Data quality
How far to trust the data
Everything here is computed from the records, not typed. The short version: the curves are traceable to their sources, but most of the simulation curves were read from figures and almost none has an independent check.
How the curves were obtained
897 of the 1636 simulation curves (55%) were read from published figures, and 0 of them are used for ranking. They are read from the vector paths in the figure files with the axes calibrated on the printed tick marks, not by eye from an image; the reading error the scripts declare is a median 0.01 dex and at most 0.05 dex. That is the error of reading the figure. Whether the figure itself is right only its authors can say.
| Kind | How obtained | Records | Used for ranking |
|---|---|---|---|
| empirical-model | a table in the paper | 40 | 40 |
| observation | computed by this project from a catalogue | 12 | 12 |
| observation | read from a figure | 16 | 0 |
| observation | the authors' table | 15 | 15 |
| observation | a published fit | 113 | 107 |
| observation | a table in the paper | 209 | 208 |
| simulation | computed by this project from a catalogue | 262 | 262 |
| simulation | read from a figure | 897 | 0 |
| simulation | the authors' table | 259 | 126 |
| simulation | a published fit | 67 | 37 |
| simulation | a table in the paper | 151 | 133 |
Independent checks
A figure-reading can be checked directly only where the same simulation, relation and epoch also exists as a table or fit with identical definitions. That holds for 11 digitized curves (1.2%). For every other digitized curve the reading has no independent check in the data; the extraction code is tested on synthetic figures with known answers, which checks the method and not each figure. The pairs differ by a median of 0.000 to 0.528 in the curve's units.
| Curve | shared points | median difference | largest difference |
|---|---|---|---|
| THESAN-1 THESAN-1 (sfrd, z = 10) | 49 | 0.528 | 1.451 |
| THESAN-1 THESAN-1 (sfrd, z = 10) | 49 | 0.354 | 1.039 |
| NIHAO zoom suite NIHAO zoom sample with black holes (per galaxy) (bh, z = 0) | 50 | 0.139 | 0.831 |
| NIHAO zoom suite NIHAO zoom sample with black holes (per galaxy) (shmr, z = 0) | 48 | 0.096 | 0.432 |
| Illustris Illustris-1, n = 0.00316 h^3 Mpc^-3, all (measured, Fig. HOD-ND left) (hod, z = 0) | 9 | 0.026 | 0.127 |
| EAGLE L0100N1504 (Ref), n = 0.00316 h^3 Mpc^-3, all (measured, Fig. HOD-ND left) (hod, z = 0) | 9 | 0.012 | 0.106 |
| Illustris Illustris-1, n = 0.01 h^3 Mpc^-3, all (measured, Fig. HOD-ND left) (hod, z = 0) | 11 | 0.014 | 0.092 |
| EAGLE L0100N1504 (Ref), n = 0.01 h^3 Mpc^-3, all (measured, Fig. HOD-ND left) (hod, z = 0) | 11 | 0.016 | 0.067 |
| Illustris Illustris-1, n = 0.0316 h^3 Mpc^-3, all (measured, Fig. HOD-ND left) (hod, z = 0) | 11 | 0.009 | 0.048 |
| EAGLE L0100N1504 (Ref), n = 0.0316 h^3 Mpc^-3, all (measured, Fig. HOD-ND left) (hod, z = 0) | 13 | 0.007 | 0.028 |
| COLIBRE m6 (L200, 50 pkpc aperture, figure) (gsmf, z = 0) | 18 | 0.000 | 0.000 |
A second reader re-digitized a random sample by a different method (a 400 dpi raster of each figure, axes recalibrated from the tick marks, the curve traced by colour; the original extraction's vector paths were not read). 20 curves from 8 papers could be checked, out of 96 tried; 10 are within tolerance and 10 are not. On logarithmic axes the median difference is 0.0006 dex. Each failure was inspected and none could be attributed to a stored curve, but the check could not see a wrong panel or run, and 76 curves could not be checked at all. One reader made it.
For the 49 papers behind the figure-read simulation curves we looked for released numbers. By paper: the full table is available for 3, part of it for 5; only the simulation's catalogues are public (the plotted numbers would be recomputed) for 24; the numbers are available on request or announced for 5 (288 curves); nothing was found for 12. Of the 10 curves for which a table matching the plotted numbers exactly was found, 10 have been replaced by the table's numbers and 0 are still read from the figure; 1 more figure-read curve duplicates numbers already in the dataset as an official table.
Where one simulation and run appears in two papers with the same axis definitions, the papers can be compared with each other; 1 such pair exists today. The full lists are in DATA-QUALITY.md.
Where the evidence is thin
The number of distinct simulations with a curve of each relation near each epoch. Amber cells have one or two: an offset against so few says little about the field. 58 cells are amber, and a grey zero means no simulation at all.
| Relation | z = 0.1 | z = 0.5 | z = 1 | z = 2 | z = 3 | z = 4 | z = 6 | z = 8 |
|---|---|---|---|---|---|---|---|---|
| bh | 11 | 3 | 9 | 9 | 6 | 9 | 1 | 2 |
| bhsigma | 2 | 1 | 1 | 1 | 0 | 1 | 0 | 1 |
| btfr | 5 | 1 | 1 | 1 | 0 | 1 | 0 | 0 |
| colour | 3 | 0 | 1 | 1 | 1 | 1 | 0 | 0 |
| elf | 1 | 1 | 1 | 1 | 1 | 1 | 2 | 1 |
| fgas500 | 2 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| fmr | 4 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| gas | 3 | 1 | 2 | 2 | 0 | 2 | 0 | 1 |
| gsmf | 11 | 8 | 9 | 9 | 7 | 11 | 12 | 11 |
| h2-fraction | 3 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| hi-fraction | 3 | 0 | 1 | 0 | 0 | 1 | 0 | 0 |
| hod | 2 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| jstar | 5 | 2 | 3 | 2 | 2 | 1 | 0 | 0 |
| ks | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 0 |
| metald | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 0 |
| mzr | 7 | 1 | 6 | 5 | 6 | 6 | 5 | 9 |
| quenched | 6 | 2 | 4 | 4 | 2 | 0 | 0 | 1 |
| sfms | 8 | 1 | 7 | 6 | 5 | 6 | 7 | 7 |
| sfrd | 0 | 0 | 0 | 10 | 1 | 0 | 1 | 3 |
| shmr | 10 | 5 | 8 | 6 | 4 | 7 | 4 | 7 |
| size | 9 | 6 | 7 | 8 | 5 | 7 | 4 | 4 |
| smd | 0 | 0 | 0 | 4 | 1 | 1 | 0 | 1 |
| ssfr | 9 | 2 | 5 | 5 | 3 | 5 | 0 | 1 |
| stfr | 3 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| uvbeta | 0 | 0 | 0 | 0 | 0 | 1 | 3 | 1 |
| uvbetam | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 0 |
| uvlf | 0 | 0 | 0 | 0 | 0 | 1 | 6 | 7 |
| wp | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| ysz | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| zstar | 6 | 2 | 5 | 5 | 4 | 6 | 5 | 5 |
Corrections
Every fix to the data, an analysis or a claim, with the release it was made in, is in CORRECTIONS.md. What was changed in the data, and what was not, is in Methods §4. Source checksums and the spelling of definitions are in DATA-QUALITY.md.
Third-party terms
370 records carry explicit third-party terms. 180 of them are values this project binned from a repository with no licence file; permission has not yet been requested, and they should be treated as not cleared for reuse until it is.