Suppose that $x,y\in\mathbb H$, where $\mathbb H$ is a separable Hilbert space. Define the tensor $x\otimes y:\mathbb H\to\mathbb H$ by setting $(x\otimes y)(z)=\langle z,y\rangle x$ for $z\in\mathbb H$. We can also define $A=x\otimes y\tilde\otimes x\otimes y: S_2(\mathbb H)\to S_2(\mathbb H)$ in an analogous way, where $S_2(\mathbb H)$ is the space of the Hilbert-Schmidt operators, which is also a Hilbert space. I am interested in the trace of $A$. On one hand, $$ \operatorname{Tr}A=\operatorname{Tr}[x\otimes y\tilde\otimes x\otimes y]=\operatorname{Tr}[x\otimes y]\operatorname{Tr}[x\otimes y]=|\langle x,y\rangle|^2. $$ On the other hand, \begin{align*} \operatorname{Tr}A &=\sum_{k,l\ge1}\langle{A(e_k\otimes e_l),e_k\otimes e_l}\rangle_{\mathrm{HS}}\\ &=\sum_{k,l\ge1}\langle\langle e_k\otimes e_l,x\otimes y\rangle_{\mathrm{HS}}(x\otimes y),e_k\otimes e_l\rangle_{\mathrm{HS}}\\ &=\sum_{k,l\ge1}\langle x\otimes y,e_k\otimes e_l\rangle_{\mathrm{HS}}\langle e_k\otimes e_l,x\otimes y\rangle_{\mathrm{HS}}\\ &=\|x\otimes y\|_{\mathrm{HS}}^2\\ &=\|x\|^2\|y\|^2. \end{align*} Where am I making a mistake? Which is the correct answer?

Any help is much appreciated!

  • $\begingroup$ For Hilbert-Schmidt operators $S,T$ one defines the scalar product by $<S,T>_{HS}=Tr (ST^{*})$ and so if You view $x\otimes y$ as a Hilbert-Schmidt operator ($S_2(\mathbb{H})\cong \mathbb{H}\otimes\mathbb{H}$) then $||x\otimes y||_{HS}^2=Tr ((x\otimes y)(x\otimes y)^{*})$. I can't see how this is equal to $||x||^2||y||^2$ $\endgroup$ Sep 24, 2018 at 11:52
  • 1
    $\begingroup$ @PeterMelech $\|x\otimes y\|_{\mathrm{HS}}^2=\sum\|(x\otimes y)(e_k)\|^2=\|x\|^2\sum|\langle e_k,y\rangle|^2=\|x\|^2\|y\|^2$ (see here). $\endgroup$
    – Cm7F7Bb
    Sep 24, 2018 at 11:56
  • $\begingroup$ Sure! I can't find a mistake either. Good question! $\endgroup$ Sep 24, 2018 at 11:59

1 Answer 1


Alright to first clear up how maps of the form $(x\otimes y)\tilde\otimes(z\otimes w)$ for arbitrary $x,y,w,z\in\mathbb H$ act on general $\varphi\in S_2(\mathbb H)$, it goes like $$ [(x\otimes y)\tilde\otimes(z\otimes w)](\varphi):=\langle \varphi,z\otimes w\rangle_\text{HS} (x\otimes y) $$ as defined by OP in another question. One readily verifies $\langle \varphi,z\otimes w\rangle_\text{HS} =\langle \varphi(w),z\rangle_{\mathbb H}$. The above map admits the trace $$ \operatorname{tr}[(x\otimes y)\tilde\otimes(z\otimes w)]=\sum_{k,l\geq 1}\langle [(x\otimes y)\tilde\otimes(z\otimes w)](e_k\otimes e_l),e_k\otimes e_l\rangle_\text{HS}=\sum_{k,l\geq 1} \langle (e_k\otimes e_l)(w),z\rangle_{\mathbb H}\langle x,(e_k\otimes e_l)(y)\rangle_{\mathbb H}=\sum_{k,l\geq 1} \langle w,e_l\rangle_{\mathbb H}\langle e_k,z\rangle_{\mathbb H}\langle x,e_k\rangle_{\mathbb H}\langle e_l,y\rangle_{\mathbb H}=\langle w,y\rangle_{\mathbb H}\langle x,z\rangle_{\mathbb H} $$ by basis expansion. On the other hand, $\operatorname{tr}(a\otimes b)=\langle a,b\rangle_{\mathbb H}$ for all $a,b\in\mathbb H$ (can also be easily seen by basis expansion). Piecing things together, one gets the somewhat peculiar identity $$ \boxed{\operatorname{tr}[(x\otimes y)\tilde\otimes(z\otimes w)]=\operatorname{tr}(x\otimes z)\operatorname{tr}(w\otimes y)} $$ so the trace identity in your answer is not correct. This is solely due to the above definition of the tensor product $\tilde\otimes$ to go from $S_2(\mathbb H)$ to operators on $S_2(\mathbb H)$ which does not coincide with the usual tensor product (as in that case one would in fact have $\operatorname{tr}(A\tilde\otimes B)=\operatorname{tr}(A)\operatorname{tr}(B)$ given $A,B$ are trace-class). To see how $(x\otimes y)\tilde\otimes (x\otimes y)$ should behave on $\mathbb H\otimes\mathbb H$ to mimic the usual tensor product of linear Hilbert space operators, check my old answer below.


To identify the Hilbert-Schmidt-Operators with $\mathbb H\otimes\mathbb H$ (as Peter Melech pointed out in his comment) and to not raise confusion regarding the actual Hilbert space tensor product $\otimes_{\mathbb H}=:\otimes$, I will "rename" your map $x\otimes y$ to $\langle \cdot,y\rangle x:\mathbb H\to\mathbb H$ (but the action is the same!). Consider the map $$ A=(\langle \cdot,y\rangle x)\otimes(\langle \cdot,y\rangle x):\mathbb H\otimes \mathbb H\to\mathbb H\otimes \mathbb H $$ defined via $$ a\otimes b\mapsto (\langle a,y\rangle x)\otimes(\langle b,y\rangle x)=\langle a,y\rangle\langle b,y\rangle\, x\otimes x=\langle a\otimes b,y\otimes y\rangle\, x\otimes x $$ and its linear extension onto all of $\mathbb H\otimes\mathbb H$, i.e., $$\boxed{A(z)=\langle z,y\otimes y\rangle x\otimes x}$$ for all $z\in\mathbb H\otimes\mathbb H$. [Edit: Thus in your notation and given your definition, the identity $$(x\otimes y)\tilde\otimes(x\otimes y)=(x\,\tilde\otimes \,x)\otimes(y\,\tilde\otimes \,y)$$ has to hold to mimic the usual tensor product of linear operators.] Now on the one hand, $$ \operatorname{tr}(A)=\operatorname{tr}(\langle \cdot,y\otimes y\rangle\, x\otimes x)=\langle x\otimes x,y\otimes y\rangle=\langle x,y\rangle^2\,. $$ On the other hand, let $\lbrace e_k\rbrace_{k\geq 1}$ be an orthonormal basis of $\mathbb H$, then $\lbrace e_k\otimes e_l\rbrace_{k,l\geq 1}$ is an orthonormal basis of $\mathbb H\otimes\mathbb H$. By definition of the trace $$ \operatorname{tr}(A)=\sum_{k,l\geq 1}\langle A(e_k\otimes e_l),e_k\otimes e_l\rangle= \sum_{k,l\geq 1}\langle e_k\otimes e_l,y\otimes y\rangle\langle x\otimes x,e_k\otimes e_l\rangle\\ =\Big\langle x\otimes x, \sum_{k,l\geq 1}\langle y\otimes y,e_k\otimes e_l\rangle e_k\otimes e_l\Big\rangle=\langle x\otimes x,y\otimes y\rangle=\langle x,y\rangle^2 $$ where in the last step we just used the basis expansion in $\mathbb H\otimes\mathbb H$.

  • $\begingroup$ Did'nt You change from $A=x\otimes y \tilde{\otimes}x\otimes y$ to $x\otimes x\tilde{\otimes}y\otimes y$? $\endgroup$ Sep 25, 2018 at 13:34
  • $\begingroup$ Curiously for that map You could as well compute $Tr(x\otimes x\tilde{\otimes}y\otimes y)=Tr(x\otimes x)Tr(y\otimes y)=<x,x><y,y>=||x||^2||y||^2$! $\endgroup$ Sep 25, 2018 at 13:48
  • $\begingroup$ Yes I guess in OP's notation the key result I used is $A=(x\otimes y)\tilde\otimes(x\otimes y)=(x\tilde\otimes x)\otimes(y\tilde\otimes y)$ ($\neq (x\otimes x)\tilde\otimes(y\otimes y)$ (!)). But this is why I personally am not too fond of this notation - it at least for me is quite easy to confuse $x\otimes y$ (the $\langle\cdot,y\rangle x$ pairing) and actual tensor products $x\tilde\otimes y$ in $\mathbb H\tilde\otimes\mathbb H$... $\endgroup$ Sep 25, 2018 at 16:07
  • $\begingroup$ However the OP's map acts like $A(e_k\otimes e_l)=<e_k\otimes e_l,x\otimes y>x\otimes y$ whereas Your map acts like $A(e_k\otimes e_l)=<e_k\otimes e_l,y\otimes y>x\otimes x$ and the point is that for both(!) maps one gets two different results for the trace using the two different methods. I don't think it is a matter of misleading notation, there must be some more subtle mistake. $\endgroup$ Sep 26, 2018 at 8:39
  • 1
    $\begingroup$ Your peculiar identity looks quite strange (but is true I think) and it is really counterintuitive that the "mutiplicity" of the trace with the "tensor product" does not hold, as well it is unclear what a usual tensor product could be. However I agree that the mistake lies exactly here:+1 $\endgroup$ Sep 27, 2018 at 10:35

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.