How DESI Works

The light path from galaxy to CCD — the Mayall telescope, 5,000 robotic fiber positioners, and three spectrograph arms

Each DESI measurement begins as a stream of photons that left a galaxy billions of years ago and ends as a distribution of electrons stored in a charge-coupled device. The distribution of electrons across the face of the CCD is proportional to the wavelength of the light at the time it reaches DESI, which has been redshifted by the expansion of the universe during its long travel time. This page follows that path through the instrument.

The telescope

The Mayall 4-meter telescope on Kitt Peak was built in 1973 and has been rebuilt around DESI: a new corrector widens the usable field, and at the focus sits a focal plane carrying 5,000 optical fibers, each on its own robotic positioner. Before an exposure the robots place every fiber on a chosen target to a few microns.

Scale drawing of the Mayall 4-meter telescope tilted at 32 degrees, the Kitt Peak latitude, on its equatorial horseshoe-yoke mount. Five ray bundles spanning a 3.2 degree field of view, each drawn in a different color to mark its field angle rather than any wavelength, enter the tube, reflect from the primary mirror, pass through the six-element DESI corrector and converge on the curved focal plane, which is labeled as having fibers pointing downward. A panel lists the specifications: 4 meter f/2.7 paraboloid, six-lens corrector, delivered f/3.9, 0.8 meter curved focal plane, 5,020 robotic fiber positioners. — The Mayall in its working pose, tilted to the Kitt Peak latitude of 32°. Five ray bundles spanning the 3.2° field of view reflect from the 4 m f/2.7 primary, pass the six-element corrector, and converge on the curved focal plane. Color distinguishes field angle, not wavelength. Geometry is to scale; the rays are traced geometrically rather than element by element.
By J. W. Rohlf. Source: mayall_desi_raytrace.py (Python, matplotlib) — interactive when run locally, with a tilt slider and a focal-plane zoom.

Zoomed view of the DESI corrector barrel and focal plane. Five ray bundles, one color per field angle rather than per wavelength, pass through six lens elements and land at different radii along a curved focal surface at the top, where short pencils mark the fibers pointing downward from each landing point. — Zoomed on the corrector and focal plane. The five field angles land at different radii on the curved focal surface — which is what the 5,000 fiber positioners exist to follow, and why the focal plane is curved rather than flat.

That curvature has a direct consequence for the hardware. Each petal carries 500 fiber positioners and, because the focal surface curves, no two of their sockets sit at the same angle — every one is precision machined to its own orientation, so that the fiber it holds points along the local normal to the surface. Ten petals, each with 500 individually angled sockets, make up the 5,000.

Making them was Boston University's contribution to the instrument. The petals were machined at BU, each with its five hundred sockets cut to its own orientation; the fiber positioners were prototyped at BU as well, along with the jigging that held everything in register for the assembly carried out at the University of Michigan.

This is unglamorous engineering that sets a ceiling on everything downstream. A fiber whose socket points a little off the local normal collects less light, and loses it before the spectrograph ever sees it. A positioner that cannot place its fiber to a few microns puts part of the wrong object — or part of the sky — onto a target that subtends about one arcsecond. Neither error can be corrected later in software: the photons are simply not there. The precision of the map begins in a machine shop.

David Kirkby's 360° panoramas of the DESI focal plane show the assembled hardware in place on the telescope.

Each fiber has a core 107 µm across, subtending about one arcsecond on the sky — which is why a DESI spectrum samples the central arcsecond of a galaxy rather than all of its light. The fibers carry the light out of the telescope to ten spectrographs housed in a temperature-controlled room below.

Assigning the fibers

Before any of that happens, something has to decide where each of the 5,000 fibers points. A single field contains far more candidate targets than there are fibers — about 30,000 against 5,000 — and the positioners cannot reach arbitrarily far or overlap one another. Assignment is therefore a scheduling problem, solved by priority: quasars first, then emission-line galaxies, luminous red galaxies, bright galaxies and finally standard stars.

Assignment revealed one priority class at a time. Left: the whole focal plane, 5,000 positioners against 30,613 candidate targets, colored by what each fiber was given. Center: one petal, with the two-arm robots drawn individually — each can only reach targets inside its own patrol disk, and neighbors that would collide are rejected. Right: completeness by class. Quasars are rare, so almost a quarter of them get a fiber; emission-line galaxies are so numerous that only 15% do, and they still take 2,845 of the fibers. This is why the survey visits each point on the sky more than once.
Animation by J. W. Rohlf. Source: desi_focal_plane.py (Python, matplotlib, scipy) — interactive when run locally, with sliders for patrol radius, collision distance and target density.

The priority order is a cosmological choice, not an engineering one. Quasars go first because they are the scarcest tracer and the only route to the Lyman-α forest; stars go last because they are calibration rather than signal. Which tracer a given fiber ends up observing is decided here, seconds before the shutter opens.

The light path

From galaxy to CCD, light reaches the Mayall primary, passes the corrector to the focal plane, enters a fiber, and is collimated; dichroic D1 peels off the blue, D2 splits red from near-infrared, and each arm is dispersed by its own grating onto its own detector — so all three CCDs are illuminated at once, covering 360 to 980 nm in a single exposure.
Animation by J. W. Rohlf; a still of the completed path appears in Results From DESI, DPF 2026. Source: desi_light_path_ltr.py (Python, matplotlib).

Inside a spectrograph

Five hundred fibers enter each spectrograph through a V-groove block that holds them in a precise line, and a collimator turns the diverging beams into parallel ones. Then the light is divided. The first dichroic, D1, reflects the blue and passes everything else; the second, D2, splits what remains into red and near-infrared. Each of the three beams meets its own grating and its own CCD:

CCD B: 360–593 nm — the blue arm, where [O II] appears for the emission-line galaxies
CCD R: 566–772 nm — the red arm, carrying Hα at low redshift and the 4000 Å break at z ≈ 0.8
CCD NIR: 747–980 nm — the near-infrared arm, which sets the upper redshift reach of every galaxy sample

Splitting the beam is what makes the survey possible. A single detector covering 360 to 980 nm would have to compromise either resolution or efficiency somewhere across that range; three optimized arms do not. It is also why the tracer samples have the redshift limits they do — each is bounded by the wavelength at which its defining feature runs off the end of the third CCD.

What lands on a CCD

Each fiber occupies its own strip of the CCD; in this view, wavelength runs up the frame. Each exposure records 500 spectra side by side on each CCD.

Simulated red-arm CCD frame. Twenty fiber columns run across the image and wavelength increases upward from 550 to 720 nanometers, with a color bar at the left. Bright horizontal bands crossing every column are sky emission lines; short isolated dashes are cosmic-ray hits. A yellow box marks fiber 1. Below, the spectrum extracted from that fiber shows four broad sky-line peaks and one narrow marked peak near 710 nanometers. — A simulated red-arm exposure, 550–720 nm, with twenty fibers rather than 500 so that the individual strips stay legible. Wavelength increases upward. The bright horizontal bands crossing every column are **sky lines** — the atmosphere, recorded identically in every fiber and later subtracted. The short isolated dashes are **cosmic rays**, which is what limits a dark-time exposure to about twenty minutes. The yellow box marks the fiber whose extracted spectrum is shown below: four sky peaks, and one narrow line at 710 nm that is Hα from a galaxy at z = 0.08 — 656.3 nm at rest, stretched by 8% due to cosmic redshift.
By J. W. Rohlf. Source: ccd_mapping.py (Python, matplotlib) — interactive when run locally, with sliders for line strength, continuum, sky level and read noise.

Reading it out

Readout is the last thing that happens inside the instrument, and it is not instantaneous. A CCD holds charge, not numbers, and turning one into the other means moving that charge physically across the chip to a single amplifier at the edge. The readout electronics do this row by row, and whatever they add along the way — the read noise — sets the floor beneath every spectrum DESI records.

Charge does not stay on the chip. Clocking voltages march it step by step towards a serial register at the edge, where it is amplified and digitized one packet at a time — so the frame empties progressively, as here. The yellow box marks the fiber whose extracted spectrum is shown below; watch it change as each new column arrives, and the cosmic ray pass through it as a flat-topped block that no real photon could produce.
By J. W. Rohlf. Source: ccd_mapping.py.

Subtracting the sky

A digitized frame is not yet a useful measurement until the night sky background, glowing in OH bands that are hundreds of times brighter than the galaxies behind them, is removed. This is why some fibers are deliberately pointed at dark sky — they measure what has to be subtracted.

A 600-second exposure accumulating photon by photon, over 738–756 nm. Top: the raw galaxy fiber — four bright OH sky lines, with the [O II] doublet at 746.2 and 746.8 nm barely a ripple beside the 747.2 nm sky line. Second: the sky model, the median of ten sky fibers, with its ±1σ uncertainty. Third: the difference — the doublet climbing out of the noise as the exposure grows, against the true noiseless signal in red. Bottom: the residual, which should stay inside the ±1σ envelope if the subtraction is unbiased. Watch the signal-to-noise figures: by the end each line is above 90.
Animation by J. W. Rohlf. Source: desi_sky_subtraction.py (Python, matplotlib).

Elsewhere on this site

More to come: the same talk includes simulations of structure formation, of the baryon acoustic oscillation feature, and of redshift-space distortions, which will be added here.