7
$\begingroup$

Let $f$ be a continuous real-valued function defined on an $n$ dimesional simplex $\Sigma\subset \mathbb{R}^n $. The classical simplicial approximation scheme provides a sequence $f_k$ of piecewise affine functions converging uniformly to $f$: namely, on each simplex $S$ of the $k$-th baricentric subdivision of $\Sigma$, one takes ${f_k} _ {|S}$ to be the affine interpolation of the data of $f$ on the vertices of $S$. Precisely, if $\omega$ is a modulus of continuity for $f$, it follows easily that $$ \big\| f - f_k \big\|_{\infty,\Sigma}\, \le\, \omega\bigg( \Big(\frac{n}{n+1}\Big )^k \operatorname{diam}\Sigma \bigg)\, .$$

Notice that $\nabla f_k(x) $ is defined and locally constant at any point $x\in\Sigma$ not in the $(n-1)$-skeleton of the $k$-th barycentric subdivisions; in particular, a.e. in $\Sigma$. However, understanding the behaviour of this sequence is not as simple as before, even for smooth $f$. For instance, the norm of $\nabla f_k(x) $ is in general not bounded by the uniform norm of $\nabla f $.

Questions. Assuming that $f$ has bounded and continuous first order derivatives, can we conclude that $\nabla f_k $ converges to $\nabla f $ in $L^1(\Sigma)$? How to dominate pointwise $|\nabla f_k (x) |$ (independently of $k$), and how to bound $\|\nabla f_k -\nabla f\|_{1,\Sigma} $ ?

$\endgroup$
1
  • $\begingroup$ It seems to me that this works using essentially the same proof you would use for proving that a piecewise linear approximation scheme for a $C^1$ function on an interval converges in $L^1$ (and maybe it even converges weakly in $L^\infty$?). The rough idea is that you are approximating an integral (of the gradient) by a Riemann sum (sum of the gradient times the area of each simplex). $\endgroup$
    – Deane Yang
    Jun 14, 2012 at 12:27

1 Answer 1

4
$\begingroup$

You can represent on every simplex $\sigma$ of your triangulation the derivative $D f_k$ by $$ D f_k = \frac{1}{\lvert \sigma \rvert} \sum_{i = 0}^n \int_{\sigma} \Bigl(\frac{1}{(1 - \beta_i (x))^n} - 1\Bigr)\, Df(x)[a_i - x]\,dx D\beta_i, $$ where $\beta_i$ are the maps that give barycentric coordinates $$ \beta_0 (x) + \dots + \beta_n (x) = 1 \quad \text{ and } \quad \beta_0 (x)a_0 + \dots + \beta_n (x)a_n = x. $$ For a proof, see J. Van Schaftingen, Approximation in Sobolev spaces by piecewise affine interpolation, [arXiv:1312.5986]).

The kernel appearing in this integral representation is absolutely integrable ($a_i - x$ controls $\beta_i (x)$), uniformly with respect to simplices that do not degenerate. This should imply in fact an $L^\infty$ approximation property under your assumptions. (If the function $f$ is merely in a Sobolev space, one can improve the convergence by translating the triangulation points.)

$\endgroup$

Your Answer

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

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