Let $D$ be a regular domain in $\mathbb R^2$. Suppose that $u \in H_0^1(D) \cap C(D)$. Does this imply $u \in C(\overline D)$ and $u|_{\partial D} = 0$?
-
$\begingroup$ It shouldn't be true if the embedding from $H_0^1(D)\to C(D)$ is not continuous by some functional analysis mumbo-jumbo, which I'm pretty sure it isn't even when $D$ is a disk, but I'll leave it to someone else to figure out the details. $\endgroup$– Aleksei KulikovCommented Jul 20 at 17:51
-
$\begingroup$ I do not agree with Aleksei Kulikov's view that a negative answer could be obtained by some "functional analysis mumbo-jumbo". Instead, it seems to me that either there should be an explicit counterexample or a proof of the positive result. I tried some "tilted cone" like functions with one finite jump discontinuity point at the boundary, but these examples failed basically because computing $\int_\Omega|\nabla u|^2$ lead to the implication $\int_0^1t^{-\frac32}{\rm\,d\,}t=+\infty\Rightarrow\int_\Omega|\nabla u|^2=+\infty\,$. $\endgroup$– TaQCommented Jul 26 at 14:20
2 Answers
The answer is no. Namely, for every nonempty open set $\Omega\subset\mathbb R\times\mathbb R$ there is a continuous function $u$ in $W^{1,2}_{\boldsymbol.}(\Omega)$ that has no continuous extension defined on the closure of $\Omega\,$. The (counter)example is constructed as follows.
Let $\langle\,x_i:i\in\mathbb N_0\,\rangle$ be any injective sequence in $\Omega$ with at least one accumulation point, and such that all accumulation points are on the boundary of $\Omega\,$. Then one can choose a stictly positive sequence $\langle\,R_i:i\in\mathbb N_0\,\rangle$ in $\ell^{\,2}(\mathbb N_0)$ such that $\langle\,{\rm B\,}(x_i,R_i):i\in\mathbb N_0\,\rangle$ is a sequence of disjoint balls in $\Omega\,$. For $s\le 0$ letting $\ln\kern.3mm s=-\infty=-\ln\kern.3mm(+\infty)$ and defining $u_i:\Omega\to[\,0\,,1\,]$ by $$ x\mapsto\inf\,\{\,\sup\,\{\,R_i\ln\kern.5mm(1- \ln\frac{|\,x-x_i\,|}{R_i}\Bigg)\,,0\,\}\,,1\,\} $$ then the function $u=\sum_{\,i\,\in\,\mathbb N_0\,}u_i$ has the claimed property.
If there are points that you do not understand, leave a comment specifying the unclear parts.
-
1$\begingroup$ @DanieleTampieri : Vai did you change my notation? Everything there was intentional. Did you find some mathematical error? You have now wasted one of the possible 10 edits before the post goes "cw", if the convention is still the same as earlier. $\endgroup$– TaQCommented Jul 27 at 19:26
-
$\begingroup$ I apologize for the edit. I just found strange the notations for the sequences (and I admit this is somewhat arbitrary) and I found not understandable the notation for the mapping (and this is still now my opinion). Said that I think there's no need to comment anymore, just revert the edit (as you've already done). $\endgroup$ Commented Jul 28 at 6:35
-
$\begingroup$ For strange reasons I see that I downvoted...I do not know why, probably I just pressed the wrong key. If you make some change, I can change, too. Sorry $\endgroup$ Commented Aug 15 at 17:55
-
$\begingroup$ @GiorgioMetafune I do not quite understand your comment; do you intend to mean that you unintentionally downvoted my answer and that you cannot (for what reason?) revert your downvote unless I somehow edit my answer? $\endgroup$– TaQCommented Aug 17 at 3:21
-
1$\begingroup$ Yes, exactly. If I try to revert my vote, the system says that I cannot do until the post has been reedited. $\endgroup$ Commented Aug 17 at 6:30
This is not true. Take a sequence of points $(x_n) \subset D$ which converge to a point on the boundary (or even worse: whose accumulation points are $\partial D$).
Then, for every point $x_n$ we can find $\varepsilon_n > 0$ and $\varphi_n \in C_c^\infty(D)$ with $\varphi_n \ge 1$ on the ball $B_{\varepsilon_n}(x_n)$, $\|\varphi_n\|_{H_0^1(D)} \le 2^{-n}$ such that $\varphi_n$ is supported on $B_{2\varepsilon_n}(x_n)$. Then, consider $$ u := \sum_{n \in \mathbb N} \varphi_n. $$
It is easy to check that $u$ has the desired properties.
-
$\begingroup$ How do you construct the functions $\varphi_n\,$? If you just take some compactly supported smooth $\varphi$ and dilate and translate it, then $\int_\Omega|\nabla\varphi_n|^2$ essentially remains unchanged. and you do not get it ${}\le2^{-n}\,$. $\endgroup$– TaQCommented Jul 26 at 14:09
-
$\begingroup$ @TaQ the "embedding" $H_0^2(\Omega)\to C(\bar{\Omega})$ is not continuous for any open domain $\Omega\subset \mathbb{R}^2$, and $C^\infty_c(\Omega)$ is dense in $H_0^2(\Omega)$, hence there are smooth compactly supported functions in $\Omega$ with arbitrarily small $H_0^2(\Omega)$ and arbitrarily large $C(\Omega)$-norm (this is more or less the solution I had in mind as well). $\endgroup$ Commented Jul 27 at 20:25
-
$\begingroup$ @AlekseiKulikov Assuming that "$H^2_0(\Omega)$" is a misprint and that you actually mean the space that e.g. in Adams' book is $W^{1,2}_0(\Omega)\,$, the logical development of the Sobolev embedding theory proceeds so that one first (e.g) gets for open $\Omega\subseteq{}^{2\boldsymbol.\,}\mathbb R$ the embedding $W^{1,p}(\Omega)\hookrightarrow C\overline{(\Omega)}$ for $2<p\le+\infty\,$, and then shows by the classical log−log example (that is also used in my answer) that this does not extend to $p=2\,$. (to be continued) $\endgroup$– TaQCommented Jul 27 at 22:07
-
$\begingroup$ (cont.) So the explicit example is a more fundamental logical step here for the nonembedding, and deducing in the manner you (seem to) suggest is like first proving A, then deducing B from it, and then deducing A from B. OK, it can be seen as a "functional analysis mumbo-jumbo" way of getting the result, but to me it is an unnecessary (logical) complication. $\endgroup$– TaQCommented Jul 27 at 22:08
-
$\begingroup$ @TaQ yes, I meant $H^1_0(\Omega)$, sorry. I agree that from the pedagogical point of view your answer is superior if we want to teach the students, but this is MathOverflow, and my opinion is that when you encounter problems in the wild you are free to drop all the nukes you have on them (so, e.g., yes, you use log--log example to show lack of continuous embedding, but it was known for a hundred years already, so I use it as a black box, and from there construct a continuous example as requested by the OP). If the OP question was just about lack of bound, it would not be appropriate for MO. $\endgroup$ Commented Jul 27 at 22:48