中文

量子学习中隐私与稳定性的等价性及其泛化保证

量子物理 2026-02-06 v2 信息论 机器学习 math.IT

摘要

我们提出了统一的信息论框架,阐明了量子学习算法中稳定性、隐私与泛化性能之间的相互作用。我们以量子互信息为基础,建立了预期泛化误差的上界,并推导了概率上界,概括了Esposito等人(2021)的经典结果。此外,我们提供了预期真实损失相对于预期经验损失的下界。此外,我们证明了 (ε,δ)(\varepsilon, \delta)-量子差分隐私学习算法是稳定的,从而确保了强大的泛化保证。最后,我们扩展了分析到作弊学习算法,引入信息论可接受性(ITA)来表征当学习算法对特定数据集实例不敏感时隐私的根本限制。

关键词

引用

@article{arxiv.2602.01177,
  title  = {Equivalence of Privacy and Stability with Generalization Guarantees in Quantum Learning},
  author = {Ayanava Dasgupta and Naqueeb Ahmad Warsi and Masahito Hayashi},
  journal= {arXiv preprint arXiv:2602.01177},
  year   = {2026}
}

备注

31 pages, 3 figures; Major revision including a new probabilistic bound on generalization error (Theorem 2) and a new complementary lower bound on the expected true loss (Theorem 3); Appendices have been expanded to include further proofs and details