Closed‑form analysis of diffusion learning dynamics uncovers an inverse‑variance law of distributional convergence; MLP‑UNets comply, while convolutional U‑Nets break it.