Building On Learning-Augmented Online Sorting

Online sorting is a problem that has gained attention recently, and is a natural translation of sorting into an online setting. The problem at hand is to place elements of an online sequence immediately and irrevocably with the goal of producing an array that is as close to sorted as possible upon inserting all elements. Recent work by Bercea et al. [ESA 2026] introduced the incorporation of learned predictions into the online sorting domain. They show lower bounds for the consistency and robustness of any algorithm, including randomization, then provide a deterministic algorithm that is $O(n/r)$-consistent and $O(r)$-robust, which are tight bounds. However, this algorithm requires prior knowledge of the optimal offline cost of the full input sequence before all elements have been revealed. We resolve this by providing an algorithm with matching bounds that does not require knowledge of the optimal cost, then extend these results to the more general problem of online metric TSP, in line with the prior work.