Compatability of Eigsolve for complex fp128

[htweinberger]

I currently don't believe that KrylovKit.eigsolve supports generic AbstractFloat types with greater than 64fp precision. I am using Quadmath.jl or DoubleFloats.jl as the element types. My matrix is currently Matrix{Complex{Double64}} (alias for Array{Complex{DoubleFloat{Float64}}, 2}) and I believe that the function fails at the schur method stage. I have the following error.

no method matching hschur!(::SubArray{Complex{…}, 2, Matrix{…}, Tuple{…}, false}, ::SubArray{Complex{…}, 2, Matrix{…}, Tuple{…}, false})

Are there any plans to introduce Eigsolve methods for fp128? Thank you !

[lkdvos]

Thanks for reaching out!

Unfortunately, for now you are correct in thinking that we only support the BLAS/LAPACK real and complex floating point types as the underlying scalars.
There are no fundamental reasons for this, except that we have to be able to decompose some matrices through some LAPACK calls. (hschur! for eigsolve).
It should be somewhat straightforward to support other element types, as long as there is an implementation of these LAPACK methods. (For example, https://github.com/JuliaLinearAlgebra/GenericLinearAlgebra.jl/)
While this is definitely something that we would eventually like to support, it hasn't been too much of a priority. If you are interested in contributing, that would definitely be greatly appreciated (and happy to help of course!)

[lkdvos]

@kshyatt perhaps something that could be part of MatrixAlgebraKit.jl?