Imaginary part of complex gradient is dropped

[sethaxen]

In a number of cases at least, it seems that the imaginary part of the gradient for complex inputs and real scalar outputs is dropped. A few examples:

julia> using FiniteDifferences

julia> x = randn(ComplexF64, 3)
3-element Array{Complex{Float64},1}:
   0.5296680539510132 + 0.508966079006392im
  0.45109029662066524 + 1.4033240412225767im
 -0.49491015411581324 + 0.15307916651762313im

julia> FiniteDifferences.grad(central_fdm(5, 1), x -> sum(abs2, x), x)[1] # expected: 2x
3-element Array{Float64,1}:
  1.0593361079020813
  0.9021805932413423
 -0.9898203082316152

julia> 2x
3-element Array{Complex{Float64},1}:
  1.0593361079020265 + 1.017932158012784im
  0.9021805932413305 + 2.8066480824451534im
 -0.9898203082316265 + 0.30615833303524626im

julia> FiniteDifferences.grad(central_fdm(5, 1), abs2, 1.0+2.0im)[1] # expected: 2.0 + 4.0im
1.9999999999999962

Also, am I supposed to see this error?

FiniteDifferences.grad(central_fdm(5, 1), abs2, 1+2im)[1] # expected: 2.0 + 4.0im
ERROR: InexactError: Int64(0.9879814839007404)

[willtebbutt]

Thanks for raising the issue. What version of FiniteDifferences are you running?

[sethaxen]

I saw this with both the most recent release (v0.9.6) and master on Julia 1.4.

[sethaxen]

I just checked a few releases, and it looks like the return type became real in v0.9.3, but before that the imaginary part was just 0 going back at least to v0.8.

[willtebbutt]

Closing this with #76 . @sethaxen please feel free to re-open if it turns out that this hasn't solved your issue.

[nickrobinson251]

in #76 (#76 (comment)) it was decided this was a breaking change (rather than a bug fix), so the change is available in a new breaking release v0.10.0 (280c4fa)