Incorrect behavior of max(d::Dual, ::Float64) when partials are -Inf

[jeffreyesun]

See the following MWE:

using ForwardDiff
a = ForwardDiff.Dual(-1.0, -Inf)
a = max(a, 0.0) #Dual{Nothing}(0.0,NaN)

The correct output should be Dual{Nothing}(0.0,0.0), as the constant 0.0 does not depend on the x. This was causing my code to fail as the solver I was using couldn't handle the NaN values.

[devmotion]

Did you try to enable NaN-safe mode? https://juliadiff.org/ForwardDiff.jl/dev/user/advanced/#Fixing-NaN/Inf-Issues

[jeffreyesun]

I tried it, but the problem persists. I should note that the derivative in this case is not undefined; it should be equal to exactly zero.

[jeffreyesun]

The code being called appears to be define_binary_dual_op in src/dual.jl, but I can't make heads or tails of it.

There is a simple workaround, which I'm using for now:

if a < 0
    a = zero(a)
end

[mcabbott]

It's knotty but I think it ends up here:
https://github.com/JuliaDiff/DiffRules.jl/blob/489e2942e10776c96ab70c5044f595951bbbcaab/src/rules.jl#L93

I don't think there's any reason those should be one(x) not just true/false, and since false is a strong zero, I believe it would fix the problem.

This is what ChainRules has, BTW:

https://github.com/JuliaDiff/ChainRules.jl/blob/5818173b31fac9d358acda21f5751978a5dcb2e5/src/rulesets/Base/fastmath_able.jl#L220

[devmotion]

I opened a PR: JuliaDiff/DiffRules.jl#87