1
$\begingroup$

I'm having lots of troubles in understanding the proof the estimation of the classical $H^1$ error using finite elements of degree $r$.

$$||u-u_h||_{H^1(\Omega)} \leq \frac{M}{\alpha} C h^r |u|_{H^{r+1}(\Omega)}$$ where $M$ and $\alpha$ are the operator norm and the coercivity constant, respectively and $C$ is a constand independent on $h$.

My book (Quarteroni - Numerical methods for differential problems) shows this by using the Cèa's lemma and the estimate on the seminorm of the interpolation error. More precisely, it states:

$$||u-u_h||_{H^1(\Omega)} \leq \frac{M}{\alpha} \inf_{v_h \in V_h} ||u - v_h||_{H^1(\Omega)} \leq \frac{M}{\alpha} ||u-\Pi^ru||_{H^1(\Omega)}$$ Now use the estimate $$|u-\Pi^r u|_{H^m(\Omega)} \leq Ch^{r+1-m} |u|_{H^{m}(\Omega)}$$ (valid for $m=0,1$ and $r\geq 1$. ) with $m=1$ to conclude

The very last step is what I cannot do: I can't see how to pass from the $H^1$ seminorm to the full $H^1$ norm. I mean, if I need to bound $||u-\Pi^ru||_{H^1(\Omega)}$, how can I use the bound on the seminorm?

$\endgroup$
1
$\begingroup$

I think $u - \Pi u$ is zero on a part of the boundary. Then you can use Poincare inequality to bound $\| u - \Pi u \|_0 \leq C| u - \Pi u|_1$.

It could also be possible to look at the case $m = 0$ and argue the $L^2$ part is of higher order and, hence, smaller in the asymptotic limit.

$\endgroup$
8
  • $\begingroup$ Poincarè inequality was my first guess indeed, but I wasn't able to say whether or not it's $0$ on the boundary. All I know is that $\Pi u$ is equal to $u$ on the dofs, by definition. Why do you think it can be $0$ on some non trivial part of the boundary? @knl $\endgroup$ Jul 10 at 22:45
  • $\begingroup$ Well, you probably have some boundary condition to make the solution unique? $\endgroup$
    – knl
    Jul 11 at 13:56
  • $\begingroup$ This is usually done by assuming that the boundary values $u|_{\partial\Omega}$ are piecewise polynomial so that $(u-\Pi u)|_{\partial\Omega}=0$. The proof can be extended to arbitrary boundary conditions, but that adds a substantial level of complexity. $\endgroup$ Jul 11 at 23:35
  • $\begingroup$ @bobinthebox Does this answer the question? $\endgroup$
    – knl
    Jul 14 at 10:10
  • $\begingroup$ Actually before accepting I was trying to verify whether or not that was an assumption on the book. I think that what the author is using is the Bramble-Hilbert lemma $$||\tau(u)||_{s,p,\Omega} \leq ||\tau|| |u|_{k+1,p,\Omega}$$ for $\tau:W^{k+1,p}(\Omega \rightarrow W^{s,p}(\Omega)$ and $s \leq k$. @knl Do you agree? $\endgroup$ Jul 14 at 17:58

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.