Negative chi^2 from Commander

Use of Healpix, camb, CLASS, cosmomc, compilers, etc.
Post Reply
Björn Sörgel
Posts: 13
Joined: April 05 2013
Affiliation: Institute of Astronomy/Kavli Institute for Cosmology Cambridge

Negative chi^2 from Commander

Post by Björn Sörgel » October 22 2013

Dear all,
when I run with feedback=3, the following LnLike and Chi2 are printed for the individual datasets:

Code: Select all

Calling CAMB
 CAMB done
 lowlike_v222.clik lnlike =    1007.28444991537     
 DR7 BAO likelihood =    1.91087739414207     
 6DF BAO likelihood =   1.449826393350034E-002
 DR9 BAO likelihood =   5.291717134685368E-002
 commander_v4.1_lm49.clik lnlike =   -3.74495218656174     
 CAMspec_v6.2TN_2013_02_26_dist.clik lnlike =    4192.16885357843     
   1007.284   2014.569   CMB: lowlike_v222.clik
      1.911      3.822   BAO: DR7
      0.014      0.029   BAO: 6DF
      0.053      0.106   BAO: DR9
     -3.745     -7.490   CMB: commander_v4.1_lm49.clik
   4192.169   8384.338   CMB: CAMspec_v6.2TN_2013_02_26_dist.clik
Why is there a negative chi2 from commander? And secondly, isn't the BAO chi2 very low?. I first discovered this in a modified Jun2013 version, but I could reproduce it with a vanilla Oct2013 one.

Cheers,
Björn

Antony Lewis
Posts: 1384
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: Negative chi^2 from Commander

Post by Antony Lewis » October 22 2013

The normalization of the likelihoods is arbitrary (constant offsets in the log like don't matter); they could have added a constant offset to the commander likelihood to make it look like a \chi^2 but didn't.

Björn Sörgel
Posts: 13
Joined: April 05 2013
Affiliation: Institute of Astronomy/Kavli Institute for Cosmology Cambridge

Negative chi^2 from Commander

Post by Björn Sörgel » October 22 2013

Dear Antony,
thanks for the quick reply. "Problem" solved.

Cheers,
Björn

Luke Hart
Posts: 40
Joined: July 13 2015
Affiliation: University of Manchester

Re: Negative chi^2 from Commander

Post by Luke Hart » August 02 2017

Antony Lewis wrote:The normalization of the likelihoods is arbitrary (constant offsets in the log like don't matter); they could have added a constant offset to the commander likelihood to make it look like a \chi^2 but didn't.
Antony, it seems that if you try and run GetDistGUI.py with this dataset, it won't let you view the .margestats or the Parameter tables due to the negative chi2 from commander? Anyway you can think of getting around this??

Luke

Antony Lewis
Posts: 1384
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: Negative chi^2 from Commander

Post by Antony Lewis » August 02 2017

You need to edit the .ranges file for the chain, and remove the 0 N range for chi2_xxx where xxx is the likelihood.

Luke Hart
Posts: 40
Joined: July 13 2015
Affiliation: University of Manchester

Negative chi^2 from Commander

Post by Luke Hart » August 02 2017

Fantastic thank you :)

Post Reply