3
$\begingroup$

While answering a question involving Harmonic numbers $H_n$, I wanted to simplify the terms $$f_n = \sum_{i=0}^{n-1}{H_i}^2 - \frac{1}{n}\sum_{0\le i,j\le n-1}H_i H_j. $$ To do so, I used SageMath to compute $f_n$ for $1\le n \le 100$ and then found the numerators of the resulting sequence to be OEIS sequence A187487; i.e., the $n$th numerator is the numerator of $n-H_n$. Indeed, all cases computed are consistent with the following being an identity for $n\ge 1$:

$$\sum_{i=0}^{n-1}{H_i}^2 - \frac{1}{n}\sum_{i=0}^{n-1}\sum_{j=0}^{n-1}H_i H_j = n - H_n $$ which is the same as $$\sum_{i=0}^{n-1}{H_i}^2 - \frac{1}{n}\left(\sum_{i=0}^{n-1}H_i\right)^2 = n - H_n. $$ Does anyone have an idea of how to prove this? Or a reference?

$\endgroup$
  • 2
    $\begingroup$ did you try looking at $f_n-f_{n-1}$ ? and expanding everything (as $\sum\sum\sum\sum \frac1{kl}$ ) ? $\endgroup$ – reuns Sep 25 '16 at 22:47
  • $\begingroup$ @user1952009 - I just edited to show a double-summation form that can be factored, so maybe that will help. $\endgroup$ – r.e.s. Sep 25 '16 at 22:51
  • $\begingroup$ ...For example, $$\sum_iH_i^2=\sum_i\sum_j\sum_k\frac1{jk}[j\leqslant i]\,[k\leqslant i]$$ and so on... $\endgroup$ – Did Sep 25 '16 at 22:51
2
$\begingroup$

After looking at this some more, I think a better way of approaching it is via summation by parts (or Abel's lemma), the discrete analog of integration by parts. This isn't Abel's formula in full generality, but it's all we need, and it's easy to verify:

$$\sum_{k=0}^{n-1} a_k = na_n-\sum_{k=1}^{n} k(a_{k}-a_{k-1}).$$

From this we can compute

\begin{align}\sum_{k=0}^{n-1} H_k &= nH_n-\sum_{k=1}^{n} k(H_{k}-H_{k-1}) \\&=nH_n-\sum_{k=1}^{n} k \frac1{k} \\&=nH_n-n \\&=n(H_n-1) \end{align}

and

\begin{align}\sum_{k=0}^{n-1} H_k^{\,2} &= nH_n^{\,2}-\sum_{k=1}^{n} k(H_{k}^{\,2}-H_{k-1}^{\,2}) \\&=nH_n^{\,2}-\sum_{k=1}^{n}k\left((H_{k-1}+\frac1{k})^2 -H_{k-1}^{\,2}\right) \\&=nH_n^{\,2}-\sum_{k=1}^{n}k\left(\frac1{k^2}+\frac{2H_{k-1}}{k} \right) \\&=nH_n^{\,2}-\sum_{k=1}^{n}\big(\frac1{k}+2H_{k-1}\big) \\&=nH_n^{\,2}-\sum_{k=1}^{n}\frac1{k}-2\sum_{k=1}^{n}H_{k-1} \\&=nH_n^{\,2}-H_n-2\sum_{k=0}^{n-1}H_{k} \\&=nH_n^{\,2}-H_n-2n(H_n-1) \\&=\frac{\big(n(H_n-1)\big)^2}{n}+n-H_n \\&=\frac1{n}\left(\sum_{k=0}^{n-1}H_k\right)^2+n-H_n, \end{align}

so

$$ \sum_{k=0}^{n-1} H_k^{\,2} - \frac1{n}\left(\sum_{k=0}^{n-1}H_k\right)^2=n-H_n,$$

as desired.


By the way, essentially the same argument, but using integration by parts instead of Abel's lemma, yields the formulas

$$\int_e^x (\ln t)^2 \,dt-\frac1{x}\left(\int_e^x \ln t\; dt\right)^2=x-e$$

and

$$\int_1^x (\ln t)^2 \,dt-\frac1{x}\left(\int_1^x \ln t\; dt-1\right)^2=x-2.$$

$\endgroup$
  • 1
    $\begingroup$ (+1) for the summation by parts. Although it's slightly more work to verify the full formula $$\sum_{k=0}^{n-1}a_k b_k = a_n\sum_{k=0}^{n-1}b_k - \sum_{k=1}^{n}(a_k-a_{k-1})\sum_{j=0}^{k-1}b_j,$$ I notice this will give an even shorter derivation, using it first with $a_k=H_k,\,b_k=1$ and then again with $a_k=b_k=H_k$. $\endgroup$ – r.e.s. Sep 28 '16 at 17:42
2
$\begingroup$

Your formula is correct. The algebra is kind of messy, but here it is:

\begin{align} \sum_{i=0}^{n-1}H_i&=\sum_{i=0}^{n-1} \sum_{j=1}^i \frac1{j} \\&=\sum_{j=1}^{n-1}\frac{n-j}{j} \quad\quad\quad\scriptsize{\text{(since }\frac1{j}\text{ occurs in the above sum }n-j\text{ times)}} \\&=n\sum_{j=1}^{n-1}\frac1{j}-\sum_{j=1}^{n}1 \\&=nH_{n-1}-n+1. \\ \\ \sum_{i=0}^{n-1}H_i^{\,2}&=\sum_{i=0}^{n-1}\left( \sum_{j=1}^i \frac1{j} \cdot\sum_{k=1}^i \frac1{k}\right) \\&=\sum_{\substack{0\le i \le n-1\\1\le j \le i\\1 \le k \le i}}\frac1{jk} \\&=\sum_{j=1}^{n-1}\sum_{k=1}^{n-1} \frac{n-\max(j,k)}{jk} \quad\quad\quad\scriptsize{\text{(since }\frac1{jk}\text{ occurs in the above sum }n-\max(j,k)\text{ times)}} \\&=\sum_{j=1}^{n-1}\left(\frac1{j}\left(\sum_{k=1}^{j-1}\frac{n-j}{k}+\sum_{k=j}^{n-1}\frac{n-k}{k} \right) \right) \\&=\sum_{j=1}^{n-1}\left(\frac1{j}\left((n-j)\,H_{j-1}+\sum_{k=j}^{n-1}\frac{n}{k} - \sum_{k=j}^{n-1}1\right) \right) \\&=\sum_{j=1}^{n-1}\frac{(n-j)\,H_{j-1}+n(H_{n-1}-H_{j-1}) - (n-j) }{j} \\&=\sum_{j=1}^{n-1}\frac{-jH_{j-1}+nH_{n-1} - (n-j)}{j} \\&=\sum_{j=1}^{n-1}\frac{nH_{n-1} - n}{j} - \sum_{j=1}^{n-1}\frac{jH_{j-1} - j}{j} \\&=n(H_{n-1}-1)H_{n-1}-\sum_{j=1}^{n-1}H_{j-1}+n-1 \\&=n(H_{n-1}-1)H_{n-1}-\left((n-1)H_{n-2}-(n-1)+1\right)+n-1\quad\quad\quad\scriptsize\text{(by the first formula)} \\&=n(H_{n-1}-1)H_{n-1}-(n-1)H_{n-2}+2n-3 \\&=n H_{n-1}^{\,2}-nH_{n-1}-(n-1)(H_{n-1}-\frac1{n-1})+2n-3 \\&=n H_{n-1}^{\,2}-nH_{n-1}-(n-1)H_{n-1}+(n-1)\frac1{n-1}+2n-3 \\&=n H_{n-1}^{\,2}+(1-2n)H_{n-1}+2n-2 \\&=H_{n-1}(nH_{n-1}+1-2n)+2n-2. \end{align}

It follows that

\begin{align} \sum_{i=0}^{n-1}{H_i}^2 - \frac{1}{n}\left(\sum_{i=0}^{n-1}H_i\right)^2 &=H_{n-1}(nH_{n-1}+1-2n)+2n-2-\frac1{n}(nH_{n-1}-n+1)^2 \\&=nH_{n-1}^{\,2} +H_{n-1}(1-2n)+2n-2-\frac{n^2H_{n-1}^{\,2}+n^2+1-2n^2H_{n-1}+2nH_{n-1}-2n}{n} \\&=nH_{n-1}^{\,2} +H_{n-1}(1-2n)+2n-2-nH_{n-1}^{\,2}-n-\frac1{n}+2nH_{n-1}-2H_{n-1}+2 \\&=H_{n-1}(1-2n+2n-2) +2n-2 -n-\frac1{n}+2 \\&=n-H_{n-1}-\frac1{n} \\&=n-H_n. \end{align}

$\endgroup$
  • $\begingroup$ :D and $f_n-f_{n-1}$ it doesn't work ? (should leave us with less terms) $\endgroup$ – reuns Sep 26 '16 at 23:44
  • $\begingroup$ I didn't try taking differences since it was clear that this direct approach was working. There's probably some way to organize things to simplify the algebra and show why this is true. $\endgroup$ – Mitchell Spector Sep 26 '16 at 23:46
  • 1
    $\begingroup$ (+1) Thank you for the nice direct derivation. I just found the paper "New harmonic number identities with applications" by Roberto Tauraso, which shows how to generate a variety of Harmonic number identities using what it calls "stuffle products". (The formula you found for $\sum_{i=0}^{n-1}H_i^{\,2}$ appears as (3) in a list of such identities on p. 3. Although I can derive it easily using the method given in that paper, I don't yet understand exactly why that method works.) $\endgroup$ – r.e.s. Sep 27 '16 at 5:16
  • $\begingroup$ I just posted a new answer, which gives a much clearer approach to proving this. It also shows how you can come up with a similar formula using integrals and logarithms instead of sums and harmonic numbers. $\endgroup$ – Mitchell Spector Sep 28 '16 at 2:34

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.