On a high level, both approaches build accurate estimates

We thank the reviewers for their comments and insightful reviews. 's is only logarithmic as the main dependency is w.r.t. VI algorithm for SSP was proved in [37] to converge in time quadratic w.r.t. the size of the considered state space This allows tuning the parameter online according to the desired behavior. A sketch of the proof of Thm. 1 is currently available in App. B. In case of acceptance we will use We will include additional experiments for varying L in the final version.