English

Sharper bounds for online learning of smooth functions of a single variable

Machine Learning 2021-06-01 v1 Discrete Mathematics Machine Learning

Abstract

We investigate the generalization of the mistake-bound model to continuous real-valued single variable functions. Let Fq\mathcal{F}_q be the class of absolutely continuous functions f:[0,1]Rf: [0, 1] \rightarrow \mathbb{R} with fq1||f'||_q \le 1, and define optp(Fq)opt_p(\mathcal{F}_q) as the best possible bound on the worst-case sum of the pthp^{th} powers of the absolute prediction errors over any number of trials. Kimber and Long (Theoretical Computer Science, 1995) proved for q2q \ge 2 that optp(Fq)=1opt_p(\mathcal{F}_q) = 1 when p2p \ge 2 and optp(Fq)=opt_p(\mathcal{F}_q) = \infty when p=1p = 1. For 1<p<21 < p < 2 with p=1+ϵp = 1+\epsilon, the only known bound was optp(Fq)=O(ϵ1)opt_p(\mathcal{F}_{q}) = O(\epsilon^{-1}) from the same paper. We show for all ϵ(0,1)\epsilon \in (0, 1) and q2q \ge 2 that opt1+ϵ(Fq)=Θ(ϵ12)opt_{1+\epsilon}(\mathcal{F}_q) = \Theta(\epsilon^{-\frac{1}{2}}), where the constants in the bound do not depend on qq. We also show that opt1+ϵ(F)=Θ(ϵ12)opt_{1+\epsilon}(\mathcal{F}_{\infty}) = \Theta(\epsilon^{-\frac{1}{2}}).

Keywords

Cite

@article{arxiv.2105.14648,
  title  = {Sharper bounds for online learning of smooth functions of a single variable},
  author = {Jesse Geneson},
  journal= {arXiv preprint arXiv:2105.14648},
  year   = {2021}
}