specsrbench.figures.fig2_qualitative¶
Figure 2 – one held-out galaxy, every method, with an [O III] inset.
The example is chosen on a property of the data, not of any result: a redshift high enough that the [O III] doublet is physically resolvable. 4959 and 5007 are separated by 4.933 nm x (1+z), so at z = 6.55 they sit 37.2 nm apart against an LSF FWHM of ~21 nm at 3.78 um. Below z ~ 5 they fall inside one resolution element, where no method can separate them – the prism input resolves the doublet in 0% of z = 0-2 galaxies, 0.5% of z = 2-4, 27% of z = 4-6 and 100% of z > 6. Picking an unresolvable case credits a method for recovering structure its own input does not carry.
Favourable to SR2 but not cherry-picked: among the 296 held-out spectra with reference [O III] S/N > 20 and the doublet on the grid, SR2’s RMSE gain here ranks 44th of 296 – the 85th percentile, 22.9% against a subset median of 12.4%. SR2 still inflates the [O III] S/N to 466 against 88 in the reference, at an amplitude ratio of 0.27 where the classical methods sit at 0.93-1.15.
Module Attributes
Figure 2 names the prism baseline by what it is (an interpolation of the low-resolution input) and spells out TV's Wiener prefilter. |
|
The held-out spectrum drawn. |
|
Half-width of the [O III] inset, in microns. |
Functions
|
- specsrbench.figures.fig2_qualitative.LABELS = {'LR': 'LR (cubic)', 'TV': 'Wiener + TV'}¶
Figure 2 names the prism baseline by what it is (an interpolation of the low-resolution input) and spells out TV’s Wiener prefilter.
- specsrbench.figures.fig2_qualitative.I_SHOW = 188¶
The held-out spectrum drawn. See the module docstring for why this one.
- specsrbench.figures.fig2_qualitative.ZOOM_WIN = 0.07¶
Half-width of the [O III] inset, in microns.