Id: arxiv/2611.19772 Url: https://arxiv.org/abs/2611.19772 Tags: stat.ML, cs.LG, math.PR Score: 0 Title: Tight Sequential Confidence Intervals for Online Kernel Regression Domain: stat.ML Source: arxiv Authors: Haichen Hu, David Simchi-Levi Reviews: 0 Subject: stat.ML Category: research Comments: 0 Disputed: false Division: stat Mentions: 0 Discussed: 0 Published: 2026-11-15 Page: https://delphi.science/r/arxiv/2611.19772 We characterize, in the worst case over kernels, the sharp minimax width of online confidence intervals for elements of a reproducing kernel Hilbert space (RKHS). This resolves the open problem posed by Vakili et al. (2021) in its original discrete-time form. A learner observes $y_i=f(x_i)+ε_i$ at adaptively chosen points, with $\|f\|_{\mathcal H_k}\le B$ and i.i.d. $R$-sub-Gaussian noise, and reports the interval $|f(x)-μ_n(x)|\leρ_n(δ)σ_n(x)$, where $μ_n$ and $σ_n$ are the posterior mean and standard deviation of a surrogate Gaussian process model with regularization parameter $λ$. Writing…