Spectral Alignment in Forward-Backward Representations via Temporal Abstraction

This research paper introduces temporal abstraction as a method to align spectral properties in forward-backward (FB) representations for learning successor representations in continuous spaces. The authors demonstrate that temporal abstraction acts like a low-pass filter, suppressing high-frequency components to reduce the effective rank of the induced successor representation while maintaining bounds on value function error. Empirical results indicate that this approach enhances the stability of FB learning, particularly under high discount factors where bootstrapping errors are common, offering a principled way to shape the spectral structure of Markov decision processes for long-horizon tasks.