6
$\begingroup$

This is a problem that I have been stuck for a few months.

Let $X$ be a Hilbert space and $A:B:X\to X$ be two non-commuting semi-positive self-adjoint bounded linear operators. Is it true that $$\|(I+A+B)^{-1}A\|\le 1.$$ If it is, can you suggest any reference to me?


I was able to show $\|(I+A+B)^{-1}A\|\le \|A\|$ and $\|(I+A)^{-1}A\|\le 1$. However, I don't know how to show the desired result without assuming $\|A\|\le 1$ or $AB=BA$.

If it helps, I can show $A=C^*D_1^*D_1C$ and $B=C^*D_2^*D_2C$ for some linear bounded $C,D_1,D_2$.

$\endgroup$
9
$\begingroup$

The claim is false. Consider the following matrix argument.

\begin{eqnarray*} \|(I+A+B)^{-1}A\| \le 1\quad\Leftrightarrow\quad \begin{bmatrix}I & (I+A+B)^{-1}A \\ A(I+A+B)^{-1} & I \end{bmatrix} \ge 0. \end{eqnarray*} The latter inequality amounts to showing $I \ge (I+A+B)^{-1}A^2(I+A+B)^{-1}$, or in other words, we must show that $(I+A+B)^2 \ge A^2$. Since the map $x \mapsto x^2$ is not operator monotone, suggets that it should be possible to make this inequality fail.

And indeed, here is an explicit counterexample: \begin{equation*} A = \begin{bmatrix}1 &5\\ 5 &61 \end{bmatrix},\quad B = \begin{bmatrix}9 &-9\\ -9 & 13\end{bmatrix},\quad C = (I+A+B)^2-A^2 = \begin{bmatrix}111 &-654\\ -654& 1895 \end{bmatrix}. \end{equation*} The matrix $C$ has a negative eigenvalue: $1003 - 2\sqrt{305845} < 0$.

$\endgroup$
8
$\begingroup$

(Unless I messed up the arithmetic, here's a counterexample.)

Let $B=\begin{bmatrix} 1 & 1 \\ 1 & 1 \end{bmatrix}$ and $A=\begin{bmatrix} 1 & 0 \\ 0 & R \end{bmatrix}$ where $R$ is a positive number that we'll choose at the end. Then $(1+A+B)^{-1}=\begin{bmatrix} 3 & 1 \\ 1 & R+2 \end{bmatrix}^{-1}=1/(3R+5)\begin{bmatrix} R+2 & -1 \\ -1 & 3 \end{bmatrix}.$

Then $(1+A+B)^{-1}A=1/(3R+5)\begin{bmatrix} R+2 & -R \\ -1 & 3R \end{bmatrix}.$

Then $||(1+A+B)^{-1}A ||\geq 1/(3R+5)|| \begin{bmatrix} 0 & -R \\ 0 & 3R \end{bmatrix} ||=\frac{\sqrt{10}R}{3R+5}$

Then just take $R$ big enough.

$\endgroup$
5
$\begingroup$

Though two answers are already posted, let me explain how to understand that it is false from general reasoning. The inequality $\|(I+A+B)^{-1}A\|\leqslant 1$ is equivalent to $\|(I+A+B)^{-1}A x\|\leqslant \|x\|$ for any vector $x$. Take $x=A^{-1}(I+A+B)y$, the inequality rewrites as $\|y\|\leqslant \|y+A^{-1}(I+B)y\|$. We are given only that $A,B$ are positive definite, thus for given vector $z$, $A^{-1}z$ may be any vector which forms an acute angle with $z$. For $z=(I+B)y$ this in general allows the vector to form an obtuse angle with $y$ (unless $y$ is an eigenvector of $B$.) So, if the direction of $A^{-1}(I+B)y$ forms an obtuse angle with $y$ and $A^{-1}$ is small enough, the inequality $\|y\|\leqslant \|y+A^{-1}(I+B)y\|$ fails.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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