A recent arXiv preprint, "Online Inference in Distributional Temporal-Difference Learning," has been withdrawn by its authors after they identified a technical error in the proof of Theorem 4. The paper had proposed methods for statistical inference on functionals of the return distribution in reinforcement learning, estimated from a single Markov trajectory using nonparametric distributional temporal-difference learning.
The authors had proved that the root-T error of the Polyak–Ruppert averaged estimator converges weakly to a centered Gaussian random element in Cramér space. They also claimed that, conditionally on the observed trajectory, the bootstrap analogue converges to the same limit, justifying inference for smooth functionals like variance, CVaR, expected shortfall, and expectiles. For nonsmooth functionals such as return quantiles, they developed a local asymptotic theory over T^{-1/2}-neighborhoods of thresholds.
The withdrawal note states that there is a technical error in "Step 3" of the proof of Theorem 4, section 4. Because the paper is withdrawn, the claims should not be treated as established results. The abstract remains available, but the full text is no longer distributed under a license, and the findings await correction or re-validation.