search

LSSTDESC / augur

DESC forecasting and inference validation tool
MIT License
4 stars 2 forks source link

Prenormalize vector parameter to use same strategy as cosmosis #68

Open fjaviersanchez opened 1 day ago

fjaviersanchez commented 1 day ago

Cosmosis pre-normalizes the input parameters before taking the numerical derivatives, so the step size depends on the parameter. Add an option to do something similar, which will allow for easier validations.

fjaviersanchez commented 1 day ago

Preliminary implementation in #69

fjaviersanchez commented 1 day ago

On top of this, there's some information in this paper: https://arxiv.org/pdf/1606.03451. The idea is to select a step in each parameter that allows to see a change in the likelihood larger than 0.1 (so the differentiation is stable), and in fact, they iterate so the step in each parameter is such that the change in the chi-square is ~1.