PSFEx native sampling leaves a g1 residual on sub-0.5″ exposures; PSF consumers assume PSF_SAMP = 1

Ouverte
  • Quelque chose ne va pas ?
#989 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub

Les mainteneurs répondent en général sous 1 jour

Personne n'a encore pris cette issue.

Évaluation

Difficulté
5/5
Temps estimé
Plus d'une semaine
Accessibilité débutants
25/100
Type d'issue
Bug
Clarté
Plutôt claire
Activité
Active
Stack technique
python
Domaine
data

Piste de recherche

Start with psfex_interp.py and the PSFExInterpolator.interpsfex entry point; inspect how it reads PSF_SAMP and find its existing tests. The proposed guard should reject sampling values other than 1, with a test that covers the rejection. Later sampling support and the bounded galaxy shear pilot require separate validation; completion criteria for those are not specified in the issue.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

Claude Opus on behalf of Cail; investigation by GPT-6.1 Sol.

On the best-seeing exposures, PSFEx at native sampling (PSF_SAMPLING 1) leaves a coherent g1 residual. Our PSF consumers assume unit sampling, so the obvious config fix would silently break them.

The residual shows up at exposure level once the mean PSF FWHM is below ≈0.5″ (≈2.7 px). 286 of the 15,523 v2 validation exposures (1.8%) are that sharp; 2,504 (16%) are below 0.6″. Above ≈0.5″ the median residual is flat at zero, so a blanket seeing cut isn't warranted.

exposure residual vs PSF FWHM

For one exposure, 2780860 (18BD97, 2 Sept 2022, 0.50″), against a same-region control at 0.70″:

  • Production gives Δg1 = −4.0×10⁻³; the control gives −0.2×10⁻³. The rerun reproduces production's PSFEx basis exactly.
  • PSFEx at 0.5-px sampling (101×101 model) gives −1.3×10⁻³ with centered validation stamps. Rendered at each validation star's sub-pixel phase, it gives −0.02 to +0.26×10⁻³, depending on the Lanczos kernel: consistent with zero.
  • Phase-matching the native model alone brings it to −1.0 to −2.7×10⁻³ (kernel-dependent). On the control, all four conventions agree, so pixel phase matters only once the stars are undersampled.

per-CCD residual under four model/validation conventions

Why PSF_SAMPLING 0 (auto) is not the fix: with SAMPLE_AUTOSELECT N, PSFEx assumes a nominal 6-px FWHM and picks 1.28 px per model pixel, which is coarser. And any sampling other than 1 breaks the consumers:

  • PSFExInterpolator.interpsfex (psfex_interp.py) returns the raw PSF_MASK grid and ignores PSF_SAMP.
  • PSF validation attaches the CCD's native WCS to that grid. On a 0.5-sampled model it measures ΔT/T = −0.75: twice the linear size, four times the area.
  • The galaxy_psf store carries the array without sampling metadata, and ngmix's make_ngmix_observation builds the PSF Jacobian from the image WCS.

Proposed:

  1. A guard, now: psfex_interp raises if PSF_SAMP ≠ 1, with a test. It protects every consumer until sampling is supported.
  2. For run 2: psfex_interp emits native-pixel effective-PSF stamps with an explicit centering convention. Then validate sampling 0.5 / size 101 as the candidate configuration.
  3. For run 1: flag the sub-0.5″ exposures for diagnostics. Measure the galaxy shear bias on a bounded pilot (a few nights, paired ngmix fits) before any rerun. The galaxy-level effect is unmeasured; the star residual does not set it.

Possibly related: #956, a ~0.4% PSF size residual on P3 tiles, measured with phase-shifted models. Not the same measurement.

Report, data and reproduction: Nibi /scratch/cdaley/psf-undersampling/REPORT.md.

Langage dominant
Python
Étoiles
18
Forks
14
Merge moyen
3 j 18 h
PR mergées (30 j)
27

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de CosmoStat/shapepipe

Toutes les issues de CosmoStat/shapepipe

Issues similaires

Plus d'issues Python

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.