🤖 AI Summary
This work investigates macroscopic statistical regularities in integer multiplication dynamics and their relationship with universal coding length. We propose the Multiplicative Turing Ensemble (MTE), a stochastic dynamical system over integers driven by i.i.d. prime multipliers. Introducing Elias Omega coding length as an energy function—novel in statistical modeling—we derive Gibbs priors over integers and primes via maximum entropy and variational calculus. Theoretically, we identify a qualitative phase transition in the statistical tail behavior distinguishing machine-adaptive generation from human-generated complexity, explaining Pareto-type heavy tails and the absence of well-defined first moments. Empirical validation on Debian and PyPI package repositories confirms that the scaled Omega prior achieves minimal KL divergence, supporting a structural bifurcation in complexity distributions. Our core contribution is a coding-length-driven statistical framework for multiplicative dynamics, yielding testable statistical criteria to discriminate machine- versus human-generated patterns.
📝 Abstract
We study integer-valued multiplicative dynamics driven by i.i.d. prime multipliers and connect their macroscopic statistics to universal codelengths. We introduce the Multiplicative Turing Ensemble (MTE) and show how it arises naturally - though not uniquely - from ensembles of probabilistic Turing machines. Our modeling principle is variational: taking Elias' Omega codelength as an energy and imposing maximum entropy constraints yields a canonical Gibbs prior on integers and, by restriction, on primes. Under mild tail assumptions, this prior induces exponential tails for log-multipliers (up to slowly varying corrections), which in turn generate Pareto tails for additive gaps. We also prove time-average laws for the Omega codelength along MTE trajectories. Empirically, on Debian and PyPI package size datasets, a scaled Omega prior achieves the lowest KL divergence against codelength histograms. Taken together, the theory-data comparison suggests a qualitative split: machine-adapted regimes (Gibbs-aligned, finite first moment) exhibit clean averaging behavior, whereas human-generated complexity appears to sit beyond this regime, with tails heavy enough to produce an unbounded first moment, and therefore no averaging of the same kind.