43
$\begingroup$

recent conjecture :Let $S_n=1+\frac{1}{2}+\frac{1}{3}+\ldots+\frac{1}{n}$, where $n$ is a positive integer. Prove that :there exist infinite many $n\in\mathbb{N^{+}}$ such that $$S_n-[S_n]<\dfrac{1}{n^2}$$ where $[x]$ represents the largest integer not exceeding $x$.

previous problem:How to prove that $a<S_n-[S_n]<b$ infinitely often

$\endgroup$
  • 4
    $\begingroup$ Can you find any example with $n \gt 1$? Having looked at this for small $n$, I would guess there are probably no other examples. $\endgroup$ – Henry Dec 18 '16 at 1:33
  • 10
    $\begingroup$ Heuristically this should happen a finite number of times. Consider the values $n_i$ where $\lfloor H_{n_i}\rfloor $ jumps from $i$ to $i+1$. Since $H_i\approx\ln i$, we have $n_i\approx e^i$ (this is of course a very rough approximation, but that's all we need for this heuristic). Now, at $n_i$ it's reasonable to assume (heuristically) that the fractional part $\{H_{n_i}\}$ is equidistributed in $(0, \frac1{n_i})$; this means that the probability it's less than $1/(n_i^2)$ is $\frac1{n_i}$. But then the total number of 'hits' is $\sum_i(n_i^{-1})\approx\sum_i e^{-i}$, which converges. $\endgroup$ – Steven Stadnicki Dec 23 '16 at 2:07
  • 1
    $\begingroup$ @alphacapture There's nothing specific about $\{H_{n_i}\}$ here - the heuristic is that we've added $1/n_i$ to $H_{n_i-1}$ to get a number that 'rolls over' into the next integer, so we know that $\{H_{n_i}\}\geq 0$ (by definition) and $\{H_{n_i}\}\leq 1/n_i$ (because otherwise the rollover point would've been earlier); the heuristic is just in the assertion it's equidistributed within its possible range (because we have no specific reason to believe otherwise). $\endgroup$ – Steven Stadnicki Dec 23 '16 at 23:25
  • 2
    $\begingroup$ Formally, we can say that the set $\{\alpha \in [0,1] : (S_n+\alpha)-[S_n+\alpha] < 1/n^2 \ \text{for infinitely many} \ n \in \mathbb{N}^+\}$ has measure $0$. But showing that $0$ does not belong to this set doesn't appear to be trivial. $\endgroup$ – JimmyK4542 Dec 24 '16 at 7:42
  • 4
    $\begingroup$ You should probably add to your post that you have now asked the same question on [MathOverflow](there exist infinite many $n\in\mathbb{N}$ such that $S_n-[S_n]<\frac{1}{n^2}$). See this answer and other discussions about cross-posting. $\endgroup$ – Martin Sleziak Jan 15 '17 at 11:42
21
$\begingroup$

This is a heuristic argument. It does not show that there are not infinitely many occurrences, just that without some unforeseen conditions, the probability of infinitely many occurrences is zero.


The Euler-Maclaurin Sum Formula says that $$ H_n=\log(n)+\gamma+\frac1{2n}-\frac1{12n^2}+\frac1{120n^4}+O\!\left(\frac1{n^6}\right)\tag{1} $$ Subtracting $\gamma$ and exponentiating gives $$ \begin{align} e^{H_n-\gamma} &=ne^{\frac1{2n}-\frac1{12n^2}+\frac1{120n^4}}+O\!\left(\frac1{n^5}\right)\\ &=\left(\frac1n-\frac1{2n^2}+\frac5{24n^3}-\frac1{16n^4}+\frac{47}{5760n^5}+\frac1{2304n^6}\right)^{-1}+O\!\left(\frac1{n^5}\right)\tag{2} \end{align} $$ Inversion of the power series $y=x-\frac{x^2}2+\frac{5x^3}{24}-\frac{x^4}{16}+\frac{47x^5}{5760}+\frac{x^6}{2304}$ says $$ x=y+\frac{y^2}2+\frac{7y^3}{24}+\frac{y^4}6+\frac{523y^5}{5760}+\frac{y^6}{20}+O\!\left(y^7\right)\tag{3} $$ where $x=\frac1n$ and $y=e^{\gamma-H_n}$. Taking the reciprocal of $(3)$ yields $$ n=e^{H_n-\gamma}-\frac12-\frac1{24}e^{\gamma-H_n}+\frac3{640}e^{3\gamma-3H_n}+O\!\left(\frac1{n^5}\right)\tag{4} $$ To find an $n$ so that the fractional part of $H_n$ is less than $\frac1{n^2}$, we need to find an $H_n$ so that the fractional part of $n$ is greater than $1-\frac1n$. $$ \begin{array}{r|r} H_n&n\\\hline 1&1.000000000000\\ 2&3.638675849525\\ 3&10.773523676598\\ 4&30.153290055642\\ 5&82.827475640215\\ 6&226.008738099299\\ 7&615.215019121592\\ 8&1673.187107043897\\ 9&4549.053308117194 \end{array}\tag{5} $$ Without a reason that the fractional part of $(4)$ not be uniformly distributed, the probability that the fractional part of $H_n$ is less than $\frac1{n^2}$ would be $$ \frac1n\approx e^{\gamma-H_n}\tag{6} $$ and since $$ \begin{align} \sum_{H_n=1}^\infty e^{\gamma-H_n} &=\frac{e^\gamma}{e-1}\\ &\lt\infty\tag{7} \end{align} $$ Borel-Cantelli says that the probability that there are infinitely many $n$ so that the fractional part of $H_n$ is less than $\frac1{n^2}$ is $0$.

$\endgroup$
  • 1
    $\begingroup$ How much chance that the Pythagorean Theorem is true? I guess someone can give a probabilistic "proof" that it is zero too. $\endgroup$ – Han de Bruijn Jan 12 '17 at 20:13
  • 3
    $\begingroup$ I explicitly said that this was not a proof, that it was only heuristic. If someone posts a proof, this would serve to enhance the wonder because the odds are against it. If nothing else, the inverse of the harmonic numbers was worth the exercise for me. $\endgroup$ – robjohn Jan 12 '17 at 20:20
  • 1
    $\begingroup$ I hope that some of the results that I have gotten, while they don't settle the question, are helpful to someone. $\endgroup$ – robjohn Jan 12 '17 at 20:40
  • $\begingroup$ Sure Rob, but I only wanted to point out that one has to be careful with transfer of probabilistic arguments to environments where they do not really belong. $\endgroup$ – Han de Bruijn Jan 12 '17 at 20:59
  • 1
    $\begingroup$ @HandeBruijn Out of curiosity, how do you (propose to) cast the Pythagorean Theorem in such a form as to be able to apply probabilistic heuristics to it? $\endgroup$ – Steven Stadnicki Jan 20 '17 at 22:33
1
$\begingroup$

Sorry for posting as answer but I think it would be more convenient here. Firstly, I am offering to change the title of the question to ''There is no $n$ such ... ''. It might be more attractive. Secondly, I have the following generalized conjecture:

Let $M$ be the set of all $n$ such that $$H_n - \lfloor{H_n\rfloor} < \frac{1}{n^{1+\epsilon}}.$$ Then $$\forall\epsilon>0 : |M| = \bar\eta(\epsilon) < \infty.$$

Also I suspect that $\bar\eta$ monotonically decreasing function on $(0,1]$. Below I presented possible values of $\bar\eta$ ('possible' because I am not able to check all $n$'s). Actually, I was checking only up to $10^5$ first $n$'s.points_eta Next, I tried all $n$'s up to $10^6$.points_eta6 Note that I assume $H_n$ starts with $\frac12$, not with 1.

$\endgroup$
  • $\begingroup$ Obviously false. $S_n-[S_n]<\frac1n$ holds about every n of $S_n-[S_n]<m<S_{n+1}-[S_{n+1}]$. $\endgroup$ – Takahiro Waki Jan 21 '17 at 6:14
  • $\begingroup$ @TakahiroWaki $\epsilon > 0$. $\endgroup$ – LRDPRDX Jan 21 '17 at 8:24
  • $\begingroup$ However even the case of $\frac1{n^2}$, now anyone can't say there doesn't exist one solution. As I say, we need to search up to $10^{100}$. Although you claim this conjecture holds even If n can be reduce consectively, there is not such a logic in your answer. Of course to find such a logic is more hard than solving this conjecture. $\endgroup$ – Takahiro Waki Jan 21 '17 at 9:36
  • $\begingroup$ So. What logic is behind the statement of this question? $\endgroup$ – LRDPRDX Jan 21 '17 at 9:52
  • $\begingroup$ Do you claim $ε$ is $0.8$? $\endgroup$ – Takahiro Waki Jan 21 '17 at 10:28

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.