Time-uniform, nonparametric, nonasymptotic confidence sequences (Q2039804)

This paper is really a continuation of the authors' work [Probab. Surv. 17, 257--317 (2020; Zbl 1456.60054)] in the context of confidence sequences. A confidence sequence is defined as a sequence of confidence sets \(CI_t\), \(t=1,2,\dots\) which satisfy \(P(\forall t \geq 1:\theta_t \in CI_t)\geq 1-\alpha, \text{ where } \theta_t\) is a quantity of interest and \(1-\alpha\) is a preassigned confidence level. The authors develop methods to derive confidence sequences by using exponential concentration inequalities based on time-uniform Chernoff bounds. They draw connections between the Cramér-Chernoff method for exponential concentration, the law of the iterated logarithm (LIL) and the sequential probability ratio test. They explore confidence sequences as time-uniform extensions of the exponential concentration, provide nonasymptotic characterizations of the second, and generalize the sequential tests to nonparametric settings. Numerous illustrations of their methods are given such as an empirical-Bernstein bound growing at a LIL rate, and upper LIL for the maximum eigenvalue of a sum of random matrices. Some applications such as covariance matrix estimation, and estimation of sample average treatment effect under the Neyman-Rubin potential outcomes model are given.