Model not fitting after changing pblum to component-coupled and dataset-coupled in multiple passband lightcurves

[kh-kr]

Hi,

I have been trying to do inverse modelling with three Gaia lightcurves. The parameter estimation and optimisation worked when I set pblum_mode to dataset-scaled for all three lightcurves. When coming to emcee, which requires setting the plumb to anything other than dataset-scaled. So I kept the Gaia G-band lightcurve to component-coupled and others to dataset-coupled. And then using some initial parameter distribution, I see that the model no longer fits the folded lightcurves. Why is it so?

The figure and code are attached below for reference.

Additionally, I have a few more doubts regarding this.

• The Gaia G passband luminosity of the primary still appears as the default one, and others are changed. So are these correct?

{'pblum@primary@lc_G': <Quantity 12.56637061 W>, 'pblum@secondary@lc_G': <Quantity 1.29840452 W>, 'pblum@primary@lc_BP': <Quantity 11.10691025 W>, 'pblum@secondary@lc_BP': <Quantity 0.85760881 W>, 'pblum@primary@lc_RP': <Quantity 4.30567077 W>, 'pblum@secondary@lc_RP': <Quantity 0.71495032 W>}

• Since I am using three passband lightcurves, is it possible to get separate temperatures from all three lightcurves? I mean, the multiple temperatures can really give a hint to the nature of the hidden companion in case it is an SB1 system.
• How trustworthy are the fitted parameters from the estimators and optimisers?

Please let me know if you need any additional information.

Looking forward to hearing from you soon.

b.set_value('pblum_mode@lc_G', 'component-coupled')
b.set_value('pblum_mode@lc_BP', 'dataset-coupled')
b.set_value('pblum_mode@lc_RP', 'dataset-coupled')

b.add_compute(compute='dyn_lc',overwrite=True)
b['enabled@lc_G@dyn_lc'] = True
b['enabled@lc_BP@dyn_lc'] = True
b['enabled@lc_RP@dyn_lc'] = True

print(b.compute_pblums(compute = 'dyn_lc'))
b.run_compute( 'dyn_lc')

b.add_distribution('teffratio', phoebe.gaussian_around(0.1), distribution='ndg_gauss',overwrite_all=True)
b.add_distribution('t0_supconj', phoebe.gaussian_around(10), distribution='ndg_gauss',allow_multiple_matches=True)
b.add_distribution('incl@binary', phoebe.gaussian_around(10), distribution='ndg_gauss',allow_multiple_matches=True)
b.add_distribution('requivsumfrac@binary', phoebe.gaussian_around(0.05), distribution='ndg_gauss',allow_multiple_matches=True)
b.add_distribution('ecosw@binary', phoebe.gaussian_around(10), distribution='ndg_gauss',allow_multiple_matches=True)

b.run_compute(compute='dyn_lc', sample_from='ndg_gauss', sample_num=10, model='from_ndg', overwrite=True)

b.plot(model='from_ndg', x='phase', show=True,legend = True)

[kecnry]

We were just about to release a bugfix that addresses some related issues - @aprsa do you know if these will be addressed there? Should we test before release or re-check after?

In the meantime, I'll answer your other questions

Since I am using three passband lightcurves, is it possible to get separate temperatures from all three lightcurves? I mean, the multiple temperatures can really give a hint to the nature of the hidden companion in case it is an SB1 system.

Do you mean different temperatures per-passband or do multiple passbands allow you to constrain per-component absolute temperatures?

How trustworthy are the fitted parameters from the estimators and optimisers?

Estimators are designed just as a shortcut to get estimated initial values based on the geometry of the data (without any model), optimizers then try to find the local minimum. You should not use either of these as a black box and expect them to give you the correct answer and should always be used in conjunction with sampling to understand uncertainties and correlations between parameters.

[kh-kr]

Thank you for the response @kecnry .

Do you mean different temperatures per-passband or do multiple passbands allow you to constrain per-component absolute temperatures?

I meant absolute temperatures per component.

I have now somewhat resolved the model fitting issue, but I still have some doubts. Below are the details:

1. Previously my fluxes were in the units of electrons per second, so using appropriate relations, I have converted them to watts per meter square and also corrected for extinction.

2. For passband luminosities: I used the median flux (unit: watt per meter square) of the lightcurve which is almost equal to the out-of-eclipse flux of the binary. I then multiplied this median flux to 4 x pi x distance^2. Here distance is in meters.

3. I then provided only Gaia G passband luminosity and set

b.set_value('pblum_mode@lc_G', 'component-coupled'), 
b.set_value('pblum_mode@lc_BP', 'dataset-coupled'), 
b.set_value('pblum_mode@lc_RP', 'dataset-coupled')

4. Now the model still does not fit the lightcurve, the model fluxes are around 2 times larger. I wonder how the model fluxes are being calculated. I have attached the figure below for reference. Here green, yellow, and magenta model light curves represent Gaia G, BP, and RP bands respectively, which are systematically higher than observed lightcurves.

How can I resolve this issue now?

Also, I am getting errors that WARNING Online passbands unavailable (reached max tries)
How can I resolve this? does this also mean that plum_mode dataset-coupled may not work?

[aprsa]

If your only goal is to get individual temperatures, then you shouldn't fiddle with passband luminosities manually; instead, set one of them to scale with a reference dataset; for example:

b.set_value('pblum_mode', dataset='lc_G', value='dataset-scaled')
b.set_value('pblum_mode', dataset='lc_BP', value='dataset-coupled')
b.set_value('pblum_mode', dataset='lc_RP', value='dataset-coupled')
b.set_value('pblum_dataset', dataset='lc_BP', value='lc_G')
b.set_value('pblum_dataset', dataset='lc_RP', value='lc_G')

That way phoebe will fit light levels (passband luminosity) for the G band automatically, and then retain the color index for BP and RP wrt G band. Note that you'll need phoebe 2.4.22+ for this to work, where a bug was fixed that prevented this scheme from working reliably in some circumstances.

As for the failing online passbands, are you behind a firewall? I'm not seeing service issues on our end. Where are you trying to connect from?

[kh-kr]

Thank you, @aprsa, for your prompt response.
I have updated Phoebe and rerun the code as you suggested. The online passband issue has been resolved as well. However, the BP and RP band models are still not fitting properly. I have attached the figure for your reference.

If I keep the lc_G scaled, will the absolute luminosities be unaffected? Will I still be able to obtain L1 and L2 for both components?

[kh-kr]

I figured out that the issue might be related to the conversion factors from e/s to w/m². I verified this against earlier e/s fluxes, and it appears to be working fine.

However, I am encountering an error that states:

WARNING: Sampling with dataset-scaled can cause unintended issues. Consider using component-coupled and marginalizing over pblum.

Does this mean that I will not be able to estimate the absolute luminosities of both components?

[aprsa]

That's correct: once you have your temperatures near-correct (and your second post hints that you're close), then you will want to set pblum@lc_G to the automatically determined value and switch to using it explicitly for sampling. Otherwise you're ignoring the covariances between pblum and other parameters, causing your posterior widths to be underestimated.