EDP Sciences logo
Subscriber Authentication Point
Search
Menu
Home All issues Volume 711 (July 2026) A&A, 711 (2026) A19 Full HTML
Issue
A&A
Volume 711, July 2026
Euclid Quick Data Release (Q1)
Article Number A19
Number of page(s) 18
Section Extragalactic astronomy
DOI https://doi.org/10.1051/0004-6361/202554591
Published online 30 June 2026

© The Authors 2026

Licence Creative CommonsOpen Access article, published by EDP Sciences, under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

This article is published in open access under the Subscribe to Open model. This email address is being protected from spambots. You need JavaScript enabled to view it. to support open access publication.

1. Introduction

Supermassive black holes (MBH > 106 M⊙) and their host galaxies are believed to grow in tandem, as postulated by theoretical (e.g., Silk & Rees 1998) and observational (e.g., Magorrian et al. 1998) arguments. At the same time, early galaxy-evolution simulations showed that in order to reproduce observed galaxy size and brightness distributions, as well as their star formation, some form of energetic feedback from the central black hole is required (Bower et al. 2006; Croton 2006). Black holes can influence their host galaxies through the release of gravitational potential energy in the form of radiation during their active phase of accretion of matter, also known as active galactic nuclei (AGNs). In addition, some AGNs are known to power energetic jets and winds, which extend their influence on their host galaxy. Therefore, AGNs hold the missing piece needed to advance our knowledge of the black-hole-galaxy co-evolution (Fabian 2012; Heckman & Best 2014).

The lack of a detailed theory of AGN feedback leads to unrealistic models, failing to capture the observed complexity of the AGN population. For example, models do not reliably predict the ratio of unobscured (type 1; face-on view of the accretion disc) to obscured (type 2; edge-on view of the accretion disc and obscuring torus) AGNs, nor the evolution of their luminosity function (Whalen et al. 2020). Habouzit et al. (2021) presented six current state-of-the-art galaxy-evolution simulations compared to the number density of AGNs derived from X-ray observations. Even though these models are successful in predicting the observed properties of normal galaxies, none of them reliably predicts the history of supermassive-black-hole growth and the corresponding AGN phase. Thus, the creation of an AGN activity model rooted in observations is needed to act as ground truth for galaxy-evolution simulations and to motivate the prescription of stochastic processes in sub-grid physics.

The most significant challenge in AGN studies is that each part of the electromagnetic spectrum captures a different aspect of the central engine, leading to major inconsistencies between detection methods (Padovani et al. 2017). This is particularly true for the obscured AGN population, nowadays suspected to also be a phase during the evolution of an AGN (Georgakakis et al. 2009; Urrutia et al. 2012; Glikman et al. 2015; Kim & Im 2018; Stacey et al. 2022), and not only an outcome of a geometric alignment of the disc/torus system towards the observer (Rose et al. 2013; Almeida & Ricci 2017).

The unification scenario (Antonucci 1993; Urry & Padovani 1995) postulated that the observed variety in the presence of broad and narrow emission lines in the spectra of active galaxies was due to the obscuration induced by a molecular torus along the line of the sight of the observer. The distinction between AGNs and quasars is largely a description of the relative luminosity of the central engine and the host galaxy, with quasars being extremely luminous and dominating over the host-galaxy emission. The definition of red quasars corresponds to sources that show broad lines (i.e. type 1 sources) but with significant absorption in their continuum (Glikman et al. 2012; Banerji et al. 2013). Red quasars seem to contradict the AGN unification scenario; this was first shown by Klindt et al. (2019) and more recently Andonie et al. (2022), Fawcett et al. (2023), Petley et al. (2024), Calistro Rivera et al. (2024), and Yue et al. (2024) showed evidence of enhanced radio detection rates from this population, which is at odds with a simple orientation-induced obscuration. Current models and several observations argue that red quasars could be an initial, short-lived stage during the onset of quasar activity within a galaxy. As the gas and dust are driven into the centre of the galaxy, the initial phase of accretion is shrouded in a dusty cocoon. Subsequently, the radiation pressure and induced winds will clear out the region around the black hole revealing a type 1, unobscured, blue quasar (Urrutia et al. 2009; Banerji et al. 2012, 2015; Temple et al. 2019; Calistro Rivera et al. 2021).

Determining the physical parameters of quasars and AGNs and contrasting them with inactive galaxies (i.e. non-AGN hosts) as a function of luminosity, stellar mass, star formation rate, and obscuration, as well as across cosmic time and large-scale environments, is crucial (e.g. Wethers et al. 2018; Laloux et al., in prep.). It will enable the creation of an evolutionary scenario for galaxies including the incidence of AGNs, which is much needed to establish the true evolutionary path of quasars and to inform the recipes used in simulations. However, AGNs are short-lived phenomena, and detailed statistical studies of this population have been hindered by the size of the available datasets.

The value of large AGN and quasar samples lies not only in the accurate determination of AGN luminosity-distribution functions, but also in allowing for detailed decomposition into AGN sub-populations. A recent decomposition of ∼150 mid-infrared-detected AGNs, split into unobscured, red quasars, and type 2 sources, shows the luminosity distribution functions to be of a double-power-law form with a break at a characteristic luminosity (Glikman et al. 2018). Similar distributions are also found at other wavelengths (e.g., Fotopoulou et al. 2016). Much larger samples are needed to draw definitive conclusions; however these studies show already differences in the number density of red and blue quasars, or type 1 and type 2 quasars between low and high redshift (z ∼ 1).

Euclid is a mission of the European Space Agency (ESA), expected to detect billions of sources in the optical and near-infrared (Laureijs et al. 2011; Euclid Collaboration: Mellier et al. 2025). The Euclid observational campaign will observe a third of the extragalactic sky through two surveys. The Euclid Wide Survey (EWS), covering more than ∼14 000 deg2, and the Euclid Deep Survey (EDS), focusing on three different areas for a total of 63.1 deg2. With its unprecedented, large dataset, Euclid offers the possibility to study and identify the largest sample to date of unobscured and obscured AGNs across all wavebands, extending to the faintest sources. A key aspect will be the determination of a corresponding selection function, which remains under ongoing development in the context of AGN studies.

This paper presents the first systematic identification of red quasars within the Euclid Quick Release Q1 (2025), comparing mock and observed sources (described in Sect. 2) anddiscussing and refining colour-selection criteria that will be functional for the EWS. The analysis was performed independently using external and Euclid photometry, exploiting the Euclid NIR stand-alone capability in recovering the target population. Section 3 provides a detailed description of the multi-step approach adopted in this study, based on both colour-selection criteria and machine-learning techniques. Section 4 presents the main findings and the performance of the proposed selection function for red quasars. Section 5 discusses the implications of the results and outlines the directions for future analyses.

2. Dataset

In this work, we studied the photometric separation between mock red quasars and observed sources in a multi-dimensional colour space. We used a synthetic catalogue for the former and the COSMOS2020 dataset (Weaver et al. 2022) for the latter. The resulting selection function was then applied to the Euclid Q1 dataset (Euclid Quick Release Q1 2025). In the following, we describe each dataset, the cuts applied, and the methods used to match their photometry.

2.1. Mock red quasars

Mock red quasars are described over a range of wavelengths extending from the optical through the near-infrared (NIR) and up to the mid-infrared (MIR): DECam g, r, i, z in the optical (Dark Energy Survey Collaboration 2016), Euclid VIS (IE) and NISP YE, JE, HE in the NIR (Euclid Collaboration: Schirmer et al. 2022; Euclid Collaboration: Cropper et al. 2025; Euclid Collaboration: Jahnke et al. 2025), VISTA J, H, Ks from the NIR to MIR (McCracken et al. 2012), and WISE W1, W2 in the MIR (Wright et al. 2010). The collection of filters used in this work and their weighted central wavelengths is shown in Fig. 1. The employed photometry is expressed in the AB magnitude system.

Thumbnail: Fig. 1. Refer to the following caption and surrounding text. Fig. 1.

Set of passbands describing the photometry of mock red quasars. The panels show the transmission efficiencies for the broad-band optical DECam griz filters, the Euclid VIS (optical) and NISP (NIR) filters, and the NIR to MIR VISTA JHKs filters. The vertical dashed lines are placed at the weighted central wavelength of each filter.

To create the mock catalogue, we used the synthetic quasar spectrum constructed in Euclid Collaboration: Lusso et al. (2024), where a sample of quasars from SDSS DR7 (Shen et al. 2011) was selected across redshifts and grouped into nine bins of full width at half maximum (FWHM) of the Hβ emission line (see Table 1 in Euclid Collaboration: Lusso et al. 2024). Specifically, quasars were organised into three equivalent-width (EW) intervals ([15, 30], [30, 60], ≥ 60 Å) and three FWHM intervals ([2000, 3000], [3000, 5000], ≤5000 km/s), yielding a grid of nine spectral stacks. Stitching together the stacked spectra, the authors created a composite spectrum with very broad wavelength coverage.

Figure 2 shows the unobscured quasar spectrum and the dramatic impact of E(B − V) = 0.25 attenuation applied to it. The coloured bars correspond to the rest-frame wavelength coverage of the red-grism of Euclid. To create a mock catalogue of red quasars, we used the first spectral stack from Euclid Collaboration: Lusso et al. (2024) and the luminosity function of red quasars black determined in Glikman et al. (2018). The latter is described by a double-power-law function, which was already presented in Lacy et al. (2015); this is characterised by a faint-end and a right-end slope and a break luminosity, where the dominance shifts from the faint to the bright end.

Thumbnail: Fig. 2. Refer to the following caption and surrounding text. Fig. 2.

Spectrum comparison between unobscured quasar stack (dark blue line; Euclid Collaboration: Lusso et al. 2024) and a reddened quasar spectrum with E(B − V) = 0.25 (light blue line). Emission lines are highlighted using coloured vertical markers along with their respective labels, pointing to different atomic species and ionisation states. The coloured bars correspond to the redshift range of the red grism of NISP.

We created a grid of bolometric luminosity (40 < log10Lbol < 48, in erg s−1) and redshift (0 < z < 7) and calculated the expected number of red quasars by integrating the luminosity function. For each mock SED, we applied reddening according to their distribution covering 0.25 < E(B − V) < 1.45 and assuming the SMC attenuation law (Prevot et al. 1984) and intergalactic medium attenuation as described in Euclid Collaboration: Lusso et al. (2024). The chosen E(B − V) corresponds to the observed range in the sample of Glikman et al. (2018).

Finally, we applied an observed magnitude cut corresponding to the expected depth of EWS; i.e. J < 24.5. Figure 3 shows the coverage of the luminosity-redshift plane of our mock catalogue, assuming 14 500 deg2 sky coverage, at the wide-depth of Euclid. The black line shows the break luminosity of red quasars from Glikman et al. (2018).

Thumbnail: Fig. 3. Refer to the following caption and surrounding text. Fig. 3.

Luminosity-redshift plane for mock sample of red quasars. The black line is the knee of the luminosity function of Glikman et al. (2018).

2.2. COSMOS2020 sample

We used the COSMOS2020 dataset as our reference sample of observed sources. The catalogue includes multi-wavelength photometry spanning from the ultraviolet (UV) to the MIR, making it a suitable benchmark for assessing the separation of red quasars and other populations.

We cleaned the COSMOS2020 sample from corrupted photometry and fake detections and excluded the objects with null entries for the photometric redshift. To do this, we applied the mask FLAG_COMBINED = 0 to remove objects near bright stars and saturated regions, ACS_MU_CLASS = 1 to identify stars and remove them to retain galaxies, and lp_type ≠ 9 to exclude fake detections. Furthermore, we applied a magnitude upper cut at 23.5 in VISTA H in both COSMOS2020 and the mocks. Among the selected sources in the COSMOS2020 catalogue, we took into account those that are flagged as AGN candidates. The selected catalogue includes 95 052 objects.

Additionally, we excluded 1493 X-ray sources observed with Chandra (Civano et al. 2016) from our analysis. X-ray-selected AGNs outline a broad range of obscuration and their optical-to-NIR colours may not fully align with the selection criteria used in this study, which is based solely on red-quasar NIR photometry. We note that this sub-sample of X-ray AGNs constitutes a small fraction of the COSMOS2020 dataset and is not expected to dominate the target red quasar population. Its inclusion will be valuable for future works (see Sect. 5.6), as their properties can be used to refine the distinction between reddened AGNs and red galaxies.

2.3. Template fitting for COSMOS2020

Each object in COSMOS2020 is described by a photometric dataset including UV measures from GALEX, optical observations from the Subaru Hyper Suprime-Cam (HSC) and the Canada-France-Hawaii Telescope (CFHT), NIR data from VISTA, MIR data from the SPLASH programme of the Spitzer Space Telescope, and optical medium-band observations from Subaru. We refer the reader to Weaver et al. (2022) for references on these individual datasets. The properties of each filter are described in Table A.1.

The mock dataset includes VISTA, DECam, and Euclid VIS and NISP photometry. In order to match the photometric datasets of mocks and COSMOS2020, we used template fitting (TF) to estimate the optical and NIR fluxes that were originally not available in the latter. TF compares the input photometric dataset with a library of spectral energy distributions (SEDs) to identify the best match. For this task, we used the Phosphoros package (Paltani et al., in prep). Phosphoros is a fully Bayesian TF algorithm, supporting flexible prior distributions across allparameters (redshift, reddening, SED index, and luminosity) and producing multi-dimensional and marginalized posterior distributions. It was successfully employed in the Euclid photo-zchallenge, presented in Euclid Collaboration: Desprez et al. (2020), which was designed to evaluate the accuracy of various methods for photometric redshift estimation against the stringent requirements of Euclid cosmic shear analyses (Cropper et al. 2013, Tarsitano et al., in prep.). In Desprez et al. (2023), Phosphoros was validated and benchmarked against a similar code, Le Phare (Arnouts & Ilbert 2011). For each input galaxy, Phosphoros provides a multi-variate posterior distribution, allowing the inference of flux estimates in the missing bands from the best fit. For additional details about its metrics and models we refer the reader to Desprez et al. (2023) and Euclid Collaboration: Tucci et al. (2026).

2.4. Euclid Q1 dataset

We analysed mocks and COSMOS2020 (methods are described in Sect. 3) to derive a selection function for candidate red quasars, and we applied it to Euclid Q1 (Euclid Collaboration: Mellier et al. 2025). Q1 consists of a first visit of the Euclid Deep Fields (EDFs), spanning across a total area of 63.1 deg2 of the extragalactic sky; this area is divided among the Euclid Deep Field North (EDF-N, 20 deg2), Euclid Deep Field Fornax (EDF-F, 10 deg2), and the Euclid Deep Field South (EDF-S, 23 deg2). More details about the Q1 release are presented in Euclid Collaboration: Aussel et al. (2026), Euclid Collaboration: McCracken et al. (2026), Euclid Collaboration: Polenta et al. (2026) and Euclid Collaboration: Romelli et al. (2026).

In this work, we focused on the EDF-F, for which we find overlap with a collection of AGN candidates from Zou et al. (2022) – selected through the flag flag_IRagn_D12 – which follows the MIR colour-based cut proposed in Donley et al. (2012), and a catalogue of radio-selected quasars (Miller et al. 2013). These two datasets were used as external reference samples to independently assess the reliability of the selection function. In this work, we refer to them as control samples, as they provide an observational benchmark for AGNs, including potentially obscured sources. Their purpose is to evaluate whether known AGNs, selected via independent criteria, are successfully recovered by our method. This comparison is valuable for assessing the completeness of our selection function and flagging potential biases. We built a first dataset of 5 301 332 EDF-F sources, obtained by matching the Euclid morphology (Euclid Collaboration: Romelli et al. 2026) and photometric redshift (Euclid Collaboration: Tucci et al. 2026) catalogues. Then, we excluded objects flagged as spurious and with unphysical photometric and redshift properties. Furthermore, we only considered sources classified either as galaxies or quasars, according to the classification assigned in the Euclid photometric redshift pipeline (Euclid Collaboration: Tucci et al. 2026). A summary of this selection function is reported in Table 1. Additionally, we applied a cut near the limiting magnitude of the HE band, corresponding to 23.5. The final sub-sample at play counts 1 331 325 sources.

Table 1.

Sample selection function applied to the Q1 sample.

3. Methods

Colour-based selection criteria were extensively studied in the literature to identify AGNs and reddened AGNs. Most notably, Lacy et al. (2004), Stern et al. (2005), Mateos et al. (2012), and Assef et al. (2018) defined empirical cuts in MIR colour space. Alongside MIR-based selections, NIR colours have been proposed in Banerji et al. (2012) and Glikman et al. (2012, 2013, 2018), to separate red quasars from stars and galaxies.

To follow-up on these studies exploiting the unique depth and high-resolution of Euclid, we introduce a novel selection method that is solely based on Euclid NIR photometry. Our goal is to enhance the systematic identification of red quasars in the EWS where MIR photometry may be incomplete or unavailable.

To achieve this, we conducted a series of statistical analyses to study the photometric selection criteria that maximise the separation between the COSMOS2020-observed dataset and the mock red quasars. In particular, we explored the separation in a multi-dimensional colour-colour space, as a function of magnitude and redshift. Our study was performed independently using VISTA+DECam and reconstructed Euclid photometry. This allowed us to assess the impact of different wavelength coverage and filter sets on the identification of red quasars and to evaluate the Euclid NIR stand-alone capability of recovering our target population.

First, we collected features that are directly transferable into the reconstructed Euclid-like photometry. More precisely, we considered VISTA J − Ks, Y − Ks and J − H to describe NIR colours, and DECam i − Ks to sharpen the selection of reddened sources based on their optical-to-NIR transition. In the Euclid-like colour space, we used YE − HE, JE − HE, and YE − JE as NIR colours, and we adopted IE − HE as a metric to estimate the optical-to-NIR excess. We proceeded with the implementation of a multi-step analysis, which is detailed as follows:

  1. Principal component analysis (PCA): This method has been successfully applied in previous studies to investigate the underlying physical properties of AGNs and their host galaxies. Boroson & Green (1992) were the first to use it in the AGN domain, to analyse optical emission lines and continuum properties of a low-redshift quasar sample. Their work was followed by Corbin (1993), Laor et al. (1994, 1997), Wang et al. (1996), and Brandt & Boller (1998). Their analyses found correlation between the primary eigenvector (or principal component) and quasar spectral features depending on physical parameters including Eddington ratio, luminosity, and black-hole spin. Beyond optical emission lines, the PCA has been applied to AGN spectral energy distributions and photometric datasets. Yip et al. (2004) applied PCA on SDSS quasar spectra and showed that the eigenvectors (named eigenspectra in their work) have the power to disentangle the contribution of the host-galaxy light, the optical continuum, and the AGN emission. Kuraszkiewicz et al. (2009) analysed a sample of red 2MASS AGNs (Cutri et al. 2002). Among their results, they found that the second principal component was correlated with optical-to-infrared colours (B − Ks, B − R, J − Ks), depending on the contribution of the host galaxy relative to the AGN emission. PCA was also used to study AGN samples in Hao et al. (2005), Stern & Laor (2012), Panda & Śniegowska (2024). In this work, we applied a PCA to the aforementioned multi-dimensional colour space to identify the most informative colours that separate mock red quasars from the observed COSMOS2020 sources. Our analysis focuses on broadband photometric selection in preparation for systematic large-scale red quasar searches in the Euclid Survey. Through a linear combination of the original features, a PCA reduces dimensionality while capturing variance in the dataset. It serves as an exploratory framework to highlight which optical and NIR colours contribute most to the identification of red quasars. Detailed information follows in Sect. 3.1.

  2. Empirical colour-colour cuts: Using the most significant colours identified via the PCA, we studied a colour-colour selection function for red quasars. Additional information on the metric adopted to evaluate the selection performance is reported in Sect. 3.2.

  3. Machine learning-based refinement: In this phase, we trained a probabilistic random-forest classifier (RF; Breiman 2001) to refine the previous selection function for red quasars and mitigate the effects of contaminants. RF has the advantage of handling non-linear relationships between the input features, so it sets complex decision boundaries that a PCA and empirical colour-based cuts cannot capture. The trained RF model was then applied to the Euclid Q1 dataset, where we selected candidate red quasars based on their predicted probabilities. We refer readers to Sect. 3.3 for a detailed description of this method.

In the analysis, we considered the discriminating power of additional features, such as compactness criteria, and we used external AGN datasets as control samples. The latter play a crucial role in identifying the proposed selection function and assessing its robustness, or revealing the risk of introducing a bias that could reduce the completeness of the red quasar selection. We used our findings to build a first census of candidate red quasars in Euclid Q1 EDF-F. Such a sample provides a framework for future spectral analyses and will serve as a training set for artificial-intelligence-based automated classification, extended to the EDF-N, EDF-S, and EWS.

3.1. Hyper-colour determination and colour selection

A PCA is a statistical method that can be used to project a dataset from a high-dimensional space into a low-dimensional space, retaining its most meaningful properties. More precisely, the technique consists of mapping the original features into a new set of uncorrelated ones, named principal components. The result is a linear combination where each coefficient represents the contribution of the corresponding original feature in forming a principal component. With this technique, the dataset is linearly transformed onto a new coordinate system whose directions – the principal components – capture as much variance in the data as possible, with the first component capturing the most variance, the second capturing the next most, and so on.

In this work, we used the module PCA available in the Python library Scikit-learn (Pedregosa et al. 2011), which follows the implementation presented in Halko et al. (2009). Applying PCA decomposition to our multi-dimensional dataset, described by the aforementioned colour features, means collapsing it into a low-dimensional one, where each principal component (PC) is a linear combination of the original colours (c):

PC = ∑ j = 1 N a j c j , Mathematical equation: $$ \begin{aligned} \mathrm{PC} = \sum _{j = 1}^{N} a_{j} c_{j} , \end{aligned} $$(1)

with aj being the coefficients of the linear combination, and N, the number of involved colours. Features with consistently high coefficients across components explain a larger portion of the variance and are more important in the transformed space. The aim is to identify principal components receiving significant contributions by a set of original colours and study their discriminating power between red quasars and the rest of the sample. We refer to such principal components as hyper-colours (HCs).

We ran a PCA on the mocks and COSMOS2020, standardising the input features to ensure comparability across different scales. By assuming three principal components, we identified a cut in the HC space that guarantees the highest discriminating power. We refer to this as the HP cut.

Furthermore, we studied the impact that the single colour features, cj, have on the HCs, based on their linear coefficients aj, and we identified a cut in the multi-dimensional colour-colour space made by them. We refer to this selection cut as a colour-colour (CC) cut.

3.2. Forecast analysis

By applying the HP or CC cut, we estimate the completeness (C) and purity (P) of the selected sample. Completeness is the fraction of mock red quasars correctly identified by the proposed selection criterion and is defined as

C = TP TP + FN , Mathematical equation: $$ \begin{aligned} C \ = \frac{\mathrm {TP}}{\mathrm {TP} + \mathrm {FN}} , \end{aligned} $$(2)

where true positives (TPs) and false negatives (FNs) are the number of red quasars correctly identified and missed by the cut, respectively. The sum TP + FN thus corresponds to the total number of red quasars in the mock sample. Purity is defined as the fraction of TPs among all selected objects. We calculated it as

P = TP TP + FP , Mathematical equation: $$ \begin{aligned} P \ = \frac{\mathrm {TP}}{\mathrm {TP} + \mathrm {FP}} , \end{aligned} $$(3)

where false positives (FPs) are the number of sources incorrectly identified as red quasars by the selection cut. In this work, we define FPs as the number of COSMOS2020-selected sources that are not classified as AGNs, plus the number of AGN candidates passing the cut with DECam i − Ks < 1.7. We assumed this metric since 1.7 is the lower limit for optical-to-NIR excess in the mock sample. Finally, we applied the colour-based selection functions to EDF-F, and we studied them with the aid of two control samples of MIR-selected and radio-selected AGN candidates in the same field.

3.3. Probabilistic random forest

A RF is a machine-learning algorithm that consists of creating an ensemble of decision trees and combines their outputs to make predictions. Each tree in the forest is trained on a random subset of the data and features, and the final classification is done by majority voting (standard RF) or by averaging the predicted probabilities assigned to each class across all the decision trees (probabilistic RF). In a binary classification case, a probabilistic RF assigns each source two values, corresponding to their probability of belonging to each of the two classes. This approach allows us to estimate the confidence level of each classification and make a probabilistic selection of candidate red quasars. The RF method can identify and rank the most important features that differentiate red quasars from other objects, and it can handle complex and non-linear relationships between features. Furthermore, RF is more robust in terms of over-fitting, which makes it suitable for noisy or imbalanced datasets, as in the case of targeting red quasars (Breiman 2001; Chen et al. 2004). In our work, we used the module RandomForestClassifier available in the Python library Scikit-learn.

First, we trained a RF classifier on three sets of features (named S1, S2 and S3) based on different combinations of Euclid-based colours and magnitudes. More precisely, S1 includes the most significant colours identified by PCA, and S2 includes all the Euclid NIR colours. S3 includes both Euclid NIR colours and magnitudes. This multi-step approach allowed us to test the impact of expanding the primary feature set in terms of purity, completeness, and classification performance.

For each set of features, we identified the best model using hyper-parameter tuning. The RF model operates within a hyper-parameter space described by a set of key parameters, whose combinations can impact the performance of the model itself. These parameters include the number of trees, the minimum amount of samples required to split a tree node (i.e. a point in the tree where the data is split based on the value of a certain input feature), and the leaf size (i.e. the minimum number of samples required to be in a leaf, which is a terminal node making a final prediction). We used the Scikit-learn module RandomizedSearchCV to explore this hyper-parameter space and identify the best performing set of key parameters. For each sampled combination of parameters, the algorithm employs a stratified k-fold cross-validation strategy with k = 5 folds. This technique divides the dataset into five partitions and trains the model on four of them, using the fifth for validation. The goal of this approach is to achieve a robust estimate of the model performance and check if over-fitting occurs at certain partitions. The best combination of parameters is selected based on the cross-validation accuracy.

Through hyper-parameter tuning and cross-validation (run with a fixed random seed for reproducibility), we obtained an optimised RF model for each initial set of features (S1, S2, and S3). We compared the three models in terms of feature importance, completeness, and purity, and we applied the best one to the Euclid Q1 EDF-F dataset. We discuss the results and our findings in Sect. 4.

4. Results

The main goal of this work is to identify a robust sample of candidate red quasars in the Euclid Q1 dataset. As described in Sect. 3, we adopted a multi-step approach involving PCA, colour-colour cuts, and machine-learning classification. In this section, we present results of each step and evaluate their effectiveness in isolating the target population.

4.1. Selection function in the HC space

The PCA on mocks and COSMOS2020, run in VISTA-defined multi-dimensional colour space, provides three HCs whose coefficients are reported in Table A.2. HC1 is a weighted average of all the input colour features, with a slight emphasis on J − Ks and i − Ks. HC2 and HC3 are dominated by H − Ks and i − Ks. Their explained variances (91%, 6.2%, and 2.7%, respectively) indicate that HC1 is related to the overall colour gradients across the feature set, while HC2 and HC3 isolate information specific to a certain feature. The combination of HC1 and HC3 provides the strongest discriminating power, as displayed in Fig. 4. The contours, normalised to their respective subsets, show that this HC space is able to effectively disentangle the populations of mock red quasars from the observed sources in COSMOS2020. The latter includes both galaxies and AGN candidates, as described in Sect. 2. The HC cut of HC1 > 0.6 and HC3 > −0.9 leads to an overall completeness level of 98%, with 81% purity. Among the COSMOS2020 AGN candidates passing the cut, 91% of them have DECam i − Ks > 1.7. As reported in Sect. 3, based on mocks we considered this as a requirement to identify possible obscured AGNs and flag them as candidate red quasars.

Thumbnail: Fig. 4. Refer to the following caption and surrounding text. Fig. 4.

Hyper colour-colour diagram obtained using VISTA+DECam photometry, showing the separation of mock red quasars from COSMOS2020 sources. The latter include galaxies (gold contours) and AGN candidates (in purple). The contours are normalised to their respective subsets. The proposed selection function is displayed via the grey dashed lines.

We ran the PCA again, transferring the original VISTA colour features into the Euclid-like colour space, yielding similar results. The coefficients of the principal components are displayed in Table A.3. Analogously to the VISTA-based HCs, the Euclid-based HC1 is dominated by YE − HE, and HC3 obtains most contribution from IE − HE, tracking the optical-to-NIR transition. Figure 5 displays the Euclid-like HC space formed by HC1 and HC3. An HC cut of HC1 > 0.3 and HC3 > −0.9 leads to an overall completeness level of 97% with 68% purity.

Thumbnail: Fig. 5. Refer to the following caption and surrounding text. Fig. 5.

Hyper colour-colour diagram obtained using Euclid-like photometry, showing separation of mock red quasars from COSMOS2020 galaxies (gold contours) and AGN candidates (in purple). The proposed selection function is displayed via the grey dashed lines.

Beyond unveiling hyper-colours, the PCA highlights the original colour features, cj in Eq. (1), bringing the most weight in the identification of candidate red quasars. More precisely, in the VISTA parameter space, the most important colours for HC1, HC2 and HC3, based on their linear coefficients, are J − Ks, H − Ks, and i − Ks. In the Euclid-like parameter space, the most important features for the tree components are YE − HE, JE − HE, and IE − HE, respectively. We further advanced our analysis with a visual and quantitative assessment of the separation of mock red quasars using these multi-dimensional colour-colour spaces.

4.2. Photometric selection in colour-colour space

Figure 6 displays the separation between mock red quasars and the COSMOS2020 dataset achieved in a multi-dimensional parameter space spanning from optical to NIR wavelengths and defined through the most important colour features according to the PCA. The two populations are compared in the J − Ks versus H − Ks diagram, and COSMOS2020 is further split in two subsets to highlight candidate AGNs. If we apply the cut J − Ks > 0.8 and H − Ks > 0.2, we obtain an overall completeness level of 99% and a purity of 78%. We find that 85% of the candidate AGNs passing this multi-dimensional colour cut responds to the mock-calibrated red excess of DECam i − Ks > 1.7.

Thumbnail: Fig. 6. Refer to the following caption and surrounding text. Fig. 6.

VISTA+DECam colour-colour diagram showing separation of mock red quasars from COSMOS2020 galaxies (gold contours) and AGN candidates (in purple). The proposed selection function is displayed by the grey dashed lines. A subset of COSMOS2020 candidate AGNs overlays the selection, colour-coded as i − Ks.

As the CC cut is defined through optical and NIR colours, the purity estimate can be influenced by dependences on redshift and magnitude. We first estimated purity as a function of redshift. In Fig. 7, we show the separation between the COSMOS2020 sample and mock red quasars as a function of redshift, based on the most relevant features (J − Ks and i − Ks). We obtain 90% purity at z < 0.5, 69% at 0.5 < z < 1.5, and 87% at higher redshifts. The observed trend arises from colour degeneracies between red quasars and massive red galaxies. We refer the reader to Sect. 5.1 for a more detailed discussion of these effects. The magnitude dependency sees purity values of 77% for H < 20, 65% for 20 < H < 22 and 86% for objects in the interval 22 < H < 23.5.

Thumbnail: Fig. 7. Refer to the following caption and surrounding text. Fig. 7.

VISTA+DECam colour-z diagram showing the separation of mock red quasars from COSMOS2020. The colour-codes adopted for the displayed populations are as in Fig. 6.

We proceeded with the forecast analysis using the same metric, but defined with Euclid-like photometry. We studied a grid of multi-dimensional CC cuts to maximise the separation between the subsets at play, and we weighted the FP rate with the unreddened COSMOS2020 AGN candidates. The results are displayed in Fig. 8, with the colour-redshift evolution shown in Fig. 9. In this case, we find that the colour cut, YE − HE > 0.7 and JE − HE > 0.3 with IE − HE > 1.8, leads to an overall 99% completeness level and 67% purity, with redshift-dependent fluctuations: 86% purity at z < 0.5, 57% and 63% at 0.5 < z < 1.5 and higher redshifts, respectively. Purity values with magnitudes are 56% for HE < 20, 50% for 20 < HE < 22, and 77% for objects in the interval 22 < H < 23.5.

Thumbnail: Fig. 8. Refer to the following caption and surrounding text. Fig. 8.

Euclid-like colour-colour diagram showing separation of mock red quasars from COSMOS2020. The proposed colour-colour selection function is displayed via the grey dashed lines. The COSMOS2020 sample passing the selection is displayed in purple.

Thumbnail: Fig. 9. Refer to the following caption and surrounding text. Fig. 9.

Euclid-like colour-redshift diagram showing separation of mock red quasars (in red) from COSMOS2020 (golden contours). The COSMOS2020 passing the proposed CC cut is displayed in purple.

4.3. Colour-colour selection applied to Euclid Q1

Having established the methods for the selection of red quasars based on mock and observed training samples in the previous sections, we then applied these methods to the Euclid Q1 EDF-F dataset. The selection is supported by control samples; MIR-selected and radio-selected AGN candidates, introduced in Sect. 2. Figure 10 display the discriminating power of the multi-dimensional CC cut, yielding an overall 98% completeness level of the mock sample. The colour-redshift evolution is displayed in Fig. 11. The MIR-selected and radio-selected AGNs have 77% and 47% reddened sources, according to the proposed cut IE − HE > 1.8. Among them, 87% pass the multi-dimensional CC cut. The selection functions are summarised in Table 2. Combining the CC cut with the RF classification probability (described in Sect. 3), we flag 151 853 sources as candidate red quasars.

Thumbnail: Fig. 10. Refer to the following caption and surrounding text. Fig. 10.

Colour-colour diagram showing separation of mock red quasars from the Euclid Q1 sample. The proposed colour-colour selection is displayed via the grey dashed lines. The control samples are colour-coded by IE − HE; dots and diamonds represent MIR-selected and radio-selected AGNs, respectively.

Thumbnail: Fig. 11. Refer to the following caption and surrounding text. Fig. 11.

Colour-z diagram showing separation of mock red quasars from the Euclid Q1 sample. The proposed colour-colour selection is displayed by the grey dashed line. The control samples are colour-coded by IE − HE; dots and diamonds represent the MIR-selected and radio-selected AGNs, respectively. The contours are coloured as in Fig. 10.

Table 2.

Forecast red quasars’ completeness (C) and purity (P) from NIR selection functions.

4.4. Random-Forest analysis

In the previous steps of the analysis, we used a PCA to identify which regions in the hyper-colour space are most effective in distinguishing red quasars from other populations. Furthermore, using the most significant features based on PCA, we defined a three-dimensional colour-selection function, achieving high completeness levels and moderate purity (see Table 2).

In this final stage of the analysis, we aim to refine our selection using a probabilistic RF classifier. As discussed in Sect. 3.3, a RF can capture non-linear decision boundaries in the feature space, which are not detected by PCA or empirical CC cuts. This makes RF well-suited to increasing the purity of the selected sample of red quasars.

We applied a probabilistic RF classifier to identify candidate red quasars starting from a first set of photometric features, S1, including the most significant colours according to PCA: JE − HE, YE − HE, and IE − HE. The hyper-parameter search lead to an optimised model with 100 trees, minimum split size of 10, and minimum leaf size equal to 1. This configuration yields a mean cross-validation accuracy level of 95%, with a standard deviation of < 1%. We recover feature-importance values of 0.35 for JE − HE, 0.39 for YE − HE, and 0.26 for IE − HE.

Applying a probability threshold of P > 0.7 for classification, we obtain a 98% completeness level and 87% purity. We determined the probability threshold so that the completeness did not fall below the value we achieved using empirical CC cuts. This criterion ensures that the RF classification is at least as conservative as the other selection methods while improving purity.

To gather further insights from our analysis, we employed the uniform manifold approximation and projection algorithm (UMAP). As described in McInnes et al. (2018), the UMAP is a non-linear dimensionality-reduction technique that preserves both the local and global structure of the data and highlights possible clusters and patterns. We used UMAP to visualise the dataset in a reduced-dimensionality space and check the separation between classes.

The UMAP visualisation of the classified objects (Fig. 12), colour-coded by probability of being a red quasar, shows that a threshold of 0.7 outlines a boundary region between the two classes. One hundred sources randomly drawn from the test set are plotted over the map, with symbols following the empirical CC cut. Stars are used for sources classified as red quasars, and circles represent sources that did not pass the selection. The overlay displays a direct comparison between the empirical colour-selection method and the RF classification. Previously selected sources are mostly present in the region a with higher probability of being a red quasar and a minority populates the transition region. The RF refinement of these boundaries goes beyond the level of accuracy achieved by the empirical CC cut, thereby reducing contamination and raising purity.

Thumbnail: Fig. 12. Refer to the following caption and surrounding text. Fig. 12.

UMAP visualisation for the results of the probabilistic RF classifier run on the most significant colour features based on PCA. Hexagons are colour-coded by probability of being a candidate red quasar. The overlaid scatter plot displays a random subset with symbols reflecting the empirical CC cut. Stars represent objects previously classified as red quasars. Circles are employed otherwise.

We repeated the analysis on extended sets of features, specifically on S2 and S3 (introduced in Sect. 3.3), in order to assess the impact of additional information on the classification performance. More precisely, S2 includes YE − JE, JE − HE, YE − HE, and IE − HE, while S3 builds upon S2 by including in the set of features the magnitudes IE, YE, JE, and HE. The RF classifier optimised for S2 yields similar results, without improving completeness and purity. The feature-importance analysis assigns to the additional colour, YE − JE, a value of 0.03. Such results align with our findings from S3. In this case, the importance of the magnitude features is < 2%, and no increase in completeness and purity is registered.

Along this multi-step analysis, the dominant importance of the PCA-selected colours remains unchanged and leads to a consolidated probabilistic RF model that we applied to the Euclid Q1 EDF-F sample.

The UMAP visualisation of the Euclid Q1 sample, displayed in Fig. 13, is colour-coded by probability of being a red quasar. The over-plotted symbols represent a fraction (25%) of the two control samples: circles denote the MIR-selected AGN candidates, while diamonds represent the radio-selected AGN candidates. White markers indicate the sources that passed the previous empirical CC cut selection, and light grey colours those that were not classified as red quasars. Analogously to our findings from the analysis of the COSMOS2020 dataset, we notice that the spatial distribution of the sources is clustered around two different populations (according to their likelihood of being red quasars) and that the boundaries obtained via RF are more robust to contaminants, compared to the empirical CC cuts.

Thumbnail: Fig. 13. Refer to the following caption and surrounding text. Fig. 13.

UMAP visualisation of Euclid Q1 dataset, colour-coded by the RF-based probability of being a red quasar. The overlaid scatter plot displays 25% of the validation samples. MIR and radio-selected candidate AGNs are marked by circles and diamonds, respectively. The symbols are filled in white if the source passed the empirical CC cut; otherwise, they are coloured in light grey.

5. Discussion

In this section, we comment and expand on our findings presented above, discussing the performance of the proposed selection function, the refinement using probabilistic RF, and the added value of the Euclid data through morphological and colour analyses. We also discuss potential future extensions, including dual AGN identification and spectroscopical analyses.

5.1. Performance of the selection function

The VISTA-based HC spaces provide effective separation between mock red quasars and observed sources. Their translation into the Euclid-like HC space preserves the structure of the primary component (HC1), while introducing small shifts in the secondary components (HC2 and HC3). These shifts are expected due to the differences in filter characteristics and photometric uncertainties. The photometric selection function, using multi-dimensional colour cuts in the Euclid optical and NIR regimes, forecasts high levels of completeness (98%) and moderate purity (78%) for candidate red quasars. The completeness of the mocks remains consistently high across redshift and magnitude intervals, suggesting that the proposed selection effectively captures the reddened quasar population. Purity decreases in intermediate redshift ranges and at faint magnitudes, where physical degeneracies between red quasars and red galaxies become more significant. However, comparing redshift and magnitude distributions of the mock red quasars and the sources in COSMOS2020 suggests there is also an observational effect at play. Figure A.1 shows that the redshift distribution of mock red quasars peaks at higher values compared to the observed sources, as they are modelled to represent a deeper survey aiming to capture the properties of obscured AGNs. Purity estimates can be affected in the mismatched higher redshift range. The lower purity at intermediate redshift could be partially mitigated by training the selection function on larger, more representative datasets that include additional sources of variation, such as differences in host-galaxy properties or environmental factors. Future work could explore the inclusion of environmental parameters, such as local density or clustering, to refine selection criteria. Larger values of purity for bright objects is also influenced by the broader photometric extent of Euclid-like mocks, which outnumber COSMOS2020 at lower magnitudes (Fig. A.2). On the opposite side, fainter sources are more prone to contamination by non-AGN populations due to increasing photometric uncertainties, an aspect that potentially reduces purity at fainter magnitudes.

5.2. Considerations on the mock-based approach

In our analysis, we used a red quasar mock catalogue based on the stacking of SDSS quasar spectra with varying levels of reddening applied. While this approach effectively models the intrinsic quasar properties, it does not account for the host-galaxy contamination. The host-galaxy flux can dominate over the AGN emission in the optical and NIR bands, especially for lower redshift sources. The consequent dilution of quasar colours may reduce the effectiveness of our HC selection criteria, which are primarily optimised for quasar-dominated SEDs. Such aspects set the seed for future works, where our mock catalogue incorporates realistic galaxy templates, either by adding scaled galaxy SEDs to the quasar spectra or by using simulations informed by empirical measurements of quasar host properties. Nevertheless, our analysis still provides valuable insights into the potential of Euclid when it comes to identifying red quasars. As reported in Sect. 4, we also note that our results remain robust for brighter sources (H < 22), which are aligned closer to SDSS, where the impact of host galaxies is expected to be minimal.

5.3. Optimisation of the selection function with probabilistic RF

The probabilistic RF classifier refines the empirical CC cut by achieving higher levels of completeness and purity (98% and 88%, respectively). This improvement is due to the RF ability to integrate non-linear relationships in the multi-dimensional feature space, which are not detected with simpler colour-colour cuts.

A key aspect in our RF approach is that is starts from a set of features including the most significant colours, according to PCA, and then expand it with additional colours and magnitudes. Through feature-importance analysis, this multi-step methodology allows us to assess how the classification performance is impacted by the information encoded in the additional features. Furthermore, such expansion did not lead to substantial improvements in completeness and purity, confirming the NIR multi-dimensional colour space (JE − HE, YE − HE, and IE − HE) as the most informative and effective to classify out target population.

The UMAP visualisation of the datasets, colour-coded by probability of a source being a red quasar, shows that the threshold P > 0.7 effectively segments the transition region between the two classes. In comparison with the overlaid validation datasets, it shows consistency and robustness against contaminants.

5.4. Visual assessment of the morphology and environment of candidate red quasars

To further clarify our selection, we made an initial visual assessment of the morphological properties of the candidate red quasars. The visual inspection of the selected sources reveals that most candidates exhibit point-like or slightly extended profiles, consistent with AGN-dominated systems. The multi-wavelength images in the first three panels of Fig. 14 show sources with bright NIR emission and a weaker component in the VIS band, consistent with significant optical light attenuation due to dust. These characteristics suggest that the sources are indeed candidate red quasars. In this case, we also highlight the importance of this dataset for future morphological analyses, performing host galaxy decomposition and adding information on the properties of the AGN component and its dusty cocoon.

Thumbnail: Fig. 14. Refer to the following caption and surrounding text. Fig. 14.

Candidate red quasars. Panels show, from left to right, the RGB image and the HE, JE, YE, and IE bands. Raw images were obtained through ESA Datalabs (Navarro et al. 2024).

In addition to examples of individuals red quasar candidates, in the fourth panel of Fig. 14 we introduce an example of a candidate dual-AGN system. The image displays two distinct compact sources with optical-to-NIR colour IE − HE = 2.4, at redshifts of 0.86 (central source) and 0.9; these are classified as candidate red quasars with RF-based probabilities of 0.82 (central source) and 0.74. Their proximity consists of a projected distance of less then 100 kpc and a difference in redshift not surpassing 0.06. Such criteria align with the definition of a dual AGN given in earlierliterature (De Rosa et al. 2019). Such systems offer a unique observational window into AGN triggering mechanisms, galaxy mergers, and the evolution of supermassive black holes. The example presented in this work will be part of future systematic searches of dual-AGN systems. This effort will use morphological analysis and spectroscopic follow-up, where available. The results of this investigation will be presented in forthcoming publications.

5.5. Comparison with VISTA+DECam-based selection

We expanded this first characterisation of the selected sources, making a comparison between candidates selected through VISTA and DECam colours and via Euclid-only colours. In the EDF-F, we identified a sample of 43528 objects observed by Euclid, the Vista Hemisphere Survey (VHS, McMahon et al. 2019), and DES. Among them, we selected 3% and 4% using solely Euclid and VISTA+DECam colours, respectively.

As shown in Fig. 15, Euclid-only-selected objects extend into redder IE − HE values. This suggests that Euclid is better at identifying the reddest sources, which might be missed by VISTA. Furthermore, we observe that the candidates identified through the VISTA+DECam system display a broader distribution towards bluer colours. This can be explained by VISTA depth and resolution, which are both inferior to Euclid, leading to misclassification and missing reddened sources. This preliminary analysis suggests that Euclid’s better resolution and NIR sensitivity enables a more complete and robust identification of red quasars.

Thumbnail: Fig. 15. Refer to the following caption and surrounding text. Fig. 15.

Optical-to-NIR colour distributions for EDF-F sample of candidate red quasars observed through both Euclid and VISTA+DECam photometry. The orange and purple distributions are for candidates selected only via Euclid and VISTA+DECam colours, respectively.

Figure 16 displays two examples of visually inspected candidate red quasars identified with colour selection based on either Euclid or VISTA+DECam only. These sources illustrate typical trends we observed across the sample, including enhanced compactness and contrast in the Euclid bands. More precisely, the first row of Fig. 16 shows an example of a red-quasar candidate selected with Euclid only. The image in the IE band shows a compact source with a bright centre and some faint surrounding structure which can be attributed to the quasar host galaxy. The presence of asymmetry in the outer structure hints at a merger history or disturbed morphology. The Euclid HE-band image shows a smooth compact core with higher emission than in the optical. This proves that the optical-to-NIR contrast is high. The VISTA and DES images are noisier, and the source is harder to distinguish. The poorer signal can explain why the VISTA+DECam system did not classify this sources as a candidate red quasar. Images are normalised and shown in flux units.

Thumbnail: Fig. 16. Refer to the following caption and surrounding text. Fig. 16.

Examples of two candidate red quasars. The first and the second rows show an Euclid-only and VISTA+DECam-only-selected candidate, respectively. Panels from left to right display the object in the IE and HE filters, in DECam i and VISTA Ks.

We repeated the visual inspection on sources that were selected as candidate red quasars through the VISTA+DECam system only. In the second row of Fig. 16, the Euclid IE band displays an extended morphology with clear structures, possibly star-forming regions or satellite companions. In the DECam i band only the overall structure is less visible due to the lower resolution compared to Euclid. Similarly, in the VISTA bands the galaxy is less resolved and appears to be more concentrated in NIR. In this case, the lower resolutions of DECam and VISTA make the system look more compact, while Euclid sees a more extended morphology with detailed substructural features. Therefore, the VISTA+DECam colours used for the selection can be biased towards the integrated light of the system rather than its true morphology. This suggests that the Euclid-selected sample of red quasars tends to be cleaner, avoiding extended galaxies being misclassified as quasars.

5.6. Implications for future studies

In this work, we flagged over 150 000 sources in the EDF-F as candidate red quasars. This population serves as a first base for future refinements, incorporating additional diagnostics such as MIR data, spectral analyses (spectra were not considered in this work), and morphological compactness. Thesefeatures should help to further improve purity without compromising completeness. In terms of number expectations, we refer to the studies by Euclid Collaboration: Bisigello et al. (2024) and Euclid Collaboration: Selwood et al. (2025), which were conducted in anticipation of the Q1 data release. The former predicts a fraction of 57% AGNs among NIR-selected sources in EDF-F. The latter estimates an obscured fraction of 26% in the expected AGN population. Applying these forecasts to our EDF-F sample, we calculate 196 992 expected obscured AGNs. This is compatible with our selected sample, which is set to include reddened AGNs, a contamination of red galaxies, and red quasars. The classification of these sources will be the subject of future refinements based on the aforementioned criteria.

In this work, we excluded X-ray sources from the COSMOS2020 sample. However, their integration will be beneficial to future analyses, as their properties are directly linked to the degree of obscuration (Ma et al. 2024). This approach, which can be tested against the Q1 catalogue of X-ray AGN counterparts by Euclid Collaboration: Roster et al. (2026), can help improve the distinction between truly reddened AGNs from red galaxies, thereby improving the purity of the selected sample.

6. Conclusions

In this work, we explored the capability of selection criteria based on Euclid optical and NIR photometry to identify and characterise red quasars. We explored the separation in a multi-dimensional colour-colour space, as a function of magnitude and redshift, and we compared it with selections based on VISTA+DECam photometry. Through a joint PCA and statistical analysis we identified selection functions achieving high completeness levels and moderate purity. Then, to refine the identification of red quasars, we used a probabilistic RF classifier, which significantly improved purity (88%) and maintained a high level of completeness (98%). This result reflects the capability of this method to identify complex, non-linear relationships in a multi-dimensional colour space, beyond the reach of simple empirical cuts. Additionally, visualisation methods such as UMAP confirm a clear separation between red quasars and contaminants, reinforcing the robustness of our selectioncriteria.

Our analysis has shown that Euclid-based selection recovers redder quasar candidates than those identified by VISTA+DECam, as evidenced by shifts in the IE − HE distribution. This suggests that Euclid’s superior depth and resolution enable a more complete detection of highly obscured sources, which might be misclassified or missed in lower resolution surveys. Conversely, we found that some VISTA+DECam-selected sources appear as extended galaxies in Euclid images, indicating that lower resolution surveys may introduce contamination by compact galaxies in the sample of red quasars.

A key insight from future morphological inspection of the selected candidates will come from investigating the capability of Euclid to resolve host-galaxy features in a subset of sources, distinguishing between true quasars and compact galaxies.

We used our findings to build a first census of candidate red quasars in Euclid Q1. The catalogue will be released as a fits table to the Euclid Collaboration. This work provides the foundation for systematic red quasar searches in the Euclid Wide Survey (EWS). The proposed selection function will be further refined with additional diagnostics, including MIR data, spectroscopic follow-ups, and analysis of the host morphology. This first census of red-quasar candidates in Euclid Q1 represents a significant step towards a more complete understanding of the dusty AGN population and its connection to galaxy evolution.

Acknowledgments

This work has made use of the Euclid Quick Release (Q1) data from the Euclid mission of the European Space Agency (ESA), 2025, https://doi.org/10.57780/esa-2853f3b. The Euclid Consortium acknowledges the European Space Agency and a number of agencies and institutes that have supported the development of Euclid, in particular the Agenzia Spaziale Italiana, the Austrian Forschungsförderungsgesellschaft funded through BMK, the Belgian Science Policy, the Canadian Euclid Consortium, the Deutsches Zentrum für Luft- und Raumfahrt, the DTU Space and the Niels Bohr Institute in Denmark, the French Centre National d’Etudes Spatiales, the Fundação para a Ciência e a Tecnologia, the Hungarian Academy of Sciences, the Ministerio de Ciencia, Innovación y Universidades, the National Aeronautics and Space Administration, the National Astronomical Observatory of Japan, the Netherlandse Onderzoekschool Voor Astronomie, the Norwegian Space Agency, the Research Council of Finland, the Romanian Space Agency, the State Secretariat for Education, Research, and Innovation (SERI) at the Swiss Space Office (SSO), and the United Kingdom Space Agency. A complete and detailed list is available on the Euclid web site (www.euclid-ec.org). This work has made use of CosmoHub, developed by PIC (maintained by IFAE and CIEMAT) in collaboration with ICE-CSIC. CosmoHub received funding from the Spanish government (MCIN/AEI/10.13039/501100011033), the EU NextGeneration/PRTR (PRTR-C17.I1), and the Generalitat de Catalunya. This research makes use of ESA Datalabs (datalabs.esa.int), an initiative by ESA’s Data Science and Archives Division in the Science and Operations Department, Directorate of Science. The VISTA Hemisphere Survey data products served at Astro Data Lab are based on observations collected at the European Organisation for Astronomical Research in the Southern Hemisphere under ESO programme 179.A-2010, and/or data products created thereof. This work has benefited from the support of Royal Society Research Grant RGS\R1\231450. This research was supported by the International Space Science Institute (ISSI) in Bern, through ISSI International Team project #23- 573 “Active Galactic Nuclei in Next Generation Surveys”. F. R., B. L. acknowledge the support from the INAF Large Grant “AGN & Euclid: a close entanglement” Ob. Fu. 01.05.23.01.14.

References

  1. Almeida, C. R., & Ricci, C. 2017, Nat. Astron., 1, 679 [Google Scholar]
  2. Andonie, C., Alexander, D. M., Rosario, D., et al. 2022, MNRAS, 517, 2577 [NASA ADS] [CrossRef] [Google Scholar]
  3. Antonucci, R. 1993, ARA&A, 31, 473 [Google Scholar]
  4. Arnouts, S., & Ilbert, O. 2011, LePHARE: Photometric Analysis for Redshift Estimate, Astrophysics Source Code Library [record ascl:1108.009] [Google Scholar]
  5. Assef, R. J., Stern, D., Noirot, G., et al. 2018, ApJS, 234, 23 [Google Scholar]
  6. Banerji, M., McMahon, R. G., Hewett, P. C., et al. 2012, MNRAS, 427, 2275 [Google Scholar]
  7. Banerji, M., McMahon, R. G., Hewett, P. C., Gonzalez-Solares, E., & Koposov, S. E. 2013, MNRAS, 429, L55 [Google Scholar]
  8. Banerji, M., Alaghband-Zadeh, S., Hewett, P. C., & McMahon, R. G. 2015, MNRAS, 447, 3368 [Google Scholar]
  9. Boroson, T. A., & Green, R. F. 1992, ApJS, 80, 109 [Google Scholar]
  10. Bower, R. G., Benson, A. J., Malbon, R., et al. 2006, MNRAS, 370, 645 [Google Scholar]
  11. Brandt, N., & Boller, T. 1998, Astron. Nachr., 319, 163 [Google Scholar]
  12. Breiman, L. 2001, Mach. Learn., 45, 5 [Google Scholar]
  13. Calistro Rivera, G., Alexander, D. M., Rosario, D. J., et al. 2021, A&A, 649, A102 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  14. Calistro Rivera, G., Alexander, D. M., Harrison, C. M., et al. 2024, A&A, 691, A191 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  15. Chen, C., Liaw, A., & Breiman, L. 2004, Using Random Forest to Learn Imbalanced Data, Technical Report 666, Department of Statistics, UC Berkley [Google Scholar]
  16. Civano, F., Marchesi, S., Comastri, A., et al. 2016, ApJ, 819, 62 [Google Scholar]
  17. Corbin, M. R. 1993, ApJ, 403, L9 [Google Scholar]
  18. Cropper, M., Hoekstra, H., Kitching, T., et al. 2013, MNRAS, 431, 3103 [Google Scholar]
  19. Croton, D. J. 2006, MNRAS, 369, 1808 [NASA ADS] [CrossRef] [Google Scholar]
  20. Cutri, R. M., Nelson, B. O., Francis, P. J., & Smith, P. S. 2002, in IAU Colloq. 184: AGN Surveys, eds. R. F. Green, E. Y. Khachikian, & D. B. Sanders, ASP Conf. Ser., 284, 127 [NASA ADS] [Google Scholar]
  21. Dark Energy Survey Collaboration (Abbott, T., et al.) 2016, MNRAS, 460, 1270 [Google Scholar]
  22. De Rosa, A., Vignali, C., Bogdanović, T., et al. 2019, New A Rev., 86, 101525 [NASA ADS] [CrossRef] [Google Scholar]
  23. Desprez, G., Picouet, V., Moutard, T., et al. 2023, A&A, 670, A82 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  24. Donley, J. L., Koekemoer, A. M., Brusa, M., et al. 2012, ApJ, 748, 142 [Google Scholar]
  25. Euclid Collaboration (Desprez, G., et al.) 2020, A&A, 644, A31 [EDP Sciences] [Google Scholar]
  26. Euclid Collaboration (Schirmer, M., et al.) 2022, A&A, 662, A92 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  27. Euclid Collaboration (Bisigello, L., et al.) 2024, A&A, 691, A1 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  28. Euclid Collaboration (Lusso, E., et al.) 2024, A&A, 685, A108 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  29. Euclid Collaboration (Cropper, M., et al.) 2025, A&A, 697, A2 [Google Scholar]
  30. Euclid Collaboration (Jahnke, K., et al.) 2025, A&A, 697, A3 [Google Scholar]
  31. Euclid Collaboration (Mellier, Y., et al.) 2025, A&A, 697, A1 [Google Scholar]
  32. Euclid Collaboration (Selwood, M., et al.) 2025, A&A, 693, A250 [Google Scholar]
  33. Euclid Collaboration (Aussel, H., et al.) 2026, A&A, 711, A1 (Euclid Q1 SI) [Google Scholar]
  34. Euclid Collaboration (McCracken, H., et al.) 2026, A&A, 711, A2 (Euclid Q1 SI) [Google Scholar]
  35. Euclid Collaboration (Polenta, G., et al.) 2026, A&A, 711, A3 (Euclid Q1 SI) [Google Scholar]
  36. Euclid Collaboration (Romelli, E., et al.) 2026, A&A, 711, A4 (Euclid Q1 SI) [Google Scholar]
  37. Euclid Collaboration (Roster, W., et al.) 2026, A&A, 711, A16 (Euclid Q1 SI) [Google Scholar]
  38. Euclid Collaboration (Tucci, M., et al.) 2026, A&A, 711, A5 (Euclid Q1 SI) [Google Scholar]
  39. Euclid Quick Release Q1 2025, https://doi.org/10.57780/esa-2853f3b [Google Scholar]
  40. Fabian, A. C. 2012, ARA&A, 50, 455 [Google Scholar]
  41. Fawcett, V. A., Alexander, D. M., Brodzeller, A., et al. 2023, MNRAS, 525, 5575 [CrossRef] [Google Scholar]
  42. Fotopoulou, S., Buchner, J., Georgantopoulos, I., et al. 2016, A&A, 587, A142 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  43. Georgakakis, A., Clements, D. L., Bendo, G., et al. 2009, MNRAS, 394, 533 [Google Scholar]
  44. Glikman, E., Urrutia, T., Lacy, M., et al. 2012, ApJ, 757, 51 [Google Scholar]
  45. Glikman, E., Urrutia, T., Lacy, M., et al. 2013, ApJ, 778, 127 [Google Scholar]
  46. Glikman, E., Simmons, B., Mailly, M., et al. 2015, ApJ, 806, 218 [Google Scholar]
  47. Glikman, E., Lacy, M., LaMassa, S., et al. 2018, ApJ, 861, 37 [NASA ADS] [CrossRef] [Google Scholar]
  48. Habouzit, M., Li, Y., Somerville, R. S., et al. 2021, MNRAS, 503, 1940 [NASA ADS] [CrossRef] [Google Scholar]
  49. Halko, N., Martinsson, P. G., & Tropp, J. A. 2009, arXiv e-prints [arXiv:0909.4061] [Google Scholar]
  50. Hao, L., Strauss, M. A., Tremonti, C. A., et al. 2005, AJ, 129, 1783 [Google Scholar]
  51. Heckman, T. M., & Best, P. N. 2014, ARA&A, 52, 589 [Google Scholar]
  52. Kim, D., & Im, M. 2018, A&A, 610, A31 [EDP Sciences] [Google Scholar]
  53. Klindt, L., Alexander, D. M., Rosario, D. J., Lusso, E., & Fotopoulou, S. 2019, MNRAS, 488, 3109 [Google Scholar]
  54. Kuraszkiewicz, J., Wilkes, B. J., Schmidt, G., et al. 2009, ApJ, 692, 1180 [NASA ADS] [CrossRef] [Google Scholar]
  55. Lacy, M., Storrie-Lombardi, L. J., Sajina, A., et al. 2004, ApJS, 154, 166 [Google Scholar]
  56. Lacy, M., Ridgway, S. E., Sajina, A., et al. 2015, ApJ, 802, 102 [NASA ADS] [CrossRef] [Google Scholar]
  57. Laor, A., Fiore, F., Elvis, M., Wilkes, B. J., & McDowell, J. C. 1994, ApJ, 435, 611 [NASA ADS] [CrossRef] [Google Scholar]
  58. Laor, A., Fiore, F., Elvis, M., Wilkes, B. J., & McDowell, J. C. 1997, ApJ, 477, 93 [Google Scholar]
  59. Laureijs, R., Amiaux, J., Arduini, S., et al. 2011, arXiv e-prints [arXiv:1110.3193] [Google Scholar]
  60. Ma, Y., Goulding, A., Greene, J. E., et al. 2024, ApJ, 974, 225 [NASA ADS] [CrossRef] [Google Scholar]
  61. Magorrian, J., Tremaine, S., Richstone, D., et al. 1998, AJ, 115, 2285 [Google Scholar]
  62. Mateos, S., Alonso-Herrero, A., Carrera, F. J., et al. 2012, in Half a Century of X-ray Astronomy, eds. I. Georgantopoulos, & M. Plionis, 148 [Google Scholar]
  63. McCracken, H. J., Milvang-Jensen, B., Dunlop, J., et al. 2012, A&A, 544, A156 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  64. McInnes, L., Healy, J., & Melville, J. 2018, arXiv e-prints [arXiv:1802.03426] [Google Scholar]
  65. McMahon, R. G., Banerji, M., Gonzalez, E., et al. 2019, VizieR On-line Data Catalog: II/359. [Google Scholar]
  66. Miller, N. A., Bonzini, M., Fomalont, E. B., et al. 2013, ApJS, 205, 13 [NASA ADS] [CrossRef] [Google Scholar]
  67. Navarro, V., del Rio, S., Diego, M. A., et al. 2024, in ESA Datalabs: Digital Innovation in Space Science, ed. A. Cortesi(Singapore: Springer Nature Singapore), 1 [Google Scholar]
  68. Padovani, P., Alexander, D. M., Assef, R. J., et al. 2017, A&A Rev., 25, 2 [NASA ADS] [CrossRef] [Google Scholar]
  69. Panda, S., & Śniegowska, M. 2024, ApJS, 272, 13 [NASA ADS] [CrossRef] [Google Scholar]
  70. Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011, JMLR, 12, 2825 [Google Scholar]
  71. Petley, J. W., Morabito, L. K., Rankine, A. L., et al. 2024, MNRAS, 529, 1995 [NASA ADS] [CrossRef] [Google Scholar]
  72. Prevot, M. L., Lequeux, J., Maurice, E., Prevot, L., & Rocca-Volmerange, B. 1984, A&A, 132, 389 [Google Scholar]
  73. Rose, M., Tadhunter, C. N., Holt, J., & Rodríguez Zaurín, J. 2013, MNRAS, 432, 2150 [Google Scholar]
  74. Shen, Y., Richards, G. T., Strauss, M. A., et al. 2011, ApJS, 194, 45 [Google Scholar]
  75. Silk, J., & Rees, M. J. 1998, A&A, 331, L1 [NASA ADS] [Google Scholar]
  76. Stacey, H. R., Costa, T., McKean, J. P., et al. 2022, MNRAS, 517, 3377 [NASA ADS] [CrossRef] [Google Scholar]
  77. Stern, J., & Laor, A. 2012, MNRAS, 423, 600 [NASA ADS] [CrossRef] [Google Scholar]
  78. Stern, D., Eisenhardt, P., Gorjian, V., et al. 2005, ApJ, 631, 163 [Google Scholar]
  79. Temple, M. J., Banerji, M., Hewett, P. C., et al. 2019, MNRAS, 487, 2594 [Google Scholar]
  80. Urrutia, T., Becker, R. H., White, R. L., et al. 2009, ApJ, 698, 1095 [Google Scholar]
  81. Urrutia, T., Lacy, M., Spoon, H., et al. 2012, ApJ, 757, 125 [Google Scholar]
  82. Urry, C. M., & Padovani, P. 1995, PASP, 107, 803 [NASA ADS] [CrossRef] [Google Scholar]
  83. Wang, T., Brinkmann, W., & Bergeron, J. 1996, A&A, 309, 81 [NASA ADS] [Google Scholar]
  84. Weaver, J. R., Kauffmann, O. B., Ilbert, O., et al. 2022, ApJS, 258, 11 [NASA ADS] [CrossRef] [Google Scholar]
  85. Wethers, C. F., Banerji, M., Hewett, P. C., et al. 2018, MNRAS, 475, 3682 [Google Scholar]
  86. Whalen, K. E., Hickox, R. C., DiPompeo, M. A., Richards, G. T., & Myers, A. D. 2020, ApJ, 888, 71 [Google Scholar]
  87. Wright, E. L., Eisenhardt, P. R. M., Mainzer, A. K., et al. 2010, AJ, 140, 1868 [Google Scholar]
  88. Yip, C. W., Connolly, A. J., Vanden Berk, D. E., et al. 2004, AJ, 128, 2603 [Google Scholar]
  89. Yue, B. H., Best, P. N., Duncan, K. J., et al. 2024, MNRAS, 529, 3939 [NASA ADS] [CrossRef] [Google Scholar]
  90. Zou, F., Brandt, W. N., Chen, C.-T., et al. 2022, ApJS, 262, 15 [NASA ADS] [CrossRef] [Google Scholar]

Appendix A: Supplementary information

Thumbnail: Fig. A.1. Refer to the following caption and surrounding text. Fig. A.1.

Redshift distributions of the mock and the COSMOS2020 datasets. For the latter, we consider the photometric redshift calculated with LePhare and available in The Classic catalogue.

Thumbnail: Fig. A.2. Refer to the following caption and surrounding text. Fig. A.2.

VISTA H magnitude distributions of the mock and the COSMOS2020 datasets.

Table A.1.

Characteristics of filters in COSMOS2020 used for template fitting.

Table A.2.

Principal Component Coefficients for each VISTA HC.

Table A.3.

Principal component coefficients for each Euclid-like HC.

All Tables

Table 1.

Sample selection function applied to the Q1 sample.

In the text
Table 2.

Forecast red quasars’ completeness (C) and purity (P) from NIR selection functions.

In the text
Table A.1.

Characteristics of filters in COSMOS2020 used for template fitting.

In the text
Table A.2.

Principal Component Coefficients for each VISTA HC.

In the text
Table A.3.

Principal component coefficients for each Euclid-like HC.

In the text

All Figures

Thumbnail: Fig. 1. Refer to the following caption and surrounding text. Fig. 1.

Set of passbands describing the photometry of mock red quasars. The panels show the transmission efficiencies for the broad-band optical DECam griz filters, the Euclid VIS (optical) and NISP (NIR) filters, and the NIR to MIR VISTA JHKs filters. The vertical dashed lines are placed at the weighted central wavelength of each filter.
In the text
Thumbnail: Fig. 2. Refer to the following caption and surrounding text. Fig. 2.

Spectrum comparison between unobscured quasar stack (dark blue line; Euclid Collaboration: Lusso et al. 2024) and a reddened quasar spectrum with E(B − V) = 0.25 (light blue line). Emission lines are highlighted using coloured vertical markers along with their respective labels, pointing to different atomic species and ionisation states. The coloured bars correspond to the redshift range of the red grism of NISP.
In the text
Thumbnail: Fig. 3. Refer to the following caption and surrounding text. Fig. 3.

Luminosity-redshift plane for mock sample of red quasars. The black line is the knee of the luminosity function of Glikman et al. (2018).
In the text
Thumbnail: Fig. 4. Refer to the following caption and surrounding text. Fig. 4.

Hyper colour-colour diagram obtained using VISTA+DECam photometry, showing the separation of mock red quasars from COSMOS2020 sources. The latter include galaxies (gold contours) and AGN candidates (in purple). The contours are normalised to their respective subsets. The proposed selection function is displayed via the grey dashed lines.
In the text
Thumbnail: Fig. 5. Refer to the following caption and surrounding text. Fig. 5.

Hyper colour-colour diagram obtained using Euclid-like photometry, showing separation of mock red quasars from COSMOS2020 galaxies (gold contours) and AGN candidates (in purple). The proposed selection function is displayed via the grey dashed lines.
In the text
Thumbnail: Fig. 6. Refer to the following caption and surrounding text. Fig. 6.

VISTA+DECam colour-colour diagram showing separation of mock red quasars from COSMOS2020 galaxies (gold contours) and AGN candidates (in purple). The proposed selection function is displayed by the grey dashed lines. A subset of COSMOS2020 candidate AGNs overlays the selection, colour-coded as i − Ks.
In the text
Thumbnail: Fig. 7. Refer to the following caption and surrounding text. Fig. 7.

VISTA+DECam colour-z diagram showing the separation of mock red quasars from COSMOS2020. The colour-codes adopted for the displayed populations are as in Fig. 6.
In the text
Thumbnail: Fig. 8. Refer to the following caption and surrounding text. Fig. 8.

Euclid-like colour-colour diagram showing separation of mock red quasars from COSMOS2020. The proposed colour-colour selection function is displayed via the grey dashed lines. The COSMOS2020 sample passing the selection is displayed in purple.
In the text
Thumbnail: Fig. 9. Refer to the following caption and surrounding text. Fig. 9.

Euclid-like colour-redshift diagram showing separation of mock red quasars (in red) from COSMOS2020 (golden contours). The COSMOS2020 passing the proposed CC cut is displayed in purple.
In the text
Thumbnail: Fig. 10. Refer to the following caption and surrounding text. Fig. 10.

Colour-colour diagram showing separation of mock red quasars from the Euclid Q1 sample. The proposed colour-colour selection is displayed via the grey dashed lines. The control samples are colour-coded by IE − HE; dots and diamonds represent MIR-selected and radio-selected AGNs, respectively.
In the text
Thumbnail: Fig. 11. Refer to the following caption and surrounding text. Fig. 11.

Colour-z diagram showing separation of mock red quasars from the Euclid Q1 sample. The proposed colour-colour selection is displayed by the grey dashed line. The control samples are colour-coded by IE − HE; dots and diamonds represent the MIR-selected and radio-selected AGNs, respectively. The contours are coloured as in Fig. 10.
In the text
Thumbnail: Fig. 12. Refer to the following caption and surrounding text. Fig. 12.

UMAP visualisation for the results of the probabilistic RF classifier run on the most significant colour features based on PCA. Hexagons are colour-coded by probability of being a candidate red quasar. The overlaid scatter plot displays a random subset with symbols reflecting the empirical CC cut. Stars represent objects previously classified as red quasars. Circles are employed otherwise.
In the text
Thumbnail: Fig. 13. Refer to the following caption and surrounding text. Fig. 13.

UMAP visualisation of Euclid Q1 dataset, colour-coded by the RF-based probability of being a red quasar. The overlaid scatter plot displays 25% of the validation samples. MIR and radio-selected candidate AGNs are marked by circles and diamonds, respectively. The symbols are filled in white if the source passed the empirical CC cut; otherwise, they are coloured in light grey.
In the text
Thumbnail: Fig. 14. Refer to the following caption and surrounding text. Fig. 14.

Candidate red quasars. Panels show, from left to right, the RGB image and the HE, JE, YE, and IE bands. Raw images were obtained through ESA Datalabs (Navarro et al. 2024).
In the text
Thumbnail: Fig. 15. Refer to the following caption and surrounding text. Fig. 15.

Optical-to-NIR colour distributions for EDF-F sample of candidate red quasars observed through both Euclid and VISTA+DECam photometry. The orange and purple distributions are for candidates selected only via Euclid and VISTA+DECam colours, respectively.
In the text
Thumbnail: Fig. 16. Refer to the following caption and surrounding text. Fig. 16.

Examples of two candidate red quasars. The first and the second rows show an Euclid-only and VISTA+DECam-only-selected candidate, respectively. Panels from left to right display the object in the IE and HE filters, in DECam i and VISTA Ks.
In the text
Thumbnail: Fig. A.1. Refer to the following caption and surrounding text. Fig. A.1.

Redshift distributions of the mock and the COSMOS2020 datasets. For the latter, we consider the photometric redshift calculated with LePhare and available in The Classic catalogue.
In the text
Thumbnail: Fig. A.2. Refer to the following caption and surrounding text. Fig. A.2.

VISTA H magnitude distributions of the mock and the COSMOS2020 datasets.
In the text

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.