9
$\begingroup$

It is claimed that the following function produces only integer values for all integer $m \geq 1$, $N \geq 2$.

$F(m,N)=\frac{N^m}{2^m}\displaystyle \sum_{j=1}^{N-1} \operatorname{cosec} ^{2m}\left(\frac{\pi j}{N}\right)$

How do I prove this?

(Source: https://twitter.com/SamuelGWalters/status/1513266704855363587 - and extensive testing numerically makes me reasonably confident it is true).

Edit to say that now I have found a solution, the following comments are not necessary to tackle the problem (but might provide further insight into F).

I found this paper: fq.math.ca/Papers1/44-3/quartgauthier03_2006.pdf. I used the method there to express $F$ solely in terms of rational numbers - in outline, this went as follows...

Writing $g_m(z) = \text{cosec}^{2m+2}(z)$ it can be shown that

$\frac{d^{2}}{dz^{2}}g_m(z) = (2m+2)(2m+3)g_{m+1}(z)-(2m+2)^2g_m(z)$

and so writing

$g_m(z)=\displaystyle\sum_{r=0}^m \phi_{r,m} \frac{d^{2r}}{dz^{2r}}[g_0(z)]$

we can establish that

$\phi_{r,m}=0$ for $r <0$ and $r>m$,

$\phi_{0,0}=1$,

$\phi_{r,m+1}=\large \frac{\phi_{r-1,m}+(2m+2)^2\phi_{r,m}}{(2m+2)(2m+3)}$

From the Mittag-Leffler theorem,

$g_0(z)=\displaystyle\sum_{n\in\mathbb{Z}}(z-n\pi)^{-2}$ so

$\frac{d^{2r}}{dz^{2r}}[g_0(z)]=(2r+1)!\displaystyle\sum_{n\in\mathbb{Z}}(z-n\pi)^{-2(r+1)}$

This enables us to write:

$F(m,N)=\displaystyle\sum_{r=1}^{m}(2r-1)!\phi_{r-1,m-1}T_{r,m}(N)$

where $T_{r,m}(N)=\displaystyle \frac{N^{m+2r}}{2^m\cdot\pi^{2r}}\displaystyle \sum_{j=1}^{N-1}\displaystyle\sum_{n\in\mathbb{Z}}( j-nN)^{-2r}$

Then using properties of the Riemann-zeta function, including for integer $r$ that $\zeta(2r)=2^{2r-1}|B_{2r}|\pi^{2r}/(2r)!$, where $(B_{2r})$ are Bernouilli numbers it can be shown that

$T_{r,m}(N) = 2^{2r-m}(N^{2r+m}-N^m)|B_{2r}|/(2r)!$

In summary,

$F(m,N)=\displaystyle\sum_{r=1}^{m}2^{2r-m-1} r^{-1} \phi_{r-1,m-1} |B_{2r}|N^m(N^r-1)(N^r+1)$

$\phi_{r,m}$ as defined above.

For reference, $B_0 = 1$, $B_1 = \frac12$, $B_r = 0$ for odd $r > 1$,

$\displaystyle \sum_{k=0}^m {m+1 \choose k}B_k=m+1$

How can I show $F(m,N)$ is always an integer for $m \geq 1, N \geq 2$?


ADDITIONAL NOTE

For reference, by numerical grind, I have found these formulae (and confirmed they give integer values):

$F(1,N) = N(N-1)(N+1)/(2 \cdot 3)$

$F(2,N) = N^2 (N-1)(N+1)(N^2+11)/(2^2 \cdot 3^2 \cdot 5)$

$F(3,N) = N^3(N-1)(N+1)(2N^4+23N^2+191)/(2^3 \cdot 3^3 \cdot 5 \cdot 7)$

$F(4,N) = N^4(N-1)(N+1)(3N^6+43N^4+337N^2+2497)/(2^4 \cdot 3^4 \cdot 5^2 \cdot 7)$

$F(5,N) = N^5(N-1)(N+1)(4N^8 + 70N^6 + 642N^4 + 4250N^2 + 29594)/(2^6 \cdot 3^5 \cdot 5 \cdot 7 \cdot 11)$

and just in case you think you've spotted a pattern:

$F(6,N) = N^6(N-1)(N+1)(1382N^{10} + 28682N^8 + 307961N^6 + 2295661N^4+13803157N^2+92427157)/k$

where $k=2^6\cdot 3^6 \cdot 5^3 \cdot 7^2 \cdot 11 \cdot 13$

Also, for all $m$,

$F(m,2) = 1$

$F(m,3) = 2^{m+1}$

$F(m,4) = 2^{m}(2\cdot2^m+1)$

$F(m,5) = 2((5+\sqrt{5})^m + (5-\sqrt{5})^m)$

$F(m,6) = (3^m + 1)(2^{2m+1}+1)-1$

$F(m,7) = $?

$F(m,8) = 2^{2m}(2^{m+1}(2+\sqrt{2})^m+2^{m+1}+2^{m+1}(2-\sqrt{2})^m+1)$

[each of these can be shown from the original definition of $F$ using known values of $\text{cosec}(\pi/N)$]

$\endgroup$
7
  • 3
    $\begingroup$ As mentioned in the link you give, this is a very particular case of the Verlinde formula. For a purely algebraic proof, see D. Zagier, Elementary Aspects of the Verlinde Formula and of the Harder-Narasimhan-Atiyah-Bott Formula, Israel Math. Conf. Proc. 9 (1996). $\endgroup$
    – abx
    Commented Apr 3, 2023 at 19:42
  • 1
    $\begingroup$ @abx - thanks for that - I've found that paper. It appears to offer lots of ways of evaluating the above function, but it's not immediate obvious that those demonstrate that the result must be an integer, so I'll take this away to digest. $\endgroup$
    – MilesB
    Commented Apr 3, 2023 at 20:33
  • $\begingroup$ See also OEIS sequence A247239. $\endgroup$ Commented Apr 4, 2023 at 15:30
  • 3
    $\begingroup$ I added some information to the OEIS page (oeis.org/A247239). In particular there is a generating function $$\sum_{m=1}^\infty (2/N)^m F(m,N) x^{2m}=1-\frac{Nx \cot(N\sin^{-1}x)}{\sqrt{1-x^2}}$$ (references for this formula to papers of Richard Stanley and me are on the OEIS page). This formula makes it easy to compute $F(m,N)$ for any fixed $m$, but I don't see that it leads to a proof of integrality. $\endgroup$
    – Ira Gessel
    Commented Apr 4, 2023 at 19:32
  • $\begingroup$ In line with Gessel's note: if $\sum_ka_k(N)x^k=\frac{x\tan(N\sin^{-1}x)}{\sqrt{1-x^2}}$ then each $a_k(N)$ also appears to be an integer. $\endgroup$ Commented Apr 5, 2023 at 15:46

2 Answers 2

4
$\begingroup$

I'll be using the generating function as pointed by Ira Gessel. Check the reference to Stanley's very interesting work, in particular on page 491. So, let $$f_N(x):=1-\frac{N\,x\,\cot(N\,\sin^{-1}x)}{\sqrt{1-x^2}}=\sum_{m=1}^{\infty}(2/N)^mF(m,N)\,x^{2m}.$$ Let's work with $g_N(x):=\log(\sin(N\sin^{-1}x))$ as it is related to $f_N=1-x\frac{d g_N}{dx}$. If $y=\cos\theta$ then Tchebychev polynomials of the $2^{nd}$-kind can be given by $U_{N-1}(\cos\theta)=\frac{\sin(N\,\theta)}{\sin\theta}$ and more explicitly $$U_{N-1}(y)=\sum_{k=0}^{\lfloor\frac{N-1}2\rfloor}\binom{N}{2k+1}\,(y^2-1)^k\,y^{N-1-2k}.$$ Tayloring for our purpose, $\sin(N\sin^{-1}x)=\sin(\sin^{-1}x)\cdot U_{N-1}(\sqrt{1-x^2})$. Hence, $$\sin(N\sin^{-1}x)=x\cdot \sum_{k=0}^{\lfloor\frac{N-1}2\rfloor}(-1)^k\binom{N}{2k+1}\,x^{2k}\,(1-x^2)^{\frac{N-1}2-k}.$$ Case 1: $N=2n+1$ is odd. Denote, for a short hand, $$P_n(x):=\sum_{k=0}^n(-1)^k\binom{2n+1}{2k+1}\,x^{2k}\,(1-x^2)^{n-k}$$ which has constant term $P_n(0)=N$. Further designate $Q_n(x):=P_n(x)-P_n(0)$. Then, we have \begin{align*} f_N&=1-x\frac{d}{dx}\left[\log x+\log\left(N+Q_n(x)\right)\right] =-x\frac{d}{dx}\left[\log N+\log\left(1+\frac1NQ_n(x)\right)\right] \\ &=x\frac{d}{dx}\left[-\log\left(1+\frac1NQ_n(x)\right)\right] =x\frac{d}{dx}\left[\sum_{m=1}^{\infty}\frac1m\cdot\left(\frac1{N}Q_n(x)\right)^m\right] \\ &=\sum_{m=1}^{\infty}\frac1{N^m}\cdot Q_n(x)^{m-1}\cdot x\frac{d}{dx}Q_n(x). \end{align*} At this stage, observe that the coefficients of $Q_n(x)$ are divisible by $2$, hence $Q_n(x)^{m-1}$ has a factor of $2^{m-1}$; while $\frac{dQ_n}{dx}$ contributes a factor of $2$ from its coefficients. All together, the polynomial $\frac{N^m}{2^m}\left(\frac{Q_n(x)^{m-1}}{N^m}\cdot x\frac{d}{dx}Q_n(x)\right)$ has integer coefficients. That means the coefficient of $\frac{N^m}{2^m}\left(\frac{2^m}{N^m}F(m,N)\right)$ are indeed integers, as required.

Case 2: $N=2n$ even. The argument is analogous (there is an extra factor which is easily handled).

$\endgroup$
2
  • $\begingroup$ It should be noted that this generating function for $f_N(x)$ (in a slightly different form) is also found in Zagier's paper, Elementary Aspects of the Verlinde Formula and of the Harder-Narasimhan-Atiyah-Bott Formula, Israel Math. Conf. Proc. 9 (1996), archive.mpim-bonn.mpg.de/id/eprint/2138/1/preprint_1994_5.pdf $\endgroup$
    – Ira Gessel
    Commented Apr 6, 2023 at 18:09
  • $\begingroup$ @T. Amdeberhan - Thanks for this. You will see that I have submitted my own solution and prefer it over yours mainly because it uses simpler techniques. Clearly your solution is pretty slick and uses some more technical ideas including generating functions, Tchebychev polys and calculus (and it is shorter!). However, having put the effort in to understand Zagier's paper and structured my own version of a solution, I will personally stick with my approach for now. $\endgroup$
    – MilesB
    Commented Apr 6, 2023 at 22:45
1
$\begingroup$

It turns out rewriting the function in terms of Bernouilli numbers etc, while useful in generating specific values of $F(m,N)$, is not an easy path to proving they are integers. The following proof is reasonably elementary (knowledge of complex $n$th roots of $1$, no calculus) and is drawn from D. Zagier, Elementary Aspects of the Verlinde Formula and of the Harder-Narasimhan-Atiyah-Bott Formula, Israel Math. Conf. Proc. 9 (1996) - see Theorem 1 part (vii). (I've given more detail on some of the steps, and structured it to focus on the proof of integral value).

PS - I'm an actuary, with no academic credentials beyond a Maths BA from 30 years ago, but I've enjoyed reading your comments and finding a solution to a tweet that I read 1 year ago (almost to the day!) and which has bugged me ever since.

APPROACH

Fix N.

(1) We will define a function $f(g,\mathbf{x})$ on integer $g \geq 0$ and $\mathbf{x}=(x_1,x_2,...,x_{N-1})$ for integer $x_i \geq 0$.

Then, defining $\mathbf{e_1}=(1,0,0,...0), \mathbf{e_2}=(0,1,0,...,0), ... \mathbf{e_{N-1}}=(0,0,...0,1)$

we will show

(2) $f(m+1,\mathbf{0}) = F(m,N)$

(3) $f(0,\mathbf{0})$ is an integer

(4) $f(0,\mathbf{e_t})$ is an integer

(5) $f(0,\mathbf{e_t+e_u})$ is an integer

(6) $f(0,\mathbf{e_t+e_u+e_v})$ is an integer

(7) $\displaystyle \sum_{t} f(0,\mathbf{x+e_t})f(0,\mathbf{y+e_t})=f(0,\mathbf{x+y})$ for all $\mathbf{x,y}$

(8) using (3) - (7), $f(0,\mathbf{x})$ is an integer for all $\mathbf{x}$

(9) $\displaystyle \sum_{t} f(g,\mathbf{x+2e_t})= f(g+1,\mathbf{x})$ for all $g$

(10) using (8) - (9), $f(g,\mathbf{x})$ is an integer for all $g$ and $\mathbf{x}$

and so conclude from (2) and (10) that $F(m,N)$ is an integer.

All summations are over $\{1,2,...,N-1\}$.

STEPS (1), (2)

(1)

Let

$s(z)=\displaystyle \frac{-2N}{(z - z^{-1})^2}$

$r_a(z) =\displaystyle \frac{z^{a} - z^{-a}}{z - z^{-1}}$

$r(z,\mathbf{x}) = \displaystyle\prod_a r_a(z)^{x_a}$

$\omega = e^{i \pi / N}$

$f(g,\mathbf{x})=\displaystyle \sum_{j=1}^{N-1} s(\omega^j)^{g-1}r(\omega^j,\mathbf{x})$

(2)

$s(\omega^j) = \displaystyle\frac{-2N}{(2i\sin(\pi j / N))^2}=\frac{N}{2\sin^2(\pi j / N)}\quad$ and $\quad r(\omega^j,\mathbf{0})=1\quad$ so

$f(m+1,\mathbf{0})=(\frac{N}{2})^m\sum\frac{1}{\sin^{2m}(\pi j/N)} = F(m,N)$

USEFUL RESULTS

For the next steps, the following will be useful.

Let $T_k=\displaystyle \sum_j \omega^{jk} + \omega^{-jk}$ and $U_k=\displaystyle\sum_j \frac{\omega^{jk}-\omega^{-jk}}{\omega^j-\omega^{-j}}$

Then (derivation in footnote):

$\begin{cases} T_k = 2(N-1), & U_k = 0 & \text{if $k$ is multiple of $2N$} \\ T_k = -2, & U_k = 0 & \text{if $k$ even, and not a multiple of $2N$}\\ T_k = 0, & U_k = N-k+2Nq & \text{if $k$ odd (some integer $q$)}\\ \end{cases}$

STEPS (3) - (6)

(3)

$f(0,\mathbf{0}) = \displaystyle \sum_j \frac{(\omega^j - \omega^{-j})^2}{-2N}=\frac{-1}{2N}\sum (\omega^{2j}+\omega^{-2j}-2)=\frac{-1}{2N}(T_2 - 2(N-1)) = 1$

(4)

$r(\omega^j, \mathbf{e_t})= \displaystyle \frac{\omega^{jt} - \omega^{-jt}}{\omega^j - \omega^{-j}}$,

so $r(\omega^j,\mathbf{e_1})=1$, so $f(0,\mathbf{e_1})=1$

For $t>1$,

$f(0,\mathbf{e_t})=\displaystyle -\frac{1}{2N} \sum_j (\omega^j-\omega^{-j})(\omega^{jt} - \omega^{-jt})=-\frac{1}{2N}(T_{t+1}-T_{t-1})=0$

(5)

$r(\omega^j,\mathbf{e_t+e_u})=\displaystyle \frac{\omega^{jt} - \omega^{-jt}}{\omega^j - \omega^{-j}}\frac{\omega^{ju} - \omega^{-ju}}{\omega^j - \omega^{-j}}$ so

$f(0,\mathbf{e_t+e_u})=-\displaystyle\frac{1}{2N}\sum_j (\omega^{j(t+u)}+\omega^{-j(t+u)}-\omega^{j(u-t)}-\omega^{-j(u-t)})$

$=-\frac{1}{2N}(T_{t+u}-T_{t-u})$ from which

$f(0,\mathbf{e_t+e_u})= \begin{cases} 1 & \text{if }t = u \\ 0 & \text{if }t \neq u \end{cases}$

(6)

$f(0,\mathbf{e_t+e_u+e_v})=\displaystyle -\frac{1}{2N}\sum_j\frac{(\omega^{jt} - \omega^{-jt})(\omega^{ju} - \omega^{-ju})(\omega^{jv} - \omega^{-jv})}{\omega^j - \omega^{-j}}$

$=-\frac{1}{2N}(U_{t+u+v}-U_{t+u-v}+U_{t-u-v}-U_{t-u+v})$

If $t+u+v$ etc are even, then this is $0$.

If $t+u+v$ etc are odd, then (since $U_k = N - k + \text{multiple of $2N$}$)

$U_{t+u+v}-U_{t+u-v}+U_{t-u-v}-U_{t-u+v}$

$=(t+u+v)-(t+u-v)+(t-u-v)-(t-u+v)+2Nq = 2Nq$ for integer $q$.

So $f(0,\mathbf{e_t+e_u+e_v})$ is an integer.

ANOTHER USEFUL RESULT

Let $a_{jk} =\displaystyle \sum_t r_t(\omega^j)r_t(\omega^k)=\frac{\sum (\omega^{jt} - \omega^{-jt})(\omega^{kt}-\omega^{-kt})}{(\omega^j-\omega^{-j})(\omega^k-\omega^{-k})}$

Along similar lines to (4), this is $0$ except

$a_{jj}=\displaystyle \frac{1}{(\omega^j - \omega^{-j})^2}(T_{2j}-2(N-1))=\frac{-2N}{(\omega^j - \omega^{-j})^2}$

So $a_{jk}= \begin{cases} s(\omega^j) & \text{if }j=k\\ 0 & \text{otherwise} \end{cases}$

STEPS (7) - (10)

(7)

$\displaystyle \sum_{t} f(0,\mathbf{x+e_t})f(0,\mathbf{y+e_t})$

$=\displaystyle \sum_t \left[ \sum_j s(\omega^j)^{-1}\prod_a r_a(\omega^j)^{x_a}\cdot r_t(\omega^j) \sum_k s(\omega^k)^{-1}\prod_b r_b(\omega^k)^{y_b}\cdot r_t(\omega^k) \right]$

$=\displaystyle \sum_j \sum_k s(\omega^j)^{-1}s(\omega^k)^{-1}\prod_a r_a(\omega^j)^{x_a}\prod_b r_b(\omega^k)^{y_b}\sum_t r_t(\omega^j) r_t(\omega^k)$

$=\displaystyle\sum_j s(\omega^j)^{-1} s(\omega^j)^{-1}\prod_a r_a(\omega^j)^{x_a+y_b}s(\omega^j)$ using the previous result for $a_{jk}$

$=f(0,\mathbf{x+y})$

(8)

Let $A_n = \{\mathbf{x}:x_i \geq 0, x_1+x_2+...+x_{N-1} = n\}$

We prove by induction on $n$, that $f(0,\mathbf{x})$ is an integer for $\mathbf{x} \in A_n$

The results from (3) - (6) show this is the case for $n \leq 3$.

So pick $n \geq 4$ and assume that the function is an integer for $A_m$ for all $m < n$.

$\mathbf{x} \in A_n \implies$ for some $a$ and $b$, $\mathbf{x-e_a-e_b} \in A_{n-2}$

so applying (7), $\displaystyle \sum_t f(0,\mathbf{x-e_a-e_b+e_t})f(0,\mathbf{e_a+e_b+e_t})=f(0,\mathbf{x})$

So $f(0,\mathbf{x})$ is an integer, because $\mathbf{x-e_a-e_b+e_t} \in A_{n-1}$ and $\mathbf{e_a+e_b+e_t} \in A_3$ so these give integer values by induction assumption.

(9)

$\displaystyle \sum_{t} f(g,\mathbf{x+2e_t})=\sum_t \sum_j s(\omega^j)^{g-1}r(\omega^j,\mathbf{x})r_t(\omega^j)^2$

$=\displaystyle \sum_j s(\omega^j)^{g-1}r(\omega^j,\mathbf{x})\sum_t r_t(\omega^j)r_t(\omega^j) = \sum_j s(\omega^j)^gr(\omega^j,\mathbf{x})$ using the $a_{jj}$ result

so $\displaystyle \sum_{t} f(g,\mathbf{x+2e_t})=f(g+1,\mathbf{x})$

(10)

It is a straightforward induction on $g$ using (8) and (9), to deduce that $f(g,\mathbf{x})$ is an integer for all $g$.

This completes the proof.


FOOTNOTE

Note that $\omega^{Nk} = (-1)^k$ and $\omega^{-1} = \omega^{2N-1}$

If $k$ is a multiple of $2N$ then $\omega^k = \omega^{-k}=1$, so $T_k = 2(N-1)$, $U_k=0$

Otherwise, summing as geometric progressions,

$T_k= \displaystyle \frac{\omega^k-\omega^{Nk}+\omega^{(N+1)k}-1}{1-\omega^k}$ which simplifies to

$-2$ if $k$ is even and $0$ if $k$ is odd.

It is straightforward to show that

$\displaystyle\frac{\omega^{jk}-\omega^{-jk}}{\omega^j-\omega^{-j}}=\omega^{j(k-1)}+\omega^{-j(k-1)}+\omega^{j(k-3)}+\omega^{-j(k-3)}+...+\omega^{jp}+\omega^{-jp}+a$

where

$p = 2$ and $a=1$ if $k$ is odd,

$p=1$ and $a=0$ if $k$ is even.

Thus $U_k=T_{k-1} + T_{k-3} + ... T_p+(N-1)a$

If $k$ is even then $U_k=0$

If $k$ is odd, then $T_{k-1} + ... + T_p$ contains $(k-1)/2$ instances of $T_{\text{even}}$ which are either $-2$ or $-2+2N$, so $U_k = -(k-1) + 2Nq + (N-1)$ where $q$ counts the occurrences of multiples of $2N$ in $\{2, 4, ... k-1\}$.

So, if $k$ is odd, $U_k = N-k+2Nq$ for some integer $q$.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .