Online Fair Division with Additional Information

A team of computer science researchers has published a significant paper titled 'Online Fair Division with Additional Information' on arXiv. The work tackles a core problem in algorithmic game theory and AI: how to fairly allocate a sequence of indivisible goods (like tasks, compute resources, or items) to multiple agents as they arrive online, with decisions being irrevocable. The researchers prove that without any future information, achieving even approximate versions of key fairness notions—envy-freeness, proportionality, and maximin share fairness—is strongly impossible. This establishes a clear baseline for the challenge.

The paper's major contribution is showing how different types of limited additional information can unlock provable fairness guarantees. First, with just 'normalization information' (knowing each agent's total value for all goods), they design an algorithm that achieves stronger fairness results than previously known and show matching impossibilities for stronger notions. Second, and more powerfully, with 'frequency predictions' (knowing the multiset of future item values but not their order), they create a meta-algorithm. This framework can lift a broad class of guarantees from offline 'share-based' algorithms directly to the online setting, matching the best-known offline bounds.

Finally, the researchers provide practical, learning-augmented versions of their models that are robust to real-world imperfections. These variants gracefully degrade fairness guarantees in proportion to the error in noisy total value predictions or noisy frequency predictions, making the approach applicable where machine learning models provide the forecasts.