Fractional Risk Analysis of Stochastic Systems with Jumps and Memory

A research team from multiple institutions, including lead authors Yimeng Sun and Zhuoyuan Wang, has published a groundbreaking paper titled "Fractional Risk Analysis of Stochastic Systems with Jumps and Memory" on arXiv. The work addresses a critical gap in safety assessment for autonomous systems like self-driving cars and robotic controllers, which often operate in environments where sudden, asymmetric events (jumps) and long-term dependencies (memory) make traditional risk models inaccurate. Their key innovation is deriving a unified space- and time-fractional Partial Differential Equation (PDE) that captures these complex dynamics in a single framework, enabling the joint evaluation of risk probabilities across different initial states and time horizons.

This new model fundamentally differs from standard, non-fractional PDEs and reveals risk behaviors that previous methods could not detect. Crucially, the team demonstrates how physics-informed machine learning can be used to solve these complex fractional PDEs efficiently. This approach bypasses the need for millions of computationally expensive, long-horizon simulations typically required to capture rare safety-critical events. The result is a method that provides accurate risk predictions across a wide range of system configurations and, importantly, generalizes well to dynamics it wasn't explicitly trained on, a key requirement for deploying AI in unpredictable real-world settings.