The classic example of a function that has a drastic cancelation when summed over divisors is $\mu(n)$, with complete cancellation for every number other than $1$. Another such function is the Liouville function $\lambda(n)$. Both of these functions have have the property that $\sum_{n=1}^{\infty}\frac{f(n)}{n}=0$. Is this a general pattern? I.e do the conditions

\begin{equation} \lim_{N\to\infty}\frac{1}{N}\sum_{n=1}^{N}\left|\sum_{d|n}f(d)\right|=0\tag{i} \end{equation}

\begin{equation} f(n)=O(1)\tag{ii} \end{equation}


\begin{equation} \sum_{n=1}^{\infty}\frac{f(n)}{n}=0\tag{iii} \end{equation}

The interesting question here is about the $\mathit{convergence}$ of the sum. If we know that the sum in (iii) converges then we simply note that


Now since the terms $\sum_{d|n}f(d)$ tend to be small, the sum $\frac{\sum_{d|n}f(d)}{n^{1+\epsilon}}$ will go to infinity slower than $\zeta(1+\epsilon)$ and so


which implies that the value of the sum must be zero. A rigorous proof of the above statement is not much harder than the handwaving done above.


Interestingly enough, it is actually enough to know

\begin{equation} \frac{1}{N}\sum_{n=1}^{N}\left|\sum_{d|n}f(d)\right|=o(1)\tag{1} \end{equation}

to deduce

\begin{equation} \sum_{n=1}^{\infty}\frac{f(n)}{n}=0\tag{2} \end{equation}

to do this, we work instead under the change of variables $g(n):=\sum_{d|n}f(d)$ which shows that the problem $"(1)\implies(2)"$ is equivalent to saying that

\begin{equation} \frac{1}{N}\sum_{n=1}^{N}\left|g(n)\right|=o(1)\tag{3} \end{equation}

implies that

\begin{equation} \sum_{n=1}^{\infty}\frac{\sum_{d|n}\mu\left(\frac{n}{d}\right)g(d)}{n}=0\tag{4} \end{equation}

by Mobius inversion. Working with this sum, we get that

\begin{align*} \sum_{n=1}^{N}\frac{\sum_{d|n}\mu\left(\frac{n}{d}\right)g(d)}{n}&=\sum_{n=1}^{N}\frac{\sum_{d|n}\mu\left(\frac{n}{d}\right)g(d)}{n}\\ &=\sum_{d=1}^{N}\sum_{n=1}^{N/d}\frac{\mu\left(n\right)g(d)}{nd}\\ &=\sum_{d=1}^{N}\frac{g\left(d\right)}{d}M\left(N/d\right) \end{align*}

Where here


is the logarithmically weighted Mertens function. Strong forms of the PNT show that

\begin{equation} |M(x)|=O\left(e^{-C\sqrt{\log(x)}}\right)\tag{5} \end{equation}

for some constant $C$. Thus, for any $L>0$

\begin{align*} \sum_{n=1}^{N}\frac{\sum_{d|n}\mu\left(\frac{n}{d}\right)g(d)}{n}&=\sum_{d=1}^{N}\frac{g(d)}{d}M(N/d)\\ &=\sum_{d=N/L}^{N}\frac{g(d)}{d}M(N/d)+\sum_{d<N/L}\frac{g(d)}{d}M(N/d)\\ &=S_1+S_2 \end{align*}

The second sum $S_2$ can be treated with the inequality (5) to be $o(L)$ as $N\to\infty$. As for $S_1$, we see that

\begin{align*} \sum_{d=N/L}^{N}\frac{g(d)}{d}M(N/d)&=\sum_{j=1}^{L}\sum_{d=N/(j+1)}^{N/j}\frac{g(d)}{d}M(N/d)\\ &=\sum_{j=1}^{L}M(j)\sum_{d=N/(j+1)}^{N/j}\frac{g(d)}{d}\\ \end{align*}

Because of (3), each inner sum $\sum_{d=N/(j+1)}^{N/j}\frac{g(d)}{d}$ tends to $0$ as $N\to\infty$. Thus, first taking $N$ and then $L$ to infinity we yield our result.


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.