This originates from the fact that 1) my inferred stellar mass distribution of GAMA is 0.7 dex higher than that of GAMA DR4 (v24); 2) for the same set of SPS parameters, my predicted u-band magnitude can be 5 mags fainter than Chang's. Then I went back and checked my emulator. Several changes are made:
I cleaned up my code in speculator.py by a lot. Now the SuperSpeculator inherits many functions from Speculator. The _predict_spec_restframe now returns spectra in unit of Lsun/AA, and transform returns spectra in unit of erg/s/cm2/AA. My code has several differences from Chang's, e.g., when redshifting, I use linear interpolation to interpolate the spectrum to the original wavelength grid. Chang uses trapz interpolation to make sure the overall flux is conserved. But this makes very little difference in the overall spectrum and magnitude.
The reason for the u-band discrepancy is SDSS transmission curves. I used sedpy and its transmission curves, which is very different from speclite. Now I use speclite SDSS-2010 transmission curves assuming airmass=1.3.
In order to make an apple-to-apple comparison between my emulator and Chang's, I trained a new emulator adopting the PROVABGS model (SFH + ZH). Now two versions are available, dubbed as NMF (a fixed metallicity) and NMF_ZH (with metallicity history). In conclusion, my emulator agrees with Chang's very well.
In the end, there is basically no problem with my emulator. It turns out that GAMA logmintsfh (or logmstar) is not consistent between DR3 (v19 and v20) and DR4 (v24). I didn't find an explanation in DR4 paper. Need to figure this out.
This originates from the fact that 1) my inferred stellar mass distribution of GAMA is 0.7 dex higher than that of GAMA DR4 (v24); 2) for the same set of SPS parameters, my predicted u-band magnitude can be 5 mags fainter than Chang's. Then I went back and checked my emulator. Several changes are made:
speculator.pyby a lot. Now theSuperSpeculatorinherits many functions fromSpeculator. The_predict_spec_restframenow returns spectra in unit ofLsun/AA, andtransformreturns spectra in unit oferg/s/cm2/AA. My code has several differences from Chang's, e.g., when redshifting, I use linear interpolation to interpolate the spectrum to the original wavelength grid. Chang uses trapz interpolation to make sure the overall flux is conserved. But this makes very little difference in the overall spectrum and magnitude.sedpyand its transmission curves, which is very different fromspeclite. Now I usespecliteSDSS-2010 transmission curves assuming airmass=1.3.NMF(a fixed metallicity) andNMF_ZH(with metallicity history). In conclusion, my emulator agrees with Chang's very well.logmintsfh(orlogmstar) is not consistent between DR3 (v19 and v20) and DR4 (v24). I didn't find an explanation in DR4 paper. Need to figure this out.