betaFeedback

WorksPhysics and Astronomy

FastSpecFit: Spectrophotometric Modeling of 38 Million Extragalactic Targets from the Dark Energy Spectroscopic Instrument Second Data Release

Abstract

Optical spectrophotometry provides a direct window into the physical conditions, kinematics, and formation histories of galaxies and quasars. FastSpecFit is an open-source Python software package that forward-models each Dark Energy Spectroscopic Instrument (DESI) spectrum jointly with its broadband photometry, convolving stellar-population synthesis and emission-line templates with each fiber's native spectral resolution. This forward-modeling approach preserves DESI's uncorrelated pixel statistics, fits a typical spectrum in ~0.1 s of single-core time, and scales to tens of millions of objects, with uncertainties estimated via Monte Carlo resampling. FastSpecFit measures stellar velocity dispersions, non-parametric star-formation histories, stellar masses, star-formation rates, dust attenuation, K-corrections, rest-frame magnitudes, and fluxes and equivalent widths for 46 emission lines, including separate narrow and broad components for the Balmer lines. We describe these algorithms in detail, validate the measurements against independent codes, and present the resulting value-added catalog for DESI Data Release 2, which provides spectrophotometric and physical properties for 38 million extragalactic targets at 0<z<4. We also release a companion photometry-only catalog and document the catalogs previously released for the DESI Early Data Release and Data Release 1.

Corresponding author: John Moustakas {sup}12` Department of Physics & Astronomy, University College London, Gower Street, London, WC1E 6BT, UK

{sup}13` Departamento de Astrofísica, Universidad de La Laguna (ULL), E-38206, La Laguna, Tenerife, Spain

{sup}14` Instituto de Astrofísica de Canarias, C/ Vía Láctea, s/n, E-38205, La Laguna, Tenerife, Spain

{sup}15` Institut d’Estudis Espacials de Catalunya (IEEC), c/ Esteve Terradas 1, Edifici RDIT, Campus PMT-UPC, 08860 Castelldefels, Spain

{sup}16` Institute of Space Sciences, ICE-CSIC, Campus UAB, Carrer de Can Magrans s/n, 08913 Bellaterra, Barcelona, Spain

{sup}17` Instituto de Física, Universidad Nacional Autónoma de México, Circuito de la Investigación Científica, Ciudad Universitaria, Cd. de México C. P. 04510, México

{sup}18` Department of Astronomy & Astrophysics, University of Toronto, Toronto, ON M5S 3H4, Canada

{sup}19` Department of Physics & Astronomy and Pittsburgh Particle Physics, Astrophysics, and Cosmology Center (PITT PACC), University of Pittsburgh, 3941 O’Hara Street, Pittsburgh, PA 15260, USA

{sup}20` European Southern Observatory, Karl-Schwarzschild-Straße 2, 85748 Garching bei München, Germany

{sup}21` Institució Catalana de Recerca i Estudis Avançats, Passeig de Lluís Companys, 23, 08010 Barcelona, Spain

{sup}22` Institut de Física d’Altes Energies (IFAE), The Barcelona Institute of Science and Technology, Edifici Cn, Campus UAB, 08193 Bellaterra (Barcelona), Spain

{sup}23` Departamento de Física, Universidad de los Andes, Cra. 1 No. 18A-10, Edificio Ip, CP 111711, Bogotá, Colombia

{sup}24` Observatorio Astronómico, Universidad de los Andes, Cra. 1 No. 18A-10, Edificio H, CP 111711 Bogotá, Colombia

{sup}25` Institute of Cosmology and Gravitation, University of Portsmouth, Dennis Sciama Building, Portsmouth, PO1 3FX, UK

{sup}26` Fermi National Accelerator Laboratory, PO Box 500, Batavia, IL 60510, USA

{sup}27` Institute of Astronomy & Kavli Institute for Cosmology, University of Cambridge, Madingley Road, Cambridge CB3 0HA, UK

{sup}28` Department of Physics, University of Wisconsin-La Crosse, La Crosse, WI 54601, USA

{sup}29` Department of Astronomy, University of Texas at Austin, 2515 Speedway, Austin, TX 78712, USA

{sup}30` Center for Cosmology and AstroParticle Physics, The Ohio State University, 191 West Woodruff Avenue, Columbus, OH 43210, USA

{sup}31` Department of Physics, The Ohio State University, 191 West Woodruff Avenue, Columbus, OH 43210, USA

{sup}32` NSF NOIRLab, 950 N. Cherry Ave., Tucson, AZ 85719, USA

{sup}33` Department of Physics and Astronomy, University of California, Irvine, 2115 Frederick Reines Hall, Irvine, CA 92697, USA

{sup}34` Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge CB3 0HA, UK

{sup}35` Department of Astronomy, The Ohio State University, 4055 McPherson Laboratory, 140 W 18th Avenue, Columbus, OH 43210, USA

{sup}36` Sorbonne Université, CNRS/IN2P3, Laboratoire de Physique Nucléaire et de Hautes Energies (LPNHE), FR-75005 Paris, France

{sup}37` Department of Astronomy and Astrophysics, UCO/Lick Observatory, University of California, 1156 High Street, Santa Cruz, CA 95064, USA

{sup}38` Department of Physics & Astronomy, University of Wyoming, 1000 E. University, Dept. 3905, Laramie, WY 82071, USA

{sup}39` Department de Física, Serra Húnter, Universitat Autònoma de Barcelona, 08193 Bellaterra (Barcelona), Spain

{sup}40` SLAC National Accelerator Laboratory, 2575 Sand Hill Road, Menlo Park, CA 94025, USA

{sup}41` Instituto de Estudios Astrofísicos, Facultad de Ingeniería y Ciencias, Universidad Diego Portales, Av. Ejército Libertador 441, Santiago, Chile

{sup}42` Department of Physics and Astronomy, University of Waterloo, 200 University Ave W, Waterloo, ON N2L 3G1, Canada

{sup}43` Perimeter Institute for Theoretical Physics, 31 Caroline St. North, Waterloo, ON N2L 2Y5, Canada

Keywords: Astronomy software (1855); Catalogs (205); Galaxy properties (615); Spectrophotometry (1556); Stellar populations (1622)

Introduction

The Dark Energy Spectroscopic Instrument (DESI; M. Levi et al. [84]; DESI Collaboration et al. [34, 37]) is conducting an eight-year survey (2021–2029) which will ultimately yield spectroscopic redshifts for more than 60 million galaxies and quasars over 17,000 deg2^2. These data are being made public in a series of staged releases. The Early Data Release (EDR; DESI Collaboration et al. [38]), comprising 1.7 million unique spectra from the survey validation (SV) campaign (2020 December–2021 May), was released in 2023 June. Data Release 1 (DR1; DESI Collaboration et al. [39]), released in 2025 March, uniformly reprocessed the SV data and added the first 13 months of the main-survey sample, for a total of 20.4 million unique spectra. Data Release 2 (DR2), the subject of this paper, extends the main-survey data through 2024 April, its first three years, includes an unprecedented 38 million unique redshifts, and is scheduled for public release in the first part of 2027. Each successive release has sharpened DESI’s baryon acoustic oscillation distance measurements to sub-percent precision and currently shows a 2.8–4.2σ\sigma preference for an evolving dark energy equation of state over the standard Λ\LambdaCDM concordance cosmological model (A. G. Adame et al. [3]; M. Abdul Karim et al. [1]).

Each spectrum DESI acquires encodes a wealth of astrophysical information beyond its redshift. The shape and absorption-line content of the stellar continuum trace the underlying stellar populations, dust content, and kinematics of a galaxy, while its emission-line spectrum constrains the star-formation rate, chemical abundances, ionization state, and gas kinematics, as well as the presence and strength of an active galactic nucleus (AGN) (C. Conroy [28], and references therein). Extracting these physical properties at DESI scale—and combining them with broadband photometry to normalize the overall spectral energy distribution (SED)—requires a fitting code that is fast, robust, and automated across an enormous range of signal-to-noise ratios (S/N), redshifts, and galaxy types.

Various publicly available codes already perform this kind of spectrophotometric modeling. Full-spectrum fitting codes—among them pPXF (M. Cappellari & E. Emsellem [21]; M. Cappellari [20]), GandALF (M. Sarzi et al. [114, 115]), STARLIGHT (R. Cid Fernandes et al. [27]), FIREFLY (D. M. Wilkinson et al. [138]), FIт3D/pyPipe3D (S. F. Sánchez et al. [111, 112]), ULySS (M. Koleva et al. [72]), alf (C. Conroy et al. [31]), and others—extract the kinematics and properties of the stellar populations and line-emitting gas; SED-fitting codes such as CIGALE (M. Boquien et al. [16]), Bagpipes (A. C. Carnall et al. [22]), Prospector (B. D. Johnson et al. [66]), and MAGPHYS (E. da Cunha et al. [32]) combine photometry (and, in some cases, spectroscopy) with physically motivated stellar population synthesis (SPS) models to infer star-formation histories, stellar masses, and dust properties; and codes like PyQSOFit ([51]) are tailored to fitting the ultraviolet (UV) and optical spectra of quasars.

Each of these codes makes different trade-offs among flexibility, physical realism, and computational speed, but none were designed to be run efficiently on the tens of millions of spectra acquired by DESI. Existing codes also do not take advantage of the DESI resolution matrix, which compactly encodes the instrumental resolution and guarantees uncorrelated uncertainties from pixel to pixel ([15, 53]).

FastSpecFit1 has been developed to meet both these challenges, scaling efficiently to tens of millions of spectra while fully exploiting DESI’s native spectral resolution. Using a forward-modeling fitting approach, FastSpecFit convolves physically motivated stellar continuum and emission-line templates with each fiber’s resolution matrix and compares the result directly to the observed spectra, preserving the uncorrelated pixel statistics that resampling would otherwise destroy (§ 3). In addition, FastSpecFit is engineered for speed; it fits a typical DESI spectrum in ≈ 0.1 s on a single CPU core (≈ 5 s including the default 50 Monte Carlo realizations; § 6.3), making it possible to process the full DESI dataset—and to refit it as the data-reduction pipeline and SPS models evolve—with modest computing resources. This speed also makes it practical to estimate parameter uncertainties from repeated fits to Monte Carlo-resampled data rather than from computationally expensive posterior sampling (§ 4.3), and, via a hybrid message passing interface (MPI) plus multiprocessing design, to scale from a single serial process on a laptop to thousands of cores on a supercomputer (§ 6.3).

Beyond speed, FastSpecFit is designed to be flexible, applying only the physical constraints the data strictly require and otherwise letting the data determine the best-fitting model. The continuum model imposes only minimal priors: non-negative stellar population coefficients, an energy-conserving dust model, and a freely fitted stellar velocity dispersion bounded widely enough to accommodate the full spectral diversity of DESI targets. In exchange for speed and robustness at low S/N, the default configuration we adopt deliberately fixes other ingredients, including the stellar metallicity, the slope of the attenuation curve, and the infrared (IR) dust-emission parameters (§ 3.1, § 3.4.3), although users can relax most of these choices with custom templates and fitting options (§ 4.4). The emission-line model is similarly flexible, supporting independent narrow and broad kinematic components and capturing the broad Balmer-line emission that distinguishes type 1 AGN whenever the data require it. Unlike codes that target a specific class or spectral type, FastSpecFit applies this same machinery uniformly to every spectrum, without assuming in advance whether the target is quiescent, star-forming, post-starburst, or hosts an AGN. This flexibility minimizes classification-driven selection effects and maximizes the discovery potential of a dataset as large and diverse as DESI, where unusual or physically ambiguous objects emerge naturally from the fit rather than requiring special-case handling.

Although this paper presents the first comprehensive description of FastSpecFit’s algorithms, measurements from the earlier EDR and DR1 value-added catalogs (VACs; see Appendix § 11) have already enabled a broad range of extragalactic science by both DESI collaboration members and the wider scientific community. We briefly highlight a small, illustrative sample of that work here. Emission-line fluxes, equivalent widths, and stellar masses from FastSpecFit have been used to identify and characterize changing-look AGN [52] and to triple the census of dwarf AGN candidates in the local universe (R. Pucha et al. 2025); to measure direct-method (TeT_e-based) chemical abundances for nearly 50,000 star-forming galaxies in DESI DR2 [120], including two of the most metal-poor star-forming galaxies ever found (J. Moustakas et al., in preparation), and to characterize the atomic gas sequence and mass-metallicity relation across five decades in stellar mass [119]; to assemble catalogs of post-starburst galaxies [127] and to infer the physical properties of galaxies in cosmic voids [105]; and to measure stellar velocity dispersions for strong gravitational lens candidates [61]. The FastSpecFit measurements have also been essential to measurements of the galaxy luminosity and stellar functions [136, 91], and to the modeling of intrinsic alignments [123] and photometric-redshift calibration [11] for the next generation of weak-lensing surveys.

Although this paper is long, we have striven to write it so it can be read in pieces as well as start to end. In § 2 we describe the DESI spectroscopic and photometric data FastSpecFit fits, together with the survey, program, and target-class structure particular to DESI needed to interpret the resulting catalogs. The technical core of the paper, § 3, presents FastSpecFit’s algorithms and the physical quantities it derives in full detail; as noted there, readers who want the output before the machinery can skip this section. In § 5 we ground that machinery in three worked examples spanning the DESI extragalactic classes, a natural starting point for readers who want to see spectra and their best-fitting models to help build intuition first. In § 6 we present the DR2 catalog itself—its construction and organization, recommended quality cuts, computational performance, and validation against independent codes. Readers who only want to use the FastSpecFit DR2 VAC will find this section largely self-contained. In § 7 we step back to illustrate the catalog’s scientific reach and describe planned improvements, before concluding with a summary in § 8. Finally, in Appendix § 10 we document a companion, photometry-only VAC, and in Appendix § 11 we place the DR2 catalog in the context of FastSpecFit’s full release history, including the earlier EDR and DR1 catalogs.

We adopt a flat Λ\LambdaCDM cosmology with Ωm=0.315\Omega_m = 0.315, consistent with Planck Collaboration et al. [100], and a G. Chabrier [23] initial mass function (IMF) spanning 0.080.08–120 M⊙120\,M_\odot; we define h≡H0/(100 km s−1 Mpc−1)h \equiv H_0/(100\ \mathrm{km\ s^{-1}\ Mpc^{-1}}) and express stellar masses and other derived quantities in units with h=1h = 1.

Spectroscopic and Photometric Data

In this section we describe the DESI spectroscopic (§ 2.1) and broadband photometric (§ 2.2) data that FastSpecFit fits, and the target classes and redshift measurements that define the extragalactic objects that DESI observes (§ 2.3).

DESI Spectroscopy

DESI is a highly multiplexed fiber spectrograph mounted on the Mayall 4-m telescope at Kitt Peak National Observatory ([35, 36]).

A wide-field optical corrector (T. N. Miller et al. [90]) provides a 3.23\mathord{.}2-diameter field of view. Within this field, 5000 robotically actuated positioners (J. H. Silber et al. [124]) place 1.51\mathord{.}5-diameter fibers (C. Poppett et al. [101]) onto target coordinates. These fibers feed ten three-arm spectrographs covering 36003600–98249824 Å in the observed frame ([36]). The focal plane is divided into ten wedge-shaped “petals”, each of which routes its 500 fibers to one of the ten spectrographs. A single pointing of the full 5000-fiber focal plane defines a “tile”, the basic unit of the survey’s observing (tiling) strategy (E. F. Schlafly et al. [117]). The three cameras sense blue (B: 36003600–59305930 Å), red (R: 56005600–77207720 Å), and near-IR (Z: 74707470–98249824 Å) light at a spectral resolution of R≡λ/ΔλFWHM≈2000\mathcal{R} \equiv\lambda/\Delta\lambda_{\mathrm{FWHM}} \approx2000–35003500, 34003400–48004800, and 38003800–52005200, respectively ([36]).

J. Guy et al. [53] describe the DESI spectroscopic pipeline in detail (see also DESI Collaboration et al. [38, 39] for shorter summaries). FastSpecFit relies on three per-fiber products from this pipeline: the one-dimensional spectrum, its corresponding inverse-variance spectrum, and the resolution matrix R\mathbf{R}, a band-diagonal, sparse matrix that quantifies the effective, wavelength-dependent instrumental resolution. R\mathbf{R} is a natural output of the “spectroperfectionism” algorithm of A. S. Bolton & D. J. Schlegel [15], which the DESI pipeline uses to extract one-dimensional spectra with uncorrelated pixels. It is also a key ingredient of FastSpecFit’s forward-modeling philosophy: convolving a perfect, infinite-resolution model spectrum with R\mathbf{R} yields the equivalent model at DESI’s effective spectral resolution (§ 3).

The per-fiber spectra are extracted onto an 0.80.8 Å linearly spaced wavelength grid (in vacuum and in the Solar System barycentric frame of reference), flux-calibrated using a network of F-type standard stars (typically 10 standard stars per petal), and repeat exposures are coadded using inverse-variance weights. J. Guy et al. [53] and S. Koposov et al. [73] report the residual wavelength-calibration uncertainty to be ≈0.025\approx0.025 Å per sky line, or ≈1 km s−1\approx1\ \mathrm{km\ s^{-1}} at 7000 Å. Meanwhile, comparison of the calibrated spectra against hydrogen-atmosphere white-dwarf models shows a residual flux-calibration scatter of about 2% over most of the 36003600–98009800 Å range (rising to ≈6%\approx6\% shortward of 3700 Å), comparable to the ±6%\pm6\% scatter between synthesized DESI spectrophotometry and imaging photometry (§ 2.2) for bright stars ([53]).

Broadband Photometry

FastSpecFit uses broadband optical and IR photometry from the DESI Legacy Imaging Surveys Data Release 9 (DR9) ((hereafter, Legacy Surveys; A. Dey et al. 2019, D. J. Schlegel et al., in preparation)), which covers the full DESI footprint using two imaging surveys and three cameras. South of a declination boundary of approximately +32∘+32^\circ, the grzgrz imaging comes from the DECam Legacy Survey (DECaLS; [41]), obtained with the Dark Energy Camera (DECam; [46]) on the Blanco 4-m telescope at the Cerro Tololo Inter-American Observatory (CTIO). North of this boundary, gg- and rr-band imaging come from the Beijing-Arizona Sky Survey (BASS; [145]), which was conducted with the 90Prime camera ([139]) on the Bok 2.3-m telescope, and the zz-band imaging comes from the Mayall zz-band Legacy Survey (MzLS), which used the Mosaic-3 camera ([40]) on the Mayall 4-m telescope. The Bok and Mayall telescopes are both located at Kitt Peak National Observatory (KPNO). Following A. D. Myers et al. [95], we track the imaging system for each DESI target via the PHOTSYS flag (N or S) to assign the appropriate filter curves from speclite.2

The Legacy Survey catalogs are generated using the Tractor forward-modeling photometry pipeline ([78 b]), which fits a common two-dimensional image model simultaneously across all available grzgrz imaging to derive the integrated flux for each source. A forced-photometry technique is used to derive fluxes in the four WISE bands—W1W1 (3.4 µm), W2W2 (4.6 µm), W3W3 (12 µm), and W4W4 (22 µm) ([140])—based on the unWISE coadds3 produced by A. M. Meisner et al. [89] for W1W1 and W2W2 (based on the six-year WISE plus NEOWISE-Reactivation imaging; [87]) and D. Lang [77] for W3W3 and W4W4. FastSpecFit fits this total photometry in all seven grzgrz and WISE W1W1–W4W4 bands together with the DESI spectrum in a joint spectrophotometric fit (§ 3); in any band where a source is undetected, the reported 1σ1\sigma flux uncertainty is used as an upper limit.

DESI Target Classes and Redshifts

This section briefly summarizes the DESI survey and program structure, the extragalactic target classes, and the redshift-determination pipeline needed to interpret the FastSpecFit VACs; for a more detailed discussion of all these topics see DESI Collaboration et al. (2024a,b, 2026).

DESI observations are organized by survey and program ([117, 95]). The primary program is the main survey, which began on 2021 May 14 as a five-year survey (DESI Run 1; 2021–2026) and has since been extended by an additional three years (DESI Run 1b; 2026–2029). The main survey was preceded by several SV phases, most notably the One-Percent Survey (SV3), which covered roughly 1% of the final DESI footprint to high spectroscopic completeness [37 b]. Within each survey, observations are assigned to the “bright” program (moonlight or reduced transparency), the “dark” program (nominal dark-sky conditions), or the “backup” program (twilight and poor-weather time; A. Dey et al. [42]) on the basis of sky conditions. A small fraction of observations that do not fit any of these three categories are assigned to an “other” program [39].

DESI photometrically selects primary extragalactic targets from the Legacy Surveys ([41]; D. J. Schlegel et al., in preparation) and organizes them into four classes spanning complementary redshift ranges [95]. The bright galaxy survey (BGS, 0<z≲0.60 < z \lesssim0.6; [54]) targets galaxies to r≲20.2r \lesssim20.2 mag in the bright program; the luminous red galaxy (LRG, 0.4<z≲1.10.4 < z \lesssim1.1; [142]), emission-line galaxy (ELG, 0.6<z≲1.60.6 < z \lesssim1.6; [103]), and quasar (QSO, 0.9<z≲40.9 < z \lesssim4; [25]) samples are all observed in the dark program; and finally, secondary or tertiary (special) targets are observed as part of a wide range of bespoke science programs and can be assigned to any program [95, 39].

Spectroscopic redshifts and spectral classifications are determined by Redrock (S. Bailey et al., in preparation; see also [53, 6]), a principal component analysis (PCA) template-fitting code that independently fits templates for stars, galaxies, and quasars over a type-dependent grid of trial redshifts and records the maximum-likelihood redshift Z and template class SPECTYPE (STAR, GALAXY, or QSO). The two primary redshift-quality diagnostics reported by Redrock are ZWARN, a per-object warning bitmask where ZWARN=0 signals no known hardware, data-quality, or redshift-fitting problems, and DELTACHI2, the difference in χ2\chi^2 between the best and second-best fits across all spectral classes. A third diagnostic, COADD_FIBERSTATUS, is a per-object bitmask that logically combines the per-exposure fiber-status flags contributing to the coadded spectrum and records hardware or data-quality problems such as a broken or stuck fiber, poor positioning, or a bad amplifier readout (D. Green et al., in preparation; S. Bailey et al., in preparation).

Finally, three after-burner processing steps follow Redrock: an emission-line fitter (emlinefit; [103, 38]) used to validate ELG redshifts; a Mg II λλ2796,2803\lambda\lambda2796, 2803 doublet fitter ([25]); and the QuasarNET neural-network QSO classifier ([18, 50]). The latter two after-burners are essential for improving the completeness of the QSO sample (§ 6.1).

Spectrophotometric Modeling

This section describes the technical, algorithmic structure of fastspec, the default spectrophotometric fitting mode within the larger FastSpecFit software package. Readers interested in examples of fastspec applied to three illustrative DESI targets may safely skip to § 5, while those interested only in the resulting DR2 VAC may skip to § 6.

Let F(λ;θ)F(\lambda;\theta) be the observed-frame UV through IR SED of a galaxy, where θ\theta denotes the physical properties—stellar mass, star-formation history, dust content, kinematics, redshift, and so on—that determine it. Our overarching goal is to find the model

M(λ;θ)=C(λ;θC)+E(λ;θE),(1)M(\lambda;\theta) = C(\lambda;\theta_C) + E(\lambda;\theta_E), \tag*{(1)}

the sum of a stellar continuum model CC (§ 3.4) and an emission-line model EE (§ 4), that best reproduces FF in a maximum-likelihood sense, where θ≡θC∪θE\theta\equiv\theta_C \cup\theta_E is the union of the continuum and emission-line parameter sets.

A key algorithmic choice we make in fastspec is to fit CC and EE sequentially rather than jointly, subtracting the best-fitting continuum model from the data before fitting the emission lines. Because CC draws on the overall continuum shape and broadband photometry while EE draws on (relatively) narrow “residual” features, decoupling them lets us fit the lines robustly and flexibly even when the continuum model is a poor fit to the data due to, for example, residual calibration errors (§ 3.4.4).

We prepare the spectroscopic and photometric data for fitting in two ways. First, we correct both datasets for Galactic extinction using the D. J. Schlegel et al. [118] dust map, recalibrated following E. F. Schlafly & D. P. Finkbeiner [116], and the E. L. Fitzpatrick [45] extinction curve with RV≡AV/E(B−V)=3.1R_V \equiv A_V/E(B-V) = 3.1. We evaluate the extinction curve at the vacuum wavelength of each spectral pixel and at the effective wavelength of each of the seven grzgrz and WISE W1–W4 bandpasses. Second, to guard against underestimated pipeline uncertainties, we add a minimum uncertainty floor in quadrature to the formal inverse variances. We adopt a 1% floor for the spectroscopy and, for the photometry, a fixed per-band floor of 2% for grzgrz (DECam, BASS, and MzLS) and 5% for WISE W1–W4. We use these floored inverse variances throughout the pipeline, including in the Monte Carlo resampling described in § 4.3.

In the remainder of this section we describe the fitting pipeline’s two primary model inputs, the SPS template library (§ 3.1) and the emission-line list (§ 3.2). We then describe the three sequential algorithmic stages fastspec executes in turn—adaptive emission-line masking via fast, preliminary fits in wavelength patches (§ 3.3), stellar continuum modeling (§ 3.4), and emission-line modeling (§ 4). We close by describing the physical quantities fastspec derives from these fits (§ 4.2), how we estimate their uncertainties (§ 4.3), and the additional fitting modes, including a “photometry-only” mode called fastphot, built on this same machinery (§ 4.4).

SPS Template Library

We generate SPS templates using Flexible Stellar Population Synthesis4 (FSPS; C. Conroy et al. [30]; C. Conroy & J. E. Gunn [29]) via a fork of the python-fsps5 interface. We construct these models using a custom version of the C3K theoretical spectral library, C3K_R3K6, the MIST stellar isochrones (v1.0; J. Choi et al. [26]; A. Dotter [43]), and a G. Chabrier [23] IMF from 0.08–120 M⊙M_{\odot}. The C3K_R3K library spans 100 Å to 1000 μ\mum. In the rest-frame optical (3500–9800 Å) it has RC3K=3000\mathcal{R}_{\mathrm{C3K}} = 3000, corresponding to a Gaussian line-spread function with σC3K=c/(2.355RC3K)=42\sigma_{\mathrm{C3K}} = c/(2.355\mathcal{R}_{\mathrm{C3K}}) = 42 km s−1^{-1}, sampled in constant-velocity Δv=50\Delta v = 50 km s−1^{-1} pixels. At shorter and longer wavelengths, our models have lower resolution and sparser wavelength sampling, which keeps the UV-to-IR template lightweight for synthetic photometry and dust-emission modeling (§ 3.4.3) without wasting resolution where the spectroscopic fit does not need it.

Following J. Leja et al. [81], we build five “continuity age” basis templates spanning fixed lookback-time bins, each assuming a constant star-formation rate over its duration. The two youngest bins span 0–30 Myr and 30–100 Myr (with upper edges at log⁡10(t/yr)=7.48\log_{10}(t/\mathrm{yr}) = 7.48 and 8.08.0, respectively), while the next two bins are logarithmically spaced between 100 Myr and 11.6 Gyr (0.1–1.1 Gyr and 1.1–11.6 Gyr) and the oldest bin spans 11.6–13.7 Gyr, the last 15% of a 13.7 Gyr cosmic age (Figure 1). We generate templates at fixed solar metallicity, log⁡10(Z/Z⊙)=0\log_{10}(Z/Z_{\odot}) = 0 (with no α\alpha-element enhancement), with and without nebular emission lines (N. Byler et al. [19]). We need both variants because the broadband photometry includes any nebular line emission, whereas we model the emission lines in the spectrum separately (§ 4). We therefore use the variant with emission lines in the joint spectrophotometric fit (§ 3.4.3), where masking line-contaminated pixels restricts its effect to the synthetic photometry, and for K-corrections (§ 4.2). We use the variant without emission lines everywhere else, including the velocity-dispersion fit (§ 3.4.1), the continuum model we subtract before fitting the emission lines, and continuum-derived quantities such as Dn(4000)D_n(4000) and the monochromatic luminosities (§ 4.2).

Figure 1. The five SPS templates spanning FastSpecFit’s lookback-time age bins (§ 3.1), computed at solar metallicity (log⁡10Z/Z⊙=0\log_{10} Z/Z_{\odot} = 0) with τV=0.1\tau_V = 0.1 and normalized to a 1010 M⊙10^{10}\,M_{\odot} galaxy at z=0.1z = 0.1. Each curve corresponds to one age bin, assuming a constant star-formation rate over its duration, ranging from 0–30 Myr (youngest) to 11.6–13.7 Gyr (oldest); the SED shifts from UV-bright continua with prominent nebular emission in the youngest populations to red, featureless spectra in the oldest. The gray band marks the DESI optical spectral range (0.36–0.98 μ\mum); the curves along the bottom of the panel show the DECam grzgrz and WISE W1–W4 broadband filter throughputs that FastSpecFit includes in the simultaneous spectrophotometric fit.

We index the resulting rest-frame templates Sn(λ)S_n(\lambda), n=1,…,ntemplaten = 1,\ldots,n_{\mathrm{template}} (with ntemplate=5n_{\mathrm{template}} = 5) by star-formation history (SFH) age bin, separately for each nebular-emission variant, and normalize each to the flux density produced at 10 pc by one solar mass of stars formed at the tabulated lookback time,

Sn(λ)=L⊙4π(10 pc)2Lλ,nL⊙,(2)S_n(\lambda) = \frac{L_{\odot}}{4\pi(10\ \mathrm{pc})^2}\frac{L_{\lambda,n}}{L_{\odot}}, \tag*{(2)}

in units of erg s−1 cm−2 A˚−1 M⊙−1\mathrm{erg\ s^{-1}\ cm^{-2}\ \mathring{A}^{-1}\ \mathcal{M}_{\odot}^{-1}}, so that each template’s fitted coefficient cn≥0c_n \ge0 (§ 3.4.3) carries units of M⊙\mathcal{M}_{\odot}.

In addition, we generate a single infrared dust emission template Sdust(λ)S_{\mathrm{dust}}(\lambda) from B. T. Draine et al. [44]. This model is parameterized by the polycyclic aromatic hydrocarbon (PAH) mass fraction, qPAHq_{\mathrm{PAH}}, the minimum starlight intensity heating the diffuse interstellar medium (ISM) dust component Umin⁡U_{\min}, and the fraction γ\gamma of dust mass exposed to a power-law distribution of starlight intensities (up to a fixed maximum Umax⁡U_{\max}) representing photodissociation regions. Because most DESI targets do not have adequately deep IR photometry to constrain a large number of dust emission parameters, we adopt fiducial values of qPAH=3.5q_{\mathrm{PAH} } = 3.5, Umin⁡=1.0U_{\min} = 1.0, and γ=0.01\gamma= 0.01 and normalize our IR template to unit bolometric luminosity.

There is no AGN-like emission (either continuum or emission lines) in these templates, a shortcoming which should be kept in mind throughout this paper (§ 5.3). However, § 7.2 discusses some of the planned improvements for FastSpecFit, including a QSO/AGN fitting mode.

Emission-Line List

The default FastSpecFit line-list, data/emlines.ecsv, includes 46 emission lines, spanning rest-frame Lyα\alpha λ1215\lambda1215 through [S III] λ9532\lambda9532 (Table 1). We select this list to span the full range of galaxy and QSO spectral types over the DESI redshift range, targeting the strongest and most diagnostically useful transitions rather than every possible detectable line; the resulting set supports target selection and scientific follow-up across a broad range of object classes, from star-forming galaxies to high-redshift quasars.

Emission LineLine Prefixa^{a}Rest Wavelengthb^{b}Initial Patchc^{c}“Strong”d^{d}
UV/QSO Lines
Lyα\alphaLYALPHA1215.670a✓
N V λ1240\lambda1240NV_12401240.14a✓
O I λ1304\lambda1304OI_13041304.35b
Si IV λ1396\lambda1396SILIV_13961396.76c
C IV λ1549\lambda1549CIV_15491549.06d✓
He II λ1640\lambda1640HEII_16401640.42e
Al III λ1857\lambda1857ALIII_18571857.40f
Si III] λ1892\lambda1892SILIII_18921892.030f✓
C III] λ1908\lambda1908CIII_19081908.734f✓
Mg II λ2796\lambda2796MGII_27962796.3511g✓
Emission LineLine PrefixaRest WavelengthbInitial Patchc“Strong”d
Mg II λ\lambda2803MGII_28032803.5324g✓
He II λ\lambda4686HEII_46864687.02o
Narrow Balmer/Helium Lines
H6H63890.1576j
HϵH\epsilonHEPSILON3971.195k
HδH\deltaHDELTA4102.892l
HγH\gammaHGAMMA4341.684m✓
He I λ\lambda4471HEI_44714472.7290n
HβH\betaHBETA4862.683p✓
He I λ\lambda5876HEI_58765877.2690r
HαH\alphaHALPHA6564.613t✓
Broad Balmer Lines
H6b_bH6_BROAD3890.1576j
HϵbH\epsilon_bHEPSILON_BROAD3971.195k
HδbH\delta_bHDELTA_BROAD4102.892l
HγbH\gamma_bHGAMMA_BROAD4341.684m✓
HβbH\beta_bHBETA_BROAD4862.683p✓
HαbH\alpha_bHALPHA_BROAD6564.613t✓
Narrow Forbidden Lines
[Ne V] λ\lambda3346NEV_33463346.7828h
[Ne V] λ\lambda3426NEV_34263426.8637h
[O II] λ\lambda3726OII_37263727.0919i✓
[O II] λ\lambda3729OII_37293729.8750i✓
[Ne III] λ\lambda3869NEIII_38693869.8611j
[O III] λ\lambda4363OIII_43634364.4351m
[O III] λ\lambda4959OIII_49594960.2937p✓
[O III] λ\lambda5007OIII_50075008.2383p✓
[N II] λ\lambda5755NII_57555756.1887q
[O I] λ\lambda6300OI_63006302.0435s
[S III] λ\lambda6312SIII_63126313.8062s
[N II] λ\lambda6548NII_65486549.8578t✓
[N II] λ\lambda6584NII_65846585.2696t✓
Emission LineLine PrefixaRest WavelengthbInitial Patchc“Strong”d
[S II] λ6716\lambda6716SII_67166718.2913u✓
[S II] λ6731\lambda6731SII_67316732.6705u✓
[Ar III] λ7135\lambda7135ARIII_71357137.77v
[O II] λ7320\lambda7320OII_73207321.731v
[O II] λ7330\lambda7330OII_73307332.199v
[S III] λ9069\lambda9069SIII_90699071.103w
[S III] λ9532\lambda9532SIII_95329533.227w

Table 1. Emission Lines

The adopted vacuum rest-frame wavelengths in Table 1 come from the NIST Atomic Spectra Database7 for the Balmer series, the helium lines, the Mg II λλ2796\lambda\lambda2796, 2803 doublet, and most forbidden transitions; the UV broad-line wavelengths come from D. E. Vanden Berk et al. [132]; and wavelengths for [N II] λλ6548\lambda\lambda6548, 84, [O III] λλ4959\lambda\lambda4959, 5007, [N II] λ5755\lambda5755, and [O III] λ4363\lambda4363 come from the CHIANTI atomic database.8

We organize these lines into four general categories—UV/QSO lines, narrow Balmer/helium lines, broad Balmer/helium lines, and narrow forbidden lines—which are used to set kinematic fitting priors (§ 4.1.1). Table 1 also lists a set of 23 spectrally adjacent “patches”, and a Boolean bit (“strong”) which flags which lines, if present, are most likely to be well-detected in a given spectrum. We describe these latter quantities in detail in the next subsection (§ 3.3).

Adaptive Emission-Line Masking

In order to fit the stellar continuum robustly, we must first identify pixels which could be affected by emission lines. The challenge is that DESI targets span an enormous range of emission-line properties, from purely absorption-line quiescent galaxies with no measurable line-emission to extreme starbursts and broad-line AGN whose emission dominates large stretches of the observed spectrum. A single, fixed masking window sized for the strongest expected lines would over-mask quiescent spectra and discard usable continuum information, while a window sized for narrow lines found in typical star-forming galaxies would under-mask galaxies and QSOs with broad emission lines and bias the stellar continuum fit.

Therefore, before fitting the stellar continuum (§ 3.4.3), we run a lightweight, per-object fitting pass—the line masker—that determines whether a spectrum contains statistically significant line-emission, and, if so, the approximate velocity widths and velocity shifts of those lines. Crucially, the line-masker also performs an initial determination of whether the spectrum contains broad Balmer line-emission (although the statistically defensible decision is made in the final fitting pass, described in § 4.1.1).

First, we activate all the spectral patches (Table 1) with at least one line in the observed spectral range and fit all the “strong” lines with the narrow-only kinematic model, together with a simple linear local continuum in each patch. The lines we initially fit are simply the subset of lines which, if present, are most likely to yield useful kinematic constraints. In the narrow-only kinematic model, all the forbidden lines are tied together into a single kinematic group (shared velocity width and velocity center) anchored on [O III] λ5007\lambda5007, while all the narrow Balmer and He I lines are anchored on Hα\alpha in a second kinematic group (see § 4.1.1 for more details). Then, we define a pixel window of ±5σline\pm5\sigma_{\mathrm{line}} centered on each line’s expected observed-frame wavelength, where σline\sigma_{\mathrm{line}} is that line’s current velocity-width estimate. We iteratively update σline\sigma_{\mathrm{line}}, beginning with 150 km s−1^{-1} and 3000 km s−1^{-1} for narrow and (candidate) broad lines, respectively, but always floored at a minimum line-width of 50 km s−1^{-1}, and grow a surrounding continuum window until at least 11 unmasked pixels are available. When this adaptive growth causes two spectrally adjacent patches’ continuum windows to overlap by more than 11 pixels, we merge them into a single patch, sharing one local continuum and the union of both patches’ lines. We run two such iterations in total,

updating the line-pixel windows with the best-fit velocity shifts and widths before the second pass.

Next, if any of the three strongest Balmer lines (Hα\alpha, Hβ\beta, or Hγ\gamma) fall within the observed spectral range, we repeat the procedure with the narrow+broad model, initializing all the broad Balmer line-widths at 1000 km s−1^{-1}. In this fitting mode, all the “strong” broad Balmer lines are kinematically tied together, separately from the two previously described narrow-line kinematic groups (see also § 4.1.1). After this fitting pass, if any broad Balmer line reaches S/N > 1.5, we adopt the narrow+broad line-widths for masking; otherwise, we retain the narrow-only widths. Here, S/N is the fitted peak amplitude divided by the robust standard deviation of the line-subtracted residuals in that patch.

With the robustly estimated line-widths, we can now build the complete line-mask. Iterating over all the lines in the observed spectral range (not just the initial set of candidate strong lines), we robustly estimate each line’s S/N by comparing its peak amplitude against the robust standard deviation of the line-subtracted residuals in its surrounding local continuum, and discard “weak” lines with S/N < 3.5 from further consideration, in order to avoid over-masking the continuum spectrum. For the surviving lines, we adopt a ±5σline\pm5\sigma_{\mathrm{line}} window, centered on each line’s expected wavelength and sized using its converged velocity width and velocity shift, to build the final pixel mask. Regardless of the outcome of the broad-line test, we floor the broad Balmer line-width at 1000 km s−1^{-1} when building this mask, so that the final emission-line fit (§ 4.1.1) can remeasure any broad component.

Figure 2 illustrates this process for a single BGS target from the bright program of the main survey, TARGETID=39627670102744387, at z=0.3292z = 0.3292. Two pairs of spectrally adjacent patches are merged in this example: the “m” (Hγ\gamma) and “p” (Hβ\beta and [O III] λλ4959,5007\lambda\lambda4959, 5007) patches merge into a single “mp” patch, and the “t” (Hα\alpha and [N II] λλ6548,84\lambda\lambda6548, 84) and “u” ([S II] λλ6716,31\lambda\lambda6716, 31) patches merge into “tu”, because their adaptive continuum windows overlap by more than 11 pixels. The line-masker also detects broad Balmer line-emission in this target, so the corresponding patches are masked with wider windows, using the narrow+broad kinematic model described above.

Figure 2. Illustration of the line-masking algorithm for an example BGS target at z=0.3292z = 0.3292 (TARGETID=39627670102744387 from the main survey bright-time program, appearing on HEALPix 17366). Top: the full per-camera DESI spectrum (B, R, Z in blue, green, and red, respectively) together with the per-patch best-fitting local linear continuum and Gaussian model line-profiles (darker blue, green, and red in B, R, and Z, respectively) for all the candidate strong lines in the observed spectral range. Bottom: Zoomed-in panels which more clearly show the per-patch fitting results. In each panel, camera-colored pixels mark the pixels in the final line-mask (and therefore excluded from the subsequent continuum-fit), while light-gray pixels are the surrounding continuum used for that patch’s local linear fit. The darker curves show the combined best-fit line-plus-continuum model, while the dashed black line and surrounding dotted band show the local linear continuum and its ±1σ\pm1\sigma noise envelope. Finally, the legends indicate the S/N value inferred for each line and the patch label. For this particular object, a significant broad Hα\alpha Balmer line was correctly identified, resulting in patches “m” and “p” and patches “t” and “u” merging into larger patches “mp” and “tu”, respectively, and wider line masks.

Stellar Continuum Model

This section describes the steps fastspec uses to find the best-fitting continuum model, C(λ;θC)C(\lambda; \theta_C). Because speed and flexibility are paramount, the default fitting mode uses a minimal parameter set, θC={c,τV,σ⋆}\theta_C = \{\mathbf{c}, \tau_V, \sigma_\star\}, where c\mathbf{c} are the non-negative template coefficients, τV\tau_V is the VV-band dust optical depth, and σ⋆\sigma_\star is the stellar velocity dispersion. We emphasize, however, that the fitting infrastructure can be easily extended to include additional parameters (e.g., redshift, a two-component dust model or QSO model spectra), which we discuss in § 7.2. For each object, we determine θC\theta_C in three stages. First, we measure σ⋆\sigma_\star from stellar absorption features (§ 3.4.1); then we derive a scalar aperture correction that places the DESI fiber spectrum on the total-flux scale of the broadband photometry (§ 3.4.2); and finally we jointly fit for dust attenuation τV\tau_V and the template coefficients c\mathbf{c}, holding zz fixed at its input (Redrock) value (§ 3.4.3). An optional nonparametric smooth-continuum correction then removes residual low-level systematics (§ 3.4.4) before the spectrum passes to the emission-line fitter.

Stellar Velocity Dispersion

The line-of-sight stellar velocity dispersion σ⋆\sigma_\star traces the depth of a galaxy’s gravitational potential well; it is a key observable for several galaxy scaling relations, such as the Fundamental Plane, the black hole mass–σ⋆\sigma_\star relation, and the halo-mass–stellar-mass relation, and a fundamental ingredient for most theoretical models of galaxy formation ([104, 12, 75, 133 and references therein]). A reliable measurement, however, requires adequate rest-frame wavelength coverage and relatively high-S/N spectroscopy of the absorption-line features most sensitive to dynamical broadening (e.g., Ca H&K, the GG-band, Hβ\beta, Mg bb). In DESI, these conditions are generally met only for BGS and LRG targets; however, because fastspec does not use targeting information as a prior, we attempt to measure σ⋆\sigma_\star for every spectroscopic target.

We determine whether a direct measurement of σ⋆\sigma_\star is possible from two conditions: rest-frame wavelength coverage and “fit significance.” First, we require the spectrum of a given target to span the full rest-frame range 3800–4900 Å, bridging Ca H&K, the 4000-Å break, and Hβ\beta. If this condition is satisfied, then we perform the actual fit using all unmasked pixels between 3800–6000 Å (in the rest-frame). In practice, this requirement means that we can only estimate σ⋆\sigma_\star for galaxies up to z≈1z \approx1. Second, we determine fit significance by performing a fast χ2\chi^2 grid scan of six trial velocity-broadening widths evenly spaced between 50 and 500 km s−1^{-1}, simultaneously solving for τV\tau_V and the template coefficients by linear least-squares at each grid point (§ 3.4.3). If this scan yields a peak-to-peak Δχ2≥25\Delta\chi^2 \ge25 (and with the minimum value away from either boundary) then we proceed with a direct measurement. In general, Δχ2<25\Delta\chi^2 < 25 implies that the spectrum has inadequate S/N in the stellar continuum (e.g., a faint ELG target) to accommodate σ⋆\sigma_\star as an additional free parameter.

Most DESI targets in DR2 fail one or both of the preceding conditions; however, we would still like to use an approximately correct value of σ⋆\sigma_\star. In this case, we estimate σ⋆\sigma_\star from an approximate stellar-mass–velocity-dispersion relation,

log⁡10(σ⋆/km s−1)=a+b(log⁡10(M⋆h−2/M⊙)−11),(3)\log_{10}(\sigma_\star/\mathrm{km\,s}^{-1}) = a + b\left(\log_{10}(\mathcal{M}_\star h^{-2}/\mathcal{M}_\odot) - 11\right), \tag*{(3)}

with (a,b)=(2.30,0.25)(a,b)=(2.30, 0.25) using a fast, preliminary estimate of the stellar mass, M⋆\mathcal{M}_\star (§ 3.4.3). If our estimate of M⋆\mathcal{M}_\star fails, we instead adopt a fallback σ⋆=150\sigma_\star= 150 km s−1^{-1} (VDISP_NOMINAL), close to the characteristic “knee” of the velocity dispersion function ([10, 129]). In either fallback case we set VDISP_IVAR = 0 to indicate that σ⋆\sigma_\star has not been independently measured.

When a direct measurement is possible, we determine σ⋆\sigma_\star by finding the best-fitting Gaussian kernel dispersion, σ⋆kern∈[50,500]\sigma_\star^{\mathrm{kern}} \in[50, 500] km s−1^{-1}, using scipy.optimize.least_squares with a Trust-Region Reflective minimization algorithm ([17]), fitting for σ⋆kern\sigma_\star^{\mathrm{kern}} and τV\tau_V simultaneously over the same range as above. At each trial value of σ⋆kern\sigma_\star^{\mathrm{kern}} we convolve the templates with a Gaussian broadening kernel using a Fast Fourier Transform (FFT), which multiplies the templates’ precomputed Fourier transforms by the kernel’s and is markedly faster than a direct pixel-space convolution. We then recover the physical velocity dispersion by adding back the intrinsic resolution of our SPS models in quadrature,

σ⋆=(σ⋆kern)2+σC3K2,(4)\sigma_\star= \sqrt{\left(\sigma_\star^{\mathrm{kern}}\right)^2 + \sigma_{\mathrm{C3K}}^2}, \tag*{(4)}

where σC3K=42\sigma_{\mathrm{C3K}} = 42 km s−1^{-1} (§ 3.1).

Aperture Correction

By default, fastspec fits the spectrum (0.36 – 0.98 μ\mum; § 2.1) of each extragalactic DESI target simultaneously with its corresponding seven-band (grzgrz and WISE W1–W4) Legacy Surveys integrated photometry (0.47 – 22 μ\mum; § 2.2). Fitting all the available spectrophotometry over the broadest possible wavelength baseline helps constrain physical properties like dust attenuation and ties the derived stellar masses, star-formation rates, and rest-frame photometry to the total light of each source rather than to the light captured by the 1.5″-diameter DESI fiber. However, we have to account for the fact that the spectrum and photometry are measured over different spatial scales before combining them.

So-called aperture effects (or aperture bias) are extremely difficult to treat in detail because galaxies exhibit genuine radial gradients (or, more generally, spatial inhomogeneities) in color, stellar populations, age, metallicity, and more ([110], and references therein). In addition, the amount of light the DESI fiber captures depends on each galaxy’s apparent (angular) size on the sky, and therefore on redshift through the angular diameter distance ([59]). These effects conspire to make aperture bias a strong function of target class, affecting low-redshift BGS targets the most and higher-redshift, angularly compact, and more spectrophotometrically homogeneous LRG and ELG targets the least.

Nevertheless, in the current version of fastspec we use the simplest possible model to account for aperture bias: we assume that the spectral shape within the fiber is representative of the galaxy as a whole, independent of wavelength or the magnitude of the aperture correction. We derive this correction by synthesizing grzgrz photometry from a fast initial fit to the DESI spectrum to get PkfibP_k^{\mathrm{fib}} for k∈{g,r,z}k \in\{g, r, z\}, and computing the per-band multiplicative factor

ηk=DkPkfib,(5)\eta_k = \frac{D_k}{P_k^{\mathrm{fib}}}, \tag*{(5)}

where DkD_k is the total Tractor flux of the object in bandpass kk. Then, we compute the median scalar aperture correction, ηˉ=median⁡(ηg,ηr,ηz)\bar{\eta} = \operatorname{median}(\eta_g, \eta_r, \eta_z), and apply the resulting multiplicative factor (falling back to ηˉ=1.0\bar{\eta} = 1.0 if any part of the preceding analysis fails) during the full spectrophotometric fit (§ 3.4.3).

Fitting the Stellar Continuum

To fit the stellar continuum we use a forward-modeling approach. We carry each rest-frame SPS template (§ 3.1) through a sequence of physical and instrumental transformations—dust attenuation, redshifting, and instrumental broadening (or filter convolution)—to produce a prediction that we can compare directly to the observed spectrophotometry. We then solve for the template weights and dust attenuation (at fixed σ⋆\sigma_\star; § 3.4.1) that minimize the inverse-variance-weighted residuals.

First, we model the attenuation of each template as a uniform foreground screen,

Snatt(λ)=Sn(λ)e−τVk(λ),(6)S_n^{\mathrm{att}}(\lambda) = S_n(\lambda)e^{-\tau_V k(\lambda)}, \tag*{(6)}

where Sn(λ)S_n(\lambda) is the unattenuated template given by equation (2), τV\tau_V is the VV-band optical depth, and

k(λ)=(λ5500 A˚)n(7)k(\lambda) = \left(\frac{\lambda}{5500\,\text{\AA}}\right)^n \tag*{(7)}

is the attenuation curve, which we model as a simple power law with fixed index n=−0.7n=-0.7 based on the S. Charlot & S. M. Fall [24] birth-cloud model. Absorbed photons are re-emitted in the IR following energy balance. We scale the B. T. Draine et al. [44] dust emission template, Sdust(λ)S_{\mathrm{dust}}(\lambda), by the luminosity absorbed by each template,

Labs,n=∫0.01 μm1000 μmSn(λ)(1−e−τVk(λ)) dλ,(8)L_{\mathrm{abs},n} = \int_{0.01\,\mu\mathrm{m}}^{1000\,\mu\mathrm{m}} S_n(\lambda)(1-e^{-\tau_V k(\lambda)})\,\mathrm{d}\lambda, \tag*{(8)}

so that the total (attenuated plus reradiated) continuum template conserves bolometric luminosity,

Sntot(λ)=Snatt(λ)+Labs,nSdust(λ).(9)S_n^{\mathrm{tot}}(\lambda) = S_n^{\mathrm{att}}(\lambda) + L_{\mathrm{abs},n}S_{\mathrm{dust}}(\lambda). \tag*{(9)}

For an object at redshift zz with luminosity distance DLD_L (in pc), we relate each rest-frame template to its observed-frame counterpart via

Snobs(λ)=TIGM(λ,z) F0 M0(10 pcDL)2Sntot(λ/(1+z))1+z,(10)S_n^{\mathrm{obs}}(\lambda) = T_{\mathrm{IGM}}(\lambda,z)\,\mathcal{F}_0\,\mathcal{M}_0\left(\frac{10\,\mathrm{pc}}{D_L}\right)^2\frac{S_n^{\mathrm{tot}}(\lambda/(1+z))}{1+z}, \tag*{(10)}

where TIGM(λ,z)T_{\mathrm{IGM}}(\lambda,z) is the A. K. Inoue et al. [63] intergalactic medium (IGM) transmission function which accounts for Lyα\alpha forest and Lyman-limit attenuation; F0=1017\mathcal{F}_0 = 10^{17} is a dimensionless scale factor converting the cgs flux densities of § 3.1 to our internal flux units of 10−17 erg s−1 cm−2 A˚−110^{-17}\,\mathrm{erg}\,\mathrm{s}^{-1}\,\mathrm{cm}^{-2}\,\text{\AA}^{-1} (J. Guy et al. [53]); and M0=1010 M⊙\mathcal{M}_0 = 10^{10}\,M_\odot is a fixed mass normalization that rescales the fitted template coefficients to order unity for a well-conditioned non-negative least-squares (NNLS) solve (see below).

We restrict the dust optical depth to the range τV∈[0,2]\tau_V \in[0, 2], and by default we include coefficients for all five SPS age bins (§ 3.1), even for stellar populations nominally older than the age of the universe at the redshift of a given object. Because our age sampling is coarse, the boundary between physical and unphysical ages is poorly defined, and excluding bins outright would introduce more redshift-dependent bias than it removes [80].

At this point, we have all the ingredients necessary to fit the stellar continuum. Let dd and II be the Galactic extinction-corrected and floor-inflated DESI spectrum and inverse-variance spectrum, respectively, and mm be the binary emission-line mask from § 3.3. Each of these arrays contain npixn_{\mathrm{pix}} elements, with mi=1m_i = 1 for pixels retained in the fit and 0 for masked pixels. Also let DD and IDI_D be the Galactic extinction-corrected and floor-inflated nband=7n_{\mathrm{band}} = 7 (grzgrz and WISE W1W1–W4W4) photometric fluxes and inverse variances, respectively.

We construct the continuum model C(λ;θC)C(\lambda; \theta_C) as the non-negative linear combination of the observed-frame templates Snobs(λ)S_n^{\mathrm{obs}}(\lambda) from equation (10),

C(λ;θC)=∑ncnSnobs(λ),(11)C(\lambda; \theta_C) = \sum_n c_n S_n^{\mathrm{obs}}(\lambda), \tag*{(11)}

holding σ⋆\sigma_\star fixed at the value determined in § 3.4.1 and solving only for τV\tau_V (which enters through SnobsS_n^{\mathrm{obs}}) and the coefficients c\mathbf{c}. To compare this model against the spectroscopic and photometric data, we transform it in two ways: (1) we broaden each template by σ⋆\sigma_\star (§ 3.4.1), convolve it with the wavelength-dependent resolution matrix R\mathbf{R} (§ 2.1), and resample (using a flux-conserving trapezoidal-resampling algorithm) onto the same npixn_{\mathrm{pix}}-pixel wavelength grid as dd, giving Ci≡C(λi;θC)C_i \equiv C(\lambda_i; \theta_C); and (2) we convolve each template with the nbandn_{\mathrm{band}} filter response functions (§ 2.2) to derive the synthetic photometry PkP_k. Collecting these transformed templates as the columns of a design matrix Ψ\mathbf{\Psi}, the continuum model at every pixel and bandpass becomes

C=Ψc,(12)\mathbf{C} = \mathbf{\Psi}\mathbf{c}, \tag*{(12)}

and we determine τV\tau_V and c\mathbf{c} by minimizing

χ2(τV,c)=∑i=1npixmiIi(di−Ciηˉ)2+∑k=1nbandID,k(Dk−Pk)2,(13)\chi^2(\tau_V,\mathbf{c}) = \sum_{i=1}^{n_{\mathrm{pix}}} m_i I_i \left(d_i - \frac{C_i}{\bar{\eta}}\right)^2 + \sum_{k=1}^{n_{\mathrm{band}}} I_{D,k}(D_k - P_k)^2, \tag*{(13)}

where ηˉ\bar{\eta} is the scalar aperture correction described in § 3.4.2.

Because equation (12) is linear in the multiple template coefficients c\mathbf{c} but nonlinear only in the single scalar τV\tau_V, rather than solve for both simultaneously we exploit the separability via variable projection (VARPRO; [49]). With this approach, we perform an outer Brent scalar minimization over τV\tau_V, with an inner NNLS solve for c\mathbf{c} at each trial τV\tau_V,

c∗=arg min⁡c≥0∥Ψc−b∥2,(14)\mathbf{c}^{*} = \underset{\mathbf{c} \ge0}{\operatorname{arg\,min}} \left\lVert\mathbf{\Psi}\mathbf{c} - \mathbf{b} \right\rVert^2, \tag*{(14)}

where b\mathbf{b} is the corresponding weighted data vector, and the mask, inverse variances, and aperture correction enter as row weights on Ψ\mathbf{\Psi} and b\mathbf{b} alike—miIi/ηˉ\sqrt{m_i I_i}/\bar{\eta} for spectral rows and ID,k\sqrt{I_{D,k}} for photometric rows. Minimizing ∥Ψc−b∥2\lVert\mathbf{\Psi}\mathbf{c}-\mathbf{b}\rVert^2 is therefore equivalent to minimizing equation (13) at fixed τV\tau_V. Because the inner NNLS solve is exact at every trial τV\tau_V, VARPRO converges an order of magnitude faster than direct optimization over the full parameter vector, which would otherwise require finite-differencing across all template coefficients.

With τV\tau_V and c\mathbf{c} now in hand, and σ⋆\sigma_\star fixed from the earlier stages, every element of θC\theta_C is determined, giving the best-fitting stellar continuum model C(λ;θC)C(\lambda;\theta_C). From this model we derive several physical properties of interest—stellar mass, star-formation rate, and stellar age—as described in § 4.2.

Smooth Continuum Correction

Despite its flexibility, our parametric continuum model C(λ)C(\lambda) (§ 3.4.3) can leave small- or even large-scale systematic residuals in the continuum-subtracted spectrum. Depending on their wavelength-dependence and amplitude, leaving these residuals unaddressed can severely bias the emission-line modeling (§ 4.1), especially for faint lines.

The unmodeled residuals can be due to instrumental effects, template mismatch, or both. For example, the noisy boundary between the three DESI cameras (§ 2.1) is especially susceptible to residual additive (bias, dark, or sky-subtraction) errors, or to multiplicative (flat-fielding or flux-calibration) errors. Template mismatch, meanwhile, is most severe for QSOs and broad-line AGN (§ 5.3). Their observed continuum typically blends genuine host-galaxy starlight with Fe II emission in the rest-frame UV (roughly 1100–3000 Å; [134]) and a featureless, non-thermal power-law continuum from accretion onto the supermassive black hole ([132]). Our SPS templates can represent the host-galaxy light but not the other two, nor cleanly separate them.

Keeping these effects in mind, the current version of fastspec treats all such residuals, irrespective of their origin, as a single, low-level additive correction to C(λ)C(\lambda). We estimate and remove these residuals with a nonparametric smooth continuum after subtracting the best-fitting, line-free continuum model. Working per camera, we slide a 75-pixel window in 125-pixel steps across the residual spectrum, applying a 2σ2\sigma iterative clip in each window and computing the robust median and standard deviation. We discard windows with fewer than 15 unmasked pixels. We then fit a weighted quadratic spline (scipy.interpolate.UnivariateSpline, k=2k=2; [135]) to the window medians, giving the smooth continuum correction Csmooth(λ)C_{\mathrm{smooth}}(\lambda). For each camera we also record SMOOTHCORR_B, SMOOTHCORR_R, and SMOOTHCORR_Z, the median of the smooth correction (using all unmasked pixels) divided by the observed flux, expressed as a percentage—a simple, per-object diagnostic of the correction’s overall amplitude relative to the source itself.

Emission-Line Model

After subtracting the best-fitting continuum model C(λ;θC)C(\lambda;\theta_C) (§ 3.4.3) and the smooth continuum correction Csmooth(λ)C_{\mathrm{smooth}}(\lambda) (§ 3.4.4) from the data did_i, we are left with a residual spectrum ri≡di−Ci−Csmooth,ir_i \equiv d_i-C_i-C_{\mathrm{smooth},i} (in erg s−1^{-1} cm−2^{-2} Å−1^{-1}) containing one or more emission lines (§ 3.2). In the following three subsections we describe how we construct and fit the emission-line model, E(λ;θE)E(\lambda;\theta_E) (§ 4.1). The full emission-line parameter set θE\theta_E consists of three parameters—an amplitude, velocity offset, and velocity width—for each of the nlinen_{\mathrm{line}} lines in our list. We then describe how we use physically motivated “kinematic groups” among different sets of lines to reduce the number of free parameters in θE\theta_E, including (optionally) decomposing the Balmer lines into broad and narrow components (§ 4.1.1), and conclude with a summary of all the quantities we measure for each detected (or undetected) emission line (§ 4.1.2).

Forward Model

We forward-model each object’s emission-line spectrum as a non-negative sum of the nlinen_{\mathrm{line}} emission lines from our line list (§ 3.2) that fall within the observed-frame wavelength range at the Redrock redshift zz,

E(λ;θE)=∑p=1nlineϕ(λ;Ap,λp∗,σℓ,p),(15)E(\lambda;\theta_E)=\sum_{p=1}^{n_{\mathrm{line}}}\phi(\lambda;A_p,\lambda_p^*,\sigma_{\ell,p}), \tag*{(15)}

where we model each individual line as a log-Gaussian in observed wavelength,

ϕ(λ;A,λ∗,σℓ)=Aexp⁡(−(ln⁡λ−ln⁡λ∗)22σℓ2).(16)\phi(\lambda;A,\lambda^*,\sigma_\ell)=A\exp\left(-\frac{(\ln\lambda-\ln\lambda^*)^2}{2\sigma_\ell^2}\right). \tag*{(16)}

Here, AA is the peak amplitude (in erg s−1^{-1} cm−2^{-2} Å−1^{-1}), λ∗\lambda^* is the observed line center in Å, including both cosmological (zz) and peculiar (Δv\Delta v) motion of the line-emitting gas,

λ∗=λrest(1+z+Δvc),(17)\lambda^*=\lambda_{\mathrm{rest}}\left(1+z+\frac{\Delta v}{c}\right), \tag*{(17)}

and σℓ=σ/c\sigma_\ell=\sigma/c is the line width in ln-λ\lambda units, where σ\sigma (in km s−1^{-1}) is the physical (Doppler) line-width.13 With this parameterization, the total flux of each line (in erg s−1^{-1} cm−2^{-2}) is

Φ=2πAλ∗σℓ=2πAλ∗σc,(18)\Phi=\sqrt{2\pi}A\lambda^*\sigma_\ell=\sqrt{2\pi}A\lambda^*\frac{\sigma}{c}, \tag*{(18)}

which is independent of the DESI wavelength grid and instrumental resolution. Comparing this profile to a DESI spectrum requires convolving it with the resolution matrix R\mathbf{R} (§ 2.1). Unlike the stellar continuum, however, which varies smoothly enough across a pixel that sampling the continuum model at pixel centers before applying R\mathbf{R} preserves its flux (§ 3.4.3), it is possible for an emission line to be narrower than a single 0.8 Å DESI pixel.14 Low-mass, low-metallicity star-forming galaxies, in particular, can exhibit strong nebular line-emission with intrinsic line-widths of tens of km s−1^{-1} or less—narrower than a pixel over most of the observed wavelength range ([120 J. Moustakas et al., in preparation; V. Manwadkar et al., in preparation]).

The obvious forward-modeling analog, m≈Rϕ\mathbf{m} \approx\mathbf{R}\boldsymbol{\phi}, therefore fails: point-sampling ϕ(λ)\phi(\lambda) at the pixel centers captures only a small, pixel-phase-dependent fraction of a narrow line’s true area, biasing the recovered flux low by a few percent up to >30%>30\%, depending on where the line center falls within a pixel (see Appendix D in J. Guy et al. [53] for additional discussion). In detail, this bias arises because R\mathbf{R}’s rows are normalized to conserve flux only for smooth inputs sampled at pixel centers, with no matching column normalization to protect sub-pixel structure ([53]). Moreover, this bias is not just a modeling artifact; simulations show that even a simple (model-independent) integration of the observed pixels (Φbox\Phi_{\mathrm{box}}; § 4.1.2) suffers the same flux loss for lines narrower than about one pixel ([73 private communication]).15

To contend with this bias, we adopt the deconvolution strategy devised by S. Koposov for RVSpecFit ([74, 73]). The key idea is that R\mathbf{R} samples a smooth underlying resolution function, which we can represent with Gaussian radial basis function (RBF) interpolation. Rather than sampling ϕ(λ)\phi(\lambda) at pixel centers, we pre-convolve our emission-line model with a fixed Gaussian kernel of width σG=0.5\sigma_G = 0.5 Å—narrow compared to the DESI resolution but wide enough to smooth the sub-pixel structure of the model—and propagate the result through a correspondingly deconvolved resolution matrix

W=RM−1,(19)\mathbf{W} = \mathbf{R}\mathbf{M}^{-1}, \tag*{(19)}

where M\mathbf{M} is the Gaussian RBF design matrix, Mij=G(λi−λj;0,σG)M_{ij} = G(\lambda_i-\lambda_j; 0, \sigma_G). The predicted line model is then m≈WF\mathbf{m} \approx\mathbf{W}\mathbf{F}, where F\mathbf{F} is the σG\sigma_G-pre-convolved profile evaluated at pixel centers. Because each FiF_i is a weighted integral of ϕ(λ)\phi(\lambda) rather than a point sample, it captures the correct line area regardless of the line’s width or its phase (position) relative to the pixel grid.

Formally, Fi=(ϕ∗GσG)(λi)F_i = (\phi* G_{\sigma_G})(\lambda_i) is the convolution of the intrinsic profile given by equation (16) with the unit-area Gaussian kernel, evaluated at pixel center λi\lambda_i. Because both ϕ\phi and the kernel are exact Gaussians in ln-λ\lambda space, their convolution is itself Gaussian, with variances adding in quadrature to give an effective width

σeff=σℓ2+σG,ℓ2,(20)\sigma_{\mathrm{eff}} = \sqrt{\sigma_\ell^2 + \sigma_{G,\ell}^2}, \tag*{(20)}

where σG,ℓ≡σG/λ∗\sigma_{G,\ell} \equiv\sigma_G/\lambda^*. In addition, because the kernel has unit area, the convolution also conserves the total line flux, given by equation (18). However, since the profile has broadened from σℓ\sigma_\ell to σeff\sigma_{\mathrm{eff}}, its peak amplitude must shrink by the same factor to keep that flux fixed. The pre-convolved profile of a single line at pixel center λi\lambda_i therefore becomes

Fi=Aσℓσeffexp⁡(−(ln⁡λi−ln⁡λ∗)22σeff2),(21)F_i = A \frac{\sigma_\ell}{\sigma_{\mathrm{eff}}} \exp\left(-\frac{(\ln\lambda_i-\ln\lambda^*)^2}{2\sigma_{\mathrm{eff}}^2}\right), \tag*{(21)}

directly analogous to equation (16). In addition, because σeff≥σG,ℓ\sigma_{\mathrm{eff}} \geq\sigma_{G,\ell}, even a formally unresolved line (σℓ→0\sigma_\ell\to0) is represented by a profile spanning several pixels, so its flux is correctly captured by the pixel grid rather than lost between samples. Collecting the pre-convolved profiles of all nlinen_{\mathrm{line}} lines into F\mathbf{F}, the pixelized counterpart of equation (15) finally becomes

E=WF.(22)\mathbf{E} = \mathbf{W}\mathbf{F}. \tag*{(22)}

With this pixelized model in hand, we determine the parameters θE\theta_E of every line by minimizing

χ2(θE)=∑i=1npixIi(ri−Ei)2,(23)\chi^2(\theta_E) = \sum_{i=1}^{n_{\mathrm{pix}}} I_i(r_i-E_i)^2, \tag*{(23)}

using the trust-region reflective algorithm (scipy.optimize.least_squares; [17]), where rir_i is the residual (continuum-subtracted) spectrum, IiI_i is the corresponding inverse variance spectrum, EiE_i is the emission-line model given by equation (22), and npixn_{\mathrm{pix}} is the number of pixels in the spectrum.

Because FastSpecFit has to fit tens of blended lines simultaneously16, this minimization involves many free parameters, and finite-difference gradients would require O(3×nline)O(3 \times n_{\mathrm{line}}) additional forward-model evaluations per iteration. We instead speed up the minimization by supplying least_squares with an analytic Jacobian: closed-form first derivatives of equation (21) with respect to each free parameter, reducing this cost to a single forward-model call per iteration.

First, because FiF_i is linear in AA, the amplitude gradient is simply

∂Fi∂A=FiA.(24)\frac{\partial F_i}{\partial A} = \frac{F_i}{A}. \tag*{(24)}

Writing xi≡ln⁡λi−ln⁡λ∗x_i \equiv\ln\lambda_i-\ln\lambda^* for brevity, the velocity-shift gradient follows from the dependence of ln⁡λ∗=ln⁡[λrest(1+z+Δv/c)]\ln\lambda^* = \ln[\lambda_{\mathrm{rest}}(1+z+\Delta v/c)] on Δv\Delta v:

∂Fi∂Δv=Fixiσeff2c(1+z+Δv/c).(25)\frac{\partial F_i}{\partial\Delta v} = \frac{F_i x_i}{\sigma_{\mathrm{eff}}^2 c(1+z+\Delta v/c)}. \tag*{(25)}

And finally, the line-width gradient, written in a form that remains finite as σℓ→0\sigma_\ell\to0, accounts for the dependence of both the amplitude factor σℓ/σeff\sigma_\ell/\sigma_{\mathrm{eff}} and the exponent on σℓ=σ/c\sigma_\ell=\sigma/c:

∂Fi∂σ=Aσeffcexp⁡(−xi22σeff2)[1−σℓ2σeff2(1−xi2σℓ2)].(26)\frac{\partial F_i}{\partial\sigma} = \frac{A}{\sigma_{\mathrm{eff}}c}\exp\left(-\frac{x_i^2}{2\sigma_{\mathrm{eff}}^2}\right)\left[1-\frac{\sigma_\ell^2}{\sigma_{\mathrm{eff}}^2}\left(1-\frac{x_i^2}{\sigma_\ell^2}\right)\right]. \tag*{(26)}

The parameters θE\theta_{\mathrm{E}} recovered from this minimization are the amplitude, velocity shift, and width of every individual line. In the next section § 4.1.1 we describe how we share Δv\Delta v and σ\sigma across physically motivated kinematic groups to reduce this parameter count.

Kinematic Groups and Broad-Line Decomposition

FastSpecFit reduces the dimensionality of its emission-line fitting § 4.1 by grouping lines into kinematic groups, within which all members share a common velocity shift Δv\Delta v and intrinsic velocity width σ\sigma. All kinematic groups, bounds, and amplitude constraints are loaded at startup from the bundled configuration file data/emline-constraints.yaml, as summarized in Table 2. This parameter file also defines two important kinematic models: an emission line model consisting entirely of narrow lines (narrow-only) and a model with both narrow and broad lines (narrow+broad).

ProfileGroupAnchorσ\sigma (km s−1^{-1})∣Δv∣|\Delta v| (km s−1^{-1})Δv\Delta v releaseda^{a}
Global (both models)
GlobalFree UV/AGN⋯\cdots1–10,000<2,500<2,500⋯\cdots
Lines (independent): Lyα\alpha, [N V] λ\lambda1240, O I λ\lambda1304, Si IV λ\lambda1396, C IV λ\lambda1549, He II λ\lambda1640, Al III λ\lambda1857, He II λ\lambda4686
GlobalC III] pairC III] λ\lambda19081–10,000<2,500<2,500N
Member: Si III] λ\lambda1892
GlobalMg II pairMg II λ\lambda28031–10,000<2,500<2,500N
Member: Mg II λ\lambda2796 (amplitude ratio free; see below)
Narrow-only profile
Narrow-onlyForbidden[O III] λ\lambda50071–750<500<500Y
Members: [Ne V] λλ\lambda\lambda3346,3426, [O II] λλ\lambda\lambda3726,3729, [Ne III] λ\lambda3869, [O III] λλ\lambda\lambda4363,4959, [N II] λ\lambda5755, [O I] λ\lambda6300, [S III] λ\lambda6312, [N II] λλ\lambda\lambda6548,6584, [S II] λλ\lambda\lambda6716,6731, [Ar III] λ\lambda7135, [O II] λλ\lambda\lambda7320,7330, [S III] λλ\lambda\lambda9069,9532
Narrow-onlyNarrow Balmer + He IHα\alpha1–750<500<500Y
Members: H6, Hϵ\epsilon, Hδ\delta, Hγ\gamma, Hβ\beta, He I λ\lambda4471, He I λ\lambda5876
Narrow-onlyFixed (A=0A = 0)⋯\cdots⋯\cdots⋯\cdots⋯\cdots
Lines: Hαb\alpha_b, Hβb\beta_b, Hγb\gamma_b, Hδb\delta_b, Hϵb\epsilon_b, H6b_b
Narrow+broad profile
Narrow+broadAll narrowHα\alpha1–750<500<500N
Members: [Ne V] λλ\lambda\lambda3346,3426, [O II] λλ\lambda\lambda3726,3729, [Ne III] λ\lambda3869, [O III] λ\lambda4363, [N II] λ\lambda5755, [O I] λ\lambda6300, [S III] λ\lambda6312, [N II] λλ\lambda\lambda6548,6584, [S II] λλ\lambda\lambda6716,6731, [Ar III] λ\lambda7135, [O II] λλ\lambda\lambda7320,7330, [S III] λλ\lambda\lambda9069,9532, H6, Hϵ\epsilon, Hδ\delta, Hγ\gamma, Hβ\beta, He I λ\lambda4471, He I λ\lambda5876
Narrow+broad[O III] doublet[O III] λ\lambda50071–750<500<500N
Member: [O III] λ\lambda4959
Narrow+broadBroad BalmerHαb\alpha_b1–10,000<2,500<2,500N
Members: Hβb\beta_b, Hγb\gamma_b, Hδb\delta_b, Hϵb\epsilon_b, H6b_b
Amplitude constraints
Fixed ratios (Aline=f⋅ArefA_{\mathrm{line}} = f \cdot A_{\mathrm{ref}})Free doublet ratios (r≡Aline/Arefr \equiv A_{\mathrm{line}}/A_{\mathrm{ref}})
[N II] λ\lambda6548 / λ\lambda6584f=1/3.033f = 1/3.033[O II] λ\lambda3726 / λ\lambda3729r0=1r_0 = 1, r∈[0,2]r \in[0, 2]
[O III] λ\lambda4959 / λ\lambda5007f=1/2.964f = 1/2.964[S II] λ\lambda6731 / λ\lambda6716r0=1r_0 = 1, r∈[0,2]r \in[0, 2]
[O II] λ\lambda7330 / λ\lambda7320f=1/1.225f = 1/1.225Mg II λ\lambda2796 / λ\lambda2803r0=1r_0 = 1, r∈[0,10]r \in[0, 10]

Table 2. Emission-Line Kinematic Groups and Amplitude Constraints

The narrow-only model ties all forbidden lines to one group anchored to [O III] λ5007\lambda5007 and a second group consisting of all H I and He I lines anchored to Hα\alpha. Separately, the narrow+broad model assigns the [O III] λλ4959,5007\lambda\lambda4959,5007 doublet to its own group anchored on [O III] λ5007\lambda5007; the remaining forbidden, Balmer, and helium lines to a second group anchored on narrow Hα\alpha; and the broad Balmer lines to a third group anchored on broad Hα\alpha. The [O III] doublet gets its own kinematic group because of the frequency with which this doublet shows an asymmetric or complex line-profile in AGN, which can bias the fitting of the fainter forbidden lines. Regardless, in both models, the lines in the global groups of Table 2 have a single component, with no narrow+broad decomposition.

Within each kinematic group, select emission-line doublets carry an additional amplitude constraint, set in the same constraints file as either fixed or free. In the default configuration, the amplitude ratios of [N II] λ6548\lambda6548 / λ6584\lambda6584, [O III] λ4959\lambda4959 / λ5007\lambda5007, and [O II] λ7330\lambda7330 / λ7320\lambda7320 are fixed to their theoretical spontaneous-transition-probability value [98], while [O II] λ3726\lambda3726 / λ3729\lambda3729, [S II] λ6716\lambda6716 / λ6731\lambda6731, and Mg II λ2796\lambda2796 / λ2803\lambda2803, carry a free (but bounded) amplitude ratio. In the fitting engine, we parameterize each doublet as a reference-line amplitude and a bounded (or fixed) amplitude ratio, which lets us keep each ratio within its bounds during the fit. Moreover, for [O II] λλ3726,29\lambda\lambda3726, 29 and [S II] λλ6716,31\lambda\lambda6716, 31 the fitted ratio, OII_DOUBLET_RATIO and SII_DOUBLET_RATIO, respectively, carries physical significance as tracers of the local electron density [98]. In addition, this parameterization often lets FastSpecFit recover a faint doublet component in noisy spectral regions, where an independent-amplitude fit might otherwise drive it to an unphysical part of parameter space.

FastSpecFit fits the narrow-only model first and then tests the narrow+broad model only when broad-line fitting is enabled (true by default, but see § 4.4) and when at least one of the first three Balmer lines (Hα\alpha, Hβ\beta, or Hγ\gamma) falls within the observed

**Note—**All kinematic constraints, optimizer bounds, and amplitude ratios are read from data/emline-constraints.yaml at startup. Lines within a group share a common velocity shift Δv\Delta v and width σ\sigma; the anchor carries the free kinematic parameters. Subscript b denotes the broad component of a Balmer line.

spectral range. When both models are fit, FastSpecFit selects the preferred solution based on four conditions that must all be satisfied:

  1. The change in χ2\chi^2 over the pixels spanning the broad Balmer lines must exceed the corresponding reduction in the number of degrees of freedom,

ΔχBalmer2>ΔNdof,Balmer,(27)\Delta\chi^2_{\mathrm{Balmer}} > \Delta N_{\mathrm{dof,Balmer}}, \tag*{(27)}

where Δχ2=χnarrow2−χbroad2\Delta\chi^2 = \chi^2_{\mathrm{narrow}} - \chi^2_{\mathrm{broad}} and ΔNdof=Ndof,narrow−Ndof,broad\Delta N_{\mathrm{dof}} = N_{\mathrm{dof,narrow}} - N_{\mathrm{dof,broad}}, equal to the number of additional free parameters introduced by the broad-Balmer group. In other words, the data must justify the additional free parameters in a maximum-likelihood sense. This criterion is only evaluated using pixels within ±2.5σbroad\pm2.5\sigma_{\mathrm{broad}} of the non-helium broad Balmer lines.

  1. The fitted velocity dispersion of the broad Hα\alpha component must exceed σbroad(Hα)>250 km s−1\sigma_{\mathrm{broad}}(\mathrm{H}\alpha) > 250\ \mathrm{km\ s}^{-1}, in order to exclude spurious broad fits driven by noise or unmodeled continuum residuals.

  1. The broad component must be wider than the narrow component, σbroad(Hα)>σnarrow(Hα)\sigma_{\mathrm{broad}}(\mathrm{H}\alpha) > \sigma_{\mathrm{narrow}}(\mathrm{H}\alpha), ensuring a physical hierarchy between the two kinematic components.

  1. The amplitude of at least one of the first two broad Balmer lines in the spectral range (e.g., Hα\alpha and Hβ\beta or, at higher redshift, Hβ\beta and Hγ\gamma) must satisfy S/Nbroad>2.5_{\mathrm{broad}} > 2.5, where we estimate the noise robustly from the interquartile range of the fit residuals in the Balmer-line pixels after subtracting the full emission-line model.

If any condition fails, we retain the narrow-only result, but always record the pixel-level χ2\chi^2 comparison in the output columns DELTA_LINECHI2 and DELTA_LINENDOF.

Once a model has been selected, FastSpecFit runs an optional final relaxation pass that frees the velocity shifts and/or line-widths of individual lines. However, we only adopt this solution if there is a statistically significant improvement in the fit given the increased number of degrees of freedom, Δχ2>ΔNdof\Delta\chi^2 > \Delta N_{\mathrm{dof}}. We record the relevant comparison parameters in the output file as DELTA_KINECHI2 and DELTA_KINENDOF.

In the default configuration (Table 2), we only relax the final velocity shifts of lines in the narrow-only model, allowing them to change by no more than ±100 km s−1\pm100\ \mathrm{km\ s}^{-1} around their best-fit value, but keeping the line-widths constrained within each kinematic group. This final round of flexibility accommodates minor wavelength-calibration residuals (§ 6.5) and is especially important when fitting star-forming galaxies with extremely narrow and strong line-emission ((J Moustakas et al., in preparation)). Doublet-tied pairs (e.g., [N II] λλ6548,84\lambda\lambda6548, 84) always retain a shared velocity shift during this final optimization round, to prevent unphysical differences between emission lines arising from the same atomic species, and by default, we do not enable a final relaxation pass for the three-group narrow+broad model (to prevent overfitting).

Summary of Measured Emission-Line Properties

For every line in our list (Table 1), whether formally detected or not, we tabulate a standard set of derived quantities in the output catalog; Table 3 summarizes their symbols, output-column names, and units.

SymbolColumna^{a}UnitsInverse variance?
AAMODELAMPerg s−1^{-1} cm−2^{-2} Å−1^{-1}⋯\cdots
AobsA_{\mathrm{obs}}AMPerg s−1^{-1} cm−2^{-2} Å−1^{-1}Y
Δv\Delta vVSHIFTkm s−1^{-1}Y
σ\sigmaSIGMAkm s−1^{-1}Y
Φ\PhiFLUXerg s−1^{-1} cm−2^{-2}Y
Φbox\Phi_{\mathrm{box}}BOXFLUXerg s−1^{-1} cm−2^{-2}Y
FcF_cCONTerg s−1^{-1} cm−2^{-2} Å−1^{-1}Y
EWEWÅY
Φlim\Phi_{\mathrm{lim}}FLUX_LIMITerg s−1^{-1} cm−2^{-2}⋯\cdots
χline2\chi_{\mathrm{line}}^2CHI2⋯\cdots⋯\cdots
npixlinen_{\mathrm{pix}}^{\mathrm{line}}NPIX⋯\cdots⋯\cdots

Table 3. Each column is prefixed by the line name (e.g., OIII_5007_FLUX); Table 1 lists the full set of modeled lines.

Note—Columns marked “Y” have a companion _IVAR column storing the inverse variance, estimated via the same per-object Monte Carlo procedure described in § 4.3. AA, Δv\Delta v, and σ\sigma are defined in § 4.1; AobsA_{\mathrm{obs}}, Φ\Phi, Φbox\Phi_{\mathrm{box}}, FcF_c, EW, Φlim\Phi_{\mathrm{lim}}, χline2\chi_{\mathrm{line}}^2, and npixlinen_{\mathrm{pix}}^{\mathrm{line}} are defined in § 4.1.2.

We infer AA, Δv\Delta v, and σ\sigma by optimizing the pre-convolution Gaussian parameters given by equation (21), as well as χline2\chi_{\mathrm{line}}^2 for the fit and the number of unmasked pixels npixlinen_{\mathrm{pix}}^{\mathrm{line}} contributing to it (both within ±3σ\pm3\sigma of the line-center). We additionally measure the observed amplitude of each line, AobsA_{\mathrm{obs}}, the peak flux density of the fully pixelized, resolution-matrix-convolved model E=WF\mathbf{E} = \mathbf{W}\mathbf{F} (equation (22)) as it actually appears on the DESI pixel grid. In general (but especially for lines narrower than one pixel), the pre-convolution of § 4.1 means AA can substantially exceed AobsA_{\mathrm{obs}}; AobsA_{\mathrm{obs}}, not AA, is what one would measure by eye in the spectrum and is therefore one of the quantities we can use to determine the statistical significance of a line (§ 6.2).

Next, we compute the Gaussian-integrated flux Φ\Phi analytically from the fitted parameters using equation (18), again, independent of the pixel grid or resolution matrix. As a model-independent check on Φ\Phi, we also derive the boxcar-integrated flux Φbox\Phi_{\mathrm{box}} by directly summing the continuum-subtracted spectrum

Φbox=∑i∈patchriΔλi,(28)\Phi_{\mathrm{box}} = \sum_{i \in\mathrm{patch}} r_i \Delta\lambda_i, \tag*{(28)}

over the unmasked pixels within ±3σ\pm3\sigma of the line center (requiring at least seven pixels, and falling back to a default width if no measurement of σ\sigma is available for that line). Naturally, Φbox\Phi_{\mathrm{box}} is a biased estimate of the emission-line flux for blended or overlapping lines (e.g., [O II] λλ3726\lambda\lambda3726, 29) and also suffers from the subpixel bias for very narrow lines discussed in § 4.1.

For each line, we estimate the local continuum flux density FcF_c, by computing the iteratively sigma-clipped mean of the observed spectrum, with the best-fitting emission-line model subtracted, over two sideband windows spanning 33–10σ10\sigma on either side of the line-center (again requiring at least seven pixels per side). This local, data-driven estimate is more robust to continuum-model mismatch at the line’s exact wavelength than simply interpolating C(λ)C(\lambda) would be. The rest-frame equivalent width (EW) then follows from

EW=ΦFc(1+z).(29)\mathrm{EW} = \frac{\Phi}{F_c(1+z)}. \tag*{(29)}

We report the 1σ1\sigma upper limit on the flux of every line with a valid continuum measurement by substituting the uncertainty in the local continuum for the amplitude in equation (18),

Φlim=2π(λ∗IFc)σc,(30)\Phi_{\mathrm{lim}} = \sqrt{2\pi}\left(\frac{\lambda^{*}}{\sqrt{I_{F_c}}}\right)\frac{\sigma}{c}, \tag*{(30)}

where IFcI_{F_c} is the inverse variance of FcF_c (discussed in § 4.3). However, because IFcI_{F_c} is the uncertainty on the mean continuum in the sidebands (not at the wavelength of the line), Φlim\Phi_{\mathrm{lim}} is only an approximate flux limit. For lines that are dropped during fitting, we use a nominal σ\sigma of 75 km s−1^{-1} for narrow lines and 1200 km s−1^{-1} for broad lines, so that we can still report a limit for every line in the spectrum. We compute Φlim\Phi_{\mathrm{lim}} as another possible metric for determining the statistical significance of a line (§ 6.2), and so that users can more easily fold upper limits into their specific scientific analyses.

Separately, for four lines—C IV λ1549\lambda1549, Mg II λλ2796\lambda\lambda2796, 2803 (summed as one blended feature), Hβ\beta, and [O III] λ5007\lambda5007—we also measure three non-parametric moments of the line profile directly from the residual spectrum, without assuming any particular functional form. Departures from a Gaussian profile, particularly asymmetric wings, can signal outflows or other physical processes, most often associated with AGN; these moments provide a model-independent starting point for a more detailed follow-up analysis of such profiles.

We sum over pixels within ±5σ\pm5\sigma of the line center λ∗\lambda^{*} (equation (17)), using the same fallback width adopted above when σ\sigma is unmeasured (75 km s−1^{-1} for narrow lines, 1200 km s−1^{-1} for broad), and requiring a window at least two pixels wide. Writing λi\lambda_i and rir_i for the observed wavelength and residual flux of the iith pixel in that window, the first moment is the flux-weighted mean wavelength,

M1=∑iλiri∑iri,(31)M_1 = \frac{\sum_i \lambda_i r_i}{\sum_i r_i}, \tag*{(31)}

and the nnth central moment is

Mn=∑i(λi−M1)nri∑iri;(32)M_n = \frac{\sum_i (\lambda_i-M_1)^n r_i}{\sum_i r_i}; \tag*{(32)}

we evaluate equation (32) for the first two non-trivial orders, M2M_2 and M3M_3. We tabulate M1M_1, M2M_2, and M3M_3 (in observed-frame A˚\mathrm{\AA}, A˚2\mathrm{\AA}^2, and A˚3\mathrm{\AA}^3, respectively), together with their Monte Carlo-derived inverse variances, as MOMENT1, MOMENT2, and MOMENT3 columns prefixed by each line’s name (e.g., OIII_5007_MOMENT1).

In § 4.3 we describe how we derive the companion inverse-variance columns listed in Table 3 via a per-object Monte Carlo procedure.

Spectrophotometric and Physical Properties

Because physical SPS templates underpin our best-fitting stellar continuum model C(λ;θC)C(\lambda; \theta_C), we can compute several secondary, albeit scientifically high-impact physical and spectrophotometric quantities. First, recall from equation (2) that each template coefficient cnc_n carries units of M⊙\mathcal{M}_{\odot} of stars formed in age bin nn, rescaled by the mass normalization M0\mathcal{M}_0 introduced in equation (10). The current stellar mass (in h−2 M⊙h^{-2}\,\mathcal{M}_{\odot}), therefore, is simply

M⋆=M0h2∑n=1ntemplatecnf⋆,n,(33)\mathcal{M}_{\star} = \frac{\mathcal{M}_0}{h^2} \sum_{n=1}^{n_{\mathrm{template}}} c_n f_{\star,n}, \tag*{(33)}

where f⋆,n∈(0,1]f_{\star,n} \in(0, 1] is the fraction of formed mass that remains in stars and stellar remnants after mass loss from stellar winds and supernovae (taken directly from FSPS). The mean star-formation rate (SFR) over the last 100 Myr also follows by summing the mass formed within our first two age bins (0–30 Myr and 30–100 Myr; § 3.1), and dividing by 100 Myr,

SFR=M0h2(108 yr)∑n=12cn,(34)\mathrm{SFR} = \frac{\mathcal{M}_0}{h^2(10^8\ \mathrm{yr})} \sum_{n=1}^{2} c_n, \tag*{(34)}

in units of h−2 M⊙/yrh^{-2}\,\mathcal{M}_{\odot}/\mathrm{yr}. Additional estimates of the SFR of course can be inferred from various other tracers, such as the Hα\alpha, [O II], or UV luminosity (where they have been reliably measured and with appropriate corrections for dust attenuation and other effects; [93, 69, 56]), but FastSpecFit does not directly report those quantities. Finally, we compute the mean stellar age as the coefficient-weighted average of the age-bin midpoints tnt_n (in Gyr),

⟨t⋆⟩=∑n=1ntemplatecntn∑n=1ntemplatecn.(35)\langle t_{\star} \rangle= \frac{\sum_{n=1}^{n_{\mathrm{template}}} c_n t_n}{\sum_{n=1}^{n_{\mathrm{template}}} c_n}. \tag*{(35)}

Because the coefficients cnc_n represent formed stellar mass rather than surviving mass or luminosity, ⟨t⋆⟩\langle t_{\star} \rangle weights older stellar populations more heavily than either a light-weighted or a surviving-mass-weighted age, and so is at least as old as both.

Next, we compute K-corrections and rest-frame absolute magnitudes by synthesizing photometry from the best-fitting continuum model C(λ)C(\lambda) (including nebular emission; § 3.1), in the observed and rest frames, following the standard formalism of D. W. Hogg et al. [60] and M. R. Blanton et al. [14]. For each output bandpass we identify the observed-frame Legacy Surveys or WISE filter whose rest-frame effective wavelength lies closest to the target band, restricted to filters detected at S/N >2> 2. This nearest-bandpass matching minimizes the magnitude of the K-correction at the cost of an expected discontinuity in the derived K-corrections as a function of redshift wherever the nearest-matching observed filter switches from one bandpass to the next. When no bandpass satisfies the S/N threshold, we instead report the absolute magnitude synthesized purely from the model, with its inverse variance set to zero.

The absolute magnitude follows the standard relation (M. R. Blanton & S. Roweis [13]),

mR=MQ+DM(z)+KQR(z)−5log⁡10h,(36)m_R = M_Q + \mathrm{DM}(z) + K_{QR}(z) - 5\log_{10} h, \tag*{(36)}

where mRm_R is the observed-frame apparent magnitude in bandpass RR (corrected for Galactic extinction), MQM_Q is the absolute magnitude in the rest-frame target bandpass QQ, DM(z)\mathrm{DM}(z) is the distance modulus, and KQR(z)=2.5log⁡10(frestsynth/fobssynth)K_{QR}(z) = 2.5\log_{10}(f_{\mathrm{rest}}^{\mathrm{synth}}/f_{\mathrm{obs}}^{\mathrm{synth}}) is the K-correction from bandpass QQ to its nearest-matching observed bandpass RR. To build band-shifted rest-frame bandpasses (M. R. Blanton et al. [14]) we blueshift the filter response itself rather than the model,

RQz0(λ)≡RQ((1+z0)λ),(37)R_Q^{z_0}(\lambda) \equiv R_Q\bigl((1+z_0)\lambda\bigr), \tag*{(37)}

so that convolving RQz0R_Q^{z_0} with the rest-frame SED reproduces what RQR_Q would measure on the same galaxy at redshift z0z_0. Table 4 reports the default set of rest-frame bandpasses and band-shifts computed by FastSpecFit via the bundled data/legacysurvey-dr9.yaml parameter file, as well as the speclite filter curve names and their respective rest-frame wavelengths.

BandFilter Curveλeffa\lambda_{\mathrm{eff}}^{a}Band-shift z0z_0
DECam ggdecam2014-g48901.0
DECam rrdecam2014-r64701.0
DECam zzdecam2014-z91961.0
Johnson UUbessell-U36180.0
Johnson BBbessell-B44420.0
Johnson VVbessell-V55360.0
2MASS JJtwomass-J124080.0
SDSS uusdss2010-u36140.1
SDSS ggsdss2010-g47490.1
SDSS rrsdss2010-r62060.1
SDSS iisdss2010-i75260.1
SDSS zzsdss2010-z89470.1
WISE W1wise2010-W1340030.1

Table 4. Rest-Frame Bandpasses for Absolute Magnitudes and K-Corrections. Note—Default rest-frame bandpasses, band-shifts z0z_0, and speclite filter curves used to infer K-corrections and absolute magnitudes (§ 4.2), as configured in the bundled data/legacysurvey-dr9.yaml parameter file. a^{a}Effective filter wavelengths in Å are reported for the unshifted filter curve.

In addition to rest-frame photometry, we also measure several rest-frame monochromatic luminosities and observed-frame continuum flux densities using the best-fitting continuum-only model, C(λ)C(\lambda), which excludes nebular emission. We report luminosities at 1450, 1700, 3000, and 5100 A˚\mathrm{\AA} in solar units (λLλ\lambda L_\lambda), and at 1500 and 2800 A˚\mathrm{\AA} in erg s−1^{-1} Hz−1^{-1} (LνL_\nu), the latter chosen to facilitate UV-continuum SFR estimates (e.g., R. C. Kennicutt [69]). We similarly report the observed-frame continuum flux density (in erg s−1^{-1} cm−2^{-2} A˚−1\mathrm{\AA}^{-1}) at the rest-frame centers of Lyα\alpha, [O II], Hβ\beta, [O III] λ5007\lambda5007, and Hα\alpha, which can be used to infer the total line-luminosity from the corresponding emission-line EW (e.g., G. Zhu et al. [143]). In all cases, we measure these quantities directly from C(λ)C(\lambda), using a robust (sigma-clipped) linear fit to the continuum within ±100\pm100 A˚\mathrm{\AA} of each reference wavelength.

Finally, we measure three different flavors of the 4000 A˚\mathrm{\AA} break, Dn(4000)D_n(4000) (M. L. Balogh et al. [9]): DN4000_OBS, computed directly from the observed spectrum, without any correction for emission-line contamination; DN4000_MODEL, from the best-fitting stellar continuum model (excluding the smooth continuum correction; § 3.4.4); and our fiducial DN4000, from the observed spectrum after subtracting the best-fitting emission-line model.

Parameter Uncertainties via Monte Carlo

FastSpecFit’s forward model chains together several nonlinear stages: an initial grid search for σ⋆\sigma_\star (§ 3.4.1), the VARPRO solve for τV\tau_V and c\mathbf{c} (§ 3.4.3), and the nonlinear emission-line kinematics (§ 4.1.1). Analytic error propagation through every downstream quantity (stellar mass, K-corrections, EWs, etc.) is therefore impractical. Monte Carlo resampling instead offers a fast, general way to sample the multi-dimensional covariance among all fitted and derived parameters. By perturbing the input data and refitting, each nonlinear stage of the pipeline propagates its own uncertainty naturally, without having to derive a bespoke error-propagation formula for every derived quantity.

By default, FastSpecFit generates nmonte=50n_{\mathrm{monte}} = 50 independent realizations of the observed spectrum and photometry by adding Gaussian noise consistent with the (floor-inflated; § 3) inverse variance,

di(m)∼N(di,Ii−1),(38)d_i^{(m)} \sim\mathcal{N}\left(d_i, I_i^{-1}\right), \tag*{(38)}

for realizations m=1,…,nmontem = 1, \ldots, n_{\mathrm{monte}}, where did_i indexes both spectral pixels and photometric bandpasses. We generate these draws once, using a reproducible per-object seed recorded in the SEED output column. This seed is itself derived deterministically from a single global run seed. We reuse the same nmonten_{\mathrm{monte}} realizations, unchanged, throughout the rest of the pipeline, so the same draws flow through the velocity-dispersion, continuum, and emission-line fits described below. The inverse variance of any fitted or derived quantity θ\theta then follows from the scatter across realizations,

σθ−2=[Var⁡(θ(1),…,θ(nmonte))]−1.(39)\sigma_{\theta}^{-2} = \left[\operatorname{Var}\left(\theta^{(1)}, \ldots, \theta^{(n_{\mathrm{monte}})}\right)\right]^{-1}. \tag*{(39)}

Throughout the pipeline, we report uncertainties as inverse variances; for parameters that are not measured for a given object, or that are held fixed at a nominal value, we assign IVAR=0 rather than a Monte Carlo estimate.

On the continuum side, the Monte Carlo realizations enter in two places. First, after determining σ⋆\sigma_\star (§ 3.4.1), we refit τV\tau_V, σ⋆\sigma_\star, and c jointly on each realization to obtain VDISP_IVAR. Second, with σ⋆\sigma_\star fixed at its adopted value, we repeat the full VARPRO solve—the outer τV\tau_V minimization and inner NNLS solve for c—on each realization, yielding nmonten_{\mathrm{monte}} independent estimates τV(m)\tau_V^{(m)} and c(m)c^{(m)} that directly give TAUV_IVAR. Because every stellar-population and photometric quantity in § 4.2 is a fixed function of c (equations 33–37), we re-evaluate those relations on each c(m)c^{(m)} and propagate the resulting scatter to determine LOGMSTAR_IVAR, SFR_IVAR, AGE_IVAR, DN4000_MODEL_IVAR, the ABSMAG*_SYNTH_IVAR, and the rest-frame luminosity and continuum-flux inverse variances.

The emission-line Monte Carlo is deliberately cheaper. The full kinematic model-selection procedure (§ 4.1.1) chooses between narrow-only and narrow+broad models and optionally relaxes velocity-width bounds; repeating it for every Monte Carlo realization would be expensive. Instead, we fix the kinematic model and velocity-width bounds at their fiducial values and refit only the line parameters, warm-started at the fiducial best-fit solution with loosened convergence tolerances. The assumption underpinning these choices is that each realization is a small perturbation of the data rather than an independent object, so the preferred kinematic model should rarely change between realizations. We build each realization’s continuum-subtracted spectrum from the corresponding continuum-fitting realization, d(m)−C(m)(λ)−Csmooth(λ)d^{(m)} - C^{(m)}(\lambda) - C_{\mathrm{smooth}}(\lambda), holding the smooth continuum correction Csmooth(λ)C_{\mathrm{smooth}}(\lambda) (§ 3.4.4) fixed at its fiducial value across all realizations rather than recomputing it per draw. This procedure yields Monte Carlo uncertainties for every fitted line parameter (§ 4.1.2 and Table 3).

Additional Fitting Modes

Everything described in this section so far uses FastSpecFit’s default fitting code, fastspec, which jointly fits each object’s DESI spectrum with its corresponding optical/IR broadband photometry using a nominal set of modeling priors. However, we designed FastSpecFit with flexibility and extensibility in mind, letting users choose different input data and model parameters while sharing a single, speed-oriented fitting engine.

In practice, FastSpecFit exposes this flexibility through command-line options that specify what data to fit. For example, the --ignore-photometry option drops the broadband photometry from the fastspec fit, effectively disabling the aperture correction (§ 3.4.2). For a scientific study focused on lower-redshift BGS targets, this option is arguably more relevant. Other comparable options such as --no-smooth-continuum disable the smooth continuum correction (§ 3.4.4), while --no-broadlinefit skips the narrow+broad kinematic fitting (§ 4.1.1), an option that may be useful when studying, for example, H II-region-dominated spectra.

A second major fitting mode is fastphot, which fits only the broadband photometric SED with the templates convolved to a nominal velocity dispersion, σ⋆=150 km s−1\sigma_{\star} = 150\ \mathrm{km\ s^{-1}}. As for fastspec, in this mode the redshift is an input quantity, taken from the DESI pipeline’s Redrock output for whichever DESI coaddition type—per-exposure, per-night, cumulative (multi-night), or per-HEALPix—is detected automatically from the input Redrock file [53, 39]. However, the user can override the pipeline redshifts via the --input-redshifts argument if externally supplied redshifts (e.g., from visual inspection) are available.

A third fitting mode is called stackfit, which fits arbitrary user-supplied stacked spectra (e.g., binned by redshift, stellar mass, or another physical property) rather than spectra of individual DESI targets (e.g., [119]). In this mode, the user is responsible for providing a “stacked” (mean) resolution matrix and for recognizing that the assumption of uncorrelated pixels no longer holds, but at DESI scale, stacked spectra can reveal significantly fainter spectroscopic features. With stackfit, the fit is purely spectroscopic (i.e., --ignore-photometry is enforced).

Finally, three bundled configuration files control the physical inputs to every fitting mode, and each can be swapped out for a custom version. The photometric-catalog file (data/legacysurvey-dr9.yaml by default; see § 4.2) can be overridden with --fphotofile, letting FastSpecFit fit photometry from surveys other than the Legacy Surveys. Users can similarly override the emission-line parameter file (data/emlines.ecsv by default; see § 3.2) with --emlinesfile to add, remove, or retune individual lines, or the kinematic-constraints file (data/emline-constraints.yaml by default; see § 4.1.1) with --constraintsfile to redefine the kinematic groups, velocity-width bounds, and amplitude ratios that link lines during fitting. Users can even build a custom set of SPS templates with bin/build-templates (§ 3.1) and use them in any fitting mode via the --templates command-line argument.

More details regarding these fitting modes and all the available options can be found in the ever-evolving online documentation for FastSpecFit.17 Looking ahead, we describe additional fitting modes and capabilities currently under development in § 7.2.

Illustrative Examples

We illustrate FastSpecFit’s ability to model DESI spectrophotometry of extragalactic targets with three example fits: a low-redshift (z≈0.1z \approx0.1) BGS galaxy dominated by an evolved stellar population but hosting a narrow-line AGN (§ 5.1); an ELG at z≈1z \approx1 with prominent [O II] λλ3726,29\lambda\lambda3726, 29, [O III] λλ4959,5007\lambda\lambda4959, 5007, and higher-order Balmer lines (§ 5.2); and a high-redshift (z≈2z \approx2) QSO with broad Lyα\mathrm{Ly}\alpha, C IV λ1549\lambda1549, and Mg II λλ2796,2803\lambda\lambda2796, 2803 emission (§ 5.3). Together, these examples highlight both the dynamic range of the code and some of its known limitations, most notably the use of purely stellar templates to model QSO continua.

Low-Redshift BGS Galaxy

Figure 3 shows the DESI spectrum and Legacy Surveys color-cutout of a low-redshift (z=0.0813z = 0.0813) BGS target (TARGETID=39628087574398347) with clear line-emission and a stellar continuum dominated by an old, evolved stellar population. The Tractor fit to the Legacy Surveys imaging (§ 2.2) yields an apparent (Galactic extinction-corrected) rr-band magnitude of 15.67 mag, an n=3.70n = 3.70 J. L. Sérsic [121] morphological type, and a half-light semi-major axis of 3\farcs89. Given this galaxy’s angular extent and the 1\farcs5 DESI fiber, we derive a median aperture correction of ηˉ=6.35\bar{\eta} = 6.35, ranging between 5.82 and 7.05 across the three grzgrz bands (§ 3.4.2). An aperture correction this large is not surprising given the galaxy’s redshift and angular size, but it is several times larger than the typical value for BGS targets.

Figure 3. FastSpecFit output for a low-redshift BGS galaxy at z=0.0813z = 0.0813 (TARGETID=39628087574398347; main survey, bright program, HEALPix 27851). Top: broadband SED showing the observed Legacy Surveys grzgrz and WISE W1–W4 photometry with 1σ1\sigma uncertainties (filled orange circles; here, the uncertainties are smaller than the symbols), the best-fitting stellar continuum model (gray curve), and the synthesized model photometry (open diamonds). The top axis gives the rest-frame wavelength, and the gray bar below the photometry marks the observed DESI wavelength range and the median aperture correction, ηˉ=6.35\bar{\eta} = 6.35. The inset in the lower right shows the Legacy Surveys grzgrz color image with the 1\farcs5-diameter DESI fiber (solid) and a 10\arcsec reference aperture (dashed) overlaid. Middle: observed-frame DESI spectrum (light colors, one per camera) with the best-fitting stellar continuum (black), the stellar continuum plus smooth-continuum correction (dashed gray), and the total model including emission lines (bold, per-camera colors). Gray labels mark every emission line in the observed wavelength range included in the fit. Bottom: zooms on four representative line complexes—[O II][\mathrm{O\,II}] λλ3726,29\lambda\lambda3726, 29, Hβ+[O III]\mathrm{H}\beta+[\mathrm{O\,III}] λλ4959,5007\lambda\lambda4959, 5007, Hα+[N II]\mathrm{H}\alpha+[\mathrm{N\,II}] λλ6548,84\lambda\lambda6548, 84, and [S II][\mathrm{S\,II}] λλ6716,31\lambda\lambda6716, 31—showing the spectrum after subtracting both the stellar continuum and the smooth-continuum correction, the individual best-fitting Gaussian components (thin curves), and the total emission-line model (bold curve).

From the fit to the stellar continuum we derive a stellar velocity dispersion of σ⋆=287±11 km s−1\sigma_\star= 287 \pm11\ \mathrm{km}\ \mathrm{s}^{-1}, and from the joint spectrophotometric fit, including the infrared WISE bands, we infer a luminosity and rest-frame color of M0.1r=−21.14M_{0.1r} = -21.14 and 0.1(g−r)=0.944{}^{0.1}(g-r) = 0.944, respectively, a stellar mass of log⁡10(M⋆/M⊙)=11.24\log_{10}(M_\star/M_\odot) = 11.24, and an approximate age of ⟨t⋆⟩=12.7\langle t_\star\rangle= 12.7 Gyr, all of which are consistent with the galaxy’s optical morphology (M. R. Blanton & J. Moustakas [12]). Formally, we also derive a VV-band optical depth of τV=0.73\tau_V = 0.73; however, this surprisingly large dust attenuation value most likely reflects the age-dust-metallicity degeneracy rather than indicating a truly dusty massive galaxy. Our relatively simple SPS modeling choices—restricted set of model ages, fixed solar metallicity, simple IR spectrum, etc. (§ 3.1)—can be susceptible to these kinds of degeneracies for certain classes of galaxies, so the SPS properties inferred by FastSpecFit should be interpreted with care.

The fit to the emission-line spectrum builds out the rest of the physical picture for this galaxy. Despite a relatively quiescent stellar continuum, we detect mul-

tiple forbidden and permitted recombination lines at high statistical significance, including Hα\alpha, Hβ\beta, Hγ\gamma, [O II] λλ3726,29\lambda\lambda3726, 29, [O III] λλ4959,5007\lambda\lambda4959, 5007, [N II] λλ6548,84\lambda\lambda6548, 84, and [S II] λλ6716,31\lambda\lambda6716, 31. We find no evidence for a broad Balmer-line component and derive average narrow-line kinematics of σnarrow=281±2 km s−1\sigma_{\mathrm{narrow}} = 281 \pm2\ \mathrm{km\ s^{-1}} and Δvnarrow=+76±30 km s−1\Delta v_{\mathrm{narrow}} = +76 \pm30\ \mathrm{km\ s^{-1}} (§ 4.1.1). Not surprisingly for this type of object, the log⁡10([O III] λ5007/Hβ)=+0.53\log_{10}([\mathrm{O\ III}]\ \lambda5007/\mathrm{H}\beta) = +0.53 and log⁡10([N II] λ6584/Hα)=−0.25\log_{10}([\mathrm{N\ II}]\ \lambda6584/\mathrm{H}\alpha) = -0.25 line-ratios place this galaxy clearly into the type 2 Seyfert/AGN region of the Baldwin, Phillips, and T erlevich (BPT; [8]) diagram (§ 7.1).

Intermediate-Redshift Emission-Line Galaxy

Figure 4 shows the broadband photometric SED, Legacy Surveys color image, and DESI spectrum of an example intermediate-redshift (z=0.9449z = 0.9449) ELG target (TARGETID=39627663165362386). This galaxy is considerably fainter than the BGS target shown in Figure 3; its Galactic extinction corrected rr-band magnitude is 21.09 mag, although it remains well-detected all the way out to ≈5 μm\approx5\,\mu\mathrm{m} in the two shortest WISE bands. The best-fitting Tractor morphology is a point-spread function (PSF); in other words, the galaxy is unresolved in the Legacy Surveys imaging. Not unexpectedly, the multiplicative aperture correction is small, ranging between 1.013 and 1.033 with a median value of ηˉ=1.026\bar{\eta} = 1.026 (§ 3.4.2).

Figure 4. Same as Figure 3, but for an intermediate-redshift ELG at z=0.9449z = 0.9449 (TARGETID=39627663165362386; main survey, dark program, HEALPix 29973) with a median aperture correction of ηˉ=1.03\bar{\eta} = 1.03. In the top panel, open circles with downward carets indicate broadband non-detections (upper limits). The bottom row zooms on [O II] λλ3726,29\lambda\lambda3726, 29, Hγ\gamma, Hβ\beta, and [O III] λλ4959,5007\lambda\lambda4959, 5007.

Although the stellar continuum is only marginally detected (S/N≈2.5\mathrm{S/N} \approx2.5--3.7 pixel−13.7\,\mathrm{pixel}^{-1} over the three DESI cameras), its shape is consistent with that of a blue, star-forming galaxy. The best-fitting FastSpecFit model yields M0.1r=−21.59M_{0.1r} = -21.59, 0.1(g−r)=0.053{}^{0.1}(g-r) = 0.053, log⁡10(M⋆/M⊙)=9.51±0.21\log_{10}(M_{\star}/M_{\odot}) = 9.51 \pm0.21, and ⟨t⋆⟩=0.06±3.9 Gyr\langle t_{\star} \rangle= 0.06 \pm3.9\,\mathrm{Gyr}. We also infer an SED-based SFR of 39±1 M⊙ yr−139 \pm1\,M_{\odot}\,\mathrm{yr}^{-1} and a dust attenuation value of τV=0.37±0.02\tau_{V} = 0.37 \pm0.02. All these properties are broadly consistent with the optical morphology and color of the galaxy, and are fairly typical for DESI ELG targets.

Despite the relatively low continuum S/N, all the strongest nebular emission lines lying within the DESI spectral range are well-detected, including [O II] λλ3726\lambda\lambda3726, 29, [O III] λλ4959\lambda\lambda4959, 5007, Hβ\beta, and Hγ\gamma, as shown in the bottom row of Figure 4. From the emission-line fits, we measure a mean narrow-line width and velocity center of σnarrow=101±5 km s−1\sigma_{\mathrm{narrow}} = 101 \pm5\,\mathrm{km\,s}^{-1} and Δvnarrow=−3±10 km s−1\Delta v_{\mathrm{narrow}} = -3 \pm10\,\mathrm{km\,s}^{-1}, respectively (§ 4.1.1) and cleanly resolve the [O II] λλ3726\lambda\lambda3726, 29 doublet. The relative strength of the two components of the [O II] doublet depends on the collisional de-excitation rate of their upper levels (D. E. Osterbrock [98]), and so the doublet ratio we measure, λ3726/λ3729=0.841±0.053\lambda3726/\lambda3729 = 0.841 \pm0.053, implies an electron density of ne=275−89+94 cm−3n_{e} = 275^{+94}_{-89}\,\mathrm{cm}^{-3} for the star-forming gas (V. Luridiana et al. [86]). Taken together, the stellar mass, rest-frame color, star-formation rate, and emission-line properties we measure are all consistent with a young, low-mass, actively star-forming galaxy.

High-Redshift Quasar

Figure 5 shows the broadband SED, Legacy Surveys color image, and DESI spectrum of our third and final example, a high-redshift (z=2.0731z = 2.0731) QSO target (TARGETID=39627657142340435) with a Galactic extinction-corrected rr-band magnitude of 18.58 mag. As expected for this class of target, the best-fitting Tractor morphology is a point source (PSF) and the aperture correction is small, ηˉ=1.14\bar{\eta} = 1.14 (§ 3.4.2). At this redshift, the DESI spectrum spans rest-frame wavelengths of ≈1170\approx1170--3200 A˚3200\,\mathring{\mathrm{A}}, revealing a blue, power-law continuum detected at high S/N (S/N≈17\mathrm{S/N} \approx17--21 pixel−121\,\mathrm{pixel}^{-1} over the three DESI cameras). Superposed on this continuum are several prominent, broad UV emission-line complexes, including Lyα\alpha+N V λ1240\lambda1240, C IV λ1549\lambda1549, C III] λ1908\lambda1908+Si III] λ1892\lambda1892+Al III λ1857\lambda1857, and Mg II λλ2796\lambda\lambda2796, 2803.

Figure 5. Same as Figure 3, but for a high-redshift QSO at z=2.0731z = 2.0731 (TARGETID=39627657142340435; main survey, dark program, HEALPix 29972) with a median aperture correction of ηˉ=1.14\bar{\eta} = 1.14. As in Figure 4, open circles with downward carets in the top panel indicate broadband non-detections (upper limits). The bottom row zooms on four broad UV line complexes, Lyα\alpha+N V λ\lambda1240, C IV λ\lambda1549, C III] λ\lambda1908+Si III] λ\lambda1892+Al III] λ\lambda1857, and Mg II λλ\lambda\lambda2796, 2803.

Although the fit to the stellar continuum is formally well-behaved, it is not physically meaningful for this object. Our SPS templates are purely stellar (§ 3.1), so they cannot represent the non-thermal power-law emission from the accretion disk, the Fe II pseudo-continuum in the rest-frame UV (§ 3.4.4), or the hot-dust emission from the torus that typically dominates the mid-IR SED of QSOs ([96 b]). Consequently, the continuum-derived physical properties reported by fastspec (stellar mass, SFR, age, dust attenuation, etc.) should not be used for QSOs. The K-corrections and rest-frame luminosities (§ 4.2), meanwhile, are less sensitive to this template mismatch because they depend mainly on how well the model reproduces the shape of the observed SED, not on its physical interpretation. At z≈2z \approx2, however, the rest-frame optical bands fall in the gap between the observed zz and W1 bands, so even these quantities depend on how the stellar templates interpolate across this gap and should not be used for precision work.

Because none of Hα\alpha, Hβ\beta, or Hγ\gamma falls within the observed spectral range, fastspec does not test the narrow+broad model for this target (§ 4.1.1). Instead, it models the UV emission lines using the global UV/AGN kinematic groups (Table 2), which allow line widths up to 10,000 km s−1^{-1} and velocity shifts up to ±2500\pm2500 km s−1^{-1} regardless of the kinematic model (§ 4.1.1). Averaged over these lines, we measure σUV≈2700±1500\sigma_{\mathrm{UV}} \approx2700 \pm1500 km s−1^{-1} and ΔvUV≈−640±1300\Delta v_{\mathrm{UV}} \approx-640 \pm1300 km s−1^{-1}, where the quoted uncertainties represent the spread among the measured lines, not the statistical uncertainties. These large variations in line-center (away from the Redrock redshift) and line-width mostly reflect the complex kinematics of the broad-line region, which affects different UV lines in a way that depends on ionization state and QSO luminosity (e.g., [122]). However, the bottom row of Figure 5 also shows that a single Gaussian is a poor representation for the strong, often asymmetric UV lines, especially Lyα\alpha, whose blue side is clearly affected by absorption in the Lyα\alpha forest. Consequently, users who require precise UV broad-line kinematics for DESI QSOs should interpret these values as approximate widths and centroids rather than a complete decomposition of the individual line-profiles. We are addressing both the continuum and emission-line limitations illustrated by this example in a fast QSO-fitting mode currently under development (§ 7.2).

DR2 fastspec VAC

In this section we present the FastSpecFit Loa/v2.0 VAC, which will be publicly released as part of DESI DR2. In DESI vernacular, public data releases include one or more spectroscopic productions, or “specprods”, which consist of a well-defined input set of spectra (usually defined by observing date range) processed with a tagged version of the DESI pipeline ([39]). DR2 includes the Loa specprod, named after Mauna Loa on the island of Hawai‘i (DESI specprods are always named after mountains or mountain ranges). In production, a tagged version of FastSpecFit takes the data from a specprod as its input and produces a complete VAC. Here, we present the v2.0 version of the Loa FastSpecFit VAC. In Appendix § 10 we describe a complementary fastphot VAC produced with the same version of the code and the same input DESI redshifts, and in Appendix § 11 we document previously released FastSpecFit VACs, including those from the EDR (Fuji specprod) and DR1 (Iron specprod).

In § 6.1 we describe the sample-selection criteria for the Loa/v2.0 VAC, our correction of a subset of Redrock QSO redshifts, and the VAC’s file organization. We recommend potential quality cuts in § 6.2, discuss the code’s computational performance at DR2 scale in § 6.3, and validate FastSpecFit measurements against indedependent codes and catalogs in § 6.4. Finally, in § 6.5 we use FastSpecFit emission-line velocity-shift residuals to verify the DESI wavelength calibration.

Sample Selection and Catalog Production

The FastSpecFit VAC includes all extragalactic objects in Loa which satisfy the following minimal selection criteria: (1) non-sky target (OBJTYPE = TGT; see A. D. Myers et al. [95]); (2) valid spectrum (ZWARN & NODATANODATA = 0, where the NODATANODATA bit (292^{9}) identifies spectra in which all extracted inverse-variance values are zero; see S. Bailey et al., in preparation); and (3) spectroscopic redshift z>10−3z > 10^{-3} (applied after correcting a small fraction of QSO redshifts; see below). We do not impose any other cuts on targeting bits or other redshift-warning bits, ensuring that the VAC includes the widest possible range of objects across every DESI target class. And although the lower redshift limit potentially excludes a handful of very low-redshift dwarf galaxies (V. Manwadkar et al., in preparation), we impose the cut to ensure a well-defined luminosity distance and distance modulus, which are required by our modeling (§ 3). For a complementary VAC of all stellar (z≈0z \approx0) targets in DESI, we refer the reader to S. Koposov et al. [73].

Unfortunately, the Redrock redshift-fitting pipeline assigns an incorrect redshift to a small but astrophysically important fraction of QSO targets ([25, 5]). The default FastSpecFit pipeline corrects these redshifts using the QuasarNet afterburner catalog (D. Green et al. [50]), which reruns Redrock with a QuasarNet redshift prior and is distributed alongside each Redrock output file ([38, 39]).18 For primary QSO targets and QSO targets observed as part of the WISE variability secondary program ([25, 95]), we replace the original Redrock redshift with this result when QuasarNet independently classifies the object as a QSO and when the maximum line-confidence across six diagnostic transitions (Lyα\alpha, C IV, C III, Mg II, Hβ\beta, and Hα\alpha) exceeds 0.99. For these objects, we use the updated Z and ZWARN values and copy the original Redrock redshift and ZWARN columns to Z_RR and ZWARN_RR, respectively. Table 5 summarizes the total number of objects in the VAC observed as part of the DESI main, special, and SV programs, totaling over 38 million objects, including the approximately 165,000 targets, or 0.4% of the full sample with corrected redshifts.

ProgramNumber of ObjectsNumber of Corrected QSO Redshiftsa^{a}
Main Survey
backup60,3720
bright13,185,4991,161
dark23,273,230157,118
total36,519,101158,279
Special Survey
backup5120
bright105,1721
dark261,706232
total367,390233
Survey Validationb^{b}
backup2,82517
bright438,45675
dark880,6426,009
other36,769106
total1,358,6926,207
Total38,245,183164,719

Table 5. Summary of the DR2/Loa FastSpecFit Value-Added Catalog. a^{a}Described in § 6.1. b^{b}Sum of CMX, SV1, SV2, and SV3 surveys.

We produce the Loa/v2.0 VAC using FastSpecFit version 3.6.1 and the ftemplates-chabrier-2.2.0.fits SPS template library (§ 3.1).19 We run on the HEALPix-coadded DESI spectra and organize the output catalogs identically to how the input spectra and redshift catalogs are organized in DESI DR1 and DR2 (see §3.3 in DESI Collaboration et al. [39] and DESI Collaboration et al., in preparation).20

Each fastspec output file contains well over 1000 columns organized across five FITS extensions:

  • HDU0 (PRIMARY): empty image with processing metadata keywords, including the software version, Monte Carlo random seed, and boolean fitting flags.

  • HDU1 (METADATA): target metadata mirroring the DESI Redrock catalog, including sky coordinates, the input redshift and ZWARN bitmask, targeting bits, and extinction-corrected Legacy Survey photometry.

  • HDU2 (SPECPHOT): derived spectrophotometric quantities, including template coefficients, stellar-population parameters (τV\tau_V, σ⋆\sigma_\star, age, log⁡10(M⋆h−2/M⊙)\log_{10}(M_\star h^{-2}/M_\odot), SFR, and stellar metallicity; § 3.1), Dn(4000)D_n(4000), synthesized photometry, K-corrections, absolute magnitudes, monochromatic luminosities, and continuum flux densities at key emission-line centers.

  • HDU3 (FASTSPEC): emission-line fitting results, including per-camera median S/N and smooth continuum corrections, aperture correction factors, doublet line ratios, and per-line measurements (amplitude, flux, equivalent width, velocity shift, line width, and upper limits) for all modeled emission lines.

  • HDU4 (MODELS): 3D image array [Npix×3×Nobj][N_{\mathrm{pix}} \times3 \times N_{\mathrm{obj}}] with the best-fitting model spectra on a common linear wavelength grid (3600–9824 A˚\mathring{\mathrm{A}}, 0.8 A˚ pixel−10.8\,\mathring{\mathrm{A}}\,\mathrm{pixel}^{-1}, vacuum wavelengths). The three planes store the full spectrophotometric model, the smooth continuum correction alone, and the emission-line model alone, respectively, coadded across the three cameras. We omit this extension from the merged catalogs described below to keep individual file sizes manageable.

For the most complete, column-by-column description of the data model, pinned to the FastSpecFit version used to produce this VAC, we refer to the online documentation.21

Because there are more than 52,000 individual files (one per HEALPix), most users will find it convenient to work with the merged catalogs included as part of the VAC, which combine all per-HEALPix files for a given SURVEY and PROGRAM into a single file:

VACDIR/catalogs/fastspec-SURVEY-PROGRAM.fits.gz

where VACDIR is the top-level VAC directory. However, the bright and dark programs of the main survey contain too many objects to be merged into a single catalog, so those have been split into 12 separate nside = 1 catalogs:22

VACDIR/catalogs/fastspec-loa-main-{bright,dark}-nside1-hpNN.fits

where NN runs from 00 to 11.

As a concrete example, the Python code snippet below shows how to retrieve the fastspec measurements for a single target given only its sky position. It uses fastspecfit.util.radec2pix to locate the nside = 1 HEALPix file covering that position, matches the nearest catalog entry, and reports its redshift, stellar mass, and HαH\alpha equivalent width—drawn, respectively, from the METADATA, SPECPHOT, and FASTSPEC extensions described above.

from astropy.table import Table
from astropy.coordinates import SkyCoord
import astropy.units as u
from fastspecfit.util import radec2pix

# for main-bright/main-dark, the merged VAC is pre-split into 12
# nside=1 HEALPixels; radec2pix locates the file for a given target
ra, dec = 143.508, 55.241  # example coordinates (I Zw 18)
hpix = radec2pix(nside=1, ra=ra, dec=dec)
fastfile = f’fastspec-loa-main-bright-nside1-hp{hpix:02d}.fits’

meta = Table.read(fastfile, hdu=’METADATA’)
specphot = Table.read(fastfile, hdu=’SPECPHOT’)
fast = Table.read(fastfile, hdu=’FASTSPEC’)

# nearest catalog entry to the requested position
coord = SkyCoord(ra * u.deg, dec * u.deg)
catalog = SkyCoord(meta[’RA’] * u.deg, meta[’DEC’] * u.deg)
indx = coord.separation(catalog).argmin()

print(f"z={meta[’Z’][indx]:.4f} "
      f"log(M*/Msun)={specphot[’LOGMSTAR’][indx]:.2f} "
      f"EW(Ha)={fast[’HALPHA_EW’][indx]:.1f} Ang")

Additional examples and tutorials for working with the data can be found in the FastSpecFit documentation.23

Recommended Quality Cuts

Which quality cuts to apply depends on the science being done, so below we describe several FastSpecFit diagnostics rather than recommend a single threshold. With that in mind, the starting point for any sample selection is a target-class-dependent redshift-quality cut using the quantities and bitmasks reported by Redrock (§ 2.3). In the remainder of this paper, we adopt the following “good redshift” criteria for the three primary DESI galaxy classes. First, we define

Q=[COADD_FIBERSTATUS&∼(RESTRICTED∣VARIABLE)=0]&(SPECTYPE!="STAR")&(Z>0.001),(40)\mathcal{Q} = [\mathtt{COADD\_FIBERSTATUS} \mathbin{\&} \sim(\mathit{RESTRICTED} \mathbin{|} \mathit{VARIABLE}) = 0] \mathbin{\&} (\mathtt{SPECTYPE!="STAR"}) \mathbin{\&} (Z > 0.001), \tag*{(40)}

where italics denote bit names rather than catalog columns, the bracketed term is a bitwise AND restricting COADD_FIBERSTATUS to the RESTRICTED and VARIABLE FIBERSTATUS bits24—informational flags that do not by themselves indicate a data-quality problem—and two additional cuts intended to reject stellar contaminants. The per-class criteria are then

BGS:  (ZWARN=0)&(DELTACHI2>40)&(Z<0.8)&Q(41)\mathrm{BGS}: \; (\mathtt{ZWARN}=0) \mathbin{\&} (\mathtt{DELTACHI2}>40) \mathbin{\&} (Z<0.8) \mathbin{\&} \mathcal{Q} \tag*{(41)}
LRG:  (ZWARN=0)&(Z<1.5)&(DELTACHI2>15)&Q(42)\mathrm{LRG}: \; (\mathtt{ZWARN}=0) \mathbin{\&} (Z<1.5) \mathbin{\&} (\mathtt{DELTACHI2}>15) \mathbin{\&} \mathcal{Q} \tag*{(42)}
ELG:  log⁡10[S/N([O II] λλ3726,29)]>0.9−0.2log⁡10(DELTACHI2)&Q(43)\mathrm{ELG}: \; \log_{10}[\mathrm{S/N}([\mathrm{O\,II}]\,\lambda\lambda3726,29)] > 0.9 - 0.2\log_{10}(\mathtt{DELTACHI2}) \mathbin{\&} \mathcal{Q} \tag*{(43)}

where the ELG cut is taken from A. Raichoor et al. [103] but using the [O II] fluxes measured by FastSpecFit (§ 6.4). Of course, although we adopt these particular criteria to generate some of the figures in this paper, per-class cuts continue to be refined by the DESI team, and so readers building their own samples should check the DESI documentation and associated data-release papers for the most recent recommendations.25

For individual emission lines, we recommend selecting on line-amplitude S/N, AMP×AMP_IVAR\mathrm{AMP}\times\sqrt{\mathrm{AMP\_IVAR}}, rather than integrated-flux S/N, FLUX×FLUX_IVAR\mathrm{FLUX}\times\sqrt{\mathrm{FLUX\_IVAR}}. Amplitude directly tests whether a line rises above the local continuum noise, while flux is a derived quantity (given by equation (18)) whose precision also depends on the fitted line width and velocity shift. FastSpecFit applies this same amplitude-based criterion internally to decide which pixels to mask around each line (§ 3.3) and to test for a significant broad-line component (§ 4.1.1). Once a line is judged significant, any analysis that uses its flux directly should instead propagate FLUX_IVAR. Our Monte Carlo procedure (§ 4.3) computes FLUX_IVAR by jointly refitting the amplitude, width, and velocity shift on every realization, so it captures their correlated contribution to the flux uncertainty in a way AMP_IVAR does not.

For continuum-derived quantities, we do not recommend a single threshold. FastSpecFit instead measures dozens of independent diagnostics, continuum and per-line χ2\chi^2 values, aperture corrections, velocity dispersion, and more, precisely so that users can build a sample selection driven by their own science case rather than inheriting ours. RCHI2_PHOT, the reduced χ2\chi^2 of the broadband photometric fit, is one such example and is a natural cut for any analysis that leans on the stellar-population parameters derived from the SED (stellar mass, SFR, K-corrections). Similarly, analyses of stellar velocity dispersions should restrict to VDISP_IVAR >0> 0, which selects objects with a direct measurement rather than a value estimated from the M⋆M_\star–σ⋆\sigma_\star relation or fixed at VDISP_NOMINAL (§ 3.4.1). Finally, we caution against using continuum-derived physical properties for QSOs and broad-line AGN (§ 5.3).

Computational Performance

Because FastSpecFit fits tens of millions of DESI spectra, each refit 50 (or more) times with Monte Carlo resampling (§ 4.3), computational speed is a first-order design requirement. We strive to meet this requirement by optimizing the per-object fitting code and by exploiting the problem’s embarrassingly parallelism.

First, we cache reused arrays, especially template FFTs, the per-object redshift and IGM prefactors, and the χ2(σ⋆)\chi^2(\sigma_\star) grid (§ 3.4.1) across the Monte Carlo realizations (§ 4.3). We implement the numerically intensive kernels—dust attenuation, IGM transmission, resolution-matrix convolution, and the emission-line forward model and its Jacobian—using Numba just-in-time (JIT) compilation to native machine code (S. K. Lam et al. [76]). Moreover, we run an explicit warm-up pass before the parallel production run, compiling every kernel once and writing the compiled artifacts to a shared, persistent on-disk Numba cache; every subsequent process, across all ranks and nodes, loads the compiled kernels from cache rather than recompiling them. With Monte Carlo resampling disabled, these optimizations enable fastspec to fit a typical DESI spectrum in ≈0.1\approx0.1 s, and fastphot (§ 4.4) to fit the broadband photometry of a typical object in ≈10−2\approx10^{-2} s. With Monte Carlo, because each realization repeats most of the same fitting stages, the total runtime scales roughly linearly with the number of Monte Carlo iterations requested (§ 4.3).

In production, we run FastSpecFit on the NERSC Perlmutter supercomputer, whose 3072 CPU nodes each provide 128 physical cores (two 64-core AMD EPYC 7763 processors).26 When fitting the full DR2 sample—38,245,183 objects in 52,325 HEALPix files—the code processes a typical DESI spectrum (with 50 Monte Carlo realizations) in a median 4.8 s (mean 6.0 s) of single-core time. For comparison, fastphot processes the same sample, also with 50 Monte Carlo realizations, in a median 0.42 s of single-core time. We use MPI for Python (mpi4py; L. Dalcin & Y.-L. L. Fang [33]) to distribute independent HEALPix files across ranks and Python multiprocessing to parallelize per-object fits within each rank. Per-object runtime varies between a median 3.9 s/object/core for main-survey dark-time targets (LRGs, ELGs, QSOs) versus 8.1 s/object/core for main-survey bright-time targets (BGS). This gap in time is primarily due to BGS targets more frequently triggering a direct velocity-dispersion measurement, which adds a χ2\chi^2-scan and nonlinear least-squares solve to the per-object cost (§ 3.4.1).

Validation

To validate FastSpecFit’s measurements, we compare our results against several independent codes and catalogs. Naturally, these comparisons are not exhaustive, and because we generate the largest possible comparison samples, it is possible for differences to be larger or smaller for narrower subsets of objects. Nevertheless, the following figures demonstrate that fastspec produces strong emission-line fluxes, stellar masses, and velocity dispersions that match reasonably well to existing independent measurements.

For the subsequent comparisons, we match our VAC to each external catalog by TARGETID when a common DESI identifier exists (A. D. Myers et al. [95]), otherwise using a 1\farcs5 positional threshold. In addition, we confirm each match by requiring redshift agreement, c∣Δz∣<1000c|\Delta z| < 1000 km s−1^{-1}, which is necessary even when comparing two DESI samples because of possible differences between specrods and because of our override of some QSO redshifts (§ 6.1). For each matched pair we also compute the pull,

pull=xref−xextδxext2+δxref2(44)\mathrm{pull} = \frac{x_{\mathrm{ref}} - x_{\mathrm{ext}}}{\sqrt{\delta x_{\mathrm{ext}}^2 + \delta x_{\mathrm{ref}}^2}} \tag*{(44)}

where xx and δx\delta x are the measured quantity—line flux in Figure 6, velocity dispersion in Figure 7—and its uncertainty in the external and fastspec (reference) catalogs, respectively.

Figure 6. Comparison of fastspec emission-line fluxes for four strong nebular emission lines ([O II] λλ3726\lambda\lambda3726, 29, summed, [O III] λ5007\lambda5007, Hβ\beta, and Hα\alpha) against (top) the DESI DR2/Loa emlinefit afterburner measurements ([103]), and (bottom) emission-line fluxes from the H. Zou et al. (in preparation) DR1/Iron VAC (see also [146, 128]). Within each panel, we only plot objects with S/N >3> 3 in both catalogs, and the shaded histogram in the inset shows the pull distribution (equation (44)) compared to a unit Gaussian (black curve). We show the logarithmically-scaled two-dimensional number density as a color map, and the contours enclose 50%, 75%, 95%, and 99.5% of the joint distribution; outside the outermost contour, we show individual galaxies as markers. The dashed line is the one-to-one relation. Finally, each panel lists the matched sample size NN, the median difference in logarithmic flux, Δmed\Delta_{\mathrm{med}} (fastspec minus external) in dex, and the NMAD scatter in dex.

Figure 7. Comparison of fastspec stellar velocity dispersions with those measured by C. E. Ross et al. [106] using pPXF ([20]), for a matched sample of main-survey, bright-time galaxies (§ 3.4.1). As in Figure 6, the color map, contours, and outlying markers show the two-dimensional density distribution, the dashed line is the one-to-one relation, and the inset shows the pull distribution (equation (44)) compared to a unit Gaussian (black curve). We find a median offset and NMAD scatter of 15.9 ± 12.9 km s−1^{-1} (fastspec yields systematically larger σ⋆\sigma_\star values) and a median pull of 1.06 ± 0.82.

In Figure 6 we compare our emission-line fluxes for [O II] λλ3726\lambda\lambda3726, 29 (summed, with the uncertainties added in quadrature), [O III] λ5007\lambda5007, Hβ\beta, and Hα\alpha against the Loa (DR2) emlinefit afterburner measurements [103] and against the Iron (DESI DR1; [39]) Stellar Mass and Emission Line Catalog27 (v1.0) from H. Zou et al. (in preparation; see also [146, 128]). The line-fluxes in all these catalogs have been corrected for foreground Galactic extinction. We require S/N >3> 3 for each line in both catalogs, and combine the SV3 bright and dark programs to maximize the matched sample. (Combining the main bright and dark samples gives nearly identical statistics, simply with many more objects.)

The agreement with emlinefit is very good for [O II], with a median difference and NMAD\mathrm{NMAD}28 scatter of −0.006±0.022-0.006 \pm0.022 dex. For Hα\alpha and Hβ\beta we find median differences of 0.029±0.0200.029 \pm0.020 dex and 0.085±0.0670.085 \pm0.067 dex, respectively—tight, but offset by roughly one NMAD\mathrm{NMAD} (fastspec is brighter). These systematic differences are most likely due to emlinefit’s treatment of the continuum around each line as a constant pedestal with no correction for stellar absorption, which would cause the Balmer lines to be underestimated (positive pull; Figure 6 insets). The differences with [O III], meanwhile, are notably larger, with an offset of −0.115±0.033-0.115 \pm0.033 dex and a distinctly non-Gaussian pull. Because the emlinefit continuum window for [O III] (nominally ±100\pm100 Å in the rest frame around each line) extends blueward to the immediate vicinity of Hβ\beta, it is possible that the same unmodeled stellar absorption depresses [O III]’s local continuum estimate and biases its line flux high.

Meanwhile, the agreement with the VAC from H. Zou et al. (in preparation) is excellent in terms of median offset (∣Δmed∣<0.005|\Delta_{\mathrm{med}}| < 0.005 dex) and scatter (NMAD<0.04\mathrm{NMAD} < 0.04 dex) for all four lines. The pull distributions for all four lines are narrower than a unit Gaussian, as expected because both catalogs measure the same spectra and therefore share the same noise. Consequently, this comparison validates the line fluxes, but it does not test their uncertainties.

Next, Figure 7 compares the fastspec σ⋆\sigma_\star measurements (§ 3.4.1) against those from the DESI DR1 (Iron) Fundamental Plane study of C. E. Ross et al. [106] for the same set of objects observed at 0.0033<z<0.10.0033 < z < 0.1 in the bright program of the main survey (see also K. Said et al. [107]). C. E. Ross et al. [106] fit each DESI spectrum with pPXF (v8.2.4; M. Cappellari [20]) using the empirical Indo-US Coudé Feed stellar library (σIndo−US=30\sigma_{\mathrm{Indo-US}} = 30 km s−1^{-1}; F. Valdes et al. [131]), restricted to the DESI B camera (≈3600\approx3600–55505550 Å, σB≈35\sigma_B \approx35–6565 km s−1^{-1}; see § 2.1). By comparison, our theoretical C3K R3K templates have σC3K=42\sigma_{\mathrm{C3K}} = 42 km s−1^{-1} (§ 3.1) and we fit σ⋆\sigma_\star over a slightly different rest-frame wavelength range, 38003800–60006000 Å (§ 3.4.1).

For N=80,693N = 80,693 objects in common, we find a median offset and NMAD scatter of Δmed=15.9±12.9\Delta_{\mathrm{med}} = 15.9 \pm12.9 km s−1^{-1} (fastspec systematically higher), with a pull of 1.06±0.821.06 \pm0.82, indicating that the offset is comparable to the combined per-object uncertainties. The residual is statistically significant but relatively small on an instrumental scale—about a third of a native DESI pixel (0.80.8 Å, or ≈53\approx53 km s−1^{-1} at 45004500 Å). However, its precise origin is not clear from this comparison alone. For context, C. E. Ross et al. [106] also compare their σ⋆\sigma_\star measurements against D. Thomas et al. [130], who measure velocity dispersions for a sample of Sloan Digital Sky Survey (SDSS; D. G. York et al. [141]) galaxies using the MILES empirical stellar library (P. Sánchez-Blázquez et al. [113]) and an earlier version of pPXF; they find very good agreement, with an offset of only ∼\sim1–2 km s−1^{-1} over most of the σ\sigma range. Certainly, there are real differences between the stellar libraries used in these analyses relative to the theoretical stellar library we adopt, especially with respect to non-solar [α\alpha/Fe] abundance patterns, which may have an effect on the derived velocity dispersion (e.g., [99] (C. Conroy et al. 2014)). On the other hand, fastspec incorporates the wavelength-dependent instrumental resolution (via the DESI resolution matrix) directly into a full forward model, with no pixel resampling; we have not tested how approximating (or neglecting) the resolution matrix, or resampling onto a common wavelength grid, changes the derived velocity dispersion. In conclusion, we find that fastspec yields precise velocity dispersions over the full dynamic range of σ⋆\sigma_\star values spanned by low-redshift galaxies, albeit with an as-yet unexplained systematic difference relative to existing measurements.

Next, in Figure 8 we compare the stellar masses from fastspec against three independent, CIGALE-based ([16]) SED-fitting catalogs, organized by DESI target class (BGS, LRG, and ELG; § 2.3). The top row compares our measurements against the DR2 (Loa) SED-fitting VAC from H. Zou et al. (in preparation; see also H. Zou et al. [146]; J. Sui et al. [128]), while the middle and bottom rows compare our measurements to the DR1 (Iron) *AGN Host Galaxies Physical Properties

Figure 8. Comparison of fastspec stellar masses (h=1h = 1, Chabrier IMF) against three independent, CIGALE-based catalogs, organized by DESI target class along columns. In each row, we compare our measurements against (top) the DR2 (Loa) VAC of H. Zou et al. (in preparation); see also [146, 128], (middle) the DR1 (Iron) CIGALE-AGN VAC of M. Siudek et al. [125], and (bottom) the GSWLC-X2 catalog of S. Salim et al. [109, 108]. As in Figure 6, the color map, contours, and markers show the two-dimensional density distribution, and the dashed line is the one-to-one relation. Overall, we find median offsets of Δmed=0.15\Delta_{\mathrm{med}} = 0.15–0.300.30 dex (fastspec systematically higher) with NMAD = 0.090.09–0.170.17 dex scatter for BGS and LRG targets, with the scatter increasing to NMAD ≈0.4\approx0.4 dex for ELG targets. The median offsets and scatter are consistent with expectations from different (non-parametric versus parametric) SFH priors (J. Leja et al. [83, 82]) and, more generally, from the stellar-library, dust-attenuation, and other modeling choices that differ across independent SED-fitting codes.

VAC from M. Siudek et al. (v1.2; 2024)29 and the GSWLC-X230 catalog of S. Salim et al. [109, 108], respectively. For the first two DESI catalogs, we compare against our combined SV3 bright plus dark catalogs, while for the GSWLC-X2 we compare just against SV3/bright (there are a negligible number of matches from SV3/dark); as before, matching to the main-survey fastspec catalogs increases the number of objects in the comparison sample but not the statistical results or our conclusions.

All three of these external catalogs perform photometric SED fits, but the Zou et al. VAC augments its grzgrz and WISE W1–W2 photometry with 10-band top-hat filter photometry synthesized from the DESI optical spectrum and scaled for aperture bias using an achromatic aperture correction analogous to our own (§ 3.4.2). In addition, all three catalogs adopt a parametric SFH—a delayed-τ\tau model, with an optional recent burst for M. Siudek et al. [125] and S. Salim et al. [108], in contrast to fastspec’s non-parametric, five-bin star-formation history (§ 3.1). M. Siudek et al. [125] also include an AGN component (J. Fritz et al. [48]) in their SED fits. In constructing these comparisons, we correct for differences in IMF and cosmology.

Within the BGS and LRG subsamples, the fastspec stellar masses are systematically higher than each CIGALE-based catalog by Δmed=0.15\Delta_{\mathrm{med}} = 0.15–0.300.30 dex, with tight scatter (NMAD =0.09= 0.09–0.170.17 dex). Both the sign and the magnitude of this offset agree well with previous comparisons between non-parametric and parametric SFH-derived masses, which attribute ≈0.1\approx0.1–0.30.3 dex systematic differences to the additional flexibility of a non-parametric model accommodating more mass in old, high-mass-to-light-ratio stellar populations ([83, 82, 85]). The scatter in the difference for ELG targets is higher, NMAD ≈0.4\approx0.4 dex, likely reflecting the fainter, lower-S/N broadband photometry available for these star-forming targets, especially in the WISE IR bands. More generally, offsets of this size are typical of SED-based stellar masses, which carry systematic uncertainties of ≈0.2\approx0.2–0.30.3 dex depending on the adopted SPS models, priors, and fitting technique ([94, 28]). Overall, the agreement between fastspec and three independent, CIGALE-based stellar-mass catalogs is reasonable, particularly given the different data (photometry only versus spectrophotometry), templates, and star-formation-history parameterizations involved.

Finally, in Figure 9 we compare our SPS-derived SFRs against the same three catalogs. Unlike the stellar masses, the four sets of SFRs are not defined identically. FastSpecFit reports the mean SFR over the last 100 Myr (equation (34)), as does GSWLC-X2 (S. Salim et al. [109]), whereas the Zou et al. and M. Siudek et al. [125] catalogs report the instantaneous SFR and the mean SFR over the last 10 Myr, respectively (DESI Collaboration et al., in preparation). Moreover, only GSWLC-X2 constrains the SFR with observed-frame UV (GALEX; D. C. Martin et al. [88]) photometry.

Figure 9. Comparison of fastspec SFRs (h=1h = 1, Chabrier IMF) against the same three CIGALE-based catalogs shown in Figure 8, organized by DESI target class along columns. We classify galaxies as star-forming or quiescent using a rest-frame 0.1(g−r)0.1(g-r) color cut (equation (45)). For star-forming galaxies, the color map, contours, and markers are as in Figure 6, while quiescent galaxies are shown as a gray density map only. The dashed line is the one-to-one relation. Each panel lists the number of star-forming galaxies NN, the median difference Δmed\Delta_{\mathrm{med}} (fastspec minus external) in dex, and the NMAD scatter in dex, all computed using the star-forming galaxies only.

Because of the difficulty of constraining the SFRs of quiescent galaxies, we follow DESI Collaboration et al. (in preparation) and compute the comparison statistics for star-forming galaxies only using the rest-frame 0.1(g−r)0.1(g-r) color. Separately for BGS, LRG, and ELG targets, we measure the color of the red-sequence ridge in bins of stellar mass and redshift, and fit these measurements with a plane. We then classify a galaxy as star-forming if it is more than 0.2 mag bluer than this ridge-line,

0.1(g−r)<0.706+0.067(log⁡10(M⋆h−2/M⊙)−10)−0.130(z−0.1).(45){}^{0.1}(g-r) < 0.706 + 0.067\left(\log_{10}(M_\star h^{-2}/M_\odot)-10\right)-0.130(z-0.1). \tag*{(45)}

This cut lies ≈0.07\approx0.07 mag blueward of the minimum between the red sequence and the blue cloud (where we can measure it), so it deliberately excludes green-valley galaxies at the cost of some genuinely star-forming galaxies. In the sample matched to the Zou et al. catalog, it classifies 49%, 13%, and 99% of the BGS, LRG, and ELG targets as star-forming, slightly below the fractions that DESI Collaboration et al. (in preparation) find for BGS and LRG targets (51% and 18%).

For star-forming galaxies, our SFRs generally track the external measurements over nearly four decades. For BGS and ELG targets, the median offsets relative to Zou et al. and M. Siudek et al. [125] range from −0.09-0.09 to +0.24+0.24 dex with NMAD = 0.28–0.41 dex. For the relatively small fraction of star-forming LRGs, meanwhile, the scatter is much larger, NMAD = 0.57–0.66 dex, with large median offsets that are opposite in sign (−0.41-0.41 and +0.20+0.20 dex), indicating that none of these catalogs constrains their SFRs well (see also DESI Collaboration et al., in preparation).

Compared to the GSWLC-X2 catalog, which is only available for BGS targets, the FastSpecFit SFRs are systematically higher by nearly a factor of three, Δmed=0.44±0.23\Delta_{\mathrm{med}} = 0.44 \pm0.23 dex. To determine which set of SFRs is more accurate, we compute an independent SFR for the subset of these galaxies with well-measured (S/N >3> 3) Hα\alpha and Hβ\beta line-emission. We correct the Hα\alpha flux for aperture effects and dust attenuation using the Balmer decrement, and convert it to SFR following C.-N. Hao et al. [56]. GSWLC-X2 agrees with these Hα\alpha-based SFRs to within ≈0.1\approx0.1 dex, whereas the FastSpecFit SFRs exceed them by ≈0.57\approx0.57 dex, with no dependence on the size of the aperture correction. The Zou et al. and M. Siudek et al. [125] SFRs also exceed the Hα\alpha-based SFRs, by 0.34–0.41 dex, so this excess is not unique to FastSpecFit.

Our stellar masses for these galaxies are also higher than GSWLC-X2, by 0.20 dex (Figure 8, bottom row), so the corresponding offset in specific SFR, SFR/M⋆\mathcal{M}_\star, is about half as large, ≈0.20\approx0.20 dex. What distinguishes GSWLC-X2 from the other three catalogs, however, is its UV photometry, which leads us to posit that the SFR discrepancy we find is due to the well-known degeneracy between dust attenuation and stellar age [e.g. 28]. At optical and near-IR wavelengths, a fit can reproduce the same red continuum with either an older, less attenuated stellar population or a younger, dustier one, and the latter solution implies a higher SFR once corrected for dust. Rest-frame UV photometry, which is dominated by young, massive stars, breaks this degeneracy. Consistent with this picture, the SFR offset relative to GSWLC-X2 correlates with the difference in inferred attenuation (Spearman ρ=0.50\rho= 0.50). Moreover, the Zou et al. SFRs, which have the largest median attenuation of the three catalogs (τV=0.64\tau_V = 0.64, versus 0.39 for fastspec and 0.28 for GSWLC-X2), track this difference even more closely (ρ=0.82\rho= 0.82). We conclude that, without rest-frame UV constraints, the SPS-derived SFRs from FastSpecFit (and from other optical and near-IR SED fits) should be used with caution, particularly in an absolute sense. The relative SFRs, on the other hand, are more robust, given that they track the external measurements over nearly four orders of magnitude. For star-forming galaxies with well-measured Hα\alpha and Hβ\beta emission, dust-corrected SFRs computed from our emission-line fluxes provide a more reliable absolute calibration.

Pipeline Diagnostics

One of the dividends of fitting every DESI spectrum with a physically motivated forward model is that the residuals carry information about possible systematic calibration errors in the reduced data. To illustrate this kind of application, in this section we use fastspec measurements to characterize wavelength-solution errors revealed by correlated per-camera emission-line velocity shifts.

FastSpecFit does not fit the three DESI cameras (B, R, Z) independently. Narrow emission lines instead share a common kinematic anchor during the primary fit—[O III] λ5007[\mathrm{O\,III}]\,\lambda5007 for forbidden lines, Hα\alpha for narrow Balmer and He I lines (§ 4.1.1). However, in the final fitting pass, we free the velocity shift, VSHIFT of individual lines and doublets and re-optimize the solution within a ±100\pm100 km s−1^{-1} window of the tied value, with the relaxed solution adopted only if it improves χ2\chi^2 (§ 4.1.1).

Because the rest wavelength and systemic redshift zz together determine each line’s nominal observed-frame wavelength, this population of quasi-independent VSHIFT measurements, aggregated across many lines and redshifts, serves as a proxy for residual wavelength-solution errors as a function of observed-frame wavelength even though no single camera is ever fit in isolation.

To investigate this effect, we select main-survey galaxies passing our baseline redshift-quality criteria (§ 6.2) for which the relaxed kinematic pass is adopted and no broad Balmer component was fit, so that every narrow line’s VSHIFT is a genuine measurement rather than a copy of its tied anchor. We select any line detected with flux S/N >7> 7 and a formal VSHIFT uncertainty below 25 km s−1^{-1}. We exclude the 0.02% of fits pinned at the optimizer’s hard ±500\pm500 km s−1^{-1} bound, which biases their formal uncertainties toward implausibly small values. These cuts yield a unique set of 16 usable emission lines (11 forbidden lines and 5 Balmer lines), spanning [O II] λλ3726,29[\mathrm{O\,II}]\,\lambda\lambda3726,29 through [S III] λ9532[\mathrm{S\,III}]\,\lambda9532.

Figure 10 shows the resulting running median VSHIFT as a function of observed-frame wavelength for forbidden lines (blue), Balmer and helium lines (orange), and all lines combined (black), as well as interquartile range of individual measurements within each bin (shaded band). The forbidden-line median is consistent with zero to within ±1\pm1 km s−1^{-1} across the full DESI wavelength range, with no discontinuity at the B/R or R/Z camera boundaries (vertical shaded gray band). The Balmer lines, by contrast, show a ≈−5\approx-5 km s−1^{-1} offset between roughly 4800 and 5800 Å, the observed-frame window populated by Hβ\beta and higher-order Balmer lines at low redshift, returning to zero once Hα\alpha dominates the sample redward of ≈6600\approx6600 Å. However, because this offset appears only in the Balmer lines and not in the forbidden lines at the same observed wavelength, we attribute it to imperfect continuum-subtraction in and around the higher-order Balmer lines rather than a wavelength-calibration error, which would affect all species equally at fixed observed wavelength. Given DESI’s per-pixel velocity sampling of 25–66 km s−1^{-1} over this wavelength range (§ 4), the <1<1 km s−1^{-1} forbidden-line residual amounts to no more than a few percent of a single pixel, indicating that the wavelength solution is calibrated well below the pixel scale across the full DESI optical range.

Figure 10. Running median velocity shift Δv\Delta v (top) and the number of contributing measurements (bottom) as a function of observed-frame wavelength for DR2 main-survey galaxies with a relaxed, per-line kinematic solution (§ 4.1.1). We compute statistics in equal-width (≈100\approx100 Å) bins spanning the full DESI wavelength range (3600–9824 Å). The blue and orange curves show the derived trends using forbidden and Balmer lines, respectively, while the black curve combines all the lines; the shaded horizontal bands show the interquartile range of the individual measurements within each bin. Gray vertical bands mark the B/R and R/Z camera-boundary transitions, and the thin horizontal dashed line marks zero offset. Although the Balmer lines show a ≈−5\approx-5 km s−1^{-1} offset between roughly 4800 and 5800 Å, we attribute this shift to imperfect stellar absorption subtraction of the higher-order Balmer lines rather than systematic wavelength-calibration errors. Using the forbidden lines, we measure a systematic error of <1< 1 km s−1^{-1} (≲\lesssim few% of a pixel) over the full DESI spectral range, including at the inter-camera boundaries.

One final notable finding: We exclude the [S III] λλ9069,9532[\mathrm{S\,III}]\,\lambda\lambda9069,9532 doublet from Figure 10 because both lines show a consistent +11+11–1212 km s−1^{-1} offset which is not corroborated by any other lines in that part of the observed-frame spectrum. Therefore, we interpret the offset as a rest-wavelength error rather than a wavelength-calibration residual. Taking the median VSHIFT of each line as a pure rest-wavelength correction, i.e., setting Δv=0\Delta v = 0 in equation (17) and solving for λrest\lambda_{\mathrm{rest}}, implies revised rest wavelengths of 9071.4 Å and 9533.6 Å for [S III] λ\lambda9069 and [S III] λ\lambda9532, respectively, ≈ 0.3–0.4 Å redward of the NIST (Ritz31) values used in this version of FastSpecFit.

Discussion

The FastSpecFit DR2 VAC provides simultaneous measurements of stellar continuum, emission-line, and broadband SED properties for more than 38 million extragalactic DESI targets spanning a broad range of intrinsic physical properties and redshift. Although a specific scientific investigation is outside the scope of the current paper, in §6.1 we briefly illustrate the scientific reach of the FastSpecFit VAC with two example figures on the stellar mass and AGN content of the DESI sample. We then conclude in §6.2 with a short discussion of some of the planned updates and improvements to the code.

Example Scientific Applications

We begin by showing the distribution of stellar mass versus redshift for BGS, LRG, and ELG targets in Figure 11. We limit the sample to the SV3 bright (BGS) and SV3 dark (LRG, ELG) programs for this illustrative figure, but the results using the DR2 main-survey data are equivalent. For each panel, we only retain objects which pass the target-class-dependent criteria listed in § 6.2, as well as requiring LOGMSTAR_IVAR >0> 0.

Figure 11. Stellar mass versus redshift for BGS (left), LRG (center), and ELG (right) targets observed during the DESI SV3 bright and dark programs. The color scale and contours in each panel show the two-dimensional number density of galaxies passing per-class redshift quality cuts (see § 6.2); contours enclose 50, 75, 95, and 99.5% of the sample, and individual galaxies outside the outermost contour are plotted as points.

This figure shows that the three target classes trace complementary regions of stellar mass and redshift. The BGS sample spans the widest mass range at z≲0.6z \lesssim0.6, and includes galaxies from well below the knee of the stellar mass function to the most massive systems in the local universe (J. Moustakas et al. [94]; C. Hahn et al. [55]); the LRG sample includes, by design, the most massive, passively evolving galaxies over 0.3≲z≲1.50.3 \lesssim z \lesssim1.5 (R. Zhou et al. [142]); and the ELG sample preferentially consists of low-to-intermediate mass star-forming systems at 0.6≲z≲1.60.6 \lesssim z \lesssim1.6, with a secondary tail of dwarf galaxies, log⁡10(M⋆h−2/M⊙)≲9\log_{10}(\mathcal{M}_{\star}h^{-2}/\mathcal{M}_{\odot}) \lesssim9, to z≈0z \approx0 (V. Manwadkar et al., in preparation; A. Raichoor et al. [103]). Together, these three samples provide expansive coverage of the galaxy population from z≈0z \approx0 to z≈1.6z \approx1.6, making the DR2 catalog a uniquely comprehensive resource for studies of galaxy evolution over the past ≈10\approx10 Gyr.

Next, in Figure 12 we construct two optical emission-line diagnostic diagrams from the FastSpecFit VAC which can be used to identify and characterize AGN. For simplicity, we show just BGS targets from the SV3 bright program with a flux S/N > 3 on all the relevant emission lines. For a more comprehensive sample of AGN in the BGS sample and other target classes, see S. Juneau et al. [67], M. Siudek et al. [126], and S. Juneau et al., in preparation. The left-hand panel of Figure 12 shows the traditional BPT (J. A. Baldwin et al. [8]) diagram, [O III] λ5007/Hβ[\mathrm{O\,III}]\,\lambda5007/\mathrm{H}\beta versus [N II] λ6584\lambda6584/Hα\alpha, with the empirical star-forming/AGN demarcation curve from G. Kauffmann et al. [68] shown as a dashed curve. Galaxies above and below this boundary have line-emission consistent with ionization from an AGN and star formation, respectively. Meanwhile, the star-formation sequence itself is a sequence in gas-phase abundance, with a second-order dependence on excitation or ionization parameter (e.g., L. J. Kewley & M. A. Dopita [70]).

Figure 12. AGN diagnostic diagrams for BGS galaxies in the SV3 bright program. In both panels, we require flux S/N > 3 in all the emission lines used to construct each diagram, and the background color map and contours show the number density of objects. (Left) Traditional BPT diagram (J. A. Baldwin et al. [8]), where the dashed curve shows the empirical star-forming/AGN demarcation boundary from G. Kauffmann et al. [68]. (Right) P1P_1–P3P_3 optical diagnostic diagram of X. Ji & R. Yan (2020), which separates star-forming galaxies (low P1P_1) from AGN (high P1P_1). P1P_1 and P3P_3 are defined in equations (46) and (47). The top axis shows the AGN fraction fAGNf_{\mathrm{AGN}} mapped from P1P_1 following G. Jin et al. [65].

The right-hand panel of Figure 12 visualizes the same BGS sample in the P1P_1–P3P_3 optical diagnostic diagram of X. Ji & R. Yan (2020), where

P1≡0.63log⁡([N II]/Hα)+0.51log⁡([S II]/Hα)+0.59log⁡([O III]/Hβ)(46)P_1 \equiv0.63 \log([\mathrm{N\,II}]/\mathrm{H}\alpha) + 0.51 \log([\mathrm{S\,II}]/\mathrm{H}\alpha) + 0.59 \log([\mathrm{O\,III}]/\mathrm{H}\beta) \tag*{(46)}
P3=−0.46log⁡([N II]/Hα)−0.37log⁡([S II]/Hα)+0.81log⁡([O III]/Hβ)(47)P_3 = -0.46 \log([\mathrm{N\,II}]/\mathrm{H}\alpha) - 0.37 \log([\mathrm{S\,II}]/\mathrm{H}\alpha) + 0.81 \log([\mathrm{O\,III}]/\mathrm{H}\beta) \tag*{(47)}

This panel contains ≈8%\approx8\% fewer objects than the left-hand panel, due to the additional S/N cut on the [S II] λλ6716\lambda\lambda6716, 31 doublet. In this projection, P1P_1 maps onto a continuous AGN fraction, fAGNf_{\mathrm{AGN}} ([65] (X. Ji & R. Yan 2020)), the fractional contribution of AGN photoionization to the total line emission. At fixed P3P_3, X. Ji & R. Yan (2020) show that contours of constant ionization parameter run nearly parallel to the P1P_1 axis, largely decoupled from variations in gas-phase abundance, with second-order variations due to differences in nitrogen abundance. Together, these diagrams show that the fastspec emission-line measurements recover the well-known structure of the star-forming and AGN populations, making the DR2 catalog a powerful resource for statistical studies of nuclear activity and ionized gas in galaxies.

Planned Improvements

FastSpecFit continues to be actively developed based on feedback from a small but growing community of users both within and outside of the DESI collaboration. We strive to fix reported bugs rapidly and to implement user requests on a somewhat slower, but regular basis. In addition, there is ongoing effort to improve and expand the documentation and to add to the number of user-facing tutorials for working with the code and VACs.

The next major development cycle will focus on making FastSpecFit’s default inputs and assumptions easier for users to modify, building on infrastructure that is already partly in place. FastSpecFit already supports a flexible line list and a fairly flexible kinematic-constraints file (§ 3.2, § 4.1.1). Users can build and load (at runtime) custom, project-specific files to add or remove lines and adjust amplitude constraints without touching the fitting code. D. Scholte et al. [120], for example, extend the default line-list to measure a larger number of electron-temperature-sensitive auroral lines in a large sample of low-redshift star-forming galaxies. The kinematic model choices, however, are more rigid; they currently only support the narrow-only and narrow+broad options documented in § 4.1.1. To address this shortcoming, we plan to generalize the configuration file so that users can assign an arbitrary mixture of Gaussian components to any line, not just the Balmer series. With this change, for example, the [O III] λλ4959,5007\lambda\lambda4959, 5007 doublet and other forbidden lines can be modeled with multiple kinematic components (with flexible constraints between them) rather than a single Gaussian line (e.g., C. M. Harrison et al. [58]; R. Pucha et al. 2025).

Two additional fitting modes (§ 4.4) are also under active development. The first, fastqso, is being designed to better model AGN/QSO spectra (§ 5.3). Instead of a non-negative sum of SPS models, its continuum model will consist of a UV power law, a velocity-broadened Fe II (11001100–31003100 \AA) template, a tunable amount of host-galaxy (stellar) light, and an infrared torus component. Combined with the more flexible emission-line modeling described above, fastqso will yield significantly better rest-frame photometry, emission-line properties, and more. The second fitting mode under development, fastbayes (for “fast Bayesian”), is being scoped as a complementary alternative to fastphot. Instead of the current maximum-likelihood method (§ 3.4.3), fastbayes will interpolate and likelihood-weight a large precomputed grid of SPS models to estimate the posterior probability distribution of stellar mass, SFR, dust attenuation, and other physical properties. Because the grid is precomputed rather than sampled on the fly, this approach will return full posterior distributions in a fraction of the time required by standard Bayesian fitting codes (e.g., B. D. Johnson et al. [66]). We defer a more complete description of these and other developments to forthcoming papers.

Summary

FastSpecFit is a fast, flexible, and physically motivated code for modeling DESI stellar continuum and emission-line spectra jointly with broadband photometry. By convolving stellar-population and emission-line templates with each fiber’s native resolution matrix (§ 3.4 and § 4) rather than resampling the data, FastSpecFit preserves DESI’s uncorrelated pixel statistics while fitting a typical spectrum in ≈0.1\approx0.1 s on a single CPU core. Its non-parametric SFH, energy-conserving dust model, and independently fitted narrow and broad emission-line kinematics let the data—rather than a pre-assumed galaxy type—determine the best-fitting model, so the same machinery applies uniformly across the full diversity of DESI extragalactic targets, from quiescent LRGs to star-forming ELGs to broad-line QSOs (§ 5). Its speed and hybrid MPI/multiprocessing design make it practical to estimate parameter uncertainties from repeated Monte Carlo fits (§ 4.3) and to scale from a single laptop process to thousands of supercomputer cores (§ 6.3).

We use FastSpecFit to produce the DR2 VAC (Loa/v2.0), the primary subject of this paper: approximately 38 million extragalactic targets spanning the DESI commissioning, survey-validation, special-program, and main-survey campaigns (§ 6.1), including corrected redshifts for roughly 165,000 QSOs. We validate the resulting stellar masses, velocity dispersions, and emission-line fluxes against independent codes and catalogs (§ 6.4), and we illustrate the catalog’s scientific reach with the stellar mass–redshift distribution of the three primary galaxy target classes and with optical emission-line diagnostics of nuclear activity (§ 7.1). Alongside the fastspec VAC we also produce a companion, photometry-only fastphot VAC using the same input dataset and template set (Appendix § 10). We describe companion catalogs for the Early Data Release and Data Release 1, produced with earlier template and FastSpecFit versions, in Appendix § 11.

Several improvements are planned for future FastSpecFit releases, including more flexible, multi-component emission-line kinematics and two new fitting modes, fastqso for QSOs and broad-line AGN and fastbayes for fast posterior estimation (§ 7.2). The code and its measurements are already enabling a broad range of extragalactic science using the public EDR and DR1 VACs (§ 1), and a growing number of studies are already drawing on the DR2 catalog described here, including direct-method chemical abundance measurements for star-forming galaxies ([120]) and a search for the most metal-poor star-forming galaxies yet found (J. Moustakas et al., in preparation); we expect this number to grow substantially once the DR2 VAC is publicly released in 2027.

Acknowledgements

JM gratefully acknowledges funding support from the U.S. Department of Energy, Office of Science, Office of High Energy Physics under Award Number DE-SC0020086, and DS acknowledges support from a UKRI Frontier Research Guarantee Grant (PI Cullen; grant reference EP/X021025/1). We also thank Ben Johnson for assistance constructing custom FSPS models, Benjamin J. Weiner for the suggestion of fitting emission-line doublets as an amplitude and a line-ratio, and Ashod Khederlarian for contributing a tutorial to an early version of FastSpecFit.

This research used data obtained with the Dark Energy Spectroscopic Instrument (DESI). DESI construction and operations is managed by the Lawrence Berkeley National Laboratory. This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of High-Energy Physics, under Contract No. DE–AC02–05CH11231, and by the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility under the same contract. Additional support for DESI was provided by the U.S. National Science Foundation (NSF), Division of Astronomical Sciences under Contract No. AST0950945 to the NSF’s National Optical-Infrared Astronomy Research Laboratory; the Science and Technology Facilities Council of the United Kingdom; the Gordon and Betty Moore Foundation; the Heising-Simons Foundation; the French Alternative Energies and Atomic Energy Commission (CEA); the National Council of Humanities, Science and Technology of Mexico (CONAHCYT); the Ministry of Science and Innovation of Spain (MICINN), and by the DESI Member Institutions: https://www.desi.lbl.gov/collaborating-institutions. The DESI collaboration is honored to be permitted to conduct scientific research on I’ologam Du’ag (Kitt Peak), a mountain with particular significance to the Tohono O’odham Nation. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the U.S. National Science Foundation, the U.S. Department of Energy, or any of the listed funding agencies.

The Legacy Surveys consist of three individual and complementary projects: the Dark Energy Camera Legacy Survey (DECaLS; Proposal ID #2014B-0404; PIs: David Schlegel and Arjun Dey), the Beijing-Arizona Sky Survey (BASS; NOAO Prop. ID #2015A-0801; PIs: Zhou Xu and Xiaohui Fan), and the Mayall zz-band Legacy Survey (MzLS; Prop. ID #2016A-0453; PI: Arjun Dey). DECaLS, BASS and MzLS together include data obtained, respectively, at the Blanco telescope, Cerro Tololo Inter-American Observatory, NSF’s NOIRLab; the Bok telescope, Steward Observatory, University of Arizona; and the Mayall telescope, Kitt Peak National Observatory, NOIRLab. Pipeline processing and analyses of the data were supported by NOIRLab and the Lawrence Berkeley National Laboratory (LBNL). The Legacy Surveys project is honored to be permitted to conduct astronomical research on I’oligam Du’ag (Kitt Peak), a mountain with particular significance to the Tohono O’odham Nation.

NOIRLab is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation. LBNL is managed by the Regents of the University of California under contract to the U.S. Department of Energy.

This project used data obtained with the Dark Energy Camera (DECam), which was constructed by the Dark Energy Survey (DES) collaboration. Funding for the DES Projects has been provided by the U.S. Department of Energy, the U.S. National Science Foundation, the Ministry of Science and Education of Spain, the Science and Technology Facilities Council of the United Kingdom, the Higher Education Funding Council for England, the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign, the Kavli Institute of Cosmological Physics at the University of Chicago, Center for Cosmology and Astro-Particle Physics at the Ohio State University, the Mitchell Institute for Fundamental Physics and Astronomy at Texas A&M University, Financiadora de Estudos e Projetos, Fundacao Carlos Chagas Filho de Amparo, Financiadora de Estudos e Projetos, Fundacao Carlos Chagas Filho de Amparo a Pesquisa do Estado do Rio de Janeiro, Conselho Nacional de Desenvolvimento Cientifico e Tecnologico and the Ministerio da Ciencia, Tecnologia e Inovacao, the Deutsche Forschungsgemeinschaft and the Collaborating Institutions in the Dark Energy Survey. The Collaborating Institutions are Argonne National Laboratory, the University of California at Santa Cruz, the University of Cambridge, Centro de Investigaciones Energeticas, Medioambientales y Tecnologicas-Madrid, the University of Chicago, University College London, the DES-Brazil Consortium, the University of Edinburgh, the Eidgenossische Technische Hochschule (ETH) Zurich, Fermi National Accelerator Laboratory, the University of Illinois at Urbana-Champaign, the Institut de Sciences de l’Espace (IEEC/CSIC), the Institut de Fisica d’Altes Energies, Lawrence Berkeley National Laboratory, the Ludwig Maximilians Universitat Munchen and the associated Excellence Cluster Universe, the University of Michigan, NSF’s NOIRLab, the University of Nottingham, the Ohio State University, the University of Pennsylvania, the University of Portsmouth, SLAC National Accelerator Laboratory, Stanford University, the University of Sussex, and Texas A&M University.

BASS is a key project of the Telescope Access Program (TAP), which has been funded by the National Astronomical Observatories of China, the Chinese Academy of Sciences (the Strategic Priority Research Program “The Emergence of Cosmological Structures” Grant # XDB09000000), and the Special Fund for Astronomy from the Ministry of Finance. The BASS is also supported by the External Cooperation Program of Chinese Academy of Sciences (Grant # 114A11KYSB20160057), and Chinese National Natural Science Foundation (Grant # 12120101003, # 11433005).

The Legacy Survey team makes use of data products from the Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE), which is a project of the Jet Propulsion Laboratory/California Institute of Technology. NEOWISE is funded by the National Aeronautics and Space Administration.

The Legacy Surveys imaging of the DESI footprint is supported by the Director, Office of Science, Office of High Energy Physics of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231, by the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility under the same contract; and by the U.S. National Science Foundation, Division of Astronomical Sciences under Contract No. AST-0950945 to NOAO.

Portions of the FastSpecFit code base, beginning with version 3.3.0, were developed with the assistance of Anthropic’s Claude (via Claude Code, models Sonnet 4.6 and 5). This assistance included documentation, unit and integration tests, packaging and automated testing and release workflows, parallel-processing and production scripts, and parts of the continuum and emission-line fitting code centered around optimization and speed. All AI-generated code was reviewed, tested, and validated by the authors, who take full responsibility for its correctness. In addition, Claude Code (models Sonnet 5 and Opus 5.5) was used in the preparation of this manuscript to generate outlines; translate the technical implementation in the code into equations; develop and refactor the figure-generation and analysis scripts and tables (all using author-provided data); suggest candidate references (all validated by the authors); check the manuscript for internal consistency (including quoted numerical values, notation, and cross-references); suggest revisions to the text for clarity and conciseness; and reformat the manuscript source for submission. The authors reviewed and revised all AI-generated text and take full responsibility for the content of this paper.

Software: Astropy ([7]), corner.py ([47]), FastSpecFit ([92]), healpy ([144]), mpi4py ([33]), NumPy ([57]), Numba ((S. K. [76], Matplotlib [62], SciPy [135], seaborn [137], speclite [71]

DR2 fastphot VAC

This appendix briefly documents the fastphot Loa/v2.0 VAC, a closely related but distinct catalog to the fastspec VAC introduced in § 6. As described in § 4.4, fastphot is an alternative fitting mode available within the larger FastSpecFit software package which models just broadband photometry, but at the fixed DESI spectroscopic redshift. The VAC has been generated using the identical input DESI dataset (Loa specprod), FastSpecFit software tag (3.6.1), and template set (ftemplates-chabrier-2.2.0.fits); the sample selection, corrected QSO redshifts, and final object counts are identical to those given in Table 5.

Because fastphot fits only broadband photometry, its per-HEALPix files contain three FITS extensions rather than fastspec’s five—PRIMARY, METADATA, and SPECPHOT—mirroring the corresponding fastspec extensions but omitting the emission-line FASTSPEC table and the MODELS best-fitting-spectra extension entirely, since there is no spectrum to model or emission lines to fit (§ 6.1). Its per-HEALPix and merged catalogs, on the other hand, follow the same SURVEY/PROGRAM/HEALPIX naming convention as fastspec, but with a “fastphot” rather than “fastspec” prefix. For a detailed column-by-column description of the fastphot data model, pinned to the FastSpecFit version used to produce this VAC, please see the online documentation.32

Besides being substantially faster to produce (median runtime is 0.42 s/object/core, an order of magnitude faster than fastspec; § 6.3), fastphot offers a purely photometric alternative for science cases that do not require the DESI spectrum itself. For example, its K-corrections and global physical properties, including the inferred stellar mass, come from the integrated broadband photometry, without the simplistic and uncertain aperture-correction fastspec uses (§ 3.4.2).

To explore just one example, in Figure 13 we compare the stellar masses we derive using fastspec and fastphot split by DESI target class. Overall, the agreement is very good, with median offsets and NMAD scatter of −0.056±0.091-0.056 \pm0.091 dex, −0.012±0.088-0.012 \pm0.088 dex, and −0.019±0.324-0.019 \pm0.324 dex for BGS, LRG, and ELG targets, respectively. Examining the residuals with redshift, BGS targets at z≲0.4z \lesssim0.4 exhibit the most pronounced differences in exactly the sense that you might expect from aperture effects. The systematically higher fastspec stellar masses are possibly due to the larger mass-to-light ratio (older stellar ages) exhibited in the centers of galaxies and traced by the central DESI spectrum. Even so, the median offset does not exceed ∣Δmed∣≈0.05|\Delta_{\mathrm{med}}| \approx0.05 dex for any of the three target classes, underscoring the overall agreement between the two fitting modes.

Figure 13. Comparison of FastSpecFit stellar masses derived from spectrophotometric fitting (fastspec) and from photometry-only fitting (fastphot), organized by target class. Top row: fastphot log⁡10(M⋆h−2/M⊙)\log_{10}(M_{\star}h^{-2}/M_{\odot}) versus fastspec log⁡10(M⋆h−2/M⊙)\log_{10}(M_{\star}h^{-2}/M_{\odot}) (h=1h = 1, Chabrier IMF) for BGS (left), LRG (center), and ELG (right) subsamples. As in Figure 8, the color map, contours, and outlying markers show the two-dimensional density distribution, and the dashed line is the one-to-one relation; each panel lists the matched sample size NN, the median mass offset Δmed\Delta_{\mathrm{med}} (fastphot minus fastspec) in dex, and the NMAD scatter. Bottom row: Residual Δlog⁡10(M⋆h−2/M⊙)\Delta\log_{10}(M_{\star}h^{-2}/M_{\odot}) (fastphot minus fastspec) as a function of redshift for the same three subsamples. Black points mark the running median in bins of redshift, with error bars enclosing the interquartile range. The dashed horizontal line marks zero offset. We attribute the systematic offset seen in BGS targets at low redshift to imperfect aperture corrections for the most extended, nearby galaxies, whose half-light radii subtend a significant fraction of the DESI fiber diameter.

Value-Added Catalog Revision History

FastSpecFit has been in near-continuous development since late-2020 (not coincidentally, the beginning of DESI on-sky observations), and so the code base and its template library have evolved progressively and, in some cases, substantially over time. Concurrently, each new DESI specprod has incorporated both minor and major improvements in the data-reduction pipeline, as well as adding spectra for millions of new objects. In order to meet the significant demand for astrophysical measurements of DESI targets beyond just their redshift—both within and outside of the DESI collaboration—FastSpecFit has been on a fairly rapid development cycle, resulting in 12 unique fastspec VACs and four unique fastphot VACs (including the ones documented in this paper) since 2022 August. In § 1 we cite just a small fraction of the publications which have resulted from these datasets.

a^{a}FastSpecFit software tag(s) used to produce the VAC.

b^{b}SPS template library, ftemplates-{name}.fits unless otherwise noted (§ 3.1).

c^{c}To be released in 2027 as part of DESI/DR2.

**Note—**No fastphot VAC was produced for the Fuji (EDR) specprod. Release notes, known issues, and the full change history for every version are maintained in the FastSpecFit online change log at https://github.com/desihub/fastspecfit/blob/main/doc/changes.rst.

In Table 6 we document this history by listing the current set of fastspec and fastphot VACs, the date they were released, and the templates and version of FastSpecFit which were used, organized by DESI specprod and data release. For additional details regarding each of these releases, especially substantive algorithmic changes, major bug fixes, and much more, we refer to the detailed FastSpecFit change log and release notes.33

VAC VersionRelease DateSoftware Tag(s)aTemplatesb
Loa (DR2) — fastspec
v2.02027 Aprc3.6.1chabrier-2.2.0
v1.02027 Aprc3.1.4chabrier-2.0.0
Loa (DR2) — fastphot
v2.02027 Aprc3.6.1chabrier-2.2.0
v1.02027 Aprc3.2.0chabrier-2.0.0
Iron (DR1) — fastspec
v4.02026 Dec3.6.1chabrier-2.2.0
v3.02025 Jun3.1.5chabrier-2.0.0
v2.12024 Jan2.5.0chabrier-1.3.0
v2.02023 Aug2.4.1, 2.4.2chabrier-1.1.0
v1.02023 Feb2.1.0, 2.1.1chabrier-1.0.0
Iron (DR1) — fastphot
v2.02026 Dec3.6.1chabrier-2.2.0
v1.02026 Jun3.2.0chabrier-2.0.0
Fuji (EDR) — fastspec
v3.22024 Jan2.5.0chabrier-1.3.0
v3.12023 Dec2.4.3chabrier-1.2.0
v3.02023 Aug2.4.1, 2.4.2chabrier-1.1.0
v2.02023 Feb2.1.0, 2.1.1chabrier-1.0.0
v1.02022 Aug1.0.0, 1.0.1SSP-CKC14z v1.0
VAC VersionRelease DateSoftware Tag(s)a^{a}Templatesb^{b}

Table 6. FastSpecFit VAC Revision History. Table 6 continued on next page

In the remainder of this appendix, we quantify the statistical differences across a subset of these releases for a handful of key quantities measured by FastSpecFit. We match Loa/v2.0 to the three earlier fastspec VACs which have been public the longest, Iron/v3.0, Iron/v2.1, and Fuji/v3.2. We omit Iron/v4.0, which we produced with the same code and templates as Loa/v2.0, as well as the earlier Iron and Fuji VACs, which were only available internally to DESI collaboration members before being superseded. We also restrict the comparison to SV3, whose targets appear in every release and whose coadded spectra include the same exposures in each specprod. After matching on TARGETID, we require each object to satisfy the good-redshift criteria for its target class (§ 6.2) in both catalogs. Among these objects, fewer than 0.2% of the BGS, LRG, and ELG targets have redshifts which differ by more than 1000 km s−1s^{-1} between releases. For QSO targets, however, this fraction rises to 12% relative to Iron and 19% relative to Fuji, which we attribute to changes in both the QuasarNet classifications and Redrock redshifts between specprods and in how FastSpecFit updates its QSO redshifts (§ 6.1). We exclude objects whose redshift changed from the following comparisons.

In Table 7 we list the median difference and NMAD of several key quantities for each target class. Relative to Iron/v3.0, the median offsets are smaller than 0.01 dex (or 0.01 mag) for every quantity except the SFR. For BGS and LRG targets, the scatter in stellar mass, absolute magnitude, velocity dispersion, and emission-line flux is also below 0.01 dex (or 0.01 mag). The small positive offset in σ⋆\sigma_\star (0.004–0.007 dex)

QuantityClassIron/v3.0 Δ\DeltaIron/v3.0 σ\sigmaIron/v2.1 Δ\DeltaIron/v2.1 σ\sigmaFuji/v3.2 Δ\DeltaFuji/v3.2 σ\sigma
Δlog⁡10M⋆\Delta\log_{10} \mathcal{M}_{\star}BGS−0.002-0.0020.0080.008+0.009+0.0090.0310.031+0.003+0.0030.0440.044
LRG0.0000.0000.0090.009+0.004+0.0040.0470.047−0.003-0.0030.0590.059
ELG+0.003+0.0030.0650.065+0.012+0.0120.2140.214−0.003-0.0030.2790.279
Δlog⁡10SFRa\Delta\log_{10} \mathrm{SFR}^{a}BGS+1.243+1.2430.8410.841+0.618+0.6180.5930.593+0.609+0.6090.5940.594
LRG+0.838+0.8381.9051.905−0.180-0.1801.2641.264−0.250-0.2501.4211.421
ELG+0.578+0.5780.5500.550+0.262+0.2620.5070.507+0.252+0.2520.5310.531
ΔDn(4000)\Delta D_n(4000)BGS+0.002+0.0020.0230.023+0.004+0.0040.0280.028+0.003+0.0030.0480.048
LRG0.0000.0000.0280.028+0.002+0.0020.0360.036+0.001+0.0010.0600.060
ELG+0.004+0.0040.0480.048+0.004+0.0040.0520.052+0.004+0.0040.0800.080
Δ 0.1Mr\Delta\,{}^{0.1}M_rBGS+0.001+0.0010.0020.002−0.001-0.0010.0050.005−0.001-0.0010.0080.008
LRG0.0000.0000.0010.001−0.001-0.0010.0050.005−0.001-0.0010.0070.007
ELG+0.002+0.0020.0190.019−0.005-0.0050.0430.043−0.005-0.0050.0550.055
Δlog⁡10σ⋆\Delta\log_{10} \sigma_{\star}BGS+0.007+0.0070.0080.008+0.001+0.0010.0520.052+0.001+0.0010.0530.053
LRG+0.004+0.0040.0070.007−0.003-0.0030.0520.052−0.002-0.0020.0540.054
Δlog⁡10F(Hα)\Delta\log_{10} F(\mathrm{H}\alpha)BGS+0.006+0.0060.0060.006+0.011+0.0110.0230.023+0.009+0.0090.0230.023
Δlog⁡10F([O III] λ5007)\Delta\log_{10} F([\mathrm{O\,III}]\,\lambda5007)BGS+0.001+0.0010.0090.009+0.004+0.0040.0320.032+0.004+0.0040.0340.034
ELG+0.001+0.0010.0050.005+0.004+0.0040.0210.021+0.002+0.0020.0230.023
Δlog⁡10F([O II] λλ3726,29)\Delta\log_{10} F([\mathrm{O\,II}]\,\lambda\lambda3726,29)ELG−0.002-0.0020.0040.004−0.001-0.0010.0210.021−0.002-0.0020.0230.023

Table 7. Comparison of Loa/v2.0 with Earlier fastspec VACs

**Note—**Median difference (Δ\Delta; Loa/v2.0 minus the earlier VAC) and normalized median absolute deviation (σ\sigma) for SV3 targets with a good redshift (§ 6.2) in both catalogs and redshifts that agree to within 1000 km s−1^{-1}. Line fluxes and σ⋆\sigma_{\star} require flux S/N >3> 3 in both catalogs.

is consistent with the correction for the intrinsic resolution of the C3K templates (equation (4)), which we introduced in FastSpecFit 3.5.0. Relative to Iron/v2.1 and Fuji/v3.2, both produced with FastSpecFit 2.5.0, the scatter is larger (typically 0.02–0.06 dex), but the median offsets remain below 0.015 dex. The notable exception is the stellar mass of ELGs, which scatters by 0.21–0.28 dex relative to these two catalogs with no significant median offset. We attribute this scatter to the substantial changes in our SPS templates since 2.5.0 (Table 6), combined with the low S/N of most ELG continua. Moreover, because Iron/v2.1 and Fuji/v3.2 share the same code and templates, the difference between their two columns isolates the effect of the data-reduction changes from Fuji to Iron, which increase the scatter only modestly.

Finally, the SFRs in earlier VACs are not directly comparable to those in Loa/v2.0. Starting with FastSpecFit 3.5.0, we report the SFR averaged over the most recent 100 Myr (equation (34)). Earlier versions instead reported the instantaneous SFR of the best-fitting model, to which only the youngest (0–30 Myr) age bin contributes. We attribute the large SFR differences in Table 7 (median offsets up to 1.2 dex and NMAD values of 0.5–1.9 dex) primarily to this change in definition, and we recommend against combining SFRs from Loa/v2.0 with those from any earlier VAC.

References

  1. [1]Abdul Karim, M., Aguilar, J., Ahlen, S., et al. 2025a, PhRvD, 112, 083514. https://arxiv.org/abs/2503.14739
  2. [3]Adame, A. G., Aguilar, J., Ahlen, S., et al. 2025a, JCAP, 2025, 008. https://arxiv.org/abs/2411.12021
  3. [5]Adame, A. G., Aguilar, J., Ahlen, S., et al. 2025c, JCAP, 2025, 124. https://arxiv.org/abs/2404.03001
  4. [6]Anand, A., Guy, J., Bailey, S., et al. 2024, AJ, 168, 124. https://arxiv.org/abs/2405.19288
  5. [7]Astropy Collaboration, Price-Whelan, A. M., Lim, P. L., et al. 2022, ApJ, 935, 167. https://arxiv.org/abs/2206.14220
  6. [8]Baldwin, J. A., Phillips, M. M., & Terlevich, R. 1981, PASP, 93, 5
  7. [9]Balogh, M. L., Morris, S. L., Yee, H. K. C., Carlberg, R. G., & Ellingson, E. 1999, ApJ, 527, 54. https://arxiv.org/abs/astro-ph/9906470
  8. [10]Bezanson, R., van Dokkum, P. G., Franx, M., et al. 2011, ApJL, 737, L31. https://arxiv.org/abs/1107.0972
  9. [11]Blanco, D., Leauthaud, A., Ulf Lange, J., et al. 2025, arXiv e-prints, arXiv:2512.15964. https://arxiv.org/abs/2512.15964
  10. [12]Blanton, M. R., & Moustakas, J. 2009, ARA&A, 47, 159. https://arxiv.org/abs/0908.3017
  11. [13]Blanton, M. R., & Roweis, S. 2007, AJ, 133, 734. https://arxiv.org/abs/astro-ph/0606170
  12. [14]Blanton, M. R., Brinkmann, J., Csabai, I., et al. 2003, AJ, 125, 2348
  13. [15]Bolton, A. S., & Schlegel, D. J. 2010, PASP, 122, 248. https://arxiv.org/abs/0911.2689
  14. [16]Boquien, M., Burgarella, D., Roehlly, Y., et al. 2019, A&A, 622, A103. https://arxiv.org/abs/1811.03094
  15. [17]Branch, M. A., Coleman, T. F., & Li, Y. 1999, SIAM Journal on Scientific Computing, 21, 1
  16. [18]Busca, N., & Balland, C. 2018, arXiv e-prints, arXiv:1808.09955. https://arxiv.org/abs/1808.09955
  17. [19]Byler, N., Dalcanton, J. J., Conroy, C., & Johnson, B. D. 2017, ApJ, 840, 44. https://arxiv.org/abs/1611.08305
  18. [20]Cappellari, M. 2023, MNRAS, 526, 3273. https://arxiv.org/abs/2208.14974
  19. [21]Cappellari, M., & Emsellem, E. 2004, PASP, 116, 138
  20. [22]Carnall, A. C., McLure, R. J., Dunlop, J. S., & Davé, R. 2018, MNRAS, 480, 4379. https://arxiv.org/abs/1712.04452
  21. [23]Chabrier, G. 2003, PASP, 115, 763. https://arxiv.org/abs/astro-ph/0304382
  22. [24]Charlot, S., & Fall, S. M. 2000, ApJ, 539, 718
  23. [25]Chaussidon, E., Yèche, C., Palanque-Delabrouille, N., et al. 2023, ApJ, 944, 107. https://arxiv.org/abs/2208.08511
  24. [26]Choi, J., Dotter, A., Conroy, C., et al. 2016, ArXiv e-prints. https://arxiv.org/abs/1604.08592
  25. [27]Cid Fernandes, R., Mateus, A., Sodré, L., Stasińska, G., & Gomes, J. M. 2005, MNRAS, 358, 363. https://arxiv.org/abs/astro-ph/0412481
  26. [28]Conroy, C. 2013, ARA&A, 51, 393. https://arxiv.org/abs/1301.7095 Conroy, C., Graves, G. J., & van Dokkum, P. G. 2014, ApJ, 780, 33. https://arxiv.org/abs/1303.6629
  27. [29]Conroy, C., & Gunn, J. E. 2010, ApJ, 712, 833. https://arxiv.org/abs/0911.3151
  28. [30]Conroy, C., Gunn, J. E., & White, M. 2009, ApJ, 699, 486. https://arxiv.org/abs/0809.4261
  29. [31]Conroy, C., Villaume, A., van Dokkum, P. G., & Lind, K. 2018, ApJ, 854, 139. https://arxiv.org/abs/1801.10185
  30. [32]da Cunha, E., Charlot, S., & Elbaz, D. 2008, MNRAS, 388, 1595. https://arxiv.org/abs/0806.1020
  31. [33]Dalcin, L., & Fang, Y.-L. L. 2021, Computing in Science & Engineering, 23, 47
  32. [34]DESI Collaboration, Aghamousa, A., Aguilar, J., et al. 2016a, ArXiv e-prints. https://arxiv.org/abs/1611.00036
  33. [35]DESI Collaboration, Aghamousa, A., Aguilar, J., et al. 2016b, ArXiv e-prints. https://arxiv.org/abs/1611.00037
  34. [36]DESI Collaboration, Abareshi, B., Aguilar, J., et al. 2022, AJ, 164, 207. https://arxiv.org/abs/2205.10939
  35. [37]DESI Collaboration, Adame, A. G., Aguilar, J., et al. 2024a, AJ, 167, 62. https://arxiv.org/abs/2306.06307
  36. [38]DESI Collaboration, Adame, A. G., Aguilar, J., et al. 2024b, AJ, 168, 58. https://arxiv.org/abs/2306.06308
  37. [39]DESI Collaboration, Abdul Karim, M., Adame, A. G., et al. 2026, AJ, 171, 285. https://arxiv.org/abs/2503.14745
  38. [40]Dey, A., Rabinowitz, D., Karcher, A., et al. 2016, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 9908, Ground-based and Airborne Instrumentation for Astronomy VI, ed. C. J. Evans, L. Simard, & H. Takami, 99082C
  39. [41]Dey, A., Schlegel, D. J., Lang, D., et al. 2019, AJ, 157, 168. https://arxiv.org/abs/1804.08657
  40. [42]Dey, A., Koposov, S. E., Najita, J. R., et al. 2026, AJ, 171, 348. https://arxiv.org/abs/2505.17230
  41. [43]Dotter, A. 2016, ApJS, 222, 8. https://arxiv.org/abs/1601.05144
  42. [44]Draine, B. T., Dale, D. A., Bendo, G., et al. 2007, ApJ, 663, 866. https://arxiv.org/abs/astro-ph/0703213
  43. [45]Fitzpatrick, E. L. 1999, PASP, 111, 63. https://arxiv.org/abs/astro-ph/9809387
  44. [46]Flaugher, B., Diehl, H. T., Honscheid, K., et al. 2015, AJ, 150, 150. https://arxiv.org/abs/1504.02900
  45. [47]Foreman-Mackey, D. 2016, The Journal of Open Source Software, 1, 24
  46. [48]Fritz, J., Franceschini, A., & Hatziminaoglou, E. 2006, MNRAS, 366, 767. https://arxiv.org/abs/astro-ph/0511428
  47. [49]Golub, G. H., & Pereyra, V. 1973, SIAM Journal on Numerical Analysis, 10, 413
  48. [50]Green, D., Kirkby, D., Aguilar, J., et al. 2025, JCAP, 2025, 087. https://arxiv.org/abs/2505.01596
  49. [51]Guo, H., Shen, Y., & Wang, S. 2018, PyQSOFit: Python code to fit the spectrum of quasars,, Astrophysics Source Code Library, record ascl:1809.008 http://ascl.net/1809.008
  50. [52]Guo, W.-J., Zou, H., Greenwell, C. L., et al. 2025, ApJS, 278, 28. https://arxiv.org/abs/2408.00402
  51. [53]Guy, J., Bailey, S., Kremin, A., et al. 2023, AJ, 165, 144. https://arxiv.org/abs/2209.14482
  52. [54]Hahn, C., Wilson, M. J., Ruiz-Macias, O., et al. 2023, AJ, 165, 253. https://arxiv.org/abs/2208.08512
  53. [55]Hahn, C., Aguilar, J. N., Alam, S., et al. 2024, ApJ, 963, 56
  54. [56]Hao, C.-N., Kennicutt, R. C., Johnson, B. D., et al. 2011, ApJ, 741, 124. https://arxiv.org/abs/1108.2837
  55. [57]Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357. https://arxiv.org/abs/2006.10256
  56. [58]Harrison, C. M., Alexander, D. M., Mullaney, J. R., & Swinbank, A. M. 2014, MNRAS, 441, 3306. https://arxiv.org/abs/1403.3086
  57. [59]Hogg, D. W. 1999, ArXiv Astrophysics e-prints
  58. [60]Hogg, D. W., Baldry, I. K., Blanton, M. R., & Eisenstein, D. J. 2002, arXiv e-prints, astro. https://arxiv.org/abs/astro-ph/0210394
  59. [61]Huang, X., Inchausti, J. C., Storfer, C. J., et al. 2025, arXiv e-prints, arXiv:2509.18089. https://arxiv.org/abs/2509.18089
  60. [62]Hunter, J. D. 2007, Computing in Science and Engineering, 9, 90
  61. [63]Inoue, A. K., Shimizu, I., Iwata, I., & Tanaka, M. 2014, MNRAS, 442, 1805. https://arxiv.org/abs/1402.0677
  62. [65]Jin, G., Dai, Y. S., Pan, H.-A., et al. 2021, ApJ, 923, 6. https://arxiv.org/abs/2109.11084
  63. [66]Johnson, B. D., Leja, J., Conroy, C., & Speagle, J. S. 2021, ApJS, 254, 22. https://arxiv.org/abs/2012.01426
  64. [67]Juneau, S., Canning, R., Alexander, D. M., et al. 2025, AJ, 169, 157. https://arxiv.org/abs/2404.03621
  65. [68]Kauffmann, G., Heckman, T. M., Tremonti, C., et al. 2003, MNRAS, 346, 1055
  66. [69]Kennicutt, Jr., R. C. 1998, ARA&A, 36, 189
  67. [70]Kewley, L. J., & Dopita, M. A. 2002, ApJS, 142, 35
  68. [71]Kirkby, D., Robitaille, T., Weaver, B. A., et al. 2024, desihub/speclite: Bug fix release: General clean-up prior to refactoring package infrastructure, v0.20 Zenodo
  69. [72]Koleva, M., Prugniel, P., Bouchard, A., & Wu, Y. 2009, A&A, 501, 1269. https://arxiv.org/abs/0903.2979
  70. [73]Koposov, S., Li, T. S., Allende Prieto, C., et al. 2026, The Open Journal of Astrophysics, 9, 55260. https://arxiv.org/abs/2505.14787
  71. [74]Koposov, S. E. 2019, RVSpecFit: Radial velocity and stellar atmospheric parameter fitting,, Astrophysics Source Code Library, record ascl:1907.013 http://ascl.net/1907.013
  72. [75]Kormendy, J., & Ho, L. C. 2013, arXiv e-prints. https://arxiv.org/abs/1308.6483
  73. [76]Lam, S. K., Pitrou, A., & Seibert, S. 2015, in Proc. Second Workshop on the LLVM Compiler Infrastructure in HPC, 1–6
  74. [77]Lang, D. 2014, AJ, 147, 108. https://arxiv.org/abs/1405.0308
  75. [78]Lang, D., Hogg, D. W., & Mykytyn, D. 2016a, The Tractor: Probabilistic astronomical source detection and measurement,, Astrophysics Source Code Library, record ascl:1604.008
  76. [80]Leja, J., Carnall, A. C., Johnson, B. D., Conroy, C., & Speagle, J. S. 2019a, ApJ, 876, 3. https://arxiv.org/abs/1811.03637
  77. [81]Leja, J., Johnson, B. D., Conroy, C., van Dokkum, P. G., & Byler, N. 2017, ApJ, 837, 170. https://arxiv.org/abs/1609.09073
  78. [82]Leja, J., Speagle, J. S., Johnson, B. D., et al. 2020, ApJ, 893, 111. https://arxiv.org/abs/1910.04168
  79. [83]Leja, J., Johnson, B. D., Conroy, C., et al. 2019b, ApJ, 877, 140. https://arxiv.org/abs/1812.05608
  80. [84]Levi, M., Bebek, C., Beers, T., et al. 2013, arXiv e-prints, arXiv:1308.0847. https://arxiv.org/abs/1308.0847
  81. [85]Lower, S., Narayanan, D., Leja, J., et al. 2020, ApJ, 904, 33. https://arxiv.org/abs/2006.03599
  82. [86]Luridiana, V., Morisset, C., & Shaw, R. A. 2015, A&A, 573, A42. https://arxiv.org/abs/1410.6662
  83. [87]Mainzer, A., Bauer, J., Cutri, R. M., et al. 2014, ApJ, 792, 30. https://arxiv.org/abs/1406.6025
  84. [88]Martin, D. C., Fanson, J., Schiminovich, D., et al. 2005, ApJL, 619, L1
  85. [89]Meisner, A. M., Lang, D., Schlafly, E. F., & Schlegel, D. J. 2021, Research Notes of the American Astronomical Society, 5, 168
  86. [90]Miller, T. N., Doel, P., Gutierrez, G., et al. 2024, AJ, 168, 95. https://arxiv.org/abs/2306.06310
  87. [91]Moore, S. G., Cole, S., Wilson, M., et al. 2026, MNRAS, 549, stag866. https://arxiv.org/abs/2511.01803
  88. [92]Moustakas, J., Buhler, J., Scholte, D., Dey, B., & Khederlarian, A. 2023, FastSpecFit: Fast spectral synthesis and emission-line fitting of DESI spectra,, Astrophysics Source Code Library, record ascl:2308.005 http://ascl.net/2308.005
  89. [93]Moustakas, J., Kennicutt, Jr., R. C., & Tremonti, C. A. 2006, ApJ, 642, 775
  90. [94]Moustakas, J., Coil, A. L., Aird, J., et al. 2013, ApJ, 767, 50. https://arxiv.org/abs/1301.1688
  91. [95]Myers, A. D., Moustakas, J., Bailey, S., et al. 2023, AJ, 165, 50. https://arxiv.org/abs/2208.08518
  92. [96]Nenkova, M., Sirocky, M. M., Ivezić, Ž., & Elitzur, M. 2008a, ApJ, 685, 147. https://arxiv.org/abs/0806.0511
  93. [98]Osterbrock, D. E. 1989, Astrophysics of gaseous nebulae and active galactic nuclei (Mill Valley, CA: University Science Books)DOI
  94. [99]Pernet, E., Boecker, A., & Martín-Navarro, I. 2024, A&A, 687, L14. https://arxiv.org/abs/2406.04394
  95. [100]Planck Collaboration, Aghanim, N., Akrami, Y., et al. 2020, A&A, 641, A6. https://arxiv.org/abs/1807.06209
  96. [101]Poppett, C., Tyas, L., Aguilar, J., et al. 2024, AJ, 168, 245
  97. [103]Raichoor, A., Moustakas, J., Newman, J. A., et al. 2023, AJ, 165, 126. https://arxiv.org/abs/2208.08513
  98. [104]Renzini, A. 2006, ARA&A, 44, 141. https://arxiv.org/abs/astro-ph/0603479
  99. [105]Rincon, H., Benzvi, S., Douglass, K., et al. 2025, ApJ, 982, 38. https://arxiv.org/abs/2411.00148
  100. [106]Ross, C. E., Howlett, C., Lucey, J. R., et al. 2025, arXiv e-prints, arXiv:2512.03226. https://arxiv.org/abs/2512.03226
  101. [107]Said, K., Howlett, C., Davis, T., et al. 2025, MNRAS, 539, 3627. https://arxiv.org/abs/2408.13842
  102. [108]Salim, S., Boquien, M., & Lee, J. C. 2018, ApJ, 859, 11. https://arxiv.org/abs/1804.05850
  103. [109]Salim, S., Lee, J. C., Janowiecki, S., et al. 2016, ApJS, 227, 2. https://arxiv.org/abs/1610.00712
  104. [110]Sánchez, S. F. 2020, ARA&A, 58, 99. https://arxiv.org/abs/1911.06925
  105. [111]Sánchez, S. F., Pérez, E., Sánchez-Blázquez, P., et al. 2016, RMxAA, 52, 21. https://arxiv.org/abs/1509.08552
  106. [112]Sánchez, S. F., Barrera-Ballesteros, J. K., Lacerda, E., et al. 2022, ApJS, 262, 36. https://arxiv.org/abs/2206.07062
  107. [113]Sánchez-Blázquez, P., Peletier, R. F., Jiménez-Vicente, J., et al. 2006, MNRAS, 371, 703. https://arxiv.org/abs/astro-ph/0607009
  108. [114]Sarzi, M., Falcón-Barroso, J., Davies, R. L., et al. 2006, MNRAS, 366, 1151. https://arxiv.org/abs/astro-ph/0511307
  109. [115]Sarzi, M., Falcón-Barroso, J., Davies, R. L., et al. 2017, GANDALF: Gas AND Absorption Line Fitting,, Astrophysics Source Code Library, record ascl:1708.012 http://ascl.net/1708.012
  110. [116]Schlafly, E. F., & Finkbeiner, D. P. 2011, ApJ, 737, 103. https://arxiv.org/abs/1012.4804
  111. [117]Schlafly, E. F., Kirkby, D., Schlegel, D. J., et al. 2023, AJ, 166, 259, doi: 10.3847/1538-3881/ad0832DOI
  112. [118]Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, ApJ, 500, 525
  113. [119]Scholte, D., Saintonge, A., Moustakas, J., et al. 2024, MNRAS, 535, 2341. https://arxiv.org/abs/2408.03996
  114. [120]Scholte, D., Cullen, F., Moustakas, J., et al. 2026, MNRAS, 551, stag1381, doi: 10.1093/mnras/stag1381
  115. [121]Sérsic, J. L. 1968, Atlas de Galaxias Australes (Cordoba, Argentina: Observatorio Astronomico)
  116. [122]Shen, Y., Brandt, W. N., Richards, G. T., et al. 2016a, ApJ, 831, 7. https://arxiv.org/abs/1602.03894
  117. [123]Siegel, J., McCullough, J., Amon, A., et al. 2025, arXiv e-prints, arXiv:2507.11530. https://arxiv.org/abs/2507.11530
  118. [124]Silber, J. H., Fagrelius, P., Fanning, K., et al. 2023, AJ, 165, 9. https://arxiv.org/abs/2205.09014
  119. [125]Siudek, M., Pucha, R., Mezcua, M., et al. 2024, A&A, 691, A308. https://arxiv.org/abs/2409.19066
  120. [126]Siudek, M., Mezcua, M., Circosta, C., et al. 2025, A&A, 700, A209. https://arxiv.org/abs/2506.09143
  121. [127]Soto, D., & French, K. D. 2025, Research Notes of the American Astronomical Society, 9, 165
  122. [128]Sui, J., Zou, H., Scholte, D., et al. 2026, AJ, 171, 351. https://arxiv.org/abs/2603.05934
  123. [129]Taylor, L., Bezanson, R., van der Wel, A., et al. 2022, ApJ, 939, 90. https://arxiv.org/abs/2210.00144
  124. [130]Thomas, D., Steele, O., Maraston, C., et al. 2013, MNRAS, 431, 1383. https://arxiv.org/abs/1207.6115
  125. [131]Valdes, F., Gupta, R., Rose, J. A., Singh, H. P., & Bell, D. J. 2004, ApJS, 152, 251. https://arxiv.org/abs/astro-ph/0402435
  126. [132]Vanden Berk, D. E., Richards, G. T., Bauer, A., et al. 2001, AJ, 122, 549. https://arxiv.org/abs/astro-ph/0105231
  127. [133]Vegetti, S., Birrer, S., Despali, G., et al. 2024, SSRv, 220, 58. https://arxiv.org/abs/2306.11781
  128. [134]Vestergaard, M., & Wilkes, B. J. 2001, ApJS, 134, 1. https://arxiv.org/abs/astro-ph/0104320
  129. [135]Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, Nature Methods, 17, 261. https://arxiv.org/abs/1907.10121
  130. [136]Wang, W., Yang, X., Jing, Y., et al. 2025, ApJ, 986, 218. https://arxiv.org/abs/2503.03317
  131. [137]Waskom, M. 2021, The Journal of Open Source Software, 6, 3021
  132. [138]Wilkinson, D. M., Maraston, C., Goddard, D., Thomas, D., & Parikh, T. 2017, MNRAS, 472, 4297. https://arxiv.org/abs/1711.00865
  133. [139]Williams, G. G., Olszewski, E., Lesser, M. P., & Burge, J. H. 2004, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 5492, Ground-based Instrumentation for Astronomy, ed. A. F. M. Moorwood & M. Iye, 787–798
  134. [140]Wright, E. L., Eisenhardt, P. R. M., Mainzer, A. K., et al. 2010, AJ, 140, 1868. https://arxiv.org/abs/1008.0031
  135. [141]York, D. G., Adelman, J., Anderson, Jr., J. E., et al. 2000, AJ, 120, 1579
  136. [142]Zhou, R., Dey, B., Newman, J. A., et al. 2023, AJ, 165, 58. https://arxiv.org/abs/2208.08515
  137. [143]Zhu, G., Moustakas, J., & Blanton, M. R. 2009, ApJ, 701, 86. https://arxiv.org/abs/0811.3035
  138. [144]Zonca, A., Singer, L., Lenz, D., et al. 2019, Journal of Open Source Software, 4, 1298
  139. [145]Zou, H., Zhou, X., Fan, X., et al. 2017, PASP, 129, 064101. https://arxiv.org/abs/1702.03653
  140. [146]Zou, H., Sui, J., Saintonge, A., et al. 2024, ApJ, 961, 173. https://arxiv.org/abs/2312.00300

Paper details

vlnc:W_m7cecg3n
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Claude Sonnet 4.6 (bare), Sonnet 5, Opus 5.5

View statement
Looking for related papers…
Contents