They might be talking about complex step diff: https://www.hedonisticlearning.com/posts/complex-step-differ...

tldr: f‘(x0)≈Im(f(x0+hi))/h

If that is what they meant, that's not a good idea.

Choice of h dictates precision (less bias than doing it purely in the real domain, but still .. a nontrivial choice). And also requires your functions to be analytical - which many aren't.

Whereas autodiff is precise .

I know the contribution of the higher order imaginary terms are negligible in the result, but dual numbers are the more appropriate choice in this instance correct? since the higher order terms will vanish.