expAv undeflow/overflow tricks

[Paul vdb]

Hello,

I'm trying to use expAv for a markov modulated poisson process with lots of states (2~000). I'm getting lots of overflow/undeflow issues. I've tried to dig deeper by reading through the paper that is referenced in the help file and through the code to try and understand it better: Direct statistical inference for finite Markov jump processes via the matrix exponential | Computational Statistics

Are there any tricks to reduce underflow/overflow errors with this function? The issue is as I try to add more time expAv((Q-Lambda)*t, v). I can do alright by discretizing time and repeated calls to expAv,  but hoping for something more robust. I tried exp(-rho)*expAv(rhoP*t, v), but I think this mirrors what is being called via normalization.

Thanks so much,

Paul

[Kasper Kristensen]

It's a problem in RTMB that needs to be fixed. I've opened a github issue.