posterior smoother subdraws

force solving the model for smoother, by relaxing qz_criterium, when taking posterior subdraws (likelihood has been already evaluated without error for the same parameter vector).

changed milestone to %5.x

@rattoma When can this happen? I thought that subsequent functions have to properly inherit the qz_criterium

I see two ways out here:

1. We indeed go for the solution outlined above:

computing the smoother with reduced state space (as for likelihood) and retrieve ex-post the remaining static variables. I have no clue how involved this would be (also in terms of the timeframe)

2. Adjust this PR to try to adjust qz_criterium (potentially iteratively in steps of factor 10), but only conditional on the value of info (I guess 3,4, and 5), regardless of whether this applies to the prior or posterior.

option 1. would be ready for PR. while doing so, I also spotted the 2 bugs, which I fixed in !1847 (closed)

option 2. I would apply it only for posterior (we do not know if likelihood is OK there: unless we check if likelihood is computable if info>0), I would not increase more than in two steps by factor 10.

Then I would prefer option 1. It seems less error prone and hackish.

Seeing the smoother_redux-implementation, I don't think it's feasible to use this as the default. In !1860 (merged) I instead fail gracefully.

agree, thanks!

requested review from @JohannesPfeifer

assigned to @sebastien

mentioned in merge request !1860 (merged)

resolved all threads

Superseded by !1860 (merged)

closed