Testing High Dimensional Mean Under Sparsity

Motivated by the likelihood ratio test under the Gaussian assumption, we develop a maximum sum-of-squares test for conducting hypothesis testing on high dimensional mean vector. The proposed test which incorporates the dependence among the variables is designed to ease the computational burden and to maximize the asymptotic power in the likelihood ratio test. A simulation-based approach is developed to approximate the sampling distribution of the test statistic. The validity of the testing procedure is justified under both the null and alternative hypotheses. We further extend the main results to the two sample problem without the equal covariance assumption. Numerical results suggest that the proposed test can be more powerful than some existing alternatives.