# An alternate definition of Sobolev space $W^{1,p}(\Omega)$ when $1<p\leq\infty$ and consequences

Suppose that we define the Sobolev space $$W^{1,p}(\Omega)$$ with $$1, where $$\Omega\subset\mathbb{R}^d$$ ($$d\geq 1$$) is an open set (not necessarily bounded), in the following manner.

Definition. We say $$u\in W^{1,p}(\Omega)$$ if there is a constant $$C$$ such that for every open set $$U$$ with $$\overline{U}\subset\Omega$$ we have $$\|\tau_hu-u\|_{L^p(U)}\leq C|h|,\;\;\forall h\in\mathbb{R}^n\,\, \text{ s.t } \,0<|h|< d(U,\partial \Omega).\label{1}\tag{*}$$

($$\tau_h$$ is the translation map). We know that the usually defined weakly differentiable functions satisfy \eqref{1}. On the other hand, any function $$u$$ satisfying \eqref{1} can be shown to satisfy $$\exists C\in\mathbb{R}, \,\text{ s.t }\left|\int_{\Omega}u\varphi'\right|\leq C\|\varphi\|_{L^{p'}(\Omega)}, \;\forall \varphi\in C^1_c(\Omega).\label{2}\tag{**}$$ By virtue of the Riesz representation theorem on $$L^p$$ and Hahn-Banach/extension by continuity, \eqref{2} implies that $$u$$ is weakly differentiable when $$1. Therefore, the definitions are equivalent.

However, \eqref{1} is easier for me to grasp (especially when $$p=\infty$$) than your usual technical definition (that requires $$u$$ to satisfy integration by parts with functions in $$C^1_c$$), as \eqref{1} seems to insinuate that "$$u$$ is differentiable on average" when $$p<\infty$$ and that "$$u$$ is Lipschitz a.e" (so differentiable a.e) when $$p=\infty$$. So here are my questions:

1. What more can be said about the intuition behind \eqref{1}, and what is a nice example of a function $$u\not\in C^1(\Omega)$$, for which we can prove \eqref{1} directly?

2. How can we use \eqref{1}, while possibly avoiding Riesz or even defining the weak derivative, to show that $$W^{1,p}(\Omega)\subset C(\overline{\Omega}) \text{ for all } \begin{cases}1
-----
Possible approach to question 2 when $$\Omega=\Bbb R$$.

I think one can use convolutions to go from the integral inequality \eqref{1} to a point-wise inequality. My intuition comes from the following inequality: if $$u\in W^{1,p}(\Bbb R)$$ (as defined by \eqref{1}) with $$1 and $$\{\rho_n\}$$ is the standard sequence of mollifiers then $$\|\rho_n\star (\tau_h u-u)\|_{L^\infty(\Bbb R)}\leq \underbrace{\|\tau_h u-u\|_{L^p(\Bbb R)}}_{\leq C|h|}\cdot\underbrace{\|\rho_n\|_{L^{p'}(\Bbb R)}}_{\leq Kn^{1/p}}.$$ Maybe this inequality will help if we also assume that $$u\in L^\infty(\Bbb R)$$, since this assumption implies $$\|\rho_n\star (\tau_h u-u)\|_{\infty}\to \|\tau_h u-u\|_\infty.$$

3. Do you think this approach might lead somewhere?

• For 1, this might be silly, but how about Lipschitz functions? For 2, if you allow yourself to interpret (**) as the distributional derivative $Du$ of $u$ being an element of the dual of $L^{p'}$, then $v(x) := \langle Du, \chi_{[a,x]}\rangle$ should bring you far for $d=1$ and I put $\Omega = (a,b)$. For $d>1$, the embedding is general not true if you do not assume boundary regularity on $\Omega$. (This is usually used to transfer the problem to the full space which opens another can of worms here..) Jul 15, 2021 at 8:34
• One practical problem I see here is that you have defined the Sobolev space, but you never actually have defined the notion of a weak derivative. I am not sure that there are many applications of functions $u \in W^{1,p}(\Omega)$ that do not at some point explicitly use that $\nabla u$ is a well defined $L^p$-function. But if you have to show that anyway, you could just have started with the usual definition in the first place.
– mlk
Jul 15, 2021 at 9:48
• (1) By $C^1$ I guess you refer to the space of continuously differentiable functions. If so, you may take functions that has graphs with finitely many kinks. Take a point $a\in\Omega$ and a ball $B(a,r)\subset\Omega$ centered at $x$. Let $f(x)=|x-a|$ on $B(a,r)$ and equal to $r$ elsewhere. (2) You are looking for the (proof of the) Sobolev embedding theorem. Jul 15, 2021 at 15:11
• thank you for the comments and the answer to question 1. I have updated what I want related to question 2. Jul 20, 2021 at 8:09

Preliminary remark. The equivalence of the usual definition of Sobolev spaces and the property \eqref{1} can for instance be found in Proposition 9.3 in this book by Brezis. It is even true in the vector-valued case under appropriate assumptions - see Theorem 2.2 in this article by Arendt and Kreuter, or Proposition 2.5.7 of this book by Hytönen, van Neerven, Veraar and Weis.

I would not say that it yields an alternative approach to the theory of Sobolev spaces, but rather that it is a useful tool the complements the classical definition via weak derivatives. I called it a tool because it is useful in the theory of Sobolev spaces on various occasions - also if you still you the classical definition.

Applications. Here are several applications of the characterization \eqref{1} of $$W^{1,p} := W^{1,p}(\Omega)$$. Let us assume, for the sake of simplicity, that $$\Omega$$ is bounded. (For more general result, please see the references provided below.) Fix $$p \in (1,\infty]$$

1. On bounded domains, the property \eqref{1} is obviously satisfied by every Lipschitz continuous function. So it follows that all Lipschitz continuous functions on $$\Omega$$ are in $$W^{1,p}$$ (as also pointed out in a comment by Hannes).

2. One can also deduce from \eqref{1} that for every Sobolev function $$u \in W^{1,p}$$ and every Lipschitz continuous $$f: \mathbb{R} \to \mathbb{R}$$, the composition $$f \circ u$$ is again in $$W^{1,p}$$.

More details about 1. and 2. can, for instance, be found in Corollary 2.7 in Arendt and Kreuter - which even deals with the vector-valued case.

1. In Remark 7 on page 269 in Brezis it is noted that \eqref{1} implies that every function in $$W^{1,\infty}$$ has a continuous representative.

2. Property \eqref{1} is used in the proof of the Rellich–Kondrachov theorem in Brezis (Theorem 9.16).

3. In Theorem 2.5.17 of Hytönen et. al., properpty \eqref{1} is used (in the vector-valued case) is used to prove an interpolation property of Sobolev spaces.

Summary. The property \eqref{1} is very useful - but I'd consideration it a complement of the classical definition rather than an alternative.

• Thank you for the answer! Ironically, I own both books (they are excellent!). However, even by referring to them, I haven't seen a satisfactory answer to my second question, which tries to prove continuity of weakly differentiable functions directly from $(*)$. Why am I seeking this? It's because $(*)$ mimics a continuity property, and if you've read my suggestion (before question 3), I think we can use convolutions in a smart way to get pointwise inequalities. I'll upvote the answer, but I'll not accept it yet! Jul 20, 2021 at 12:16
• @UserA: Thanks for your response! I think I understand a bit better now what you're after. Jul 20, 2021 at 19:02