Id: arxiv/2608.25055 Url: https://arxiv.org/abs/2608.25055 Tags: math.AP Score: 0 Title: A Mean-Field Theory of Transformers: Well-Posedness of the Coupled Data--Parameter Dynamics and Global Convergence of Training Domain: math.AP Source: arxiv Authors: Michael Herty, Hailiang Liu Reviews: 0 Subject: math.AP Category: research Comments: 0 Disputed: false Division: math Mentions: 0 Discussed: 0 Published: 2026-08-25 Page: https://delphi.science/r/arxiv/2608.25055 We develop a rigorous mean-field theory for transformer networks that captures two large-scale limits inherent in the architecture: the number of tokens $N\to\infty$ in the input sequence and the number of attention heads $H\to\infty$ in each layer. It further considers an infinite number of layers leading to a time-continuous formulation. The resulting framework couples two interacting mean-field objects: a token distribution $μ_t\in\mathcal P(\mathbb R^d)$, which evolves through network depth $t\in[0,T]$ according to a McKean--Vlasov transport equation, and the attention-parameter…