Triggered by this thread I was wondering whether there exists a case study showcasing the superior scaling of the adjoint ODE method and / or reference models / posteriors and configurations?
I’m guessing a good candidate would be some kind of linear chemical reaction network with unknown initial states, parameters and measurement accuracy? It could be made abitrarily large, would scale quite badly using the forward methods and we would know the exact solution (via matrix exponentials)?
I have already coded up a fake data example, which can be parametrically increased in size. With that example I did run benchmarks showing the far superior scaling of adjoint ODE solver and I also ran SBC with it. Everything went ok…but I haven’t repeated these experiments with the final production version of the solver. These codes are buried in the PR comments on the merged adjoint ODE thread, which is huge… but it’s there in the digital nirvana.
Your suggestion to use a system which we can benchmark against a matrix exponential sounds great.