specsrbench.figures.fig4_mae_summary¶
Figure 4 – global reconstruction fidelity, five ways.
Five bar panels over the eight methods. The top row is the paper’s central argument read left to right: raw MAE, where SR2 leads by 30%; the amplitude ratio, where it sits at 0.54 of the reference’s scale; and the scale-free MAE, where every method is level. Absolute error against a noisy reference falls when you shrink toward zero, whatever the reconstruction quality, and the third panel is the same comparison with that route closed. The bottom row is MAE normalised by the reference’s own flux uncertainty, and per-spectrum RMSE.
These panels replaced the manuscript’s global-fidelity table on 2026-10-02, so
the numbers a reader used to look up there are printed beside the bars of the
amplitude and scale-free panels. tests/test_paper_consistency.py checks
compute() against summary_final.csv row by row.
Module Attributes
|
Functions
|
|
|
One row per method, plus the two reference lines the panels draw. |
- specsrbench.figures.fig4_mae_summary.GAUSSIAN_ABS_FLOOR = 0.7978845608028654¶
E[|N(0, sigma)| / sigma]– where the uncertainty-normalised panel would sit for a reconstruction that is perfect up to the reference’s own noise.