A covariance matrix with a special pattern (e.g., sparsity or block structure) is essential for conducting multivariate analysis on high-dimensional data. Recently, a block covariance or correlation pattern has been observed in various biological and biomedical studies, such as gene expression, proteomics, neuroimaging, exposome, and seed quality, among others. Specifically, this pattern partitions the population covariance matrix into uniform (i.e., equal variances and covariances) blocks. However, the unknown mathematical properties of matrices with this pattern limit the incorporation of this pre-determined covariance information into research. To address this gap, we propose a block Hadamard product representation that utilizes two lower-dimensional "coordinate" matrices and a pre-specific vector. This representation enables the explicit expressions of the square or power, determinant, inverse, eigendecomposition, canonical form, and the other matrix functions of the original larger-dimensional matrix on the basis of these "coordinate" matrices. By utilizing this representation, we construct null distributions of information test statistics for the population mean(s) in both single and multiple sample cases, which are extensions of Hotelling's $T^2$ and $T_0^2$, respectively.
翻译:针对高维数据的多元分析,具有特殊模式(如稀疏或块结构)的协方差矩阵至关重要。最近,各种生物和生物医学研究中观察到了块协方差或相关模式,如基因表达、蛋白组学、神经影像、暴露组和种子质量等。具体而言,此模式将总体协方差矩阵分为一致(即等方差和协方差)的块。然而,这种模式的未知数学属性限制了该预先确定的协方差信息在研究中的应用。为了解决这一问题,我们提出了块Hadamard乘积表示,利用两个低维“坐标”矩阵和一个预先指定的向量。该表示基于这些“坐标”矩阵,使得原始较高维矩阵的平方或乘幂、行列式、逆、特征分解、规范形式和其他矩阵函数的显式表达成为可能。通过利用该表示,我们构建了单样本和多样本情况下的总体均值信息检验统计量的空分布,分别是Hotelling's $T^2$和$T_0^2$的扩展。