Let for $i\in [n]$, $P_i$ be some orthogonal projectors defined on the Hilbert space $W$ such that they commute on subspace $V < W$ (i.e, for any $i, j \in [n]$ and $v \in V$: $P_iP_j(v) = P_jP_i(v)$). Are there $n$ projectors $\hat{P}_i$ such that they operate the same as $P_i$ on $V$ (i.e, for any $i\in [n]$ and $v\in V$, $\hat{P}_i(v)= P_i(v)$) and they commute on the Hilbert space $W$?
The generalization of commutative property of orthogonal projectors on a subspace to the whole space
-
$\begingroup$ Is $W$ a Hilbert space? Otherwise I don't know what you mean by "orthogonal projector". $\endgroup$– Nik WeaverCommented Nov 15, 2017 at 18:07
-
$\begingroup$ Is $V$ finite dimensional? (Or is $W$ closed?) $\endgroup$– Alex DegtyarevCommented Nov 15, 2017 at 19:15
1 Answer
This works for two projections, and then it follows from Halmos' two projections theorem. I originally thought the same argument would address the general case, but as Dominique pointed out, that's not so clear.
This says that if $P$ and $Q$ project onto $M$ and $N$, respectively, then we can decompose the Hilbert space as $$ W = (M\cap N) \oplus (M\cap N^{\perp}) \oplus (M^{\perp}\cap N) \oplus (M^{\perp}\cap N^{\perp}) \oplus (S\oplus T) , $$ and this reduces both $P$ and $Q$, and after a unitary transformation, the action on the last two components (if present) is $$ P = \begin{pmatrix} 1 & 0 \\ 0 & 0 \end{pmatrix}, \quad Q = \begin{pmatrix} 1-H & W \\ W & H \end{pmatrix} , $$ with $0<H<1$ and $W=(H(1-H))^{1/2}$. (This of course generalizes the trivial fact that a projection in $\mathbb C^2$ looks like $\left( \begin{smallmatrix} \cos^2\alpha & \sin\alpha\cos\alpha \\ \sin\alpha\cos\alpha & \sin^2\alpha\end{smallmatrix}\right)$.)
Now if $PQv=QPv$ and $v$ has a component $(v_1,v_2)\in S\oplus T$, then it follows that $Wv_1=Wv_2=0$, so $v_1=v_2=0$. In other words, $V$ must be contained in the sum of the first four summands, and thus the desired modifications can be obtained by simply replacing the part on $S\oplus T$ by zero.
-
$\begingroup$ I had no idea this was due to Halmos! I always considered it folklore. $\endgroup$ Commented Nov 16, 2017 at 3:23
-
$\begingroup$ @NikWeaver: The article I linked to says they'll call it the Halmos 2 projections theorem because Halmos's presentation is so well written, but many others had this or similar versions before... $\endgroup$ Commented Nov 16, 2017 at 3:51
-
$\begingroup$ Oh, I see. Well, Halmos's writing is unusually clear. $\endgroup$ Commented Nov 16, 2017 at 12:26
-
$\begingroup$ I do not think the generalization to many projectors works with the same argument. Say $P,Q,R$ pairwise commute on $V$. Then, using your proof, we get $\hat P,\hat Q$ such that $\hat P=P$ on $V$ and $\hat Q=Q$ on $V$, and $\hat P,\hat Q$ commute. Now we apply want to apply the argument to the next pair, say $P,R$. But if we apply it to $P,R$, we get a different $\hat P$ that might not commute with $\hat Q$ any more. Alternatively, we may try to apply the argument to $\hat P,R$. But that doesn't work because we don't know whether $\hat PRv=R\hat Pv$ for all $v\in V$. $\endgroup$ Commented Nov 16, 2017 at 12:35
-
$\begingroup$ @DominiqueUnruh: Yes, I was sloppy. I thought originally for $n$ projections $v$ would still have to be in the sum of the now $2^n$ summands formed with the spaces we project on, but it's not so clear how that would follow or if it's true. $\endgroup$ Commented Nov 16, 2017 at 17:38