betaFeedback

WorksPhysics and Astronomy

Towards the LSST DESC pipeline for cosmology with galaxy clusters – design and validation on simulations

Abstract

We present the current implementation of an automated framework developed within the Legacy Survey of Space and Time Dark Energy Science Collaboration (LSST DESC) for galaxy cluster cosmology, in preparation for the first LSST data releases of the NSF-DOE Vera C. Rubin Observatory. The cluster cosmology pipeline integrates DESC-developed tools dedicated to galaxy/cluster catalog ingestion, cluster number count and weak lensing measurements, modeling, and parameter inference through posterior sampling, enabling reproducible end-to-end cluster cosmology analyses. We test and validate the pipeline using synthetic mock datasets and the LSST DESC cosmoDC2 simulation, performing a joint analysis of redMaPPer cluster abundance and stacked weak lensing profiles measured in bins of redshift and richness. We constrain two cosmological parameters, ΩcΩ_c and σ8σ_8, along with six cluster mass-observable relation parameters. Our baseline analysis recovers the fiducial cosmoDC2 values of ΩcΩ_c and σ8σ_8 within 0.51σ0.51σ and 0.34σ0.34σ, respectively. We also showcase the flexibility of the overall pipeline by assessing the impact of fiducial cosmology and alternative modeling prescriptions on the parameter constraints. Lastly, we also discuss the current limitations of the pipeline and the developments needed for LSST data.

Introduction

Galaxy clusters are hosted by the most massive dark matter halos, which form through the gravitational collapse of large matter overdensities and are the most massive gravitationally bound structures in the Universe (Kravtsov & Borgani 2012). Their mass function and spatial distribution are sensitive to both the expansion history and the growth of large-scale structure over cosmic timescales, making galaxy clusters a powerful cosmological probe complementary to other probes such as the cosmic microwave background, Type Ia supernovae, baryon acoustic oscillations, and weak gravitational lensing. This sensitivity enables constraints on cosmological parameters and the nature of dark energy and dark matter in the Universe (see Allen et al. 2011; Weinberg et al. 2013 for reviews). The primary cosmological probe using galaxy clusters is their abundance in the Universe, which has been used in cosmological analyses of major surveys including Planck (Ade et al. 2016), SPT (de Haan et al. 2016; Bocquet et al. 2019), eROSITA (Ghirardini et al. 2024), DES (Abbott et al. [1]; Abbott et al. 2025), HSC (Chiu et al. 2024), and KiDS (Lesci et al. 2022; Lesci et al. 2025).

One of the dominant limiting factors in optical cluster cosmology is the combined uncertainty and scatter of the mass-observable relation (MoR) (Allen et al. 2011), which relates cluster halo mass to observable proxies, as halo mass cannot be directly measured and must instead be inferred (e.g., Murata et al. 2018; Payerne et al. 2025). As a result, cluster abundance alone provides limited constraints on cosmological parameters, due to strong degeneracies with the scaling-relation parameters. Cluster weak gravitational lensing, in which the projected images of distant galaxies are sheared by the gravitational field of foreground clusters, provides a nearly unbiased mass calibration that helps break this degeneracy (Murata et al. 2019; Abbott et al. 2025). However, there are other limiting factors to this approach that may bias the scaling relation, such as the correlation between cluster observable and the cluster weak lensing profile (e.g., Zhang et al. 2024). Other combinations are also being explored, for instance adding cluster clustering and the galaxy two-point statistics in a joint 6×2pt+N6 \times2\mathrm{pt}+\mathrm{N} analysis ([85]). The joint analysis of cluster abundance and cluster weak lensing is one of the standard approaches adopted by the community to constrain cosmological parameters (Mantz et al. 2015; Chiu et al. 2023; Bocquet et al. 2024; Fumagalli et al. 2024; Ghirardini et al. 2024). Carrying it out in a reliable, reproducible, and automated way requires a dedicated analysis pipeline, modular enough to accommodate evolving approaches, a need that becomes especially pressing in the context of the upcoming Stage-IV surveys.

We are transitioning from the successful Stage-III galaxy surveys such as the Dark Energy Survey (DES, [1, 85] (Costanzi et al. 2021; Abbott et al. 2025)), the Hyper Suprime-Cam survey (HSC, Murata et al. 2019; Sunayama et al. 2024; Chiu et al. 2024) and the Kilo-Degree Survey (KiDS, Lesci et al. 2022), to a new era of Stage-IV lensing surveys with the Legacy Survey of Space and Time (LSST, LSST Dark Energy Science Collaboration 2012), the ESA satellite Euclid (Euclid Collaboration et al. 2025), and NASA’s Nancy Grace Roman Space Telescope (Bailey et al. 2023). These next-generation surveys will deliver unprecedentedly large galaxy samples, enabling the detection of tens of thousands of clusters and high-precision weak gravitational lensing measurements. This work focuses on LSST, which will map 18,000 square degrees of the southern sky over 10 years, allowing measurements of shapes and redshifts of a few billion galaxies up to z∼3z \sim3. LSST will permit the detection and mapping of the distribution of ∼100 000\sim100\,000 galaxy clusters, as well as characterizing their dark matter content through lensing ([87] (LSST Dark Energy Science Collaboration 2012)).

The LSST Dark Energy Science Collaboration (LSST-DESC, LSST Dark Energy Science Collaboration 2012) is preparing for the cosmological analysis of LSST data, and has been developing a suite of software tools covering all stages of the analysis chain, from catalog construction and weak lensing measurement to theoretical modeling and parameter inference. Cluster constraints will eventually be combined with the other DESC probes, which requires the cluster analysis to be built on the same tools and calibrations as the rest of the collaboration, so that systematics and covariances are treated consistently across probes. Combining these tools into a validated, end-to-end framework is one of the key challenges in preparing for LSST Data Release 1 (DR1). The significantly reduced statistical uncertainties of LSST compared to Stage-III surveys place stringent requirements on the control of systematic effects in galaxy cluster cosmology. Key examples include galaxy shape calibration (Sheldon & Huff 2017; McClintock et al. 2019), selection effects (Aguena & Lima 2018; Rykoff et al. 2014), and photometric redshift uncertainties in cluster weak lensing (Applegate et al. 2014; Payerne et al. 2025), all of which become more critical in the LSST regime. While these effects will need to be addressed progressively as LSST data arrive, a necessary first step is to establish a validated baseline: a working end-to-end analysis framework that correctly combines all the relevant DESC tools at their current version, spanning data ingestion, data vector generation, theoretical modeling, covariance estimation, and Bayesian inference. Each stage must be benchmarked to provide a reliable baseline for future analyses. Within DESC, individual tools have already been presented and validated in dedicated code papers, including CCL (Chisari et al. 2019), CLMM (Aguena et al. 2021, 2026), TXPipe (Prat et al. 2023), RAiL (van den Busch et al. 2026) and Augur (Rogozenski et al. 2026), but no cluster cosmology analysis validating the DESC tools as an integrated system has yet been carried out.

In this work, we introduce CLPipe1, the current implementation of the LSST-DESC Cluster Pipeline, designed to unify these DESC tools into a single automated and reproducible analysis chain, in which individual components can be swapped or extended as the analysis evolves, minimizing human error and ensuring reproducibility. We demonstrate its capability to perform a cosmological analysis, from catalog-level data to parameter inference using the simulated cosmoDC2 extragalactic catalog (Korytov et al. 2019; Abolfathi et al. 2021). To this end, we analyze the abundance and lensing of redMaPPer-detected clusters in cosmoDC2 (Rykoff et al. 2014). This work constitutes the first end-to-end cluster cosmology analysis performed entirely within the LSST-DESC Cluster Pipeline framework, establishing a validated baseline in preparation for the first cosmological data release of LSST.

This paper is organized as follows. § 2 presents the ingredients needed for a counts and weak lensing cluster analysis. § 3 presents each stage of the LSST-DESC Cluster Pipeline (connectivity to other DESC tools and performance). We validate the pipeline inference framework with mocks in § 4. § 5 presents the analysis of the cosmoDC2 simulated dataset that is used to test and validate the pipeline. In § 6 we discuss our results and perspectives.

Cosmology with galaxy clusters

Extracting cosmological constraints from a cluster catalog requires connecting theory predictions to the estimated data vectors. In this work, we consider the combination of cluster abundance and cluster weak lensing. We describe the general ingredients and steps of this joint analysis below, providing the background needed to understand the construction of the pipeline presented in § 3.

The analysis starts with a cluster catalog containing a set of properties for each cluster, in particular its redshift and an observable that serves as a mass proxy. The abundance is then the distribution of the detected clusters over this mass proxy–redshift plane. Its prediction depends on the survey volume and halo mass function (e.g., (Tinker et al. 2008; Despali et al. 2016)), with the mass–observable relation connecting the theoretical halo mass to the observed mass proxy (e.g., (Giodini et al. 2013)). This prediction may also account for the selection function, describing the incomplete and impure detections of clusters (e.g., (Aguena & Lima 2018)), and for other systematics such as projection effects (e.g., (Costanzi et al. 2019; Sunayama et al. 2020)) and cluster photometric redshift uncertainties.

We then estimate the cluster weak lensing data vectors, the cluster shear profiles, which provide a complementary probe to cluster counts. Galaxy clusters act as lenses for the galaxies behind them, distorting their observed shapes, such that the measured average tangential shear of background galaxies around clusters provides information on their underlying mass distribution ([13] (Schneider 2005)). Individual cluster shear profiles may be dominated by noise from intrinsic ellipticity of source galaxies (shape noise) ((Hoekstra 2003; Wu et al. 2019)), and thus they can be stacked over samples of clusters to increase the signal-to-noise ratio ((Mandelbaum et al. 2005)). Modeling such profiles requires a description of the matter density around halos ((Navarro et al. 1997; Aguena et al. 2021)), which can include systematic effects such as cluster miscentering ((Zhang et al. 2023)) and the contamination of the source sample by cluster member galaxies ((Applegate et al. 2014)).

Once the abundance and lensing data vectors are built, we extract cosmological parameters jointly with the mass–proxy relation parameters and any additional analysis nuisance parameters. This is usually done through likelihood-based inference, where the posterior distribution of the parameters is sampled with Markov Chain Monte Carlo (MCMC) methods ((Abbott et al. 2025; Artis et al. 2022; Ghirardini et al. 2024)). The form of the data vectors depends on the approach adopted. In the binned approach, clusters are grouped into fixed intervals of mass proxy and redshift: the abundance data vector is the number of clusters in each interval, and the lensing data vector is a stacked profile per interval ([1] (Lesci et al. 2022)). In the unbinned approach, the data vector instead retains the mass proxy and redshift of each detected cluster, with the lensing signal entering at the level of individual cluster profiles ((Cash 1979; Hu & Cohn 2006; Bocquet et al. 2019)). This avoids the information loss from binning and stacking, at the cost of computationally expensive likelihood and covariance evaluations ((Payerne et al. 2024)) for the large catalogs expected from optical surveys.

Finally, cluster abundance and weak lensing are combined by multiplying their individual likelihoods to extract their cosmological information. Each likelihood requires a covariance matrix. For the abundance, the scatter arises from the Poisson noise and from the super-sample covariance ((Hu & Kravtsov 2003; Lacasa et al. 2018; Payerne et al. 2023)). For the lensing profiles, the dominant contribution on small scales is the shape noise of the source galaxies ((Wu et al. 2019)). Both covariances can be computed from analytical models or estimated from the data. As the theoretical covariances depend on the parameters themselves, they can either be held fixed during the inference or recomputed at each step. The LSST-DESC Cluster Pipeline assembles these analysis steps into a single automated framework, described in the next section.

The LSST-DESC Cluster Pipeline

This section describes the structure of the LSST-DESC Cluster Pipeline, from catalog ingestion to cosmological parameter inference. We first outline the main workflow and processing stages (§ 3.1), and then detail the pipeline management and configuration with Ceci (§ 3.2). Next, we describe its components (§ 3.3–§ 3.7): CCL and CLMM backends, data vector construction with TXPipe, modeling with CROW and TJPCov, and inference framework with Firecrown. We also discuss the pipeline performance for our test case in § 3.8. All the codes above and the pipeline development are done with the Python programming language.

Workflow

The LSST-DESC Cluster Pipeline aims to be an end-to-end framework for galaxy cluster cosmological analysis for LSST. The pipeline, hosted in the CLPipe repository, integrates a suite of DESC tools into a unified workflow managed by Ceci2.

The pipeline, whose overall structure is illustrated in Figure 1, follows three main stages: (i) catalog ingestion and data vector estimation; (ii) covariance matrix evaluation; and (iii) likelihood construction and parameter inference.

Diagram of the DESC Cluster Cosmology pipeline, showing Galaxy Catalog and Cluster Catalog inputs flowing into TXPipe, with Data Retrieval and CL Ensemble components, and connections among CLMM, CROW, CCL, TJPCov, and Firecrown

Figure 1. Diagram showing the structure of the DESC Cluster Cosmology pipeline. Purple rectangles represent input and output data products, while blue rectangles denote software modules. Solid arrows indicate data passed as inputs, while dashed arrows represent imported modules.

The workflow goes as follows: two configuration files define the stages to run and the analysis choices. The pipeline then begins with a catalog of detected galaxy clusters (giving at least their sky positions, the cluster redshifts, and the mass proxies) and a catalog of galaxies (giving at least their sky positions, shapes, redshift or photometry, and shear calibration information), representing the raw lensing dataset to assess the cluster lensing profiles.

These two catalogs are processed by TXPipe3 (Prat et al. 2023), which computes cluster abundance measurements and individual/stacked cluster weak lensing profiles in bins of mass proxy and redshift. Specifically for cluster lensing, TXPipe imports the cluster-related data management routines of CLMM4 (Aguena et al. 2021, 2026) to compute lensing profiles and their covariance matrices using data resampling techniques (jackknife, bootstrap, etc.). The output data vectors and data-driven lensing covariances are stored in the format handled by SAC5, the DESC standard for the storage of data vectors, covariances, and metadata. The SACC products are then read with TJPCov6, which computes theoretical covariances. ance matrices for cluster abundance measurements using CROW7. Finally, Firecrown8 performs likelihood evaluation and MCMC sampling, using the SACC files as input. It also combines the data vectors and covariances with theoretical predictions for cluster abundance and weak lensing profiles, computed using CROW, built upon the modeling capabilities of CCL (Chisari et al. 2019) and CLMM9, to derive posterior constraints on cosmological and mass-observable parameters. Although the pipeline performs and saves the computation of unbinned cluster lensing profiles, the inference is currently developed only for a binned cluster analysis.

The specific analysis configuration for LSST-DESC cluster cosmology is still to be defined, as it must be optimized against the upcoming LSST data and its systematics. The modular design of the pipeline’s components makes it straightforward to extend or reconfigure the analysis as the science develops. Incorporating the unbinned likelihood, weak lensing mass, or a broader combination of cluster probes such as the 3×2pt+N3 \times2\mathrm{pt}+N analysis would not require redesigning the pipeline.

Pipeline integration with Ceci

To run the pipeline, we use the Ceci pipeline manager package, which connects all of the analysis components discussed in the previous section. In Figure 2, we show a possible pipeline configuration for our baseline analysis presented on cosmoDC2 in § 5 as an example. Our implementation uses several ceci.PipelineStage objects10, each responsible for a single operation in the analysis workflow, shown as red ellipses.

Figure 2. Flow chart of the Ceci pipeline structure described in § 3.2, generated automatically with Ceci. Yellow rectangles denote the pipeline’s overall input files; sky-blue rectangles denote its output files; red ellipses denote individual pipeline stages, labeled with their stage names (TX– and BPZ– prefixes for TXPipe/RAIL stages, CLCluster– for cluster-specific TXPipe stages, and CLP– for CLPipe stages). Arrows indicate file-based dependencies between stages.

The stages in Figure 2 come from three code bases. Stages with the TX prefix are general TXPipe stages, also used in other DESC analyses such as 3×2pt3 \times2\mathrm{pt}. CLCluster stages are the cluster-specific measurement stages, developed for this work within TXPipe. Stages with the BPZ prefix belong to the RAIL11 (van den Busch et al. 2026) pipeline for redshift estimation. CLP stages were developed in CLPipe to connect the outputs of the previous stages to TJPCov and Firecrown. All stages are configured and run together by CLPipe as a single workflow. Stages are connected through input and output files, so execution order follows data dependencies rather than an explicit call sequence.

This file-based connection is what makes the pipeline modular. Any stage can be replaced with another that produces compatible outputs, and new stages can be added for extra computations, without changing the rest of the pipeline. The overall run is set through two YAML files: (i) one defining the PipelineStage objects to be used, their connections, and inputs/outputs paths; and (ii) a configuration file containing meta-parameters for each PipelineStage.

This second configuration file is where most of the modeling choices are set. As an example, for the CLPFirecrown module, the configuration file controls both the model choices (mass definition, halo mass function, scaling relation parameters) and the inference setup (which probe to use, priors, and sampler settings). Listing 1 shows a representative subset of this configuration12. Changing a model choice, such as the halo mass function or mass definition, requires editing a single string in this file, with no changes to the pipeline code itself.

Ceci also enables MPI parallelization of computationally intensive stages, distributing them across multiple processes. The posterior sampling itself is performed subsequently using CosmoSIS with the Firecrown likelihood, executed outside the Ceci workflow, since the chains can run for hours or days and are often tested or restarted independently of the other stages.

Backends: CCL and CLMM

This section describes the two backends used in the pipeline. Both CCL and CLMM are existing DESC tools, not developed as part of this work.

The Core Cosmology Library (CCL) is the LSST DESC library that computes fundamental cosmological quantities, supporting many cosmologies in addition to Λ\LambdaCDM. In this pipeline, CCL is not used as a standalone analysis stage, but rather as a common cosmology engine accessed consistently by all major modules, including TXPipe, CLMM, CROW, TJPCov, and Firecrown, ensuring internal consistency across all stages of the pipeline.

In this work, we restrict the analysis to Λ\LambdaCDM. Extensions whose parameters are already supported by CCL and Firecrown, such as w0waw_0w_aCDM, can be sampled by adding them to the CLPFirecrown configuration, although their propagation through all modules remains to be validated. Since CCL already supports several other cosmological models and serves as the shared cosmology engine for every major module, extending the pipeline to other models would mainly involve configuring or extending CCL itself. The theory modules (CROW, TJPCov, CLMM), which already rely on CCL for their cosmological calculations, would then need to be checked for consistency, while Firecrown would additionally need to support sampling the parameters of the new cosmological model.

The Cluster Lensing Mass Modeling (CLMM) library is the weak lensing backend of the pipeline, used both at the data level within TXPipe (see § 3.4) and at the theory level within CROW (§ 3.5), and it relies on CCL for cosmological calculations13. Here we use several features

Listing 1 Example configuration of a CLPipe module.

CLPFirecrown: hmf: ’despali16’ mass_def: ’200c’ survey_name: ’test_survey’ use_shear_profile: true is_deltasigma: true use_cluster_counts: true

introduced in version 2 of CLMM, which was developed to include new implementations supporting the first weak-lensing analyses of LSST clusters. Details of CLMM v2 are presented in Aguena et al. (2026).

Data vector construction with TXPipe

In the LSST-DESC Cluster Pipeline, TXPipe performs the measurements needed for the cluster abundance and weak lensing analyses. Below we describe the TXPipe stages grouped in different blocks. The first block uses existing stages from the general TXPipe framework (TX– stages, together with the RAIL BPZ– stages in Figure 2), while the cluster abundance and weak lensing measurement stages, along with the output SACC format, were developed as part of this work (CLCluster– stages).

Data ingestion, galaxy-redshift assignment, and shear calibration: This first block follows the framework described in Prat et al. (2023). It begins by ingesting pre-processed galaxy and cluster catalogs that will be saved in the HDF5 data format, which may include multi-band photometry, shape measurements, and associated calibration information. The cluster catalog contains positions, observable properties (e.g., richness, a count of member galaxies typically weighted by their membership probabilities), identifiers assigned by the cluster finder, and redshifts. Following ingestion, TXPipe assigns redshifts to galaxies either by estimating them internally from photometry using RAIL or from external inputs. When redshift point estimates are provided, TXPipe uses these directly. Otherwise, the pipeline operates on full photometric redshift probability distributions. Finally, the galaxy catalog undergoes selection cuts in redshift and signal-to-noise ratio, with shear response calibration and selection response corrections applied following the Metadetection and Metacalibration formalism ((Sheldon & Huff 2017)), depending on catalog type.

Cluster abundance and weak lensing measurements: For cluster weak lensing, the measured quantity is the ellipticity of each source galaxy, whose tangential component around a cluster traces the shear induced by the cluster’s projected mass ([13]). Averaging this tangential component in radial bins around the cluster center gives the cluster weak lensing profile. Building this profile requires source galaxies’ redshift information, used both to select galaxies behind the cluster and to set the lensing efficiency of each cluster–source pair ((Mandelbaum et al. 2018)).

When individual source redshifts are available, we identify source galaxies around each cluster using their individual redshift information and a neighbor search within a maximum angular radius. This maximum is set per cluster by converting a fixed physical search radius into an angle at the cluster’s redshift, using a fiducial cosmology specified in the configuration file and computed with CCL. The sensitivity of the measured data vectors to this cosmological choice is explored in App. B. We are also implementing a tomographic bin-based source selection for the case where only redshift bin membership is known, avoiding the need for individual photometric redshift point estimates in future analysis.

CROW: Theoretical prediction for cluster number counts and stacked lensing profiles

We compute the cluster weak lensing probes using CLMM as the underlying library. The pipeline currently supports two lensing probes: the reduced tangential shear (gtg_{\mathrm{t}}), the tangential distortion of background galaxy shapes induced by the cluster’s gravitational potential, and the excess surface density (ΔΣ\Delta\Sigma), the shear rescaled by the critical surface density of each cluster–source pair. We refer the reader to Aguena et al. (2026) for the full estimator formalism.

We compute individual cluster profiles from the cluster-galaxy pairs, providing unbinned lensing measurements for each cluster. In parallel, we bin clusters in redshift and cluster observable, which directly gives the cluster abundance in each bin. For this same binning scheme, we radially bin the individual profiles and combine them to construct the stacked weak lensing profiles. We then apply CLMM’s bootstrap or jackknife resampling methods to derive a data-driven covariance of the stacked lensing profiles.

Output products: At the end of the TXPipe framework, in CLClusterSACC, we write both cluster counts and stacked weak lensing profiles per bin to a SACC file, together with their covariance matrices and binning information. For cluster counts, the covariance is just the number of counts NijN_{ij} in each bin, saved as a placeholder in SACC, to be replaced by the full theoretical covariance in the next step of the pipeline. We also save individual cluster lensing profiles as pipeline outputs but these are not currently included in the SACC file.

We developed the Cluster Reconstruction of Observables Workbench (CROW), a DESC tool for computing theoretical predictions of galaxy cluster probes. In practice, it combines the halo mass function, mass-observable relation, and selection function within a halo-model framework to predict cluster counts and cluster weak lensing profiles using CCL and CLMM. It is used both in TJPCov for cluster counts covariance calculation and in Firecrown, where its predictions are compared against the data vectors at the likelihood level.

Formally, for bins of redshift and cluster observable (i,j)(i,j), this halo-model framework predicts the number counts and stacked weak lensing profile as

Nij=ΩS∫zizi+1dz∫ln⁡xiln⁡xj+1dln⁡x∫Mmin⁡Mmax⁡dM×d2V(z)dzdΩdn(M,z)dMP(ln⁡x∣M,z)Φ(M,x,z),(1)N_{ij} = \Omega_{\mathrm{S}} \int_{z_i}^{z_{i+1}} dz \int_{\ln x_i}^{\ln x_{j+1}} d\ln x \int_{M_{\min}}^{M_{\max}} dM \times\frac{d^2 V(z)}{d z d\Omega} \frac{d n(M,z)}{d M} P(\ln x \mid M,z) \Phi(M,x,z), \tag*{(1)}
Θij(R)=ΩSNij∫zizi+1dz∫ln⁡xiln⁡xj+1dln⁡x∫Mmin⁡Mmax⁡dM×d2V(z)dzdΩdn(M,z)dMP(ln⁡x∣M,z)Φ(M,x,z)Θ(R∣M),(2)\Theta_{ij}(R) = \frac{\Omega_{\mathrm{S}}}{N_{ij}} \int_{z_i}^{z_{i+1}} dz \int_{\ln x_i}^{\ln x_{j+1}} d\ln x \int_{M_{\min}}^{M_{\max}} dM \times\frac{d^2 V(z)}{d z d\Omega} \frac{d n(M,z)}{d M} P(\ln x \mid M,z) \Phi(M,x,z) \Theta(R \mid M), \tag*{(2)}

where the integrals run over (i) zz, the cluster’s redshift, (ii) MM, the halo mass15, and (iii) xx, the cluster observable, related to the halo mass through the mass-observable relation P(ln⁡x∣M,z)P(\ln x \mid M, z). We made no distinction between the halo and associated cluster redshifts, though this could be accounted for by marginalizing over the cluster’s redshift estimate and will be done in the future. ΩS\Omega_{\mathrm{S}} represents the survey footprint area in steradians Lacasa et al. (2018), d2V(z)/dzdΩ\mathrm{d}^{2}V(z)/\mathrm{d}z\mathrm{d}\Omega is the comoving volume element, dn(M,z)/dM\mathrm{d}n(M,z)/\mathrm{d}M is the halo mass function, denoting the mass and redshift distribution of the dark matter halos (e.g. Tinker et al. 2008; Despali et al. 2016), whose most massive ones host clusters of galaxies, and Φ(M,x,z)\Phi(M,x,z) is the selection function, accounting for the cluster selection systematics. In Eq. (2), Θ\Theta denotes the weak lensing probe, (ΔΣ,gt)(\Delta\Sigma, g_{\mathrm{t}}), consistent with the two probes supported by our pipeline (§ 3.4). The code also supports computation of the mean halo mass in each bin, ⟨M⟩ij\langle M\rangle_{ij}, obtained by replacing Θ(R∣M)\Theta(R\mid M) with MM in Eq. (2).

Each ingredient entering Eqs. (1) and (2) is instantiated in CROW as a separate, modular object: a CROW.ClusterAbundance or CROW.ClusterShearProfile theory object (both built from a CCL cosmology and halo mass function), a mass-observable relation object implementing P(ln⁡x∣M,z)P(\ln x \mid M,z), selection function objects jointly forming the selection function Φ(M,x,z)\Phi(M,x,z), and a redshift object relating the observed cluster redshift to the true halo redshift, either directly (no integral) or through a photometric redshift kernel (future work). CROW.ClusterShearProfile additionally controls every lensing-specific choice, the predicted probe (ΔΣ\Delta\Sigma or gtg_{\mathrm{t}}), halo concentration, one-vs. two-halo contributions, boost factor corrections, and miscentering. For reduced shear specifically, it also sets the lensing-efficiency approximation and whether it is interpolated. These objects are combined by a CROW.recipe object, which performs the integrals described above.

Currently, the halo mass functions are computed via CCL, the mass-observable relation implements a logarithmic mass-richness relation as in Muratov et al. (2019), Eq. (6), and the selection function is modeled through completeness (missed detections) and purity (false detections) terms as in Aguena & Lima (2018), Eqs. (9)–(10). Since each ingredient above is a standalone object, any of these can be altered independently16, and other parametrizations will be tested in future analyses.

CROW supports two modes for evaluating the multi-dimensional integrals in Eqs. (1) and (2): a fast, tabulated mode using precomputed grids and Simpson’s rule from SciPy Virtanen et al. (2020), and a direct, non-tabulated mode using the NumCosmo integration framework Dias Pinto Vitenti & Penna-Lima [30]. The tabulated mode is ∼100\sim100 times faster while maintaining accuracy. We show in Figure 3 the relative difference between the CROW tabulated approach and non-tabulated results, reaching ∼10−5%\sim10^{-5}\% for both counts and lensing profiles. For this benchmark in Figure 3, we used binned abundance plus stacked ΔΣ\Delta\Sigma predictions with: Muratov et al. (2019) scaling relation, Despali et al. (2016) halo mass function; Aguena & Lima (2018) selection function; 4 redshift bins z∈[0.2,0.8]z \in[0.2,0.8] and 6 richness bins λ∈[20,200]\lambda\in[20,200]; 10 radial bins R∈[1.0,10]R \in[1.0,10] Mpc for lensing profiles; grid resolution of 40 points in mass with M∈[1012,1015.5]M⊙M \in[10^{12},10^{15.5}]M_{\odot}, and 20 points for each redshift and richness bins respectively. The tabulated mode is recommended for inference while the non-tabulated mode remains available for validation.

Figure 3. Left: precision plot for the CROW count prediction comparing tabulated to non-tabulated results, assuming the latter to be the truth. Right: Same as left, but for ΔΣ(R=2 Mpc)\Delta\Sigma(R = 2\ \mathrm{Mpc}). Colors and markers denote the richness bins.

Theoretical covariance with TJPCov

TJPCov is DESC’s general-purpose covariance calculator. We developed the theoretical cluster count covariance within TJPCov, using CROW and CCL as backends, which we describe below. This covariance matrix CovN\mathrm{Cov}_{N} is given by

[CovN]ijkl=δikKδjlKNij+[CovSSC]ijkl,(3)[\mathrm{Cov}_{N}]_{ijkl} = \delta^{K}_{ik}\delta^{K}_{jl}N_{ij} + [\mathrm{Cov}_{\mathrm{SSC}}]_{ijkl}, \tag*{(3)}

where i,ki,k run over redshift bins and j,lj,l over cluster observable bins (δikK\delta^{K}_{ik} is the Kronecker symbol). In Eq. (3), the first term is the Poisson shot-noise variance contribution, since it is a count experiment. The second term is the Super Sample Covariance (SSC), arising from fluctuations of modes larger than the survey volume (Hu & Kravtsov 2003; Lacasa et al. 2018; Fumagalli et al. 2021), and which couples the local halo number densities across space and mass–redshift bins. The SSC term adds variance and correlation between the abundance of galaxy clusters measured in different redshift-richness bins and is given by (Lacasa et al. 2018)

[CovSSC]ijkl=NijNkl⟨b⟩ij⟨b⟩klSik,(4)[\mathrm{Cov}_{\mathrm{SSC}}]_{ijkl} = N_{ij}N_{kl}\langle b\rangle_{ij}\langle b\rangle_{kl}S_{ik}, \tag*{(4)}

where ⟨b⟩ij\langle b\rangle_{ij} is the halo bias averaged over the cluster population in the (i,j)(i,j) redshift–observable bin, and Sik=⟨δm,iδm,k⟩S_{ik} = \langle\delta_{m,i}\delta_{m,k}\rangle is the covariance of the survey-window-smoothed matter overdensity between the ii-th and kk-th redshift bins. Compared to the Poisson shot noise, the SSC contribution is proportional to Nij2N_{ij}^{2}, which makes it non-negligible for LSST-sized cluster catalogs (Payerne et al. 2023, 2024).

We use TJPCov within CLPipe in the CLPCovariance stage (Figure 2), where the SACC file produced by TXPipe is used to extract the binning and survey area information of the data vectors. In addition, fiducial cosmological parameters and fiducial mass-observable relation parameters are provided as inputs to compute the theoretical cluster counts covariance.

The resulting theoretical covariance matrix replaces the data-driven cluster-count covariance in the SACC file, while the weak-lensing covariance provided by CLClusterEnsembleProfiles is left unchanged. A theoretical covariance model for cluster weak lensing is currently under development.

Likelihood-based inference with Firecrown

Firecrown is DESC’s likelihood framework, providing the infrastructure for implementing likelihoods. In the LSST-DESC Cluster Pipeline, it is the final stage combining data and theoretical predictions for Bayesian parameter inference. Within the existing Firecrown structure, we developed statistical objects and their likelihood files for the binned cluster abundance and stacked weak lensing profile (an unbinned, individual-profile likelihood is left for future work). All cluster-related Firecrown configuration files are generated at the CLPFirecrown stage of CLPipe. We describe the resulting Firecrown flow below.

Data and parameter input: In Firecrown, the binned cluster counts and stacked weak lensing statistic objects read the SACC file produced by TJPCov, loading the data vector and associated metadata. The covariance itself is read using Firecrown’s existing infrastructure. The parameters to be constrained, including cosmological parameters, mass-observable relation parameters, and others, are specified along with their prior distributions in configuration files.

Likelihood: We create a Firecrown likelihood Python file that combines both probes using our developed statistics. Cluster abundance and the cluster lensing profile measurement are correlated, as they arise from the same underlying matter density field and share some systematics. It is generally assumed, however, that any cross-covariance between them is negligible compared to the proper covariance terms, as justified in Costanzi et al. (2019). We adopt this assumption in the current implementation, which can be extended to a joint likelihood including the cross-correlation between the two probes. The likelihood then reads

Ljoint({N^ij},{Θ^ij})=LN({N^ij})×LΘ({Θ^ij}).(5)\mathcal{L}_{\mathrm{joint}}(\{\widehat{N}_{ij}\},\{\widehat{\Theta}_{ij}\}) = \mathcal{L}_{N}(\{\widehat{N}_{ij}\}) \times\mathcal{L}_{\Theta}(\{\widehat{\Theta}_{ij}\}). \tag*{(5)}

In the above, the hat overscript relates to the measured quantities, while its absence denotes their predictions. The abundance likelihood LN\mathcal{L}_{N} is described with a Poisson likelihood for small samples, and with a Gaussian likelihood once the super-sample covariance becomes comparable to the shot noise (Payerne et al. 2024). In our binned implementation, we adopt the latter, modeled as a multivariate Gaussian across all redshift-richness bins (Hu & Kravtsov 2003; Payerne et al. 2023), using the covariance matrix of Eq. (3). We similarly adopt a Gaussian likelihood for the cluster weak lensing probe Θ∈{ΔΣ,gt}\Theta\in\{\Delta\Sigma,g_{\mathrm{t}}\} [1] (McClintock et al. 2019). At each point in parameter space, Firecrown updates all sampled parameters, with cosmological parameters specifically propagated through its interface with CCL. Our statistical objects then evaluate the theoretical predictions for cluster abundance and weak lensing profiles using CROW. For the moment, the covariance is held fixed throughout inference.

Firecrown further interfaces with several statistical frameworks, including Cobaya (Torrado & Lewis 2021), NumCosmo, and CosmoSIS (Zuntz et al. 2015), each providing its own set of native samplers and cosmological backends, allowing the use of different sampling algorithms and parameterizations. The pipeline integration framework generates all files required for the inference, while the final execution of the sampler is performed separately.

Bayesian inference and output: Once all configuration files are in place, we run the MCMC inference with the Firecrown likelihood and the chosen sampler, producing posterior chains from which parameter constraints and posterior covariances are derived.

Computing performance

The LSST-DESC Cluster Pipeline has been tested at the National Energy Research Scientific Computing Center (NERSC), DESC’s primary computing facility, and at the IN2P3 Computing Center (CC-IN2P3), which hosts the Rubin Observatory French Data Facility. We discuss its performance on the current implementation validated on simulations covering 440 deg2440\,\mathrm{deg}^{2} of the sky (see § 5). The long-term goal is a pipeline that scales to Y10 LSST data volumes, expected to be 40 times bigger than this test case.

Data processing

Within TXPipe, we load large catalog files in chunks rather than reading entire files into memory, a task made straightforward by the HDF5 format. MPI parallelization, managed by Ceci, distributes computations per chunk to reduce wall-clock time. In the specific case of our analysis, most output files (Metadetect shear catalog, cluster catalog, lensing profiles, source photo-zz probability distributions) have modest storage footprints, ranging from 2 to 80 GB for the 36 million galaxies selected. However, scaling such data products to Y10 LSST statistics remains an open challenge. TXPipe requires a peak of ∼\sim40 GB of RAM in parallelized mode and completes in approximately 10 minutes for our test case17 (∼\sim16 minutes when photo-zz estimation with RAIL is included, and ∼\sim1.5 hours with source catalog ingestion). This computation time depends strongly on parallelization and on the number of galaxies carried through the pipeline at each stage, set first by the ingest step and again by the source selection both in the shear calibration stage and at the cluster source pairing level. Assessing the computational requirements of repeated catalog production and validation at LSST-DESC scale remains an area for future work.

One current limitation of the Ceci pipeline manager is uniform MPI resource allocation: requesting 60 MPI jobs per CPU applies that allocation to all stages, even those requiring no parallelization, leaving resources underutilized. The overhead is negligible, and the results are unaffected, but this represents suboptimal resource use. Addressing this through dynamic job allocation via Parsl is under study.

Computing theoretical prediction and inference

For the given analysis binning scheme, the count covariance calculation with TJPCov runs in ∼\sim1 minute with no memory bottlenecks. However, recomputing the covariance at every MCMC iteration is currently impractical. Developing a more dynamic covariance evaluation scheme is an important direction for future development. The covariance is computed once at the desired fiducial cosmology and held fixed.

Then, the primary computational bottleneck lies in the theoretical prediction stage. In the baseline analysis, the prediction of galaxy cluster probes (counts and ΔΣ\Delta\Sigma) involves nested three-dimensional integrals (Eqs. (1) and (2)). When evaluated directly using numerical integration routines within CROW, it requires in total ∼\sim20 seconds, making MCMC analyses computationally impractical. This evaluation time is expected to grow when adding more integrals (e.g., by marginalizing over cluster photometric redshift uncertainties).

To tackle this problem, we adopt a grid-based implementation of the nested integrals. For the baseline analysis presented in § 5, the total computation of the Firecrown likelihood evaluation requires $\sim$0.7 s, with the CAMB power spectra computation contributing a further ∼\sim1.5 s18. Parallelizing across MPI workers matched to the sampler’s ensemble batch size, ∼\sim300,000 posterior samples are generated in ∼\sim7$ hours.

Moreover, at the inference stage, parallelization is available for CosmoSIS-based runs, though Firecrown computations are currently serial. Parallelizing the likelihood evaluation itself offers a natural avenue for future optimization as the parameter space expands.

Validation with synthetic mock datasets

Before applying the LSST-DESC Cluster Pipeline to realistic simulations, we first test its inference machinery on mock data for which the underlying model is exactly known. Generating mocks with the same model we will use for the likelihood inference isolates any bias arising from the inference chain itself (CROW, TJPCov, and Firecrown) from systematics associated with real data. If the pipeline cannot recover the input parameters when the underlying model is known exactly, no result built on top of it can be trusted.

We perform this test using the same cluster counts and stacked excess surface density (ΔΣ\Delta\Sigma) analysis, binned in redshift and richness, presented in § 5, and show that the fiducial cosmological and mass-observable parameters are recovered consistently when the analysis model matches the data-generation model. Note that this test does not exercise the data measurement part of the pipeline, mainly TXPipe, since mock data vectors are obtained directly from the mock generation code19. We describe the mock data generation and its agreement with the pipeline’s theoretical predictions in § 4.1 and present the resulting parameter constraints in § 4.2.

Synthetic mock data generation

To produce the binned mock data vectors, we need realistic cluster and galaxy source catalogs. Directly predicting the binned counts and lensing profiles with CROW would not be sufficient, as it bypasses the unbinned, per-halo and per-galaxy scatter that gives rise to a realistic data-vector covariance in the first place. We therefore generate mock catalogs at the unbinned, individual level, using the Python code capish (Payerne et al. 2026) together with CLMM, before binning the resulting profiles and counts identically to the baseline analysis (see § 5).

Using capish, we generate the underlying halo population (true masses M200cM_{200\mathrm{c}} and redshifts ztruez_{\mathrm{true}}) over 436.6 deg2^2 through an SSC log-normal and Poisson sampling of the Despali et al. (2016) halo mass function at the unbinned level. Cluster richnesses λobs\lambda_{\mathrm{obs}} are then generated as log-Gaussian sampling of the mass-richness relation (Lima & Hu 2005; Murata et al. 2019)

P(ln⁡λ∣M,z)∝1σln⁡λexp⁡{−12(ln⁡λ−μln⁡λσln⁡λ)2},(6)P(\ln\lambda\mid M,z) \propto\frac{1}{\sigma_{\ln\lambda}} \exp\left\{-\frac{1}{2}\left(\frac{\ln\lambda-\mu_{\ln\lambda}}{\sigma_{\ln\lambda}}\right)^2\right\}, \tag*{(6)}

where the mean cluster’s richness λ\lambda is linked to MM via a power law relation. We define the mean and standard choices leave the likelihood numerically unchanged. deviation richness μln⁡λ\mu_{\ln\lambda} and σln⁡λ\sigma_{\ln\lambda}, such that

μln⁡λ=μ0+μmln⁡(MMp)+μzln⁡(1+z1+zp).(7)\mu_{\ln\lambda} = \mu_0 + \mu_m \ln\left(\frac{M}{M_p}\right) + \mu_z \ln\left(\frac{1+z}{1+z_p}\right). \tag*{(7)}
σln⁡λ=σ0+σmln⁡(MMp)+σzln⁡(1+z1+zp).(8)\sigma_{\ln\lambda} = \sigma_0 + \sigma_m \ln\left(\frac{M}{M_p}\right) + \sigma_z \ln\left(\frac{1+z}{1+z_p}\right). \tag*{(8)}

In the above, MpM_p and zpz_p are pivot values, equal to the ones in Table 1 caption. The mean relation is parametrized by an offset richness μ0\mu_0, a mass dependence μm\mu_m, and a redshift dependency μz\mu_z, while the scatter is modeled through the analogous parameters σ0\sigma_0, σm\sigma_m, σz\sigma_z. For simplicity, we do not apply any cluster finder selection effects (purity and/or completeness) on the resulting cluster catalog.

ParameterPriorMockCosmoDC2
Ωc\Omega_{\mathrm{c}}U(0.10,0.50)U(0.10, 0.50)0.220.22
σ8\sigma_8U(0.5,1.1)U(0.5, 1.1)0.800.80
μ0\mu_0U(2,10)U(2, 10)3.343.3446±0.01313.3446 \pm0.0131
μm\mu_mU(0.5,2.0)U(0.5, 2.0)1.00.9587±0.02250.9587 \pm0.0225
μz\mu_zU(−2,2)U(-2, 2)0.0−0.0449±0.1160-0.0449 \pm0.1160
σ0\sigma_0U(0.1,2.0)U(0.1, 2.0)0.560.5626±0.00980.5626 \pm0.0098
σm\sigma_mU(−0.6,0.3)U(-0.6, 0.3)0.00.0454±0.00920.0454 \pm0.0092
σz\sigma_zU(−0.5,2.0)U(-0.5, 2.0)0.0−0.0458±0.0614-0.0458 \pm0.0614

Table 1. Priors and fiducial values for the sampled parameters, for pivot values Mpivot=1014.3 M⊙M_{\mathrm{pivot}} = 10^{14.3}\,M_\odot and zpivot=0.5z_{\mathrm{pivot}} = 0.5. Second column: uniform prior distributions, shared between the idealized-mock analysis (§ 4) and the cosmoDC2 analysis (§ 5), in addition to σln⁡λ>0\sigma_{\ln\lambda} > 0. Third column: injected values used to build the idealized mocks. Fourth column: fiducials obtained from posterior means from the MCMC fit to the matched redMaPPer cosmoDC2 halo catalog (see App. § 9) and from the cosmoDC2 fiducial cosmology.

For the galaxy sources, we use the mock generation modules of CLMM to (i) generate background galaxy catalogs20 for each galaxy cluster; (ii) lens these galaxies with a spherical NFW (Navarro et al. 1997) profile evaluated at the true mass M200cM_{200c}, redshifts ztruez_{\rm true} and concentration c200c=3.8c_{200c} = 3.8; (iii) estimate the perfectly-centered individual cluster excess surface density profile in physical radial bins; and (iv) compute stacked lensing profiles in the same richness-redshift bins as the baseline redMaPPer analysis. Mock galaxy cluster abundance is then measured in the same bins.

Since the mock data generation is stochastic, the level of agreement between a single realization’s data vector and the theoretical prediction depends strongly on the specific random seed used. We therefore generate 1 000 independent mock realizations and compute the mean cluster counts and stacked lensing profiles across them. Figure 4 shows the percent-level agreement between the mean mock data vector and the fiducial prediction using CROW, for one representative redshift bin across all richness bins. Next, we will use this mean data vector to perform the parameter inference.

Figure 4. Validation of the mock data vectors against the pipeline’s theoretical predictions, for the representative redshift bin 0.5<z<0.60.5 < z < 0.6 and all richness bins. Left: cluster counts per richness bin, comparing the mean over the 1 000 mock realizations (points) with the fiducial CROW prediction evaluated at the input parameters of Table 1 (step lines). Right: same comparison for the stacked excess surface density profiles ΔΣ(R)\Delta\Sigma(R), with the mean mock data vectors as points and the fiducial predictions as lines, one per richness bin. Error bars show the 1σ1\sigma uncertainty of a single realization, from the covariance of the 1 000 mocks. The mean mock data vectors agree with the fiducial predictions at the percent level. The other redshift bins show a similar level of agreement.

Parameter posterior inference

We perform a joint analysis of cluster abundance and stacked cluster lensing profiles with a combined Gaussian likelihood as described in § 3.7. The counts covariance is assumed to be the theoretical Poisson+SSC covariance of Eq. (3) evaluated at the fiducial parameters. The lensing covariance entering this likelihood is estimated directly from the scatter of the 1 000 individual mock realizations about their mean, without further rescaling by the number of realizations. Pairing this single-realization-sized covariance with the (noise-free) mean data vector follows the Asimov-dataset approach of Cowan et al. (2011). It tests whether the fiducial parameters are recovered with a single-realization-sized posterior.

We sample the joint posterior distribution of an eight-dimensional parameter space (μ0,μm,μz,σ0,σm,σz,Ωc,σ8\mu_0,\mu_m,\mu_z,\sigma_0,\sigma_m,\sigma_z,\Omega_c,\sigma_8) with the emcee sampler with the CosmoSIS framework21, with over 2.8×1062.8 \times10^6 samples and the first 15% discarded as burn-in. The priors and fiducial values adopted for the sampled parameters are shown in Table 1.

We show in Figure 5 the posterior distribution of cosmology and scaling relation parameters, along with the fiducial values. Using the mean data vector over the 1 000 mock realizations, the posterior is well centered on the fiducial values, although residual shifts away from the exact fiducial point arise from parameter degeneracies (e.g., between the mass-richness relation scatter and the cosmological parameters). Nonetheless, we still recover all fiducials within the 1σ1\sigma credible region, with all posterior-mean offsets <0.4σ< 0.4\sigma from the fiducials. The maximum-posterior sample of the chain recovers all eight input parameters to better than 0.08σ0.08\sigma.

Figure 5. Posterior constraints on the six mass–richness relation parameters and the cosmological parameters Ωc\Omega_{\mathrm{c}} and σ8\sigma_{8}, for the mean mock data vector. Contours show the 68% and 95% credible regions. Dashed lines indicate the fiducial cosmology and mass–richness parameters used to generate the mock catalog.

This test confirms statistically consistent recovery of the input values when the modeling choices for the inference match the assumptions behind the mock data. It also demonstrates that parameter updates propagate correctly through each stage at every sampling step, and that the modular pieces of the pipeline are correctly assembled into a single, functioning workflow.

Application to the LSST-DESC cosmoDC2 simulation

To validate the LSST-DESC Cluster Pipeline, we perform an end-to-end cosmological analysis of the redMaPPer cluster catalog derived from the LSST-DESC cosmoDC2 simulation. The analysis combines binned cluster abundance with the stacked excess surface density profile (ΔΣ\Delta\Sigma), using richness λ\lambda as the cluster observable and the mass-richness relation (Eq. (6)), to jointly constrain two cosmological parameters, Ωc\Omega_{\mathrm{c}} and σ8\sigma_8 from Λ\LambdaCDM, together with the six associated mass-richness relation parameters.

Dataset

The LSST DESC Data Challenge 2 (DC2, Abolfathi et al. 2021) provides a suite of cosmological and image simulations designed to test and refine the analysis tools that will be used for LSST. From this suite, we use the cosmoDC2 extragalactic catalog.

The cosmoDC2 extragalactic catalog

The cosmoDC2 simulation22 (Korytov et al. 2019) is a large synthetic galaxy catalog covering ΩDC2=440 deg2\Omega_{\mathrm{DC2}} = 440\ \mathrm{deg}^2 and built on the Outer Rim NN-body simulation (Heitmann et al. 2019); dark matter halos were identified via a Friends-of-Friends (FoF) algorithm and assigned both FoF masses (MFoFM_{\mathrm{FoF}}) and spherical overdensity masses (M200cM_{200\mathrm{c}}). Galaxies are then painted on halos using Gal-Sampler (Hearin et al. 2020) based on the Galacticus semi-analytic model (Benson 2012). cosmoDC2 compiles "truth" galaxy properties (along with host dark matter halo position, shape, and masses) such as stellar mass, intrinsic ellipticities, and local shear/convergence evaluation (obtained via ray-tracing), and broadband filter "true" magnitude to a redshift z∼3z \sim3 and r≃28r \simeq28. It has been extensively tested to fulfill science requirements for LSST preparation (see e.g. Kovacs et al. 2022), including cluster properties.

redMaPPer galaxy cluster catalog

redMaPPer is a cluster finder algorithm that identifies clusters through overdensities of red-sequence galaxies (Rykoff et al. 2014). It has been applied to the cosmoDC2 extragalactic catalog, and each cosmoDC2 redMaPPer detection is assigned a sky position, photometric redshift (computed from the photometric bands of identified member galaxies), and a richness23 λ\lambda. The full cosmoDC2-redMaPPer cluster catalog contains ∼880,000\sim880,000 detections24 within 5<λ<2705 < \lambda< 270 and 0.1<z<1.150.1 < z < 1.15.

To assess the selection systematics of the cluster catalog, we match redMaPPer-detected clusters to the latent cosmoDC2 halo population using a membership-based approach25. We use the matched cluster-halo catalog to establish (i) the fiducial mass-richness relation in Eq. (6) (considering Mpiv=1014.3M⊙M_{\mathrm{piv}} = 10^{14.3}M_\odot and zpiv=0.5z_{\mathrm{piv}} = 0.5) and (ii) the calibration of the redMaPPer selection function. Let us note that the halo-to-cluster distance26 accessible through this matching procedure is essential to assess miscentering in weak lensing profile prediction. The matching procedure, fiducial scaling relation, selection function, and miscentering systematic studies are detailed in App. § 9.

Galaxy cluster binned probes

Measured abundance and stacked lensing profiles

We measure the cluster abundance and stacked weak lensing profiles using the CCLCluster stages (see § 3.4), configured for this analysis as follows.

We adopt the binning scheme in redshift-richness following the edges z∈{0.2,0.3,0.4,0.5,0.6,0.7,0.8}z \in\{0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8\} and λ∈{20,35,70,100,200}\lambda\in\{20, 35, 70, 100, 200\}, leaving 2658 clusters. The binning scheme is inspired from standard binned lensing analyses such as DES (Abolfatti et al. 2020; Abolfatti et al. 2025) and KiDS (Lesci et al. 2022), and accounts for the degradation of the redMaPPer performance at low27 and high redshifts (McClintock et al. 2019; Cui et al. 2025), while the richness threshold λ>20\lambda> 20 mimics the high-purity cuts generally used to ensure clean cluster samples. The resulting cluster abundance is shown in Figure 6.

Figure 6. Cluster abundance as a function of richness λ\lambda in redshift bins for the baseline redMaPPer cosmoDC2 catalog. Each panel shows a redshift slice with (i) the measured cluster counts from TXPipe (points); (ii) the fiducial prediction (dashed step lines), i.e. the CROW prediction evaluated at the true cosmoDC2 cosmology and the fiducial redMaPPer scaling-relation and selection-function parameters (App. § 9); and (iii) the baseline fit (solid step lines), i.e. the same prediction with the eight sampled parameters set to the maximum posterior parameter values of the joint N+ΔΣN+\Delta\Sigma baseline analysis (§ 5.3.1), all non-sampled quantities kept at their fiducial values. Lower panels show the residuals (N−Npred)/σ(N-N_{\rm pred})/\sigma for the fiducial (dashed) and baseline fit (solid) predictions, and the grey band marks ±1σ\pm1\sigma. Data errors are computed with the theoretical covariance described in § 5.2.2.

For the source selection, we use the TXPipe ingestion and source-selection stages (TXCosmoDC2Mock, TXSourceTomography and TXSourceSelectorMetadetect) with a mag-i≤24.25i \leq24.25 cut (truth cosmoDC2 photometry) and a global redshift cut 0.1<ztrue,gal<30.1 < z_{\mathrm{true,gal}} < 3. The resulting source sample

yields a mean source density of ngal≈25 arcmin−2n_{\mathrm{gal}} \approx25\ \mathrm{arcmin}^{-2}, consistent with 10 years of observations with LSST (Mandelbaum 2018; Chang et al. 2013). Next, for each redMaPPer detection with assigned redshift zclz_{\mathrm{cl}}, galaxies are selected according to ztrue,gal>zcl+0.2z_{\mathrm{true,gal}} > z_{\mathrm{cl}} + 0.2, identifying galaxies behind the cluster28. Lensed ellipticities are computed by combining the local shear/convergence evaluation and an intrinsic ellipticity, taken to be a Gaussian variable with shape noise σSN=0.26\sigma_{\mathrm{SN}} = 0.26 per component, matching observational estimates (Chang et al. 2013).

We configure TXPipe to bin the source galaxies within seven logarithmically spaced radial bins within 1.1 Mpc<R<3.7 Mpc1.1\ \mathrm{Mpc} < R < 3.7\ \mathrm{Mpc}. The lower bound avoids signal attenuation observed at small scales in cosmoDC2 (Korytov et al. 2019; Kovacs et al. 2022; Payerne et al. 2025), while the upper bound restricts to the one-halo regime. We show in Figure 7 the measured ΔΣ\Delta\Sigma stacked lensing profiles in the different richness-redshift bins. In App. § 11, we also compare the measured redMaPPer cosmoDC2 stacked lensing profiles to the measurements from Payerne et al. (2025) and find good agreement.

Figure 7. Stacked excess surface density profiles for the baseline redMaPPer cosmoDC2 catalog. Each panel shows a redshift slice with (i) the measured stacked lensing profiles from TXPipe (points); (ii) the fiducial prediction (dashed lines), as defined in Figure 6; and (iii) the baseline fit (solid lines), i.e. the prediction evaluated at the maximum posterior parameter values of the same joint fit, as defined in Figure 6. Lower panels show the residuals (ΔΣ−ΔΣpred)/σ(\Delta\Sigma-\Delta\Sigma_{\rm pred})/\sigma for each richness bin, with the same line styles, and the grey band marks ±1σ\pm1\sigma. Error bars come from the bootstrap covariance of § 5.2.2.

Comparison to fiducial cosmological prediction

We discuss the baseline modeling choices we use in this paper and compare data to our fiducial predictions.

For both count and lensing predictions in Eqs. (1) and (2), we use the Despali et al. (2016) halo mass function prescription (set according to the M200cM_{200\mathrm{c}} overdensity mass definition). For the scaling relation, we use the Murata et al. (2019) mass-richness relation (Eq. (6)). The selection function follows the completeness parametrization of Aguena & Lima (2018) (Φ(M,x,z)=c(M,z)\Phi(M,x,z)=c(M,z))29, with fiducial parameters obtained from the matching between cosmoDC2 halos and redMaPPer detections (see Table 2). We do not include projection effects in the baseline modeling.

ParameterFiducial Value
c0c_01.13
c1c_10.77
log⁡ac\log a_c13.31
log⁡bc\log b_c0.20
p0p_01.98
p1p_10.81
ln⁡ap\ln a_p2.22
ln⁡bp\ln b_p−0.66-0.66

Table 2. Fiducial selection function parameters of the matched redMaPPer–cosmoDC2 catalog, following the model of Aguena & Lima (2018) (Eqs. (9)–(11)).

For lensing, we model ΔΣ\Delta\Sigma with the one-halo term using a Navarro–Frenk–White (NFW, Navarro et al. 1997) profile with concentration c200c=3.8c_{200\mathrm{c}}=3.8, representative of the mean concentration observed in the cluster sample30. We fix the cluster concentration to a fiducial mean value for simplicity. Tests with a free concentration parameter showed negligible impact on the inferred constraints. Given our radial range (R<3.7 MpcR < 3.7\ \mathrm{Mpc}), we model only the one-halo term in Eq. (2). We do not consider boost factor corrections, since member contamination in cosmoDC2 is below 0.1% even for photo-zz-selected sources (Payerne et al. 2025). We show in App. § 9 that miscentering31 has negligible impact on the modeling, due to (i) the low fraction of miscentered clusters and (ii) the low radial cut at R>1 MpcR > 1\ \mathrm{Mpc} we consider in this work. In the same Appendix, we show that the scatter in the redMaPPer cluster photometric redshifts also has a low impact on the modeling.

the modeling. The theoretical cluster count covariance matrix (Eq. (3)) combines a Poisson shot-noise term with the SSC contribution, assuming the Tinker et al. (2010) halo bias prescription. This covariance is evaluated at the fiducial cosmoDC2 cosmology and held fixed throughout inference (see § 3.7). For weak lensing, we adopt a bootstrap covariance with 100 resamples in each redshift-richness bin of the redMaPPer cluster lensing profiles (we neglect off-diagonal covariance terms since they are very noisy). Between different richness-redshift bins, the covariance between stacked profiles and between different radial scales is generally ignored, as it is significantly smaller than the per-bin components.

We compare in Figure 6 and Figure 7 the fiducial and best-fit predictions to the data vectors, with their residuals in the lower panels. Most residuals lie within 2σ2\sigma, which shows that the baseline model reproduces the overall amplitude and shape of both data vectors. The lensing error bars can be improved with a theoretical covariance matrix, which is in development (see § 6).

Cosmological inference

We sample the joint (N+ΔΣ)(N+\Delta\Sigma) posterior distribution of an eight-dimensional parameter space (μ0,μm,μz,σ0,σm,σz,Ωc,σ8)(\mu_0,\mu_m,\mu_z,\sigma_0,\sigma_m,\sigma_z,\Omega_c,\sigma_8) with emcee within the CosmoSIS framework, using the Gaussian likelihood described in § 3.7.32 We adopt flat priors and fiducials described in Table 1. Each chain contains over 2.2×1062.2 \times10^6 samples generated with 100 walkers, with the first 15% discarded as burn-in. We present the baseline analysis in § 5.3.1, and, in § 5.3.2, vary one ingredient of the analysis at a time (the halo mass function and the fiducial cosmology input) to demonstrate the pipeline’s flexibility.

redMaPPer CC+WL: Baseline analysis

Figure 8 shows the posterior constraints obtained from cluster abundance alone, lensing alone, and their combination. For the cosmological parameters, the lensing-only fit remains prior-dominated, while cluster counts alone provide slightly tighter constraints on Ωc\Omega_c and σ8\sigma_8. The overlap between contours and the different degeneracy directions between count-only and lensing-only constraints (especially in the σz−[μ0,μm,μz]\sigma_z-[\mu_0,\mu_m,\mu_z] plane) make their combination powerful to significantly increase parameter precision. The complementarity between probes is especially visible in the σ8−Ωc\sigma_8-\Omega_c plane, where posteriors break free from the prior-dominated square contours. For the joint case, the fiducial cosmoDC2 cosmology is recovered well within 1σ1\sigma, with posterior means offset by only 0.51σ0.51\sigma in Ωc\Omega_c and 0.34σ0.34\sigma in σ8\sigma_8 from the fiducial values. The corresponding baseline-fit (maximum posterior values) predictions are over-plotted on data vectors in Figure 6 and Figure 7.

Figure 8. Posterior constraints on the six mass–richness relation parameters and the cosmological parameters Ωc\Omega_c and σ8\sigma_8 for the cosmoDC2 redMaPPer cluster dataset. Contours show the 68% and 95% credible regions obtained from cluster abundance alone (blue), stacked weak lensing alone (red), and their combination (black). Dashed lines indicate the fiducial cosmoDC2 cosmology and the fiducial mass–richness parameters from Table 1.

We also compare in App. § 11 our fixed-cosmology results (sampling only the posterior of the cluster scaling relation parameters) to the work of Payerne et al. (2025)

who inferred, independently of this work, the redMaPPer cluster scaling relation from the cosmoDC2-redMaPPer abundance and lensing measurements. The comparison shows overall good agreement despite differences in modeling choices and data processing.

redMaPPer CC+WL: Variations across different analysis setups

The LSST-DESC Cluster Pipeline’s modular design allows individual ingredients of the analysis, from the halo mass function to the fiducial cosmology, to be swapped without changing the rest of the pipeline. We illustrate this flexibility here by testing the robustness of our fiducial cosmoDC2 constraints against alternative analysis configurations, summarized in Figure 9.

Figure 9. Summary of the marginalized constraints on Ωc\Omega_c (left) and σ8\sigma_8 (right) for the baseline analysis and the validation tests described in § 5.3.2. For each case, the point marks the marginalized posterior mean and the dark (light) bars the 1σ1\sigma (2σ2\sigma) intervals, both derived from the posterior chains; the corresponding mean ±1σ\pm1\sigma values are annotated on each row. Vertical lines indicate the fiducial cosmoDC2 values, and colors group the different analysis variations. All cases recover the fiducial cosmology within 1σ1\sigma.

Varying fiducial cosmological parameters: Estimating the stacked weak lensing profiles and the theoretical cluster counts covariance requires a priori knowledge of the fiducial background cosmology. We show in App. B the ratio between the lensing estimators computed at the fiducial cosmology (taking Ωc=0.22\Omega_{\mathrm{c}} = 0.22) and for different Ωc={0.20,0.24}\Omega_{\mathrm{c}} = \{0.20, 0.24\}. Varying Ωc\Omega_{\mathrm{c}} can shift the lensing profiles up to ∼5%\sim5\%. This dependence can be accounted for by applying cosmology-dependent corrections to the lensing data vector ([1] (Sereno 2015; Lesci et al. 2022)). Since we analyze a simulation with known input cosmology, we do not apply these corrections and leave their implementation to the pipeline for future work.

Moreover, the cluster count covariance in Eq. (3) is computed once at the fiducial cosmoDC2 cosmology. In principle, the covariance matrix should be updated at each MCMC evaluation, and the consequences of keeping it fixed have already been explored (Fumagalli et al. 2021; Lettieri et al. 2026). To test the impact of this choice, we update our fiducial σ8\sigma_8 value and perform our baseline parameter estimation, which leaves the computation of the lensing estimators unaffected. Changing the fiducial cosmology only requires editing the cosmological parameter values in a single shared configuration file, which is then propagated through every stage of the pipeline.

We vary σ8∈{0.70,0.90}\sigma_8 \in\{0.70, 0.90\} and estimate the corresponding parameter posterior distributions. Figure 9 shows a summary of the resulting cosmological constraints. Relative to the baseline cosmology, the σ8=0.70\sigma_8 = 0.70 case produces mean posterior shifts of 0.34σ0.34\sigma and 0.27σ0.27\sigma for Ωc\Omega_{\mathrm{c}} and σ8\sigma_8, respectively, while the shifts are only 0.06σ0.06\sigma and 0.03σ0.03\sigma for σ8=0.90\sigma_8 = 0.90. Changes in the fiducial cosmology also modify the covariance and are therefore expected to affect both the posterior means and constraining power. Despite these variations, the recovered cosmological parameters remain consistent with the true cosmoDC2 values within 1σ1\sigma in all cases.

Varying the halo mass function: In our analysis, we use Despali et al. (2016) halo mass function (HMF). Several other HMFs (Tinker et al. 2008, 2010; Bocquet et al. 2019), calibrated on simulations, are available in the literature to predict the underlying halo mass-redshift distribution in the Universe, and can differ from Despali et al. (2016) prediction for a variety of halo mass and redshift ranges33. We want to test the impact of using different HMFs on parameter constraints, namely Tinker et al. (2008), Tinker et al. (2010) and Bocquet et al. (2019). In our pipeline, switching between them requires editing a single entry in the CLPfirecrown configuration file (the hmf field in Listing 1), with no changes to the pipeline code.

Figure 9 shows the results as we perform the inference for different HMFs. Although the HMF choice affects our predictions, part of this effect is reduced by degeneracies with the mass–richness relation parameters. As a result, the different HMFs produce similar posterior contours, and we recover the true cosmoDC2 cosmology within 1σ1\sigma in all cases. Tinker et al. (2008) shows the best agreement with the baseline, with less than 0.05σ0.05\sigma shifts to the mean posterior on both parameters. The largest shift in Ωc\Omega_{\mathrm{c}} is obtained for Tinker et al. (2010) (∼0.19σ\sim0.19\sigma from the baseline), and Bocquet et al. (2019) for σ8\sigma_8 (∼0.11σ\sim0.11\sigma from the baseline).

Discussion and perspectives

In this work, we present the first version of the fully integrated LSST-DESC Cluster Pipeline for analyzing LSST cluster data. In its current version, it enables cosmological analyses starting from galaxy and cluster catalogs to parameter posteriors and relies on several tools developed within the collaboration: Ceci for pipeline integration, TXPipe for data ingestion and formatting, CCL, CROW, and CLMM for modeling cluster probes, TJPCov for theoretical covariance computation, and Firecrown for Bayesian inference.

This pipeline represents a significant collaborative achievement, developed jointly by members of the DESC cluster working group and the broader LSST-DESC collaboration. The pipeline is modular by construction, such that the data measurement, theoretical modeling, covariance, and inference stages can each be modified or replaced separately, allowing us to adapt to different analysis choices. Ceci continuously integrates these stages into a single coherent pipeline, and cosmological parameters are propagated consistently across all stages through a shared CCL backend. Major modeling choices can be changed by editing a single string in a configuration file. New data formats can similarly be accommodated by updating the relevant input entry points. Finally, the pipeline is built to run on batch computing farms, with Ceci-managed MPI parallelization distributing computationally intensive stages across multiple processes to reduce wall-clock time and memory usage.

We first tested the inference part of the pipeline with a mean mock dataset well described by our probe predictions, and found statistically consistent recovery of the fiducial parameters. We then performed a first full run of the pipeline by analyzing the binned abundance and ΔΣ\Delta\Sigma lensing profile of galaxy clusters detected with redMaPPer on the cosmoDC2 truth galaxy catalog. Using cosmoDC2 allows us to test the pipeline in ideal conditions, as the true galaxy properties are known, especially shapes and redshifts, with no survey systematics. The outcome of the pipeline run is the posterior distribution of the cosmological parameters and of the redMaPPer cluster scaling relation parameters, which showed good overall agreement with the fiducial values of the simulated dataset. We also showed the modularity of the analysis for different configurations (different fiducial cosmology and HMF prescription), recovering fiducials within the credible region in all cases.

Beyond this idealized setup, which constitutes a first-step validation of the workflow, further work is needed for analyses with real cluster and lensing datasets. We discuss below the observational challenges of moving to real LSST data, what remains specific to the binned N+ΔΣN + \Delta\Sigma baseline validated here, how the pipeline extends beyond this baseline, and, finally, planned improvements to covariance and inference.

Observational systematics. Real LSST data introduces several systematics relevant for cluster cosmology. In our cosmoDC2 analysis, clusters’ background galaxies are selected via their true redshift, which is also used to compute the lensing-source geometrical weights. LSST galaxy redshifts, by contrast, will be estimated from noisy ugrizyugrizy photometry ((Zhang et al. 2026)), potentially biasing the lensing profile through outliers34 and cluster members35 entering the source sample, and through biased lensing-source distance weights that cannot be mitigated simply by increasing sample size (Simet et al. 2017; Applegate et al. 2014). Although unused in our test case, this pipeline is already capable of handling photometric redshift inputs in TXPipe for data vector construction, but is under development on the modeling side. Additionally, a tomographic bin-based source selection is already under development in TXPipe as an alternative, replacing individual point estimates with source-redshift distributions per tomographic bin.

Shear calibration is not exercised in the cosmoDC2 idealized test, since true galaxy shapes are directly provided. For LSST, shapes will instead be measured from images, where PSF modeling ([10]), blending (Ramel et al. 2026), and the calibration method itself (Sheldon & Huff 2017; Zhou et al. 2026) introduce complex, potentially non-linear relationships between observed and true shapes. Currently, DESC plans to provide different shear catalogs for LSST, such as Metadetect (Sheldon et al. 2020, 2023) and AnaCal (Li & Mandelbaum 2023; Li et al. 2024). The TXPipe component of the LSST-DESC Cluster Pipeline is already adapted to use Metadetect and is being extended to AnaCal catalogs.

Finally, cluster finders will rely on realistic LSST multi-band photometry and survey properties (PSF, depth variations, blending, etc.), which can introduce several sources of systematic bias in the cluster catalogs (miscentering, Ding et al. 2025; uncertain cluster photometric redshift estimates, selection function biases, Aguena & Lima 2018; etc.). The selection function is modeled within CROW as a standalone, modular object, so further implementations can be easily incorporated. Projection effects are also being added to the theoretical predictions. The richness–shear covariance induced by the cluster selection may also need to be modeled (Zhang et al. 2024).

Refining the binned counts and stacked ΔΣ\Delta\Sigma baseline. Our current lensing baseline deliberately includes only the one-halo regime, excluding scales where miscentering, member contamination (Gruen et al. 2015; McClintock et al. 2019), and neighboring-cluster contributions become important. At small scales (∼1\sim1 Mpc), where the lensing amplitude is strongest, these effects tend to bias or dilute the signal, while at larger scales the two-halo contribution and the fiducial-cosmology dependence of ΔΣ\Delta\Sigma both grow (Sereno 2015). CROW already supports a two-halo term and individual shear profile miscentering corrections via CLMM, as configurable modeling choices, although the current implementation to the binned theoretical predictions is computationally slow and must be optimized. Extending our radial range requires validating and optimizing these existing capabilities and further modeling development to incorporate refined halo concentration modeling, cluster triaxiality (Fu et al. [34]), halo orientation (Dietrich et al. 2014), and the fiducial cosmology correction to the shear profile predictions (Sereno 2015).

Beyond this baseline. The LSST-DESC Cluster Pipeline is designed to support analysis choices beyond the binned N+ΔΣN+\Delta\Sigma combination presented here. The pipeline can also run end-to-end with the reduced tangential shear gtg_{\mathrm{t}} profile instead of ΔΣ\Delta\Sigma. Individual, unbinned cluster lensing profiles are already computed and saved in TXPipe, but the complete pipeline workflow is not yet fully developed for an unbinned cluster analysis. Combining cluster probes with other LSST-DESC probes in a joint 6×2pt+N6 \times2\mathrm{pt}+N analysis requires additional pipeline stages and further development at the likelihood level. This development should be easier for probe combinations already implemented within the same infrastructure, such as the standard LSST-DESC 3×2pt3 \times2\mathrm{pt} analysis (weak lensing and galaxy clustering) that uses the same DESC tools (Prat et al. 2023).

Covariance and inference. The cluster count covariance used in this work is evaluated once, at a single fiducial cosmology, and held fixed throughout inference. In principle, it should instead be recomputed at each parameter configuration explored by the sampler. Additionally, a theoretical covariance model for cluster weak lensing, beyond the data-driven bootstrap covariance used in this work, is also still to be implemented and will be needed to capture contributions such as intrinsic profile scatter and correlated large-scale structure (Wu et al. 2019). Given the pipeline’s modular design, it can also be extended to support Simulation-Based Inference (SBI) as an alternative inference approach (Payerne et al. 2026). This could be achieved by introducing an additional inference stage while leaving the data measurement and theoretical modeling stages described here unchanged, with the theory modules providing the forward simulator required by the SBI framework.

Data and software availability

The data products required to reproduce the results presented in this work, including the SACC data vectors, the posterior chains, and the mock realizations, are hosted on the NERSC Science Gateway portal at https://portal.nersc.gov/cfs/lsst/clpipe/data/. Regarding the DESC tools, the results of this work were obtained with CLPipe v1.0.0, which runs TXPipe (tag clpipe-paper-v1), CROW v1.0.12, TJPCov v0.6.0 and Firecrown v1.15.0a1.dev187 (commit 7098afc). These codes rely on CCL v3.2.1, CLMM v1.16.1 and SACC v2.1.2.

The stages are connected with Ceci v2.5. All DESC codes are publicly available at https://github.com/LSSTDESC.

Acknowledgments

General acknowledgments. EJB acknowledges support from the Enigmass+ research federation (CNRS, Université Grenoble Alpes, Université Savoie Mont-Blanc). CA was supported by DOE grant DE-SC009193 and the Leinweber Foundation. CGG was supported by the César Nombela Research Talent Attraction grant from the Community of Madrid (Ref. 2025-T1/TEC-36302). This work is supported by the French National Research Agency in the framework of the “France 2030” (ANR-15-IDEX-02). The DESC acknowledges ongoing support from the Institut National de Physique Nucléaire et de Physique des Particules in France; the Science & Technology Facilities Council in the United Kingdom; and the Department of Energy and the LSST Discovery Alliance in the United States. DESC uses resources of the IN2P3 Computing Center (CC-IN2P3–Lyon/Villeurbanne - France) funded by the Centre National de la Recherche Scientifique; the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231; STFC DiRAC HPC Facilities, funded by UK BEIS National E-infrastructure capital grants; and the UK particle physics grid, supported by the GridPP Collaboration. This work was performed in part under DOE Contract DE-AC02-76SF00515. This paper has undergone internal review in the LSST Dark Energy Science Collaboration. We are grateful to the internal reviewers, Ian Dell’Antonio and Rance Solomon, who provided helpful feedback and suggestions on the manuscript. We thank Matthew R. Becker for his work on the cluster prediction module of Firecrown, which later became CROW.

Author contributions. EJB: Leading the analysis, development lead of CLPipe, co-lead developer of CROW, contributions to TXPipe, Firecrown, TJPCov and SACC, original draft and manuscript writing. CP: Conceptualization, methodology, CLMM development, DC2-like mocks with capish and CLMM, data analysis, manuscript writing and review. MAG: CROW co-developer, co-lead of CLMM development, review of the early draft. CC: CLMM development, initial cluster weak lensing measurement in TXPipe, early definition of the SACC cluster tracers, pipeline design discussions, extensive feedback on the original draft. SE: CLPipe development, analysis templates and code organization, installation of the conda environments for Firecrown and TXPipe that enabled the analysis. TG: Project conceptualization and management, discussions of the analysis steps, help with the analysis of the results, review of the draft. MR: DESC Builder, Cluster Working Group convener, cluster modules in TXPipe and use of ceci, co-lead of CLMM development, review of the early draft. SDPV: Firecrown development, initial cluster counts framework, numerical integration backend, cross-checks with NumCosmo. JZ: TXPipe maintainer, cluster modules in TXPipe, help with the analysis, review of the manuscript. NA: Scientific discussions, debugging, interpretation of the results, manuscript feedback. MAp: Contributions to Firecrown and design of the cluster counts framework. CA: Early development of the cluster code in TXPipe. DB: DESC Builder, precursor studies, scientific discussions, early testing and debugging. NEC: DESC Builder, CCL development, comments on the final draft. AF: Implementation of the super-sample covariance in TJPCov. CGG: Review of the cluster covariance implementation in TJPCov. NLM: Scientific discussions, testing and debugging, interpretation of the results, review of the draft. AIM: DESC Builder, early contributions to RAIL and CLMM conceptualization, funding acquisition, project administration, software and validation. MP: Firecrown development. EMP: TXPipe maintainer. MPL: NumCosmo halo and cluster codes used to validate Firecrown, design of the initial cluster counts code. ESR: redMaPPer development, redMaPPer run on cosmoDC2. NŠ: DESC Builder, CCL and TJPCov development. RS: Internal reviewer of this manuscript. KT: Explanatory notebooks for the TXPipe analysis stages. YZ: Firecrown development, initial cluster counts framework.

cosmoDC2-redMaPPer cluster catalog

To derive fiducial parameters and analyze cluster properties, we match detected clusters to their true halo counterparts using membership-based matching via the ClEvaR36 software. The matching procedure is performed after applying cuts of M200c>1013M⊙M_{200c} > 10^{13}M_\odot to the halo catalog and λ>5\lambda> 5 to the redMaPPer cluster catalog. Each cross-matched cluster is then assigned a true spherical overdensity mass M200cM_{200c}, a true RAtrue\mathrm{RA}_{true}, DECtrue\mathrm{DEC}_{true}, and a true redshift ztruez_{true} from its associated cosmoDC2 halo.

Mass-Richness

With the matched catalog, we fit the six parameters of the mass–richness relation (Eq. (6)) using MCMC sampling with Gaussian likelihood and truncation correction (Kelly (2007)). The fit uses pivot values Mp=1014.3 M⊙M_p = 10^{14.3}\,M_\odot and zp=0.5z_p = 0.5, with cuts M200c>4×1013 M⊙M_{200c} > 4 \times10^{13}\,M_\odot, 10<λ<40010 < \lambda< 400, 0.2<z<0.80.2 < z < 0.8. The mass and richness cuts follow Payerne et al. (2025), who test the stability of this fiducial relation against these choices. The mass cut excludes low-mass halos that would bias the fit, while the redshift cut reflects the degraded red-sequence detection of redMaPPer at higher redshifts (Rykoff et al. (2014)).

The quality of the fit is illustrated in Figure 10. Both panels demonstrate that the fitted mass–richness relation provides an accurate description of the observed richness distribution. The posterior mean values are listed in Table 1 along with the used priors.

Figure 10. Validation of the fitted mass–richness relation. Left: mass–richness scatter plot for the matched redMaPPer cosmoDC2 catalog. The dashed line shows Eq. (6) evaluated at the best-fit parameter values for z=0.4z = 0.4. Right: normalized distributions of ln⁡λ\ln\lambda in three halo mass bins (histograms) compared with the corresponding model predictions (dashed lines) obtained using the best-fit parameter values.

Cluster vs. Halo Redshift

We compare the redMaPPer photometric cluster redshifts to the true halo redshifts from the cosmoDC2 catalog for all matched pairs. The fractional redshift difference, (zcluster−zhalo)/(1+zhalo)(z_{\mathrm{cluster}}-z_{\mathrm{halo}})/(1+z_{\mathrm{halo}}), has a mean of −3×10−4-3 \times10^{-4} and a scatter of 0.01050.0105. The absence of a significant systematic bias indicates that no redshift correction is required, and we therefore adopt the redMaPPer redshifts throughout the analysis.

Selection Function

For the baseline cosmoDC2 analysis, we model the selection function Φ\Phi in Eqs. (1) and (2), following the prescription of Aguena & Lima (2018)37, such that Φ(M,λ,z)=c(M,z)/p(λ,z)\Phi(M,\lambda,z) = c(M,z)/p(\lambda,z), where we model completeness (c(M,z))(c(M,z)) and purity (p(λ,z))(p(\lambda,z)) as

c(M,z)=NmatNhalos=(M/M0)nc(M/M0)nc+1,(9)c(M,z) = \frac{N_{\mathrm{mat}}}{N_{\mathrm{halos}}} = \frac{(M/M_0)^{n_c}}{(M/M_0)^{n_c}+1}, \tag*{(9)}
p(λ,z)=NmatNclusters=(ln⁡λ/ln⁡λ0)np(ln⁡λ/ln⁡λ0)np+1,(10)p(\lambda,z) = \frac{N_{\mathrm{mat}}}{N_{\mathrm{clusters}}} = \frac{(\ln\lambda/\ln\lambda_0)^{n_p}}{(\ln\lambda/\ln\lambda_0)^{n_p}+1}, \tag*{(10)}

with

nc=c0+c1(1+z)(11)n_c = c_0 + c_1(1+z) \tag*{(11)}
np=p0+p1(1+z)(12)n_p = p_0 + p_1(1+z) \tag*{(12)}
log⁡M0=log⁡ac+log⁡bc(1+z)(13)\log M_0 = \log a_c + \log b_c(1+z) \tag*{(13)}
ln⁡λ0=ln⁡ap+ln⁡bp(1+z),(14)\ln\lambda_0 = \ln a_p + \ln b_p(1+z), \tag*{(14)}

and NmatN_{\mathrm{mat}}, NhalosN_{\mathrm{halos}}, and NclustersN_{\mathrm{clusters}} denote the number of matched cluster–halo pairs, the total number of halos, and the total number of detected clusters, respectively.

A detailed characterization of the cosmoDC2 redMaPPer cluster catalog is presented in Ricci et al., in prep. The fitted fiducial values are reported in Table 2. For λ>20\lambda> 20, the catalog is ∼100%\sim100\% pure and so we consider p(λ,z)=1p(\lambda,z) = 1 above this richness threshold. Its completeness is approximately 20%20\%, 70%70\%, and 100%100\% for 1013≲M200c/M⊙<1013.610^{13} \lesssim M_{200c}/M_\odot< 10^{13.6}, 1013.6<M200c/M⊙<1014.210^{13.6} < M_{200c}/M_\odot< 10^{14.2}, and higher masses, respectively.

Miscentering

We measure the projected offset RoffR_{\mathrm{off}} between each redMaPPer cluster center and its matched halo center, assuming that the lensed positions of the central galaxies of the cosmoDC2 halos define the true cluster centers. We consider cluster–halo pairs with M200c≥4×1013M⊙M_{200c} \geq4 \times10^{13}M_\odot, λ≥20\lambda\geq20, and 0.2≤z≤1.00.2 \leq z \leq1.0. The resulting offset distribution is shown in the left panel of Figure 11. Approximately 86% of clusters are perfectly centered (i.e., redMaPPer selects the true central galaxy, giving a null offset), while the full sample has a mean projected offset of ⟨Roff⟩=0.033 Mpc\langle R_{\mathrm{off}}\rangle= 0.033\,\mathrm{Mpc}.

Offset distributions and miscentering impact on excess surface density

Figure 11. Left: Distribution of projected offsets RoffR_{\rm off} between redMaPPer cluster centers and their matched cosmoDC2 halo centers, on a logarithmic scale. Center: Best-fit offset distribution (Eq. (15)) for the non-zero offset subsample, given as σoff=0.12\sigma_{\rm off} = 0.12 Mpc. Right: Relative difference between individual ΔΣ\Delta\Sigma profiles computed with and without miscentering at z=0.5z = 0.5, using the measured offset fraction and best-fit σoff\sigma_{\rm off}.

To characterize the non-zero offset population, we follow Appendix A of Payerne et al. (2025) and model the offset distribution with a Gamma distribution as (Gicoli et al. (2021); McClintock et al. (2019); Zhang et al. (2019)),

p(Roff∣σoff)=Roffσoff2exp⁡(−Roffσoff),(15)p(R_{\mathrm{off}} \mid\sigma_{\mathrm{off}}) = \frac{R_{\mathrm{off}}}{\sigma_{\mathrm{off}}^2} \exp\left(-\frac{R_{\mathrm{off}}}{\sigma_{\mathrm{off}}}\right), \tag*{(15)}

which gives a best-fit scale radius of σoff=0.12 Mpc\sigma_{\mathrm{off}} = 0.12\,\mathrm{Mpc} (center panel of Figure 11).

To assess the impact of miscentering on the weak-lensing signal, we use CROW to compute individual cluster ΔΣ\Delta\Sigma profiles at z=0.5z = 0.5 with and without the miscentering correction, using the measured offset fraction and best-fit σoff\sigma_{\rm off}. The resulting change in the predicted signal is only 1.2%, 0.4%, and 0.2% at R=1R = 1, 2, and 3 Mpc respectively (right panel of Figure 11). We therefore neglect miscentering effects in this work.

Stacked excess surface density profile measurement: impact of fiducial cosmology

The estimation of the ΔΣ(R)\Delta\Sigma(R) profile depends on the choice of a fiducial cosmology, which enters the computation of the geometric lensing weights and the cluster–source galaxy associations. Variations in the fiducial cosmology therefore affect the measured data vector, with inferred weak-lensing masses changing by approximately 5–10% (Meneghetti et al. 2017). This dependence can be accounted for at the modeling level during cosmological inference (Sereno 2015). To illustrate the impact of this effect in our analysis, we estimate the stacked weak-lensing profile from our baseline using a few different fiducial cosmologies. For simplicity, we combine all clusters in the redshift range 0.2<z<0.80.2 < z < 0.8 and richness range 20<λ<20020 < \lambda< 200 into a single stacked profile. Figure 12 shows that changing the fiducial cosmology modifies the measured ΔΣ(R)\Delta\Sigma(R) profile by ∼5%\sim5\% relative to the baseline cosmology.

Figure 12. Stacked weak lensing profiles estimated from data with different fiducial cosmologies to search for galaxy sources. The lower panel shows the fractional difference with respect to the baseline profile, and the grey band marks ±5%. The profiles were calculated on a single redshift-richness bin.

Cross-Validation Against an Independent cosmoDC2-redMaPPer Analysis

As a validation of the TXPipe procedure, Figure 13 shows stacked lensing profiles around redMaPPer clusters measured by TXPipe in the LSST DESC cosmoDC2 simulated datasets (see full details in § 5). Here we extend the radius binning for comparison purposes. When comparing to Payerne et al. (2025) who analyzed the same cluster dataset with the LSST DESC CLCosmoSim code (with different source selection), it shows an overall good agreement.

Figure 13. Comparison between redMaPPer cosmoDC2 TXPipe lensing data vectors used in this work and CLCosmoSim lensing data vectors from Payerne et al. (2025). Dataset and configurations are described in § 5. Filled contours represent CLCosmoSim error bars.

We perform a fixed-cosmology parameter inference with the same configuration as in § 5. Figure 14 shows the corresponding posterior distribution when combining cluster count and lensing. Our constraints are consistent within 2σ2\sigma with the scaling relation fiducial values. The comparison also shows good agreement with the results obtained in Payerne et al. (2025), with a slight difference in μz\mu_z. This can be explained by a difference in the selection function modeling between the two analyses. Payerne et al. (2025) model purity as a non-trivial function of redshift and richness, whereas we fix it to unity above our richness threshold (λ≥20\lambda\ge20; App. § 9.§ 3), which could directly impact the contours in the μz\mu_z–ln⁡λ0\ln\lambda_0 plane. Other differences between the two analyses include the source selection used in the data measurement, and the injection of artificial shape noise in our pipeline versus the use of intrinsic galaxy shapes in Payerne et al. (2025).

Figure 14. Posterior distribution of the mass–richness relation parameters using abundance and weak lensing profiles. LSST-DESC Cluster Pipeline posteriors are shown in blue and Payerne et al. (2025) chains are shown in dashed black lines. Dashed crosses represent the fitted fiducial values from Table 1.

This paper was built using the Open Journal of Astrophysics \LaTeX template. The OJA is a journal which provides fast and easy peer review for new papers in the astro-ph section of the arXiv, making the reviewing process simpler for authors and referees alike. Learn more at http://astro.theoj.org.

References

  1. [1]Abbott T. M. C., et al., 2020, Phys. Rev. D, 102, 023509
  2. [10]Arbeláez D. A. H., Léget P.-F., Malagón A. A. P., 2026, Revista eSpectra, 4
  3. [13]Bartelmann M. Schneider P., 2001, Phys. Rep., 340, 291
  4. [30]Dias Pinto Vitenti S., Penna-Lima M., 2014, Astrophysics Source Code Library, app asl–1408
  5. [34]Fu S., et al., 2026, arXiv e-prints, arXiv:2605.06587arxiv.org/abs/2605.06587
  6. [85]To C., et al., 2021, Phys. Rev. Lett., 126, 141301
  7. [87]Vedder C. J. G., Chisari N. E., 2021, MNRAS, 500, 5561

Paper details

vlnc:W_q3vrg26q
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents