Drift Estimation for Stochastic Differential Equations with Denoising Diffusion Models

A team of researchers has published a novel paper, 'Drift Estimation for Stochastic Differential Equations with Denoising Diffusion Models,' introducing a surprising new application for generative AI. The work, authored by Marcos Tapia Costa, Nikolas Kantas, and George Deligiannidis, tackles a core problem in stochastic calculus: estimating the deterministic 'drift' component of a stochastic differential equation when only multiple, high-frequency observed trajectories are available.

The technical innovation lies in reformulating the statistical estimation problem as a conditional denoising task. Instead of using diffusion models to generate images or text, the team trains a conditional diffusion model on the observed SDE trajectories. The model learns to denoise the stochastic process, and a key byproduct of this training is a direct estimator for the unknown drift function. The method assumes the diffusion coefficient is known and focuses on time-homogeneous drift. In empirical tests across different drift function classes, the diffusion-based estimator performed on par with established classical methods in low-dimensional settings. More notably, it maintained consistent competitiveness in higher-dimensional problems, with performance gains that the authors note cannot be solely attributed to neural network architecture choices.

This research is significant because it bridges two distinct fields: generative AI and mathematical finance/engineering. Stochastic differential equations are fundamental for modeling everything from stock prices and interest rates to biological systems and physical phenomena. Accurate drift estimation is critical for prediction, control, and understanding these systems. By demonstrating that a powerful tool from generative AI can be repurposed for this foundational mathematical task, the paper opens a new pathway for applying deep learning to complex scientific computing problems, potentially offering more scalable solutions for high-dimensional SDEs where traditional methods struggle.