{"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","summary":"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…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-11-15","reactions":{}}