Distribution/correlation-free test for two-sample means in high-dimensional functional data with eigenvalue decay relaxed

We propose a methodology for testing two-sample means in high-dimensional functional data that requires no decaying pattern on eigenvalues of the functional data. To the best of our knowledge, we are the first to consider such a problem and address it. To be specific, we devise a confidence region for the mean curve difference between two samples, which directly establishes a rigorous inferential procedure based on the multiplier bootstrap. In addition, the proposed test permits the functional observations in each sample to have mutually different distributions and arbitrary correlation structures, which is regarded as the desired property of distribution/correlation-free, leading to a more challenging scenario for theoretical development. Other desired properties include the allowance for highly unequal sample sizes, exponentially growing data dimension in sample sizes and consistent power behavior under fairly general alternatives. The proposed test is shown uniformly convergent to the prescribed significance, and its finite sample performance is evaluated via the simulation study and an implementation to electroencephalography data.

Automated summary

We propose a methodology for testing two-sample means in high-dimensional functional data that does not require a decaying pattern on eigenvalues of the functional data. To the best of our knowledge, we are the first to consider and address such a problem. Specifically, we devise a confidence region for the mean curve difference between two samples, which directly establishes a rigorous inferential procedure based on the multiplier bootstrap. In addition, the proposed test permits the functional observations in each sample to have mutually different distributions and arbitrary correlation structures, which is regarded as the desired property of distribution/correlation-free, leading to a more challenging scenario for theoretical development.