I am recently looking into automatic differentiation, as implemented by
ForwardDiff.jl. The manual says functions that call blas are not supported, therefore any functions containing
eig are not supported. I think I partly understand the reason behind this statement.
However, I noticed that
eigh (eigen decomposition of Hermitian matrix) is supported by
AlgoPy (https://github.com/b45ch1/algopy), a package providing automatic differentiation in python. This is quite unexpected since derivatives of eigenvectors are not well defined since eigenvectors can be multiplied by arbitrary number. I wonder how
AlgoPy supports eigen decomposition and whether it’s doable in julia.