Principled Generative Modeling: Theoretical Error Bounds and Applications to Missing Values

18.09.2026 11:15 – 12:15

RESEARCH INSTITUTE FOR STATISTICS AND INFORMATION SCIENCE SEMINARS

ABSTRACT

In this talk, I will present two recent works on generative modeling. The first one deals with non-asymptotic error bounds for probability flow ODEs, a deterministic reformulation of diffusion models. While probability flow ODEs have been highly successful in practice, most theoretical guarantees rely on restrictive regularity assumptions on the target distribution, such as strong log-concavity. In our work, we established convergence bounds in the 2-Wasserstein distance under considerably weaker assumptions. In particular, our results now also cover non-log-concave distributions such as Gaussian mixtures. Notably, the asymptotic behavior of the error bound matches that obtained under stronger assumptions, while the non-asymptotic rates explicitly quantify the detrimental effect of the non-log-concavity.
In the second part of the talk, I will present a new generative modeling approach to handling missing values. In current practice, missing-data workflows frequently rely on ad-hoc imputation strategies without any theoretical backing. We propose FLOWGEM, a principled iterative method for generating a complete dataset from a dataset with values Missing at Random (MAR). The core idea is to employ an approximate Wasserstein gradient flow to minimize the expected Kullback–Leibler (KL) divergence between the observed data distribution and the distribution of the generated sample over different missingness patterns. Notably, this target functional is minimized by the true data distribution. Simulation studies and real-data benchmarks demonstrate that FLOWGEM performs competitively with state-of-the-art methods across a range of settings, and it also appears well-suited for future theoretical analysis and guarantees.

Lieu

Bâtiment: Uni Mail

Boulevard du Pont-d'Arve 40
1205 Geneva

Room M 5220, 5th floor

entrée libre