Slightly Conservative Bootstrap for Maxima of Sums

Abstract: We study the bootstrap for the maxima of the sums of independent random variables, a problem of high relevance to many applications in modern statistics. Since the consistency of bootstrap was justified by Gaussian approximation in Chernozhukov et al. (2013), quite a few attempts have been made to sharpen the error bound for bootstrap and reduce the sample size requirement for bootstrap consistency. In this paper, we show that the sample size requirement can be dramatically improved when we make the inference slightly conservative, that is, to inflate the bootstrap quantile $t_{\alpha}^*$ by a small fraction, e.g. by $1\%$ to $1.01\,t^*_\alpha$. This simple procedure yields error bounds for the coverage probability of conservative bootstrap at as fast a rate as $\sqrt{(\log p)/n}$ under suitable conditions, so that not only the sample size requirement can be reduced to $\log p \ll n$ but also the overall convergence rate is nearly parametric. Furthermore, we improve the error bound for the coverage probability of the standard non-conservative bootstrap to $[(\log (np))^3 (\log p)^2/n]^{1/4}$ under general assumptions on data. These results are established for the empirical bootstrap and the multiplier bootstrap with third moment match. An improved coherent Lindeberg interpolation method, originally proposed in Deng and Zhang (2017), is developed to derive sharper comparison bounds, especially for the maxima.