Let $n \in \mathbb{N}$ and $p \in [1,\infty]$ be fixed and endow $\mathbb{C}^n$ with the $p$-norm $\|\cdot\|_p$. For every matrix $A \in \mathbb{C}^{n \times n}$ we denote the operator norm of $A$ as an operator on $\mathbb{C}^n$ by $\|A\|_p$, too. Moreover, let $|A|$ denote the matrix whose entries are the absolute values of the entries of $A$.

The number $\| \,|A|\,\|_p$ is sometimes called the regular norm of $A$ (in particular in Banach lattice theory, where a norm on the space of regular operators is constructed this way).

We clearly have $\|A\|_p \le \| \,|A|\,\|_p$, and equality holds for $p = 1$ and $p = \infty$. For general $p$, Mark Meckes explained in his answer to this question that the Riesz-Thorin theorem implies the estimate \begin{align*} \| \,|A|\,\|_p \le c_{n,p} \|A\|_p, \tag{E} \end{align*} where $c_{n,p} = n^{\frac{2}{p}(1 - \frac{1}{p})}$.

Question: Is the estimate (E) sharp? If not, what is the optimal constant for the estimate (depending on $n$ and $p$)?


  • Clearly, the inequality (E) is sharp (and actually an equality) for $p = 1$ and $p = \infty$.

  • Let $U \in \mathbb{C}^{n\times n}$ be such that $\frac{1}{\sqrt{n}}U$ is a unitary matrix and such that each entry of $U$ has modulus $1$ (such a matrix exists for every $n$; for instance, we can choose $U$ to be the transformation matrix of the Fourier transform on $\mathbb{Z}/n\mathbb{Z}$). Then one can prove by employing Hölder's inequality several times that $\|\, |U| \, \|_p = n$ for each $p \in [1,\infty]$, and that $\|U\|_p = n^{1/p}$ for $p \in [1,2]$, while $\|U\|_p = n^{1-1/p}$ for $p \in [2,\infty]$. Thus, we obtain $\|\,|U|\,\|_p = \tilde c_{n,p} \|U\|_p$, where \begin{align*} \tilde c_{n,p} = \begin{cases} n^{1-1/p} \quad & \text{for } p \in [1,2], \\ n^{1/p} \quad & \text{for } p \in [2,\infty]. \end{cases} \end{align*} We have $\tilde c_{n,p} \le c_{n,p}$, and the existence of $U$ shows that the optimal constant for (E) is somewhere in between.

  • Note that $\tilde c_{n,2} = c_{n,2} = n^{1/2}$, so we know that the estimate (E) is sharp for $p=2$; this is also mentioned in Mark Meckes' answer quoted above.

Note: A loosely related question concerned with the Schatten $p$-norm of a matrix can be found here.

Edits made 2018-06-03:

  • I've replaced the real scalar field with $\mathbb{C}$ since this seems to be more appropriate for the question.

  • I've added the example matrix $U$ to show that the number $\tilde c_{n,p}$ is a lower bound for the optimal constant.


It is trivially $n^{1/p}$ for $p>2$ (and, therefore, $n^{1-\frac 1p}$ for $1<p<2$ because the norm of the adjoint operator in the dual space is the same as the norm of the operator in the space itself). To see it, choose the vector $x$ of $\ell^p$ norm $1$ so that $$ \sum_i\left|\sum_j |a_{ij}|x_j\right|^p=\||A|\|^p $$ Now choose $i$ so that $$ \left|\sum_j |a_{ij}|x_j\right|^p\ge \frac 1n\||A|\|^p $$ Rotate $x_j$ to $y_j$ preserving the absolute values so that for this particular $i$, we have $$ \sum_j |a_{ij}|x_j=\sum_j a_{ij}y_j $$ Then $$ \|A\|^p\ge \sum_i\left|\sum_j a_{ij}y_j\right|^p\ge \frac 1n\||A|\|^p $$ and we are done.

What "Riesz-Thorin" are we talking about? Much less trivial questions get closed in no time and this one got 8 upvotes and 2 favorites...

  • $\begingroup$ +1 Thank you very much for your answer. The whole story is indeed much simpler than I (and several colleagues I asked) thought. Concerning your last paragraph: Sometimes, whether a question is trivial or not simply depends on the perspective from which we look at it. As I wrote, one can indeed use the Riesz-Thorin interpolation theorem together with Hölder's inequality to prove the (suboptimal) estimate I stated in the question. Your argument, though, shows that this interpolation approach is not only an overkill; it even misses the target :-). $\endgroup$ – Jochen Glueck Jun 3 '18 at 20:32

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.