with_logabsdet_jacobian(::InverseFunctions.FunctionWithInverse, x)

[cscherrer]

with_logabsdet_jacobian(f, x) currently breaks if f::InverseFunctions.FunctionWithInverse. Today @oschulz and I discussed the need for such a method. Could this be added? It will require a dependency on InverseFunctions.jl, but that's very light.

[devmotion]

Sure that could be done.

But I think it does not scale and points to the design issues I raised in JuliaMath/InverseFunctions.jl#25. By wrapping a function in a FunctionWithInverse you break all downstream dispatches for the unwrapped function but it seems undesirable and unfeasible to add support for FunctionWithInverse to every single package. So I think the two general solutions are to (a) remove FunctionWithInverse and setinverse or (b) to unwrap it every time you call into another package that might dispatch on the function (with an unwrap function as suggested in JuliaMath/InverseFunctions.jl#25 (comment)). Since I assume (a) might not be an option for you, probably (b) is the best alternative.

[oschulz]

@cscherrer, we could dispatch on PushforwardMeasure{<:InverseFunction}, right? We just need it for the density method(s) after all.

[cscherrer]

I kind of like the unwrap idea - then we don't need much in the way of special cases, and it's very little additional code. Ok to close this unless either of you think there's more to discuss. Thanks!