# Arithmetic Functions and Dirichlet Convolution

Published:

$\def\bb{\mathbb}$An arithmetic function is a sequence $f\colon\bb N\to\bb C$. Some of the important arithmetic functions we study in number theory are:

• $\omega(n)$: the number of distinct prime divisors of $n$.
• $\mu(n)$: the Möbius function, defined by \begin{align*} \mu(n) = \begin{cases} (-1)^{\omega(n)} & \text{if $n$ is square-free},\\[.5em] \hfil 0 \hfil & \text{otherwise}. \end{cases} \end{align*} (An integer $n$ is square-free if it has no square factors other than $1$.)
• $\varphi(n)$: the (Euler) totient function, which counts the number of integers between $1$ and $n$ that have no common divisors with $n$, i.e., \begin{equation*} \varphi(n) = \sum_{\substack{k\leqslant n\\\operatorname{hcf}(k,n)=1}}1. \end{equation*}
• $\Lambda(n)$: the (von) Mangoldt function, which outputs $\log p$ if $n$ is a prime power $p^k$, and is zero otherwise, i.e., \begin{align*} \Lambda(n) = \begin{cases} \log p & \text{if $n=p^k$ for some prime $p$ and integer $k\geqslant 1$},\\[.5em] \hfil 0 \hfil & \text{otherwise}. \end{cases} \end{align*}

Here is a table of the first few values of these functions.

$n$$\texttt 1$$\texttt 2$$\texttt 3$$\texttt 4$$\texttt 5$$\texttt 6$$\texttt 7$$\texttt 8$$\texttt 9$$\texttt{10}$
$\omega(n)$$\texttt 0$$\texttt 1$$\texttt 1$$\texttt 1$$\texttt 1$$\texttt 2$$\texttt 1$$\texttt 1$$\texttt 1$$\texttt 2$
$\mu(n)$$\texttt 1$$\texttt{-1}$$\texttt{-1}$$\texttt 0$$\texttt{-1}$$\texttt 1$$\texttt {-1}$$\texttt 0$$\texttt 0$$\texttt 1$
$\varphi(n)$$\texttt 1$$\texttt 1$$\texttt 2$$\texttt 2$$\texttt 4$$\texttt 2$$\texttt 6$$\texttt 4$$\texttt 6$$\texttt 4$
$\Lambda(n)$$\texttt 0$$\texttt{log 2}$$\texttt{log 3}$$\texttt {log 2}$$\texttt {log 5}$$\texttt 0$$\texttt {log 7}$$\texttt{log 2}$$\texttt {log 3}$$\texttt 0$

## Some divisor sums

The only non-zero values the Möbius function takes on are $\pm1$. It might be informative to study the cancellation which we obtain when we sum over its values. Intuitively, we might try to study the sum \begin{equation*} M(x) = \sum_{n\leqslant x} \mu(n) \end{equation*} of its consecutive values, but this turns out to be a difficult sum to understand. In fact, it was conjectured that $|M(x)|<\sqrt x$, which seems reasonable if we plot $M(x)$ and $\pm\sqrt x$ on the same axes, but it’s actually false, i.e., there exists a large value of $x$ for which $M(x) > \sqrt x$.

It turns out the sum we should study is that over the divisors of a given $n$. Let $\varepsilon(n)$ denote the arithmetic function which is $1$ when $n=1$, and $0$ everywhere else. It turns out that $$\label{eq:muDivisorSum} \sum_{d\mid n} \mu(d) = \varepsilon(n),$$ where the sum is over all divisors $d$ of $n$. For example, we see that for $n=6$, we have \begin{equation*} \mu(1) + \mu(2) + \mu(3) + \mu(6) = 1 -1 -1 + 1 = 0 = \varepsilon(6). \end{equation*}

The proof of this fact is quite straightforward.

Proof of \eqref{eq:muDivisorSum}: In the case when $n=1$, we clearly have $\sum_{d\mid 1}\mu(d)=\mu(1)=1$. For $n\geqslant 2$, factorise $n=\prod_{i=1}^{k}p_i^{n_i}$ by the fundamental theorem of arithmetic, where $k=\omega(n)$. Observe that the sum has non-zero contributions from square-free divisors only, i.e., \begin{align*} \sum_{d\mid n}\mu(d) &= \mu(1)+\mu(p_1)+\cdots+\mu(p_k) + \mu(p_1p_2) + \cdots + \mu(p_1\cdots p_k)\\[-7pt] &= 1+(-1)^1 + \cdots + (-1)^1 + (-1)^2 + \cdots + (-1)^k\\[1pt] &= 1+ \binom k1 (-1)^1 + \binom k2(-1)^2 + \cdots + \binom kk (-1)^k = (1-1)^k = 0, \end{align*} by the binomial theorem, which completes the proof. $\tag*{\Box}$

We have a similarly nice result for the divisor sum of the totient function. For any integer $n\geqslant 1$, $$\label{eq:phiDivisorSum} \sum_{d\mid n} \varphi(d) = n.$$ To prove this formally, one would partition $\bb Z/n\bb Z$ according to the value of $\operatorname{hcf}(n,r)$ for each residue $r$, obtaining $n=\#(\bb Z/n\bb Z)=\sum_{d\mid n}\varphi(d)$. But we can do it in a more informal way which conveys the essential idea behind the proof.

Proof of \eqref{eq:phiDivisorSum}: Consider the $n$ fractions $1/n, \dots, n/n$, and simplify them so that they are in reduced form. Each fraction, when reduced, will have a denominator which is a divisor of $n$. But for each divisor $d$ of $n$, there are precisely $\varphi(d)$ fractions with denominator $d$. Thus the total number of fractions is $\sum_{d\mid n}\varphi(d)$. $\tag*{\Box}$

Finally, a result which relates the two functions $\mu$ and $\varphi$ is the following. $$\label{eq:phiMuDivisorSum} \varphi(n) = \sum_{d\mid n}\mu (d)\,\frac nd,$$ or written differently, $\varphi(n)= \sum_{ab = n} \mu(a)\,b$.

Proof of \eqref{eq:phiMuDivisorSum}: Let the symbol $\boldsymbol 1_{\mathcal P}$ be equal to $1$ if $\mathcal P$ is true, and $0$ otherwise. By definition of $\varphi$ and \eqref{eq:muDivisorSum}, we have \begin{align*} \varphi(n) = \sum_{\substack{k\leqslant n\\\operatorname{hcf}(k,n)=1}}1=\sum_{k=1}^n \varepsilon(\operatorname{hcf}(k,n)) &= \sum_{k=1}^n\sum_{d\mid(k,n)}\mu(d)\\[0pt] &= \sum_{k=1}^n\sum_{\substack{d\mid n\\d\mid k}}\mu(d)\\[0pt] &= \sum_{k=1}^n\sum_{\substack{d\mid n}}\mu(d)\, \boldsymbol 1_{d\mid k}\\[0pt] &= \sum_{d\mid n}\mu(d)\sum_{k=1}^n \boldsymbol 1_{d\mid k} = \sum_{d\mid n}\mu(d)\,\frac nd, \end{align*} which completes the proof. $\tag*{\Box}$

## Dirichlet Convolution

In general, sums of the form \begin{equation*} \sum_{d\mid n} f(d)\,g\Big(\frac nd\Big) \qquad \text{or, written differently,} \qquad \sum_{ab=n} f(a)\,g(b), \end{equation*} are quite common in number theory. This leads to the following definition.

Definition. Let $f,g\colon\bb N\to\bb C$ be two arithmetic functions. Their (Dirichlet) convolution, denoted by $f*g$, is the function defined by \begin{equation*} (f*g)(n) = \sum_{d\mid n} f(d)\,g\Big(\frac nd\Big) \end{equation*} for all $n\in\bb N$.

In terms of this notation, we may express \eqref{eq:phiMuDivisorSum} as \begin{equation*} \varphi = \mu * \operatorname{id}, \end{equation*} where $\operatorname{id}$ denotes the identity function $\operatorname{id}(n) = n$ for all $n$.

It turns out that the set of arithmetic functions with $f(1)\neq 0$ forms an Abelian group under the operation of Dirichlet convolution. Indeed, if $f,g, h\colon\bb N\to\bb C$ are arithmetical functions and $f(1) \neq 0$, then:

1. $f* g$ is an arithmetical function,
2. $f*(g*h) = (f*g)* h$,
3. $f*\varepsilon = \varepsilon* f = f$,
4. there exists $f^{-1}$ such that $f^{-1}* f = f* f^{-1} = \varepsilon$,
5. $f*g = g*f$.

The proofs of these properties are very straightforward, the only interesting one is the fourth one. It turns out the inverse $f^{-1}$ is given by the recursive formula $$\label{eq:inverse} f^{-1}(1) = \frac 1{f(1)} \qquad \text{and} \qquad f^{-1}(n) = -\frac{1}{f(1)}\sum_{\substack{d\mid n\\d\neq n}}f\Big(\frac nd\Big) f^{-1}(d),$$ and we can prove this by induction. Clearly for the base case, we want $f^{-1}$ to satisfy $(f*f^{-1})(1) = \varepsilon(1)$, which expands to $f(1)f^{-1}(1) = 1$, and this forces us to take $f^{-1}(1)$ as above.

Now for the inductive step, let $n\geqslant 2$, and suppose $f^{-1}$ is well-defined for all values less than $n$. Then we want to ensure that $(f*f^{-1})(n) = \varepsilon(n) = 0$, i.e., \begin{equation*} \sum_{d\mid n}f\Big(\frac nd\Big)f^{-1}(d)=0, \end{equation*} or, taking out the term where $d=n$, \begin{equation*} f(1)f^{-1}(n) + \sum_{\substack{d\mid n\\d\neq n}}f\Big(\frac nd\Big)f^{-1}(d) = 0. \end{equation*} Since we assumed that $f^{-1}(d)$ is defined for all $d<n$, we can solve this to get a unique value for $f^{-1}(n)$ as defined in \eqref{eq:inverse} above, this establishes the existence and uniqueness of $f^{-1}$. $\tag*{\Box}$

By \eqref{eq:muDivisorSum}, we have that $\mu*\boldsymbol 1=\varepsilon$, so $\mu$ and $\boldsymbol 1$ are Dirichlet inverses of each other, i.e., $\mu^{-1} = \boldsymbol 1$ and $\boldsymbol 1^{-1} = \mu$. This fact, together with the group structure of arithmetic functions under convolutions, gives us a one line proof of the famous result known as the Möbius inversion formula: $$\label{eq:mobiusInversion} g = f * \boldsymbol 1 \iff f = g* \mu.$$ This follows by convolving the equation on the left with $\mu$ on both sides.

## An Application to $\Lambda(n)$ and Cyclotomic Polynomials

We haven’t said much about $\Lambda(n)$ yet. Notice that $$\label{eq:LambdaSum} \sum_{d\mid n} \Lambda(d) = \log n,$$ or in terms of convolutions, $\boldsymbol 1*\Lambda = \log$. To see why this holds, simply factorise $n = \prod_{i=1}^k p_i^{n_i}$ using the fundamental theorem of arithmetic, to get \begin{equation*} \sum_{d\mid n}\Lambda(d) = \sum_{i=1}^k \sum_{\substack{\ell \geqslant 1\\p_i^\ell \mid n}} \log p_i = \sum_{i=1}^k n_i \log p_i = \log n, \end{equation*} as required. Now by Möbius inversion, we get a formula for $\Lambda(n)$ for free: \begin{equation*} \Lambda(n) = - \sum_{d\mid n}\mu(d)\,\log d. \end{equation*} Indeed, \eqref{eq:mobiusInversion} gives us $\bb 1*\Lambda = {\log} \implies \Lambda = {\log} * \mu,$ i.e., \begin{equation*} \Lambda(n) = \sum_{d\mid n}\mu(d)\log\Big(\frac nd\Big) = \log n \sum_{d\mid n}\mu (d) - \sum_{d\mid n}\mu(d)\,\log d, \end{equation*} and by \eqref{eq:muDivisorSum} the first sum is $\varepsilon(n)\log n$ which equals 0 for all $n$.

Now let’s see another application of Möbius inversion, this time to cyclotomic polynomials. The $n$th cyclotomic polynomial $\Phi_n$ is the unique monic irreducible polynomial that divides $x^n-1$ and does not divide $x^k-1$ for any $k<n$. Its roots are precisely the primitive $n$th roots of unity, i.e., $$\Phi_n(x) = \prod_{\substack{k\leqslant n\\\operatorname{hcf}(k,n)=1}}\big(x - e^{2\pi i\,k/n}\big).$$ First, observe that $$\label{eq:cyclotomicFactorisation} x^n - 1 = \prod_{d\mid n}\Phi_d(x).$$ Indeed, the roots of the left-hand side are precisely the $n$th roots of unity. Any root $\zeta$ of the left-hand side must have order $d\mid n$, i.e., $\zeta^d = 1$ for some $d\mid n$ (and $d$ is the smallest $d\geqslant 1$ which does this). Thus $\zeta$ is a root of $\Phi_d$, which appears on the right-hand side. Conversely, any root of the right-hand side clearly satisfies $\zeta^n = 1$. Thus the two polynomials have the same roots, and are therefore equal.

We will now prove the formula $$\label{eq:cyclotomicFormula} \Phi_n(x) = \prod_{ab=n} (x^a-1)^{\mu(b)}.$$ Indeed, taking logs of \eqref{eq:cyclotomicFactorisation}, we see that \begin{equation*} \log(x^n-1) = \sum_{d\mid n}\log(\Phi_d(x)), \end{equation*} and so by Möbius inversion, \begin{align*} \log(\Phi_n(x)) &= \sum_{ab=n} \mu(b)\,\log(x^a-1) = \log\Big(\prod_{ab=n}(x^a-1)^{\mu(b)}\Big). \end{align*} Taking exponentials of both sides yields \eqref{eq:cyclotomicFormula}.

Tags: