1
$\begingroup$

Are there any known function spaces where the gradients are uniformly bounded? For a problem I’m working on, I’ve been able to show my functions are bounded in a ball within an RKHS (reproducing kernel Hilbert space), but I can’t show the gradients are bounded.

Just wondering if there are any such spaces that constrain the size of the gradient as well.

The reason I ask is because in this paper https://arxiv.org/pdf/1905.11882.pdf, the authors rely on showing the optimal "potential" functions lie in a set with a bounded Holder norm. They're able to show specific bounds on the Holder norms (lines 11 and 12 in that paper).

However, in my setting, I have only been able to show that the norm of my optimal potential function is bounded in an RKHS. But I can't find any way of controlling the norms of the gradients similar to what the authors did there. Or if I can't do that, are there standard bounds for the covering numbers if the functions are bounded and lie in an RKHS?

$\endgroup$
4
  • $\begingroup$ On what kind of domain? Sobolev embedding should say that an appropriate Sobolev space would do the trick. As a trivial example, on an interval $I \subset \mathbb{R}$, a function in $H^1(I)$ is bounded (even absolutely continuous), so a function in $H^2(I)$ has bounded derivative (and is even $C^1$). $\endgroup$ Mar 29, 2021 at 4:18
  • $\begingroup$ Just a real valued domain. $\endgroup$
    – Kashif
    Mar 29, 2021 at 4:30
  • 1
    $\begingroup$ Why not $C^1_b(\Omega)$? (=bounded functions on the open set $\Omega$ with continuous bounded partial derivatives) $\endgroup$ Mar 29, 2021 at 8:08
  • $\begingroup$ Added edits to the question to give more detail. $\endgroup$
    – Kashif
    Mar 29, 2021 at 11:49

1 Answer 1

2
$\begingroup$

Maybe I'm misinterpreting your question but no nontrivial vector space of differentiable functions has uniformly bounded gradients. If $\inf_x \|\nabla f(x)\|= M>0$, then for $c>1$, $\inf_x\|\nabla cf(x)\|= cM>M>0$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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