6
$\begingroup$

Let $1 \leq p < \infty$ and $u \in W^{1,p}(\mathbb{R}$). Set $$ D_{h}u(x) = \frac{1}{h}(u(x+h) - u(x)), \ \ x \in \mathbb{R}, h> 0 $$ Show that $D_{h}u \to u'$ in $L^{p}(\mathbb{R}$) as $h \to 0$.

I'm trying to use the fact that $C_{c}^{1}(\mathbb{R}$) is dense in $W^{1,p}(\mathbb{R}$)

$\endgroup$
6
  • $\begingroup$ The Mean Value Theorem and Dominated Convergence Theorem should do it, if I understand your question correctly. $\endgroup$ Commented May 2, 2022 at 20:04
  • $\begingroup$ cross posted in math stackexchange. math.stackexchange.com/questions/4441553/exercise-8-13-brezis $\endgroup$
    – kodlu
    Commented May 2, 2022 at 20:14
  • 1
    $\begingroup$ @DavidHughes Did not help!!! $\endgroup$
    – user253963
    Commented May 2, 2022 at 20:53
  • 1
    $\begingroup$ @kodlu Did not help $\endgroup$
    – user253963
    Commented May 2, 2022 at 20:53
  • 1
    $\begingroup$ @Yuval Peres yes you are correct. I misread the question and thought it was only for smooth $u$. $\endgroup$ Commented May 2, 2022 at 22:30

2 Answers 2

18
$\begingroup$

The proof is not short, because it is done from first principles, without using any theorems about Sobolev space except its definition.

By the definition of $W^{1,p}$, there exist $v_n \in C_{c}^{1}(\mathbb{R})$ and $w \in L^p(\mathbb{R})$ such that $v_n \to u$ in $L^p(\mathbb{R})$ and $v_n' \to w$ in $L^p(\mathbb{R})$. In this case we write $u'=w$. Note that $v_n'$ is a classical derivative, so $$D_h v_n(x)=\frac{v_n(x+h)-v_n}{h}= I_h(v_n')(x) \,, \tag{1} $$ where for $f\in L^p(\mathbb{R})$, we write $$I_h(f)(x):=\int_0^h \frac{f(x+t)}{h} \,dx \,.$$ By Jensen's inequality [1], for all $n \ge 1 $ and $x \in \mathbb{R}$, we have $$|I_h(v_n')(x)-I_h(u')(x)|^p \le \frac{1}{h} \int_0^h |v_n'(x+t)-u'(x+t)|^p \,dt \,. $$ Integrating both sides $\,dx$ and using Fubini on the right-hand side, we obtain $$\|I_h(v_n') -I_h(u') \|^p \le \frac{1}{h}\int_0^h \|v_n'(\cdot+t)-u'(\cdot+t)\|_p^p \,dt= \|v_n' -u' \|_p^p \,.\tag{2} $$

Given $\epsilon>0$, find $k$ such that $$ \|v_k'-u'\|_p<\epsilon \,. \tag{3} $$ Let $M$ denote the Lebesgue measure of the support of $v_k$. Since $v_k'$ is uniformly continuous, there exists $h_0\in(0,1)$ such that $$\forall t\in [0, h_0], \quad \sup_{x \in \mathbb{R}} |v_k'(x+t)-v_k'(x)|<\epsilon/(M+1) \,,$$ so for $h\in [0, h_0]$ and all $x$, we have $|I_h (v_k')(x)-v_k'(x)|<\epsilon/(M+1)$, whence $$\|I_h (v_k') -v_k'\|_p^p \le (M+h) (\epsilon/(M+1))^p <\epsilon^p \,.$$ In conjunction with $(2)$ and $(3)$, this gives $$\|I_h(u')-u' \|_p \le \|I_h(u')-I_h(v_k') \|_p + \|I_h(v_k')-v_k' \|_p + \| v_k' -u'\|_p <3\epsilon \,. \tag{4}$$

Next, fix $h\in [0, h_0]$, and choose $m=m(h,\epsilon)$ such that $$\|u-v_m \|_p<\epsilon h \quad \text{and} \quad \|u'-v_m'\|_p<\epsilon \,. $$ The first inequality implies that $\|D_h(u) -D_h(v_m) \|_p<2\epsilon$. Therefore, by $(1),\, (2)$ and $(4)$, \begin{eqnarray} \|D_h(u)-u'\|_p &\le& \|D_h(u) -D_h(v_m) \|_p+\|I_h(v_m')-I_h(u')\|_p+\|I_h(u')-u' \|_p \\ &<& 2\epsilon+\epsilon+3\epsilon=6\epsilon \,. \end{eqnarray} This completes the proof.

[1] https://en.wikipedia.org/wiki/Jensen%27s_inequality#Measure-theoretic_and_probabilistic_form

$\endgroup$
2
  • $\begingroup$ Thank you very much!! $\endgroup$
    – user253963
    Commented May 2, 2022 at 22:58
  • 5
    $\begingroup$ @AGL. I hope you find it clear. This took me a couple of hours to write down carefully without using any prior theorems. I assume you know how to accept an answer? $\endgroup$ Commented May 2, 2022 at 23:00
6
$\begingroup$

$\newcommand{\R}{\mathbb R}\newcommand{\ep}{\varepsilon}$Re-define, if needed, the function $u$ on a set of Lebesgue measure zero so that \begin{equation} u(y)-u(x):=\int_x^y dt\, u'(t) \end{equation} for all real $x$ and $y$ such that $x\le y$. Then we can write \begin{equation} (D_hu)(x) =\frac{u(x+h)-u(x)}h =\int_0^1 ds\,u'(x+sh), \end{equation} so that
\begin{equation} \begin{aligned} \|D_hu-u'\|_p^p&=\int_\R dx\,\Big|\int_0^1 ds\,(u'(x+sh)-u'(x))\Big|^p \\ &\le\int_\R dx\,\int_0^1 ds\,|u'(x+sh)-u'(x)|^p \\ &=\int_0^1 ds\,\int_\R dx\,|u'(x+sh)-u'(x)|^p=:I_h(u'). \end{aligned} \end{equation}

Take any real $\ep>0$. Since $C_c(\R)$ is dense in $L^p(\R)$, there is a function $v\in C_c(\R)$ such that $\|u'-v\|_p\le\ep$. Using the Jensen inequality \begin{equation} \Big(\frac{a+b+c}3\Big)^p\le\frac{a^p+b^p+c^p}3 \end{equation} for nonnegative $a,b,c$, we get \begin{equation} |u'(x+sh)-u'(x)|^p\le3^{p-1}(|v(x+sh)-v(x)|^p+|u'(x+sh)-v(x+sh)|^p+|u'(x)-v(x)|^p), \end{equation} whence \begin{equation} \begin{aligned} I_h(u')\le3^{p-1}(I_h(v)+2\ep^p). \end{aligned} \end{equation} It remains to show that $I_h(v)\to0$ as $h\to0$. But this follows because $v$ is in $C_c(\R)$, so that $v$ has a compact support and is uniformly continuous.

$\endgroup$
7
  • $\begingroup$ The first identity in your answer, $$(D_hu)(x) =\frac{u(x+h)-u(x)}h =\int_0^1 ds\,u'(x+sh)$$ needs some justification, as $u$ is just a function in Sobolev space, and is formally just defined a.e. In particular, $u'$ is not a classical derivative. $\endgroup$ Commented May 2, 2022 at 22:14
  • 1
    $\begingroup$ I want to thank everyone $\endgroup$
    – user253963
    Commented May 2, 2022 at 22:59
  • $\begingroup$ @YuvalPeres : We can just re-define the function $u$ appropriately on a set of measure $0$ so as to make $u$ absolutely continuous -- and this is what I meant. I have now added the corresponding detail. I think this only simplifies the proof a bit. $\endgroup$ Commented May 3, 2022 at 0:40
  • 1
    $\begingroup$ @Iosif Pinelis That is correct, but this identity does not follow from the definition of the weak derivative in Sobolev space- it needs an additional argument or reference. Even after adding that, your proof is still likely to be shorter than mine. One place where this is discussed is the book by Evans and Gariepy, Theorem 4.9.1 $\endgroup$ Commented May 3, 2022 at 3:06
  • $\begingroup$ @YuvalPeres : I think this is a well-known fact. E.g., Wikipedia has this: "In the one-dimensional problem it is enough to assume that the ${\displaystyle (k{-}1)}$-th derivative ${\displaystyle f^{(k-1)}}$ is differentiable almost everywhere and is equal almost everywhere to the Lebesgue integral of its derivative" (en.wikipedia.org/wiki/Sobolev_space#One-dimensional_case) $\endgroup$ Commented May 3, 2022 at 3:11

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .