Large prime gaps and probabilistic models (Q6113258)

In this article, the authors introduce a new probabilistic model of the primes. The authors provide a comparison of the new model with Cramér's and Granville's models of the primes. The novelty in this model is how the random sets are formed by sieving random residue classes modulo primes up to a certain threshold. Under this model, certain heuristic bounds are obtained for the size of the largest gap between primes in the interval \([1, x]\). Consider the random set \[ \mathcal{S}_z:=\mathbb{Z}\setminus \bigcup_{p\leq z} (a_p\bmod{p}) \] where \(a_p\bmod{p}\) are jointly independent random variables for each prime \(p\). In this new model of primes, the random set is defined as \[ \mathcal{R}=\{n\geq e^2:\, n\in\mathcal{S}_{z(n)}\}, \] here \(z(n)\) is chosen to match the proportion of primes as closely as possible, in particular, \(z(n)\) is chosen to be the largest prime with \(\prod_{p\leq z(n)}\left(1-\frac1{p}\right)^{-1}\leq \log n.\) Under this model, the authors prove upper and lower bounds on the largest gap between elements of \(\mathcal{R}\). Let \( \mathcal{G}_{\mathcal{R}}(x)=\max\{b-a:\,[a,b]\subset [1,x]\,\,\text{and}\,\,[a,b]\cap\mathcal{R}=\emptyset\}\). They show that for any \(\epsilon>0\), with probability one, \[ (2e^{-\gamma}-\epsilon)\log^2 x\leq\mathcal{G}_{\mathcal{R}}(x)\leq (2e^{-\gamma}+\epsilon)\frac{\log^2 x\log\log x}{\log\log\log x} \] where \(\gamma\) denotes Euler-Mascheroni constant. The analysis here leads to the conjecture that the largest gap between primes is asymptotic to \(2e^{-\gamma}\log^2 x\), which matches the lower bound suggested by Granville's model. Furthermore, they prove that an analogue of Hardy Littlewood conjecture holds almost surely for the random set \(\mathcal{R}\). More precisely, they show that for a fixed \(c\in[1/2,1)\) and \(\epsilon>0\), almost surely \[ |\{n\leq x:\,n+h\in\mathcal{R} \,\,\text{for all \(h\in \mathcal{H}\)}\}|=\mathfrak{S}(\mathcal{H})\int_{2}^{x}\frac{dt}{\log^{|\mathcal{H}| }t}+O\left(x^{1-\frac{1-c}{8c^2-2c}}+\epsilon\right) \] uniformly for all admissible tuples \(\mathcal{H}\) with \(\mathcal{H}\subset[0, \exp\left(\log^{1-c}x/\log\log x\right) ]\) and \(|\mathcal{H}|\leq \log^{c}x\). In the particular case when \(\mathcal{H}=\{0\},\) they prove a more precise statement that for \(A>3/2\), almost surely \[ |\{n\leq x:\,n\in\mathcal{R}\}|=\int_{2}^{x}\frac{dt}{\log t}+O\left(x^{1/2}\log^Ax\right) \] holds, recovering the ``Riemann hypothesis'' under this model. The authors also show that if \(\mathcal{A}\) is any set of integers on which a form of Hardy Littlewood conjecture holds with specific error term, then a lower bound for the largest gap between the elements of \(\mathcal{A}\) could be obtained in terms of this asymptotic formula.