Internal-state criticality in Bayesian-inverse-Bayesian inference

Abstract

We propose Bayesian-inverse-Bayesian (BIB) inference in repeated games as a minimal generative model linking Bayesian inference, statistical mechanics, and heavy-tailed statistics. As a concrete instantiation we simulate repeated $N$-hand cyclic-dominance rock-paper-scissors, a discrete setting in which Nash-targeting algorithms collapse to uniform random play, so that any non-trivial dynamics must originate internally. Across a multi-axis sweep of design, window, and opponent conditions, the BIB dynamics remain in the same internal critical state, the argmax-persistence distribution staying a heavy-tailed power law with exponent $α\approx 1.43$ at the canonical window. Along the window and alphabet axes the exponent is not constant but drifts toward the universal $3/2$ as the finite-sample residual $(N-1)/(2m)$ vanishes. Bayes-only inference, which lacks the inverse step, shows no analogous universality and no power law. Because the argmax and laminar observables are first-passage reads of one driftless log-posterior walk, what is robust across conditions is the critical, zero-drift state itself, evidenced by the cross-design data collapse rather than by any particular exponent value. The state is also invariant across the hypothesis count $N_h$, with the cutoff time and posterior spread obeying finite-size scaling. Adding an inverse-Bayesian relaxation step (hypothesis renewal) to ordinary Bayesian inference is by itself enough to render the dynamics critical, with no external parameter adjustment. Rather than self-organizing toward an absorbing state, BIB reaches criticality by continually reconstructing the hypothesis-space boundary, a mechanism complementary to self-organized criticality that makes the criticality robust across a natural parameter range.