FiniteDiff.finite_difference_gradient changes input when it errors?

Specific Domains
1

I ran into a surprising thing today. Consider the function sumlog:

function sumlog(x)
    sum(log(i) for i in x)
end

I’m interested in taking the gradient of sumlog at x:

using FiniteDiff

x = [1e-6, 1e-6]

FiniteDiff.finite_difference_gradient(sumlog, x)
#DomainError with -5.05545445239334e-6:
#log will only return a complex result if called with a complex argument. Try log(Complex(x)).

Naturally it errors, what surprised me is that now x changed:

x
#2-element Array{Float64,1}:
# -5.05545445239334e-6
#  1.0e-6

Is that expected? I’m using FiniteDiff v2.5.0.

The gradient error arises because I think it tries to compute the symmetric difference quotient and in doing so it reaches for a point less than 0, so is there a way to request the gradient using simply the difference quotient (where the derivative is approximated using points greater than x)?

2

Can you open an issue on this?

3

just did, do you know if FiniteDiff can compute non-symmetric difference quotients?

4

Not for the gradient IIRC