An elementary proof of the prime number theorem with a remainder term (Q2539253)

An elementary proof of the prime number theorem is given in the form \[ |\psi(x) -x| \leq x \exp\{-(\log x)^\alpha (\log \log x)^\beta\} \tag{*} \] for \(x\geq \exp \exp 100\) and \(\alpha=1/7\) and \(\beta= -2\). The method is elementary in the sense that it uses neither Laplace transforms nor properties of the Riemann zeta function. The proof depends on the following generalization of Selberg's formula (cf. \textit{E. Bombieri} [Riv. Mat. Univ. Parma, II. Ser. 3, 393--440 (1962; Zbl 0121.28307)]): There is an absolute constant \(A\) such that for each positive integer \(n\) and all \(x\geq 1\) \[ \begin{multlined} \sum_{j\leq x} (\log j)^{2n-1} \Lambda(j) +B(n,n)^{-1} \sum_{jk\leq x} (\log j)^{n-1} \Lambda(j) (\log k)^{n-1} \Lambda(k) = \\= -2\int_1^x (\log t)^{2n-}\, dt+ \theta x\{An^4\log x\}^n,\end{multlined} \] where \(B(n,n)=\Gamma(n)^2/\Gamma(2n)\) and \(| \theta|\leq 1\). Let \(\mathcal F_n\) denote the \(n\)th such formula. \((\mathcal F_1\) is weaker than Selberg's formula, because the method of proof wastes one factor of \(\log x)\). These formulas are based on an algebraic identity suggested by P. J. Cohen. It is most easily stated for \(2n\) times differentiable functions \(F: \mathbb R\to\mathbb R\) which do not vanish on an interval: For each positive integer n there exists a polynomial \(P_n(Y_0,Y_1,\ldots,Y_{2n})\), independent of \(F\), such that \[ \frac 1{\Gamma(2n)}\left(\frac{F'}{F}\right)^{(2n-1)}+\left\{\frac 1{\Gamma(n)}\left(\frac{F'}{F}\right)^{(n-1)}\right\}^2= P_n(F, F',\ldots,F^{(2n)}/F^n. \] The essential point is that the denominator on the right hand side contains only \(n\) factors of \(F\). The formulas \(\mathcal F_n\) are established from the algebraic identity by means of many estimates of the character of Dirichlet's hyperbola method for the divisor problem. To determine the constants occurring in the main terms of \(\mathcal F_n\), \textit{E. Wirsing}'s estimate of the remainder term in the prime number theorem [J. Reine Angew. Math. 214/215, 1--18 (1964; Zbl 0166.31004)], which is elementary in the above mentioned sense, is cited. The prime number theorem is extracted from the formulas \(\mathcal F_n\), by a Tauberian argument modeled upon that of Wirsing (loc. cit.). For the while let \(n\) be fixed, and define a function \(R\), depending on \(n\), on \([1,\infty)\) by \[ R(X) = \sum_{j\leq x} (\log j)^{n-1} \{1-\Lambda(j)\}. \] . The formula \(\mathcal F_n\) is rewritten in terms of \(R\), and \(R\) is shown to be slowly oscillating. An absolutely continuous function \(R_0\) is found which approximates \(R\) closely and satisfies the inequality \(| R'_0(x)| \leq 1.01 (\log x)^{n-1}\). Functions \(R_1,R_2,\ldots\) are defined recursively by the relations \[ \int_1^x (\log t)^n\, dR_{\nu+1}(t) = B(n,n)^{-1} \int_1^x R_\nu(x/t)\,dR_\nu(t) \] and \(R_{\nu+1}(1)=0\), \(\nu=0,1,2,\ldots\). On the one hand it is shown that \(R_\nu(s)\) is reasonably close to \(R(x)\). On the other hand, by a generalization of Wirsing's inequality (loc. cit.), it is shown that \(| R'_\nu(x)| (\log x)^{1-n}\) becomes small as \(\nu\) and \(x\) get large. The two facts imply that \(| R(x)|\) is small compared with \(x(\log x)^{1-n}\). Finally, for a given \(x\) an optimal choice is made for \(n\) to give (*).