FOMO-Z uses galaxy colors to estimate redshifts and classify populations

Learning from reference galaxies

How much can astronomers infer about a galaxy from its brightness through different filters? FOMO-Z, a new open-source Python package presented by G. F. Lesci, addresses that question using reference galaxies with known properties. It supports galaxy selection, estimates probabilities for individual objects, and reconstructs distributions for entire populations.

The study is an arXiv preprint, first submitted on October 29, 2026, and revised on November 6. Its listing says it was submitted to Astronomy & Astrophysics; it does not establish acceptance or completed peer review.

The motivation is practical. Surveys including Euclid, Rubin and Roman will observe billions of galaxies, but spectroscopy will cover only a small fraction. Estimates derived from imaging therefore matter, as does understanding their uncertainty. For cosmological analyses, researchers also need the underlying redshift distribution of a selected population, not merely an estimated redshift for each galaxy.

Two ways to select a sample

FOMO-Z works with colors: differences between magnitudes measured in different filters. Its reference sample connects those observed colors to properties such as redshift or star-formation activity, without assuming a particular mathematical shape for the color distributions.

One approach calibrates straight-line boundaries in diagrams comparing two colors. This is especially useful for selecting galaxies behind a foreground object for weak gravitational lensing measurements. Another divides color–color diagrams into grids, allowing more flexible selection regions where enough reference information exists.

Both approaches seek to recover more target galaxies while meeting a chosen minimum purity. Completeness describes the fraction of target galaxies recovered; purity describes the fraction of selected galaxies that actually belong to the target population. Improving one can come at the expense of the other.

In comparisons using the COSMOS2020 catalog, the linear selections achieved equal or higher completeness than four published selections when matched to their filter sets, redshift thresholds and measured purity. This is a result under those comparison conditions, not proof of universal superiority.

From classifications to probabilities

Rather than assigning only a yes-or-no label, FOMO-Z can find reference galaxies whose colors match a test galaxy within tolerances set by their photometric uncertainties. Their weighted memberships then determine the estimated probability that the test galaxy meets a specified criterion.

A feature-selection procedure removes colors that contribute little additional discrimination. This matters because matching must succeed in every retained usable color: redundant constraints can unnecessarily reduce the available reference matches.

For continuous properties such as redshift, the software estimates the probability of lying below successive thresholds. It fits a smooth, non-decreasing cumulative distribution and differentiates it to obtain a probability density. The resulting density is normalized over the sampled interval.

The individual-redshift demonstration used COSMOS2020 as both reference and test sample. Agreement with catalog medians and uncertainty intervals therefore measures consistency with the reference information, not predictive accuracy on independent galaxies. Those examples also used reference redshift point estimates rather than propagating their full uncertainties, although the framework supports reference redshift probability distributions.

What the population tests showed

For quiescent and star-forming galaxies, FOMO-Z used COSMOS2015 reference labels and was applied to COSMOS-Web galaxies. Its probability patterns broadly resembled those from an existing gradient-boosting classifier, but with a sharper transition between populations. Lesci explicitly cautions that sharper separation does not establish greater accuracy; that requires independent, reliable labels.

For population reconstruction, each contributing test galaxy distributes its assigned weight among compatible reference galaxies. Normalizing that transfer prevents objects with many reference matches from contributing more simply because more analogs are available.

Applied to the KiDS-Legacy shear sample with the spectroscopic KiDZ compilation as reference, the method produced broadly similar redshift distributions to self-organizing maps across six photometric-redshift bins. The largest visible discrepancy was in the highest-redshift bin. Inputs and weighting differed between methods, and neither comparison distribution included simulation-based bias corrections, so the differences cannot be attributed to reconstruction algorithms alone.

What remains uncertain

Reliability depends on representative reference galaxies, adequate color coverage and consistently calibrated photometry. Current matching does not model covariance between colors sharing a filter. Some demonstration uncertainties were approximated or imputed, introducing further qualifications.

The present framework uses colors as inference features; adding magnitudes and sizes is future work. Its demonstrations support a flexible, inspectable approach to galaxy analysis, but further accuracy and calibration tests are needed before drawing stronger conclusions about precision cosmological applications.