PSF testing and validation notes

[sfarrens]

The objective of this issue is to gather notes on the various methods for testing and validating PSF models discussed during the reading group.

[sfarrens]

@sfarrens notes:

Home | Shear Reading Group

Notes

• Sky coordinates
  • Axel: Why is this a big deal?
  • Common reference for everything
  • Dependent on astrometric solution
• Full focal plane
  • Communication between CCDs
  • Axel: Better for any survey
• Tree rings
  • Due to impurities in silicon
  • Larger scale rings not tree rings, due to modelling effects
• Parametric model-based methods
  • PIFF has many models
  • Tobias: pixel-based models robust to different conditions
  • Parametric model is hard, even for satellites, changes in temperature etc. over time
  • Need enough stars to use a non-parametric model
  • Best approach use "best of both worlds"
  • Andre: May not need perfect representation of the PSF
  • Axel: Different story for metacalibration where PSF is deconvolved
  • Toby: May not need full precision on the wavefront
  • Alex: bandpass size is important
  • mixing several PSF models for a large bandpass
  • Andre: image quality drops off for narrow bands
• Interpolation
  • Most methods using polynomial (degree ~2)
  • MCCD uses RBF
  • Need more stars to increase the degree of the polynomial
  • Axel: Should keep a homogeneous PSF interpolation (i.e. not based on the number stars in a given region)

Validation Tests

• PSFEx (for HSC)
  • Look at size dispersion
  • Residual in e1
  • Co-add PSFs perform worse than single exposures
  • Worse PSF when the seeing is good
  • Axel: PSF size close to the pixel size -> undersampling
  • Axel: difficult to capture intra-pixel shift
  • Euclid will use multiple exposures
  • Axel: HSC validation done on stacked PSF
• KiDS
  • Shapelet expansion
  • PSF "Gaussianization"
  • PSF shear field -> reGaussianized -> residuals
  • Check 2pt correlation function for each exposure
  • Axel: surprisingly small error bars
  • Axel: This shows that reGaussianization works but not that the PSF model is good
• DES
  • Qualitative analysis
  • Rho statistics can be related to cosmology requirements
  • Correlations between residuals and PSF model
  • 5 cross-correlations

[sfarrens]

• MCCD
  • Small sample of stars
  • Metrics (pixel RMS, noise in models)
  • Found that PSFEx overestimates the size more than MCCD
  • François: What should we look for in the pixel residuals?
• Discussion
  • André: How do I know how many stars I need for modelling the PSF?
  • Tobias: Not easy to find an answer, can use empirical model
  • Axel: Need to look at of impact of this, rho stats perhaps not the best
  • Need good simulations to look into this
  • Axel: hard to know that PSF is locally good
  • François: Why is it pessimistic to assume T_PSF/T_gal = 1? [DES Y1, fig.9]
  • alpha is PSF leakage
  • Rho statistics
  • Axel: ellipticity of the PSF or of the stars -> different results
  • Axel: Ideal validation would also test local PSF features
  • Axel: Survey-specific tests in case of e.g. brighter-fatter effect
  • Look at residuals as function of magnitude
  • Axel: Polychromatic PSF model (e.g. Euclid)

[sfarrens]

Proposed Subsections

• Organised by the various tests of the PSF model
  • Visual examples from various surveys for each test
  • Notes on what features to look out for
• Qualitative vs quantitative tests
• PSF-only validation vs correlation with galaxy shapes
  • Plan A: PSF-only validation (i.e. tests that can be made of the PSF model directly without galaxy shape measurements)
• Astrometry in the book?
  • intra-pixel shift
  • PIFF more precise astrometric solution
  • Put in preprocessing section
• Rho statistics
  • include derivation
  • Paulin method?
  • interactive notebook?
  • link to requirements
  • fake ellipticity for a PSF model -> small rho stats
  • correlated errors -> impact on rho stats
• alpha coefficient