1
$\begingroup$

Let $X$ be a finite dimensional Banach space and define a matrix norm $\| \cdot \|_{X}$ by

$$ \| A \|_{X} = \sup_{x \ne 0} \frac{\|A x\|_{X}}{\|x\|_{X}} $$

where the matrix $A$ is interpreted as an operator $X \to X$ in the obvious way.

I'm looking to define a minimal norm, namely

$$ f(A) = \inf_{X} \| A \|_{X} $$

What can be said about $f$? Does it define a norm? Can it be computed?

$\endgroup$
  • $\begingroup$ I'm not sure I can follow your argument for positive definiteness. Why should $c_1$ and $c_2$ be independent of $X$? $\endgroup$ – Jochen Glueck Mar 3 at 10:24
  • $\begingroup$ You are right, perhaps that is not at all true. I do think we could find such $c_1$, $c_2$ though, I'll have to look around for a proof. We would also have to "normalize" away a constant since the proof obviously fails by considering the family of norms $\|x\| = a \|x\|_2$ as $a \to \infty$. $\endgroup$ – Jonas Adler Mar 3 at 10:47
  • $\begingroup$ Well, I think the situation is actually a bit more involved. Multiplying the norm by a constant $a > 0$ is not really a problem since the constant $a$ cancels in the definition of the induced matrix norm. $\endgroup$ – Jochen Glueck Mar 3 at 11:11
  • $\begingroup$ By the way, sub-additivity is probably also going to be a problem since infima do not respect subadditivity, in general. $\endgroup$ – Jochen Glueck Mar 3 at 11:14
  • $\begingroup$ I updated the proof of positive definiteness, the new one also gives an explicit bound. $\endgroup$ – Jonas Adler Mar 3 at 11:17
5
$\begingroup$

For finite matrices, your 'norm' is the spectral radius of $A$. Indeed, one can construct for each matrix $A$ a matrix norm induced by a vector norm such that $\|A\| \leq \rho(A) + \varepsilon$ for each $\varepsilon>0$. (And, on the other hand, $\|A\|\geq \rho(A)$ for each norm induced by a vector norm).

  1. for each matrix $A$ you can construct a Jordan-like form $A=VJV^{-1}$ in which the strictly upper diagonal part contains $\varepsilon$'s instead of 1 (you can get it by taking the Jordan form and conjugating by $\operatorname{diag}(1,\varepsilon,\varepsilon^2,\dots,\varepsilon^{n-1})$).
  2. then define the norm $\|M\|:=\|V^{-1}MV\|_2$ (where $\|K\|_2$ is the operator norm of $K$), which is induced by the vector norm $\|x\| = \|V^{-1}x\|$. Then $\|A\| = \rho(A) + O(\varepsilon)$. On the other hand, each norm is greater than the spectral radius.

This is a 'standard' proof trick that I learned in my courses.

$\endgroup$
  • 3
    $\begingroup$ Jordan form is too heavy machinery for this question. And the real case is not covered, I am afraid. Instead, we may fix any $r>\rho$, choose $m$ for which $\|A^m\|<r^m$ and define new norm as $\|x\|+\|Ax\|+\dots+\|A^{m-1} x\|$. $\endgroup$ – Fedor Petrov Mar 3 at 11:57
  • $\begingroup$ Beautiful proof! Thank you. $\endgroup$ – Jonas Adler Mar 3 at 12:19
  • 2
    $\begingroup$ I forgot to add powers of $r$: $\|x\|+r^{-1}\|Ax\|+\dots+r^{1-m}\|A^{m-1}x\|$. $\endgroup$ – Fedor Petrov Mar 3 at 17:06

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.