Research map / Image generation / Fast diffusion sampling
Conditional Diffusion Models are Minimax-Optimal and Manifold-Adaptive for Conditional Distribution Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Theoretical analysis of classifier-free conditional forward-backward (score-based) diffusion models for conditional distribution estimation, proving minimax-optimal TV rates in Euclidean density regression and manifold-adaptive Wasserstein rates depending only on intrinsic dimensions.
Problem
Statistical convergence guarantees for conditional generative models (conditional distribution estimation / distribution regression) using diffusion models, especially whether they achieve minimax rates and adapt to low-dimensional manifold structure in covariates and responses.
Approach
Models the conditional score with a piecewise time-dependent ReLU network class trained by conditional score matching on a classifier-free conditional OU forward-backward diffusion, and derives finite-sample bounds: minimax-optimal TV rate for smooth densities and W1 rates depending only on intrinsic dimensions (d_X, d_Y).
Key results
- Under smooth-density assumptions, the expected TV error of the conditional diffusion estimator is O~(n^{-1/(2+D_X/alpha_X+D_Y/alpha_Y)}), claimed minimax-optimal in the regime alpha_X in [0,1].
- W1 error bound depends only on intrinsic dimensions d_X and d_Y under manifold assumptions, with rate O~(n^{-1/(2+d_X/alpha_X)} v n^{-(1+1/alpha_Y)/(2+d_X/alpha_X+d_Y/alpha_Y)}).
- Specializing to D_X=0 recovers minimax unconditional distribution estimation rates in TV and W1; alpha_Y to infinity recovers the L2 nonparametric regression rate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the conditional forward-backward diffusion estimator for Euclidean density regression with α_X ∈ [0,1] | Conditional distribution estimation · Euclidean density regression | total variation | Helped |
| Use the conditional forward-backward diffusion estimator when covariates and responses have low-dimensional manifold structure | Conditional distribution estimation · Distribution regression on manifolds | 1-Wasserstein | Helped |
Surprising findings
- TV distance is always 1 between the estimator and target when both are mutually singular manifold-supported measures, so only W1 is meaningful in the manifold setting.
Limitations
- Results are rates/upper bounds with high-probability statements; no empirical experiments on real data in main text.
- Requires lower bound on conditional density and smoothness assumptions (C^{alpha_Y,alpha_X}, manifold smoothness).
Experimental setup
Metrics: total variation distance 1-Wasserstein distance
Baselines: kernel-based conditional density estimators GAN-based minimax estimators
Builds on: score-based diffusion theory (Oko et al. 2023) classifier-free conditional score matching Girsanov-based analysis of score-based models
Closest papers by meaning
- A Likelihood Based Approach to Distribution Regression Using Conditional Deep Generative Models (ICML 2025 · similarity 0.91)
- Guidance with Spherical Gaussian Constraint for Conditional Diffusion (ICML 2024 · similarity 0.88)
- Manifold Preserving Guided Diffusion (ICLR 2024 · similarity 0.88)
- Theory on Score-Mismatched Diffusion Models and Zero-Shot Conditional Samplers (ICLR 2025 · similarity 0.87)
- Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models (NeurIPS 2024 · similarity 0.87)
- On Statistical Rates of Conditional Diffusion Transformers: Approximation, Estimation and Minimax Optimality (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.