Large-deviation efficient confidence intervals for the sub-Gaussian mean
We study confidence intervals for the mean of iid observations with known sub-Gaussian variance proxy \(s^2\). The usual Chernoff interval is neither Gaussian-efficient (as $n \to \infty, \alpha$ fixed) nor large-deviation efficient \(n^{-1}\log(1/\alpha_n)\to c\in(0,\infty)\), and it does not adapt to the distribution at hand. Recent work produced the first Gaussian-efficient sub-Gaussian confidence sets, and the current work presents the first large-deviation efficient ones, making instance-optimality feasible and practical. The first of our constructions is a simple discretization and application of the method of types. As the bins refine with sample size, both endpoints converge almost surely to the endpoints of the level-\(c\) set of a population Kullback--Leibler projection. A standard change-of-measure argument shows that every uniformly valid interval has limiting width at least this large. A second construction uses the empirical KL projection (so-called empirical $\mathrm{KL}_\mathrm{inf}$), for which we prove a sharp cutoff of order \(\sqrt{\log(en)}+\log(1/\alpha)\). This shows a distinction from its bounded mean subclass, where recent work has shown that the $\sqrt{\log(en)}$ term can be omitted.
Authors
- Aaditya Ramdas (ORCID: https://orcid.org/0000-0003-0497-311X)
Institutions
- Stanford University (US)
Publication Details
- Journal
- Stanford Digital Repository
- Published
- 2026-11-20
- DOI
- https://doi.org/10.25740/zm998gx5428
- Primary Topic
- Probability and Statistical Research
- Type
- article
- Field-Weighted Citation Impact
- 0.00