---
title: "Read EAZY and a Euclid calibration paper with the selection visible"
description: "Follow pinned EAZY sections and Euclid 2026 figures to separate template fitting, posterior summaries and population redshift calibration."
canonical_url: "https://sandboxphysics.com/stories/galaxy-colors/reading-photoz-reliability/"
language: "en"
page_type: "story"
alternate_languages:
  zh-CN: "https://sandboxphysics.com/zh/stories/galaxy-colors/reading-photoz-reliability/"
last_modified: "2026-10-28"
---

[Galaxy colors from photons to redshift distributions](https://sandboxphysics.com/stories/galaxy-colors/)

L4 · Galaxy photometry

# Read EAZY and a Euclid calibration paper with the selection visible

Sandbox Physics · 2026-10-28

Open EAZY, arXiv:0807.1533v1, and the Euclid deep-to-wide transfer paper, arXiv:2601.02005v1. The first explains a widely used route from colors to redshift estimates; the second asks how a deep reference sample can calibrate the redshift distribution of a shallower survey. Reading them together separates a plausible answer for one galaxy from a reliable distance distribution for a whole population.

## Before you begin

Read the finite-model posterior chapter and reproduce the SDSS exercise first. These notes guide source reading; they do not reproduce either paper’s catalogue, software pipeline or cosmological result.

## Start with the model in EAZY section 2.1

In section 2.1, identify observed fluxes, errors and template combinations before looking at a redshift scatter plot. EAZY permits nonnegative linear combinations of templates. In sections 2.2–2.3, note that template construction and a rest-wavelength template-error function are separate ingredients. Section 2.4 adds a magnitude-dependent redshift prior.

Make a small comparison sheet for any code you inspect: how a spectrum is normalized, which template coefficients are optimized, which parameters are integrated, how observational and model errors enter, and which prior multiplies the result. Our course integrates one positive amplitude and sums a finite age/dust prior. EAZY’s optimized combinations are a different calculation. A shared label such as probability curve does not make two inference procedures identical.

When a variance depends on a trial template or redshift, write the full Gaussian density, including the determinant or product of standard deviations. Merely inflating a residual denominator can make a noisy model look artificially attractive. The software’s exact weighting and normalization define what its reported curve means.

Reading: Brammer · EAZY manual and source

## Match a redshift column to its definition

Read section 2.5 and the pinned manual’s Output files section. The manual distinguishes a minimum-chi-square redshift before the prior, a maximum after the prior, and redshift means obtained from normalized probability curves. Its z\_m1 and z\_m2 are means, with the latter including the prior. Record the actual column and version when comparing catalogues.

A useful exercise is to return to the course’s two-peaked example and compute its mean, median and highest peak. Ask whether each summary points to a dense part of the distribution. Also integrate the probability near both boundaries. A mean shifts when a distribution is cut off, even if its tallest interior peak barely moves. A redshift range is therefore part of the inference.

Reading: EAZY manual and source · Hogg et al.

## Use Figures 8–11 to ask who has a spectrum

Figure 8 compares estimates with and without the prior for the full photometric sample. Figure 9 relates residuals to a quality statistic. Figures 10–11 compare changed calibration or filter coverage, displaying the full sample as well as the spectroscopic subset. Section 4.3 explains why a reassuring spectroscopic comparison can miss shifts elsewhere in the sample.

For each panel, trace one group far from its comparison diagonal. Is that group represented by measured spectra? What magnitude selection gave those spectra? If only a few bright galaxies occupy it, a low global scatter on the spectroscopic subset may leave most of the group untested. This is the same selection issue that prevents our twenty-four bright SDSS objects from calibrating faint imaging galaxies.

Design a follow-up before choosing a favorite algorithm. Target a color-and-brightness region where the algorithms disagree and the spectroscopic coverage is thin. Keep that test separate from the sample used to tune zero points or priors. If obtaining spectra fails for a particular kind of galaxy, record that failure too; successful spectra alone can hide a second selection.

Reading: Brammer · SDSS DR17 imaging fluxes and flux calibration

## Move from one galaxy to a survey distribution

The survey question is a distribution $n(z)$: how many selected galaxies lie in each redshift interval. Even narrow individual posteriors can give a biased population if their calibration is wrong or the reference sample misses certain galaxies. A weak-lensing source bin depends on intrinsic spectra, image detection and selection.

Imagine two galaxies with identical intrinsic colors observed once deeply and once near a detection limit. The shallow measurements scatter farther, and some observations fail selection. A reference sample with precise colors will occupy different measured-color regions from those noisy detections. Matching only the ideal colors does not repair the difference.

A color map groups similar multiband measurements into cells; a self-organising map is one method of arranging those cells. The calibration problem is to connect their occupancy in the survey to reliable redshift information in the reference sample. Empty or weakly measured cells identify where a claimed distribution relies most on extrapolation.

Reading: Euclid Collaboration · Brammer

## Read the Euclid transfer test before its performance figure

In arXiv:2601.02005v1, read sections 2.1–2.3 for the single-band and multiband transfer and the subsequent selection. Sections 3.1–3.2 use Flagship 2.1 mock galaxies for redshift-distribution validation, including a specified noise construction. Figure 5 compares calibration configurations; Figure 6 shows a selected reconstructed distribution against its mock truth.

The Figure 6 caption gives true mean 1.045, Scenario C mean 1.048 and Scenario B mean 1.079 for the illustrated bin. Subtract before dividing: the normalized mean offsets are approximately 0.00147 and 0.0166. These two calculations concern a population mean. They are not the scatter of individual galaxy redshifts and cannot be compared directly with the SDSS exercise’s normalized median absolute deviation.

Read the widths and selection alongside those means. Ask which differences were fixed by changing the transfer, which by assigning tomographic bins differently, and which assumptions remained in the mock. A successful controlled simulation demonstrates behavior under those assumptions; applying it to a sky catalogue also needs checks of the observed flux, error and selection distributions.

$$
\Delta\bar z_{\mathrm{norm}}=\frac{\bar z_{\mathrm{reconstructed}}-\bar z_{\mathrm{true}}}{1+\bar z_{\mathrm{true}}}
$$

For the caption example, use the true mean in the denominator. A small population mean offset can coexist with individual redshift ambiguity.

Reading: Euclid Collaboration · SDSS DR17 imaging fluxes and flux calibration

## Carry these questions to the next paper

Name the target population and the reference selection. Record the flux system, covariance assumptions, template families and prior. Separate optimized coefficients from integrated parameters. Identify the redshift summary and the probability range. Check which colors and magnitudes have independent truth, then inspect residuals, interval calibration and the population distribution.

Finally state which experiment you actually repeated. Here the executable reproduction is the frozen SDSS finite-SSP exercise. The EAZY and Euclid pages are source-reading guides with explicit sections and a caption arithmetic check. A full reproduction would additionally need their exact catalogues, filters, templates, configuration, selection transfer and independent validation products.

Reading: EAZY manual and source · Euclid Collaboration

## Try the observation

### [Compare a point estimate with its surviving alternatives](https://sandboxphysics.com/labs/galaxy-colors-photometric-redshift/)

1. Turn off the u detector and note the highest peak, central interval and secondary basin probability.
2. Use this controlled ambiguity to explain why a point-estimate scatter plot and a population-calibration test ask different questions.

## Exercises and explained answers

A new method lowers the scatter on bright spectra but changes faint-galaxy bin counts substantially. Which result should decide whether its population calibration improved?

Neither result decides alone. Obtain representative independent redshift information or a justified calibration-transfer test for the affected faint color regions, with selection and uncertainties included. A smaller bright-sample scatter can coexist with worse faint-population calibration.

[Previous: Reproduce redshift distributions for twenty-four real SDSS galaxies](https://sandboxphysics.com/stories/galaxy-colors/sdss-five-band-tutorial/)

## Sources and calculation scope

1. [Hogg et al., The K correction, astro-ph/0210394v1 (2002)](https://arxiv.org/abs/astro-ph/0210394v1)

   Equations (2)–(12): spectral densities, photon response, redshift factors and the AB K correction. We use luminosity distance and define wavelength densities per Angstrom.

2. [SDSS DR17 imaging fluxes and flux calibration](https://www.sdss4.org/dr17/algorithms/fluxcal/)

   The AB-offset and physical-flux sections state the approximate u shift, tentative z shift and their uncertainties. The tutorial freezes TOP 24 clean primary GALAXY spectra with zWarning=0 ordered by specObjID, keeping linear modelFlux values and inverse variances.

3. [Brammer, van Dokkum & Coppi, EAZY: A Fast, Public Photometric Redshift Code, arXiv:0807.1533v1 (2008)](https://arxiv.org/abs/0807.1533v1)

   Sections 2.1–2.5: template combinations, template error, priors and output redshifts. Section 3.3 and section 4, especially Figures 8–11, connect the error model and priors to real validation. Our finite SSP posterior is a separate teaching calculation.

4. [EAZY manual and source, commit f651e94d91042a84906090a66d2a3fec75504e21](https://github.com/gbrammer/eazy-photoz/tree/f651e94d91042a84906090a66d2a3fec75504e21)

   Read doc/eazy\_manual.tex alongside the pinned paper. Record the filter file indices and the selected definition of z\_peak or marginalized output before comparing catalogues. The course does not redistribute its HDFN catalogue or default templates.

5. [Euclid Collaboration, Improving redshift distribution reconstruction using a deep-to-wide transfer function, arXiv:2601.02005v1 (2026)](https://arxiv.org/abs/2601.02005v1)

   Sections 2.1–2.3 define transfer and selection; sections 3.1–3.2 define the Flagship 2.1 mock validation and color-cell calibration. Figures 5–6 compare distribution calibration. These are simulation tests, not a reproduced Euclid sky measurement.