Let $A$ be a C*-algebra. Suppose that $P:A \rightarrow A$ is a contractive completely positive projection. Does the range $P(A)$ is completely order isomorphic to a $C^*$-algebra?

In the case where the answer is false: Is it true in supposing, in addition, that the range $P(A)$ is an operator system, $A$ unital and $P(1)=1$?


Yes. $P(A)$ is abstractly a C$^\ast$-algebra $B$ under the Choi-Effros product: $a\circ b := P(ab)$; and $P(A)$ is completely order isomorphic to $B$. See (the proof of) Theorem 3.1 in [M.-D. Choi and E. G. Effros; Injectivity and operator spaces. J. Functional Analysis 24 (1977), 156-–209]. Moreover, there is a surjective $\ast$-homomorphism from $C^\ast(P(A))$ onto the C$^\ast$-algebra $B$.


I would add to Taka's answer, since the Choi-Effros paper is not widely available online, that the nontrivial parts of the proof are to show the "C$^*$-identity" $\|x\|^2=\|x^*x\|$, associativity, and to check the completeness of $P(A)$.

Indeed, the algebra $(P(A),+,\circ)$ is a $*$-algebra, normed with the norm inherited from $A$. For the C$^*$-identity, since $P$ is contractive we have $$ \|x^*\circ x\|=\|P(x^*x)\|\leq\|x^*x\|=\|x\|^2. $$ On the other hand, since $P$ is cp and contractive, it satisfies the Schwarz inequality $P(x)^*P(x)\leq P(x^*x)$, and so, for any $x\in P(A)$, $$ \|x^*\circ x\|=\|P(x^*x)\|\geq \|P(x)^*P(x)\|=\|x^*x\|=\|x\|^2. $$

Completeness follows from the fact that $P$ is a bounded projection. If $\{x_j\}$ is a Cauchy sequence in $P(A)$, then by the completeness of $A$ the sequence converges to some $x$ in $A$. As $P$ is bounded, $P(x_j)\to P(x)$; but $P(x_j)=x_j$ (since $P$ is a projection and $x_j\in P(A)$) and so $P(x)=x$, that is $x\in P(A)$. So $P(A)$ is closed.

For the associativity of the product, one needs to check that $P(aP(bc))=P(P(ab)c)$ for any $a,b,c\in A$. For this, since $P$ is selfadjoint, it is enough to show that $P(xP(a))=P(xa)$ for any $x\in P(A)$, $a\in A$. The trick that Choi-Effros use (in Paulsen's version here) is to apply the Schwarz inequality to the ccp map $P^{(2)}$ and the operator $y=\begin{bmatrix}x^*&a\\ 0&0\end{bmatrix}$: thus $P^{(2)}(y)^*P^{(2)}(y)\leq P^{(2)}(y^*y)$ reads $$ \begin{bmatrix}xx^*&xP(a)\\ P(a^*)x^*&P(a^*)P(a) \end{bmatrix} \leq \begin{bmatrix}P(xx^*)&P(xa)\\P(a^*x^*)&P(a^*a)\end{bmatrix} $$ If we now apply $P^{(2)}$ to this inequality, since it preserves positivity, we get $$ \begin{bmatrix} 0&P(xa)-P(xP(a))\\ P(a^*x^*-P(P(a^*)x^*))& P(a^*a)-P(P(a^*)P(a))\end{bmatrix} \geq0; $$ the 0 in the 1,1 entry forces the off-diagonal entries to be zero, so $P(xa)=P(xP(a))$.

  • 4
    $\begingroup$ This is unbelievable... 15 minutes after I posted here the phrase "the Choi-Effros" paper is not widely available online" (which is still probably the case), one of my colleagues, for reasons completely unrelated to this question, emailed me a copy of this very same paper! $\endgroup$ Oct 14 '10 at 6:25
  • $\begingroup$ You are welcome! I think this is an awesome result. $\endgroup$ Oct 16 '10 at 2:35
  • 2
    $\begingroup$ For future visitors to this page: it seems that the Choi-Effros paper is now available via "Open Archive" online: dx.doi.org/10.1016/0022-1236%2877%2990052-0 $\endgroup$ Jun 9 '14 at 18:53

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.