Distance-based and RKHS-based dependence metrics in high dimension (Q1996774)

The authors consider conditions under which \[\mathrm{d}\mathrm{Cov}_n^2\left(\mathbf{X},\mathbf{Y}\right)\approx\frac{1}{\tau}\sum_{i=1}^p\sum_{j=1}^q\mathrm{cov}_n^2\left(\mathcal{X}_i,\mathcal{Y}_j\right).\] Here \(\mathrm{d}\mathrm{Cov}_n^2\left(\mathbf{X},\mathbf{Y}\right)\) denotes the umbiased sample distance covariance, \[ \mathbf{X}=\left(X_1,X_2,\ldots,X_n\right)^\intercal =\left(\mathcal{X}_1,\mathcal{X}_2,\ldots,\mathcal{X}_p\right), \] \[ \mathbf{Y}=\left(Y_1,Y_2,\ldots,Y_n\right)^\intercal=\left(\mathcal{Y}_1,\mathcal{Y}_2,\ldots,\mathcal{Y}_q\right) \] are the sample matrices, where \(X_k\mathop{=}\limits^{d}X\) and \(Y_k\mathop{=}\limits^{d}Y\) are independent samples of two random vectors \(X=(x_1,x_2,\ldots,x_p)\in\mathbb{R}^p\) and \(Y=(y_1,y_2,\ldots,y_q)\in\mathbb{R}^q\) with finite componentwise second moments, and \(\mathcal{X}_i\), \(\mathcal{Y}_j\) are the componentwise samples. In addition, \(\tau\) denotes a quantity depending on the marginal distributions of \(X\) and \(Y\) as well as \(p\) and \(q\), and \(\mathrm{cov}_n\left(\mathcal{X}_i,\mathcal{Y}_j\right)\) is an unbiased sample estimate of \(\mathrm{cov}(x_i,y_j)\). The above approximate equality is considered as \(p,q\) tend to infinity, and \(n\) can either be fixed or grows to infinity at a slower rate. The paper is the first work on the connection between sample distance covariance and sample covariance.