11
$\begingroup$

This question is post on MSE a week ago. I move it here to draw more attention.

Let $u\in C^\infty(\bar I)$ be given where $I=(0,1)$. Define $$ t(\alpha):=\left(\int_I\int_I \frac{|u(x)-u(y)|^\alpha}{|x-y|^{1+s\alpha}}\right)^{\frac1\alpha} $$ where $1<\alpha<2$, $0<s<1$ is fixed. Note the $t(\alpha)$ above defines the fractional order sobolev seminorm. See the Wikipedia article on "Sobolev space", section "Sobolev spaces with non-integer $k$".

We know that for usual $L^p$ space, we have, for $p<q$, $\|u\|_{L^p(I)}\leq \|u\|_{L^q(I)}$ by using Hölder inequality. So I am wondering whether similar properties hold for sobolev fractional seminorm with non-integer $k$.

That is, I am wondering for $1<\alpha_1<\alpha_2<2$, do we have $$ t(\alpha_1)\leq Ct(\alpha_2) $$ hold, where $C$ is a constant does not depends on $u$. Just like what we usually have for $L^p$ norm. However, I tried to prove it by using Minkowski inequality or something like Hölder, but I can't do it… I think the domain $I=(0,1)$ would be important and I also tried to use Jensen inequality, but no lucky…

Any ideas? Thank you!

$\endgroup$
2
  • $\begingroup$ sorry, you are right, there was something wrong with that $\endgroup$
    – user83457
    Nov 10, 2016 at 16:03
  • $\begingroup$ The embedding properties mentioned in the wikipedia article should give something like this, though one probably has to replace $t(\alpha)$ by $\|u\|_{\alpha}+t(\alpha)$. $\endgroup$ Nov 11, 2016 at 18:54

1 Answer 1

9
$\begingroup$

Your question can be rephrased by asking whether one has a Hölder estimate $$ |u|_{W^{s, p}} \le C |u|_{W^{s, q}}, $$ when $p < q$ or whether $W^{s, q} \subset W^{s, p}$.

There is no such embedding or inequality.

For proofs of this fact you can have a look at the paper Mironescu, Sickel, A Sobolev non embedding, Atti Accad. Naz. Lincei Rend. Lincei Mat. Appl. 26 (2015), no 3, 291—298.

One way to see this is to take $$ u(x) = \sum_{j \in \mathbb{N}} c_j e^{i 2^j x}, $$ and to observe that $$ |u|_{W^{s, p}}^p \simeq \sum_{j \in \mathbb{N}} |2^{js} c_j|^p, $$ and then choose suitably $c_j$.

$\endgroup$
3
  • $\begingroup$ I understand. Btw, would it make any difference if I restrict my $p$ and $q$ is a small interval? For example, if I assume that $1<p<q<1+\epsilon$ where $\epsilon>0$ could be arbitrary small. Would this make any different? $\endgroup$
    – JumpJump
    Nov 17, 2016 at 0:02
  • 1
    $\begingroup$ No, in fact. The integral can be finite for some $q$ and infinite for every $p < q$. The only possibility is to decrease slightly the smoothness in the left-hand side of the inequality by taking a $W^{r, p}$ norm with $r <s$ and $q < p$. $\endgroup$ Nov 17, 2016 at 8:23
  • 2
    $\begingroup$ I learned about this non-embedding result about a couple of years ago and I was quite surprised. $\endgroup$ Jan 28, 2019 at 20:48

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.