I know this is probably a easy question, but some steps in the proofs I found almost everywhere contained some parts or assumptions which I think may not be that trivial, so I would like to make it rigorous and clear enough. Here is the question:

Let $C$ be the Cantor set with delete the middle third of the interval and go on. The general Cantor can be considered similarly. We want to proof the Hausdorff dimension of $C$ is $\alpha:=\log2/\log3$. So we calculate the $d$-dimensional Hausdorff measure $H^d(C)$ for all $d$ to determine the Hausdorff dimension. Let $C(k)$ be the collection of $2^k$ intervals with length $1/3^k$ in the $k^{th}$-step of construction of Cantor set.

It is rather easy to show that $H^{\alpha}(C)<\infty$ by showing that for any covering $\{E_j\}_{j=1}^{\infty}$of $C$ the set $C(k)$ also covers $C$ for $k$ large enough, so we can bound $H^{\alpha}(C)$ from above. Which implies that the Hausdorff dimension of $C$ is less than $\alpha$.

To show the dimension is actually equal to $\alpha$, it suffices to show $H^{\alpha}(C)>0$.

Now let $\{E_j\}_{j=1}^{\infty}$ be any covering of $C$ with diameter $diam(E_j)\le \delta$ for all $j$. How do we show that $$\sum_j diam(E_j)^{\alpha}>constant$$

One author (see this link) made the following assumption: $E_j$ be open, so one can find the Lebesgue number of this covering, and when $k$ large enough, any interval in $C(k)$ will be contained in $E_j$ for some $j$. Hence one can bound the $\sum_j diam(E_j)^{\alpha}$ from below by the ones of $C(k)$.

I got confused here: First why we can assume $E_j$ to be open?


2 Answers 2


I can elaborate abit why it is sufficient to show it for open ones. It might be that this is not exactly what the author was going after. If you wish, I may also continue this answer into a full proof which shows that $H^{\alpha}(C)\geq \frac{1}{2}$.

Choose for starters a $\delta$-cover $\{E_{j}\}_{j=1}^{\infty}$ of $C$ with $\sum_{j=1}^{\infty}\mathrm{diam}(E_{j})^{\alpha}\leq H^{\alpha}(C)+\delta$. Then for each $j$ we may choose a closed interval $I_{j}$ with $E_{j}\subset \mathrm{int}I_{j}$ and $\mathrm{diam}(I_{j})<(1+\delta)\mathrm{diam}(E_{j})$. Hence $\{\mathrm{int}I_{j}\}_{j=1}^{\infty}$ is an open cover of $C$, and in particular \begin{align*} H^{\alpha}(C)+\delta\geq \sum_{j=1}^{\infty}\mathrm{diam}(E_{j})^{\alpha}\geq (1+\delta)^{-\alpha}\sum_{j=1}^{\infty}\mathrm{diam}(I_{j})^{\alpha}. \end{align*} Thus to establish a lower bound for $H^{\alpha}(C)$ it suffices to establish a lower bound for $\sum_{j=1}^{\infty}\mathrm{diam}(I_{j})^{\alpha}$. (You may consider $\mathrm{int}(I_{j})$'s instead, which are open, as their diameter is the same as $I_{j}$'s.)

  • $\begingroup$ great, Thanks Thomas $\endgroup$
    – Sun
    Commented May 30, 2012 at 19:53

It is a fact in general (i.e. true in any metric space and Hausdorff measures of any dimension) that you can assume your covering sets to be open or closed. See Federer. For closed version, it is easier: diameter of $\bar{S}$ and diameter of $S$ are equal, so, if a collection of $S$'s cover your set and each has diam less than $\delta$, then you can instead consider the collection of $\bar{S}$'s, which again have diam bounded by $\delta$ ans still cover your set.

For open version, you need some sacrifice! At a cost of arbitrarily small $\sigma$, you can enlarge every set of diam less than $\delta$ to an open one with diam less than $(1+\sigma)\delta$. The latter can be used to estimate $\mathcal{H}^s_{(1+\sigma)\delta}$ within $(1+\sigma)^s$ accuracy of $\mathcal{H}^s_{\delta}$. Since for Hausdorff measure $\mathcal{H}^s$, you will send $\delta \to 0$, and $(1+\sigma)\delta$ will as well, your sacrifices will not affect the ultimate measure. (However, as expected, for one fixed $\delta$, $\mathcal{H}^s_\delta$ can be different if you only allow open coverings versus all coverings.)

  • 3
    $\begingroup$ The point is that you didn't "chase down all the $\epsilon$'s and $\delta$'s", thereby making a notational mess that no one can follow (at least, in the sense of seeing the "big picture"). Your second paragraph in particular could easily have been written in a way that would be difficult to grasp the underlying idea (by making it mostly symbolic with many unnecessary "math-grammar" details), but you did not do this. Instead, you concisely described the underlying idea. Just now I tried to find an overly-cluttered answer and didn't, but if I see one in the near future, I'll provide a link. $\endgroup$ Commented Feb 14, 2020 at 7:19
  • 1
    $\begingroup$ I see! I completely see what you mean (without you finding the example!) Writing down explicit variables and sums and inequalities will not necessarily help much with seeing why a certain propositions is true. It may in fact obscure the "reason" why it holds. I am slightly flattered by your observation as well :) $\endgroup$ Commented Feb 14, 2020 at 16:20
  • 1
    $\begingroup$ The fact that I didn't even notice the original "typos" in your explanation (which you fixed 2 hours before I saw your latest comment, just now) is also an example of the what Tao meant by: "The distinction between the three types of errors can lead to the phenomenon (which can often be quite puzzling to readers at earlier stages of mathematical development) of a mathematical argument by a post-rigorous mathematician which locally contains a number of typos and other formal errors, but is globally quite sound $\ldots$" (continued) $\endgroup$ Commented Feb 14, 2020 at 18:33
  • 1
    $\begingroup$ There is an amusing anecdote about the algebraic topologist Solomon Lefschetz that is relevant to what Tao is talking about: "... a joke which circulated about Lefschetz, namely that he never wrote a correct proof or stated an incorrect theorem ...". $\endgroup$ Commented Feb 14, 2020 at 19:07
  • 1
    $\begingroup$ Just now I tried to find an overly-cluttered answer and didn't, but if I see one in the near future, I'll provide a link. --- Here's one that I dealt with as an answer. I happened to recently cite it in this answer, so it was on my mind when I saw and commented on this answer. $\endgroup$ Commented Nov 8, 2022 at 20:09

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .