10
$\begingroup$

Let for $a>0$, $\displaystyle S_n(a)=\sum_{k=1}^{n-1}(1-\frac{k}{n})^{-a}\left(\frac{\log\left(k\right)}{k}-\frac{\log\left(k+1\right)}{k+1}\right)$.

I conjecture that there exist a number $c_a\in\mathbb R$ , such that : $$ \displaystyle S_n(a)\sim c_a \big(\sum_{k=1}^{n-1}{1\over k^a}\big)n^{a-2} \ln n\quad (n\to +\infty).$$

I found by elementary calculus that

When $a=1$, $ \displaystyle S_n(1)\sim \frac 32 \frac{\ln^2 n}n$ , so $c_1=\frac 32$ and when $a=2$, $ \displaystyle S_n(2)\sim \frac{\pi^2}6\ln n$, so $c_2=1$

I need some help for général case


I give this non-rigorous proof $\displaystyle S_n(a) = \sum_{k=1}^{n-1} \Big(1-{k \over n}\Big)^{-a} \Big({\ln k \over k} - {\ln (k+1) \over k+1} \Big). $ with change of index : $\displaystyle S_n (a)= \sum_{k=1}^{n-1} \Big({n \over k}\Big)^a \Big({\ln (n-k) \over n-k} - {\ln (n-k+1) \over n-k+1} \Big).$

I believe we have this asymptotic development : $\displaystyle {\ln (n-k) \over n-k} - {\ln (n-k+1) \over n-k+1} = {\ln n -1\over n^2}\Big(1+ {(2k-1)\over n } {2 \ln n -3 \over 2 \ln n-2} + \big({1\over n^3}\big)\epsilon_{k,n}\Big)$,

Thus ? $\displaystyle S_n (a) \sim c_a \Big(\sum_{k=1}^{n-1}\Big({n \over k}\Big)^a \Big) {\ln n \over n^{2}}$ with a constant $c_a$ to be determined


Explanations at the request of Dr. Wolfgang Hintze

A- Let us establish that Thank's for my friend Lou $\displaystyle \lim_{n\to + \infty}S_n(a)=\left\{\begin{array}{cl} 0 & \text { si } \: 0\leqslant a <2\\+\infty &\:\text {si } \: a\geqslant 2. \end{array}\right.$

Let $f: x\mapsto \dfrac{\log x}x - \dfrac {\log(x+1)}{x+1}.\qquad f(x)\underset{x\to + \infty}\sim \dfrac {\log x}{x^2}.$ $1)\quad\forall a\geqslant 2, \quad S_n(a)\geqslant \dfrac {n^{a}}{(n-1)^{a}}f(1)+\dfrac {n^{a}}{(n-2)^{a}}f(2) +0+0+\dots +0+\dfrac {n^{a}}{(1)^{a}}f(n-1)\underset{n\to + \infty}\sim n^{a-2}\log n.$ $2)\quad $ It's easy to prove that : $\:\:\forall x\in [3;n], \:\: 0<f(x)<\dfrac{ \log n}{x^2} \quad(1).$ If $a\in [0;2[.\quad S_n(a) =-\dfrac{\log n}n+ \displaystyle \sum _{k=1}^{n-1}\left(\dfrac {n^{a}}{(n-k)^{a}} -1\right) f(k).\quad$ write : $\:T_n(a):=\log n\displaystyle \sum _{k=3}^{n-1}\left(\dfrac {n^{a}}{(n-k)^{a}} -1\right) \dfrac 1{k^2}.$ From $ (1) $, it therefore suffices to prove that : $ \displaystyle \lim_{n\to + \infty} T_n(a) =0.$ $\displaystyle T_n(a)=\log n \sum_{k=1}^{n-3}\dfrac{n^{a} - k^{a}}{k^{a}(n-k)^2} =\dfrac{\log n}{n^2} \displaystyle \sum_{k=1}^{n-3}(n^{a}- k^{a})\left( k^{-a}+ \dfrac {2 k^{1-a}}{n-k} + \dfrac {k^{2-a}}{(n-k)^2}\right).\quad $ We also have inequalities: $\:\:\forall a \in \mathbb R^+,\: \forall k\in [1;n]: $ $0 \leqslant a<2 \implies 0\leqslant n^{a}- k^{a}\leqslant an^{a-1}(n-k).\quad (2), \quad 0\leqslant a \leqslant1 \implies 0\leqslant n^{a}- k^{a}\leqslant ak^{a-1}(n-k).\quad (3).$ They lead to: $\bullet \:\text{Si }1\leqslant a<2, \:\text {alors}\:\:0<T_n(a) < \log n \left(n^{a-2}\displaystyle \sum_{k=1}^{n-3}k^{-a} +2a n^{a-3}\sum_{k=1}^{n-3}k^{1-a}+an^{-1}\sum_{k=1}^{n-3} (n-k)^{-1}\right). \qquad (4)$ $\bullet \:\text{Si }0\leqslant a \leqslant1, \:\text {alors}\:\: 0<T_n(a) < \log n \left(n^{a-2}\displaystyle \sum_{k=1}^{n-3}k^{-a} +2a n^{-2}\sum_{k=1}^{n-3}1+an^{-1}\sum_{k=1}^{n-3} (n-k)^{-1}\right).\qquad (5)$ We check that the right-hand side of the inequalities (4) and (5) have a zero limit when $ n \to + \infty.$

B- Show that $\:\:\boxed{S_n(1) \underset{n\to + \infty}\sim \dfrac {3(\log n)^2}{2n}.}\qquad$ Let $g: x\mapsto x^2f(x) -\log x.\quad $ Then : $\:\:\displaystyle \lim_{+\infty} g=-1, \quad g \text { is bounded on } [1;+\infty[\:\: (1).\qquad$ Let $\:\:T_n:= \displaystyle \sum_{k=1}^{n-1} \dfrac k{n-k} f(k). $ Then: $ \:\: S_n(1) = T_n- \dfrac {\log n}n\:\:(2).\:\:\quad T_n=\dfrac 1n \displaystyle \sum_{k=1}^{n-1}\left( \dfrac 1k + \dfrac 1{n-k}\right) k^2f(k) =U_n +V_n +W_n\:\:$ with $U_n,V_n,W_n $ are defined by : $\displaystyle U_n:=\dfrac 1n\sum_{k=1}^{n-1} \dfrac {\log k}k, \quad V_n:= \dfrac 1n\sum_{k=1}^{n-1} \dfrac {\log k}{n-k}, \quad W_n:= \dfrac 1n\sum_{k=1}^{n-1}\left( \dfrac 1k + \dfrac 1{n-k}\right) g(k).\:\:\:$ According to $(1): \quad W_n=\mathcal O\left (\dfrac{\log n}n \right)\:\:(3).\quad \:\:U_n \underset{n\to + \infty}\sim \dfrac{(\log n)^2}{2n}\:\: (4). \qquad V_n =\displaystyle \dfrac {\log n}n\sum_{k=1}^{n-1}\dfrac 1{n-k} +\dfrac 1{n^2}\displaystyle\sum_{k=1}^{n-1}\dfrac {\log(k/n)}{1-(k/n)}. $ The Function $h:x\mapsto \dfrac{\ln x}{1-x}\text{ is continuous, monotonic, integrable on }\:]0;1[,\quad $ so $\:\:\displaystyle \lim_{n\to +\infty}\dfrac 1n\sum_{k=1}^{n-1} \dfrac {\log(k/n)}{1-(k/n)} =\int _0 ^1 h .$ This information, combined with the fact that: $\: \displaystyle \dfrac {\log n}n\sum_{k=1}^{n-1}\dfrac 1{n-k} \underset{n\to + \infty}\sim \dfrac {(\log n)^2}n,\:$ leads to: $\:\:V_n\underset{n\to + \infty}\sim \dfrac {(\log n)^2}n.\quad (5)$ $(2), (3), (4) $ et $(5)$ provide the equivalent of $ S_n (1) $ announced.


I can explain the case a = 2 if needed

$\endgroup$
3
  • $\begingroup$ have you tried to split your sum at some point ($k_s=\sqrt n$ or something like that) and estimate both contributions seperatly (the main contribution will be from the part where $n>k>k_s$, the other part should give something $\sim\log(k_s)/k_s$ which is small in comparison). i don't have the time to write something up as an answer, but my premiliminary results look rather promising. $\endgroup$
    – asgeige
    Dec 22, 2021 at 13:55
  • $\begingroup$ I don't know what u mean $\endgroup$
    – Jane
    Dec 22, 2021 at 20:43
  • $\begingroup$ @Jane Thank you very much for your explanation (+1) $\endgroup$ Jan 7 at 22:11

2 Answers 2

3
$\begingroup$

I will deal with the case that $a> 3$. We have $$ n^{ - a} S_n (a) = \sum\limits_{k = 1}^{n - 1} {\frac{1}{{k^a }}\left( {\frac{{\log (n - k)}}{{n - k}} - \frac{{\log (n - k + 1)}}{{n - k + 1}}} \right)} . $$ We split the sum at $k=\lfloor \sqrt n \rfloor $. Using the mean value theorem and some trivial estimation, we find \begin{align*} &\left|\sum\limits_{k = \lfloor \sqrt n \rfloor }^{n - 1} {\frac{1}{{k^a }}\left( {\frac{{\log (n - k)}}{{n - k}} - \frac{{\log (n - k + 1)}}{{n - k + 1}}} \right)} \right| \leq \sum\limits_{k = \lfloor \sqrt n \rfloor }^{n - 1} \frac{1}{{k^a }}\left|\frac{{\log (n - t_k ) - 1}}{{(n - t_k )^2 }}\right| \\ & \le C\frac{{\log n}}{{n^2 }}\sum\limits_{k = \left\lfloor {\sqrt n } \right\rfloor }^{n - 1} {\frac{1}{{k^{a - 2} }}} \le C\frac{{\log n}}{{n^2 }}\int_{\left\lfloor {\sqrt n } \right\rfloor - 1}^{ + \infty } {\frac{{dt}}{{t^{a - 2} }}} \le \frac{{C'}}{{a- 3}}\frac{{\log n}}{{n^{(a + 1)/2} }}, \end{align*} where $k-1<t_k<k$ is a suitable number and $C,C'>0$ are appropriate constants. Consequently, $$ \frac{{n^{2 - a} }}{{\log n}}S_n (a) = \sum\limits_{k = 1}^{\lfloor \sqrt n \rfloor - 1} {\frac{1}{{k^a }}\frac{{n^2 }}{{\log n}}\left( {\frac{{\log (n - k)}}{{n - k}} - \frac{{\log (n - k + 1)}}{{n - k + 1}}} \right)} + \mathcal{O}\!\left( {\frac{1}{{n^{(a - 3)/2} }}} \right). $$ If $1\leq k<\sqrt{n}$, then by the mean value theorem and the monotonicity of $x \mapsto (\log x-1)/x^2$, we find \begin{align*} 0<\frac{{n^2 }}{{\log n}}\left( {\frac{{\log (n - k)}}{{n - k}} - \frac{{\log (n - k + 1)}}{{n - k + 1}}} \right) & \le \frac{{n^2 }}{{\log n}}\frac{{\log (n - k) - 1}}{{(n - k)^2 }} \\ & \le \frac{{n^2 }}{{\log n}}\frac{{\log (n - \sqrt n ) - 1}}{{(n - \sqrt n )^2 }} < 2 \end{align*} for large $n$. Therefore, using the asymptotics $$ \frac{{\log (n - k)}}{{n - k}} - \frac{{\log (n - k + 1)}}{{n - k + 1}} \sim \frac{{\log n}}{{n^2 }} $$ for fixed $k$ and large $n$, and Tannery's theorem, we deduce $$ \mathop {\lim }\limits_{n \to + \infty } \frac{{n^{2 - a} }}{{\log n}}S_n (a) = \zeta (a) $$ provided $a>3$.

$\endgroup$
2
  • 1
    $\begingroup$ I want to extend my thanks and appreciation for your time and effort. I will try to understand your reasoning $\endgroup$
    – Jane
    Jan 7 at 16:14
  • 2
    $\begingroup$ @Gary. Great proof (1+). As an expert for asymtotic expansions, can you perhaps also solve the problem using my integral representation? $\endgroup$ Jan 7 at 23:05
1
$\begingroup$

This is not an answer but an extended comment which provides an integral representation of the sum which might serve as a starting point for further investigations.

We have to study the sum

$$S(a,n) =\sum_{k=1}^{n-1} f(k)\left(c(k) - c(k+1)\right)\tag{1}$$

Here $f(k) = (1-\frac{k}{n})^{-a}$ and $c(k)=\frac{\log(k)}{k}$

The peculiar form of the second factor of the summand suggests immediately to write it as an integral

$$c(k)-c(k+1) = -\int_{k}^{k+1} c'(x)\;dx\tag{2}$$

Then our sum becomes

$$S(a,n) = -\sum_{k=1}^{n-1} f(k) \int_{k}^{k+1} c'(x)\;dx\tag{3}$$

Notice now (this is a useful trick) that we can incorporate the factor $f(k)$ in the integrand writing it with the floor function as $f (\lfloor x\rfloor)$ for $k \lt x \lt k+1$ so that

$$S(a,n) = -\sum_{k=1}^{n-1} \int_{k}^{k+1} f (\lfloor x\rfloor) c'(x)\;dx\tag{4}$$

the sum can be done using $\sum_{k=1}^{n-1}\int_{k}^{k+1}=\int_{1}^{n}$ and using $c'(x) = -\frac{\log(x)-1}{x^2}$ leads to the integral representation of the sum

$$S(a,n) = I(a,n) := \int_{1}^{n} (1-\frac{\lfloor x\rfloor}{n})^{-a} \frac{\log(x)-1}{x^2}\;dx\tag{5}$$

Notice that, due to the floor function, the integrand is a discontinuous but finite function.

I have checked the equivalence $I=S$ symbolically and numerically in Mathematica.

$\endgroup$
3
  • $\begingroup$ Thank's for formula 5, but it's not easy with this formula to show that for example $ \int_{1}^{n} (1-\frac{\lfloor x\rfloor}{n})^{-1} \frac{\log(x)-1}{x^2}\;dx \sim \frac 32 \frac{\ln^2 n}n$ or $ \int_{1}^{n} (1-\frac{\lfloor x\rfloor}{n})^{-2} \frac{\log(x)-1}{x^2}\;dx \sim \frac{\pi^2}6\ln n $ $\endgroup$
    – Jane
    Jan 3 at 18:35
  • $\begingroup$ Ok, I agree, but could you please explain how you did the calculation for $a=1$ and $a=2$ $\endgroup$ Jan 4 at 12:53
  • $\begingroup$ I explained the case a = 1, do you also want the case a = 2? $\endgroup$
    – Jane
    Jan 6 at 20:50

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.