Bernstein-type bounds for beta distribution

Abstract: This work obtains sharp closed-form exponential concentration inequalities of Bernstein type for the ubiquitous beta distribution, improving upon sub-gaussian and sub-gamma bounds previously studied in this context. The proof leverages a novel handy recursion of order 2 for central moments of the beta distribution, obtained from the hypergeometric representations of moments; this recursion is useful for obtaining explicit expressions for central moments and various tail approximations.

Review text

Consider the distribution of \(X\sim B(\alpha,\beta)=Beta(\alpha,\beta)\) with parameters \(\alpha,\beta>0\) by \[ P\{X\leq\varepsilon\}=\int\frac{x^{\alpha-1}(1-x)^{\beta-1}}{B(\alpha, \beta)}dx, \quad 0\leq\varepsilon\leq1, \] where \(B(\alpha,\beta):=\int x^{\alpha-1}(1-x)^{\beta-1}dx\). The author provides as central result a Bernstein-type tail bound with numerically optimal parameters. More precisely, this statement is given in the following theorem: Theorem 1. Let \(X\sim B(\alpha,\beta)\) and \[ v:=\frac{\alpha\beta}{(\alpha+\beta)^2(\alpha+\beta+1)},\quad c:=\frac{2(\beta-\alpha)}{(\alpha+\beta)(\alpha+\beta+2)}. \] Then the upper tail of \(X\) is bounded as \[ P\{X>E(X)+\varepsilon\}\leq \begin{cases} \exp\left(-\frac{\varepsilon^2}{2(v+\frac{c\varepsilon}{3})}\right), &\text{if }\beta\geq\alpha,\\ \exp\left(-\frac{\varepsilon^2}{2v}\right), &\text{if }\beta<\alpha, \end{cases} \] and the lower tail of \(X\) is bounded as \[ P\{X<E(X)-\varepsilon\}\leq \begin{cases} \exp\left(-\frac{\varepsilon^2}{2(v+\frac{c\varepsilon}{3})}\right),&\text{if }\alpha\geq\beta,\\ \exp\left(-\frac{\varepsilon^2}{2v}\right), &\text{if }\alpha<\beta. \end{cases} \] For small deviations \(\varepsilon\), the following theorem shows that the parameters \(c\) and \(v\) are best possible for the exponential moment method: Theorem 2 (Best Cramér-Chernoff bound). Let \(c\) and \(v\) be as in Theorem 1. Then the best bound that can be obtained from the Cramér-Chernoff method is \[ P\{X>E(X)+\varepsilon\}\leq\exp\left(-\frac{\varepsilon^2}{2v}+\frac{c\varepsilon^3}{6v^2}+O(\varepsilon^4)\right), \] as \(\varepsilon\to 0\). Comparing this result with the tail from Theorem 1 as \(\varepsilon\to 0\), the author obtains that the values of constants \(c\) and \(v\) in Bernstein-type inequality in Theorem 1 are optimal. He gives and proves an optimal Bernstein-type concentration inequality and uses a recursion formula for central moments of the beta distribution. The author proposes numerical computations to validate numerical comparison and provides a numerical evaluation to illustrate the new bounds.

Recommendations

Cites work

Cited in

(7)

This page was built for publication: Bernstein-type bounds for beta distribution

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6157635)