Jouni Helske

Thanks, @kingaa for your helpful comments. I have now updated to the package based on your suggestions, mainly by improving the documentation and adding a new plot method. I opted...

Yes this might be nice addition, base R has an C function for this [invpartrans](https://github.com/wch/r-source/blob/af7f52f70101960861e5d995d3a4bec010bc89e6/src/library/stats/src/arima.c#L525) and it could be ported easily to Fortran (or called as such). At the moment...

Are you using the latest CRAN version or Github version? There are some memory issues with the current CRAN version related to these recent changes (already resubmitted the fix), the...

Strange, everything looks ok above. Are you sure you are using the same model you uploaded to the dropbox? Changing the prior variance of the states from diffuse to something...

Are there differences in operating systems? What if you change H_tol to a very large value like 1e300 or something similar? Or to a very small value just to see...

Very strange, I am also running Win10. I reran all the checks with valgrind etc. to check memory errors, but those didn't find any issues. What does this return: ```...

Thanks, `logLik(model)` does not need to be same as `logLik(approxSSM(model))`, the former is the approximate log-likelihood of the non-Gaussian model, whereas the former is the likelihood of the approximating Gaussian...

Ok, this was tricky. There is a check in the diffuse filtering recursions regarding whether the prediction error variance F is zero but in order to deal with numerical issues...

There are often multiple local optimums in the likelihood surfaces of state space models, so in principle, some of those spikes could be real as well, and thus it is...

Unfortunately KFAS does not directly support these as they need a bit different type of approach. However, for t-distribution I think in Durbin & Koopman (2012) book there is a...