3
$\begingroup$

Let $$V = \left\{ {v \in {H^2}(0,1):{v_x}\left( 0 \right) - v\left( 0 \right) = {v_x}\left( 1 \right) + v\left( 1 \right) = 0} \right\}.$$

Question: Is $V$ dense in ${H^1}\left( {0,1} \right)$?

I know ${C^\infty }\left( \mathbb{R} \right)$ is dense in ${H^1}\left( {0,1} \right)$. But I can't prove that $V$ is dense in ${H^1}\left( {0,1} \right)$ .

$\endgroup$
4
  • $\begingroup$ Welcome to MathOverflow! May I ask what the background of this question is? I'm asking because the imperative mood in which the question is phrased suggests that it might be a homework problem. $\endgroup$ Sep 13, 2019 at 10:29
  • 1
    $\begingroup$ It's not homework problem. I consider the nonlinear wave equation with boundary conditions $${u_x}\left( {0,t} \right) - u\left( {0,t} \right) = {u_x}\left( {1,t} \right) + u\left( {1,t} \right) = 0,$$ and inital conditions $$u\left( {x,0} \right) = {u_0}\left( x \right),\,{u_t}\left( {x,0} \right) = {u_1}\left( x \right).$$ I proved the existence of weak solution with ${u_0} \in V$. Now i want to prove the existence of weak solution with ${u_0} \in {H^1}$ by by a density argument. $\endgroup$ Sep 13, 2019 at 14:52
  • $\begingroup$ Thanks for your response! I slightly changed the wording of your post and replaced the imperative with a question. $\endgroup$ Sep 13, 2019 at 19:35
  • 1
    $\begingroup$ Concerning your question itself: Yes, $V$ is dense in $H^1(0,1)$, but the proof that I just came up with is quite weird (the Laplace operator with the same boundary conditions can be define via form methods with form domain $H^1(0,1)$ and domain $V$; then one uses the spectral theorem for self-adjoint operators to see that the domain is dense in the form domain). I'll try to find a simpler proof (I strongly suspect that there is one); if I don't succeed, I'll post the details of the proof that I just sketched. $\endgroup$ Sep 13, 2019 at 19:40

2 Answers 2

6
$\begingroup$

The space $V$ is dense in $H^1(0,1)$. Here are three (more or less) different proofs:

Proof 1 (the pedestrian way):

We set $H^1 := H^1(0,1)$ and $H^2 := H^2(0,1)$.

Fix $h \in H^1$. Since $H^2(0,1)$ is dense in $H^1(0,1)$, there exists a sequence $(f_n)_{n \in \mathbb{N}} \subseteq H^2(0,1)$ that converges to $h$ with respect to the $H^1$-norm.

Now, for each index $n$ we can construct a $H^2$-function $g_{n,0}$ which is zero on $[1/2,1]$ and satisfies \begin{align*} g_{n,0}(0) = 0, \quad g_{n,0}'(0) = f_n'(0) - f_n(0) \quad \text{and} \quad \|g_{n,0}\|_{H^1} \le 1/n \end{align*} (see below for details). Similary we can construct, for each index $n$, a $H^2$-function $g_{n,1}$ which is zero on $[0,1/2]$ and satisfies \begin{align*} g_{n,1}(1) = 0, \quad g_{n,1}'(0) = f_n'(1) + f_n(1) \quad \text{and} \quad \|g_{n,1}\|_{H^1} \le 1/n. \end{align*} Now $(f_n - g_{n,0} - g_{n,1})_{n \in \mathbb{N}}$ is a sequence in $V$ that converges to $h$ with respect to the $H^1$-norm.

Note. In the above proof we used the following

Lemma. For each (real or complex) number $\alpha$ and each positive integer $n$ there exists an $H^2$-function $g$ which is $0$ on $[1/2,1]$ and satisfies \begin{align*} g(0) = 0, \quad g'(0) = \alpha \quad \text{and} \quad \|g\|_{H^1} \le 1/n. \end{align*}

Proof of the lemma. Consider any number $\delta \in (0,1/3)$. Let $u$ be the piecewise linear function on $[0,1]$ which is defined by the following values: \begin{align*} \begin{cases} u(0) & = 1, \\ u(\delta) & = 0, \\ u(2\delta) & = -1/2, \\ u(3\delta) & = 0, \\ u(1) & = 0. \end{cases} \end{align*} Then $u$ is an $H^1$-function, so the function $g$ defined by $g(x) = \alpha \int_0^x u(y) \, dy$ is in $H^2$. Moreover, we have $g(0) = 0$ and $g'(0) = \alpha u(0) = \alpha$.

The modulus of $g'(x) = \alpha u(x)$ is always in $[0,|\alpha|]$ for $x \in [0,3\delta]$ and it is $0$ for $x > 3\delta$; this shows that $\|g'\|_{L^2} \le |\alpha| \cdot \sqrt{3\delta}$. Further we have $|g(x)| \in [0,|\alpha|\cdot 3\delta]$ for $x \in [0,3\delta]$, and $g(x)=0$ for $x \in [3\delta,1]$, so $\|g\|_{L^2} \le |\alpha| \cdot \sqrt{3\delta}^3$. Hence, if we choose $\delta$ sufficiently small we can achieve $\|g\|_{H^1} \le 1/n$.

Finally, as $g$ vanishes on $[3\delta,1]$, we can also ensure that $g$ is zero on $[1/2,1]$ by choosing $\delta \le 1/6$. This proves the lemma.

Proof 2 (via abstract nonsense):

For people who like abstract functional analysis more than computations, here is proof which I would consider to be a bit more conceptual. It is based on the following general observation:

Proposition. Let $W$ be a normed space (not necessarily complete!) over $\mathbb{K} \in \{\mathbb{R}, \mathbb{C}\}$ and let $\varphi: W \to \mathbb{K}$ be discontinuous linear functional. Then $\ker \varphi$ dense in $W$.

Proof. First show that $\ker \varphi$ is not closed, so assume the contrary. Then the quotiend space $W / \ker \varphi$ is a normed space, the quotient map $q: W \to W/\ker \varphi$ is continuous, and $\varphi$ induces a linear map $\tilde \varphi: W/\ker \varphi \to \mathbb{K}$ such that $\varphi = \tilde \varphi \circ q$. As $W / \ker \varphi$ is one-dimensional, $\tilde \varphi$ is continuous and hence, so is $\varphi$. This contradicts our assumption.

Hence, $\ker \varphi$ is a non-closed subspace of $W$ of co-dimension $1$. As the closure of $\ker \varphi$ has to be strictly larger than $\ker \varphi$, it follows that the closure coincides with $W$. This proves the proposition.

To apply this result to the question, choose $W = H^2(0,1)$ and endow this space with the $H^1$-norm (!). Since point evaluations are continuous with respect to the $H^1$-norm, but point evaluations of derivatives are not, it follows that $\varphi: H^2(0,1) \ni u \mapsto u'(0) - u(0) \in \mathbb{K}$ is not continuous with respect to the $H^1$-norm, so $\ker \varphi$ is dense in $H^2(0,1)$ with respect to the $H^1$-norm.

Similary, we consider the functional $\psi: H^2(0,1) \ni u \mapsto u'(1) + u(1) \in \mathbb{K}$. Since $\ker \varphi$ contains all $H^2$-functions which are constantly $0$ on $[0,1/2]$, we can see that the restriction of $\psi$ to $\ker \varphi$ is also discontinuous.

Hence, $V = \ker \psi \cap \ker \varphi$ is dense (with respect to the $H^1$-norm) in $\ker \varphi$, and the latter space is dense in $H^2$ (with respect to the $H^1$-norm), and $H^2$ is dense in $H^1$. This proves that $V$ is dense in $H^1$.

Disclaimer. When we compare Proof 2 with Proof 1, one might argue that Proof 2 looks simpler only on a superficial level, since I left out that details which show that point evaluations of the derivatives are not continuous with respect to the $H^1$-norm. The main advantage that I see in Proof 2 is that it reduces the assertion to a fact for which one usually has a good intuition ("point derivatives are not continuous with respect to the $H^1$-norm") even if one does not have much experience with Sobolov spaces.

Proof 3 (form methods and the Robin Laplace operator):

I'll only sketch this argument.

Define the bilinear form $a: H^1 \times H^1 \to \mathbb{C}$ by \begin{align*} a(u,v) = \int_0^1 u'(y) \overline{v'(y)} \,dy + u(0) \overline{v(0)} + u(1) \overline{v(1)}. \end{align*} Then $a$ is a symmetric and coercive bilinear form which induces a self-adjoint unbounded operator $A$ on $L^2(0,1)$. One can check that $A$ is equal to $-\Delta$ with domain $D(A) = V$, i.e. $A$ is minus the Laplace operator with so-called Robin boundary conditions.

By using the spectral theorem for self-adjoint operators one can see that, for each self-adjoint operator induced by a coercive form, the domain of the operator is dense in the corresponding form domain. Hence, $V$ is dense in $H^1(0,1)$.

Reference: For details about bilinear forms and operators induced by them, see for instance the lecture notes of the 18th Internet Seminar on Evolution Equations.

$\endgroup$
1
  • 1
    $\begingroup$ I like them all, yet I think the true proof of the plain density claim is number 2. We may also apply the lemma to $u\mapsto u'(0)$ as a discontinuous linear functional say on $H^2\cap H^1_0([0,1/10])$ with the $H^1$-norm. As a consequence, arbitrarily close to any function $v\in H^2(0,1)$ wrto the $H^1$ distance there is a function $w\in H^2(0,1)$ with arbitrary derivative at $0$ (and with $v=w$ on $[1/10,1]$ so that we can also perturb on the other side independently and find an approximation of $v$ in $V$). $\endgroup$ Sep 14, 2019 at 14:54
3
$\begingroup$

Consider, for $m\ge2$, the function $\varphi_m(x):=x(1-x)^m$ . So $$\varphi_m(0)=0\qquad \varphi_m(1)=0$$ $$ \varphi_m'(0)=1\qquad \varphi_m'(1)=0\ .$$ It is also easy to see that $\| \varphi_m'\|_2=O(m^{-{1/2}})$ (in fact, we may compute exactly both $\| \varphi_m\|_2^2$ and $\| \varphi_m'\|_2^2$ in terms of some rational functions of $m$, by means of a few Beta function integrals).

For any $v\in H^2(0,1)$ define

$$v_m(x):=v(x)+\big(v(0)-v'(0)\big)\varphi_m(x)+\big(v(1)+v'(1)\big)\varphi_m(1-x)$$

so $$v_m(0)=v(0)\qquad v_m(1)=v(1)$$ $$v_m'(0)=v(0)\qquad v_m'(1)=-v(1)$$

that is $v_m\in V$; clearly $\|v-v_m\|_{1,2}=O(m^{-{1/2}})$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.