Cf. the the title, consider a separable infinite-dimensional Hilbert space H, and the short exact sequence $$0 \to \mathcal{K}(H) \to \mathcal{B}(H) \to \mathcal{Q}(H) \to 0,$$where $\mathcal{K}(H)$ is the compact operators and $\mathcal{Q}(H)$ is the quotient, also known as the Calkin algebra.

Let $u$ be a unitary in $\mathcal{Q}(H)$. This means that $u = T + \mathcal{K}$, for a $T \in \mathcal{B}(H)$ where the differences $TT^* - I$ and $T^*T - I$ are compact. How can one show that $u$ lifts to an isometry, or a co-isometry? That is, why is there an isometry (or co-isometry) $S \in \mathcal{B}(H)$ such that $\pi (S) = u$, where $\pi$ denotes the quotient mapping? (Equivalently $S - T$ is compact)

This is an exercise in Rørdam's book on K-theory. In the previous part of the exercise one shows that whenever $E$ and $F$ are projections in $\mathcal{B}(H)$ with $Rank(E) \leq Rank(F)$ there exists a partial isometry $V$ with $V^*V = E$ and $VV^* \leq F$. I am assuming that one should use this somehow, possibly together with some results about the index map $\delta_1$, but I am really stuck on this one. Any help is appreciated!


Here is an argument that does not appeal to index theory and uses the suggestion in the book.

Let $T$ such that $\pi(T)=u$. Write $T=V|T|$ for the polar decomposition of $T$. Then, $V$ is a partial isometry, i.e. $V^*V$ and $VV^*$ are projections.

Since $\pi(T^*T)=\pi(I)$ and $\pi$ is a $*$-homomorphism, $\pi(|T|)=\pi ((T^*T)^{1/2})=\pi(I)$. It follows that $\pi(V)=\pi(T)$.

We have that $\pi (V) $ is a unitary and $V^*V$ is a projection, so $I-V^*V$ is a compact projection; thus, finite-rank. Similarly with $I-VV^*$.

Now, if $V^*V=I$ or $VV^*=I$, then $V$ is, respectively, an isometry or a co-isometry. If $I-V^*V$ and $I-VV^*$ are both nonzero, we have two cases:

  • $\dim\text{Rank}(I-V^*V)\leq\dim\text{Rank}(I-VV^*)$. By the previous exercise in the book, let $W$ be a partial isometry with $W^*W=I-V^*V$ and $WW^*\leq I-VV^*$. By conjugating the inequality with $V^*$, we get $$ 0=V^*(I-VV^*)V=V^*WW^*V. $$ It follows that $W^*V=0$. Then $V+W$ is an isometry, since $$ (V+W)^*(V+W)=V^*V+W^*W+W^*V+V^*W=V^*V+W^*W=I. $$

  • $\dim\text{Rank}(I-V^*V)\geq\dim\text{Rank}(I-VV^*)$. Similar to the previous case, now the isometry will be $V^*+W^*$.

Finally, note that since $W^*W$ is finite rank, $$0=\pi(W^*W)=\pi(W)^*\pi(W),$$ so $\pi(W)=0$. Then $$ \pi(V+W)=\pi(V)=\pi(T)=u. $$

  • $\begingroup$ Both answers are very nice, but I am accepting this one since it uses the hints in the book, and thus might be more helpful to others. $\endgroup$ – Pjolter Nov 4 '16 at 7:44

If $u \in Q(H)$ is unitary, let $T_1 \in B(H)$ such that $\pi(T_1) = u$, then consider $$ T = T_1R^{-n} $$ where $R$ denotes the left or right shift (which exists on your separable infinite dimensional Hilbert space) so that $T$ has index zero. Now $\exists S_1 \in GL(B(H))$ such that $$ T-S_1 \in K(H) $$ Hence, $S:=S_1R^n$ satisfies $$ T_1 -S\in K(H) $$ and is an isometry or co-isometry.

  • $\begingroup$ This is very nice stuff! Thank you. I do not immediately see why $S_1$ exists, but I will ponder a bit on this. (I have no previous experience with Fredholm operators and index.) $\endgroup$ – Pjolter Nov 3 '16 at 14:57
  • $\begingroup$ If $\text{index}(T) = 0$, then there is an isomorphism $V: \ker(T)\to \ker(T^{\ast})$ whose domain may be extended to all of $H$. This should help you in defining $S_1$. $\endgroup$ – Prahlad Vaidyanathan Nov 4 '16 at 3:45
  • $\begingroup$ Yes, I think I will manage that! $\endgroup$ – Pjolter Nov 4 '16 at 7:49
  • $\begingroup$ I've been looking over this argument again and there is an important point I can't seem to fathom. Where is it used that u is a unitary, i.e. that $I - T_1^*T_1$ and $I - T_1T_1^*$ are compact? To me it seems like this is never used, so this argument would then imply that every element in the Calkin algebra lifts to a (co-)isometry, but this is clearly impossible. $\endgroup$ – Pjolter Nov 11 '16 at 12:15

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.