Structure-Preserving Invariant Interpolation Schemes for Invertible Second-Order Tensors (2211.16507v1)

Abstract: Tensor interpolation is an essential step for tensor data analysis in various fields of application and scientific disciplines. In the present work, novel interpolation schemes for general, i.e., symmetric or non-symmetric, invertible square tensors are proposed. Critically, the proposed schemes rely on a combined polar and spectral decomposition of the tensor data $\boldsymbol{T}!!=!!\boldsymbol{R}\boldsymbol{Q}^T !! \boldsymbol{\Lambda} \boldsymbol{Q}$, followed by an individual interpolation of the two rotation tensors $\boldsymbol{R}$ and $\boldsymbol{Q}$ and the positive definite diagonal eigenvalue tensor $\boldsymbol{\Lambda}$ resulting from this decomposition. Two different schemes are considered for a consistent rotation interpolation within the special orthogonal group $\mathbb{SO}(3)$, either based on relative rotation vectors or quaternions. For eigenvalue interpolation, three different schemes, either based on the logarithmic weighted average, moving least squares or logarithmic moving least squares, are considered. It is demonstrated that the proposed interpolation procedure preserves the structure of a tensor, i.e., $\boldsymbol{R}$ and $\boldsymbol{Q}$ remain orthogonal tensors and $\boldsymbol{\Lambda}$ remains a positive definite diagonal tensor during interpolation, as well as scaling and rotational invariance (objectivity). Based on selected numerical examples considering the interpolation of either symmetric or non-symmetric tensors, the proposed schemes are compared to existing approaches such as Euclidean, Log-Euclidean, Cholesky and Log-Cholesky interpolation. In contrast to these existing methods, the proposed interpolation schemes result in smooth and monotonic evolutions of tensor invariants such as determinant, trace, fractional anisotropy (FA), and Hilbert's anisotropy (HA)...{continued see pdf}