2

Is there a standard method for finding expansions in $N$ of sums like

$$S(N)=\sum_{n=0}^N \sqrt{N^2-n^2}$$

beyond the first term?

It is easy to compute here that $$S(N)=N^2 \int_0 ^1 \sqrt{1-x^2} \mathrm d x + \mathcal O(N) = \pi N^2/4 + \mathcal O(N)$$

but finding $S(N) \approx \pi N^2 /4 + N/2$ is harder (I can do this by an unverified exchange of limits which fails at higher order), and at the moment I find the fact that the next correction seems to be a multiple of $\sqrt N$ pretty perplexing, and don't have an exact form for the coefficient. (Numerically I find $S \sim \pi N^2/4 + N/2 -0.2939955\ldots\sqrt N$, though this may not in fact be the correct form; I went up to $N=3,\!000,\!000$ plotting the ratio with $\sqrt N$).

Please note that I'm interested in generally applicable techniques for such a computation rather than this specific summation.

not all wrong
  • 16,178
  • 2
  • 35
  • 57

1 Answers1

5

Usually one can use the Euler-Maclaurin sum formula to get information like this. Its first iteration takes the form

$$ \begin{align} &\frac{1}{n} \sum_{k=0}^{n-1} f\left(\frac{k}{n}\right) \\ &\qquad = \int_0^1 f(x)\,dx + \frac{f(0)-f(1)}{2n} - \frac{1}{n} \int_0^1 \left(-nx - \lceil -nx\rceil - \frac{1}{2}\right)f'(x)\,dx, \end{align} $$

and when applied to your sum we get

$$ \frac{S(n)}{n^2} = \frac{\pi}{4} + \frac{1}{2n} + \frac{1}{n} \int_0^1 \left(-nx - \lceil -nx\rceil - \frac{1}{2}\right)\frac{x}{\sqrt{1-x^2}}\,dx. $$

Now it just remains to estimate the integral. To do this, we first expand the periodic factor in Fourier series as

$$ -nx - \lceil -nx\rceil - \frac{1}{2} = \frac{1}{\pi} \sum_{k=1}^{\infty} \frac{\sin(2\pi k n x)}{k}, $$

so that

$$ \begin{align} &\int_0^1 \left(-nx - \lceil -nx\rceil - \frac{1}{2}\right)\frac{x}{\sqrt{1-x^2}}\,dx \\ &\qquad = \frac{1}{\pi} \sum_{k=1}^{\infty} \frac{1}{k} \int_0^1 \sin(2\pi k n x) \frac{x}{\sqrt{1-x^2}}\,dx \\ &\qquad = \frac{1}{\pi} \sum_{k=1}^{\infty} \frac{1}{k} \left[\left. -\sin(2\pi k n x) \sqrt{1-x^2} \right|_0^1 + 2\pi k n \int_0^1 \cos(2\pi k n x) \sqrt{1-x^2}\,dx\right] \\ &\qquad = \frac{1}{\pi} \sum_{k=1}^{\infty} \frac{1}{k} \cdot 2\pi k n \int_0^1 \cos(2\pi k n x) \sqrt{1-x^2}\,dx \\ &\qquad = \frac{1}{2} \sum_{k=1}^{\infty} \frac{J_1(2\pi k n)}{k}, \end{align} $$

where $J_1$ is the Bessel function of the first kind of order $1$. This Bessel function has the asymptotic form

$$ J_1(x) = - \sqrt{\frac{2}{\pi x}} \cos\left(x+\frac{\pi}{4}\right) + O\left(x^{-3/2}\right) $$

as $x \to \infty$, and since $\cos(2\pi k n + \pi/4) = 1/\sqrt{2}$ we get

$$ J_1(2\pi k n) = - \frac{1}{\pi \sqrt{2 k n}} + O\left((kn)^{-3/2}\right). $$

It follows that

$$ \begin{align} \sum_{k=1}^{\infty} \frac{J_1(2\pi k n)}{k} &= -\frac{1}{\pi \sqrt{2n}} \sum_{k=1}^{\infty} \frac{1}{k^{3/2}} + O\left(n^{-3/2}\right) \\ & = -\frac{\zeta(3/2)}{\pi\sqrt{2n}} + O\left(n^{-3/2}\right) \end{align} $$

and hence that

$$ \int_0^1 \left(\lceil nx\rceil - nx - \frac{1}{2}\right)\frac{x}{\sqrt{1-x^2}}\,dx = -\frac{\zeta(3/2)}{\pi2^{3/2}\sqrt{n}} + O\left(n^{-3/2}\right). $$

Therefore

$$ \frac{S(n)}{n^2} = \frac{\pi}{4} + \frac{1}{2n} - \frac{\zeta(3/2)}{\pi 2^{3/2} n^{3/2}} + O\left(n^{-5/2}\right). $$

or

$$ S(n) = \frac{\pi}{4}n^2 + \frac{1}{2}n - \frac{\zeta(3/2)}{\pi 2^{3/2}} n^{1/2} + O\left(n^{-1/2}\right). $$

Note that the constant you estimated on the $n^{1/2}$ term is

$$ - \frac{\zeta(3/2)}{\pi 2^{3/2}} \approx -0.293\ 995\ 518\ 793\ 519\ 291. $$

  • Daniel Fischer also gave a great answer here which considers the case when $f \in C^1([0,1])$. – Antonio Vargas Jan 28 '14 at 06:01
  • Lovely, thanks! – not all wrong Jan 28 '14 at 08:39
  • For what it's worth, I assumed that using the explicit form for the difference (the second integral in your first line) with its mildly pathological behaviour made using it impossible; didn't occur to expand in a Fourier basis. This is a very nice general approach for even these non-analytic parts. – not all wrong Jan 28 '14 at 10:08
  • You're very welcome! It was an interesting problem. Sometimes you get lucky when you expand an integral in series like that--sometimes after expanding some part of the integrand in Fourier series you immediately end up with an asymptotic series, like in this question. That didn't happen here, but the resulting series was nice enough that we could estimate each term individually and combine them to get an estimate for the full sum. A similar thing happened in this answer between equations 3 and 5. – Antonio Vargas Jan 28 '14 at 16:58