Nicholas Krämer
Nicholas Krämer
I think the simple solution here is to just commit to the mean in matrix format. I also have a feeling that life gets easier if for the reference solvers...
It turns out that the JVP *kind of* works, but the tests are very flaky. Therefore we skip them (#503) until this bug has been resolved for good.
It _might_ have been resolved by the linked PR. Worth checking at least.
Todo: mention the initial draft in the changelog.
Maybe the easiest way of implementing this would be a `ODESolution.discretise_offgrid(ts)` method, which returns another ODE solution but with different states and conditionals. On this, one could then use `.sample()`...
My initial guess would be the following: the part of the code you linked computes a QR decomposition of a stack of covariance matrices, and this (or any) QR decomposition...
Hi :) This issue has not received any updates for a while, and without any open todos, I'll close it. But please feel free to reopen if something is still...
Hi! Do I understand correctly that you want to have two solves with two different vector fields, but the second solve starts at the first one's terminal value? That should...
Good to hear! Please keep me posted on how it goes. If it works for the filter, it works for the smoother; the solver returns all backwards transitions, and after...
Hi, Thanks for the update! I am not entirely sure I understand correctly. Do you want to solve an ODE with a discontinuous vector field? If so, plugging the (discontinuous)...