5
$\begingroup$

Denote by $\mathbb{P}_n$ be the set of real symmetric positive definite $n \times n$ matrices. In $\mathbb{P}_n$, we define an partial order as follow: for $X, Y \in \mathbb{P}_n$, we say $X \prec Y$ if $Y-X$ is also a positive definite matrix. Let $\lambda = 2^{1-1/p}$ with $p>1$ be a real number.

Question: For $0\prec X \prec Y \prec \lambda X$, do there exist positive definite matrices $A, B$ such that

$$X = \dfrac{A+B}{2}, \qquad Y = \left(\dfrac{A^p+B^p}{2}\right)^{1/p}?$$


Remark: The answer in the case of a scalar is Yes, and $\lambda = 2^{1-1/p}$ comes from this case to ensure that the function $$f(a) = \left(\dfrac{a^p + (2x-a)^p}{2}\right)^{1/p}$$ is surjective from $[0,2x]$ to $[x, \lambda x]$.

$\endgroup$
1

1 Answer 1

10
$\begingroup$

Edit: The original answer only dealt with $p=2$, now the case $1<p<\infty$ is also dealt with. I leave the $p=2$ case as it is simpler.

No. A positive answer to your question for $p=2$ would in particular imply that $$ 0<X<Y<\lambda X \implies X^2 \leq Y^2,$$ which is well-know to be false (the square map is not operator monotone), see below for an explicit counterexample.

Indeed, if $$X = \dfrac{A+B}{2}, \qquad Y = \left(\dfrac{A^2+B^2}{2}\right)^{1/2},$$ then we have $Y^2 - X^2 = \dfrac{(A-B)^2}{4}$.

For a concrete counterexample, take $X=\begin{pmatrix} 1 &0\\0& \frac 1 4\end{pmatrix}$ and $Y = X+ t \begin{pmatrix} 1&1\\1&1\end{pmatrix}$. For small positive $t$, the inequalities $0<X<Y<\lambda X$ are satisfied, but $Y^2-X^2$ has determinant $- (3t/4)^2<0$, so is never positive.


When $1<p<\infty$, we can apply a similar argument. We shall use the following theorem by Loewner (I think): the map $t \mapsto t^r$ is operator concave if $0<r \leq 1$, and operator convex for $1\leq r \leq 2$. See for example Bhatia's book on matrix analysis.

So if we take some $1<r \leq \min(2,p)$, we have for every positive $A,B$, and $$X = \dfrac{A+B}{2}, \qquad Y = \left(\dfrac{A^p+B^p}{2}\right)^{1/p},$$ $$ X^r \leq \frac{A^r+B^r}{2} \leq \left(\frac{A^p+B^p}{2}\right)^{r/p} = Y^{r}.$$ (the first inequality is the operator convexity of $t^r$, the second is the operator concavity of $t^{r/p}$.

So a positive answer to your question would imply that $$ 0<X<Y<\lambda X \implies X^r \leq Y^r,$$ which should contradict the non-operator monotonicity of $t^r$. Something needs to be checked when $r<2$ as the non-operator monotonicity of $t^r$ only says that $0<X<Y \implies X^r \leq Y^r$ is not true, but I guess (did not check) that the same concrete counterexample as for $r=2$ above should work.

$\endgroup$
5
  • $\begingroup$ Thank you very much. Do you believe that the answer is NO for all $p > 1$? Your answer gives us that the answer is NO for all even integer $p$. $\endgroup$
    – The Nguyen
    Mar 24, 2021 at 8:46
  • $\begingroup$ @NguyễnVănThế I do not know. But if I had to guess, I would say that the answer is no for all $p>1$. $\endgroup$ Mar 24, 2021 at 9:12
  • $\begingroup$ Oh, I do have an argument for $1<p\leq 2$, I think. $\endgroup$ Mar 24, 2021 at 9:15
  • $\begingroup$ Oh, I really hope to see it. $\endgroup$
    – The Nguyen
    Mar 24, 2021 at 9:21
  • $\begingroup$ How about if we replace arithmetic mean with geometric mean, namely, do there exist $A, B > 0$ such that $ X = A \sharp B$ and $Y = \left(\dfrac{A^p+B^p}{2}\right)^{1/p}$ with $p > 1$? $\endgroup$
    – The Nguyen
    Apr 8, 2021 at 9:16

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.