Highly Efficient Selection of High-Redshift Emission-Line Galaxies for future DESI-like surveys with Deep Multi-band Imaging

Technical Summary: Highly Efficient Selection of High-Redshift Emission-Line Galaxies for Future DESI-like Surveys with Deep Multi-band Imaging

Problem Statement
Emission-line galaxies (ELGs) are critical tracers for measuring Baryon Acoustic Oscillations (BAO) and large-scale structure at redshifts $z > 1$. However, current selection methods for the Dark Energy Spectroscopic Instrument (DESI) face significant limitations in the redshift range $1.1 < z < 1.6$. The existing DESI ELG selection, based on the shallower $g, r, z$ photometry from the Legacy Imaging Surveys, suffers from low efficiency and significant density fluctuations. Specifically, the current DESI sample yields only $\sim 660$ robust ELGs per square degree in the target redshift range, with a low redshift range success rate (34%) and a redshift measurement success rate of 69%. This density places the sample in a shot-noise limited regime, hindering precise cosmological constraints on the dark energy equation of state. The paper posits that deeper, multi-band imaging (similar to the Rubin Observatory LSST or Hyper Suprime-Cam) can overcome these systematics by providing better spectral feature identification.

Methodology
The authors developed and optimized a new target selection algorithm designed to increase the ELG surface density yield by a factor of $\sim 3$ compared to current DESI targets. The methodology involved three primary stages:

1. Data Sources:

   • Photometry: The team utilized deep $ugrizy$ photometry from the COSMOS2020 catalog (Farmer version) to train selection criteria. This dataset offers depths $\sim 2$ magnitudes deeper than the Legacy Survey and includes $i$ and $y$ bands.
   • Spectroscopy: To empirically test and refine the selection, the authors used a "spec-truth" sample. This sample consists of 87,512 ELG-like objects observed with DESI in the COSMOS field, selected using broad color cuts from deep DECam and HSC imaging to capture all potential scatterers into the DESI sample.
   • Redshifts: Photometric redshifts ($z_{photo}$) were derived using the LePHARE code on the COSMOS2020 many-band data. Spectroscopic redshifts ($z_{spec}$) were obtained from DESI observations with effective exposure times between 700 and 1400 seconds.

2. Algorithm Development:

   • Initial Classification: A Random Forest (RF) classifier was trained on COSMOS2020 $ugrizy$ photometry to distinguish high-redshift ELGs ($z=1.1-1.6$) from other galaxies. The RF achieved an Area Under the Curve (AUC) of 0.960.
   • Feature Importance: Permutation importance analysis revealed that the color combinations $r-i$ and $i-y$ were the most influential features, with $i-z$ being the next most significant.
   • Optimization: The authors defined a set of color cuts and a limiting magnitude ($g_{fiber}$) based on the RF insights. These cuts were optimized using the spec-truth sample to maximize the "redshift range success rate" (the fraction of targets with secure $z_{spec}$ in $1.1 < z < 1.6$) while targeting a net surface density yield of $\sim 1370$ deg$^{-2}$. The optimization minimized a loss function balancing the success rate against the deviation from the target density.

3. Metrics:

   • Redshift Measurement Success Rate ($f_{reliable}$): Fraction of targets with secure $z_{spec}$ (based on [O II] flux S/N and $\Delta\chi^2$).
   • Redshift Range Success Rate: Fraction of targets with secure $z_{spec}$ specifically within $1.1 < z < 1.6$.
   • Net Surface Density Yield: The product of target density and the redshift range success rate.

Key Results
The optimized selection criteria, consisting of limits on $g_{fiber}$, $r-i$, $i-y$, and $i-z$, demonstrated significant improvements over the standard DESI ELG selection:

• Redshift Measurement Success Rate: Increased from 69% (DESI) to 89% (This work).
• Redshift Range Success Rate: Increased from 34% (DESI) to 84% (This work).
• Net Surface Density Yield: Increased from 660 deg$^{-2}$ (DESI) to 1372 deg$^{-2}$ (This work).
• Target Density: The new sample achieves a target density of 1632 deg$^{-2}$, which is comparable to the DESI target density (1940 deg$^{-2}$) but yields a much higher fraction of useful objects.
• Contamination: The new cuts effectively reject low-redshift contamination ($z < 1.1$) and objects at $z > 1.6$ where the [O II] doublet falls outside the DESI spectral window.

Significance and Claims
The paper claims that combining the current DESI ELG sample with this optimized sample would increase the net ELG number density by a factor of $\sim 3$. Based on Fisher matrix forecasts (FishLSS code), this increase would:

1. Move the ELG sample out of the shot-noise limited regime for the $1.1 < z < 1.6$ redshift bin.
2. Reduce the uncertainties on the BAO distance scale parameters ($\alpha_{\perp}$ and $\alpha_{\parallel}$) by a factor of $\sim 2$ at these redshifts.

The authors emphasize that this approach is feasible for future spectroscopic surveys (e.g., DESI-II) utilizing deeper, multi-band imaging like that from the Rubin Observatory or HSC. They note that the selection is robust even when applied to slightly shallower "LSST Year 2" depth photometry. Furthermore, the paper suggests that such a high-efficiency sample could be observed during gray or bright time (when the moon is present), as the [O II] line at $z > 1.1$ shifts to wavelengths ($> 7800$ Å) where night-sky emission lines dominate over the moonlight continuum, potentially increasing the efficiency of future survey operations. Finally, the authors propose that this selection method could generate purely photometric samples with well-defined redshift distributions for CMB cross-correlation and lensing analyses at higher redshifts than currently possible with LRGs.