Discrepancy in probability/chi^2

CQMP / Maxent

GNU General Public License v2.0

16 stars 14 forks source link

Discrepancy in probability/chi^2 #51

Open ryanlevy opened 3 years ago

ryanlevy commented 3 years ago

Reported by Manuel Zingl and Khaldoon Ghanem (thanks both!) there are some potential subtle problems with definitions. The code seems to use Q=chi^2 / Ndat - alpha*S in the minimization routine rather than the reported Q=1/2chi^2 - alpha*S

Current options:

Use the Ndat version of Q and fix Q and chi2 functions in the code
- we should test how things will change under this
Remove the (re)scaling
- this seems to reasonably match e.g. MEM for the Bryan method

jpfleblanc commented 3 years ago

So what you think is most correct. Would be nice to see the impact of this on the example test cases as a minimum.

ryanlevy commented 3 years ago

case1.pdf case2.pdf case3.pdf

Khaldoon graciously provided three test cases along with their outputs from MEM i.e. Jarrell Code. to be more explicit:

Original Code is the current release
Patched (1/2chi^2) is Khaldoon's proposed patch which removes the extra 2/ndat on lines 89 and 99 of maxent_helper.cpp
Patched (chi^2/ndat) carries through the extra factor of 2/ndat to the objective function Q so its consistent everywhere (changing the Q function )

egull commented 3 years ago

This looks remarkably clean. I'm ok with changing the convention.