On the bias of a probability law for brittle integers (Q6072957): Difference between revisions

The author studies the set~\(S(x, y)\) of those integers in~\([1, x]\) whose prime factors are all at most~\(y\). It is known that \[ |S(x, y)| = x^{\alpha + o(1)} \] where~\(\alpha = \alpha(x, y) \in [0, 1]\) is an implicitly defined quantity known as the saddle-point. Define a random variable~\(N\) on~\(\mathbb{N}\) by \[ \mathbb{P}(N = n) = c_\alpha \frac{\mathbf{1}[p\mid n\implies p\leq y]}{n^{\alpha(x, y)}}, \] where~\(c_\alpha = \prod_{p\leq y}(1-p^{-\alpha})\) is a normalizing factor. In view of the above, it is a natural question to wonder to what extent the random variable~\(N\) is concentrated on the integers in~\([1, x]\). The paper under review investigates this question. The author estimates in a wide range the ratio \[ P(x, y, z) := \mathbb{P}(N \leq z). \] In particular, he proves that there is a sharp transition about~\(z\sim x\), and more precisely \[ P(x, y, x y^{\Theta h}) \approx \int_{-\infty}^h \frac{e^{-v^2} d v}{\sqrt{2\pi}} \] where \(\Theta = \Theta_{x,y}\) satisfies \[ \Theta \sim \Big(\frac{\log x}{\log y}\Big(1 + \frac{\log x}{y}\Big)\Big)^{1/2}, \qquad (\min(y, \tfrac{\log x}{\log y})\to \infty). \] The author uses this result to study the bias in the ``semi-asymptotic'' estimate \[ \Psi(x/d, y) \sim d^{-\alpha} \Psi(x, y). \] Define \[ \Delta(x, y) := \frac{\sum_{d\in S(x, y)} \Psi(x/d, y)}{\sum_{d\in S(x, y)} d^{-\alpha} \Psi(x, y)}. \] The author establishes: \begin{itemize} \item an asymptotic estimate for~\(\Delta(x, y)\) in the domain~\(\exp (\log\log x)^{5/3+\varepsilon} \leq y \leq x\), whose main term is expressed in terms of quantities related to the Dickman function, \item an asymptotic estimate for~\(\log \Delta(x, y)\) in the whole range~\(2\leq y \leq x\). \end{itemize} As a conclusion, it transpires that~\(\Delta(x, y)\) tends quite rapidly to~\(0\) unless \(\log y\) is close to~\(\log x\). The proofs involve various asymptotic estimates known on~\(\Psi(x, y)\) in terms of the De Bruijn function; the saddle-point estimates of Hildebrand-Tenenbaum; and the semi-asymptotic estimates of La Bretèche-Tenenbaum.