Error estimates from forcepho

[jensmelinder]

I'm using Forcepho to estimate photometry for high redshift galaxies using NIRCam+MIRI filters. The uncertainties from Forcepho (taken as 16/ 84 percentiles of the PDF from the marginalised posterior of the flux) seems to sometimes be very small. While I expect to get much better SNR from Forcepho than from aperture photometry it seems that you sometimes get SNR improvement factors on the order of hundreds, which seems too good?

In particular sources that are more or less undetected suffer from this, causing them to go from SNR<1 to SNRs up to 5-10.

Is there a good way to estimate more believable errors? Or would this perhaps point to an issue with how the fitting is setup?

[bd-j]

You might try a single band to make things a little clearer. Otherwise I would make sure to check the flux priors and the chains to make sure you aren't hitting a prior.

[bd-j]

Also jjst to be clear, you are comparing to aperture photometry on individual exposures rather than stacks/mosaics?

[jensmelinder]

Thanks!

We (I'm working together with my postdoc on this) haven't tried running on a single filter, but will do some tests with this.

We did identify a problem with the flux priors which solved the issue for some galaxies but not all of them. Using an uncertainty catalogue for the flux prior widths helped, but some galaxies annoyingly still give tiny errors.

We're running on stacked images rather than on individual frames. I have not gotten around testing this, mainly because it would take us a lot of time to get NIRCam individual frames processed for all of the filters. The aperture errors (which is now used for the flux priors with a factor of 3 for nsig_pix) are from the stacked error images (with scaling to account for resampling/drizzling correlations).

I've looked a bit further into when HMC/NUTs might give too low uncertainties so we have some stuff to try out there as well. From what I read, one of the common issues is "chain sticking", which I think results in a low ESS value for the fit. Is the ESS for the chain stored in the patch h5 file? In that case it may be that we need to increase the burn-in sample size.

[bd-j]

ESS is not stored in the outputs, sorry.