Hi,

I have code that makes use of SelfAdjointEigenSolver<Eigen::MatrixXcd>. I seem to be getting numerically different results for the eigenvectors (very small) depending on whether I compile on the Linux or Windows platforms. Is this expected, and if not, how can I get better agreement?

Thanks

Martin

Very small differences could be explained if one is a 32 bits system while the other is 64 bits (or two 32 bits systems but one with SSE2 enabled). This is because of the extended precision (80bits) of the FPU registers (used in 32 bits) compared to 32 (float) or 64 (double) bits for SSE. Agressive optimisation flags could also explain some differences.