isnan(Dual(1.0,NaN)) == false

[goretkin]

julia> isnan(Dual(1.,NaN))
false

This didn't come up as an issue for me, but I'm curious if this is a deliberate choice.

compare

julia> isnan(1. + (NaN)im)
true

[mlubin]

Not a deliberate choice, but just following the convention that comparison operations should apply only to the real part. If there's a solid reason to change the behavior, we can definitely do so.

[mlubin]

This isn't a great citation, but http://www.coin-or.org/CppAD/Doc/doxydoc/html/nan_8hpp_source.html states that isnan(x) should be true iff x != x is true. Since we override !=, the current behavior is consistent with that definition.