Let $\Omega$ be an open (non empty) set and $u:\Omega \subset \mathbb{R}^N \to \mathbb{R}^M$ be a function such that the Hausdorff dimension of its graph is $N$.

Let $\tilde u = u$ a.e. Is it true that the Hausdorff dimension of the graph of $\tilde u$ is also $N$?

  • $\begingroup$ Maybe you would like $\Omega$ to have strictly positive Lebesgue measure, e.g. be nonempty and open. Otherwise the equality $\tilde u = u$ a.e. might always be true (and I think the Hausdorff dimensions might differ in this case). $\endgroup$ – Skeeve Apr 6 '19 at 22:19
  • $\begingroup$ @Skeeve You're right, of course. $\endgroup$ – Riku Apr 7 '19 at 11:35
  • $\begingroup$ Can you do the case $N=M=1$? $\endgroup$ – Gerald Edgar Apr 9 '19 at 11:56
  • $\begingroup$ @GeraldEdgar No. $\endgroup$ – Riku Apr 9 '19 at 12:14
  • 4
    $\begingroup$ OK maybe a counterexample can be done like this. $f : \mathbb R \to \mathbb R$, $f$ is zero except on the Cantor set, but on the Cantor set $f$ is so wild that its graph has dimension ${}\gt 1$. $\endgroup$ – Gerald Edgar Apr 9 '19 at 12:26

In addition to a very concise answer by Alex Ravsky let me address the case $N=M=1$, following the comment by Gerald Edgar. I am going to omit some technical details (which can be added if necessary).

The classical Cantor function is generated by the function $g$ which is defined by $g(0)=0$, $g(\frac13) = \frac12$, $g(\frac23) = \frac12$, $g(1)=1$ (and interpolated linearly). Now repeat the construction for a different function $g$, defined by $g(0)=0$, $g(\frac13) = \frac23$, $g(\frac23) = \frac13$, $g(1)=1$ (and interpolated linearly).

For instance, at the third step of Cantor's iterative construction we obtain the following function:


Let $u$ denote the limit function, redefined to be zero outside of the Cantor set $C$ (which explains why some pieces of the graph are dotted). Then for $\tilde u \equiv 0$ it is evident that $u = \tilde u$ a.e.

However $\alpha = \frac{\ln 4}{\ln 3} > 1$ is the Hausdorff dimension of the graph of $u$.

Indeed, it is possible to cover the graph $\Gamma_C = \{(x,u(x)) : x\in C\}$ with $4^n$ balls with of $3^{-n}$, which gives the upper estimate $\dim_H \Gamma_C \le \alpha$. For the lower estimate one can consider the image $\mu$ of Cantor measure on $C$ (which is the weak derivative of the classical Cantor function) under the mapping $x\mapsto (x,u(x))$. The measure $\mu$ is supported on $\Gamma_C$ and it is possible to show that there exists a constant $\kappa>0$ such that for any $x\in \Gamma_C$ it holds that $\mu(B_r(x)) \le \kappa \cdot r^\alpha$. Then by Lemma 1.2.8 from [1] it holds $\dim_H \Gamma_C \ge \alpha$.

Dimensions of graphs of some other functions are also discussed in [1].

[1] Bishop C.J., Peres Y. Fractals in probability and analysis. Cambridge; New York: Cambridge university press, 2017.

  • $\begingroup$ Thank you. Can you show that $u$ is not of bounded variation? $\endgroup$ – Riku Apr 10 '19 at 17:53
  • $\begingroup$ @Riku Strictly saying $u$ has zero total variation because it is zero a.e., read my answer more carefully. But if you refer to the limit function $f$ before redefining it to be zero outside of the Cantor set then it is easy to see that $f$ has infinite total variation. $\endgroup$ – Skeeve Apr 10 '19 at 18:06
  • $\begingroup$ Yes, I was thinking about the limit function. Why does it not have bounded total variation? $\endgroup$ – Riku Apr 10 '19 at 18:11
  • $\begingroup$ Well, simply by definition. The total variation is at least $2^n \frac{2^n}{3^n}$ for every $n\in\mathbb N$ (look at the blue pieces on my picture). $\endgroup$ – Skeeve Apr 10 '19 at 18:15
  • $\begingroup$ I see; thank you. Also, what software&code did you use to draw that picture? $\endgroup$ – Riku Apr 11 '19 at 11:45

Put $N=1$, $M=2$, $\Omega=\Bbb R^N$, and $u(x)=(x,0)$ for each $x\in\Bbb R^N$. Then the graph of $u$ is a straight line, so it has Hausdorff dimension $1=N$. On the other hand, let $C\subset [0,1]$ be a Cantor set and $f:C\to [0,1]^2$ be a surjective map. For each $x\in\Bbb R^N $ put $\tilde u(x)=f(x)$, if $x\in C$ and $\tilde u(x)=u(x)$, otherwise. Since a projection $\pi$ of the graph $\Gamma(\tilde u)$ onto the image $\tilde u(\Omega)$ is a non-expanding map (that is the distance between $\pi(x)$ and $\pi (y)$ is not bigger than the distance between $\pi(x)$ and $\pi (y)$ for each $x,y\in \Gamma(\tilde u)$), we have that $\dim_H \Gamma(\tilde u)\ge \dim_H \tilde u(\Omega)\ge \dim_H [0,1]^2=2$.


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.