The spatial sign covariance operator: asymptotic results and applications (Q1733274)

Let $\mathcal{H}$ be a separable Hilbert space with inner product $\langle\cdot,\cdot\rangle$ and norm $\Vert u\Vert=\langle u,u\rangle^{1/2}$. Let $X$ be a random element taking values in $\mathcal{H}$, the main object of study in the present paper is the spatial sign covariance operator of $X$ centred at $\nu$, defined by \[ \Gamma^s(\nu)=E\left\{s(X-\nu)\otimes s(X-\nu)\right\}, \] where $s(u)=u/\Vert u\Vert$ for $u\not=0$, $s(0)=0$, and $\otimes$ is the tensor product on $\mathcal{H}$. This spatial sign covariance operator is designed to offer a robust alternative to the usual sample covariance operator in functional data analysis. The corresponding sample version of $\Gamma^s(\nu)$ is given by \[ \widehat{\Gamma}^s(\nu)=\frac{1}{n}\sum_{i=1}^ns(X_i-\nu)\otimes s(X_i-\nu), \] where $X_1,\dots,X_n$ are independent, each having the same distribution as $X$. Let $\mu$ be a given location parameter of the underlying process, the authors prove the convergence of $\widehat{\Gamma}^s(\widehat{\mu}_n)$ to $\Gamma^s(\mu)$ in a suitable sense, as $n\rightarrow\infty$, assuming appropriate convergence of $\widehat{\mu}_n$ to $\mu$. This framework covers several different choices of $\widehat{\mu}_n$ which have previously been considered in the literature. Under some boundedness assumptions, the authors also establish the asymptotic distribution of $\sqrt{n}\left\{\widehat{\Gamma}^s(\widehat{\mu}_n)-\Gamma^s(\mu)\right\}$, and in particular asymptotic normality of this difference. Two applications are given. The first is to asymptotic properties of the spherical principal direction estimators. The second application is to a test for equality between spatial covariance operators for two different populations. That is, a test of a null hypothesis of the form $H_0: \Gamma_1^s=\Gamma_2^s$ against an alternative of the form $H_1: \Gamma_1^s\not=\Gamma_2^s$. The test statistic used is $n\Vert \widehat{\Gamma}_1^s-\widehat{\Gamma}_2^s\Vert_\mathcal{F}$, where $\mathcal{F}$ is the Hilbert space of Hilbert-Schmidt operators over $\mathcal{H}$. The asymptotic distribution of this test statistic is derived, and a simulation study is carried out to illustrate the performance of the proposed testing procedure.