13
$\begingroup$

This was previously posted to Math StackExchange. I was originally unsure whether it is suitable for posting here, but I've yet to get an answer there, so I'm just trying to see if people here can help.

As in the question title, let $A, B$ be a partition of the unit circle $S^1$, equipped with the Haar measure. Here, we do not require $A, B$ to be measurable. Also, assume neither $A$ nor $B$ is of measure zero, so they are either both non-measurable or both of positive measure. (Equivalently, they both have positive outer measure.) Is it possible, then, for $R_\theta(A) \cap B$ to be measurable and have Haar measure zero for all $\theta$, where $R_\theta$ is the rotation by degree $\theta$?

$\endgroup$
2

3 Answers 3

14
$\begingroup$

Assuming the Continuum Hypothesis, the answer is yes. On CH, the $\mathbf Q$-dimension of $\mathbf R$ has the cardinality of the first uncountable ordinal $\omega_1$, so we can find a $\mathbf Q$-linear basis $e_\alpha, \alpha \in \omega_1$ of ${\bf R}$, and we can normalize $e_0=1$. Thus every irrational real number $x$ has a unique representation of the form $$ x = \sum_{\alpha \leq \beta} q_\alpha e_\alpha\tag{1}\label{1}$$ where $0 < \beta < \omega_1$ and $q_\alpha, \alpha \leq \beta$ are rational numbers with all but finitely many $q_\alpha$ non-zero, and $q_\beta$ also non-zero. In particular, $q_\beta$ is the final non-zero coefficient of $x$. We then assign $e^{2\pi i x}$ to lie in $A$ if $x$ is rational, or $x$ is irrational with representation \eqref{1} with $\nu_2(q_\beta)$ even, where $\nu_2(q_\beta) \in \mathbf Z$ is the number of times two divides the final non-zero coefficient $q_\beta$, and $e^{2\pi i x}$ to lie in $B$ if $x$ is irrational with representation \eqref{1} with $\nu_2(q_\beta)$ odd. Since the final non-zero coefficient $q_\beta$ is unaffected by shifts by integers (or even rationals) on the irrationals, this is a well-defined partition of $S^1$. By construction, if $z$ is not a root of unity, then $z$ lies in $A$ if and only if $z^2$ lies in $B$. Because of this, neither $A$ nor $B$ can have measure zero (otherwise $S^1 = A \cup B$ would also have measure zero).

On the other hand, if $\theta = e^{2\pi i s}$ is an angle, then $s$ is a finite $\mathbf Q$-linear combination of the $e_\alpha$. Such rotations do not affect the final non-zero coefficient $q_\beta$ in \eqref{1} unless $\beta$ is less than or equal to one of these $\alpha$ (or $x$ is rational), which only occurs countably often. Thus $R_\theta(A) \cap B$ is countable (hence null) for all $\theta$.

$\endgroup$
1
  • 5
    $\begingroup$ The same argument seems to work under Martin's Axiom, as it implies that every set of reals of cardinality less than continuum is null. $\endgroup$
    – Yair Hayut
    Commented May 31, 2023 at 16:04
9
$\begingroup$

This question is explored in great generality by Laczkovich in

Laczkovich, Miklós, "Two constructions of Sierpiński and some cardinal invariants of ideals", Real Anal. Exch. 24(1998-99), No. 2, 663-676 (1999), MR1704742, Zbl 0964.03056 (the paper is Open access at ProjectEuclid).

Sierpiński proved such a partition exists under CH, by an argument which immediately extends to any model where every set of reals of cardinality less than continuum is null (as Terry and Yair observed in the other answer). Laczkovich shows that there is no such partition in random real models (e.g., $L$ adjoined with $\omega_2$ random reals), so this problem is independent from ZFC.

$\endgroup$
2
$\begingroup$

$\newcommand\th\theta$This is a partial answer, showing that, if $R_\th(A)\cap B$ is measurable and has Haar measure $|\cdot|$ zero for all $\th$, then $A$ and $B$ cannot be measurable with $|A|>0$ and $|B|>0$.

Indeed, suppose the contrary (we do not need to assume here that $B$ is the complement of $A$). By the regularity of the Haar measure, there exist (say open) arcs $I$ and $J$ of the circle such that $|A\cap I|>|I|/2$ and $|B\cap J|>|J|/2$. Since $|A\cap I|,|I|,|B\cap J|,|J|$ depend continuously on the endpoints of the arcs $I$ and $J$, without loss of generality (wlog) $|I|$ and $|J|$ are rational numbers, say $p/q$ and $r/s$, respectively, for some natural $p,q,r,s$. Partitioning each of the arcs $I$ and $J$ into arcs of the same length $1/(qs)$, we see that wlog $|I|=|J|=1/(qs)$.

Renormalizing the Haar measure, assume wlog that $|I|=|J|=1$, so that $|A\cap I|>1/2$ and $|B\cap J|>1/2$. Also, the condition $|I|=|J|$ implies that $I=R_{-\th}(J)$ for some real $\th$, so that $|R_\th(A)\cap J|=|A\cap R_{-\th}(J)|=|A\cap I|$. So, $$ \begin{split} |R_\th(A)\cap B|&\ge|(R_\th(A)\cap J)\cap(B\cap J)| \\ &=|R_\th(A)\cap J|+|B\cap J|-|(R_\th(A)\cap J)\cup(B\cap J)| \\ &=|A\cap I|+|B\cap J|-|(R_\th(A)\cap J)\cup(B\cap J)| \\ &\ge|A\cap I|+|B\cap J|-|J| \\ &>1/2+1/2-1=0. \end{split} $$ So, $|R_\th(A)\cap B|>0$, a contradiction. $\quad\Box$

$\endgroup$
17
  • $\begingroup$ I know that a similar fact is true about positive-measure subsets of the real line. It's no surprise that a fact about the real line would have an analogue on the circle; but do you know whether there is some general analogue of this result, say, for general compact groups? \\ By the way, your answer reminds me of techniques that seemed relevant for A different way to try to define a measure on the unit-circumference circle. I gave a partial answer; maybe you can knock out the whole thing? $\endgroup$
    – LSpice
    Commented May 31, 2023 at 14:21
  • $\begingroup$ @LSpice : Thank you for your comment. I will have it in mind. $\endgroup$ Commented May 31, 2023 at 15:05
  • $\begingroup$ You are welcome, @IosifPinelis $\endgroup$ Commented May 31, 2023 at 17:01
  • 1
    $\begingroup$ Also, from the Fubini-Tonelli theorem one has $\int |R_\theta(A) \cap B|\ d\theta = |A| |B|$ for measurable $A,B$ (if we normalize the unit circle to have total mass one), which is another way to handle this case. (The counterexample in my answer is inspired in part by one of the standard counterexamples to the Fubini-Tonelli theorem based on the first uncountable ordinal.) $\endgroup$
    – Terry Tao
    Commented May 31, 2023 at 18:25
  • 1
    $\begingroup$ @LSpice That seems right. Though we probably still need to alter the definitions of $A_{S^1}$ and $B_{S^1}$ a bit, as we could have $A_{S^1} = A \cap [0, 1)$ or $B_{S^1} = B \cap [0, 1)$ is null. Since it cannot be the case that the intersection of $A$ with any interval is null as $A$ is not null and similarly for $B$, we could replace $[0, 1)$ by a large enough interval in the definition to fix the issue. $\endgroup$
    – David Gao
    Commented May 31, 2023 at 22:26

You must log in to answer this question.

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