mabitbol
commented
5 years ago Note: You can click on the figures for a bigger version. I have higher resolution ones if you want them (github has 10MB limit). Please let me know if you use any figures.
Results so far 2019-10-08 Data model 1
- [x] Baseline
- [x] Marginalized gain and shift with 1% and 3% Gaussian priors -- no impact on r vs baseline.
- [x] Marginalized gain with 10% Gaussian prior -- 4x10^-5 increase on 95% CL bound on r. This is on the edge of our chain variability so no solid detection of an issue yet.
- [x] AHWP and sinuous antenna biases -- 95% CL bound within baseline noise.
- [x] Marginalized over angle per band with self-calibration on -- 95% CL bound within baseline noise.
- [x] Marginalized shift over 10% Gaussian prior.
Other things:
- [ ] The baseline 13 parameter MCMC is run using emcee with 256 walkers x 20,000 steps for a total of 5 million likelihood evaluations. I ran it again with 512 walkers x 50,000 steps and now I can't figure out how to use get_dist to plot it nicely as it starts resolving noise features in the posterior (probably need to modify the smoothing scale or kde params?). Anyway just FYI the larger run with 25 million samples has r < 0.00377 at 95%.
- [x] I am now plotting density normalized PDFs, e.g. area under curves are the same instead of peak normalized.
Baseline constraints as printed by get_dist (can anyone help with the formatting here? Github inline latex or a different get_dist print function?)
\begin{tabular} { l c}
Parameter & 95\% limits\ \hline {\boldmath$A{lens} $} & $0.991^{+0.056}{-0.055} $\
{\boldmath$r $} & $< 0.00379 $\
{\boldmath$\betad $} & $1.591^{+0.032}{-0.032} $\
{\boldmath$\epsilon $} & $0.000^{+0.040}_{-0.040} $\
{\boldmath$\alpha^{EE}d $} & $-0.422^{+0.077}{-0.077} $\
{\boldmath$A^{EE}d $} & $10.03^{+0.71}{-0.66} $\
{\boldmath$\alpha^{BB}d $} & $-0.202^{+0.079}{-0.080} $\
{\boldmath$A^{BB}d $} & $5.01^{+0.37}{-0.35} $\
{\boldmath$\betas $} & $-3.005^{+0.065}{-0.067} $\
{\boldmath$\alpha^{EE}s $} & $-0.61^{+0.25}{-0.26} $\
{\boldmath$A^{EE}s $} & $4.02^{+0.43}{-0.40} $\
{\boldmath$\alpha^{BB}s $} & $-0.41^{+0.28}{-0.30} $\
{\boldmath$A^{BB}s $} & $2.01^{+0.26}{-0.24} $\ \hline \end{tabular}
Baseline (r pdf):
Bandpass marginalized (r pdf):
Angles marginalized (r pdf):
$\phi(\nu)$ in data but not model (r pdf):
Baseline full triangle plot:
All of the above triangle plot (CMB + FG params shown):
damonge
Here we begin to address bandpass and polarization angle systematics. I will break the results into two categories (1) addressing biases from an incomplete model and (2) assessing the increase in sigma r from marginalizing over systematics.
1) Check for biases from incomplete models. (If we cannot include in the model or have a poor model then we would like to assess the potential for bias) (i) Impact of $\phi(\nu)$: the in-band frequency dependent polarization response/rotation. $\phi(\nu)$ can be generated from the broadband 3-layer AHWP or the sinuous antenna design. (ii) Other? Intrinsic foreground EB? (iii) Can we create any intuition or rules of thumb to help understand what shapes and levels of systematics are most critical?
TODO list
2) Assess impact on ability to constrain r. How much does our constraint on r degrade when marginalizing over additional foreground and systematic parameters. We are considering bandpass and polarization angle systematics here in particular. (i) Bandpass uncertainties. Gain, shift to parameterize uncertainty. +Jeff's model of slopes, shifts, ripples. (ii) Polarization angle uncertainties. Including $\phi(\nu)$ uncertainties. Intrinsic EB.
TODO list
General TODO list
Other issues:
Another important note: These results do not agree with the SO forecast paper because of the simplified foreground model. There is no residual FG marginalization here that drove up the errors in the forecast! Things might be more realistic and agree with the forecast when we go to data model 3.