Overcoming the Curse of Dimensionality: Convergence of Diffusion Models
skoularidou · x · 2026-08-10
This paper investigates Denoising Diffusion Probabilistic Models (DDPM) under the manifold hypothesis. The authors prove that the convergence rates for score learning and sampling complexity (w.r.t. Wasserstein distance) are independent of the ambient dimension, developing a new framework that connects diffusion models to the theory of Gaussian Process extrema.
More from Research
- How Frontier Models Train on Outcomes: A History of RL Post-Training — SergioPaniego · 2026-08-10
- Tsinghua Reveals 'Referential Dangling' Flaw in Hard Prompt Compression — Tsinghua · 2026-08-10
- Explainer: Masked Diffusion as a Discrete Analog to Continuous Diffusion — alec_helbling · 2026-08-10
- Nathan Lambert Releases RLHF Textbook on LLM Post-Training — Interconnects (Nathan Lambert) · 2026-08-10
- Interpretability Researcher Neel Nanda Shares Guide for LLM Research Beginners — NeelNanda5 · 2026-08-10
- Replacing MiniMax H3's 32B Text Encoder with 4B/8B Models: An Update — Fit_Ad7343 · 2026-08-10