Let $A$ be a $n\times n$-matrix. We let $\|A\|_p$ denote the norm of $A$ when considered as a linear operator on $\ell^p(\{1,2,\ldots,n\})$, that is, $$ \|A\|_p = \sup_{x\neq 0}\frac{\|Ax\|_p}{\|x\|_p}. $$

By the Riesz-Thorin theorem, the $p$-norm of $A$, as a function of $p$, is log-convex, meaning that the function $[1,\infty]\ni p\mapsto \log(\|A\|_p)$ is convex.

Question: Is the p-norm of a matrix strictly log-convex (unless it is constant)?

The proof of the Riesz-Thorin theorem uses holomorphic functions, so it seems plausible that if $\log(\|A\|_p)$ is constant on some non-trivial interval, then $\log(\|A\|_p)$ has to be globally constant.

Edit (following Terry Tao's answer):

Can $\|A\|_p$ be constant on some non-trivial interval, yet not be globally constant?

  • $\begingroup$ Why not just prove it directly? $\phi\left(\tfrac{p+q}{2}\right) < \sqrt{\phi(p)\phi(q)}$, where $\phi(p)=\|A\|_p$. $\endgroup$ – Suvrit May 15 '14 at 16:40
  • $\begingroup$ Do you have a proof in mind? The inequality is not true in general, for example for the identity matrix. $\endgroup$ – Hannes Thiel May 15 '14 at 16:46
  • $\begingroup$ What I meant was trying to just provide sufficient conditions for the direct inequality, by using closedness of log-cvxity under addition and multiplication. Clearly, it does not hold in general :-) $\endgroup$ – Suvrit May 15 '14 at 16:49

The answer is "no".

First, a trivial counterexample: let $A_n$ be the $n \times n$ matrix with all $1$s on the first row and zeroes elsewhere, then $\frac{1}{p} \mapsto \|A_n\|_p = n^{1/p}$ is log-convex but non-constant and not strictly log-convex.

One can deal with this example by replacing "constant" with "log-linear". However, if we consider the $(n+m) \times (n+m)$ matrix

$$ B := \begin{pmatrix} a A_n & 0\\ 0 & b A_m \end{pmatrix} $$ for some real parameters $a,b > 0$ and distinct natural numbers $n,m > 0$, then $\frac{1}{p} \mapsto \|B\|_p = \max( a n^{1/p}, b m^{1/p})$, which is log-convex but not log-linear and not strictly log-convex.

It seems to me that a similar construction allows any log-convex function to arise as $\frac{1}{p} \mapsto \|A\|_p$, at least if one allows A to be infinite-dimensional.

  • $\begingroup$ Thanks, I realize that my question was imprecise. I will edit it. What I want to know is: Can $\|A\|_p$ be constant on some non-trivial interval, yet not globally constant? $\endgroup$ – Hannes Thiel May 15 '14 at 17:02
  • 2
    $\begingroup$ Yes: take $m=1$ or $n=1$ in the above example. $\endgroup$ – Terry Tao May 15 '14 at 19:39

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.