Exercise :

Let $H$ be a Hilbert space and $A,B \in \mathcal{L}(H)$ be self-adjoint operators with $0 \leq A \leq B$ and $B \in \mathcal{L}_c(H)$. Show that $A \in \mathcal{L}_c(H)$.

Thoughts :

Relying only on the definition of a compact operator, we essentialy need to conclude that $A$ transfers bounded sets to relatively compact sets (compact closure).

Now, since $B$ is compact and self adjoint, I know that also $B^*B$ is compact. This may be of use since the property of $A$ and $B$ being self adjoint is noted in the exercise.

I think that $A \leq B \implies \|A\| \leq \|B\|$ since they are both bounded and we could take $\mathbf{1} \in H$ which yields that $$\|A(\mathbf{1})\| \leq \|A\|\|1\| \equiv \|A\| \quad \text{and} \quad \|B(\mathbf{1})\| \leq \|B\|\|1\| \equiv \|B\|$$ and since $0 \leq A \leq B$ implies that their values follow the inequality for any $x \in H$ thus the implied result.

Request : Beyond these points, I sadly do not have an intuition for a head-start, so I would really appreciate any hints or elaborations.

  • $\begingroup$ I think that this can be proven by using the eigenvalue decomposition of the compact, self-adjoint operator $B$. However, I feel that there is a simpler solution. $\endgroup$ – gerw May 11 at 11:09

I can think of two arguments. None of them is entirely elementary, though, in the sense that they depend on having square roots available.

  • Let $\{x_n\}$ be a sequence in the unit ball of $H$. Since $B$ is compact, there exists a subsequence $\{x_{n_j}\}$ such that $\{Bx_{n_j}\}$ converges. Now (using twice that $A$ is positive) \begin{align} \|A(x_{n_j}-x_{n_k})\|^2&=\langle A^2(x_{n_j}-x_{n_k}),(x_{n_j}-x_{n_k})\rangle \leq \|A\|\,\langle A(x_{n_j}-x_{n_k}),(x_{n_j}-x_{n_k})\rangle\\ &\leq \|A\|\,\langle B(x_{n_j}-x_{n_k}),(x_{n_j}-x_{n_k})\rangle\to0. \end{align} So $\{Ax_{n_j}\}$ is Cauchy and thus convergent. So $A$ is compact.

  • Since $0\leq A\leq B$, there exists $C$ with $A^{1/2}=CB^{1/2}$. Since $B$ is compact, so is $B^{1/2}$; then $A^{1/2}$ is compact and so is $A$.

The first argument depends on the inequality $A^2\leq \|A\|\,A$ for $A$ positive. This can be seen for instance by $$ \langle A^2x,x\rangle=\langle Ax,Ax\rangle=\|Ax\|^2=\|A^{1/2}A^{1/2}x\|^2\leq\|A^{1/2}\|^2\|A^{1/2}x\|^2 =\|A^{1/2}\|^2\langle Ax,x\rangle. $$ And $\|A^{1/2}\|^2=\|A\|$.

For the second argument, we first note that if $B^{1/2}x=0$, then $A^{1/2}x=0$. So we have $H=H_1\oplus H_1^\perp$, where $H_1=\overline{\operatorname{ran}B^{1/2}}$. We have $$\tag1 \|A^{1/2}x\|^2=\langle Ax,x\rangle\leq \langle Bx,x\rangle=\|B^{1/2}x\|^2, $$ so $A^{1/2}=0$ on $H_1^\perp$. On $\operatorname{ran}B^{1/2}$, we define $$ CB^{1/2}x=A^{1/2}x. $$ This is well defined by $(1)$. We have $$ \|CB^{1/2}x\|^2=\|A^{1/2}x\|^2=\langle Ax,x\rangle\leq\langle Bx,x\rangle=\|B^{1/2}x\|^2, $$ so $C$ is bounded on $\operatorname{ran}B^{1/2}$, so it extends by density to $H_1$. We put $C=0$ on $H_1^\perp$. So $C$ is bounded and $CB^{1/2}=A^{1/2}$.

  • $\begingroup$ Great answer as well, I really like the second one as well, since in a previous exercise I had proven that $B$ is compact if and only if $B^{1/2}$ is compact. My question is this though : We say that if $0\leq A \leq B$ then there exists C such that $A^{1/2} = CB^{1/2}$. Why use the square root operator ? Couldn't we imply that there exists a $C$ such that $A = CB$ ? $\endgroup$ – Rebellos May 12 at 7:15
  • 1
    $\begingroup$ For starters, the proof wouldn't work! Yes, in finite dimension, one can always get $A=CB$; but, as opposed to the construction above, you cannot guarantee that $C$ is a contraction. In infinite dimension, I'm not sure. $\endgroup$ – Martin Argerami May 12 at 8:02
  • $\begingroup$ Okay, got it ! It really wasn't as simple an exercise as it thought it would be, but your guidance (and also DisintegrateByParts' guidance) has taught me some very nice elaborations and handlings. I couldn't appreciate it more ! $\endgroup$ – Rebellos May 12 at 8:07

Because $B-A \ge 0$, then $[x,y]=\langle (B-A)x,y\rangle$ is a pseudo inner product, lacking only positive definiteness. As such, the Cauchy-Schwarz inequality holds $$ |[x,y]|^2 \le [x,x][y,y] \\ |\langle (B-A)x,y\rangle|^2 \le \langle (B-A)x,x\rangle\langle (B-A)y,y\rangle. $$ Now set $y=(B-A)x$ in the above to obtain $$ \|(B-A)x\|^4 \le \langle (B-A)x,x\rangle\langle(B-A)(B-A)x,(B-A)x\rangle \\ \|(B-A)x\|^4 \le \langle (B-A)x,x\rangle\|B-A\|\|(B-A)x\|^2 \\ \|(B-A)x\|^2 \le \|B-A\|\langle(B-A)x,x\rangle \\ \|(B-A)x\|^2 \le \|B-A\|\langle Bx,x\rangle. $$ Suppose $\{ x_n \}$ is a bounded sequence. Because $B$ is compact, there is a subsequence $\{ x_{n_k} \}$ such that $\{ Bx_{n_k} \}$ converges. By the above, $\{ (B-A)x_{n_k}\}$ is a Cauchy sequence and, hence, converges to some $y$. But $\{ Bx_{n_k} \}$ also converges. Therefore $\{ Ax_{n_k} \}$ converges, leading to the conclusion that $A$ is compact.

  • $\begingroup$ Great answer. The pseudo-inner product statement, we never mentioned something like this. Could I still make such claim, lets say ? I do understand that if it was $B-A >0$ it would be a standard inner product process with no issues. $\endgroup$ – Rebellos May 12 at 7:10
  • $\begingroup$ @Rebellos : If it is not definite, then consider $B-A+\epsilon I$, obtain the inequality, and then let $\epsilon \downarrow 0$. You'll have the inequality you want. $\endgroup$ – DisintegratingByParts May 12 at 15:43
  • $\begingroup$ Thanks for clearing up. The answer is great, I only accepted the other (great one as well), since its closer to the handlings I am familiar with. $\endgroup$ – Rebellos May 12 at 15:48

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.