I'm facing a problem in measure theory and I need to prove the following conjecture to move on. Attention: I'm not sure the following statement is true.

Let $A \subset \mathbb{R}$ be a measurable set such that $m(A)>0$ and $H$ be a countable, dense subset of $\mathbb{R}$. If $A+H=\{a+h: a \in A, h \in H\}$, prove that $m((A+H)^c)=0$.

I'm totally stuck. It's easy to see that $A+H=\displaystyle{\bigcup_{h \in H} A+h}$, so it's definitely a measurable set, but that's the only progress I've been able to make. Any help would be greatly appreciated!

  • $\begingroup$ Regularity is your friend . . . $\endgroup$ – Noah Schweber Mar 25 '17 at 21:16
  • $\begingroup$ I've already tried using theorems that I know of that are relative to regularity. Could you please expand on this? $\endgroup$ – JustDroppedIn Mar 25 '17 at 21:18
  • 2
    $\begingroup$ First, can you show that the statement is true if $A$ is open? If so, think about how $A+H$ is a union of countably many sets, and - assuming $A+H$ had positive measure complement - try to approximate each of those sets by open sets in an appropriate way. $\endgroup$ – Noah Schweber Mar 25 '17 at 21:21
  • $\begingroup$ Thanks, I'll give it a shot and I'll let you know on my progress. Your comment is greatly appreciated $\endgroup$ – JustDroppedIn Mar 25 '17 at 21:32
  • $\begingroup$ Okay, i can even show that the statement is true if $A$ has an interior point. I can't see where this is going though and this makes me unable to see the appropriate approximation with open sets.. what is it that is going to lead to a contradiction? $\endgroup$ – JustDroppedIn Mar 25 '17 at 22:09

Use Lebesgue density theorem (LDT) which has an elementary proof.

Towards a contradiction, suppose $B = \mathbb{R} \setminus (A + H)$ has positive measure. Using LDT, choose open intervals $I, J$ of same length such that $B \cap I$ has $\geq 99$ percent measure of $I$ and $A \cap J$ has $\geq 99$ percent measure of $J$. Choose $h \in H$ (using the density of $H$) such that $J + h$ meets $I$ on a set of measure $\geq 99$ percent of $I$ (which has same length as $J$). Do you see a problem now?

  • $\begingroup$ Awesome! If one already knows the Lebesgue differentiation theorem, this is very nice! $\endgroup$ – PhoemueX Mar 27 '17 at 17:37
  • $\begingroup$ I must admit, this is a great answer. I'm not familiar with this theorem yet, but I read it and its proof and it is indeed elementary. Thank you very much. $\endgroup$ – JustDroppedIn Mar 27 '17 at 18:05

Below, I provide two proofs: A short and elegant proof using Fourier techniques and a more elementary but longer probabilistic proof.

Fourier proof

Since $A$ has positive measure, so has $A \cap [-n,n]$ for some $n$. Thus, we can assume without loss of generality that $A$ is bounded.

Now assume towards a contradiction that $M := A + H$ is not conull. Then there is an $L^1$ function $f$ with $f = 0$ on $M^c$ and such that $f$ does not vanish almost everywhere (take e.g. $f = 1_{M^c \cap [-k,k]}$ for a suitable $k$). Then we have for all $h \in H$ that $$ 0 = \int f(x) 1_{A+h}(x) dx = \int f(x) 1_{-A}(h-x) dx = (f \ast 1_{-A}) (h). $$ But since $1_{-A} \in L^\infty$ and since $f \in L^1$, the convolution from above is a continuous function. Since $H$ is dense, this means $f \ast 1_{-A} \equiv 0$.

Now, take the Fourier transform to get $0 \equiv \widehat{f} \cdot \widehat{1_{-A}}$. But since $1_{-A}$ has compact support, the Fourier transform $\widehat{1_{-A}}$ is a (nonvanishing) analytic function and thus has only isolated zeros. In particular, the set where $\widehat{1_{-A}}$ does not vanish is dense. Hence, we get $\widehat{f} \equiv 0$ by continuity and then $f = 0$ almost everywhere, a contradiction.

Probabilistic proof

Since $A$ has positive measure, so has $A\cap\left[-n,n\right]$ for a suitable $n\in\mathbb{N}$, so that we can assume $A\subset\left[-n,n\right]$. Let $\left(Y_{\ell}\right)_{\ell\in\mathbb{N}}$ be an iid sequence of random variables with $Y_{\ell}\sim U\left(\left[-2n,2n\right]\right)$ (the uniform distribution on $\left[-2n,2n\right]$). Fix $x\in\left[-n,n\right]$ and consider the event $$ E_{x}:=\bigcap_{\ell\in\mathbb{N}}\left(x\notin A+Y_{\ell}\right). $$ Because of $x-A\subset\left[-2n,2n\right]$ and by translation inavriance of the Lebesgue measure, the probability of this event is \begin{align*} \mathbb{P}\left(E_{x}\right) & =\prod_{\ell=1}^{\infty}\mathbb{P}\left(x\notin A+Y_{\ell}\right)=\prod_{\ell=1}^{\infty}\mathbb{P}\left(Y_{\ell}\notin x-A\right)\\ & =\prod_{\ell=1}^{\infty}\frac{1-\lambda\left(\left[-2n,2n\right]\cap\left(x-A\right)\right)}{4n}\\ & =\prod_{\ell=1}^{\infty}\frac{1-\lambda\left(A\right)}{4n}=0. \end{align*}

Up to now, we have shown for every $x\in\left[-n,n\right]$ that $\mathbb{P}\left(E_{x}\right)=0$. By Fubini's theorem, this shows that for almost every realization $y=\left(y_{\ell}\right)_{\ell\in\mathbb{N}}$ of $\left(Y_{\ell}\right)_{\ell\in\mathbb{N}}$, the set $$ N_{y}:=\left\{ x\in\left[-n,n\right]\,:\,\forall\ell\in\mathbb{N}:\, x\notin A+y_{\ell}\right\} $$ is a null-set. Fix one such realization $y$.

Since $A$ is bounded, the map $\mathbb{R}\to L^{1}\left(\mathbb{R}\right),t\mapsto1_{A+t}$ is continuous. By density of $H$, this allows us to choose for arbitrary $\ell,m\in\mathbb{N}$ some $h_{\ell,m}\in H$ with $\lambda\left(\left[A+y_{\ell}\right]\setminus\left[A+h_{\ell,m}\right]\right)\leq\frac{2^{-\ell}}{m}$. Thus, we get \begin{align*} \lambda\left(\left[-n,n\right]\setminus\bigcup_{h\in H}\left(A+h\right)\right) & \leq\lambda\left(\left[-n,n\right]\setminus\bigcup_{\ell,m}\left[A+h_{\ell,m}\right]\right)\\ & \leq\lambda\left(\left[-n,n\right]\setminus\bigcup_{\ell\in\mathbb{N}}\left(A+y_{\ell}\right)\right)+\lambda\left(\bigcup_{\ell\in\mathbb{N}}\left(A+y_{\ell}\right)\setminus\bigcup_{\ell,m}\left(A+h_{\ell,m}\right)\right)\\ \left(\text{with arbitrary }m\in\mathbb{N}\right) & \leq\lambda\left(\bigcup_{\ell\in\mathbb{N}}\left(A+y_{\ell}\right)\setminus\left(A+h_{\ell,m}\right)\right)\leq\sum_{\ell=1}^{\infty}\frac{2^{-\ell}}{m}=\frac{1}{m}. \end{align*} Since $m\in\mathbb{N}$ can be chosen arbitrarily, this implies $\lambda\left(\left[-n,n\right]\setminus\bigcup_{h\in H}\left(A+h\right)\right)=0$.

We are almost done: Above, $H\subset\mathbb{R}$ was an arbitrary dense subset. Now, if $N\in\mathbb{Z}$ is arbitrary, then $H+N\subset\mathbb{R}$ is also dense, so that we get $$ 0=\lambda\left(\left[-n,n\right]\setminus\bigcup_{h\in H+N}\left(A+h\right)\right)=\lambda\left(\left(\left[-n,n\right]-N\right)\setminus\bigcup_{h\in H}\left(A+h\right)\right)=\lambda\left(\left[-n-N,n-N\right]\setminus\bigcup_{h\in H}\left(A+h\right)\right). $$ Since $\bigcup_{N\in\mathbb{Z}}\left[-n-N,n-N\right]=\mathbb{R}$, this easily implies $\lambda\left(\mathbb{R}\setminus\bigcup_{h\in H}\left(A+h\right)\right)=0$, as desired.

  • $\begingroup$ I'm afraid I have no knowledge of Fourier analysis, so I can't judge your answer. Even so, I'm grateful for your answer. Thing is, the reason I'm asking this question is because I need it in order to solve an exercise in measure theory. So if you have a different proof in mind, please let me know. $\endgroup$ – JustDroppedIn Mar 26 '17 at 15:57
  • $\begingroup$ @authunchained: I just added another proof :) $\endgroup$ – PhoemueX Mar 26 '17 at 16:53
  • $\begingroup$ Do you think that the Fourier proof work in any dimension? It should, only that I do not know how to make up for the fact that real analytic functions might have nonisolated zeroes (e.g., $\widehat {1_{-A}}(\xi_1, \xi_2)$ might be equal to $\xi_1^2+\xi_2^2-1$). $\endgroup$ – Giuseppe Negro Mar 27 '17 at 16:55
  • $\begingroup$ @GiuseppeNegro: Yes, it will work in every dimension. It is still true that the zero-set of an analytic function is a null-set. This can be seen as follows: Fix $x_1,\dots ,x_{n-1}$ and consider $f(x_1, \dots, x_{n-1}, \cdot)$. This is either identically zero or only has isolated zeroes. By Fubini (to calculate the measure of the zero-set), it suffices to show that the "identically zero" case (say, on the set $E \subset \Bbb{R}^{n-1}$) only happens on a set of measure zero. In dimension $2$, this is easy: If not, then the set of $x_1$ with $f(x_1,\cdot) \equiv 0$, has a cluster point $\endgroup$ – PhoemueX Mar 27 '17 at 17:31
  • $\begingroup$ so that $f(\cdot, x_2) \equiv 0$ for every $x_2$. In general, we proceed by induction: For arbitrary $x_n$ and $g(x_1,\dots,x_{n-1}) := f(x_1,\dots ,x_n)$, we have $g = 0$ on the set $E$ which has positive measure by assumption. Hence, $g \equiv 0$. Since this holds for arbitrary $x_n$, we get $f \equiv 0$.... I think my writeup is a bit confusing, but hopefully one can understand the idea :) $\endgroup$ – PhoemueX Mar 27 '17 at 17:34

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.