\documentclass[11pt]{amsart} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage{parskip} \usepackage[margin=1in]{geometry} \usepackage{amsmath,amssymb,amsthm} \usepackage[expansion=false]{microtype} \usepackage[colorlinks=true,linkcolor=blue,citecolor=blue,urlcolor=blue]{hyperref} \newtheorem{theorem}{Theorem} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \newcommand{\eps}{\varepsilon} \newcommand{\dt}{\,\frac{\mathrm{d}t}{t}} \newcommand{\dx}{\,\frac{\mathrm{d}x}{x}} \newcommand{\pos}[1]{\left(#1\right)^+} \title[Daboussi's elementary proof, reconsidered]{Daboussi's elementary proof of the prime number theorem, reconsidered} \author{Paul Pollack} \address{Department of Mathematics, University of Georgia, Athens GA 30602} \email{pollack@uga.edu} \subjclass[2020]{Primary 11N05; Secondary 11N25} \keywords{Prime number theorem, M\"obius function, smooth numbers, elementary methods} \date{} \begin{document} \begin{abstract} In 1984, Daboussi gave an ``elementary'' proof of the prime number theorem whose central actors are smooth numbers (numbers with no large prime factors). We explain a variant of his argument that might be considered even more elementary. \end{abstract} \maketitle \section{Introduction} Let $\mu$ denote the M\"obius function, and let $M(x):=\sum_{n\le x}\mu(n)$ be the corresponding summatory function. We will give a simple proof of the following 1897 theorem of von Mangoldt \cite{mangoldt}. \begin{theorem}\label{thm:main} $M(x)=o(x)$, as $x\to\infty$. \end{theorem} Landau showed that Theorem \ref{thm:main} is ``elementarily equivalent'' to the prime number theorem (PNT). (His 1899 doctoral thesis proves Theorem \ref{thm:main} from the prime number theorem \cite[pp.~6--16]{landau-thesis}; Landau published the converse implication in 1911 \cite[pp.~973--979]{landau-equivalence}. See also the elegant textbook treatment by Shapiro \cite[\S10.3, pp.~408--414]{shapiro}.) Here ``equivalent'' means that if either statement is assumed, there is a quick proof of the other, significantly shorter than a proof of either statement from scratch. The adverb ``elementarily'' draws a contrast with the more typical complex function-theoretic methods employed in prime number theory: The quick proofs in question involve only real variables, with Riemann's zeta function never (explicitly) mentioned. It was quite a sensation when, in 1948, Erd\H{o}s and Selberg announced that the prime number theorem itself could be established by elementary means \cite{erdos,selberg}. For historical context, see Spencer and Graham's article \cite{spencer-graham}, which reproduces notes by Ernst Straus, and Goldfeld's account \cite{goldfeld}, which also draws on letters from files Selberg made available to him (cf.~the interview \cite{BS08}). Both Erd\H{o}s's and Selberg's arguments use Selberg's ``symmetry formula''; see \cite[eq.~(10.4.10), p.~419]{shapiro}. In 1984, Daboussi \cite{daboussi} gave a very different elementary proof of the PNT, using smooth numbers. Further elementary proofs were later given by Hildebrand \cite{hildebrand}, via a large sieve inequality, and --- quite recently --- Richter \cite{richter}, through a general result on the distribution of $\Omega(n)$, the number of prime factors of $n$ counted with multiplicity. In this note, we present a simplification of Daboussi's argument for Theorem~\ref{thm:main}. Let $P^{+}(n)$ denote the largest prime factor of $n$, adopting the convention that $P^{+}(1):=1$. We say $n$ is $y$-smooth if $P^{+}(n)\le y$. For $x\ge1$ and $y\ge 2$, put \[ \Psi(x,y):=\sum_{\substack{n\le x \\ P^{+}(n)\le y}} 1, \qquad M(x,y):=\sum_{\substack{n\le x \\ P^{+}(n)\le y}} \mu(n).\] Set \[ a:=\limsup_{x\to\infty}\frac{|M(x)|}{x},\] so that the PNT is equivalent to proving that $a=0$. A central component of Daboussi's proof is a comparison between $|M(x,y)|$ and $a\Psi(x,y)$, based on slight variants of the ``logarithmic'' relations \begin{equation}\label{eq:mobius-log0} |M(x,y)|\log x \le \sum_{\substack{p\le y\\p^k\le x}}\log p\,|M(x/p^k,y)| +\int_1^x |M(t,y)|\dt, \end{equation} \begin{equation}\label{eq:psi-log0} \Psi(x,y)\log x =\sum_{\substack{p\le y\\p^k\le x}}\log p\,\Psi(x/p^k,y) +\int_1^x\Psi(t,y)\dt. \end{equation} (We prove both of these in \S\ref{sec:excess} below.) Daboussi extracts what he needs from \eqref{eq:mobius-log0} and \eqref{eq:psi-log0} by integrating against the same carefully constructed kernel function; see \cite[Lemma~2, p.~162, and \S III.3 and \S IV, pp.~163--164]{daboussi}. Our proof avoids any need to introduce a special kernel function by working directly with the inequality for $\bigl(|M(x,y)|-a\Psi(x,y)\bigr)\log x$ that results from subtracting $a$ copies of \eqref{eq:psi-log0} from \eqref{eq:mobius-log0}. In their lovely book, Tenenbaum and Mend\`es France describe a variant of Daboussi's argument that, like ours, avoids Daboussi's kernel function \cite[Chapter~4, \S\S5--7, pp.~92--104]{tmf}. The tradeoff is the need to develop certain ground-floor estimates from the theory of smooth numbers, including that $\Psi(x^u,x)/x^u \approx \rho(u)$ (Dickman's $\rho$-function) for bounded $u$ (with a key step in their proof being that $|M(x^u,x)|/x^u \lessapprox a\rho(u)$ in the same range). Our arrangement requires no nontrivial estimates on $\Psi(x,y)$. Of course, we do need \emph{something}. We will make free use of the following two estimates of Mertens, valid for $x\ge2$: \[ \sum_{p\le x}\frac{\log p}{p}=\log x+O(1),\qquad \prod_{p\le x}\left(1-\frac1p\right)^{-1}\ll\log x. \] For (elementary) proofs, see for example \cite[Chapter~9]{shapiro}. Throughout the paper, the variable $p$ is reserved for primes, and the exponent $k$ on a prime power $p^k$ is always assumed to be a positive integer. \subsection*{Use of LLMs} This is a work of exposition. It is offered as a (hopefully!) human-readable ``digestion'' of output produced by ChatGPT 6 Astra, as prompted by the author. P.\,P. takes no credit for the new mathematics but stands with, by, and beside the final manuscript. Lean formalizations of Theorem~\ref{thm:main} and the smooth-number argument presented here, together with verification records and instructions for reproducing the checks, are available at \url{https://github.com/twonth/daboussi-pnt-lean}. The formalized results use only Lean's standard axioms of propositional extensionality (\texttt{propext}), classical choice (\texttt{Classical.choice}), and quotient soundness (\texttt{Quot.sound}). P.\,P.~thanks OpenAI and Anthropic for providing complimentary access through the ChatGPT for Academic Researchers program and the Claude Team plan for scientists, respectively. \subsection*{Other acknowledgements} P.\,P. presented this argument in a Fall 2026 graduate course offered at the University of Georgia. He thanks the class, and the university, for allowing him this opportunity to workshop the proof. \section{A Fundamental Lemma and the deduction of Theorem \ref{thm:main}}\label{sec:fundamental} The M\"{o}bius function is important in number theory precisely because it provides an elegant, number-theoretically friendly answer to the question of ``who is number $1$'': \[ \mathbf1_{n=1} = \sum_{d\mid n}\mu(d).\] Letting $x\ge 1$ and summing this identity over $n\le x$, \[ 1=\sum_{d\le x}\mu(d)\left\lfloor\frac xd\right\rfloor =x\sum_{d\le x}\frac{\mu(d)}d -\sum_{d\le x}\mu(d)\left\{\frac xd\right\}. \] Consequently, \[ \left|\sum_{d\le x}\frac{\mu(d)}d\right| <\frac{1+\lfloor x\rfloor}{x}\le2\qquad(x\ge1). \] Expressing the same sum by partial summation, we have \[ \sum_{n\le x}\frac{\mu(n)}n =\frac{M(x)}x+\int_1^x\frac{M(t)}t\dt, \] and hence \begin{equation}\label{eq:signed} \left|\int_1^x\frac{M(t)}t\dt\right|<3\qquad(x\ge1). \end{equation} Our proof of Theorem \ref{thm:main} combines \eqref{eq:signed} with the following estimate, whose proof occupies \S\S\ref{sec:smooth}--\ref{sec:excess}. \begin{lemma}[Fundamental Lemma]\label{lem:fundamental} As $x\to\infty$, \[ \int_1^x\pos{a-\frac{|M(t)|}{t}}\dt=o(\log x). \] {\rm (}Here $v^+=\max\{v,0\}$ denotes the positive part of $v$.{\rm )} \end{lemma} Roughly speaking, Lemma \ref{lem:fundamental} says that $|M(t)|/t$ spends most of its time close to its limit superior $a$, with time measured logarithmically (i.e., with respect to the logarithmic measure $\mathrm{d}t/t$). \begin{proof}[Proof of Theorem~\ref{thm:main}, assuming the Fundamental Lemma] Suppose for a contradiction that $a>0$, and fix $B:=\exp(12/a)$. We first use \eqref{eq:signed} to show that every interval $[X,BX]$, with $X\ge1$, contains a $u$ for which \begin{equation}\label{eq:small} |M(u)|\le\frac a2u. \end{equation} Indeed, if there is no such $u$, then $M(t)$ is nonzero throughout $[X,BX]$. Since $M$ takes integer values and jumps by at most $1$, it has constant sign there. This gives \[ \left|\int_X^{BX}\frac{M(t)}t\dt\right| =\int_X^{BX}\frac{|M(t)|}t\dt \ge\frac a2\log B=6. \] But \eqref{eq:signed}, applied at $BX$ and at $X$, bounds the left-hand side strictly below $6$. Now let $X\ge X_0:=8/a$, and choose $u\in[X,BX]$ satisfying \eqref{eq:small}. To obtain a lower bound for the integral in the Fundamental Lemma, we extend this estimate to nearby values of $t$. For $u\le t\le u(1+a/8)$, the bound $|\mu(n)|\le1$ gives \[ \begin{aligned} |M(t)| &\le |M(u)|+\bigg|\sum_{u0. \] We apply this bound with $X=(2B)^{j} X_0$, for $j=0,\ldots, N-1$, noting that the corresponding intervals $(X,2BX]$ are disjoint and contained in $(1,(2B)^{N} X_0]$. Adding these estimates gives \[ \int_1^{(2B)^NX_0}\pos{a-\frac{|M(t)|}t}\dt \ge \left(\frac a4\log(1+a/8)\right)N. \] But the Fundamental Lemma implies that our left-hand side is $o(\log((2B)^NX_0))=o(N)$ as $N\to\infty$. This contradiction forces $a=0$. \end{proof} \section{Splitting off bounded prime factors}\label{sec:smooth} Recall from the introduction that $\Psi(x,y)=\sum_{n\le x,\,P^{+}(n)\le y} 1$ while $M(x,y)=\sum_{n\le x,\,P^{+}(n)\le y} \mu(n)$. Fix $y\ge 2$, and put \[ P=\prod_{p\le y}p. \] We have $|M(t,y)|\le\Psi(t,y)$ and, since every positive integer at most $y$ is $y$-smooth, \[ M(t,y)=M(t),\qquad \Psi(t,y)=\lfloor t\rfloor\qquad(1\le t\le y). \] At the other end of the range, only the squarefree $y$-smooth integers contribute to $M(t,y)$, and these are precisely the divisors of $P$. Thus, for $t\ge P$, \begin{equation}\label{eq:support} M(t,y)=\sum_{d\mid P}\mu(d)=0. \end{equation} Following Daboussi, we compare $|M(x,y)|$ with $a\Psi(x,y)$. The following estimate gets us started. \begin{proposition}\label{prop:comparison} For every $y\ge2$, \begin{equation}\label{eq:comparison} \int_1^\infty \left(\frac{|M(t,y)|}t-a\frac{\Psi(t,y)}t\right)\dt\ge0. \end{equation} \end{proposition} \begin{proof} Each positive integer admits a unique decomposition in the form $n=bc$, where $P^+(b)\le y$ and $(c,P)=1$. Since $(b,c)=1$, the multiplicativity of $\mu$ gives \[ M(x)=\sum_{\substack{c\le x\\(c,P)=1}}\mu(c)M(x/c,y). \] The terms with $c\le x/P$ vanish by \eqref{eq:support}. Taking absolute values in the remaining terms and grouping according to the value of $j:=\lfloor x/c\rfloor0$, choose $T\ge1$ so that $|M(t)|/t\le a+\eps$ for $t\ge T$. Then, for $T\le t\le y$, \[ E_y(t) =\pos{\frac{|M(t)|}t-a+a\frac{\{t\}}t} \le\eps+\frac at. \] Integrating from $1$ to $y$, using $E_y(t)\le1$ below $T$, we get that \[ \int_1^yE_y(t)\dt \le\log T+\eps\log(y/T)+a \] for all $y\ge T$. Hence, \[ \limsup_{y\to\infty} \frac{1}{\log{y}}\int_1^yE_y(t)\dt \le \eps.\] As $\eps > 0$ is arbitrary, \begin{equation}\label{eq:below} \int_1^yE_y(t)\dt=o(\log y). \end{equation} To handle the remaining range $(y,\infty)$, we need the ``logarithmic'' relations quoted in the introduction. We begin their derivation with some simple consequences of M\"{o}bius inversion. Writing $\Lambda$ for the von Mangoldt function and starting from \begin{equation}\label{eq:chebyshevlog} \log{n} = \sum_{p^k \mid n} \log{p} = \sum_{d \mid n} \Lambda(d), \end{equation} we find that \begin{align*} \Lambda(n) &= \sum_{d\mid n} \mu(d)\log{(n/d)} \\&= \log{n} \sum_{d\mid n} \mu(d) - \sum_{d \mid n}\mu(d)\log{d}\\ &=- \sum_{d \mid n}\mu(d)\log{d}. \end{align*} Inverting this last formula for $\Lambda(n)$ gives \begin{equation}\label{eq:muchebyshev} -\mu(n)\log n=\sum_{p^k\mid n}\log p\,\mu(n/p^k). \end{equation} Summing \eqref{eq:muchebyshev} over $y$-smooth integers $n\le x$, and reversing the order of summation, \[ \begin{aligned} -\sum_{\substack{n\le x\\P^+(n)\le y}}\mu(n)\log n &=\sum_{\substack{p\le y\\p^k\le x}}\log p \sum_{\substack{m\le x/p^k\\P^+(m)\le y}}\mu(m)\\ &=\sum_{\substack{p\le y\\p^k\le x}}\log p\,M(x/p^k,y). \end{aligned} \] On the other hand, partial summation gives \[ \sum_{\substack{n\le x\\P^+(n)\le y}}\mu(n)\log n =M(x,y)\log x-\int_1^xM(t,y)\dt. \] Substituting this into the immediately preceding display and rearranging, \begin{equation}\label{eq:needabs} M(x,y)\log x =-\sum_{\substack{p\le y\\p^k\le x}}\log p\,M(x/p^k,y) +\int_1^xM(t,y)\dt. \end{equation} Carrying out the same calculation with \eqref{eq:chebyshevlog} replacing \eqref{eq:muchebyshev} gives \begin{equation}\label{eq:psi-log} \Psi(x,y)\log x =\sum_{\substack{p\le y\\p^k\le x}}\log p\,\Psi(x/p^k,y) +\int_1^x\Psi(t,y)\dt, \end{equation} which is precisely the identity \eqref{eq:psi-log0} from the introduction. Taking absolute values in \eqref{eq:needabs} and appealing to the triangle inequality, \begin{equation}\label{eq:mobius-log2} |M(x,y)|\log x \le \sum_{\substack{p\le y\\p^k\le x}}\log p\,|M(x/p^k,y)| +\int_1^x |M(t,y)|\dt, \end{equation} and so we also have \eqref{eq:mobius-log0}. Subtracting $a$ times \eqref{eq:psi-log} from \eqref{eq:mobius-log2}, we obtain \begin{align*} \bigl(|M(x,y)|-a\Psi(x,y)\bigr)\log x &\le\sum_{\substack{p\le y\\p^k\le x}}\log p \bigl(|M(x/p^k,y)|-a\Psi(x/p^k,y)\bigr)\\ &\qquad+\int_1^x\bigl(|M(t,y)|-a\Psi(t,y)\bigr)\dt. \end{align*} Replacing each difference on the right by its positive part only increases the right-hand side and so preserves the inequality. Since the resulting right-hand side is nonnegative, we may also replace the left-hand side by its positive part. Noting that $(|M(t,y)|-a\Psi(t,y))^+=tE_y(t)$, we deduce after dividing by $x$ that \begin{equation}\label{eq:excess-inequality} E_y(x)\log x \le\sum_{\substack{p\le y\\p^k\le x}} \frac{\log p}{p^k}E_y(x/p^k) +\int_1^x\frac tx E_y(t)\dt. \end{equation} We integrate \eqref{eq:excess-inequality} with respect to $\mathrm{d}x/x$, over $[1,\infty)$. For each prime power $p^k$, substituting $x=p^kt$ gives \[ \int_{p^k}^\infty\frac{\log p}{p^k}E_y(x/p^k)\dx =\frac{\log p}{p^k}\int_1^\infty E_y(t)\dt. \] For the remaining term, we reverse the order of integration: \[ \begin{aligned} \int_1^\infty\left(\int_1^x\frac tx E_y(t)\dt\right)\dx &=\int_1^\infty E_y(t) \left(\int_t^\infty\frac tx\dx\right)\dt\\ &=\int_1^\infty E_y(t)\dt. \end{aligned} \] (Here all manipulations of sums and integrals are justified by nonnegativity.) Combining the two contributions, \[ \int_1^\infty E_y(t)\log t\dt \le\left(1+\sum_{p\le y}\sum_{k\ge1}\frac{\log p}{p^k}\right) \int_1^\infty E_y(t)\dt. \] The coefficient on the right is estimated using Mertens' theorem: \[ \begin{aligned} \sum_{p\le y}\sum_{k\ge1}\frac{\log p}{p^k} &=\sum_{p\le y}\frac{\log p}{p} +\sum_{p\le y}\frac{\log p}{p(p-1)} =\log y+O(1). \end{aligned} \] Hence, for some absolute constant $C>0$, \[ \int_1^\infty E_y(t)\log t\dt \le(\log y+C)\int_1^\infty E_y(t)\dt. \] Rearranging and applying \eqref{eq:excess-bigO}, we find that \begin{equation}\label{eq:centered} \int_1^\infty E_y(t)\log(t/y)\dt \le C\int_1^\infty E_y(t)\dt=O(\log y). \end{equation} With the factor $\log(t/y)$ omitted, we already had an $O(\log{y})$ bound; the advantage here is that the new factor is sensitive to \emph{where} the excess occurs. Let's mop up. Recall that $\int_1^{y} E_y(t)\, \mathrm{d}t/t = o(\log{y})$. Slightly larger values of $t$, say $y\le t \le y^{1+\eps}$ with $\eps > 0$ fixed, pose no trouble. Since $E_y(t)\le1$, \[ \int_y^{y^{1+\eps}}E_y(t)\dt \le\int_y^{y^{1+\eps}}\dt =\eps\log y, \] which will be acceptable to us upon taking $\eps$ small. It remains to consider $t$ that are substantially larger than $y$, meaning $t\ge y^{1+\eps}$. Moving the contribution from $[1,y]$ in \eqref{eq:centered} to the right and using \eqref{eq:below}, we obtain \begin{align} \int_y^\infty E_y(t)\log(t/y)\dt &\le\int_1^y E_y(t)\log(y/t)\dt+O(\log y)\notag\\ &\le(\log y)\int_1^y E_y(t)\dt+O(\log y)\notag\\ &=o((\log y)^2).\label{eq:weighted-tail} \end{align} On the other hand, $\log(t/y)\ge\eps\log y$ when $t\ge y^{1+\eps}$, so \[ \int_y^\infty E_y(t)\log(t/y)\dt \ge\int_{y^{1+\eps}}^\infty E_y(t)\log(t/y)\dt \ge\eps\log y\int_{y^{1+\eps}}^\infty E_y(t)\dt. \] Comparing with \eqref{eq:weighted-tail}, we find that \[ \int_{y^{1+\eps}}^\infty E_y(t)\dt=o(\log y). \] Combining the estimates for the three ranges gives $\int_1^\infty E_y(t)\,\mathrm{d}t/t \le\eps\log y+o(\log y)$. Dividing by $\log y$ and letting $y\to\infty$, we obtain \[ \limsup_{y\to\infty}\frac1{\log y} \int_1^\infty E_y(t)\dt\le\eps. \] But $\eps>0$ is arbitrary, and so $\int_1^\infty E_y(t) \,\mathrm{d}t/t=o(\log y)$, as desired. This proves \eqref{eq:excess-goal} and hence the Fundamental Lemma. As shown in \S\ref{sec:fundamental}, it follows that $M(x)=o(x)$. \end{proof} \begin{thebibliography}{99} \bibitem{BS08} N.\,A. Baas and C.\,F. Skau, \emph{The lord of the numbers, Atle Selberg. On his life and mathematics}, Bull. Amer. Math. Soc. (N.S.) \textbf{45} (2008), no.~4, 617--649. An extended version of the interview is available at \url{https://www.math.ntnu.no/Selberg-interview/}. \bibitem{daboussi} H. Daboussi, \emph{Sur le th\'eor\`eme des nombres premiers}, C. R. Acad. Sci. Paris S\'er. I Math. \textbf{298} (1984), no.~8, 161--164. \bibitem{erdos} P. Erd\H{o}s, \emph{On a new method in elementary number theory which leads to an elementary proof of the prime number theorem}, Proc. Nat. Acad. Sci. U.S.A. \textbf{35} (1949), 374--384. \bibitem{goldfeld} D. Goldfeld, \emph{The elementary proof of the prime number theorem: an historical perspective}, in \emph{Number Theory} (New York, 2003), D. Chudnovsky, G. Chudnovsky, and M. Nathanson (eds.), Springer, New York, 2004, 179--192. \bibitem{hildebrand} A. Hildebrand, \emph{The prime number theorem via the large sieve}, Mathematika \textbf{33} (1986), 23--30. \bibitem{landau-thesis} E. Landau, \emph{Neuer Beweis der Gleichung $\sum_{k=1}^{\infty}\mu(k)/k=0$}, Inauguraldissertation, Berlin, 1899. \bibitem{landau-equivalence} \bysame, \emph{\"Uber die \"Aquivalenz zweier Haupts\"atze der analytischen Zahlentheorie}, Sitzungsber. Kaiserl. Akad. Wiss. Wien, Math.-Naturwiss. Kl., Abt.~IIa \textbf{120} (1911), 973--988. \bibitem{mangoldt} H. von Mangoldt, \emph{Beweis der Gleichung $\sum_{k=1}^{\infty}\mu(k)/k=0$}, Sitzungsber. K\"onigl. Preu\ss. Akad. Wiss. Berlin (1897), 835--852. \bibitem{richter} F.\,K. Richter, \emph{A new elementary proof of the Prime Number Theorem}, Bull. London Math. Soc. \textbf{53} (2021), 1365--1375. \bibitem{selberg} A. Selberg, \emph{An elementary proof of the prime-number theorem}, Ann. of Math. (2) \textbf{50} (1949), 305--313. \bibitem{shapiro} H.\,N. Shapiro, \emph{Introduction to the Theory of Numbers}, John Wiley \& Sons, New York, 1983; reprinted by Dover Publications, Mineola, NY, 2008. \bibitem{spencer-graham} J. Spencer and R. Graham, \emph{The elementary proof of the prime number theorem}, Math. Intelligencer \textbf{31} (2009), no.~3, 18--23. \bibitem{tmf} G. Tenenbaum and M. Mend\`es France, \emph{The Prime Numbers and Their Distribution}, translated by P. G. Spain, Student Mathematical Library, vol.~6, American Mathematical Society, Providence, RI, 2000. \end{thebibliography} \end{document}