基于投影的高维两样本分布相等的检验方法研究

两样本检验是统计学中最重要的问题之一。经典的非参数检验方法，如著名的Kolmogorov-Smirnov检验，Cramér-von Mises检验和Anderson-Darling检验，是基于经验分布函数构造的。尽管这类检验在一维两样本框架下具有很多优良的性质如稳健性质和“distribution-free”性质，但在高维数据下，因理论性质的缺失及所遭遇到的维数诅咒问题，使这类经典的检验方法备受冷落。本课题旨在借助于统计学中著名的投影追踪思想去解决这些问题。主要研究内容有：一是在总体意义下利用投影构造能够刻画两样本分布差异的非参数度量，研究新度量的总体性质；二是在样本意义下，基于新度量构造检验统计量，并在样本量趋于无穷大但维数固定情况下，研究新统计量的大样本性质；三是在样本量固定但维数发散的情况下，研究新统计量的高维性质，并进一步考虑如何在高维情形下对新检验进行修正去提高检验功效。

Testing for homogeneity of two independent random samples is one of the most important hypothesis testing problems in statistics. The classical nonparametric tests such as the well-known Kolmogorov-Smirnov, Cramér-von Mises and Anderson-Darling tests are built on the empirical distribution functions. Although such tests have many desirable properties such as robustness and “distribution-free” properties in the one-dimensional two-sample framework, they are far less popular in high dimensions, mainly due to the lack of theoretical properties and the curse of dimensionality. This project aims to solve these problems by means of the well-known idea of projection pursuit in the statistics community. The main contents of the research are as follows. First, we will use projections to construct nonparametric measures that characterize the differences between two populations in distributions, and further study population properties of the measures. Second, we will use the measures to formulate test statistics and further study the asymptotic behaviors of the proposed test statistics under the “large sample size, fixed dimension” paradigm. Third, we will study the asymptotic behaviors of the proposed test statistics under the “fixed sample size, large dimension” paradigm and further consider how to correct the new tests in high-dimensional situations to enhance their power performance.