We know that if $f\in L^1(\mathbb{R})$, then $\|f(\cdot+1/n)-f(\cdot)\|_{L^1}\to 0$ as $n\to \infty$, which implies that there exists a subsequence $f_{n_k}=f(x+\frac{1}{n_k})$ such that $f_{n_k}\to f$ a.e.

My problem is that Is there any $f\in L^1(\mathbb{R})$ function such that the almost convergence result $$ f(x)=\lim_{n\to\infty}f(x+1/n)\quad \text{a.e.} $$ is NOT valid? I intend to believe that there exits such function. Otherwise, there would be a beautiful statement such that $f(x)=\lim_{n\to\infty}f(x+1/n)$ a.e. for any $f\in L^1(\mathbb{R})$.

Thank you for the long commenting below @David C. Ullrich. This question has been reduce to find a compact set $K$ such that for any $x\in K$, there are infinite $n$ such that $x+\frac{1}{n}\not\in K$.

  • 1,525
  • 4
  • 21
  • There must be - seems the example might be complicated... – David C. Ullrich Jan 19 '18 at 16:03
  • @DavidC.Ullrich I agree. Otherwise, the textbook would introduce such beautiful result. – Yuhang Jan 19 '18 at 16:04
  • I believe there is likely none – ibnAbu Jan 19 '18 at 16:41
  • @Abdullahi_A_Ibrahim Then, How to prove it? – Yuhang Jan 19 '18 at 16:42
  • such construction will require constructing an uncountable set $A$ such that for every $n$ and $x \in A$ , $x+1/n$ is not in $A$, – ibnAbu Jan 19 '18 at 16:46
  • I guess some kind of fat cantor set would be useful. I didn't think long enough tho – Shashi Jan 19 '18 at 16:47
  • 1
    @Abdullahi_A_Ibrahim We don't need that for every $n$, we only need that for every $x\in A$ there are infinitely many $n$ such that $x+1/n$ is not in $A$. Getting such an uncountable set is not too difficult (I believe the usual Cantor set works). To have such a set with positive measure is going to be trickier - I also think some fat Cantor set might work, but have no proof for that. – Wojowu Jan 19 '18 at 16:50
  • @Abdullahi_A_Ibrahim Not necessarily, you only need for infinite $n$ and $x\in A$, $x+1/n$ is not in $A$ – Yuhang Jan 19 '18 at 16:51
  • @Wojowu I also guess some fat cantor set may work. But I fail to prove it. – Yuhang Jan 19 '18 at 16:51
  • @Wojowu Yes, the usual Cantor set $C$ would work, except that $m(C)=0$: For almost every $x\in C$ we have $x+1/3^n\notin C$. I was just thinking some sort of fat Cantor set might do it... – David C. Ullrich Jan 19 '18 at 16:52
  • @Yuhang you are right – ibnAbu Jan 19 '18 at 17:15
  • But $L^1$ function is a limit of simple functions – ibnAbu Jan 19 '18 at 17:16
  • @Abdullahi_A_Ibrahim Wait a minute. Even by Lusin's theorem, how could you make sure that $f(x+1/n)=g(x+1/n)$ ? – Yuhang Jan 19 '18 at 18:02
  • @Yuhang $g(x)$ ? – ibnAbu Jan 19 '18 at 18:16
  • I mean that g(x)=1_A – Yuhang Jan 19 '18 at 18:19
  • @Abdullahi_A_Ibrahim Although $\mu(\{x;f(x)\neq 1_A\})<\epsilon$, it does mean that for any $x\in B:=\{x\in \{x;f(x)=1_A\}$, $x+1/n\in B$ – Yuhang Jan 19 '18 at 18:22
  • @Abdullahi_A_Ibrahim Thx anyway, I need to have my dinner. See you – Yuhang Jan 19 '18 at 18:26
  • Let us [continue this discussion in chat](http://chat.stackexchange.com/rooms/71940/discussion-between-abdullahi-a-ibrahim-and-yuhang). – ibnAbu Jan 19 '18 at 21:14
  • @Wojowu how does a nowhere dense set helps in constructing a function discontuous relative to every measurable set ? – ibnAbu Jan 19 '18 at 21:19
  • @Shashi how does a nowhere dense set helps in constructing a function discontinuous relative to every measurable set ? – ibnAbu Jan 19 '18 at 21:21
  • @Abdullahi_A_Ibrahim It was just something I thought the ones working on this could help. I did not try it myself. – Shashi Jan 19 '18 at 21:25
  • @Abdullahi_A_Ibrahim No, i don't think so. I don't think that you can't get such strong result unless $B$ is an open set. – Yuhang Jan 20 '18 at 08:49
  • @Abdullahi_A_Ibrahim Then, there may be another way to construct it. If this result is true, then it should be mentioned in some textbooks. It is a natural question to ask. Isn't it ? – Yuhang Jan 20 '18 at 10:14
  • @Yuhang I agreed it is a good question – ibnAbu Jan 20 '18 at 10:18
  • 1
    @Abdullahi_A_Ibrahim It is quite wired that nobody write anything about this question.Maybe, there is some paper concerning it. I just don't know where to find. – Yuhang Jan 20 '18 at 10:26
  • @Yuhang i reformulated your problem in https://math.stackexchange.com/questions/2613251/constructing-a-fat-cantor-set-with-certain-property – ibnAbu Jan 20 '18 at 10:56
  • @Yuhang you didnt indicate in your question $n \in N$ where $N$ is set of natural numbers – ibnAbu Jan 21 '18 at 23:18
  • 1
    I asked my master Taras Banakh, who recently wrote a paper “[Null-finite sets in metric groups and their applications]( https://arxiv.org/pdf/1706.08155)” a question on a compact set $K$ of positive measure such that for any $x\in K$, there are infinite $n$ such that $x+\frac{1}{n}\not\in K$. He guessed that it should have a negative answer, which he will try to write and to send today. – Alex Ravsky Jan 22 '18 at 11:05

7 Answers7


Not an answer - two long comments for the benefit of anyone who's looking for that counterexample that "must" exist:

First, I realized the other day that the sort of construction I'd been trying cannot possibly work. To save others from attempting the impossible:

Say $f$ is a counterexample. Say $(n_j)$ is a bad sequence for $x$ if $$|f(x)-f(x+1/n_j)|>\epsilon$$for all $j$. The sorts of constructions I'd been thinking about, fat Cantor sets, etc., would have had this property if successful:

There exists a single sequence $S=(n_j)$ such that $S$ is a bad sequence for $x$, for all $x$ in some set of positive measure.

That's impossible. Given $S$, the set of $x$ such that $S$ is bad for $x$ must have measure zero. Because there does exist a subsequence of $S$ giving convergence almost everywhere.

(To be specific: I noticed that if $C$ is the standard Cantor set then for all but countably many $x\in C$ we have $x+1/3^n\notin C$. So $\chi_C$ would be a counterexample, except $m(C)=0$. This got me thinking about fat Cantor sets. The above shows that if a fact Cantor set does give a counterexample the explanation cannot be that simple; it cannot be that "uniform".)

In the other direction:

If there is a counterexample then there is a counterexample of the form $\chi_K$ for some compact $K$.

This follows from Lusin's theorem. Say $f$ is a counterexample. There exists $[a,b]$ such that if $E$ is the set of bad points in $[a,b]$ then $m(E)>0$. Suppose that $K\subset[a,b]$ is compact, $m(K)+m(E)>b-a$, $g$ is continuous, and $g=f$ on $K$. If $x\in K\cap E$ then we must have $x+1/n\notin K$ for infinitely many $n$. Since $m(K\cap E)>0$ this shows that $\chi_K$ is a counterexample.

David C. Ullrich
  • 84,497
  • 5
  • 61
  • 132

This question has been reduce to find a compact set $K$ such that for any $x\in K$, there are infinite $n$ such that $x+\frac{1}{n}\not\in K$".

As I have understood David C. Ullrich's answer, it is also required that $m(K)>0$.

I just received a letter from Taras Banakh with the following

Corollary. For any decreasing sequence $(a_n)_{n=1}^\infty$ of positive real numbers with $\lim_{n\to\infty}a_n=0$ and $\lim_{n\to\infty}\frac{a_{n+1}}{a_n}=1$, there exists a compact subset $K\subset\mathbb R$ of positive Lebesgue measure such that $\bigcap_{n=m}^\infty(K-a_n)=\emptyset$.

Corollary follows from Theorem, see below.

Let $\mathcal K(\mathbb R)$ be the space of non-empty compact subsets of the real line, endowed with the Vietoris topology (which is generated by Hausdorff metric). It is well-known that the space $\mathcal K(\mathbb R)$ is locally compact (more precisely, each closed bounded subset of $\mathcal K(\mathbb R)$ is compact.

For every $c>0$ let $\mathcal K_c(\mathbb R)=\{K\in\mathcal K(\mathbb R):\lambda(K)\ge c\}$ be the subspace of consisting of compact sets $K$ of Lebesgue measure $\lambda(K)\ge c$. The regularity (or countable additivity) of the Lebesgue measure $\lambda$ implies that $\mathcal K_c(\mathbb R)$ is a closed subspace in $\mathcal K(\mathbb R)$. Consequently the space $\mathcal K_c(X)$ is locally compact and Polish.

Theorem. Let $(a_n)_{n=1}^\infty$ be a decreasing sequence of positive real numbers such that $\lim_{n\to\infty}a_n=0$ and $\lim_{n\to\infty}\frac{a_{n+1}}{a_n}=1$. Then for every $c>0$ the set $$\{K\in\mathcal K_c(X):\forall m\in\mathbb N\;\bigcap_{n=m}^\infty(K-a_n)=\emptyset\}$$ is dense $G_\delta$ in $\mathcal K_c(X)$.

Proof. It suffices to prove that for every $m\in\mathbb N$ the set $$\mathcal K_m=\{K\in\mathcal K_c(X):\bigcap_{n=m}^\infty(K-a_n)=\emptyset\}$$ is open and dense in $\mathcal K_c(X)$.

To see that $\mathcal K_n$ is open, take any compact set $K\in\mathcal K_m$. By the compactness of $K$, there exists $l>m$ such that $\bigcap_{n=m}^l(K-a_n)=\emptyset$. Then for every $x\in \mathbb R$ there exists $n_x\in[m,l]$ such that $x\notin K-a_{n_x}$. So, we can find a symmetric neighborhood $U_x\subset[-1,1]$ of zero such that $x+U_x$ is disjoint with $U_x+K-a_{n_x}$. By the compactness of the set $L=[-1,1]+(K-a_m)$, there exists a finite subset $F\subset L$ such that $L\subset\bigcup_{x\in F}(x+U_x)$. We claim that the open neighborhood $\mathcal U=\{K'\in\mathcal K_c(X):K'\subset\bigcap_{x\in F}(K+U_x)\}$ of $K$ in $\mathcal K_c(X)$ is contained in the set $\mathcal K_m$. Assuming that $\mathcal U\not\subset\mathcal K_m$, we can find a compact set $K'\in\mathcal U\setminus \mathcal K_m$. Since $K'\notin\mathcal K_m$, there exists a point $z\in \bigcap_{n=m}^\infty (K'-a_n)$. It follows that $z\in K'-a_m\subset [-1,1]+K-a_m=L$ and hence $z\in x+U_x$ for some $x\in F$. Then $$z\in (x+U_x)\cap\bigcap_{n=m}^\infty K'-a_n\subset (x+U_x)\cap (K'-a_{n_x})\subset (x+U_x)\cap (U_x+K-a_{n_x})=\emptyset,$$ which is a desired contradiction, showing that $\mathcal U\subset\mathcal K_m$ and the set $\mathcal K_m$ is open in $\mathcal K_m$.

Next, we show that $\mathcal K_m$ is dense in $\mathcal K_c(\mathbb R)$. Given any $K\in\mathcal K_c(\mathbb R)$ and $\varepsilon>0$ we need to find a set $K'\in\mathcal K_m$ on the Hausdorff distance $d_H(K',K)<\varepsilon$ for $K$.

Choose $k\ge m$ so large that $\frac1k<\frac12\varepsilon$. Consider the cover $\mathcal I_k=\big\{\big[\frac{n}k,\frac{n+1}k\big]:n\in\mathbb N\big\}$ of $\mathbb R$ by closed intervals of length $\frac1n$. The choice of $k$ ensures that the compact set $\tilde K=\bigcup\{I\in\mathcal I_n:I\cap K\ne\emptyset\}$ has $d_H(\tilde K,K)\le \frac1k<\frac12\varepsilon$. Also it is clear that $\tilde K$ has Lebesgue measure $\lambda(\tilde K)>\lambda(K)\ge c$. Choose $p\ge k$ so large that $(1-\frac1p)\cdot\lambda(\tilde K)\ge c$.

Since $\lim_{n\to\infty}\frac{a_{n+1}}{a_n}=1$, there exists $q>p$ such that $\frac{a_{n+1}}{a_{n}}>1-\frac1{4p}$ for all $n\ge q$. Finally, take $l>q$ such that $\frac2{lk}<a_q$. Consider the open set $$Z:=\bigcup_{z\in\mathbb Z}{\big]}\tfrac{z}{kl},\tfrac{z}{kl}+\tfrac1{pkl}{\big[}$$ and observe that for every interval $I\in\mathcal I_k$ the set $I\setminus Z$ has Lebesgue measure $\lambda(I\setminus Z)=(1-\frac1p)\lambda(I)$ and Hausdorff distance $d_H(I,I\setminus Z)<\frac12\varepsilon$. Consequently, the compact set $K'=\tilde K\setminus Z$ has Lebesgue measure $\lambda(K')=(1-\frac1p)\cdot\lambda(\tilde K)\ge c$ and $d_H(K',\tilde K)=\frac1{2pkl}<\frac12\varepsilon$. Then $d_H(K',K)\le d_H(K,\tilde K)+d_H(\tilde K,K)<\varepsilon$. It remains to prove that $K'\in\mathcal K_m$. Assuming that $K'\notin\mathcal K_m$, we could find a point $x\in\bigcap_{n=m}^\infty (K'-a_n)$. Then $x+a_n\in K'\subset\mathbb R\setminus Z$ for all $n\ge m$. Let $z\in Z$ be the unique integer number such that $\frac{z-1}{lk}<x\le \frac{z}{lk}$. The choice of $l$ guarantees that $\frac2{lk}<a_q$. Then $\frac{z+1}{lk}=\frac{z-1}{lk}+\frac2{lk}<x+a_q$. Let $i\ge q$ be the smallest number such that $x+a_i\ge \frac{z}{lk}+\frac1{plk}$. Then $x+a_{i+1}<\frac{z}{lk}+\frac1{plk}$ and hence $a_{i+1}<\frac{z}{lk}+\frac1{plk}-x<\frac{z}{lk}+\frac1{plk}-\frac{z-1}{lk}= \frac1{lk}+\frac1{plk}<\frac2{lk}$.

On the other hand, $$a_i-a_{i+1}=a_{i+1}\tfrac{a_i}{a_{i+1}} (1-\tfrac{a_{i+1}}{a_i})<\tfrac2{lk}(1-\tfrac1{4p})^{-1}\tfrac1{4p}<\tfrac1{lkp}$$ and hence $x+a_{i+1}>x+a_i-\frac1{lkp}\ge\frac{z}{lk}+\frac1{plk}-\frac1{plk}=\frac{z}{kl}$. Then $\frac{z}{kl}<x+a_{i+1}<\frac{z}{lk}+\frac1{lkp}$ implies that $x+a_{i+1}\in Z$, which is a desired contradiction, completing the proof of $K'\in\mathcal K_m$.$\square$

PS. Also Taras Banakh sent to me the reference to the talk “Theorems of H. Steinhaus, S. Picard and J. Smital” by W. Wilczyński from Ger-Kominek Workshop in Mathematical Analysis and Real Functions 20-21.11.2015.

At the last page are stated the following results.

Theorem. If $A\subset\Bbb R$ is a measurable set, then for each sequence $\{x_n\}_{n\in\Bbb N}$ convergent to $0$ the sequence of characteristic functions $\{\chi_{A+x_n}\}_{n\in\Bbb N}$ converges in measure to $\chi_A$.

Remark. There exists a measurable set $A\subset\Bbb R$ and a sequence $\{x_n\}_{n\in\Bbb n}$ convergent to $0$ such that $\{\chi_{A+x_n}\}_{n\in\Bbb N}$ does nor converge almost everywhere to $\chi_A$.

Theorem. If $A$ has the Baire property, then $\{\chi_{A+x_n}\}_{n\in\Bbb N}$ converges to $\chi_A$ except on a set of the first category.

Alex Ravsky
  • 78,203
  • 4
  • 45
  • 145
  • Three years later but it does not matter. There is a recent [posting](https://math.stackexchange.com/questions/4106171/finding-f-s-t-the-sequence-of-functions-f-nx-f-x-%E2%88%92-a-n-is-not-a-e-con/4112809#4112809) to which there result you presented seems to fit right in. What is puzzling, is that the OP on the link I mentioned, is from a textbook and gives a seemingly simple hint of a fat Cantor set. Looking at this now, I believe the problem is much harder than the simple hint seems to indicate. – Oliver Díaz May 01 '21 at 23:58

I submit that we can show that the a.e. convergence property $$\tag{FALSE} f(x+\tfrac1n)\to f(x)\qquad x\text{-a.e.}\quad \forall\, f\in L^1(\mathbb T). $$ does NOT hold by using Stein's maximal principle (thank you David C.Ullrich for pointing this out). This has the advantage of a simpler proof at the expense that we won't produce an explicit counterexample. (EDIT. See David C.Ullrich's answer for such counterexample).

Definitions. For $f\in L^1(\mathbb T)$ we define the maximal function $$ f^\star(x)=\sup_{n\in \mathbb N} |f(x+\tfrac1n)|.$$ To show that Stein's maximal principle does not hold we need to construct a sequence $(f_k)_{k\in \mathbb N}$ in $L^1(\mathbb T)$ such that $$ \frac{\|f_k^\star\|_{1,\infty}}{\|f_k\|_1}\to +\infty, $$ where the weak $L^1$ norm is defined as $$ \|g\|_{1,\infty} =\sup_{t>0} t|\{x\in\mathbb T\ :\ |g(x)|>t\}|,$$ and $|A|$ for a set $A\subset \mathbb T$ denotes Lebesgue measure.

Proof. The sequence $f_k$ is defined as $$ f_k(x)=k^2\mathbf 1_{\{|x|\le \frac1{k^2}\}}.$$ We have $$f_k^\star(x)\ge \sup_{1\le n \le k} |f_k(x+\tfrac1n)|=\sum_{n=1}^k k^2\mathbf1_{\{|x+\tfrac1n|\le\frac1{k^2}\}}=:k^2\mathbf1_{A_k}.$$ Notice that $A_k$ is the union of $k$ disjoint intervals of length $1/k^2$. Using this observation we can bound from below the weak $L^1$ norm as $$ \|f_k^\star\|_{1,\infty}\ge \|k^2\mathbf1_{A_k}\|_{1,\infty}=k. $$ Since $\|f_k\|_1=1$ for all $k$, the proof is complete. $\square$

Remark. Here we considered functions on the torus $\mathbb T=\mathbb R/\mathbb Z$, because the Stein maximal principle needs a compact group structure. This is not a real limitation, because the statement (FALSE) is local in nature. If such a statement were true for $f\in L^1(\mathbb R)$ it would also be true for functions supported on bounded intervals, and this we disproved.

Giuseppe Negro
  • 28,987
  • 6
  • 59
  • 199
  • 1
    Fabulous - I can't believe I missed this. Btw getting from here to an explicit counterexample is easier than the proof of Stein's theorem, because we can arrange things so that $\sum m(supp(f_k))<1$. See newest answer... – David C. Ullrich Jan 23 '18 at 15:36
  • 2
    For anyone interested, this argument extends easily to arbitrary sequences of (distinct) offsets $a_n$ instead of $1/n.$ Just replace the $1/k^2$ by $\min_{1\leq n,m\leq k} |a_n-a_m|.$ – Dap Jan 23 '18 at 17:27
  • Although it does not construct any counterexample, I have to say this is the most beautiful answer. That is why I have to give you bounty. Thx. – Yuhang Jan 24 '18 at 01:30
  • @Yuhang: I am so glad you liked it. Thank you for reading and for your kind words. – Giuseppe Negro Jan 24 '18 at 06:43
  • @GiuseppeNegro: It seems that I keep stumbling upon solutions you wrote a couple of years later. There is a couple of recent (as of 2021) [problems](https://math.stackexchange.com/a/4116159/121671) and [there](https://math.stackexchange.com/questions/4106171/finding-f-s-t-the-sequence-of-functions-f-nx-f-x-%E2%88%92-a-n-is-not-a-e-con) that are in a sense the same as the one you worked on here. In my solution to the former I also used switched to $\mathbb{T}$ and used Stein-Sawyer's maximal principle; but later on I found a pure real line theorem (Stein) that I used in the other problem. – Oliver Díaz May 01 '21 at 23:35
  • @GiuseppeNegro: my point being is that Stein's maximal principle for convolutions with measures with common support allows to keep the problem within $\mathbb{R}$. In any case, I do like the construction of your function $f_k$. I used a slightly different argument and showed only existence. – Oliver Díaz May 01 '21 at 23:37
  • @oliverdiaz: I have noticed your comment, I am just struggling to find the time to read your answer carefully. That is very interesting. The question is roughly the same as this one. I liked your observation linking pointwise convergence to convolution with deltas. – Giuseppe Negro May 06 '21 at 12:02
  • @GiuseppeNegro: thanks for your message. It is always a pleasure to discuss mathematics with you (if only through MSE). I was looking at the example of David C. Ulrich. I have not digested it yet, but I can see that it is a probabilistic argument of your construction. The interesting thing, is that in his use of "random translations" to apply a Borel-Cantelli type of argument, is exactly what Stein (in terms of beautiful analysis) does in his proof of the theorem of convolutions. I tried to explain that it my [posting](https://math.stackexchange.com/a/4116159/121671). – Oliver Díaz May 06 '21 at 14:38
  • @GiuseppeNegro:The source I used is Stein's Harmonic Analysis, p.p. 440-444. – Oliver Díaz May 06 '21 at 14:40

This is similar to Alex Ravsky's answer, but just answering exactly the question and nothing more. Define

$$K_d=\bigcup_{k=0}^{2^d-1} (k2^{-d}+2^{2-2d},(k+1)2^{-d})\subset [0,1]\qquad\text{for }d\geq 10$$ $$K=\bigcap_{d\geq 10}K_d.$$

It is easy to compute $\mu([0,1]\setminus K_d)=2^{2-d}$ which gives $\mu(K)\geq 1/2.$ I claim that for all $x\in[0,1]$ and for all $d\geq 10$ there is some $n$ with $2^{d-1}\leq n\leq 2^d$ such that $x+1/n\notin K_d$ - this would imply that $x+1/n\notin K$ for infinitely many $n,$ which suffices to answer the question.

Geometrically/intuitively: There is a distance of less than $2^{2-2d}$ between consecutive elements in the list $x+1/2^d, x+1/(2^d-1), \dots, x+1/2^{d-1}.$ This means that as we go along the list we cannot jump over a gap of length $2^{2-2d}.$ And the difference between the first and last elements of the list is $2^{-d}.$ Since $K_d$ has gaps of size $2^{2-2d}$ at every multiple of $2^{-d},$ not every element in the list can be in $K_d.$

Algebraically: set $n=\lfloor 1/(k2^{-d}-x)\rfloor$ where $k=\lceil x2^d\rceil+1.$ Note that $2^{d-1}\leq n\leq 2^d,$ because $1\leq k-x2^d\leq 2.$ Since $1/(n+1)\leq k2^{-d}-x\leq 1/n$ and $1/n-1/(n+1)=1/n(n+1)<1/n^2\leq 2^{2-2d},$ we have $k2^{-d}-x\leq 1/n\leq k2^{-d}+2^{2-2d}-x.$ So $x+1/n\notin K_d$ as required.

  • 24,775
  • 15
  • 47
  • I found this proof easier – ibnAbu Jan 23 '18 at 10:34
  • Your proof also works if you replace {$n$} with another countable set {$E_n$} for which $E_n \to \infty$ as $n \to \infty$ – ibnAbu Jan 23 '18 at 13:09
  • @Abdullahi_A_Ibrahim: I don't think it does. We need that the set $E=\{E_n\}$ can be split into sets $D_d$ such that the sumset $D_d+C_d$ contains $[0,1],$ for some sets $C_d$ of measure tending to zero as $d\to\infty.$ For $E_n=1/n$ the sets $D_d=\{1/2^d,1/(2^d-1),\dots,1/2^{d-1}\}$ have the property of having similar-sized gaps between consecutive elements, which allows a particularly simple $C_d$ - just a union of evenly-spaced intervals. For $E=\{2^{-n}\}$ for example, it's not so obvious what to use as $D_d$ and $C_d.$ – Dap Jan 23 '18 at 16:45
  • The pairs $C_d,D_d$ are essentially the "complementary sets" described in the abstract of https://link.springer.com/article/10.1007/s006050300001, so I think they do exist, but it takes more work. I can't access that article, but discrete analogues are proved in https://arxiv.org/abs/0910.3815 using a greedy algorithm. – Dap Jan 23 '18 at 16:48
  • If you are interested in arbitrary offsets, Giuseppe Negro's proof is likely the easiest way to go. You might need the full Stein theorem rather than the cheap version in David C. Ullrich's answer. – Dap Jan 23 '18 at 17:46
  • i noticed for all $ x \in K_d $ there some $ p $ , $ 2^{2-2d} \le p < 2^{-d} $ such that $ x+p \notin K_d $ , $ p $ not need to be $\in N $ – ibnAbu Jan 23 '18 at 20:57
  • * $ p $ not need to be $ p=\frac {1}{m} $ where $ m \in N $ – ibnAbu Jan 24 '18 at 05:23
  • And $$ \bigcup_{d \ge 10}[2^{2-2d} , 2^{-d})=[0, 2^{-10}) $$ – ibnAbu Jan 24 '18 at 09:23
  • @Dap It seems clear to me that the cheap arguument suffices in the more general case, Say $I_k$ and $A_k$ are the sets analogous to $I_k$ and $A_k$ in his and my constructions. We need to have $m(I_k)/m(A_k)\to0$ or the argument showing the maximal function does not satisfy the appropriate bound doesn't work.Now if $m(I_k)/m(A_k)\to0$ then passing to a subsequence we have $\sum m(I_k)<\infty$, $\sum m(A_k)=\infty$, and the randomization argument I gave works with no change. – David C. Ullrich Jan 24 '18 at 13:25
  • @DavidC.Ullrich: it's not clear to me why $\sum m(A_k)=\infty.$ I can see why we could assume $m(A_k)\geq \tfrac12 k \cdot m(I_k).$ But what about $m(I_k)=2^{-k}$ for example? – Dap Jan 24 '18 at 14:46
  • @Dap Fair point. It appears that I lied; $m(I_k)/m(A_k)\to0$ does not imply what I said about those sums. Now we have a new problem to think about... – David C. Ullrich Jan 24 '18 at 15:04
  • @DavidC.Ullrich: I think your argument can be generalized by replacing $I_k$ by the union of $\lceil (k^2c_k)^{-1}\rceil$ independent copies, where $c_k$ is the replacement for "$1/k^2$" – Dap Jan 24 '18 at 16:43

Heh, people banging head against wall for days, then overnight three apparently correct answers appear. On the one hand, Giuseppe Negro's answer seems the simplest, but on the other hand it depends on Stein's theorem about maximal functions.

The point to this answer is to show that that other answer can easily be converted to an explicit counterexample, based on a vague recollection of how the proof of Stein's theorem goes.

Say $X_1,\dots$ are iid random variables, uniformly distributed on $[0,1]$. Define $$I_k=[X_k,X_k+1/k^2].$$Choose $N$ so $$\sum_N^\infty\frac1{k^2}<1,$$and let $$E=\bigcup_N^\infty I_k$$and $$f=\chi_E.$$

Let $$A_k=\bigcup_{\frac k2\le n\le k}\left(I_k-\frac1n\right).$$

As Giuseppe explains, $$m(A_k)\ge c/k.$$ So Borel-Cantelli shows that almost surely almost every point lies in infinitely many $A_k$. In particular, if $B$ is the set of $x\in[0,1]\setminus E$ such that $x$ lies in infinitely many $A_k$ then $m(B)>0$. But if $x\in B$ then $f(x)\ne\lim f(x+1/n)$.

(In case it's not clear, the point to adding the restriction $n\ge k/2$ to GN's definition of $A_k$ is to ensure that if $x\in A_k$ for infinitely many $k$ then $x\in I_k-1/n$, where $n$ is unbounded.)

Note the appplication of Borel-Cantelli is a little iffy, because $A_k$ need not be a subset of $[0,1]$. If that bothers you, regard the above as a contruction on the torus $\Bbb T=\Bbb R/\Bbb Z$. As Guiseppe points out, a counterexample on $\Bbb T$ implies a counterexample on $\Bbb R$.

Edit: It's been objected that the events $A_k$ are not independent. I didn't intend to give the impression I was saying that they were. In fact $A_k\subset\Bbb T$; if we regard $\Bbb T$ as a probability space then the $A_k$ are in fact "events", but regardless they're not the independent events that I was applying Borel-Cantelli to.

With apologies for not explaining the following at the start - I thought it was well known that if $\sum m(A_k)=\infty$ and you consider "random" translates then you almost surely cover almost every point infinitely often:

For each $k$ and each $x\in\Bbb T$ let $E_{k,x}$ be the event $$x\in A_k.$$That's an honest event, a subset of the probability space where the $X_k$ live. Now if $x$ is fixed it's clear that the events $E_{k,x}$ are independent (since, again for fixed $x$, there exists a set $S_k$ such that $x\in A_k$ if and only if $X_k\in S_k$). And since $X_k$ is uniformly distributed on $\Bbb T$ it's clear that $P(E_{x,k})=m(A_k)$. So for each $x$ it is almost surely true that $x$ lies in infinitely many $A_k$, by Borel-Cantelli. Now Fubini shows that it is almost surely true that almost every $x$ lies in infinitely many $A_k$.

David C. Ullrich
  • 84,497
  • 5
  • 61
  • 132
  • The events $\{A_k\}$ don't appear to be independent since the intervals $[k/2,k]$v are not disjoint. How do you conclude that limsup of $A_k$ has measure 1? – Kavi Rama Murthy Jan 24 '18 at 01:10
  • It's a standard construction - see edit. – David C. Ullrich Jan 24 '18 at 03:07
  • Dear @DavidC.Ullrich. I just stumbled upon this nice OP after having worked on two recent (as of 2021) similar problems[1](https://math.stackexchange.com/questions/4106171/finding-f-s-t-the-sequence-of-functions-f-nx-f-x-%E2%88%92-a-n-is-not-a-e-con), [2](https://math.stackexchange.com/questions/1920408/convergence-in-norm-but-not-almost-everywhere-of-f-t-n-f-cdot-t-n-to-f/4116159#4116159). Like Guiseppe, I used a maximal principle (either Stein-Sawyer's or Stein's for convolutions). Anyway, do you know if a similar example can be constructed for general series $a_n\searrow0$? – Oliver Díaz May 01 '21 at 23:44
  • @DavidC.Ullrich: also, there is another explicitly construction for this problem in Bogachev's Measure Theory, problem 2.12.67. Interestingly enough, there is a [posting](https://math.stackexchange.com/questions/4122624/measure-of-a-cantor-like-set) appear recently with the construction related in that problem. The OP does not make reference to the problems discussed here, but there you go. – Oliver Díaz May 01 '21 at 23:51

The solution I have is a bit lengthy, so I will define f and leave out details. For $0<x<1$ $a_n(x)=[(n!)x]-n[(n-1)!x]$ ,[.] being the greatest integer function.The $x=\sum \frac {a_n(x)} {n!}$ and this representation is unique. Let $\epsilon >0$, $\frac 1 {2^{N}} <\epsilon$ and $B_k=\{x:a_{n_k} =0\}$ where $n_k =2^{N+k}$. Let $f=I_A$ where A is the union of the sets $B_k$, $k \geq 1$. Then $f(x-1/n)$ fails to converge to f(x) on $A^{c}$ whose measure is at least $1-\epsilon$.

Kavi Rama Murthy
  • 302,216
  • 20
  • 77
  • 154
  • Why does $f(x-1/n)$ fail to converge on $A^c$? – Dap Jan 23 '18 at 18:18
  • Please visit my site statmathbc.wordpress.com and click on "Solution to a problem posted on S-E" to get a detailed answer. David Ullrich's solution is neat. Mine is useful for people not familiar with probability theory. Laziness is the reason why I did not type out the entire solution here. – Kavi Rama Murthy Jan 23 '18 at 22:47
  • Looks good to me. Wikipedia calls $a_n(x)$ the ["Factorial number system"](https://en.wikipedia.org/wiki/Factorial_number_system#Fractional_values). – Dap Jan 23 '18 at 23:23

This is also not an answer but too long for a comment as well:

I dont know whether you can always find a measurable set $C$ with $\mu(C) > 0$ such that for almost every $x \in C$ , for infinite $n$ , $x+\frac{1}{n} \notin C$ where $n \in N$

But any constructive proof relying on removal of countable union of intervals $\displaystyle I=\bigcup_{i\in N} I_i$ from [a, b] to construct $C$ is not going to work because at each step we dont know which elements we have to remove from $[a,b]$ because we havent defined $C$ yet.

This means the only chance we have is to take a measurable set $C$ that has already been defined and check whether $\mu(C) > 0$ and for almost every $x \in C$ , for infinite $n$ , $x+\frac{1}{n} \notin C$ where $n \in N$

  • 1,862
  • 7
  • 25