Given a norm $N$ over ${\bf M}_n(\mathbb C)$, it is a natural question to find the best constant $C_N$ such that $$N([A,B])\le C_N N(A)N(B),\qquad\forall A,B\in{\bf M}_n(\mathbb C).$$

The answer is known at least in the following cases:

  • the operator norm $\|A\|_2=\sup\frac{\|Ax\|_2}{\|x\|_2}$ where the norm over $\mathbb C^n$ is the standard Hermitian $\|x\|_2^2=\sum_j|x_j|^2$. Then $$\|[A,B]\|_2\le2\|A\|_2\|B\|_2$$ is optimal for $n\ge2$.
  • the Frobenius norm $\|A\|^2_F=\sum_{i,j}|a_{ij}|^2$. Then a theorem by Böttcher & Wentzel (2008) tells us that $$\|[A,B]\|_F\le\sqrt2\|A\|_F\|B\|_F,$$ and again this is optimal.

I have a third norm in mind, yet of a different nature: the numerical radius $$r(A)=\sup_{x\ne0}\frac{|x^*Ax|}{\|x\|^2}.$$ This is the smallest radius of a disk $D(0;r)$ containing the numerical range (or Hausdorffian) of the matrix. What is the optimal constant $C_{nr}$ such that $r([A,B])\le C_{nr}r(A)r(B)$ for all $A,B$ in ${\bf M}_n(\mathbb C)$ ?

Let me point out that $r$ is not submultiplicative. We have at best $r(MN)\le 4r(M)r(N)$, which gives by the triangle inequality $r([A,B])\le8r(A)r(B)$, but this is certainly not optimal. However, it is a super-stable norm, in the sense that $r(M^k)\le r(M)^k$ for every $k\ge1$.

This question naturally extends to $n$-commutators, in the spirit of my previous question Standard polynomials applied to matrices .

Edit. See below Piotr Migdal's answer and my adaptation of it. It gives $C_{nr}=4$.

  • 1
    $\begingroup$ For random gaussian matrices, some experimental figures: $C_2 = 2.8409$, $C_3=2.4264$, $C_4=2.3061$, etc., that conjecture of mine is probably wrong because of numerical issues. For uniformly random matrices, it seems $C_2 = 1.367$, $C_3=0.8188$, etc.; but since I cannot trust these results, this brings up the question of how to reliably compute / estimate the numerical radius? $\endgroup$ – Suvrit Feb 29 '12 at 3:16
  • $\begingroup$ One has ${\rm Tr}[A,B]=0$. If $n=2$, we may apply to $[A,B]$ the following formula, valid whenever ${\rm Tr}M=0$ : $$4r(M)^2=\|M\|_F^2+2|\det M|.$$ $\endgroup$ – Denis Serre Mar 6 '12 at 17:14

I got $$r([A,B])\leq 4\sqrt{2} r(A) r(B).$$

It is lower than $8$ but still higher than the conjuncture $C_{nr}=4$.

I used the following facts:

  • For normal (i.e. $X^\*X=XX^\*$) matrices we have $r(X)=\sigma_1(X)$ (the largest singular value of X).

  • $\sigma_1(XY-YX)\leq 2 \sigma_1(X)\sigma_1(Y)$

  • Also, note that for $X$ and $Y$ hermitian (i.e. $X^\*= X$ and $Y^\* = Y$ ) we have

$$r(X+iY) \geq \max\left(\sigma_1(X),\sigma_1(Y)\right),$$ $$r(X+iY) \leq \sqrt{\sigma_1^2(X) + \sigma_1^2(Y)}.$$

Lets decompose $A$ and $B$ in their hermitian and antihermitian parts, $$A = A_h + i A_a, \quad B= B_h+i B_a.$$

Then $$r^2([A,B]) \leq \left( \sigma_1([A_h,B_h]-[A_a,B_a])\right)^2 + \left( \sigma_1([A_h,B_a]+[A_a,B_h])\right)^2$$ $$\leq\left( 2\sigma_1(A_h)\sigma_1(B_h)+2\sigma_1(A_a)\sigma_1(B_a) \right)^2 + \left( 2\sigma_1(A_h)\sigma_1(B_a)+2\sigma_1(A_a)\sigma_1(B_h) \right)^2$$ $$=4 ( \sigma_1^2(A_h) + \sigma_1^2(A_a) )( \sigma_1^2(B_h) + \sigma_1^2(B_a) ) +16 \sigma_1(A_h) \sigma_1(A_a) \sigma_1(B_h) \sigma_1(B_a)$$ $$\leq 8( \sigma_1^2(A_h) + \sigma_1^2(A_a) )( \sigma_1^2(B_h) + \sigma_1^2(B_a) )$$ $$\leq 32 r^2(A)r^2(B).$$

  • 1
    $\begingroup$ nice idea to use the cartesian decomposition here. $\endgroup$ – Suvrit Mar 8 '12 at 0:46
  • $\begingroup$ @Piotr. Oops, sorry! Your result seems to be valid in every dimension. let me see it in details. $\endgroup$ – Denis Serre Mar 8 '12 at 6:12
  • $\begingroup$ This could lead to a better bound: Let $k$ be the best constant such that $\sigma_1([X,Y])\le k\sigma_1(X)\sigma_1(Y)$ for every Hermitian matrices $X,Y$. Then your calculation gives a bound with constant $2k\sqrt2$. Does anyone know the constant $k$ ? If we drop the restriction that $X$ and $Y$ are Hermitian, then the best constant is $2$. $\endgroup$ – Denis Serre Mar 8 '12 at 8:02
  • 1
    $\begingroup$ @Denis Even for Hermitian $X$ and $Y$ we get the same constant, $k=2$. Just take $X=[[1,0],[0,-1]]$ and $Y=[[0,1],[1,0]]$. $\endgroup$ – Piotr Migdal Mar 8 '12 at 8:18
  • 2
    $\begingroup$ @Piotr. See below, I have adapted your proof to get the optimal constant $4$. $\endgroup$ – Denis Serre Mar 8 '12 at 12:37

Here is a partial result.

Claim. $4 \le C_{nr} \le 8$.

Proof. The upper bound has already been shown by the OP. The lower-bound follows by \begin{equation*} A = \begin{bmatrix} 0 & 1\\\\ 0 & 0 \end{bmatrix},\qquad B = \begin{bmatrix} 0 & 0\\\\ -1 & 0 \end{bmatrix} \end{equation*} for which $$\frac{r([A,B])}{r(A)r(B)} = \frac{1}{\frac 12\times\frac 12}=4.$$ (Note: Slightly more generally, the $1$ in the above matrices can be replaced by an nonzero scalar).

Based on some experiments mentioned in my comments to Denis, I am led to the following attractive conjecture.

Conjecture: $C_{nr}=4$.

Additional Remarks.

Define $$X := \begin{bmatrix} 0 & 1\\\\ 0 & 0 \end{bmatrix},$$

and let $A$, $B$ be arbitrary. Then, it is easy to see that we have the commutator inequality:

\begin{equation*} r(X \otimes [A,B]) \le 4 r(X \otimes A) r(X \otimes B), \end{equation*} where $\otimes$ denotes the Kronecker product.

  • $\begingroup$ @Suvrit. See my partial answer below. $\endgroup$ – Denis Serre Mar 7 '12 at 14:28

The answer by Piotr Migdal can be modified to give the accurate inequality $$r([A,B])\le4r(A)r(B),\qquad\forall A,B\in{\bf M}_n(\mathbb C).$$ The only new argument is that for every matrix $M$, there exists an angle $\theta$ such that $r(M)=\|{\rm Re}(e^{-i\theta}M)\|_2.$

Actually, we do have $$r(M)=\sup_\alpha\|{\rm Re}(e^{-i\alpha}M)\|_2.$$ Hereabove, the real part is defined as ${\rm Re} N=\frac12(N+\bar N^T)$. Notice that I employ the notation $\|\cdot\|$ (operator norm) which coincides with $\sigma_1$.

Let us apply this to $M=[A,B]$. With $\theta$ as above, let us decompose $e^{-i\theta}A=A_{\theta h}+iA_{\theta a}$. Then let us proceed as Piotr did: $$r([A,B])=\|{\rm Re}[e^{-i\theta}A,B]\| = \|i[A_{\theta h},B_a]+i[A_{\theta a},B_h]\|\le2(\|A_{\theta h}\|\cdot\|B_a\|+\|A_{\theta a}\|\cdot\|B_h\|),$$ which gives $$r([A,B])\le4r(e^{-i\theta}A)r(B).$$ Hence the result.

  • $\begingroup$ @Denis: Maybe it is very simple, but why the equality with the angle $\alpha$ does hold? $\endgroup$ – Piotr Migdal Mar 8 '12 at 14:57
  • $\begingroup$ @Piotr. You take the supremum of a continuous function of $\alpha\in[0,2\pi]$. It is achieved at some $\theta$. $\endgroup$ – Denis Serre Mar 8 '12 at 17:20
  • $\begingroup$ @Denis Actually, I was concerned with $r(M)=\sup_a \Vert\Re(e^{i\alpha}M)\Vert_2$. But now I see it: $r(M) = \sup_\alpha \sup_{|x|=1} \Re (x^\* e^{i\alpha}M x)$ and $\Vert (e^{i \alpha}A + e^{-i \alpha}A^\*)/2 \Vert_2 = \sup_{|x|=1} \Re(x^* e^{i \alpha}A x)$. Nice trick to remove the complex part anyway. $\endgroup$ – Piotr Migdal Mar 9 '12 at 12:42
  • $\begingroup$ And one remark: for a commutator ($X$ and $Y$ are Hermitian), the real part is with $i$ as $(i[X,Y])^\* = -i [Y,X] = i [X,Y]$, whereas $[X,Y]^\* = [Y,X] = - [X,Y]$. $\endgroup$ – Piotr Migdal Mar 9 '12 at 17:21
  • $\begingroup$ @Piotr. Yes! This is always puzzling. Only skew-symmetric matrices form a Lie algebra. $\endgroup$ – Denis Serre Mar 9 '12 at 19:29

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.