Sharp noise transitions for an adaptive single-copy stabilizer-learning process

Abstract

We determine the noise-dependent behavior of the adaptive single-copy stabilizer-learning update of Bittel, Eisert, Gong, Mele and Schatzki on globally depolarized stabilizer states. For every fixed prime local dimension p, the number of diagonal stabilizers still missing follows an exact birth–death chain. Writing c for the retained pure-state weight, starting with all n diagonal stabilizers missing, the expected time to record the optimizing stabilizer is linear in n when c²>p/(p+1), quadratic at equality, and exponential when c²<p/(p+1). We obtain the exact supercritical speed, subcritical exponential prefactors, and a critical-window mean profile uniform over all initial missing ranks, with an O(n) remainder and the exact leading recording penalty. An exact regeneration identity distinguishes first exposure of the computational-basis target from its actual recording. A finite-time bound shows that below threshold the unchanged algorithm has exponentially small success probability in any polynomial round budget, even on this simple noise family. The obstruction concerns this update and candidate generation rule.