Measuring distributional asymmetry with Wasserstein distance and Rademacher symmetrization (Q1657945)
From MaRDI portal
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Measuring distributional asymmetry with Wasserstein distance and Rademacher symmetrization |
scientific article |
Statements
Measuring distributional asymmetry with Wasserstein distance and Rademacher symmetrization (English)
0 references
14 August 2018
0 references
The paper under review aims to improve an ubiquitous symmetrization inequality by Wasserstein distance. \textit{S. Arlot} et al. [Ann. Stat. 38, No. 1, 51--82 (2010; Zbl 1180.62066)] conjectured that an inequality holds under an assumption less restrictive than symmetry. Motivated by this conjecture, the author aims to improve the inequality with a explicit constant \(C(\mu)\) depending on the symmetry of the underlying measures \(\mu\) of those i.i.d \(X_i\) such that \[ E \left\| \frac{1}{n} \sum_{i=1}^n (X_i - EX_i) \right\| \leq E \left\|\frac{1}{n} \sum_{i=1}^n \varepsilon_i (X_i - EX_i) \right\| +\frac{C(\mu)}{\sqrt{n}}, \] where \(\varepsilon_i\) is a Bernoulli random variable with \(P(\varepsilon_i = \pm 1) = \frac{1}{2}\). Section 2 considers the practical issue of computing the norm of the Rademacher sum \(R_n = \sum_{i=1}^n \varepsilon_i (X_i - \overline{X})\) with the average \(\overline{X} =\frac{1}{n}\sum_{i=1}^nX_i\). The term \(S_n = \sum_{i=1}^n (X_i - EX_i)\) and \(E\| S_n\|\) are dealt with the bootstrap estimators. The key of the paper is to tract the small correction term \(C_n(\mu)\) for \(E\|S_n\| \leq \frac{1}{M}\sum_{m=1}^M \| R_n^{(m)}\| +\frac{C_n(\mu)}{\sqrt{2n}}\). Section 3 is devoted the symmetrization result by adding the asymmetric Wasserstein distance contribution, and the main Theorem 3.4 shows that averaging a collection of random variables has an inherent smoothing and symmetric effect following the central limit theorem. Section 4 is to establish an empirical estimate of the Wasserstein distance, and the rate of convergence of empirical estimate as well as the bootstrap estimator for multiple bootstrap samples. Bootstrap permutation test procedures are given in Section 5.
0 references
concentration inequality
0 references
generalized bootstrap
0 references
high dimension confidence set
0 references
symmetrization inequality
0 references
Wasserstein distance
0 references
Rademacher sum
0 references
permutational Rademacher complexity
0 references
(co)type
0 references
Nemirovski inequality
0 references
central limit theorem
0 references
0 references
0 references