Let $\Omega$ be a sufficiently regular domain, for example $\Omega=B(0,r)$, $r>0$, and $m(x)=\textrm{dist}(x,\partial \Omega)$ be the distance to boundary function.

Suppose $\mathcal{F}$ is a given family of smooth functions, which is dense in $C^m(\bar\Omega)$ (with corresponding $m$-th order sup-norms) for arbitrary $m$. Is it possible to explicitly construct another family $\mathcal{G}$ of functions based on $\mathcal{F}$ and the function $m(x)=\textrm{dist}(x,\partial \Omega)$, such that $\mathcal{G}$ is dense in $W^{2,2}(\Omega) \cap W^{1,2}_0(\Omega)$ (i.e., any function $g\in \mathcal{G}$ satisfies that $g|_{\partial \Omega}=0$)?

A natural candidate is the set $$ \mathcal{G}=\{ mf\mid f\in \mathcal{F}\}, $$ where I directly multiply the function $f$ by the distance function. Since $\Omega$ has a sufficient regular boundary, the function $m$ is sufficiently smooth, hence for any $f\in C^{2}(\bar{\Omega})$, $mf\in C^2(\bar{\Omega})\cap C_0(\bar{\Omega})$. I also know $C^{2}(\bar{\Omega})\cap C_0(\bar{\Omega})$ is dense in $W^{2,2}(\Omega) \cap W^{1,2}_0(\Omega)$ (in $W^{2,2}$ norm). But I was not able to proceed further to show $\mathcal{G}$ is dense in $C^2(\bar{\Omega})\cap C_0(\bar{\Omega})$.


This is a partial answer (too long for a comment). The idea is to reduce to the case when the normal derivative of $f$ vanishes on the boundary.

In the following let us use the notation from the book Sobolev spaces by R.A. Adams. In particular, $\|f\|_{s,p,\Omega}$ is the norm of $f$ in $W^{s,p}(\Omega)$, let $T= \prod_{k=0}^{s-1} W^{s-k-1/p, p}(\partial \Omega)$ and let $\gamma \colon W^{s,p}(\Omega) \to T$ denote the trace map, i.e. the extension from $C_0^\infty(\mathbb{R}^n)$ to $W^{s,p}(\Omega)$ of the map $f \mapsto (\gamma_0(f),\dots, \gamma_{s-1}(f)) := (f|_{\partial \Omega}, \dots, \frac{\partial^{s-1} f}{\partial n^{s-1}}|_{\partial \Omega})$, for more details see Section 7.52 in the cited book. In our case $s=p=2$. By theorem 7.53 the operator $\gamma$ is bounded and for any $g\in T$ there exists $E(g)\in W^{s,p}(\Omega)$ such that $\gamma(E(g)) = g$ and $\|E(g)\|_{s,p,\Omega} \le C \|g\|_T$. (All constants are denoted below with $C$.)

Let $f\in W^{2,2}(\Omega) \cap W^{1,2}_0(\Omega)$. Take $\phi \in \mathcal{F}$ such that $\|f - \phi\|_{2,2,\Omega} < \epsilon$.

Define $\tilde \phi:=E(0,\gamma_1(\phi))$, i.e. take a function $\tilde\phi$ with the same normal derivative as $\phi$ but zero value on $\partial \Omega$.

Then $\gamma_0(f-\tilde \phi)=0$ and $\|\gamma_1(f-\tilde \phi)\| = \|\gamma_1(f - \phi)\| \le C \epsilon$. Hence there exists a function $r = E(\gamma(f-\tilde \phi))\in W^{2,2}(\Omega)$ such that $\gamma(r) = \gamma(f-\tilde \phi)$ and $\|r\| \le C \epsilon$. Therefore $\gamma(f - \tilde \phi - r) = (0,0)$, that is $f-\tilde \phi-r \in W^{2,2}_0(\Omega)$. Such functions can easily be approximated by functions of the form $\zeta \cdot \psi$, where $\zeta$ is a cutoff function (0 at $\partial \Omega$, 1 sufficiently far from $\partial \Omega$) and $\psi \in \mathcal{F}$.

Thus we have written $f$ in the form $f = \tilde \phi + \zeta \cdot \psi + \tilde r$, with $\tilde \phi + \zeta \cdot \psi \in W^{2,2}(\Omega)\cap W^{1,2}_0(\Omega)$ and $\|\tilde r\|_{2,2,\Omega}$ is small. Therefore $$ \mathcal{G} = \{\tilde \phi + \zeta\cdot \psi: \phi,\psi\in \mathcal{F}, \text{ $\zeta$ boundary cutoff function}\} $$ where $\tilde \phi = E(0,\gamma_1(\phi))$.

If one were able to write $E(0,\gamma_1(\phi))$ explicitly using the distance to the boundary (I believe this is possible), then this construction would be more explicit.


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.