Adversarial Resilience of Poisson-Process Submodular Maximization over Matroids

A new research paper on arXiv (2608.12134) presents a theorem on adversarial resilience for the Spiteful Greedy Swap Poisson Process (SGS-Poisson) in nonnegative submodular maximization subject to a general matroid. The algorithm, when given a controlled value oracle with bounded error, retains its approximation guarantees: 1/e for non-monotone objectives and 1-1/e for monotone objectives, even under adversarial perturbations. The implementation uses O~(nk^2 ε^{-2}) oracle calls and achieves expected value at least (1/e - ε)OPT - O(kξ) and (1-1/e - ε)OPT - O(kξ) respectively. This result enables offline-to-online reduction, yielding full-bandit combinatorial multi-armed bandit (CMAB) algorithms for general matroid-constrained submodular rewards. The paper is categorized as a cross announcement and is authored by researchers in the field of algorithms and optimization. The findings are significant for robust optimization and online learning, particularly in settings where function evaluations are noisy or adversarial.