Successor-measure spectral basis (Laplacian/proto-value eigenvectors)

Trains a deep encoder to estimate the successor measure (discounted future-state occupancy) of a real RL agent's transitions, then extracts the top-d singular/eigenvectors of the resulting matrix as a task-agnostic auxiliary basis — provably equal to graph-Laplacian eigenvectors (proto-value functions) under a symmetric transition matrix.