Closed frankschae closed 2 years ago
I must admit, it's still not clear to me whether it makes sense to differentiate through an SDE solver but I relaxed the constraints in 30046f2930c740afc5d105bfabfed5d3ccd13e26 and it seems to work now.
It makes sense to differentiate through the SDE solver for parameter estimation purposes. Though I'm not sure the Levy area calculation needs to be differentiated for the same reason why dropping the RNG on the randn works (something @frankschae and I are making much more precise in an upcoming paper BTW), so the right answer may be to drop duals before hitting this code. But anyways, since it can be supported that's good.
(requires https://github.com/SciML/StochasticDiffEq.jl/pull/459 or the overload from https://github.com/SciML/StochasticDiffEq.jl/issues/458)
AbstractFloat in https://github.com/stochastics-uni-luebeck/LevyArea.jl/blob/68c5cb08ab103b4dcd3178651f7a5dd9ce8c666d/src/fourier.jl#L21 is probably a bit too constrained.