Convergence Analysis of Discrete Diffusion Model: Exact Implementation through Uniformization

Hongrui Chen,Lexing Ying

2024-02-14

Abstract:Diffusion models have achieved huge empirical success in data generation tasks. Recently, some efforts have been made to adapt the framework of diffusion models to discrete state space, providing a more natural approach for modeling intrinsically discrete data, such as language and graphs. This is achieved by formulating both the forward noising process and the corresponding reversed process as Continuous Time Markov Chains (CTMCs). In this paper, we investigate the theoretical properties of the discrete diffusion model. Specifically, we introduce an algorithm leveraging the uniformization of continuous Markov chains, implementing transitions on random time points. Under reasonable assumptions on the learning of the discrete score function, we derive Total Variation distance and KL divergence guarantees for sampling from any distribution on a hypercube. Our results align with state-of-the-art achievements for diffusion models in $\mathbb{R}^d$ and further underscore the advantages of discrete diffusion models in comparison to the $\mathbb{R}^d$ setting.

Machine Learning

What problem does this paper attempt to address?

The paper aims to address the theoretical properties of diffusion models in discrete state spaces. Specifically:

1. **Theoretical Analysis of Discrete Diffusion Models**: Compared to diffusion models in continuous spaces, there is currently a lack of in-depth theoretical research on diffusion models in discrete state spaces (such as data appearing in tasks like natural language processing, image representation, etc.). This paper conducts a detailed theoretical analysis of discrete diffusion models by modeling the forward and reverse processes as Continuous-Time Markov Chains (CTMC).

2. **Exact Implementation Algorithm**: An algorithm based on the uniformization technique is proposed, which can accurately simulate the reverse CTMC process without introducing discretization errors. This method overcomes the discretization error issues present in continuous diffusion models and demonstrates the effectiveness of the algorithm.

3. **Complexity and Error Analysis**: Under reasonable assumptions, the paper derives guarantees for the total variation distance and KL divergence of sampling from distributions on arbitrary hypercubes. Compared to diffusion models in continuous spaces, the proposed method exhibits better linear dependency in terms of complexity, especially in high-dimensional data scenarios.

In summary, this paper fills the gap in the theoretical analysis of diffusion models in discrete state spaces and proposes an efficient and accurate simulation algorithm, demonstrating its potential in practical applications.