SLIDE 1
Tests of the Scarlet and Multi-object Fitting Deblenders for Weak - - PowerPoint PPT Presentation
Tests of the Scarlet and Multi-object Fitting Deblenders for Weak - - PowerPoint PPT Presentation
Tests of the Scarlet and Multi-object Fitting Deblenders for Weak Lensing Shear Recovery Erin Sheldon and Lorena Mezini Brookhaven National Laboratory, Stony Brook University August 15, 2018 Testing Scarlet for Shear Recovery Motivation was
SLIDE 2
SLIDE 3
Scarlet Deblender Introduction
◮ See Melchior et al. https://arxiv.org/abs/1802.10157
and talks yesterday.
◮ The number of objects, and nominal center for each object,
are inputs.
◮ Every pixel is a parameter. ◮ Use constraints to reduce the dimensionality, e.g.
◮ Positivity (required) ◮ Monotonicity ◮ Symmetry
SLIDE 4
Example Scarlet Deblend from HSC
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 Data 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
SCARLET
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 Residual
SLIDE 5
Multi-object Fitting Deblender (MOF)
◮ New Implementation of the Dark Energy Survey MOF
deblender (original Matt Becker using my ngmix code base)
◮ Multi-band deblender ◮ Fit all objects in a blend simulateously with flexible models
(original MOF was iterative)
◮ Model is bulge+disk, with both components the same size,
ellipticity and center.
◮ Regularize some parameters to make the process stable.
◮ Shape ◮ Bulge fraction ◮ Center ◮ No regularization of size or fluxes needed.
SLIDE 6
Multi-object Fitting
◮ Can fit large groups simultaneously. ◮ In simulations, very stable, sub-percent failure rate even for
large groups.
◮ I have fitted up to 225 objects. All of the 24 such groups I
tried converged
SLIDE 7
Controlled Deblend Simulations for Shear Recovery
◮ Simulate pairs of galaxies at various separations ◮ For what I will show, there is a “central” used to measure
the shear.
◮ There is a “neighbor” which is twice as big and 33%
brighter.
◮ We have tried many models for these objects. Shown here
are examples for
◮ bulge+disk ◮ bulge has a different sizes, ellipticity, and is offset from the
disk center.
◮ Disk is scattered with knots of star formation. ◮ These are not a good fit for the MOF model generally.
SLIDE 8
Example Simulated Galaxies (Large)
SLIDE 9
Example MOF de-blended pair. Low Noise.
Left is original, right is MOF deblended. Can see residuals from “knots” of star formation.
SLIDE 10
Example MOF de-blended pair. Medium Noise (S/N∼ 50.)
Left is original, right is MOF deblended.
SLIDE 11
Multiplicative bias m as a function of separation
SLIDE 12
Multiplicative bias m (zoomed in)
SLIDE 13
Main Shear Test Results
◮ MOF has expected behavior
◮ No bias for high S/N objects ◮ No bias at large separations ◮ Small bias for close separations and moderate S/N.
◮ Scarlet shows surprising behavior
◮ Large bias at high S/N ◮ Unpredictable bias as a function of separation at low S/N
◮ We have been working closely with Peter and Fred to figure
this out.
SLIDE 14
A Clue: Recovered Position Offsets, Relative to Truth
MOF Scarlet
SLIDE 15
Centering issues
◮ The MOF behavior is as expected: the best-fit positions
are centered on the truth with some Gaussian scatter.
◮ Scarlet positions look erratic. ◮ Not shown: sometimes the center can jump far outside the
image, like 1010 pixels away.
◮ Fred and Peter think they understand this behavior and
are working on a fix.
SLIDE 16
Summary
◮ Scarlet is currently showing some surprising behavior for
shear recovery.
◮ We are working with Fred and Peter to fix it. ◮ Lessons learned
◮ For weak lensing, simpler may be better. Should we adopt
different constraints for Scarlet when using it for weak lensing?
◮ Scarlet is very promising but if you want it to work for your
science case, then you need to test it!
◮ The scarlet developers are very responsive and helpful.
SLIDE 17
Extra Slides
SLIDE 18
Detection vs Deblending
◮ For faint blobs detection is more of an issue than
deblending.
◮ If you know the positioins of the objects in the field, MOF
will produce a model with small residuals
◮ If the detection is imperfect, there will be large residuals,
independent of the deblender or model that is fit.
SLIDE 19
Example Deblend
256 objects, 220 found by SExtractor
SLIDE 20