🔭 astrophysics

DESI Data Release 2 ELGs: Property-dependent subsamples, imaging systematics, and clustering

Using DESI Data Release 2 emission-line galaxies, this study demonstrates that implementing property-dependent systematic weights on distinct subsamples, particularly within the deeper Dark Energy Survey footprint, effectively mitigates spurious clustering signals and improves the accuracy of clustering measurements compared to the standard fiducial approach.

Original authors: T. Hagen, K. S. Dawson, Z. Zheng, J. Aguilar, S. Ahlen, D. Bianchi, D. Brooks, T. Claybaugh, A. de la Macorra, B. Dey, S. Ferraro, J. E. Forero-Romero, S. Gontcho A Gontcho, G. Gutierrez, J. Guy, C. H

Published 2026-06-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: T. Hagen, K. S. Dawson, Z. Zheng, J. Aguilar, S. Ahlen, D. Bianchi, D. Brooks, T. Claybaugh, A. de la Macorra, B. Dey, S. Ferraro, J. E. Forero-Romero, S. Gontcho A Gontcho, G. Gutierrez, J. Guy, C. Hahn, M. Ishak, R. Joyce, S. Juneau, A. Kremin, O. Lahav, C. Lamman, M. Landriau, L. Le Guillou, M. Manera, A. Meisner, R. Miquel, J. Moustakas, A. D. Myers, S. Nadathur, J. A. Newman, G. Niz, W. J. Percival, C. Poppett, F. Prada, I. Perez-Rafols, A. J. Ross, G. Rossi, S. Saito, E. Sanchez, D. Schlegel, J. Silber, G. Tarle, B. A. Weaver, H. Zou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a perfect group photo of millions of galaxies to understand how the universe is structured. You want to count exactly how many galaxies are in each part of the sky and how they are clustered together. However, your camera (the telescope and its sensors) isn't perfect. Sometimes the lens is a bit smudged, sometimes the lighting is uneven, and sometimes the background is too bright or too dark. These "camera glitches" can make it look like there are more or fewer galaxies in certain areas than there really are. If you don't fix these glitches, your final photo will be distorted, and your conclusions about the universe will be wrong.

This paper is about fixing those camera glitches for a specific group of galaxies called Emission-Line Galaxies (ELGs) using data from the Dark Energy Spectroscopic Instrument (DESI).

Here is the story of what they did, explained simply:

1. The Problem: The "Smudged Lens"

DESI is a massive project taking pictures of the sky to map the universe. To do this, it uses images taken by different cameras over many years.

The North vs. The South: The sky is split into a Northern and Southern hemisphere. The Northern part was photographed by one set of cameras, and the Southern part by another.
The "Deep" Camera: In a specific part of the South, there is an even deeper, higher-quality camera (from the Dark Energy Survey, or DES). It sees fainter, more distant objects better than the others.
The Glitch: Because the cameras are different, the "rules" for spotting a galaxy change. In the deep South, the camera sees so much detail that it misses some of the "weirdly colored" galaxies that the other cameras catch. This creates a fake pattern in the data. It looks like the galaxies are clustering in a certain way, but it's actually just an artifact of the different cameras.

2. The Old Solution: The "AI Filter"

To fix these glitches, the DESI team usually uses a very smart computer program (a Neural Network, or "AI") to learn how the camera glitches affect the data. It's like having a super-smart editor who looks at the photo and says, "Ah, this area looks too bright because of the lens, so I'll darken it to make it look real."

The Risk: While this AI is powerful, it's a bit of a "black box." It might get too creative and accidentally erase real patterns along with the fake ones, or it might not be flexible enough to handle every specific type of galaxy.

3. The New Idea: The "Specialized Filters"

The authors of this paper asked: What if we didn't use one giant AI for everyone? What if we sorted the galaxies into different groups based on their properties (like their mass, brightness, or color) and gave each group its own specific filter?

They tried two things:

Sorting by Property: They split the galaxies into groups based on things like how heavy they are (stellar mass) or how bright they are.
Sorting by "Weirdness": They looked at how far a galaxy's color was from the "normal" range of stars. Some galaxies were so far off the normal color chart that they were likely just measurement errors caused by the shallow cameras.

4. The Big Discovery: The "Deep South" Needs Its Own Rule

The most important finding was about the Deep South (DES) region.

The Analogy: Imagine you are counting fish in a pond. In the shallow part of the pond, you can see the weird, colorful fish easily. In the deep part, the water is so clear and deep that you only see the normal fish; the weird ones are too faint to be picked up by the sensors there.
The Result: The authors found that the "weirdly colored" galaxies were almost entirely missing from the Deep South. Because the standard method treated the whole South as one big area, it got confused.
The Fix: They realized they needed to treat the Deep South as a completely separate region from the rest of the South. By splitting the data this way, they could create a much more accurate map. It's like realizing you need two different counting rules for the shallow water and the deep water, rather than trying to use one rule for the whole pond.

5. The Outcome: A Clearer Picture

When they applied these new, specialized filters and split the Deep South into its own category:

For the whole group: The old AI method was still the best overall.
For specific groups: For about 10% of the galaxy groups (specifically the heavy ones, the dim ones, and the "weirdly colored" ones), the new specialized method worked much better. It removed the fake patterns that the old AI missed.
The "Weird" Galaxies: They also found that the galaxies with the most extreme colors were likely just measurement errors from the shallow cameras. They probably don't exist in the deep universe at all.

Summary

Think of this paper as a team of photo-editors realizing that their "one-size-fits-all" filter wasn't working for a specific type of photo. They discovered that the "Deep South" part of the sky was so different from the rest that it needed its own special rules. By splitting the data and giving specific groups of galaxies their own custom corrections, they cleaned up the map of the universe, removing the fake patterns caused by the different cameras and leaving a clearer view of how the universe is actually structured.

Technical Summary: DESI Data Release 2 ELGs: Property-dependent subsamples, imaging systematics, and clustering

Problem Statement
Galaxy redshift surveys like the Dark Energy Spectroscopic Instrument (DESI) aim to measure cosmological signals such as baryon acoustic oscillations (BAO) and redshift space distortions (RSD). However, inhomogeneities in imaging data—driven by variations in survey depth, seeing, and foreground Milky Way properties—induce artificial fluctuations in the projected galaxy density field. If uncorrected, these imaging systematics create spurious clustering signals that can bias cosmological parameter estimation.

While the DESI collaboration employs systematic weights to mitigate these effects, the standard approach for Emission-Line Galaxies (ELGs) relies on a fully connected Feed Forward Neural Network (SYSNet). Although flexible, machine learning approaches risk overcorrection and the removal of true cosmological signal. Furthermore, ELGs are the faintest tracer in DESI, with fluxes near the detection limits of the imaging surveys, making them uniquely sensitive to imaging depth variations. Previous analyses of DESI Year 1 data indicated residual spurious signals on scales beyond the BAO, suggesting the current weighting scheme may be insufficient for complex samples.

Methodology
This work proposes an alternative derivation of systematic weights for DESI Data Release 2 (DR2) ELGs, focusing on the redshift range $0.8 < z < 1.1$ . The methodology involves three primary components:

Property-Dependent Subsampling: Instead of treating all ELGs as a single homogeneous sample, the authors define subsamples based on five specific galaxy properties:
- Stellar mass ( $\log M_*$ )
- [OII] luminosity ( $\log L_{[OII]}$ )
- g-band fiber magnitude ( $g_{fib}$ )
- Perpendicular distance from the redshift-optimizing boundary in the $g-r$ vs. $r-z$ plane ( $D_{grz}^z$ )
- Perpendicular distance from the stellar locus boundary in the $g-r$ vs. $r-z$ plane ( $D_{grz}^s$ )
Linear Regression for Systematic Weights: For each subsample, the authors rederive systematic weights using a multilinear regression method (identical to that used for other DESI tracers like LRGs and quasars) rather than the fiducial neural network. This approach models the projected galaxy density as a linear function of ten imaging properties (including stellar density, HI column density, and depth/seeing in $g, r, z$ filters) and inverts the model to compute weights. This provides physical flexibility without the overfitting risks associated with machine learning.
Refined Photometric Region Treatment: The authors identify that the standard North/South photometric distinction is insufficient for ELGs in the Southern sky. They propose splitting the Southern region into "South-DES" (Dark Energy Survey footprint) and "South-DECaLS" (Legacy Survey). This is motivated by the observation that the deeper DES imaging produces different ELG number densities and redshift distributions compared to the shallower DECaLS regions, particularly for subsamples far from the stellar locus.

Key Results

Subsample Variability: The analysis reveals that ELG properties significantly influence their response to imaging systematics. Subsamples defined by high stellar mass, low [OII] luminosity, and large distance from the stellar locus ( $D_{grz}^s$ ) exhibit the most significant deviations when using the new linear regression weights compared to the fiducial SYSNet weights.
Mitigation of Spurious Signal: For approximately 10% of the subsamples (specifically those with high stellar mass, low [OII] luminosity, and high $D_{grz}^s$ ), the property-dependent linear weights further mitigate spurious clustering signals on scales around $r_p \sim 10 \, h^{-1}\text{Mpc}$ , bringing the projected two-point correlation function (2PCF) slopes closer to theoretical expectations.
Impact of DES Footprint: The study demonstrates that ELGs far from the stellar locus are largely absent from the deep DES footprint; their presence in the shallower Legacy Survey regions is likely due to measurement scatter. Consequently, treating the South as a single region leads to inaccurate random catalogs and regression outliers. Separating South-DES from South-DECaLS preserves the distinct density and redshift distributions of these regions, preventing the artificial suppression of the DES overdensity by the weighting scheme.
Full Sample Behavior: When applied to the full ELG sample, the property-dependent linear weights (even with the refined region treatment) result in excess large-scale signal compared to the fiducial SYSNet approach. The authors conclude that while the neural network remains optimal for the full sample, the linear, property-dependent approach is superior for specific subsamples.

Significance and Claims
The paper claims that a "one-size-fits-all" approach to systematic weights is insufficient for the diverse population of DESI ELGs. The primary contributions are:

Validation of Subsample-Specific Weights: Demonstrating that customizing systematic weights based on galaxy properties can improve clustering measurements for specific subsets of the data, particularly those most vulnerable to imaging systematics.
Necessity of Region-Specific Treatment: Establishing that the DES footprint requires separate treatment from the rest of the Southern survey in the LSS catalog production pipeline to accurately model random catalogs and systematic weights.
Physical Insight into Target Selection: Highlighting that ELGs with extreme colors (far from the stellar locus) are likely artifacts of measurement scatter in shallow imaging rather than physical populations, suggesting that cuts on $D_{grz}^s$ could produce more consistent samples.

The authors maintain that while their method offers improvements for subsamples, the fiducial neural-network-based weights remain the optimal choice for the full ELG sample in the context of DESI's primary cosmological analyses. They suggest that future DESI releases and other surveys (such as the Vera Rubin Observatory) should consider property-dependent treatments and refined photometric region distinctions to enhance the accuracy of clustering measurements.