General question first: upper/lower bound a sum of Kronecker products by its components. More specifically, how is $$ \Vert\sum_{\alpha}S_{\alpha}\otimes B_{\alpha}\Vert$$ bounded by the operator norms of $d_{S}\times d_{S}$ dimensional matrices $\{ S_{\alpha}\}$ and $d_{B}\times d_{B}$ dimensional matrices $\{ B_{\alpha}\}$? The operator norm $\Vert\cdot\Vert$ denotes the (usual) max abs singular value norm. The symbol $\otimes$ denotes the Kronecker or tensor product of matrices. Do we have a Cauchy-like inequality?

More specific question:

We have a "basis" set $\{ S_{\alpha} \}$ of $d_{S}^{2}-1$ Hermitian and traceless matrices: they satisfy $$\text{Tr}(S_{\alpha}S_{\beta}^\dagger)=d_{S}\delta_{\alpha\beta}$$ where $\delta_{\alpha\beta}$ is the Kronecker symbol and are all normalized with respect the operator norm: $\Vert S_{\alpha}\Vert=1$. In short, they would be an orthonormal basis for the Hilbert-Schmidt inner product if the identity matrix was included and they were normalized properly. For $d_s=2$ the $\{S_\alpha\}$ can be taken to be the Pauli matrices. The matrices $B_{\alpha}$ are also Hermitian and traceless but otherwise arbitrary. I would like to bound $\Vert\sum_{\alpha}S_{\alpha}\otimes B_{\alpha}\Vert$ from up and down by a factor of $\max_{\alpha}\Vert B_{\alpha}\Vert$. The best I have done so far is rather miserable:$$ \frac{1}{d_S}\max_{\alpha}{\left\Vert B_{\alpha}\right\Vert }\leq\Vert\sum_{\alpha}S_{\alpha}\otimes B_{\alpha}\Vert\leq(d_S^{2}-1)\max_{\alpha}{\left\Vert B_{\alpha}\right\Vert }$$

The closest reference that I have found on this material is the work by Chansangiam et al [ScienceAsia 35, 106 (2009)] which I am still going through. This question is related to a question in physics for bounding fidelity of evolutions in open quantum systems. In short I want to bound the maximum energy scale in a combined system with that of its components.

  • $\begingroup$ What happens when all the $B_\alpha$ are equal? Your upper-bound seems to be a good one then. $\endgroup$
    – Suvrit
    Feb 8 '11 at 5:46
  • $\begingroup$ @Suvrit: In that case I think the upper bound would be $\sqrt{d_s^2-1}=O(d_s)$ and would be far from tight. The $S_\alpha$ are as spread out as they can be. $\endgroup$ Feb 8 '11 at 15:32
  • $\begingroup$ By the way a lower bound is probably more challenging. $\endgroup$ Feb 8 '11 at 15:33

I do not know if this helps, but here are some upper bounds, standard in operator space theory. The first inequality, attributed to Haagerup, is an analog of the Cauchy-Schwarz inequality in your setting:

$$\Vert\sum_{\alpha}S_{\alpha}\otimes B_{\alpha}\Vert\leq \Vert\sum_{\alpha}S_{\alpha}\otimes \overline{S_{\alpha}}\Vert^{1/2} \Vert\sum_{\alpha}B_{\alpha}\otimes \overline{B_{\alpha}}\Vert^{1/2}.$$

Here for a matrix $A = (A_{i,j})$, $\overline{A}$ denotes the matrix $(\overline{A_{i,j}})$. The expressions appearing on the right-hand side of this inequality are the norms of $(S_\alpha)$ and $(B_\alpha)$ in the operator Hilbert space OH. For a proof, see for example page 123 in Pisier's Introduction to Operator Space Theory.

Another inequality (no longer symmetric) that reduces to the usual Cauchy-Schwarz inequality when the matrices are of size $1$ is the following (and the same with the role of S and B reversed): $$\Vert\sum_{\alpha}S_{\alpha}\otimes B_{\alpha}\Vert\leq \Vert\sum_{\alpha}S_{\alpha}S_{\alpha}^*\Vert^{1/2} \Vert\sum_{\alpha}B_{\alpha}^* B_{\alpha}\Vert^{1/2}.$$

Now the terms appearing on the left are, in the language of operator spaces, the row (resp. column) norm of $(S_\alpha)$ (resp. $(B_\alpha)$). The row (resp. column) norm of $S=(S_\alpha)$ is just the norm of the matrix $ROW(S)$ (resp. $COLUMN(S)$) obtained, in a block-decomposition, by putting $S_\alpha$'s on the first row (resp. column) and $0$'s on the other rows (resp. columns).

This last inequality is very easy to prove, and more generally we have $\Vert\sum_i a_i b_i\Vert\leq \Vert\sum_i a_i a_i^*\Vert^{1/2} \Vert\sum_i b_i^* b_i\Vert^{1/2}$ for any matrices $a_i$ and $b_i$. Indeed, the LHS of this inequality is $\Vert ROW(a) COLUMN(b)\Vert$, and its RHS is $\Vert ROW(a)\Vert \Vert COLUMN(b)\Vert$. This inequality is thus just expressing that the operator norm is sub-multiplicative.

Edit (for a lower bound, without the typo this time). In the case when the $S_\alpha$'s form an orthonormal family for the scalar product $\langle A,B\rangle = Tr(B^* A)/d_S$, you get the following lower bound:

$$\Vert\sum_{\alpha}S_{\alpha}\otimes B_{\alpha}\Vert\geq \max(\Vert\sum_{\alpha}B_{\alpha}^* B_\alpha\Vert^{1/2}, \Vert \sum_{\alpha}B_{\alpha} B_\alpha^*\Vert^{1/2}).$$

This is because $\sum_{\alpha}B_{\alpha}^* B_\alpha$ is $1/d_S Tr \otimes id$ applied to $X^*X$, where $X=\sum_{\alpha}S_{\alpha}\otimes B_{\alpha}$. And since $1/d_s Tr$ is a state, $1/d_S Tr \otimes id$ is of norm $1$ from $M_{d_S} \otimes M_{d_B}$ to $M_{d_B}$.

In the specific situation of your problem, here are the bounds one actually gets. Your assumptions on the $S_\alpha$'s imply that they all are unitary, and orthonormal for $\langle A,B\rangle = Tr(B^* A)/d_S$. Therefore you have that $$\Vert\sum_{\alpha}S_{\alpha}\otimes \overline{S_{\alpha}} \Vert = \Vert\sum_{\alpha}S_{\alpha}^*S_{\alpha}\Vert= \Vert\sum_{\alpha}S_{\alpha}S_{\alpha}^*\Vert = N$$ where $N$ is the number of terms in the sum (you want to take $N=d_S^2-1$).

You therefore get $$ \max(\Vert\sum_{\alpha}B_{\alpha}^* B_\alpha\Vert^{1/2}, \Vert \sum_{\alpha}B_{\alpha} B_\alpha^*\Vert^{1/2}) \leq \Vert\sum_{\alpha}S_{\alpha}\otimes B_{\alpha} \Vert $$ and $$\Vert\sum_{\alpha}S_{\alpha}\otimes B_{\alpha} \Vert \leq \sqrt N \min(\Vert\sum_{\alpha}B_{\alpha}^* B_\alpha\Vert^{1/2}, \Vert \sum_{\alpha}B_{\alpha} B_\alpha^*\Vert^{1/2}).$$

The lower bound is tight (take all the $B_\alpha$'s but one equal to zero). The upper bound too (take $B_\alpha = \overline{S_{\alpha}}$), and implies the one you gave in your question.

  • $\begingroup$ Thanks. I have to check these out... but any ideas about a lower bound esp. for the special problem where S_\alpha are some sort of basis? $\endgroup$ Feb 8 '11 at 15:35
  • $\begingroup$ A last comment: if you really want to compare with $max_\alpha\Vert B_\alpha\Vert$, the best constants one gets is $1$ for the lower bound and $N = d_S^2-1$ for the upper bound, which are both tight. $\endgroup$ Feb 8 '11 at 16:21
  • $\begingroup$ I think the upper bound will scale with $\sqrt{d_s^2-1}$ actually, as the $S_\alpha$ are well spread out in a sense but I have no proof... How do you get the lower bound of 1, by the way? $\endgroup$ Feb 8 '11 at 17:46
  • $\begingroup$ I do not understand your first sentence (but notice that for any unitary matrix $S_1,\dots S_N$, if $B_\alpha = \overline{S_\alpha}$ you get $\Vert \sum_[\alpha=1}^N S_\alpha \otimes B_\alpha \Vert =N$). For your question, the lower bound $1$ follows from my answer and the inequality $\sum B_\alpha^* B_\alpha \geq B_i^* B_i$, which implies $\Vert \sum B_\alpha^* B_\alpha \Vert \geq \Vert B_i^* B_i\Vert = \|B_i\|^2$ for all $i$. $\endgroup$ Feb 8 '11 at 17:55
  • $\begingroup$ Thanks Mikael for the update. Consider a case where $B_\alpha$ are all the same $B$. Then the upper bound must be larger than $\Vert \sum S_\alpha\Vert\Vert B\Vert$ which despite having $d_s^2-1$ term, does not scale with $d_s^2-1$... Or I have no clue! Funny, the way I got my lower bound also used a similar method (we call it partial trace in physics) but I arrived at a different bound! I still need to check things more carefully. $\endgroup$ Feb 8 '11 at 23:18

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.