On 23/06/14 10:25, Lawrence Mitchell wrote:
One possible scenario where you might wish for bit-reproducibility in an actual simulation run is computing adjoints of chaotic systems with large Lyapunov exponents, where you really would like your replayed forward model to be deterministic. However, I am by no means an expert, Patrick (if he's lurking) may have other opinions.
If your forward model is chaotic, your adjoint solve will blow up exponentially backwards in time. (After all, the tangent linearisation blows up forwards in time.) In that case, there may be sensible derivatives of functionals, but using the standard linearisation argument won't compute them, and since the whole endeavour is pointless I don't think that matching bit-for-bit reproducibility in the forward model is necessary or recommended. There's been some very interesting work by Qiqi Wang from MIT on the topic of computing gradients of chaotic systems, see e.g. http://arxiv.org/abs/1204.0159. Cheerio, Patrick