Best-fit chi^2?
a-e-cole opened this issue · 2 comments
Hi all,
I am wondering how this python/cobaya implementation of hillipop compares with previously published results. I had a look at https://arxiv.org/abs/1609.09730, which quotes a chi2 of 9995.9 (with fewer degrees of freedom that the implementation in this repo, I guess at some point something changed).
When I use cobaya's minimizer with the TT component of this likelihood, I find a best-fit
weight minuslogpost theta_s_1e2 logA ns omega_b omega_cdm tau A_planck cal100A cal100B cal143B cal217A cal217B Aradio Adusty AdustTT Asz Acib Aksz Aszxcib H0 sigma8 minuslogprior minuslogprior__0 chi2 chi2__planck_2020_hillipop.TT
1 5698.4544 1.0417533 3.0456148 0.96222611 0.022164104 0.12010398 0.056121184 0.99974758 -6.3060996e-05 0.0027286476 -0.0014550786 0.00017878047 0.00018272773 1.6739364 0.78062506 0.8803056 1.2316963 0.7894615 3.3306691e-16 1.6577574 67.601084 0.82496051 -44.766851 -44.766851 11486.442 11486.442
The minimum chi2 of 11486.442 seems quite a bit higher than that reported in 1609.09730 (even when normalized to # d.o.f.). Any ideas as to where the discrepancy comes from? Is there a more up-to-date reference for best-fit/etc. with hillipop that I should be looking at?
(For the theory code I am using CLASS v3.0 rather than CAMB, but I don't expect this to affect the situation dramatically.)
Cheers,
Alex
@acole1221 Indeed, we extended the multipole range with respect to 2015 Planck release. The current ndof is then 10816 in TT.
I hope to be able to publish a paper including the detail of this new version soon.
Hi @acole1221,
Hillipop v1.1 should have minimum chi2 values more reasonnable.
Multipole range have slightly changed as well as PS model.
But the major impact comes from a small bug corrected in the covariance matrix estimate.
Thanks for your input,
Matt.