Partition-theoretic model of prime distribution

Abstract: We make an application of ideas from partition theory to a problem in multiplicative number theory. We propose a deterministic model of prime number distribution, from first principles related to properties of integer partitions, that predicts the prime number theorem and suggests new conjectures regarding predictable variations in prime gaps. The model posits that, for $n\geq 2$, $$p_{n}\ =\ 1\ +\ 2\sum_{j=1}^{n-1}\left\lceil \frac{d(j)}{2}\right\rceil\ +\ \varepsilon(n),$$ where $p_k$ is the $k$th prime number, $d(k)$ is the divisor function, and $\varepsilon(k)\geq 0$ is an error that is negligible asymptotically; both the main term and error are enumerative functions in our conceptual model. We refine the error to give numerical estimates of $\pi(n)$ similar to those provided by the logarithmic integral, and much more accurate than $\operatorname{li}(n)$ up to $n=10{,}000$ where the estimate is almost exact. We also test a seemingly dubious prediction suggested by the model, that prime-indexed primes $p_2, p_3, p_5, p_7,$ etc., are more likely to be involved in twin prime pairs than are arbitrary primes; surprisingly, computations show prime-indexed primes are more likely to be twin primes up to at least $n=10{,}000{,}000{,}000$, with the bias diminishing but persistent as $n$ increases.