The exact order of the convergence rate in the central limit theorem for additive functions (Q1897878)

Let \(h_n\) be a sequence of real-valued additive functions normalized so that \(\sum_{p \leq n} h_n (p)^2 p^{-1} (1 - p^{- 1}) = 1\), and write \(\beta_n = \sum_{p \leq n} p^{-1} |h_n (p)^3 |\). Let \(F_n\) denote the distribution function of the numbers \(h_n (m) - A\), \(m \leq n\) with the usual centering \(A = \sum_{p \leq n} p^{-1} h_n (p)\). It follows from the classical Kubilius theory that \(F_n\) tends to the standard normal distribution \(\Phi\) if \(\beta_n \to 0\). \textit{J. Kubilius} [Litov. Mat. Sb. 19, 115-128 (1979; Zbl 0421.10030)] conjectured that the rate of convergence is \(O (\beta_n)\). Here this problem is completely solved. The main result asserts that \(F_n (u) = \Phi (u) + {D_n \over 2 \sqrt {2 \pi}} ue^{- u^2/2} + O (\beta_n)\), where \(D_n = \sum_{p,q \leq n, pq > n} (pq)^{-1} h_n (p) h_n (q)\). Thus in general we have a second main term and the Kubilius conjecture is valid only under the additional assumption that \(D_n = O (\beta_n)\). It is also shown that \(D_n \ll \beta_n^{2/3}\), and the exponent \(2/3\) here cannot be improved. The author also proposes the more complicated renormalizing condition \(\sum_{p \leq n} h_n (p)^2 p^{-1} (1 - p^{-1}) - D_n = 1\), under which the remainder in the distribution law becomes \(O (\beta_n)\). The proofs employ the classical analytic methods of number theory. The main tool is a novel estimate for the mean value of a multiplicative function \(f\) satisfying \(|f(n) |\leq 1\) which involves the quantity \(\sum_{p,q \leq n, pq > n} (pq)^{-1} (f(p) - 1)(f(q) - 1)\).