Abstract:We study the Inexact Langevin Dynamics (ILD), Inexact Langevin Algorithm (ILA), and Score-based Generative Modeling (SGM) when utilizing estimated score functions for sampling. Our focus lies in establishing stable biased convergence guarantees in terms of the Kullback-Leibler (KL) divergence. To achieve these guarantees, we impose two key assumptions: 1) the target distribution satisfies the log-Sobolev inequality (LSI), and 2) the score estimator exhibits a bounded Moment Generating Function (MGF) error. Notably, the MGF error assumption we adopt is more lenient compared to the $L^\infty$ error assumption used in existing literature. However, it is stronger than the $L^2$ error assumption utilized in recent works, which often leads to unstable bounds. We explore the question of how to obtain a provably accurate score estimator that satisfies the MGF error assumption. Specifically, we demonstrate that a simple estimator based on kernel density estimation fulfills the MGF error assumption for sub-Gaussian target distribution, at the population level.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the convergence problem of Inexact Langevin Dynamics (ILD), Inexact Langevin Algorithm (ILA) and Score - based Generative Models (SGM) when using the estimated score function for sampling. Specifically, the author focuses on establishing stable bias convergence guarantees under the Kullback - Leibler (KL) divergence. To achieve this goal, the author proposes two key assumptions:
1. **The target distribution satisfies the Log - Sobolev Inequality (LSI)**: This ensures that the target distribution has a certain smoothness and concentration.
2. **The score estimator has a bounded moment - generating function (MGF) error**: This is more relaxed than the existing \(L^\infty\) error assumption, but stronger than the \(L^2\) error assumption used in recent work.
Through these assumptions, the author aims to prove the stable convergence of ILD, ILA and SGM under the KL divergence when using the inexact score function. In addition, the author also explores how to obtain a score estimator that satisfies the MGF error assumption and shows that a simple estimator based on Kernel Density Estimation (KDE) satisfies the MGF error assumption under the sub - Gaussian target distribution.
### Main contributions
1. **Bias convergence guarantee under KL divergence**: The author proves the bias convergence guarantee of ILD and ILA under the KL divergence under the MGF error assumption. These results are stable, that is, the upper bound remains bounded at all time points.
2. **Convergence guarantee under Rényi divergence**: Under the stronger \(L^\infty\) error assumption, the author proves the convergence guarantee of ILA under the Rényi divergence. These results are also stable, which is very important for applications such as differential privacy.
3. **Convergence guarantee of SGM under KL divergence**: The author derives the convergence guarantee of SGM under the KL divergence when the target distribution satisfies LSI and the score estimator has a bounded MGF error. Compared with previous literature, these results are stable.
4. **Accuracy of KDE - based score estimator**: The author shows that when the target distribution is sub - Gaussian, the simple score estimator based on KDE satisfies the MGF error assumption at the population level, thus obtaining the iterative complexity guarantee of ILA.
### Conclusion
By introducing the MGF error assumption, this paper successfully establishes stable convergence guarantees of ILD, ILA and SGM under the KL divergence when using the inexact score function. These results are not only theoretically significant, but also provide theoretical support for practical applications. However, the author also points out the directions for future research, such as whether similar stable convergence guarantees can be obtained under the \(L^2\) error assumption, and the error analysis of score estimation in the finite - sample case.