Unravelling the mystery of #861 about powermodels.jl

[odow]

So Julia 1.9 is noticeably less-accurate at competing inv of a complex number. And this causes problems with the susceptance matrix. It's probably also the cause of the numeric issue on Windows.

Julia 1.9

julia> imag(inv(0.00064 + 0.0064im))
-154.70297029702968

julia> x = imag(inv(big"0.00064" + big"0.0064" * im))
-154.7029702970297029702970297029702970297029702970297029702970297029702970297017

julia> x - imag(inv(0.00064 + 0.0064im))
-2.335645428439141205041715414217202970297029702970297029702970170069474548472957e-14

Julia 1.6

julia> imag(inv(0.00064 + 0.0064im))
-154.7029702970297

julia> x = imag(inv(big"0.00064" + big"0.0064" * im))
-154.7029702970297029702970297029702970297029702970297029702970297029702970297017

julia> x - imag(inv(0.00064 + 0.0064im))
5.065255146012595384427816560952970297029702970297029702970298299305254515270434e-15

Diff

julia> -154.70297029702968 - -154.7029702970297
2.842170943040401e-14

[odow]

Upstream change seems to be JuliaLang/julia#47255

[odow]

The two Float64s are neighbors

julia> prevfloat(-154.70297029702968)
-154.7029702970297

so it seems unlikely that they'll revert the upstream change. I wonder if we should compute the matrix in better precision on the PowerModels side.