7
$\begingroup$

Suppose $a,b$ are two positive-definite linear operators on (say) $\mathbb R^n$. For $p\in(0,1)$, do we then have $(a+b)^p\leq a^p+b^p$ (with respect to the Loewner order)?

$\endgroup$
6
  • 1
    $\begingroup$ What do you mean by 'operator' here? And then what does the power mean? $\endgroup$
    – David Roberts
    Nov 1, 2022 at 12:02
  • 2
    $\begingroup$ @DavidRoberts : I think "operator" here is supposed to mean "linear operator". The matrix functions (here, matrix power functions) can then be defined in a standard manner -- cf. e.g. nhigham.com/2020/06/09/what-is-a-matrix-function $\endgroup$ Nov 1, 2022 at 13:45
  • 2
    $\begingroup$ I am also curious what is meant by $\le$ relation on operators. Is it supposed to be a lexicographic order on matrices with respect to a certain fixed basis? $\endgroup$ Nov 1, 2022 at 13:46
  • 1
    $\begingroup$ Please clarify your specific problem or provide additional details to highlight exactly what you need. As it's currently written, it's hard to tell exactly what you're asking. $\endgroup$
    – Community Bot
    Nov 1, 2022 at 13:47
  • 2
    $\begingroup$ @GrishaTaroyan : Since positive operators were mentioned, it appears that the order in question is the Loewner order. I have taken the liberty to edit the post accordingly. $\endgroup$ Nov 1, 2022 at 13:53

2 Answers 2

6
$\begingroup$

Iosif's example can be given a more conceptual description. Take $a=P$, $b=Q$ as projections. Then $P^p=P$, $Q^p=Q$, so the desired inequality becomes $$ (P+Q)^p \le P+Q . $$ Now $T^p\le T$, for $0<p<1$, is equivalent to $T$ having no eigenvalues in $(0,1)$. However, it's easy to give $P+Q$ an eigenvalue in this range.

For example, if $R(P),R(Q)$ span the whole space and there is a $v$ with $Pv=0$, $Qv\not= v$, then the smallest eigenvalue lies in $(0,1)$, by min-max.

$\endgroup$
2
  • 2
    $\begingroup$ This is very nice! $\endgroup$ Nov 1, 2022 at 19:55
  • 3
    $\begingroup$ @IosifPinelis: Thank you, and sorry for misspelling your name (fixed now). $\endgroup$ Nov 1, 2022 at 20:53
5
$\begingroup$

No. E.g., (identifying linear operators with matrices in a standard manner) let $$a=\left( \begin{array}{cc} 1 & 0 \\ 0 & 0 \\ \end{array} \right),\quad b=\frac12 \left( \begin{array}{cc} 1 & 1 \\ 1 & 1 \\ \end{array} \right). $$ Then (by straightforward but somewhat tedious calculations) for all $p\in(0,1)$ $$d(p):=\det(a^p+b^p-(a+b)^p) \\ =2^{-p-3/2} \\ \times\left(\sqrt{2} \left(2^p+2\right)-2 \left(\sqrt{2}-1\right) \left(2+\sqrt{2}\right)^p-2 \left(1+\sqrt{2}\right) \left(2-\sqrt{2}\right)^p\right)<0 \tag{1}\label{1},$$ and hence $(a+b)^p\not\le a^p+b^p$.


To prove the inequality in \eqref{1}, let $$d_0(p):=d(p)2^{p+3/2},\quad d_1(p):=\frac{d_0'(p)}{(2+\sqrt2)^p},\\ d_2(p):=d_1'(p)(1+1/\sqrt2)^p. $$
Then $$d_2'(p)=\left(1+\sqrt{2}\right) 2^{1-p} \left(2-\sqrt{2}\right)^p \ln\left(3-2 \sqrt{2}\right) \ln\left(2-\sqrt{2}\right) \ln\left(2+\sqrt{2}\right)>0,$$ $d_2$ is increasing (on $(0,1)$), $d_2(1)=\sqrt{2} \ln\left(2-\sqrt{2}\right) \ln\left(6+4 \sqrt{2}\right)<0$, $d_2<0$, $d_1$ is decreasing, $d_1(1)=0$, $d_1>0$, $d_0$ is increasing, $d_0(1)=0$, $d_0<0$, $d<0$. $\quad\Box$


Here is the graph $\{(p,d(p)(1-p)^{-2})\colon0<p<1\}$:

enter image description here

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.