Machine-Guided Recurrence Boundary Theory for Nahm Sums

Abstract

We develop a proof-carrying recurrence--boundary method for affine families of positive-definite Nahm sums. A universal coordinate-contiguous relation provides exact algebraic cells for finite certificates; a tropical face-limit theorem identifies parameter directions along which a higher-rank Nahm sum degenerates to a lower-rank theta or theta--hypergeometric boundary value; and a recurrence--boundary principle recovers the initial sums from sufficiently many independent boundary limits. Machine learning and reinforcement learning are used only for discovery. Evolutionary symbolic search proposes recurrences and asymptotic rays from exact algebraic data, while a $Q$-learning agent searches for short sequences of legal contiguous-cell identities. No learned output is accepted as proof: every successful candidate is replaced by an exact symbolic certificate. As the main application, we prove Shi and Wang's Conjecture~3.8 (arXiv:2607.23257) for the Nahm sums dual to Zagier's twelfth rank-three example. The search finds a second-order recurrence for a one-parameter family and a five-cell certificate for it, together with two asymptotic rays leading to binary and unary theta series. These give two linear equations for the two initial sums. Solving the resulting $2\times2$ system reduces the conjecture to two generalized-eta identities, certified by valence arguments on $\Gamma_1(300)$ and $\Gamma_1(100)$. Consequently the dual of Zagier's twelfth example is modular, and every member of the affine family is an explicit $\mathbb{Z}[q,q^{-1}]$-combination of the two base products. Complete verification artifacts accompany the paper.