Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2502.09306 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866912231246004224 |
|---|---|
| author | Cordero-Encinar, Paula Akyildiz, O. Deniz Duncan, Andrew B. |
| author_facet | Cordero-Encinar, Paula Akyildiz, O. Deniz Duncan, Andrew B. |
| contents | We investigate the theoretical properties of general diffusion (interpolation) paths and their Langevin Monte Carlo implementation, referred to as diffusion annealed Langevin Monte Carlo (DALMC), under weak conditions on the data distribution. Specifically, we analyse and provide non-asymptotic error bounds for the annealed Langevin dynamics where the path of distributions is defined as Gaussian convolutions of the data distribution as in diffusion models. We then extend our results to recently proposed heavy-tailed (Student's t) diffusion paths, demonstrating their theoretical properties for heavy-tailed data distributions for the first time. Our analysis provides theoretical guarantees for a class of score-based generative models that interpolate between a simple distribution (Gaussian or Student's t) and the data distribution in finite time. This approach offers a broader perspective compared to standard score-based diffusion approaches, which are typically based on a forward Ornstein-Uhlenbeck (OU) noising process. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_09306 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Non-asymptotic Analysis of Diffusion Annealed Langevin Monte Carlo for Generative Modelling Cordero-Encinar, Paula Akyildiz, O. Deniz Duncan, Andrew B. Machine Learning Probability Computation We investigate the theoretical properties of general diffusion (interpolation) paths and their Langevin Monte Carlo implementation, referred to as diffusion annealed Langevin Monte Carlo (DALMC), under weak conditions on the data distribution. Specifically, we analyse and provide non-asymptotic error bounds for the annealed Langevin dynamics where the path of distributions is defined as Gaussian convolutions of the data distribution as in diffusion models. We then extend our results to recently proposed heavy-tailed (Student's t) diffusion paths, demonstrating their theoretical properties for heavy-tailed data distributions for the first time. Our analysis provides theoretical guarantees for a class of score-based generative models that interpolate between a simple distribution (Gaussian or Student's t) and the data distribution in finite time. This approach offers a broader perspective compared to standard score-based diffusion approaches, which are typically based on a forward Ornstein-Uhlenbeck (OU) noising process. |
| title | Non-asymptotic Analysis of Diffusion Annealed Langevin Monte Carlo for Generative Modelling |
| topic | Machine Learning Probability Computation |
| url | https://arxiv.org/abs/2502.09306 |