As far as I can tell, nodal interpolation estimates proven with Bramble / Hilbert require higher regularity of the functions being interpolated. E.g. with linear elements in 2 dimensions, one needs $v \in W^{m,p}$ with $m p > 2, p \geq 1$ in order to have

$$ |v - I(v)|_{W^{r,p}} \leq C h^{m-r} |v|_{W^{m,p}}, $$

where $I$ is the nodal interpolant, $| \cdot |$ are seminorms in the indicated spaces and $0 \leq r \leq m$. In particular note that the bound is using the $m$-th derivative of $v$.

For my application I only have $m = 1$ and $p \leq 2$. More specifically this is what I want:

Let $V_h$ be a continuous $P_1$ space over a polygonal domain $U \subset \mathbb{R}^2$ with a quasi-uniform mesh, and $(v_h) \subset V_h$ a sequence bounded in $W^{1,2}(U)$. For each $h$ let $I_h:C^0 \rightarrow V_h$ be some interpolant (I'd use the nodal interpolant but I'm not working with $W^{2,2}$) and $q$ some quadratic function. To simplify maybe take $q(x)=x^2$. I want to show

$$ \|q(v_h) - I_h(q(v_h))\|_{L^2(U)} \leq C h^\alpha \|D q(v_h) \|_{L^2(U)}$$

for any $\alpha > 0$ which makes this work. Reading some papers, I see people change $V_h$ to be discontinuous and claim to have bounds using just the first derivative of $q \circ v_h$, but I don't know where to find precise statements and proofs of the properties of the interpolant in this case. I'm happy to make $V_h$ discontinuous, I just want the damn thing to converge :)

Any help, anyone?

  • $\begingroup$ It's not exactly clear to me what you are looking for. If $v$ is really only in $W^{1,2+\varepsilon}$, then the estimate $\|v-I_h v\|_{L^{2+\alpha}} = {\cal O}(h^{1+\alpha})$ is likely sharp. You will of course get something better if you knew that $v$ is quadratic, but the point is that quadratic functions are far smoother than just in $W^{1,2+\alpha}$. $\endgroup$ Jan 26, 2018 at 0:13
  • $\begingroup$ I don't know if I understand you well, but I am not looking for faster rates. I am looking for a bound on $\|q(v_h) - I_h(q(v_h)) \|$ which uses the first derivative $D q(v_h)$ instead of $D^2 q(v_h)$, because I don't have it. Since $v_h \in V_h$ I do have it element-wise and I thought maybe one can reason in an element, then use some sort of reverse-Poincaré / Caccioppoli inequality, then patch everything together but I don't have any equation that $v_h$ solves locally and there were more issues... well, it didn't get me far. $\endgroup$
    – Miguel
    Jan 26, 2018 at 7:55
  • $\begingroup$ I think I'm confused because on the right hand side of your first equation, you have $|v|_{W^{m,p}}$, which only has first derivatives if you choose $m=1$. $\endgroup$ Jan 26, 2018 at 13:30
  • $\begingroup$ Indeed, but the problem is that I can't just plug in $m=1$ because I need $m p \gt 2$ and have $p=2$. Oh! I see now that I forgot to specify that and only wrote it in the final paragraph, sorry. $\endgroup$
    – Miguel
    Jan 26, 2018 at 14:10
  • 1
    $\begingroup$ I know you stated $mp>2$, but you can choose $m=1$, $p=2+\alpha$. A small $alpha$ isn't going to change much in your approach, and you still only need first derivatives. $\endgroup$ Jan 26, 2018 at 16:59

1 Answer 1


What I was looking for is Clément interpolation, which is defined by averaging a target function $v \in L^1$ over cells sharing a node and assigning this value to the interpolant at that node (or equivalently by local $L^2$ projections). It is explained in detail in §6.3 of these lecture notes of Volker John at the WIAS.

Specifically, the estimate obtained for $v \in W^{1,q}$ is

$$ \| D^k(v - I_h^{\operatorname{cle}}(v)) \|_{L^q} \leq C h^{l-k} \| D^l v\|_{L^q}$$

for all $0 \leq k \leq l \leq 2 $ and $1 \leq q \leq \infty$, assuming a nice enough mesh.

Alternatively, I guess I could just use global $L^2$ projection instead of interpolation, which guarantees the convergence I want, but I need to use the same operator in other places and I don't want my code to be performing costly projections.


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.