Finite-sample characterization of Distributional Active Inference (DAIF)

Establish a rigorous characterization of the finite-sample characteristics of the Distributional Active Inference (DAIF) algorithm within the distributional reinforcement learning framework.

Background

The paper introduces DAIF, a model-free algorithm that integrates active inference into distributional reinforcement learning via push-forward constructions and shows improved contraction properties compared to standard distributional RL.

While asymptotic convergence inherits known guarantees from distributional RL, the authors note that finite-sample behavior remains theoretically uncharacterized. They suggest their framework may support future analyses, potentially by extending sample complexity tools from linear-quadratic control to well-behaved nonlinear systems.