• Corpus ID: 231802300

Online Discrepancy Minimization via Persistent Self-Balancing Walks

@article{Arbour2021OnlineDM,
  title={Online Discrepancy Minimization via Persistent Self-Balancing Walks},
  author={David T. Arbour and Drew Dimmery and Tung Mai and Anup B. Rao},
  journal={ArXiv},
  year={2021},
  volume={abs/2102.02765}
}
Komlós setting where the norm of each x is bounded from above by one, i.e., ‖xi‖2 ≤ 1. In the most general, fully adversarial setting, where the values of subsequent x are allowed to change in response to the assignment of ǫ, the best possible bound is Ω( √ n) since an adversary can always choose xt+1 orthogonal to the current partial sum wt = ∑t 1 ǫixi at each step. We consider the setting of an oblivious adversary, where the sequence of x can be arbitrarily set by the adversary, but is not…