In many real-world learning systems, ground-truth labels do not arrive immediately, and may never arrive if they come too late. A new arXiv preprint formalizes this as a PAC learning problem with random label delays and censoring beyond a fixed observation window W. The central question is how many samples are required under this observation model.
The paper proposes a distribution-dependent quantity—the delay complexity dimension—that governs the sample complexity. Rather than treating delay as a mere nuisance, this dimension appears to encode how the timing of labels interacts with the difficulty of the underlying classification problem.
The abstract does not include the full theorem statements, so the exact form of the bounds is not available from the announcement. As this is a single source, there are no differing accounts to reconcile; the summary reflects only the arXiv abstract.