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 to test Scarlet statistically for weak lensing shear recovery. ◮ Test in regime most interesting for weak lensing: objects are small blobs after pixelization and smearing by the PSF. ◮ Use the Multi-Object Fitter (MOF) from Dark Energy Survey as a benchmark. Needed a new implementation because original is not a library (Matt Becker using my ngmix package).

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

Example Scarlet Deblend from HSC 33 33 33 Data 31 SCARLET 31 Residual 31 30 30 30 32 32 32 28 28 28 29 29 29 18 18 18 17 17 17 16 16 16 23 24 23 24 23 24 26 26 26 15 15 15 14 14 14 27 27 27 22 22 22 20 13 20 13 20 13 25 25 25 21 21 21 12 12 12 10 19 10 19 10 19 9 9 9 11 11 11 5 6 5 6 5 6 7 7 7 4 4 4 2 2 2 3 3 3 8 8 8 1 1 1 0 0 0

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.

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

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.

Example Simulated Galaxies (Large)

Example MOF de-blended pair. Low Noise. Left is original, right is MOF deblended. Can see residuals from “knots” of star formation.

Example MOF de-blended pair. Medium Noise (S/N ∼ 50.) Left is original, right is MOF deblended.

Multiplicative bias m as a function of separation

Multiplicative bias m (zoomed in)

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.

A Clue: Recovered Position Offsets, Relative to Truth MOF Scarlet

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 10 10 pixels away. ◮ Fred and Peter think they understand this behavior and are working on a fix.

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.

Extra Slides

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.

Example Deblend 256 objects, 220 found by SExtractor

Example Deblend 256 objects, 220 found by SExtractor