Hoeffding and Bernstein inequalities for weighted sums of exchangeable random variables

Rina Foygel Barber
2024-08-15
Abstract:The aim of this paper is to establish Hoeffding and Bernstein type concentration inequalities for weighted sums of exchangeable random variables. A special case is the i.i.d. setting, where random variables are sampled independently from some distribution (and are therefore exchangeable). In contrast to the existing literature on this problem, our results provide a natural unified view of both the i.i.d. and the exchangeable setting.
Statistics Theory,Probability
What problem does this paper attempt to address?
The main objective of this paper is to establish Hoeffding and Bernstein-type concentration inequalities for weighted sums of the form \(w_1X_1 + \ldots + w_nX_n\), where \(X_1, \ldots, X_n\) are exchangeable random variables. Specifically, the paper addresses the following issues: 1. **Establishing new concentration inequalities**: The authors propose new versions of Hoeffding and Bernstein inequalities that are applicable to exchangeable random variables. These inequalities are particularly useful for the case of weighted sums and provide approximately tight bounds for a fixed number of terms \(n\). 2. **Recovering classical results under the i.i.d. setting**: As the sequence length \(N\) tends to infinity, these newly proposed inequalities can revert to the classical Hoeffding and Bernstein inequalities under the independent and identically distributed (i.i.d.) setting. 3. **Symmetry of weights**: For ease of interpretation, the obtained bounds need to exhibit symmetry with respect to the weights \(w_1, \ldots, w_n\). By addressing the above issues, this paper fills a critical gap in the existing literature, where the current concentration inequalities cannot be directly applied to exchangeable data when the weights \(w_1, \ldots, w_n\) are different. Moreover, these results have broad potential applications in various fields of statistics, such as permutation tests, conformal prediction, and regression inference.