I kind of like this idea. It has the major advantage that code that happened to work for some data won’t break on some other data where a numerical instability gave slightly inexact results since you must signal the intention that a matrix be hermitian explicitly. I also like the idea of having some kind of debug vs. release flags to turn on/off checking approximate hermitianness. Want to open an issue to discuss this as a potential LinearAlgebra 2.0 API change?
3 Likes