Let $\Omega$ denote a open subset of $\mathbb{R}^n$, and $W^{1,p}(\Omega)$ the Sobolev space of weakly differentiable functions $u\in L^p(\Omega)$ (that is, for which $D_iu$ exists and belongs to $L^p(\Omega)$ as well, for every $i\in\left\{1,\ldots,n\right\}$).

I'm studying boundary regularity of a solution of the Dirichlet problem for the circle $D\subseteq\mathbb{R}^2$, and the weak maximum principle is needed. For that, we need to give a proper meaning for $u\leq v$ in $\partial\Omega$ (the boundary of $\Omega$), where $u,v\in W^{1,2}(\Omega)$. The definition is: $u\leq v$ in $\partial\Omega$ iff $(u-v)^+\in W_0^{1,2}(\Omega)$, where:

  • $w^+=\max(w,0)$ denotes the positive part of $w$; and
  • $W_0^{1,2}(\Omega)$ is the closure (in Sobolev norm) of $C_0^\infty(\Omega)=\left\{w\in C^\infty(\Omega):\text{supp}(w)\text{ is compact}\right\}$

Now, a lot of statements relating to that concept need the following lemma (or something similar), which I'm unable to prove:

Lemma: Let $u\in W_0^{1,2}(\Omega)$, $v\in W^{1,2}(\Omega)$ such that $v\geq 0$ pointwise. Show that $(u-v)^+\in W_0^{1,2}(\Omega)$.

Intuitively, that should be true. If we think of continuous functions $u$ and $v$ such that $u\in C_0(\Omega)$, then $(u-v)^+\in C_0(\Omega)$, since $|(u-v)^+|\leq|u|$ (where $C_0(\Omega)$ denotes the set o compactly supported continuous functions from $\Omega$ to $\mathbb{R}$). Using mollifiers, it's easy to see that $C_0(\Omega)\subseteq W_0^{1,2}(\Omega)$, and the lemma is true in that case. I tried to give the following proof:

The case $u\in C_0^\infty(\Omega)$ is easy enough. Now, let $\left\{u_n\right\}_{n\in\mathbb{N}}\subseteq C_0^\infty(\Omega)$ be a sequence converging in $W^{1,2}(\Omega)$ to $u$. Taking subsequences if necessary, I tried making pointwise convergence (almost everywhere in $\Omega$) of $(u_n-v)^+$ and $D_i(u_n-v)^+$ to $(u-v)^+$ and $D_i(u-v)^+$, respectively, so I'd apply Lebesgue's Dominated Convergence Theorem, and conclude convergence in $W^{1,2}(\Omega)$, hence proving the lemma. The problem is exactly with the derivatives: Since $Dw^+(x)=0$ if $w(x)\leq 0$ and $Dw^+(x)=Dw(x)$ if $w(x)>0$ ($\forall w\in W^{1,2}(\Omega)$: this follows from the weak chain rule), we cannot garantee pointwise convergence if $u(x)=v(x)$.

Any idea would be of great value. Even if one must assume $\Omega$ bounded and/or $v\in C^\infty(\Omega)\cap C^0(\overline{\Omega})$, it would suffice for what I need.

Thank you in advance.


3 Answers 3


I found a nice elementary proof for this: All the equalities are meant to hold a.e..

Lemma: Let $u\in W^{1,p}(\Omega)$ and $N=\left\{x\in\Omega:u(x)=0\right\}$. Then $D_iu=0$ a.e. in $N$.

Proof: Let $u^+=\max(u,0)$ and $u^-=\min(u,0)$. By the chain rule, $u^+,u^-\in W^{1,p}(\Omega)$ and $D_iu^+(x)=D_iu(x)$ if $u(x)>0$ and $D_i u^+(x)=0$ otherwise, and $D_i u^-(x)=D_iu(x)$ if $u(x)<0$ and $D_iu^-(x)=0$ otherwise. Since $D_iu(x)=D_i(u^++u^-)(x)=D_iu^+(x)=D_iu^-(x)$, the lemma holds.$\square$

Proof of the statement in the question: First suppose $u\in C_0^\infty(\Omega)$. Then $\text{supp}(u-v)^+\subseteq\text{supp}(u)$, and a sequence of mollificationsof $(u-v)^+$, converging to it in $W^{1,p}(\Omega)$ will eventually be in $C^\infty_0(\Omega)$, and the statement holds.

Now suppose $u\in W_0^{1,p}(\Omega)$. Let $\left\{u_n\right\}_{n\in\mathbb{N}}$ be a sequence of $C_0^\infty(\Omega)$ functions converging to $u$. Taking subsequences if needed, we can assume $u_n\rightarrow u$ and $D_iu_n\rightarrow D_iu$ a.e. in $\Omega$. Then $(u_n-v)^+\rightarrow (u-v)^+$ a.e. pointwise (for $x\in\Omega$, consider the cases $u(x)>v(x)$, $u(x)<v(x)$ and $u(x)=v(x)$), and the dominated convergence theorem implies that $(u_n-v)^+$ converges to $(u-v)^+$ in $L^p(\Omega)$.

Now, let $x\in\Omega$ (such that $D_iu_n(x)\rightarrow D_iu(x)$). If $u(x)>v(x)$ or $u(x)<v(x)$, use the chain rule to see that $D_i(u_n-v)^+(x)\rightarrow D_i(u-v)^+(x)$. If $u(x)=v(x)$, by the lemma we can assume that $D_iu(x)=D_iv(x)$, and it's easily verified that $D_i(u_n-v)^+(x)\rightarrow D_i(u-v)^+(x)$. Again by dominated convergence, $D_i(u_n-v)^+\rightarrow D_i(u-v)^+$ in $L^p(\Omega)$. That means that $(u_n-v)^+\rightarrow(u-v)^+$ in $W^{1,p}(\Omega)$. Since $(u_n-v)^+\in W_0^{1,p}(\Omega)$, which is closed, the result follows.$\square$


Going through fine pointwise properties of Sobolev functions and associated nonlinear potential theory looks like doing it the hard way. Here is another approach.

Claim. Let $1< p<\infty$. If $u\in W_0^{1,p}(\Omega)$, $w\in W^{1,p}(\Omega)$, and $0\le w\le u$, then $w\in W_0^{1,p}(\Omega)$.

We need a standard truncation lemma. Here $u^+=\max(u,0)$.

Lemma 1 If $u \in W^{1,p}(\Omega)$, then $u^+ \in W^{1,p}(\Omega)$ and $|\nabla u^+|\le |\nabla u|$ almost everywhere.

Proof. Use the ACL characterization of $W^{1,p}$. Note that $$|u^+(a)-u^+(b)|\le |u(a)-u(b)|,\qquad a,b\in \Omega \tag1$$ Since $u$ is absolutely continuous on almost every coordinate-parallel segment, so is $u^+$ — just look at the definition of absolute continuity and at (1). Therefore, $u^+$ on such segments $u^+$ is differentiable a.e. Again, (1) implies that its partial derivatives cannot exceed the partial derivatives of $u$. $\Box$

Corollary 2. If $u,v \in W^{1,p}(\Omega)$, then $\min(u,v) $ and $\max(u,v)$ are in $W^{1,p}(\Omega)$ and their gradients are bounded by $|\nabla u|+|\nabla v|$.

Proof. Apply Lemma 1 to $\dfrac{u+v}{2}\pm \dfrac{(u-v)^+}{2}$. $\Box$

Proof of the Claim: Let $(w_n)$ and $(u_n)$ be smooth approximations to $w$ and $u$ in the Sobolev norm, with $u_n$ compactly supported in $\Omega$. Then $f_n:=\min(w_n^+,u_n)$ is a sequence of Lipschitz compactly supported functions. By construction, $f_n\to w$ a.e. From Lemma 1 and Corollary 2 we see that $(f_n)$ is a bounded sequence in $W^{1,p}$ norm. Therefore, it has a subsequence $f_{n_k}$ that converges to some $g\in W^{1,p}_0(\Omega)$ weakly in $W^{1,p}$ and strongly in $L^p$ (the latter by Rellich-Kondrachov). Recalling that $f_n\to w$ a.e., we conclude that $g=w$. Thus, $w\in W^{1,p}_0(\Omega)$. $\Box$

Notice how the "soft" tools of functional analysis save us the headache of dealing with boundary values. The set $W^{1,p}_0(\Omega)$ is closed in $W^{1,p}(\Omega)$ by definition, and is also convex. From functional analysis, it's weakly closed — and this is how we get $g\in W^{1,p}_0(\Omega)$, not from pointwise analysis of boundary values.

  • $\begingroup$ Just for reference: this nice fact is contained in Lemma 1.25 (ii) from [Heinonen, Kipelainen, Martio, "Nonlinear potential theory of degenerate elliptic equations"]. $\endgroup$
    – Voliar
    Aug 29 at 7:58

There is a characterization of $W_0^{1,2}$ in terms of pointwise traces: a function $u \in W^{1,2}(\Omega)$ belongs to $W_0^{1,2}(\Omega)$ if and only if $$ \lim_{r \to 0^+} \int_{B(x,r) \cap \Omega} |u(y)| \, dy = 0 $$ for $(1,2)$-quasievery point $x \in \partial \Omega$. Since you assume $u \in W^{1,2}_0(\Omega)$ and point out that $|(u-v)^+| \le |u|$, you can conclude that $(u-v)^+ \in W^{1,2}_0(\Omega)$ also.

  • 3
    $\begingroup$ What does it mean "for (1,2)-quasievery point $x\in\partial\Omega$", and, if possible, where can I find the proof if this fact? $\endgroup$ Apr 28, 2013 at 16:11

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.