A Riemannian covariance for manifold-valued data

Meshal Abuqrais,Davide Pigoli
2024-10-09
Abstract:The extension of bivariate measures of dependence to non-Euclidean spaces is a challenging problem. The non-linear nature of these spaces makes the generalisation of classical measures of linear dependence (such as the covariance) not trivial. In this paper, we propose a novel approach to measure stochastic dependence between two random variables taking values in a Riemannian manifold, with the aim of both generalising the classical concepts of covariance and correlation and building a connection to Fréchet moments of random variables on manifolds. We introduce generalised local measures of covariance and correlation and we show that the latter is a natural extension of Pearson correlation. We then propose suitable estimators for these quantities and we prove strong consistency results. Finally, we demonstrate their effectiveness through simulated examples and a real-world application.
Statistics Theory
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is: how to define and estimate the dependence between random variables in non - Euclidean spaces, especially on Riemannian manifolds. Specifically, the paper proposes a new method to measure the stochastic dependence between two random variables taking values on Riemannian manifolds, aiming to generalize the concepts of classical covariance and correlation coefficient and establish a connection with the Fréchet moments of random variables on manifolds. ### Background of the Paper - **Complexity of the Problem**: The non - linear nature of non - Euclidean spaces makes the generalization of classical linear dependence measures (such as covariance) non - trivial. - **Limitations of Existing Methods**: Existing methods (such as distance covariance and dependence measures in spherical regression) either depend on specific manifold geometries or are unable to distinguish between linear and nonlinear dependence. - **Research Motivation**: In many practical applications, data naturally belongs to non - Euclidean spaces, such as shape data in medical imaging, network data in linguistics, etc. Therefore, it is of great significance to develop statistical methods applicable to these data. ### Contributions of the Paper - **New Method**: The paper introduces Riemannian covariance and Riemannian correlation, which not only generalize the concepts of classical covariance and correlation coefficient but also provide a clear geometric interpretation. - **Theoretical Results**: It is proved that the proposed estimators are consistent under certain conditions. - **Application Verification**: The effectiveness of the new method is verified through simulation experiments and real data (such as vector electrocardiogram data). ### Formula Representation - **Riemannian Covariance Matrix**: \[ \Sigma_p(X, Y)=\mathbb{E}[\log_p X(\log_p Y)^T]-\mathbb{E}[\log_p X]\mathbb{E}[\log_p Y]^T \] - **Riemannian Covariance**: \[ \text{Rcov}_p(X, Y)=\text{tr}(\Sigma_p(X, Y)) \] - **Riemannian Correlation Matrix**: \[ R_p(X, Y)=\frac{\Sigma_p(X, Y)}{\sqrt{\text{tr}(\Sigma_p(X, X))\text{tr}(\Sigma_p(Y, Y))}} \] - **Riemannian Correlation Coefficient**: \[ \text{Rcorr}_p(X, Y)=\text{tr}(R_p(X, Y)) \] ### Conclusion The paper successfully proposes a new method to measure and estimate the dependence between random variables on Riemannian manifolds, fills this gap in this field, and provides a new tool for dealing with non - independent samples (such as time - series data). Through theoretical analysis and experimental verification, this method has good consistency and effectiveness.