Following discussions with @dgondelach, I’ve opened a couple of issues (1462 and 1463) regarding options for state transition matrix performance, starting with drag and geopotential. The idea is to use a simpler model for the derivatives, e.g. a less complex atmospheric model as some are very expensive with automatic differentiation.

Could you provide a discussion of the limitations of that approach? Perhaps a reference that discusses the impacts on accuracy and convergence?

Rice tried the “hybrid” approach (different models for state and STM) in 1967 in [1] and concluded that it causes “rapid divergence”. It’s a good read. Based on his work, I wouldn’t recommend including the hybrid approach in a library like Orekit. But perhaps in the last 50+ years someone else has figured out how to address the divergence issues.

It’s a good idea, but we need to be very careful with current orekit users.
If it’s just an option and the default behavior (i.e. calculation) remains as it is now, I think it’s acceptable to introduce this feature for users who favor calculation time over accuracy.
On the other hand, current users shouldn’t have to do anything - nothing should change for them. If an option is available, the action to be taken (i.e. a method to be called or other) must be taken by the users wishing to limit performance and increase computation speed.

I agree with Evan here.
From my experience, simplifying the force models when computing derivatives has a bad influence on convergence of orbit determination. It may help at start, while we are still far from the solution and just want to come closer. But when we are near convergence and globally derivatives are close to zero (because we are looking for the minimum of a cost function), it seems to me that having simplified derivatives just either makes the algorithm lose its path as the Jacobian is inconsistent with the evolution of the value or it makes the algorithm converge to the wrong solution.

First of all, as Bryan said, it would be an option, and certainly not the default one.

As for accuracy, well it’s always a trade-off right? I’m pretty sure I’ve seen the option to cut off the geopotential order for STMs in GMAT, although I can’t get my hand on a proper link at the moment. As for drag, the proposed approximation is only on the density, which at the moment is with finite differences by default so I wouldn’t say it’s much more accurate. Besides, orbit determination is not the only place where one needs covariance matrices and/or STMs, and it’s not always in an iterative process requiring convergence. I believe users are not all doing operational flight dynamics. Anyhow if it doesn’t make consensus I’ll ditch the issues from the 12.2 milestone for now.