I have been using some stuff from Distributions.jl and ConjugatePriors.jl to (amongst other things) sample from a NormalWishart distribution. I am getting a whole bunch of problems that trace back to Base.LinAlg.ishermitian detecting a matrix that is *definitely Hermitian* as being asymmetric. When I investigate it I see something like:

julia> A

2×2 Array{Float64,2}:

0.113637 -0.0338744

-0.0338744 0.0124418

julia> Hermitian(A)

2×2 Hermitian{Float64,Array{Float64,2}}:

0.113637 -0.0338744

-0.0338744 0.0124418

julia> A-Hermitian(A)

2×2 Array{Float64,2}:

0.0 0.0

-6.93889e-18 0.0

julia> ishermitian(A)

false

Are there no checks on relative error in ishermitian? This seems like something that would cause problems elsewhere, and be relatively simple to fix.