Given a set $\mathcal{X}$ and RKHS $\mathcal{H}$ of functions on $\mathcal{X}$, we can recover a (pseudo)metric on $\mathcal{X}$ by $d(x,y)=||\phi_x-\phi_y||_{\mathcal{H}}$, where $\phi_x=k(x,\cdot)$.

It is straightforward to see that any function $f \in \mathcal{H}$ which has RKHS norm less than $L$ is Lipschitz (with respect to our metric above) with constant $L$:

$$|f(x)-f(y)|=|\langle f,\phi_x-\phi_y\rangle|\leq ||f||_{\mathcal{H}}d(x,y), $$ for any $x,y \in \mathcal{X}$.

I am very interested in the following question: if we have a function $f\in \mathcal{H}$ which is Lipschitz with constant $L$, is there anything we can say about it's norm $||f||_{\mathcal{H}}$?

  • 1
    $\begingroup$ by "anything we can say", do you mean bounding it from above? If yes, then you want your Hilbert space to be isomorphic to a subspace of a Lipschitz space. I would expect a subspace of the Lipschitz space isomorphic to a Hilbert space necessarily be finitely dimensional, but I am not sure. $\endgroup$ – erz Jul 4 at 9:27
  • $\begingroup$ What happens when $\mathcal{X}=\mathbb{Z}$ and $\mathcal{H}=l^2$? $\endgroup$ – DCM Jul 4 at 11:39
  • $\begingroup$ @erz yes, bounding it above was precisely what I was looking for $\endgroup$ – Tyler6 Jul 4 at 13:09

I think that in general $L(.)$ and $\Vert.\Vert_\mathcal{H}$ measure quite different things.

Writing $L(f)$ for

$$ \inf\{ M>0:|f(x)-f(x')| \leq Md(x,y) \;\forall \;x,x'\in \mathcal{X}\} $$

let $\mathcal{X}=\mathbb{Z}$ and $\mathcal{H}=l^2$. Then (unless I've made an embarrassing mistake...) setting $f_n=1_{[-n,n]}$ gives you a sequence in $\mathcal{H}$ with $L(f_n)=1/2$ and $\Vert f_n\Vert_\mathcal{H}=2n$.

| cite | improve this answer | |
  • $\begingroup$ Hmm, correct me if I’m wrong but I thought $l^2$ wasn’t an RKHS? Either way it’s an interesting counterexample $\endgroup$ – Tyler6 Jul 4 at 13:06
  • 1
    $\begingroup$ What's your definition of a RKHS? I think the usual one is just "a Hilbert space of functions on a set whose evaluation functionals are norm continuous", no? $\endgroup$ – DCM Jul 4 at 13:37
  • 1
    $\begingroup$ I think another example is $\mathcal{H} = H^1_0((0,1])$, i.e. the absolutely continuous functions $f : [0,1] \to \mathbb{R}$ with $f' \in L^2([0,1])$, under the norm $\|f\|_{\mathcal{H}}^2 = \int_0^1 |f'|^2$. Then the "Lipschitz constant" is the Holder norm of exponent 1/2, $L = \sup |f(s)-f(t)|/\sqrt{|s-t|}$. So now let $f_n$ approach $f(x) = \sqrt{x}$ which has infinite $\mathcal{H}$-norm. $\endgroup$ – Nate Eldredge Jul 4 at 17:19

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.