3
$\begingroup$

In a 1973 paper by Martinet, Deshouilliers and Cohen, $A(x)$ is defined as $$A(x)=\lim_{N\to\infty}\frac{\#\{n\leq N\mid \frac{\sigma(n)}{n}≥x \}}{N}$$ where $\sigma(n)$ is the "sum-of-divisors" function $\sigma(n)=\sigma_1(n)=\sum_{d \mid n}d$. It is then stated that the following is "easily provable": $$\int_{0}^\infty x^s A(x)\, \mathrm{d}x=\frac{1}{s+1}\lim_{N\to\infty}\frac{1}{N}\sum_{n≤N}\Big(\frac{\sigma(n)}{n}\Big)^{s+1}$$ However, I cannot see the so-called "easy" proof of this statement. How is this link established?

$\endgroup$
  • $\begingroup$ This is a good question. The short answer is "this is a Mellin transform." I'm writing a longer answer now. $\endgroup$ – davidlowryduda Jun 28 '16 at 19:15
1
$\begingroup$

The relation $$ \phi(s)=\int_0^\infty x^s f(x) \frac{dx}{x} \tag{1}$$ is an example of a Mellin transform. This is something like the Fourier transform, except in different corrdinates. Just as there is an inverse Fourier transform, there is also an inverse Mellin transform. For $f$ and $\phi$ "nice" and satisfying $(1)$, they also satisfy $$ f(x)=\frac{1}{2\pi i} \int_{c - i\infty}^{c + i \infty} x^{-s} \phi(s) ds. \tag{2}$$ Similarly, nice $f$ and $\phi$ satisfying $(2)$ also satisfy $(1)$.

So your question is claiming that the Mellin transform of $xA(x)$ is $\frac{1}{s+1} \lim_{N\to \infty} \frac{1}{N} \sum (\sigma(n)/n)^{s+1}$.

I think it's actually easier to show that the inverse Mellin transform of $\frac{1}{s+1} \lim_{N\to \infty} \frac{1}{N} \sum (\sigma(n)/n)^{s+1}$ is $x A(x)$, which is what I'll do below. But to do this, I need to mention one particularly well-known inverse Mellin transform.

Generally, $$ \frac{1}{2\pi i} \int_{c - i\infty}^{c + i\infty} y^{-s} \frac{ds}{s} = \begin{cases} 1 & \text{ if } 0 \leq y < 1 \\ \frac{1}{2} & \text{ if } y = 1 \\ 0 & \text{ if } y > 1. \end{cases} \tag{3}$$ I say this is particularly well-known because this is the key idea behind Perron's Formula, which has played a pivotal role in analytic number theory. It is somewhat common to ignore the $y = 1$ case by simply not considering the case when $y = 1$.

We are now ready to prove their claim. We want to show that $$ x A(x) = \frac{1}{2\pi i} \int_{c - i \infty}^{c + i \infty} x^{-s} \lim_{N\to \infty} \frac{1}{N} \sum_{n \leq N} \left( \frac{\sigma(n)}{n}\right)^{s+1} \frac{ds}{s+1},$$ or equivalently (with some convergence conditions that I ignore in order to get to what's really going on), $$ A(x) = \lim_{N \to \infty} \frac{1}{N} \sum_{n \leq N} \frac{1}{2\pi i} \int_{(c)} \left( \frac{x}{\sigma(n)/n}\right)^{-(s+1)} \frac{ds}{s+1}.$$ The integral here is essentially $(3)$ but with the change of variables $s \mapsto s + 1$ and with $x (\sigma(n)/n)^{-1} = y$. So we have $$ \frac{1}{2\pi i} \int_{(c)} \left( \frac{x}{\sigma(n)/n}\right)^{-s-1} \frac{ds}{s+1} = \begin{cases} 1 &\text{ if } \sigma(n)/n > x \\ \frac{1}{2} &\text{ if } \sigma(n)/n = x \\ 0 &\text{ if } \sigma(n)/n < x.\end{cases}$$ Supposing for ease that we don't choose an $x$ which happens to be equal to $\sigma(n)/n$ for any $n$, then the integral is exactly the indicator function for $\sigma(n)/n > x$. But then the right hand side becomes $$ \lim_{N \to \infty} \frac{1}{N} \sum_{n \leq N} [1 \text{ if } \sigma(n)/n > x],$$ which is exactly the definition of $A(x)$.

Since $\frac{1}{s+1} \lim_{N\to \infty} \frac{1}{N} \sum (\sigma(n)/n)^{s+1}$ is the inverse Mellin transform of $xA(x)$, we also get that $xA(x)$ is the Mellin transform of $\frac{1}{s+1} \lim_{N\to \infty} \frac{1}{N} \sum (\sigma(n)/n)^{s+1}$, as we wanted to show. $\spadesuit$

$\endgroup$
1
$\begingroup$

Let $B_N(x) = \# \{ n \le N, \sigma(n)/n < x\}$. Integrating by parts $$\int_0^\infty B_N(x) x^{-s-1}dx = B_N(x)\frac{x^{-s}}{-s} |_0^\infty +\int_0^\infty \frac{x^{-s}}{s}\underbrace{dB_N(x)}_{=B_N'(x)dx} = \frac{1}{s}\sum_{n \le N} (\sigma(n)/n)^{-s}$$

and hence for $\Re(s)$ large enough $$\int_0^\infty B(x) x^{-s-1}dx=\lim_{N \to \infty} \int_0^\infty \frac{B_N(x)}{N} x^{-s-1}dx = \frac{1}{s}\lim_{N \to \infty}\frac{1}{N}\sum_{n \le N} (\sigma(n)/n)^{-s}$$ and for $-\Re(s)$ large enough, assuming it converges $$\int_0^\infty A(x) x^{-s-1}dx=\lim_{N \to \infty} \int_0^\infty \frac{N-B_N(x)}{N} x^{-s-1}dx = \frac{1}{-s}\lim_{N \to \infty}\frac{1}{N}\sum_{n \le N} (\sigma(n)/n)^{-s}$$

$\endgroup$
  • $\begingroup$ +1. Mellin inversion is overkill. You can think of this as how partial summation allows one to write $\sum_{n = 1}^{\infty} \frac{a_n}{n^s} = s \int_{1}^{\infty} \sum_{n \leq x} a_n x^{-s - 1} \, dx$. Mellin inversion (or more precisely the Perron inversion formula) is the "harder" bit, which is inverting this relation. $\endgroup$ – Peter Humphries Jul 3 '17 at 2:56

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.