[General]: Use np.linalg.eigh instead of np.linalg.eig for symmetric matrices #439

Feedback or Inquiry

CodeEntropy uses np.linalg.eig on matrices that are always symmetric: the moment of inertia tensors used for principal axes, and the force/torque covariance matrices. eig is the general-purpose solver, and for symmetric input it is the wrong tool. In particular, for (near-)degenerate eigenvalues it does not guarantee orthogonal eigenvectors, so the "principal axes" can be skewed, and the result can differ between machines and LAPACK builds.

Proposal: use np.linalg.eigh wherever the input is symmetric, including in place of MDAnalysis's principal_axes() (which also uses eig in 2.10) until we depend on a version that uses eigh.