8
$\begingroup$

Let $C$ be the Cantor middle-thirds set. Let $\mu$ be a finitely-additive isometrically-invariant measure on all subsets of $\mathbb R$. Then $\mu(3C)=2\mu(C)$, where $aB = \{ ax : x \in B \}$. Thus if $a$ is a power of $3$, $\mu(aC) = a^{\log_3 2} \mu(C)$.

Question 1: Is it the case for all $a\in (0,\infty)$ that $\mu(aC)=a^{\log_3 2} \mu(C)$, if $\mu$ is a finitely-additive isometrically-invariant measure on $\mathcal P\mathbb R$?

The remaining questions are predicated on an affirmative answer, though currently I'm suspecting a negative answer to 1.

Question 2: Is this still true if isometric-invariance is replaced by translation-invariance?

Question 3: Has there been any work on using the the relationship between $\mu(aC)$ and $\mu(C)$ for invariant measures $\mu$ on $\mathbb R^n$ (perhaps not defined on all of the powerset, but just on the Borel sets) to define a dimension and comparing it to Hausdorff and box dimensions?

$\endgroup$
  • $\begingroup$ Interesting question. For Question 3 a problem is that many sets are given zero or non-$\sigma$-finite mass for any translation-invariant Borel measure on $\mathbb{R}$ (for example, the set of Liouville numbers, see bolyai.cs.elte.hu/analysis/keleti/papers/liouville.pdf). $\endgroup$ – Pablo Shmerkin Sep 27 '13 at 14:11
  • $\begingroup$ Thanks for the reference. It's good that I specified the measures to be finitely additive, since then by supramenability of the isometries of $\mathbb R$ every subset of $\mathbb R$ has an invariant measure that normalizes it. $\endgroup$ – Alexander Pruss Sep 27 '13 at 16:07
2
$\begingroup$

Negative to the first question, so the others are moot.

Let $G_1$ be the isometries of $\mathbb R$.

First note that no finite number of translates (or reflections, but that doesn't add anything) of $(1/2)C$ covers $C$. This can be seen by playing around with base three expansions.

Let $I$ be the ideal in $\mathcal P(\mathbb R)$ generated by $(1/2)C$ and its translates. Let $\mathcal B$ be the quotient boolean algebra $\mathcal P(\mathbb R)/I$. Because $C$ isn't covered by translates of $(1/2)C$, $[C]\ne 0$. The action of $G_1$ on $\mathbb R$ induces an action of $G_1$ on $\mathcal B$. Since $G_1$ is supramenable (i.e., doesn't have nonempty paradoxical subsets), it follows from a theorem of Mycielski ("Finitely additive invariant measures. I", Colloq. Math. 42 (1979), 309–318) that there is a finitely additive measure $\mu_1$ on $\mathcal B$ invariant under $G_1$ with $\mu_1([C])=1$. Let $\mu(A)=\mu_1([A])$ for $A\subseteq\mathbb R$. Then $\mu(C)=1$ and $\mu((1/2)C)=0$ and the answer to Question 1 is negative for $a=1/2$.

This uses the Axiom of Choice.

| cite | improve this answer | |
$\endgroup$
1
$\begingroup$

This is too long for a comment. The following measure (defined on Borel sets) might be a counterexample to Question 1: let $\mathcal{I}_N$ be the collection of all left-closed, right-open intervals of length $3^{-k}$ for some $k\in\mathbb{N}, k\ge N$, and define $$ \mu(A) = \lim_{N\to\infty} \inf\left\{ \sum_i |I_i|^{\log_3 2}: A\subset \bigcup_i I_i, I_i\in\mathcal{I}_N \right\}. $$

This is well-known to be a Borel-regular measure on the Borel $\sigma$-algebra (is constructed using the "Type 2" method from Rogers' classical book "Hausdorff measures"), and it's obviously isometry-invariant. It is clear that $\mu(C)=1$. Since we are covering only with intervals of length $3^{-k}$, it seems to me the scaling law should break down ($\mu(aC)$ should be strictly larger than $a^{\log_3 2}$ for $1<a<3$). I can't prove this at the moment though.

Regarding Question 3, in addition to my comment another problem is that for some sets $E$ there exist two (again, Borel) isometry-invariant measures $\mu,\nu$ which give $E$ positive and finite mass, and distinct numbers $s,t>0$ such that $\mu(a E)=a^s \mu(E)$ and $\nu(a E)=a^t(E)$ for all $a>0$. For example, we may choose a set $E$ of different Hausdorff and packing dimensions, and which has positive and finite Hausdorff/packing measure in their respective dimensions. It is not immediately clear to me if it is possible to get 3 or more different scaling exponents.

Essentially the approach in Question 3 is implicitly used in the definition of Hausdorff and packing dimensions, but it only works if one restricts the measures to the family Hausdorff and packing measures, respectively.

I know very little about finitely additive measures on $\mathcal{P}(\mathbb{R})$ so I don't have much to say on the actual questions.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ An invariant f.a. measure on the Borel sets (or on any other algebra of sets) will extend to an invariant f.a. measure on all subsets of $\mathbb R$ (it's somewhere in Wagon's BT Paradox book), so if this works for the Borel sets, it answers the question in the negative. $\endgroup$ – Alexander Pruss Sep 28 '13 at 1:56
  • $\begingroup$ You can probably get an arbitrarily large finite number of different measure-based "dimensions" by using the trick in my answer. Let $C_1,C_2,...,C_n$ be self-similar Cantor sets of different dimensions, rescaled in such a way that no one of the sets can be covered by finitely many translates of the others. I don't have a proof that this can be done, but it wouldn't surprise me. I hereby conjecture it. Then for each $i$ there will be a f.a. invariant measure on $\mathcal P\mathbb R$ that's nonzero on $C_i$ but zero on each $C_j$, $j\ne i$. $\endgroup$ – Alexander Pruss Sep 28 '13 at 13:02

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.