Measuring redshift from DESI spectra, without Redrock

Determining spectroscopic redshift and object type from flux and inverse variance alone — and, just as much, learning when the answer cannot be determined and should be refused.

The type is read from an unlabelled image of the spectrum. That reading then selects which template family is fitted, and the fit gives the redshift.

---

Start here

METHOD.md describes how the type is read and the redshift is measured. Each rule comes with the measurement behind it, along with the negative results and what was tried and rejected.

RULE.md records what is enforced, and in code, and the corrections log traces every rule to the error that bought it. It is not a checklist to read before judging a spectrum. Reading it first was measured to spread a wrong rule more than reading without it did.

DESIGN.md describes the model being built (zfm). That covers its architecture, loss, and evaluation, plus how its training labels are cleaned. Every constraint carries the measurement that forces it.

Results

Nine sealed blind tests on fresh DESI DR1 spectra — 400 spectra, no target ever repeated across cohorts. In each, the reference redshifts and types were written to a file at mode 600 during the build and opened once, after every call was already on disk.

Tests IV–IX are main; tests X–XIII are on SV3, a deeper survey. The median depth ratio (TSNR2) is 1.29× for ELG and LRG, 1.77× for BGS and 2.23× for QSO; only its deepest tail reaches 2–4×. The SV3 columns are therefore not directly comparable to the main ones and are kept in a separate table.

	IV	V	VI	VII	VIII	IX	tests VI–IX
spectra	100	50	50	50	50	50	200
redshift within ±1000 km/s	95/100	47/50	48/50	46/50	50/50	47/50	191/200
redshift within |Δz| < 0.1	—	—	50/50	49/50	50/50	48/50	197/200
type, 5-way vs target class	89/100	46/50	47/50	45/50	43/50	49/50	184/200
type, 3-way vs SPECTYPE	98/100	50/50	50/50	48/50	49/50	50/50	197/200
the target bits, same comparison	99/100	50/50	50/50	46/50	47/50	50/50	193/200

SV3 — four tests, 200 spectra

	X	XI	XII	XIII	all four
spectra	50	50	50	50	200
redshift within ±1000 km/s	48/50	49/50	43/50	44/50	184/200
redshift within |Δz| < 0.1	50/50	49/50	43/50	44/50	186/200
type, 5-way vs target class	47/50	47/50	42/50	41/50	177/200
type, 3-way vs SPECTYPE	50/50	50/50	44/50	44/50	188/200
the target bits, same comparison	50/50	49/50	45/50	42/50	186/200

Test XIII was read with RULE.md deliberately not re-read beforehand, to see what the checklist was carrying rather than what the reader had absorbed. It scored the same as test XII, which was read with the checklist open, so the drop from X and XI is not the checklist's absence. It is the selection: tests XII and XIII were built with --zmax 1.6, X and XI without one. Under no cut, all 20 QSO targets are quasars at z = 1.61–2.50 with three or four broad UV lines in the window, and they were read 20/20. Under the cut the QSO targets are 9 quasars, 7 stars and 4 galaxies at z ≤ 1.46, and the surviving quasars have one or two lines. Errors fall 5 of 10 QSO targets against 2 of the other 40. Grouped by lines in the window: 4 lines 9/9, 3 lines 13/14, 2 lines 3/5, 1 line 0/4. Both of test XIII's type failures were caught afterwards by a guard that now refuses to run: the strongest line filter peaked above 8σ at a wavelength the candidate record never mentioned.

Two redshift criteria, because they measure different things. ±1000 km/s is the standard catastrophic-failure threshold and is a statement about precision; |Δz| < 0.1 is roughly 20,000 km/s at z = 0.5 and is a statement about identification — did the right object, at roughly the right distance, come out. Under the looser bar the score is 197/200, and five of the six newly-forgiven cases are quasar systemic offsets of 1100–1900 km/s, which were never misidentifications: broad lines are blueshifted by outflows relative to the systemic redshift. The sixth, OBJ034 at |Δz| = 0.098, is a genuine line misidentification that happens to land 2% inside the bar — the loose criterion is not a clean misidentification detector either.

Of the three that fail both criteria, two are the reference being wrong (see The reference has an error rate of its own in METHOD.md). That leaves one genuine catastrophic error in 200 objects, OBJ044, where the template fit chose a solution whose 4000 Å break measures −0.5σ over one measuring 5.5σ.

The target-bit row is the other one to read. The DESI photometric target bit is a prediction of what the object is, made before any spectrum existed. Over the 200 objects of tests VI–IX it agrees with Redrock's spectroscopic classification 193 times; reading the picture agrees 197 times. A large share of the 5-way "type errors" are therefore the target bit being wrong, not the reading.

Part V is exact on the three-way classification — every one of the fifty gets the same STAR / GALAXY / QSO type Redrock assigns.

BGS went from 13/20 to 10/10 between the two tests, from one change: BGS-vs-LRG is a redshift cut (z ≈ 0.42, worth 92.6% over 20,000 objects), not something visible in a spectrum, so it is assigned after the redshift is measured instead of guessed from the break position. The BGS↔LRG cell that dominated Part IV's errors is now empty.

Of Part V's three redshift misses: one is a true information limit (every line past the 9824 Å edge), one is a 1226 km/s quasar near-miss against a 1000 km/s mark — not a misidentification — and one is a 23σ line I mistook for a sky residual.

---

Which Redrock redshifts to trust: cleaning the training labels

zfm is trained on Redrock redshifts, and about 1% of ZWARN == 0 labels are wrong. A model that learns them copies them. code/rrwrong/ learns to say when a Redrock redshift is wrong, and the training data is rebuilt without the labels it distrusts. The full record is in DESIGN.md §6d-3.

• Truth it learns from. DESI EDR visual inspection, 18,653 targets (Lan et al. 2022, arXiv:2208.08516; Alexander et al. 2022, arXiv:2208.08517), with Redrock re-run at SV1 per-night depth, which is close to main-survey depth.

• Inputs. Redrock's own fit quality (DELTACHI2, χ²/pixel, ZERR, template coefficients, z, SPECTYPE), imaging photometry, Gaia, and the target bits.

• Checked against data it never saw. Each test compares the model with DELTACHI2 alone, keeping the same number of rows:

  test	DELTACHI2 alone	rrwrong
  1.0M main targets on two tiles: wrong-pair rate	0.73%	0.16%
  2,022 Gaia-confirmed stars Redrock placed at z > 0.1: removed	31%	100%
  110 blind-read Redrock errors: removed	14%	78%

• The clean-2 training set removes a row if any of three rules holds: a repeat observation disagrees or Gaia says it is a star; DELTACHI2 < 25; or P_WRONG > 0.01. It deletes 6.6% of main (17.43M rows kept). The backup program is exempt: no inspected spectra cover it, and there the model flagged 72% of Gaia-confirmed stars.

Scored on 60,032 validation spectra, zfm disagrees with Redrock on 727 (1.21%). On a fresh 60,032 drawn from the filtered set it disagrees on 36 (0.060%), and NMAD of Δz/(1+z) stays 2.0e-05. What changed is the reference, not zfm. Most of the disagreements sat on doubtful Redrock labels. Where Redrock is proven wrong, zfm is right on 80% of the Gaia-confirmed stars and 77% of the blind-read errors.

The rule-by-rule breakdown and the per-class figures (STAR, BGS, LRG, ELG, QSO) are in figures/clean2_s3_filter.png, clean2_s4_before_by_class.png and clean2_s5_after_by_class.png.

---

What's here

path	what it is
METHOD.md	how the type is read and the redshift is measured — start here
report/redshift_report.html	the full report, 36 sections, 24 figures. Open in a browser
DESIGN.md	the model to be built: architecture, loss, refusal and evaluation, each constraint carrying the measurement that forces it; §6d covers training-label cleaning
RULE.md	what is enforced and the corrections log (a record, not a pre-read checklist)
report/redshift_report.ipynb	the same, executed, with all code
figures/	c2_*, c3_*: blind-test headline figures; clean2_*: the label-cleaning results
code/	the measurement path and the blind-test tooling
code/rrwrong/	the learned Redrock-error filter, its validation, and the clean-2 builder (see its README)
code/build_h5/	builds the two DR1 training files from the public DESI release: ZWARN == 0 selection and official B/R/Z camera coaddition (below)

Training data: code/build_h5/

Every model is trained and tested on two HDF5 files built from DESI DR1 (iron) by these two scripts. Clean-2 (code/rrwrong/make_clean2_h5.py) is a row-deleted copy of the same two files, not a separate source.

script	output	rows
build_dr1_sv3_brz_h5.py	dr1_sv3_primary_tgt_zwarn0_brz.h5	992,821
build_dr1_main_brz_h5_parallel.py	dr1_main_primary_tgt_zwarn0_brz.h5	18,659,804

• source: zcatalog/v1/zall-pix-iron.fits for the selection, and the healpix coadd-<survey>-<program>-<healpix>.fits files for the spectra, both under /global/cfs/cdirs/desi/public/dr1/spectro/redux/iron/
• selection: SURVEY == "sv3" (or "main"), ZCAT_PRIMARY, OBJTYPE == "TGT", ZWARN == 0; the scripts refuse to run if the selection contains a repeated TARGETID
• spectra: B, R and Z are combined with desispec.coadd_cameras (desispec 0.68.1) into one 3,600–9,824 Å array, 0.8 Å per pixel, 7,781 pixels; BRZ_FLUX, BRZ_IVAR and BRZ_MASK are stored per row, the wavelength grid once (its sha256 is checked across every coadd file)
• catalogue columns: TARGETID, Z, ZERR, SPECTYPE, ZWARN, DELTACHI2, target bits and the rest of ZALL_FIELDS in the sv3 script; the main script renames SV3_*_TARGET to *_TARGET and SV_PRIMARY / SV_NSPEC to MAIN_PRIMARY / MAIN_NSPEC
• main is parallel: 64 processes write separate shards, which are merged by copying compressed chunks and then verified chunk by chunk, with two rows per shard recomputed from the source FITS. It imports the sv3 script for selection and coaddition, so the two must sit in the same directory
• provenance: each h5 records the sha256 of the script that built it (BUILD_SCRIPT_SHA256, ORIGINAL_SV3_SCRIPT_SHA256); the files here match those values

These files are a labelled export, not a blind cohort: ZCAT_PRIMARY and ZWARN are Redrock-derived, and Redrock's Z is the training label. The blind protocol below governs the reading cohorts, not these files.

Run on NERSC in an environment with desispec, fitsio, astropy and h5py. --dry-run (sv3) or --prepare-only (main) stops after the catalogue selection; both builds resume from where they stopped.

Code

file	role
speclib.py	official BRZ camera coaddition; continuum, significance and break statistics; rendering
engine.py	the FFT redshift scan — redshift is a pure translation in ln λ, so one transform covers every z
templates.py	per-class template families and their redshift ranges
methodc.py	two-stage fit: candidate generation, then decision on a common data vector
d4000.py	4000 Å break candidate generation
make_cohort.py	builds a blind cohort and writes the sealed key
make_cohort_h5.py, h5pick.py	the same from the DR1 BRZ h5 files ("sv3" → SV3 file, otherwise main), with non-repeating draws
measure3.py	fits the family the visual call selected; the call is five-way (STAR / BGS / LRG / ELG / QSO), a trailing ? marks a guess
final3.py	opens the sealed key once, scores, draws the figures
zoom.py	four-width diagnostic for spectra too faint to read from flux
check.py	asserts the cohort invariants; exits non-zero if broken
slideguard.py	refuses figure text that breaks a reporting rule
scaling.py	FLOP cost and run plan for the scaling-law fit

---

The blind protocol

No Redrock information reaches selection, inputs or figures, with one declared exception noted below. That excludes Z, SPECTYPE, DELTACHI2, and the ZCAT_PRIMARY / MAIN_PRIMARY dedup flags — those rank by ZWARN and would smuggle Redrock's confidence ordering in through the selection. Those values appear only when scoring work that is already finished.

• selection columns: TARGETID, SURVEY, PROGRAM, HEALPIX, COADD_FIBERSTATUS, DESI_TARGET, TSNR2_*, and ZWARN as a == 0 filter only
• filters: SURVEY=="main", TARGETID>0, COADD_FIBERSTATUS==0, ZWARN==0
• class: DESI_TARGET bits, priority QSO(2) > LRG(0) > ELG(1) > BGS_ANY(60) > MWS_ANY(61)=STAR. These are photometric target classes, not spectroscopic truth — what the spectrum turns out to be is a finding
• repeated TARGETIDs: keep the highest TSNR2_<class> row
• SNR: measured from the spectrum, median(|flux|·√ivar) — never TSNR2, which describes the observation and not the object
• renders: an opaque id and nothing else, shuffled across classes

The one exception: ZWARN == 0

From 2026-09-02 every cohort selects only objects with ZWARN == 0. This is a deliberate hole in the rule above, and it is worth being explicit about both halves of it.

The reason is that a score is meaningless against an answer key that has disowned its own number. ZWARN != 0 is Redrock flagging its own redshift as unreliable, so a disagreement there is two uncertain answers disagreeing, not a measurement of the reader. The n = 50 test below contains exactly this case: OBJ011 is counted as a 144,966 km/s failure against a reference carrying ZWARN 4.

The cost is that this is not a random cut. ZWARN != 0 concentrates on faint, high-redshift, feature-poor spectra — the hardest objects — so filtering them out removes hard cases by construction and makes every later cohort easier in a way the score does not show. A result from such a cohort must be read as "of objects Redrock is confident about", never as a result over DESI as a whole.

Tests IV and V predate the rule and were not filtered this way; code/check.py reports their ZWARN != 0 counts (3 of 100 and 1 of 50) without failing them, and holds later cohorts to the rule via the data/selection_rules.json each build writes.

---

Things worth knowing before reading the report

"SNR" names three different quantities. TSNR2 is a property of the observation; SNR per pixel is how bright the object is; feature significance decides whether the redshift is measurable. TSNR2 correlates with how bright a star is at −0.002 — exactly zero, because exposure depth cannot change that. The survey equalises evidence, not photons: in dark time an LRG's per-pixel SNR halves while its break significance is unchanged, and the extra exposure buys distance (median z 0.41 → 0.81) rather than precision.

Different classes carry the redshift in structurally different places. Half of an ELG's redshift information lives in 7 pixels of 7781; an LRG needs 1425. That is a difference in kind, not difficulty, and it is why one instrument cannot serve both — a line filter on an LRG returns 3.9σ, a break statistic on an ELG returns 1.4σ.

The 4000 Å break is a shape, not a feature. A local-step detector finds it 1% of the time and locks onto Hα or Na D instead; reading the whole continuum shape by eye beat it 8:1.

Every remaining error is discrete. Correct answers sit 1.9–2.9 σ_z from the truth; wrong ones are 3,470–13,362 σ_z away, and not one error in 12,000 falls within 5σ_z. The two populations do not overlap, so the remaining problem is which peak, not how precisely.

---

Status and honest limits

• The blind tests are small. Each has n = 50–100, and all of them together come to 400. Several results in this project looked strong at n ≈ 40 and moved substantially at n ≈ 20,000; the report flags each case where that happened.
• The label-cleaning comparison (v2 against v1 training data) rests on 2 seeds per arm, so it gives a direction, not a verdict. Its decisive test, scoring on visually inspected spectra, has not been run yet.
• Scores are agreement with Redrock, which is itself a measurement. Where ZWARN != 0, the two disagree without either being demonstrated wrong.
• The target class is photometric intent. Some disagreements with it are the imaging label being wrong, and the report separates those from genuine errors.
• The report records what failed as carefully as what worked, including conclusions of mine that later evidence overturned.