5

Let $P(z)$ and $Q(z)$ be polynomials such that the degree of $Q(z)$ is exactly one degree more than the degree of $P(z)$.

And assume that $ \displaystyle \sum_{n=-\infty}^{\infty} (-1)^{n} \frac{P(n)}{Q(n)}$ converges.

If you try to evaluate the above series by integrating the function $ \pi \csc (\pi z) \frac{P(z)}{Q(z)} $ around a square contour with vertices at $\pm(N + \frac{1}{2}) \pm i (N+ \frac{1}{2})$, does the integral vanish as $N \to \infty$ through the positive integers?

I know that $\csc(\pi z)$ is uniformly bounded on the contour, but that's not enough to conclude that the integral vanishes.

If you assume that the integral does vanish, you seem to get correct results, such as $$\sum_{n=-\infty}^{\infty}\frac{(-1)^{n}}{3n+1} = \frac{2 \pi}{3\sqrt{3}}.$$

EDIT: Read both of Marko Riedel's answers. They go together.

2 Answers2

5

We introduce $$f(z) = \pi \csc(\pi z) \frac{P(z)}{Q(z)} $$ and in order to evaluate the sum in question using residues as proposed, we need to show that the integral along the square goes to zero. To do this, assume the degree of $Q(z)$ is some integer $m$ plus the degree of $P(z),$ where $m$ is at least two.

I will do two of the four sides of the contour. On the left side, call it $\Gamma_1$, we have $z = N + \frac{1}{2} + it$, with $-(N+\frac{1}{2}) \le t \le N+\frac{1}{2}.$

This yields $$\left|\int_{\Gamma_1} f(z) dz \right| \le (2N+1) \max_{\Gamma_1} |f(z)|.$$ Now $$|\csc(\pi z)| = \frac{2} {\left|e^{\pi i (N + \frac{1}{2}) -\pi t}-e^{- \pi i (N + \frac{1}{2}) +\pi t}\right|} = \frac{2}{\left|(-1)^N i e^{-\pi t} + (-1)^N i e^{\pi t}\right|} = \frac{2}{\left|e^{-\pi t} + e^{\pi t}\right|}\le 1,$$ because the maximum of the norm occurs at $t=0,$ where the denominator is minimal. It follows that the bound on the integral is in fact $$(2N+1) \max_{\Gamma_1} |f(z)| \le (2N+1) \max_{\Gamma_1} \left|\frac{P(z)}{Q(z)}\right| \in \Theta\left((2N+1)\frac{1}{N^m}\right).$$ But this is$$\Theta\left(\frac{1}{N^{m-1}}\right) \to 0 \quad \text{as} \quad N\to\infty.$$

On the top side, call it $\Gamma_2$, we have $z = t + i(N + \frac{1}{2})$, with $-(N+\frac{1}{2}) \le t \le N+\frac{1}{2}.$

This yields $$\left|\int_{\Gamma_2} f(z) dz \right| \le (2N+1) \max_{\Gamma_2} |f(z)|.$$ The cosecant term along $\Gamma_2$ has $$|\csc(\pi z)| = \frac{2} {\left|e^{\pi i t -\pi (N + \frac{1}{2})}-e^{- \pi i t +\pi (N + \frac{1}{2})}\right|} \le \frac{2}{e^{\pi (N + \frac{1}{2})} - e^{-\pi (N + \frac{1}{2})}} \in \Theta\left(e^{-\pi N}\right).$$

It follows that the bound on the second integral is $$(2N+1) \max_{\Gamma_2} |f(z)| \le (2N+1) \frac{2}{e^{\pi (N + \frac{1}{2})} - e^{-\pi (N + \frac{1}{2})}} \max_{\Gamma_2} \left|\frac{P(z)}{Q(z)}\right| \in \Theta\left((2N+1) e^{-\pi N}\frac{1}{N^m}\right).$$ But this is$$\Theta\left(\frac{e^{-\pi N}}{N^{m-1}}\right) \to 0 \quad \text{as} \quad N\to\infty.$$

This concludes the argument and shows that $$ \sum_{n=-\infty}^\infty (-1)^n \frac{P(n)}{Q(n)} = -\sum_{z_0\in S} \operatorname{Res}\left(\pi\csc(\pi z) \frac{P(z)}{Q(z)}; z=z_0\right),$$ where $S$ is the set of poles of $f(z)$ other than at the integers and we assume for simplicity that the quotient of the two polynomials does not have any poles at the integers.

Here we have made use of the fact that with $$P(n) = p_0 + p_1 n + p_2 n^2 + \cdots + p_\alpha n^\alpha$$ and $$Q(n) = q_0 + q_1 n + q_2 n^2 + \cdots + q_\beta n^\beta$$ and $\beta\ge\alpha,$ we have $$ \frac{P(n)}{Q(n)} = \frac{p_\alpha + \cdots + p_0/n^\alpha}{q_\beta n^{\beta-\alpha} + \cdots + q_0/n^\alpha} \in \Theta\left( \frac{1}{n^{\beta-\alpha}} \right) \quad \text{as} \quad n \to\infty.$$

Marko Riedel
  • 61,317
  • 1
    Thanks for responding. But my question is about about when $m=1$ which I don't think this answers. – Random Variable Apr 12 '13 at 23:10
  • 1
    Yes I missed that. I hope it can still be useful. We see that the argument for $m=1$ goes through on the two horizontal segments but not on the vertical ones. What I usually do in such cases is to group consecutive terms of the series and sum with the multiplier $\pi \cot(\pi z).$ – Marko Riedel Apr 12 '13 at 23:20
  • 1
    What I mean is to work with $$\frac{P(2m)}{Q(2m)}-\frac{P(2m+1)}{Q(2m+1)} = \frac{P^(m)}{Q^(m)}$$ and $\pi\cot(\pi z).$ Then you need cancellation of the leading coefficients to get a degree difference of at least two for the new polynomials. – Marko Riedel Apr 12 '13 at 23:26
  • Actually it would seem that this cancellation always occurs, the leading coefficient in $m$ of $P(2m)Q(2m+1)$ and of $P(2m+1)Q(2m)$ is $p_\alpha 2^\alpha q_\beta 2^\beta$ in both cases. Then the new value of $m$ is at least $2\beta-(\alpha+\beta-1) = \beta-\alpha+1$, which is what we need. – Marko Riedel Apr 12 '13 at 23:35
  • So after you find the sum of this new series, how do you go backwards to find the sum of the original series? – Random Variable Apr 12 '13 at 23:36
  • No need to go backwards. The new polynomials $P^(n)$ and $Q^(n)$ in conjunction with $\pi\cot(\pi z)$ produce an exact answer. No additional computation required. – Marko Riedel Apr 12 '13 at 23:40
  • OK. So you're just regrouping terms. But isn't there danger in regrouping terms of a conditionally convergent series? – Random Variable Apr 12 '13 at 23:45
  • You are definitely right, I don't know how I could have overlooked this. The classic example is the alternating harmonic series which by the Riemann series theorem may be re-arranged to produce anything at all. So grouping terms only works for $m\ge 2,$ where we don't really need it because we have absolute convergence anyway. – Marko Riedel Apr 13 '13 at 00:25
  • 1
    @RandomVariable: I fixed the calculation and confirmed your conjecture, see my post below. – Marko Riedel Apr 13 '13 at 20:41
3

We can in fact fix the calculation from the previous post for vertical segments so that it goes through even when $m=1,$ namely by using a tighter bound.

We have $$\left|\int_{\Gamma_1} f(z) dz \right| = \left|i \int_{-(N+1/2)}^{N+1/2} \pi\csc\left(\pi\left(N+\frac{1}{2} +it \right)\right) \frac{P\left(N+\frac{1}{2} +it\right)}{Q\left(N+\frac{1}{2} +it\right)} dt \right| \\ \le \pi \int_{-(N+1/2)}^{N+1/2} \frac{2}{e^{\pi t} + e^{-\pi t}} \frac{C}{N^m} dt \le \frac{C}{N^m} \int_{-\infty}^\infty \frac{\pi}{\cosh (\pi t)} dt \\= \frac{C}{N^m} \left[ 2 \arctan(\exp(\pi t)) \right]_{-\infty}^\infty = \frac{C}{N^m} \left(2\frac{\pi}{2} - 0\right) \\= \pi \frac{C}{N^m} \in \Theta\left(\frac{1}{N^m}\right) \to 0 \quad \text{as} \quad N\to \infty,$$ even when $m=1.$ This answers the question from the OP in the affirmative. The existence of $C$ is assured for $N$ large enough.

Marko Riedel
  • 61,317