Bayesian Component Separation for DESI LAE Automated Spectroscopic Redshifts and Photometric Targeting

1. Problem Statement

Lyman-Alpha Emitters (LAEs) are high-redshift, star-forming galaxies crucial for studying the epoch of reionization and large-scale structure. However, identifying them and determining their spectroscopic redshifts using ground-based instruments like the Dark Energy Spectroscopic Instrument (DESI) presents significant challenges:

• Sky Residuals: Ground-based spectroscopy suffers from atmospheric sky emission lines. Even after standard sky subtraction, residual lines often remain.
• Line Confusion: LAEs typically exhibit only one narrow, relatively faint emission line (Ly$\alpha$). Automated pipelines often struggle to distinguish this single line from residual sky lines, leading to misclassified redshifts.
• Scalability: Current DESI redshift determination relies heavily on visual inspection (VI), which is not scalable for the millions of spectra expected in future surveys (e.g., DESI-2, which aims to observe $\approx$1.5 million LAEs).
• Pipeline Limitations: Existing automated pipelines (e.g., Redrock) use Principal Component Analysis (PCA) templates designed for stars, galaxies, and quasars, but they fail to accurately identify LAEs or marginalize over sky residuals effectively.

2. Methodology: MADGICS

The authors propose a Bayesian component separation technique called MADGICS (Marginalized Analytic Dataspace Gaussian Inference for Component Separation). Instead of blindly decomposing data, MADGICS uses data-driven priors to separate a spectrum into specific physical components.

Core Framework

The observed spectrum $x_{tot}$ is modeled as a sum of $k=3$ components:

1. LAE Component: The target Ly$\alpha$ emission.
2. Sky Residual Component: Systematic residuals from imperfect sky subtraction.
3. Residual Component: Random noise and unmodeled features.

The method assumes the data is a linear combination of these components with known covariance structures. The total covariance matrix is:
$$C_{tot} = C_{sky} + C_{LAE} + C_{res}$$

The individual components are extracted using the formula:
$$x_i = C_i C_{tot}^{-1} x_{tot}$$

Prior Generation (Data-Driven)

• Sky Residual Prior ($C_{sky}$): Constructed from 19,657 DESI sky residual spectra (from ODIN targets). Prominent sky lines are masked, and the covariance is approximated using the top 50 eigenvectors. A flux rescaling is applied to account for noise variations at different wavelengths.
• LAE Prior ($C_{LAE}$): Constructed from 4,005 high-quality, visually inspected ODIN LAE spectra (redshifts $z=2.3-4.5$). These spectra are shifted to a common redshift ($z=2.45$), sky residuals are removed via decomposition, and the covariance is calculated using the top 2 eigenvectors to capture the Ly$\alpha$ line shape (including blue cutoffs and red wings).
• Residual Prior ($C_{res}$): A diagonal matrix derived from the inverse variance of the specific target spectrum.

Redshift Determination & Confidence

• $\Delta\chi^2$ Metric: The pipeline scans redshifts ($z=2$ to $4$) and calculates the improvement in fit ($\Delta\chi^2$) when an LAE component is added:
  $$\Delta\chi^2(z) = \chi^2(\text{sky} + \text{res}) - \chi^2(\text{sky} + \text{LAE}(z) + \text{res})$$
  A large negative $\Delta\chi^2$ (or high $|\Delta\chi^2|$) indicates a confident detection.
• Uncertainty Quantification: Redshift uncertainties are derived from the curvature of the $\Delta\chi^2$ profile around the minimum.
• Calibration: The authors use injection-recovery tests (injecting simulated LAE signals into real sky residuals) to calibrate the $\Delta\chi^2$ surface and uncertainty estimates, correcting for model mismatches.

3. Key Contributions

1. Automated LAE Redshift Pipeline: Development of a scalable, automated method to determine spectroscopic redshifts for LAEs, replacing the need for visual inspection.
2. Marginalization over Sky Residuals: Unlike traditional methods that subtract a sky model, this method jointly infers the sky residual and LAE components, effectively marginalizing over systematic sky line uncertainties.
3. Data-Driven Priors: Creation of robust, data-driven covariance priors for both sky residuals and LAE spectral shapes, avoiding reliance on theoretical templates that may not match reality.
4. Confidence Metric: Introduction of $|\Delta\chi^2|$ as a proxy for detection Signal-to-Noise Ratio (SNR) and a tool for filtering false positives.
5. Survey Design Recommendations: Analysis of exposure time requirements and photometric selection cuts (using medium-band filters) to optimize future LAE targeting.

4. Results & Validation

The method was tested on 881 DESI spectra of LAE candidates selected from Subaru Suprime-Cam medium-band imaging (a held-out test set not used for prior generation).

• Accuracy: The pipeline determined redshifts within 0.005 of the visual inspection (VI) ground truth for 92.3% (812/881) of the targets.
• Detection Threshold: A threshold of $|\Delta\chi^2| \approx 25$ (roughly $5\sigma$) effectively separates true LAEs from non-detections or contaminants.
• Injection Tests:
  • In simulated spectra with uniform noise, 50% of redshifts were recovered correctly at SNR $\approx 3$, and 100% at SNR $\approx 6$.
  • Injections into real sky residuals showed similar performance, with 100% recovery at SNR $\approx 5$.
• Uncertainty Calibration: After applying a calibration factor ($s=2.16$) derived from injection tests, the redshift uncertainties were well-calibrated. The pipeline achieves accurate redshifts within $3\sigma$ for 81.7% of targets.
• Computational Efficiency: The method processes a single spectrum in approximately 0.4 seconds. Processing all 50 million DESI spectra would take $\sim$6,000 CPU hours.

5. Significance and Implications for DESI-2

This work provides a critical foundation for the upcoming DESI-2 mission, which will focus heavily on high-redshift targets.

• Scalability: The automated pipeline enables the analysis of millions of spectra, a necessity for the planned 1.5 million LAE observations.
• Targeting Strategy: The authors correlate $|\Delta\chi^2|$ with color-magnitude diagrams to recommend selection cuts for medium-band (MB) photometry (e.g., $MB_{mag} < 24$ and $(MB-g) < -1$). This suggests that MB photometry (wider filters, $\sim$250 Å) can be used effectively to pre-select LAE candidates, potentially replacing narrower filters and increasing survey efficiency.
• Exposure Time Optimization: The study quantifies the exposure time needed to converge on a redshift based on Ly$\alpha$ flux. It highlights that fainter sources ($< 10^{-16}$ erg s$^{-1}$ cm$^{-2}$) require significantly longer exposures (>3.5 hours) to achieve reliable redshifts.
• Generalizability: The MADGICS framework is not limited to LAEs; it can be applied to other spectral features (e.g., Lyman Break Galaxies, AGN, gravitational lensing candidates) and even photometric data for star-galaxy separation.

In conclusion, the paper presents a robust, Bayesian solution to the "sky line confusion" problem in LAE spectroscopy, offering a path toward fully automated, high-precision cosmological surveys of the early universe.