2
$\begingroup$

Suppose that, for some finite-dimensional real vector space $\Bbb R^n$, that $n_1(v)$, $n_2(v)$, ..., $n_k(v)$ are a set of norms on the space.

Given some $v$, then, we can look at the "vector of norms", which I will denote $v_n = (n_1(v), n_2(v), ..., n_k(v))$.

We can then look at norms on this vector of norms. For example, we could take the $\ell_1$ norm of the vector, which would be the sum of norms. It is easy to see that this will also be a norm.

  • Question 1: Is any norm on this "vector of norms" also a norm?
  • Question 2: Likewise, if we replace with seminorms, is any seminorm on the "vector of seminorms" also a seminorm?
  • Question 3: If not, for what norms do these things hold? (Do they at least hold for $\ell_p$ norms on the vector of norms?)

It is easy to see that you get homogeneity and positive-semidefiniteness, so the question is really about convexity. Does taking a "norm of norms" preserve convexity? Equivalently, does taking a norm of convex functions preserve convexity, or does taking a strictly increasing multivariate convex function of multiple convex functions preserve convexity?

EDIT - as per the answer from "mihaild" below, this isn't true for general norms, but would still like to know when it is true (in particular if it's true for $\ell_p$ norms without changing the basis).

$\endgroup$
2
$\begingroup$

At least for the first (and so for the second) question the answer is "no".

Take two norms $\|\cdot\|_1$ and $\|\cdot\|_2$ and two vectors $x$, $y$ such that $\|x\|_1 = \|x\|_2 = \|y\|_1 = \|y\|_2 = 1$, $\|x + y\|_1 \approx 2$, $\|x + y\|_2 \approx 0$.

Let $n(a, b) = \max (\frac{1}{3} |a + b|, |a - b|)$ (equal to $l_\infty$ norm in some scaled and rotated basis).

Then $n(\|x + y\|_1, \|x + y\|_2) \approx n(2, 0) = 4$, but $n(\|x\|_1, \|x\|_2) + n(\|y\|_1, \|y\|_2) = 2\cdot n(1, 1) = \frac{4}{3} < 4$.

For when it holds - at least if $n$ is such that for any $a_1 > 0, a_2 > 0, \ldots a_k > 0$ and $q_i \in [-a_i, a_i]$ we have $n(q_1, \ldots, q_n) \leqslant n(a_1, \ldots, a_n)$, then it holds: $$n(\|x + y\|_1, \ldots, \|x + y\|_n) = \\ n(\|x\|_1 + (\|x + y\|_1 - \|x\|_1), \ldots, \|x\|_n + (\|x + y\|_n - \|x\|_n)) \leqslant\\ n(\|x\|_1, \ldots, \|x\|_n) + n(\|x + y\|_1 - \|x\|_1, \ldots, \|x + y\|_n - \|x\|_n) $$ If $a_i = \|y\|_i$ and $q_i = \|x + y\|_i - \|x\|_i$, then we have $$n(\|x\|_1, \ldots, \|x\|_n) + n(\|x + y\|_1 - \|x\|_1, \ldots, \|x + y\|_n - \|x\|_n) \leqslant\\ n(\|x\|_1, \ldots, \|x\|_n) + n(\|y\|_1, \ldots, \|y\|_n) $$

It holds at least for all $l_p$ norms. I think it is equal to unit ball defined by $n$ to be contained in hypercube bounded by hyperplanes $x_i = \pm p_i$, where $p_i$ is such that $n(0, 0, \ldots, p_i, \ldots, 0) = 1$.

This condition if definitely not necessary: for example, if all $\|\cdot\|_i$ coincide, then any $n$ will work.

$\endgroup$
  • $\begingroup$ Yes, this is correct. Do you know when this does hold? Does it hold for Lp norms (without rotating the basis)? $\endgroup$ – Mike Battaglia Apr 23 '19 at 21:57
  • $\begingroup$ I wrote my thoughts about it. I don't know and could'n come up with any good criteria. $\endgroup$ – mihaild Apr 23 '19 at 22:14
  • $\begingroup$ I think the outer norm (we call it $n$) needs to be increasing in a coefficient-wise way, i.e., it satisfies: $|v| \le |w|$ (coefficientwise) implies $n(v) \le n(w)$. $\endgroup$ – gerw Apr 24 '19 at 6:18
  • $\begingroup$ I am not sure it's enough: my deduction uses that, for example, if $a > b > 0$ then $n(a, c) \geqslant n(-b, c)$. Though I can't think of any counterexample for your hypothesis either. $\endgroup$ – mihaild Apr 24 '19 at 9:59

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.