Researchers pinpoint double descent origins in reduced order models

A new research paper by Andrei A. Klishin, J. Nathan Kutz, and Krithika Manohar (University of Washington) investigates the origins of double descent in reduced order modeling (ROM). Double descent is a counterintuitive phenomenon where reconstruction error first decreases, then sharply increases, then decreases again as model complexity grows. The team shows that this error spike occurs due to catastrophic amplification of a pathological signal in the reconstruction process. Using a unified Data-Noise Averaging theory, they provide both qualitative criteria for when double descent emerges and quantitative predictions of the full risk curve at minimal computational cost.

The researchers demonstrate their theory on two case studies: static reconstruction of Sea Surface Temperature (SST) patterns and time integration of a reduced order model of a partial differential equation (PDE). They trace the instability to individual sensors and combinations, enabling targeted regularization to mitigate the effect. The work bridges machine learning theory with practical engineering applications, offering concrete mechanisms to stabilize sparse sensing systems in environmental monitoring and physics simulations.