New Extended PFOF Filter Tackles Nonlinear State Estimation with Linear Operators

Nonlinear state estimation is notoriously difficult: methods like the extended Kalman filter (EKF) struggle with non-Gaussian noise and complex dynamics. Now, researchers from Japan introduce the extended Perron-Frobenius Operator Filter (PFOF), published on arXiv. The key insight: the Perron-Frobenius operator is an infinite-dimensional linear operator that completely preserves the properties of a nonlinear system. By learning this operator through extended Dynamic Mode Decomposition (eDMD), the PFOF converts a nonlinear estimation problem into a linear one — without sacrificing fidelity to the true dynamics.

The new filter explicitly accounts for non-Gaussian probability distributions by allowing users to choose basis functions in the eDMD framework. This flexibility means the method can adapt to different system characteristics, from robotics to climate models. In two numerical examples, the PFOF demonstrated both high computational efficiency and high estimation accuracy, outperforming conventional approaches. The authors submitted the paper to IEICE's Nonlinear Theory and Its Applications (NOLTA), and while still awaiting peer review, the preprint suggests a practical alternative for real-time control systems and sensor fusion where traditional filters fall short.