (A) + f (B). Linear Algebra Appl, 2007, 423(2-3): 512–518. DOI:10.1016/j.laa.2007.02.019
[10] Bhatia R. Matrix Analysis. New York:Springer, 1997
[11] Hiai F, Zhan X. Inequalities involving unitarily invariant norms and operator monotone functions. Linear Algebra Appl, 2002, 341(1-3): 151–169. DOI:10.1016/S0024-3795(01)00353-6
[12] Lee E Y. How to compare the absolute values of operator sums and the sums of absolute values?. Oper Matrices, 2012, 6(3): 613–619.
[13] Kittaneh F, Manasrah Y. Improved Young and Heinz inequalities for matrices. J Math Anal Appl, 2010, 361(1): 262–269. DOI:10.1016/j.jmaa.2009.08.059