DESI: how spectra become a map of cosmic expansion

The Dark Energy Spectroscopic Instrument’s first-year map placed 6.4 million galaxies and quasars in a three-dimensional account of cosmic history. The results, published on April 4, 2024, covered seven epochs extending eleven billion years into the past. Dark energy’s behavior might change with time, but the survey alone agreed with the standard cosmological model. Hints of evolution emerged when researchers combined its measurements with supernova observations, and their strength depended on the supernova dataset.

Dark energy’s equation of state relates its pressure to its density. A changing equation of state would differ from the cosmological constant assumed in the standard model. The first combined results were below discovery significance and left unresolved the disagreement between different measurements of today’s expansion rate. Physicist Michael Wood-Vasey of the University of Pittsburgh remained cautious about the hints, while DESI collaborator Nathalie Palanque-Delabrouille expected further observations to test them.

Choosing what to observe

Before its main survey, DESI spent six months in 2020 and 2021 validating how to select and measure targets. Its June 2023 early release made almost two million objects, 2,480 exposures and approximately eighty terabytes of data public. Researchers checked how long galaxies of different brightness required to observe and adjusted target selection before routine collection began. The target populations included stars, bright galaxies, luminous red galaxies, emission-line galaxies and quasars. A concentrated effort called the One Percent Survey examined twenty directions, collecting approximately 700,000 objects across about one percent of the survey’s intended volume. The main survey subsequently filled the gaps between those observations.

Raw images from DESI’s ten spectrometers passed through software that decoded the recorded light into astronomical measurements. The processing code used Python, including its computationally demanding operations. The early release was freely accessible through the National Energy Research Scientific Computing Center. By June 2023, the continuing science run had catalogued more than 26 million objects and was adding over a million per month. Its first two months of routine observations were separate from the validation release.

A ruler preserved in clustering

Pressure waves in the early universe moved through a mixture of ordinary matter and light. Hydrogen and helium nuclei participated in the ionized plasma. As the universe cooled and neutral atoms formed, the waves stopped, leaving a three-dimensional ripple pattern. Galaxies retain that statistical preference in their separations. Researchers measure its apparent size at different distances to reconstruct cosmic expansion. DESI’s first-year measurement, released in April 2024, reached a combined precision of about 0.5 percent across eleven billion years. Its oldest interval, covering roughly eight to eleven billion years ago, reached 0.82 percent. Individual measurements in seven time intervals had different uncertainties, approximately one to three percent.

At the greatest distances, galaxies become difficult to measure directly. DESI instead used light from about 450,000 quasars, examining absorption by intervening hydrogen. Quasars are bright galactic cores powered by black holes. Dense gas clouds absorb their light, leaving a pattern that traces distant matter. Andreu Font-Ribera of the Institute for High Energy Physics described this technique as using the quasars to backlight intervening gas. The hydrogen measurements extended the acoustic ruler to eleven billion years ago.

Adding a test of gravity

A November 19, 2024 analysis used the clustering of nearly six million galaxies and quasars to examine gravity over eleven billion years. The growth of cosmic structure agreed with general relativity across the tested scales. Gravity’s attractive influence amplifies small differences in matter density into the galaxy strands observed today. The full-shape analysis measured how galaxies and matter were distributed across different spatial scales, extending the earlier analysis of the characteristic acoustic separation. Researchers spent months on cross-checks. They concealed the cosmological outcome until the analysis was finalized, as in the earlier acoustic study.

Pauline Zarrouk, who co-led the analysis at the Laboratory of Nuclear and High-Energy Physics, emphasized testing gravity beyond Solar System scales. Dragan Huterer of the University of Michigan co-led the collaboration’s cosmological interpretation group. He identified the measurement as DESI’s first examination of the growth of cosmic structure. The experiment was then in its fourth observing year. Researchers were analyzing the first three years of data and expected updated expansion and dark-energy measurements in spring 2025. The November results also constrained modified-gravity models proposed as alternatives to dark energy.

What three years changed

On March 19, 2025, the collaboration released results from its first three years, using nearly fifteen million of its best-measured galaxies and quasars. DESI’s measurements by themselves remained consistent with the standard model containing a cosmological constant and cold dark matter. Combinations with other observations favored a model in which dark energy changes over time. The external measurements included the cosmic microwave background, supernovae and weak gravitational lensing. The microwave background is light remaining from the early universe; weak lensing measures gravitational distortion of light from distant galaxies. Depending on the dataset combination, the preference for evolving dark energy ranged from 2.8 to 4.2 sigma, below the conventional five-sigma discovery threshold.

The analysis used blinding and additional tests beyond those performed on the first-year sample. Seshadri Nadathur of the University of Portsmouth said the extra checks increased the team’s confidence that an unaccounted-for data effect was not driving the results. The collaboration presented its findings at the American Physical Society’s Global Physics Summit in Anaheim, California. Further analyses of the existing dataset and additional observations were planned.

Making the observations available

DESI’s first major public data release, DR1, opened on March 19, 2025. Its approximately 270 terabytes contained 18.7 million unique objects: about four million stars, 13.1 million galaxies and 1.6 million quasars. The release included survey validation and observations from May 2021 through June 2022. It covered seven times the sky area of the early release and contained roughly ten times as much data. Each night, observations traveled over ESnet to the National Energy Research Scientific Computing Center, where software measured redshifts before the following morning. The processing software ran on the Perlmutter supercomputer’s GPUs nearly forty times faster than on previous NERSC systems, according to scientific data architect Daniel Margala.

Researchers regularly improved the code and reprocessed complete datasets at milestones. DR1 reprocessing took about a month on Perlmutter. The data were freely accessible through NERSC, with documentation for researchers unfamiliar with the project. Objects could also be explored through the Legacy Survey Sky Browser. Under clear conditions, DESI could gather light from a new set of five thousand objects roughly every twenty minutes, exceeding a hundred thousand galaxies in a night.

Finishing an observing plan

DESI finished its originally planned sky map ahead of schedule in April 2026, observing more than 47 million galaxies and quasars alongside twenty million stars. The final night covered a patch near the Little Dipper. Five thousand optical fibers collect light for ten spectrographs, with positioning accuracy of approximately ten microns. The spectra supply information about objects’ positions, velocities and chemical composition. The completed program comprised five bright-time passes and seven dark-time passes, covering about two-thirds of the northern night sky. Bright-time observations use periods when moonlight interferes with measurements of faint, distant targets. Roughly eighty gigabytes of nightly data traveled over ESnet for processing and quality checks.

The collaboration planned to continue observing through 2028, extending coverage from about fourteen thousand to seventeen thousand square degrees. New areas include directions closer to the Milky Way’s plane and farther south. Researchers would also revisit existing coverage for denser samples of distant, faint luminous red galaxies. The first dark-energy results from the completed five-year dataset were expected in 2027, while analysis of the first three years continued.

Reading hydrogen between the galaxies

The Lyman-alpha forest consists of absorption features left as quasar light passes through hydrogen at different redshifts. Each gas cloud absorbs light at a different observed wavelength as the universe expands. Comparing absorption along different quasar sightlines measures how gas is clustered across the distant universe. DESI’s July 30, 2026 DR2 analysis modeled the full correlation pattern, extracting information beyond the acoustic feature alone. The acoustic ruler has a characteristic comoving scale of approximately 150 megaparsecs. The Alcock–Paczynski effect supplies additional geometrical information: an incorrect assumed cosmology stretches inferred distances along and across the viewing direction by different amounts. Researchers measure that anisotropy throughout the correlation function.

DR2 was roughly twice the size of DR1. Its full-shape analysis tightened constraints and shifted the central value toward the standard cosmological model. The collaboration concluded that earlier evolving-dark-energy hints might fade, or that a more complex model might be needed to reconcile all the measurements. Supporting validation work included four hundred simulated Lyman-alpha catalogues. Researchers also varied analysis choices to test their effects on the geometrical measurement relative to its statistical uncertainty.

What a map’s coordinates mean

Comoving coordinates divide out the universe’s overall expansion, allowing a simulation box to retain a fixed coordinate size. Galaxies also have peculiar velocities associated with their local surroundings. Those motions affect observed redshifts and can distort inferred positions along the viewing direction. DESI measures these redshift-space distortions alongside other clustering properties. Nico Schuster’s Cosmic Web Explorer, presented in July 2026, runs a simplified gravitational simulation in a browser. Tens of thousands of tracers follow matter from an initial redshift of ten, with their number held fixed. Users can alter the matter density and the dark-energy equation-of-state parameters, then watch filaments and voids develop.

The simulation displays a thin, flattened slice and uses short-range gravity. Its void finder starts from empty points and grows regions toward surrounding dense structures, producing irregular boundaries. A separate mode artificially amplifies the acoustic feature so that rings become visible. A redshift-space mode switches velocity distortions on and off. Schuster designed the tool for teaching and intuition rather than absolute precision; even substantially different dark-energy models could produce visually similar simulated webs.

Using spare fibers for a changing sky

DESI also collects spectra of transient events with fibers left unassigned by the main cosmology program. Initial observations of alerts from ATLAS and the Zwicky Transient Facility produced approximately a thousand transient spectra. By May 4, 2026, a DESI-Rubin program was pointing specifically toward Rubin Observatory’s deep drilling fields. DESI could cover almost an entire field, including COSMOS, in one pointing. The team had targeted as many as three hundred transients simultaneously.

The Babamul broker filtered the alert stream using custom criteria, machine-learning scores and catalogues to exclude variable stars. Reduced spectra generally arrived the next day. Next Generation Superfit software helped classify supernovae, while most followed-up transients proved to be active galactic nuclei. Around four hundred Rubin targets had yielded approximately ten supernovae at the interview date.

The team shared supernova classifications and spectra through the Transient Name Service. Researcher Xander Hall hoped to revisit COSMOS every two to three weeks for spectral sequences. DESI’s resolving power of about five thousand could reveal narrow features beyond those accessible in customary lower-resolution supernova follow-up. Additional transient work was being proposed for DESI-II.

Stellar motions in Andromeda

Measurements of nearly 7,500 stars in Andromeda’s inner halo revealed patterns in their positions and velocities associated with a merger about two billion years ago. The results were released on February 8, 2023. These stars had formed in another galaxy before joining Andromeda, also called Messier 31. Arjun Dey of NOIRLab led the research using DESI on the Mayall telescope.

Coauthor Sergey Koposov of the University of Edinburgh interpreted the inner halos of Andromeda and the Milky Way as being dominated by a single immigration event in each galaxy. Most Milky Way halo stars came from another galaxy through a merger eight to ten billion years ago. The team planned further measurements of Andromeda’s outlying stars and hoped to survey its entire halo.

Dey said a few hours of DESI observations surpassed more than a decade of spectroscopy with much larger telescopes. The Mayall telescope achieved first light in 1973 and subsequently received upgraded instrumentation. Harvard undergraduates Gabriel Maxemin and Joshua Josephy-Zack participated through the Radcliffe Institute for Advanced Study, where coauthor Joan Najita had been a fellow in 2021–2022.

Keeping expectations out of the result

DESI masks the cosmological answer during analysis by shifting galaxy positions along the viewing direction. The shifts imitate a randomly selected alternative cosmology while preserving the statistical properties of the data. The method applies to acoustic-ruler and redshift-space-distortion measurements. Researchers Sam Brieden, Uendert Andrade and Juan Mena-Fernández discussed its operation in March 2025.

Andrade described three stages:

1. Construct a concealed modification that gives analysts scientifically plausible data.
2. Develop, test and finalize the analysis without knowing the true cosmological answer.
3. Remove the modification and apply the finalized procedure to the real measurements.

The full-shape analysis was unblinded on June 12, 2024, before publication on November 19. Brieden developed both the blinding and ShapeFit full-shape methods during his doctoral work. Unblinding typically takes place during a collaboration video call, with prepared scripts producing plots and final measurements. Under the procedure Andrade described, no further changes to the analysis are allowed after unblinding. Mena-Fernández said some blinding methods allow measurements while hiding their results from the analysts. Andrade emphasized that working this way requires trust in the processing pipeline and tests of the blinding procedure itself.