PRISM-Astro/h2f_delight_radius_norm_p4

Description

This is the dataset from DELIGHT [1], a model that locates the host galaxy of a transient from multi-resolution images. The difference with the original is that the images are not the cutouts used in the paper: they were regenerated from hips2fits (PanSTARRS DR1 HiPS, served by CDS).

It contains 12,610 examples, built by autolabeling: transient positions are simulated, not observed. For each of the 7,093 host galaxies in h2f_delight_train, positions are drawn uniformly inside the elliptical footprint of its r-band Sérsic fit, truncated at 3·Re.

This is exactly how h2f_delight_autolabeling_1pct was built. The only difference is how many positions each galaxy gets, and that is what the radius_norm name refers to.

This variant uses the exponent p = 4. Its sibling h2f_delight_radius_norm uses p = 6; see the comparison at the end.

How many examples per galaxy: radius_norm

The problem

A galaxy can only place a transient as far as 3·Re from its own center. A small galaxy simply cannot produce a large transient–host distance.

Most galaxies are small. So if every galaxy contributes roughly the same number of examples, nearly all of them land at short distances and the long ones are almost missing. In 1pct, the fullest distance bin holds 338 times more examples than the emptiest one.

The rule

Bigger galaxies generate more examples:

n = max(1, round((Re / Re0) ** p))        Re0 = 8 arcsec,  p = 4

Re0 is the size at which a galaxy starts generating more than one example. Below it, a galaxy contributes exactly one. No galaxy is ever dropped: all 7,093 are present, and 90.6 % of them contribute a single example. Only 667 galaxies (9.4 %) contribute more, up to 95 for the largest one.

Why those numbers

The count of galaxies large enough to reach a given distance drops very fast: 2,912 reach 10 arcsec, 499 reach 30, and only 27 reach 60. Raising the count to a power compensates that drop, and p = 4 with Re0 = 8 is the pair that flattened the histogram best in a grid search over both.

They are not universal constants: they were fitted to the size distribution of this catalog and must be recomputed for a different survey.

Result

	1pct	radius_norm_p4
examples	535,743	12,610
examples per galaxy	75.5 avg	1.78 avg
median distance	13.85″	9.69″
90th percentile	30.42″	35.79″
fullest / emptiest distance bin	147×	43×

Measured over 5 linear bins between 0″ and 70″, the grid used to report per-distance metrics. Neither dataset leaves an empty bin.

p = 4 versus p = 6

	radius_norm (p=6, Re0=12)	radius_norm_p4 (p=4, Re0=8)
examples	9,246	12,610
fullest / emptiest bin	56×	43×
same, after balanced sampling	8×	7×
effective sample size kept by that sampler	63.4 %	70.6 %
worst-case repeats per epoch	15	12

p = 4 is better on every measure; it just costs 3,364 more examples. The balanced sampler referenced above is the KDE sampler in prism/sampler.py (configs/sampler/distance_kde.yaml).

How it was built

Every row here is a row of 1pct, copied unchanged. Because both use the same uniform-in-area law, keeping a random subset of each galaxy's rows is a valid sample of that same law, so nothing had to be regenerated or re-downloaded.

Every cutout is centered on the simulated transient position, and the host_pos label is the offset from there to the host galaxy. The fold column is inherited from h2f_delight_train, so the 5-fold cross-validation split is preserved and no host galaxy leaks between training and validation.

Note that folds are balanced by galaxy count, not by galaxy size, and Re**4 amplifies any size asymmetry between them: fold 4 ends up with 2,934 examples and fold 5 with 2,250.

Image format

image has shape (12610, 5, 1, 30, 30), which maps to (examples, resolution levels, photometric bands, height, width). Only the r band was downloaded, hence the size-1 band dimension.

Columns

Column	Type	Description
Image and label
image	(5,1,30,30) float32	Multi-resolution PanSTARRS r-band cutout, centered on the simulated transient position.
host_pos	float32 × 2	Offset [x, y] from the simulated transient to the host galaxy, in level-0 pixels.
Identification
oid	string	Identifier of the real transient this galaxy came from. Repeats across the draws of the same galaxy.
host_name	string	Catalog name of the host galaxy.
Host galaxy
host_redshift	float64	Redshift of the host; NaN when unknown.
hostsize	float64	Angular size of the host galaxy from the catalog.
host_ra	float64	Right ascension of the host, in degrees.
host_dec	float64	Declination of the host, in degrees.
Real transient position
meanra	float64	Mean right ascension of the real transient, in degrees. Not the center of this cutout.
meandec	float64	Mean declination of the real transient, in degrees. Not the center of this cutout.
sn_ra	float64	Right ascension of the real supernova, in degrees.
sn_dec	float64	Declination of the real supernova, in degrees.
Sérsic fit — r band, from CasJobs
rSerRadius	float64	Effective radius Re of the r-band Sérsic fit, in arcsec. Defines the sampling region.
rSerMag	float64	Magnitude of the r-band Sérsic fit.
rSerAb	float64	Axis ratio b/a of the fitted ellipse, between 0 and 1. Defines the sampling region.
rSerPhi	float64	Position angle of the fitted ellipse, in degrees. Defines the sampling region.
rSerRa	float64	Right ascension of the Sérsic fit center, in degrees.
rSerDec	float64	Declination of the Sérsic fit center, in degrees.
Cross-validation
fold	int64	Cross-validation fold, from 1 to 5, inherited from the source galaxy.

References

[1] Förster, F., Muñoz Arancibia, A. M., Reyes-Jainaga, I., et al. 2022, DELIGHT: Deep Learning Identification of Galaxy Hosts of Transients using Multiresolution Images, The Astronomical Journal, 164, 195. doi:10.3847/1538-3881/ac912a

@article{Forster_2022,
  doi = {10.3847/1538-3881/ac912a},
  url = {https://doi.org/10.3847/1538-3881/ac912a},
  year = {2022},
  month = {oct},
  publisher = {The American Astronomical Society},
  volume = {164},
  number = {5},
  pages = {195},
  author = {F\"orster, Francisco and Mu\~noz Arancibia, Alejandra M. and Reyes-Jainaga, Ignacio and Gagliano, Alexander and Britt, Dylan and Cuellar-Carrillo, Sara and Figueroa-Tapia, Felipe and Polzin, Ava and Yousef, Yara and Arredondo, Javier and Rodr\'iguez-Mancini, Diego and Correa-Orellana, Javier and Bayo, Amelia and Bauer, Franz E. and Catelan, M\'arcio and Cabrera-Vives, Guillermo and Dastidar, Raya and Est\'evez, Pablo A. and Pignata, Giuliano and Hern\'andez-Garc\'ia, Lorena and Huijse, Pablo and Reyes, Esteban and S\'anchez-S\'aez, Paula and Ram\'irez, Mauricio and Grand\'on, Daniela and Pineda-Garc\'ia, Jonathan and Chabour-Barra, Francisca and Silva-Farf\'an, Javier},
  title = {DELIGHT: Deep Learning Identification of Galaxy Hosts of Transients using Multiresolution Images},
  journal = {The Astronomical Journal},
}