Optimal Oracle Complexity of Dual Halpern Iteration in Normed Spaces

Abstract

In this work, we study the convergence and error bounds of dual Halpern iteration for nonexpansive operators in normed spaces. First, we show that dual Halpern iteration exhibits the same optimal residual rate as Halpern iteration in the deterministic setting through a direct convergence analysis. With inexact operator evaluations, we also conduct a convergence analysis in normed spaces and derive the robust error bound of inexact dual Halpern iteration. Based on this robustness result, we study the oracle complexity of stochastic dual Halpern iteration in $(\mathcal{R}^d,\|\cdot\|_p)$ for $p\in(1,\infty]$, using an unbiased stochastic oracle with bounded variance $σ^2$. Notably, for $p\in[2,\infty]$, our method achieves an oracle complexity of $O\left(ε^{-1}+σ^2ε^{-3}\right)$ in its dependence on $ε$ and $σ$, which matches the lower bound for the case $p=\infty$ without a logarithmic gap in the dependence on $ε$.