search

halasadi / MAPS

software for the MAPS method
MIT License
24 stars 6 forks source link

eems2 likelihood produces nan #2

Closed halasadi closed 9 years ago

halasadi commented 9 years ago

(1) poisson likelihood produces nan values -- should not happen.

(2) In general, need to test the entire eems2 code throughly.

halasadi commented 9 years ago

I found the "bug". Actually, it is not a bug. The krylov subspace approximation (for k = 10) was not working so well and produced negative values. This is a bit worrying. For k = 20, it worked well. Will post more thorough analysis of the Krylov approximation.

halasadi commented 9 years ago

Tried a technique for Krylov approximation from Sidje, 1998. Significantly slower but more accurate. For large migration rates, the simple approach I implemented previously was in-accurate. I think the accuracy of the Krylov approx depends on how the long the markov chain reached stationarity so for large m, it took much longer. It seems like this new technique works reasonably well for bigger migration rates but not too big (m < 0.1). Probably OK for realistic parameters.