Photometric Redshift

The idea: colors are a coarse spectrum

A spectroscopic redshift is conceptually simple: spread a galaxy's light into hundreds or thousands of wavelength bins, find a known emission or absorption line — the calcium H and K doublet, the [O II] 3727 Å line, hydrogen-alpha — and measure how far it has shifted from its rest wavelength. The redshift is z = (λ_obs − λ_rest) / λ_rest, and it is good to four or five significant figures. The catch is photons: to resolve a line you have to collect enough light to fill many narrow bins, and that takes minutes of telescope time per object even for a bright galaxy.

A photometric redshift throws away almost all of that wavelength resolution and keeps only the throughput. Instead of thousands of bins you measure the galaxy's brightness through perhaps five or six broad filters — the ugrizy set spanning the ultraviolet to the near-infrared — capturing the whole spectrum in a single deep exposure of the entire field. Each filter integrates the flux over a band hundreds of angstroms wide. You no longer see individual lines, but you do see the gross shape of the spectrum, encoded in the ratios of flux between filters. Those ratios are the galaxy's colors, and it turns out the colors alone carry enough information to pin the redshift to a few percent.

The reason is that real galaxy spectra are not featureless. They have a small number of broad, powerful features — chiefly the 4000-angstrom break and the Lyman break — that survive being blurred into broadband colors. As a galaxy recedes, those features slide to longer wavelengths in lockstep with (1+z), sweeping past the edges of the filters and changing the colors in a predictable, monotonic way. Read the colors, invert the relation, and you have the redshift. This is the trade at the heart of every modern imaging survey: precision sacrificed for the ability to redshift every galaxy in the sky.

The 4000-angstrom break: a redshift clock

The workhorse feature for galaxies at z < 1.5 is the 4000-Å break, a step-down in flux just blueward of rest-frame 4000 Å. It arises because the atmospheres of cool stars are crowded with ionized-metal absorption lines — most importantly the Ca II H and K lines at 3968 and 3934 Å — that blanket the spectrum below 4000 Å, while the Balmer break of hotter stars adds to the discontinuity. The strength of the break, often measured by the D4000 index (the ratio of flux just redward to just blueward of the break), tracks stellar age: an old, red, quiescent elliptical has a strong break (D4000 ≈ 2), while a young, blue, star-forming galaxy has a weak one (D4000 ≈ 1.2).

Because the break is intrinsically sharp but broad enough to register in broadband photometry, its redshifted position acts as a clock. At z = 0 the break sits at 4000 Å, near the boundary between the u and g filters. By z = 0.5 it has moved to 6000 Å, between g and r. By z = 1 it reaches 8000 Å, between r and i:

λ_observed = 4000 Å × (1 + z)

z = 0.0   →  4000 Å   (u/g edge)
z = 0.5   →  6000 Å   (g/r edge)
z = 1.0   →  8000 Å   (r/i edge)
z = 1.5   → 10000 Å   (i/z edge)

So the photo-z problem reduces, in its simplest form, to "between which two filters does the galaxy's color suddenly change?" The galaxy is bright redward of the break and faint blueward of it; the filter pair that brackets the flux drop tells you where 4000 Å has landed, and dividing by 4000 gives (1+z). Real estimators do this more carefully — fitting the whole color vector, not just one break — but this is the physical core.

The Lyman break: photo-z for the early universe

Beyond z ≈ 3, the 4000-Å break has redshifted into the infrared where ground-based imaging is hard, and a different feature takes over: the Lyman break at rest-frame 912 Å. Neutral hydrogen — both within the galaxy and in the intergalactic medium along the line of sight — absorbs essentially all photons shortward of the Lyman limit. The result is a galaxy that is invisible in any filter sampling below the redshifted 912 Å and abruptly visible above it.

This is the basis of the Lyman-break technique (the "dropout" method) pioneered by Charles Steidel in the 1990s: a galaxy at z ≈ 3 "drops out" of the u-band but appears in g and redder. At z ≈ 4 it drops out of g; at z ≈ 6 out of i. JWST has pushed dropout selection past z = 10, identifying galaxies whose Lyman break sits beyond 1 micron. The Lyman break gives much cleaner photo-z's than the 4000-Å break because the flux contrast across it is total — but it also introduces the most dangerous degeneracy, discussed below.

Two ways to invert colors into redshift

Given a galaxy's measured colors, there are two established strategies to turn them into a redshift.

Template fitting. Take a library of galaxy spectral energy distributions (SEDs) — observed or synthetic spectra spanning ellipticals, spirals, starbursts. Redshift each template across a fine grid of z, convolve it with the survey's filter curves to predict the fluxes, and find the (template, redshift) combination that best reproduces the observed fluxes by minimizing a χ²:

χ²(z, T) = Σ_filters  [ F_obs,i − a · F_template,i(z, T) ]² / σ_i²

Marginalizing over template and the normalization a yields a full posterior P(z) — not just a single number but a probability distribution, which is what cosmology actually needs. Codes like BPZ (Bayesian Photometric Redshift), EAZY, and LePhare implement this. The advantage is that it requires no training data and extrapolates naturally to the faintest galaxies; the weakness is that it is only as good as the templates and the photometric calibration.

Machine learning. Alternatively, take a subset of galaxies that already have spectroscopic redshifts, treat their colors as inputs and their spectroscopic z as the label, and train a regressor — a random forest, a neural network, a Gaussian process, or a self-organizing map. The trained model then predicts photo-z for the far larger photometric sample. Where the training set is dense this is more accurate than template fitting and needs no SED library. Its fatal flaw is extrapolation: it cannot reliably predict redshifts for galaxies fainter, redder, or more distant than anything in its training set — precisely the regime where the biggest surveys live and where spectroscopic training data are sparsest.

Spectroscopic vs photometric redshift

Property	Spectroscopic redshift	Photometric redshift
What is measured	Position of resolved lines	Fluxes in ~5–9 broad filters
Precision (σ_z)	0.0001 – 0.001	0.02 – 0.05 × (1+z)
Distance error at z = 1	≲ 1 Mpc	~100–300 Mpc
Outlier rate	< 0.1 %	~1–10 % (Δz > 0.15)
Time per object	Minutes of slit/fiber time	Free — one image redshifts all
Throughput (modern)	~10⁷ redshifts (DESI, years)	~10⁹–10¹⁰ (LSST, Euclid)
Faint-end reach	Limited by S/N per line	Reaches the imaging depth limit
Output	Single sharp value	Full posterior P(z)

The two are complementary, not competing. Spectroscopy provides the ground truth that calibrates and trains photometric estimators; photometry provides the sheer numbers that statistical cosmology demands. A weak-lensing survey, for example, does not care that any single galaxy's redshift is uncertain by ±0.05 — it cares that the mean redshift of a tomographic bin of ten million galaxies is known to ±0.002, which a well-calibrated photo-z pipeline can deliver.

Real surveys and their numbers

Photometric redshifts are not a niche technique; they are the backbone of survey cosmology.

Survey	Filters	Galaxies with photo-z	Typical σ/(1+z)
SDSS	u g r i z	~2 × 10⁸	~0.03
DES (Dark Energy Survey)	g r i z Y	~3 × 10⁸	~0.03
KiDS	u g r i (+ VIKING NIR)	~10⁸	~0.02–0.04
COSMOS (30+ bands)	UV → mid-IR	~2 × 10⁶	~0.007
Rubin LSST (planned)	u g r i z y	~2 × 10¹⁰	~0.02–0.05
Euclid (planned)	VIS + Y J H	> 10⁹ (lensing)	< 0.05

The COSMOS field is the instructive extreme: with more than thirty photometric bands — effectively a very low-resolution spectrum — its photo-z scatter drops to σ ≈ 0.007 (1+z), approaching the precision of a spectrum. This illustrates the general rule that photo-z accuracy is set by how finely you sample the spectrum. Five broad optical filters give σ ≈ 0.03–0.05; adding the u-band and near-infrared JHK bands roughly halves the scatter and slashes the outlier rate, because they bracket both the 4000-Å break and the Lyman break from both sides.

The degeneracy that wrecks photo-z

The single most important pitfall is the break degeneracy. A broadband filter set sees a "step" in the spectrum between two filters, but it cannot, by itself, tell which step it is. A low-redshift galaxy at z ≈ 0.2 with a 4000-Å break can produce nearly identical optical colors to a high-redshift galaxy at z ≈ 3 whose Lyman break (912 Å) has redshifted to the same observed wavelength. With only optical data, the χ² fit finds two nearly equal minima — a bimodal P(z) — and choosing the wrong one produces a catastrophic outlier with Δz ≈ 2–3.

The cure is wavelength leverage. Adding a u-band detection rules out the Lyman-break interpretation for a low-z galaxy (which should still be visible in u), and near-infrared photometry samples the spectrum redward of the 4000-Å break for genuinely high-z galaxies. This is exactly why Euclid pairs its optical imaging with YJH infrared bands and why LSST's u-band, though shallow, is disproportionately valuable for photo-z. A few percent of objects will still scatter catastrophically — from blends, photometric noise, or unmodeled emission lines boosting a single band — and a major part of any photo-z pipeline is estimating and propagating that outlier fraction into the cosmological error budget.

Worked example: bracketing a redshift from two filters

Suppose you image an elliptical galaxy and find it is faint in g (centered ~4750 Å) and abruptly bright in r (centered ~6250 Å). The flux jumps between these two bands, so the 4000-Å break must lie between them — somewhere around the g/r boundary near 5500 Å. Then:

1 + z = λ_observed / λ_rest = 5500 Å / 4000 Å = 1.375
z ≈ 0.38

To check whether this is sensible, convert to a distance. With H₀ = 70 km/s/Mpc, a redshift of 0.38 corresponds to a recession velocity of roughly cz ≈ 1.1 × 10⁵ km/s (the low-z Hubble approximation), and dividing by H₀ gives a comoving distance of order 1.6 thousand megaparsecs — about 5 billion light-years (a full ΛCDM calculation gives a slightly smaller ≈1.5 Gpc). Now propagate the uncertainty: because the break is only localized to within the ~1500-Å-wide gap between the g and r filter centers (i.e. ±750 Å about the midpoint), the redshift is constrained to only roughly z = 0.38 ± 0.19 from these two bands alone — a distance uncertainty of order a gigaparsec. Adding the i and z bands to confirm the galaxy stays bright redward, and the u band to confirm it stays faint blueward, tightens the estimate and rules out the Lyman-break alias at z ≈ 3. This is the entire photo-z workflow in miniature: locate the break, divide by the rest wavelength, and use every available band to suppress the degeneracies.

Where photometric redshifts power discovery

• Weak-lensing cosmology. Surveys like DES, KiDS, and HSC measure the subtle distortion of billions of galaxy shapes by foreground dark matter. The signal depends on the distances of source and lens, so each galaxy must be placed in a redshift bin — only photo-z can supply distances for that many galaxies. Photo-z calibration is now the leading systematic in lensing-derived values of the S₈ parameter.
• Baryon acoustic oscillations. The ~150 Mpc standard ruler imprinted in galaxy clustering can be measured photometrically by binning galaxies in photo-z shells, trading some radial resolution for vastly more sky coverage.
• Galaxy cluster finding. Algorithms like redMaPPer identify clusters as overdensities of red-sequence galaxies sharing a common photo-z, then use the tight color–redshift relation of those red galaxies to estimate cluster distances to σ ≈ 0.01.
• The high-redshift frontier. Lyman-break dropout selection from deep imaging built the first samples of z > 6 galaxies; JWST photo-z's now flag candidates beyond z = 10 for spectroscopic confirmation, having reshaped expectations about how early galaxies assembled.
• Transient triage. When a gravitational-wave event or a kilonova alert arrives, photo-z catalogs of galaxies inside the localization volume let observers prioritize which host galaxies to follow up in real time.

Common misconceptions and edge cases

• "A photo-z is just a worse spectroscopic redshift." Better to think of it as a probability distribution P(z), often broad or bimodal. Cosmological analyses use the full P(z), not a single point estimate; collapsing it to one number throws away the information that prevents biased results.
• "More filters always help equally." What matters is wavelength coverage, not filter count. Adding a sixth optical band near existing ones helps little; adding a single near-infrared or u-band band that brackets a break helps enormously by breaking the Lyman/4000-Å degeneracy.
• "Machine-learning photo-z is strictly better than templates." Only inside the training distribution. Faint, distant survey galaxies often have no spectroscopic analogues, so ML methods silently extrapolate and bias the result, while template methods degrade more gracefully.
• "Photo-z gives a velocity, so it gives a peculiar velocity too." No. Photo-z scatter (hundreds of Mpc) dwarfs peculiar-velocity signals (hundreds of km/s), so photometric redshifts carry essentially no usable peculiar-velocity information — that requires spectroscopy.
• "Emission lines don't matter for broadband photometry." A strong line such as Hα or [O III] can fall inside one filter and boost its flux by tens of percent, shifting a color and skewing the fit. Surveys with medium bands or careful templates must model this; ignoring it produces a characteristic ridge of outliers.