🔭 astrophysics

Galaxy cluster count cosmology with simulation-based inference

This study demonstrates that simulation-based inference, applied to galaxy cluster surveys, mitigates systematic uncertainties and accurately recovers cosmological parameters, highlighting that mass-scale calibration is the critical factor, whereas the $S_8$ parameter proves less sensitive to such uncertainty.

Original authors: M. Regamey, D. Eckert, R. Seppi, W. Hartley, K. Umetsu, S. Tam, D. Gerolymatou

Published 2026-04-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: M. Regamey, D. Eckert, R. Seppi, W. Hartley, K. Umetsu, S. Tam, D. Gerolymatou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine being a cosmic detective. Your task is to understand how the universe is made, how big it is, how old it is, and what it consists of. But there is a problem: you cannot directly see the "glue" that holds things together (dark matter) nor can you directly measure the force pushing the universe to expand (dark energy).

How do you proceed? You look at galaxies. But not single galaxies; rather, their "clusters": galaxy clusters. These are the largest objects bound by gravity in the universe, like enormous cosmic cities filled with hundreds of galaxies swimming in a sea of extremely hot gas.

Here is what this article is about, explained simply:

1. The Problem: Measuring the Unmeasurable

To understand the universe, astronomers count these clusters. The more clusters there are, the more we know the universe is made of matter and that its fluctuations are strong.
But there is a huge obstacle: we cannot weigh a cluster directly. It is like trying to figure out how much an elephant weighs by looking only at its shadow or hearing the sound of its footsteps. We must use "tricks" (called scaling relations), such as saying: "If the gas inside the cluster is very hot, then the cluster must be very massive."

The problem is that these tricks are not perfect. If we make even a small mistake in estimating the mass, we get the entire calculation of the universe's history wrong. Furthermore, telescopes do not see everything: they have "filters" (selections) that cause them to see only the brightest or closest clusters, creating a distorted image.

2. The Solution: The "Reality Simulator"

The authors of this study said: "Enough with trying to correct the tricks by hand. Let's build a simulator that recreates the universe from beginning to end."

Imagine having an ultra-realistic video game (like The Sims, but for the universe):

Insert the rules: You tell the computer: "Make a universe with this amount of matter (Ωm) and this initial explosion force (σ8)."
The game runs: The computer generates millions of galaxy clusters, calculates how hot they are, how bright they shine, and where they are located.
The filter: Then you apply the same "filter" of the real telescope to the game (as if you were looking at the game through a dirty window or from a certain viewing angle).
The comparison: Now you compare what the computer "saw" with what the real telescope actually saw.

3. Artificial Intelligence: The Translator

Here comes the magical part: Simulation-Based Inference (SBI).
Instead of performing complicated mathematical calculations to understand why the data do not match, we use a neural network (a type of artificial intelligence).

Training: We show the AI thousands of different simulations. We tell it: "Look, if the universe is made this way, the telescope will see this. If it is made that way, it will see that."
Learning: The AI learns to connect the dots. It learns to say: "Ah, this specific distribution of hot and luminous clusters corresponds to a universe with these precise rules."
The result: When we give it the real telescope data, the AI tells us: "Hey, these data correspond almost perfectly to a universe with these characteristics."

4. What did they discover?

They tested their method on fake (simulated) data created by supercomputers, and it worked perfectly. Then they analyzed what happens if we make a mistake:

Mass is everything: If we make a mistake calibrating the "mass" of the clusters (the mass scale) by even just 10%, our calculations about the universe become wrong. It is as if there were a wrong label on the elephant's scale: everything else collapses.
The good news: There is a combination of numbers (called $S_8$ ) that seems to be much more robust. Even if we make a small mistake about the mass, this specific number remains stable. It is as if, even if we did not know the exact weight of the elephant, we could still say with certainty how tall it is.
Size matters: The larger the telescope (and the more clusters it can see), the more precise the results become. For small fields of view, random errors dominate; for large fields (like those that the future telescope eROSITA or the satellite Euclid will cover), the method becomes extremely powerful.

In summary

This article tells us that instead of trying to "fix" real data with imperfect mathematical formulas, it is better to build a virtual universe, run it with AI, and compare it with reality. It is a modern, powerful, and much more flexible approach that will allow us to better understand the secrets of the universe when new telescopes begin collecting data.

It is like moving from trying to guess the contents of a box by shaking it, to having a transparent box that shows us exactly what is inside, thanks to an intelligent simulator.

1. Problem Statement

Galaxy clusters are powerful cosmological probes because their abundance and mass distribution are highly sensitive to key parameters, specifically the matter density ( $\Omega_m$ ) and the amplitude of matter fluctuations ( $\sigma_8$ ). However, extracting precise cosmological constraints from cluster counts is currently limited by several systematic uncertainties:

Mass-Observable Relation: The inability to directly measure halo mass requires reliance on scaling relations (e.g., between X-ray luminosity/temperature and mass), which introduce intrinsic scatter and calibration uncertainties.
Selection Functions: Survey selection effects are often difficult to model analytically, leading to approximations that bias results.
Halo Mass Function (HMF) Accuracy: Theoretical models of the HMF have inherent uncertainties.
Degeneracies: Parameters like $\Omega_m$ and $\sigma_8$ are often degenerate, and the absolute mass scale calibration significantly impacts their recovery.

Traditional methods often rely on analytic likelihoods and simplified assumptions that struggle to marginalize over these complex systematics simultaneously.

2. Methodology: Simulation-Based Inference (SBI)

The authors propose a forward modeling approach using Simulation-Based Inference (SBI) to bypass the need for an explicit analytic likelihood. The core workflow involves:

A. The Forward Modeling Pipeline

The authors constructed a numerical pipeline that generates mock cluster catalogs from input cosmological and scaling relation parameters:

Cosmology & Volume: Input parameters ( $\Omega_m, \sigma_8, H_0$ ) define the survey volume and the differential comoving volume.
Mass Function: Clusters are drawn from a Halo Mass Function (HMF) (specifically the Tinker et al. 2008 model) within defined redshift slices.
Mass Correction: Since the HMF masses depend on the cosmology (via critical density $\rho_c$ ), the pipeline converts simulated masses to the cosmology assumed for the scaling relations (calibrated via Weak Lensing) to ensure consistency.
Scaling Relations:
- Mass-Temperature (M-T): A known relation is used to assign temperatures to halos, including intrinsic scatter.
- Mass-Luminosity (M-L): A power-law relation ( $L \propto M^{\alpha} E(z)^{7/3}$ ) is applied. The slope ( $\alpha$ ), normalization, and scatter are treated as free parameters to be marginalized over.
Observable Generation:
- K-correction: Rest-frame luminosities are converted to observer-frame fluxes accounting for redshift and spectral shifts.
- Aperture Correction: A $\beta$ -model is used to rescale fluxes from the theoretical $R_{500}$ aperture to a fixed physical aperture (e.g., 300 kpc), mimicking real survey extraction methods.
Selection Function: The pipeline applies the exact survey selection function (cuts in flux, temperature, redshift) to the generated observables, producing a final mock catalog directly comparable to real data.

B. Inference Engine

Algorithm: The authors use the sbi Python package implementing Density Estimation Likelihood-Free Inference (DELFI) and Sequential Neural Posterior Estimation (SNPE).
Training: A neural network is trained to map the summary statistics of the mock data (histograms of flux, temperature, and redshift) back to the input parameters.
Data Vector: The method uses a 2D histogram of Flux vs. Temperature combined with a 1D redshift histogram as the summary statistic.
Optimization: A non-uniform proposal distribution (based on a Fisher matrix derived from an initial uniform run) is used to improve sampling efficiency in the 6-dimensional parameter space ( $\Omega_m, \sigma_8, H_0$ , and M-L scaling parameters).

3. Key Contributions

End-to-End Forward Modeling: The method generates observables directly from cosmological parameters, allowing for the exact numerical implementation of complex selection functions that are difficult to express analytically.
Joint Inference: The framework simultaneously recovers cosmological parameters and the parameters of the Mass-Luminosity scaling relation, naturally marginalizing over scaling relation uncertainties.
Mass Calibration Independence: By treating the absolute mass scale normalization as a free parameter (or calibrating it externally), the method isolates the impact of mass calibration errors on cosmological constraints.
Scalability: The approach is easily adaptable to include additional observables (e.g., weak lensing shear, richness) without re-deriving complex analytic likelihoods.

4. Results

The method was validated using mock samples extracted from the UNIT1i N-body simulation (Planck 2015 cosmology).

Parameter Recovery:
- The SBI method successfully recovered the input cosmological parameters ( $\Omega_m, \sigma_8, H_0$ ) and M-L scaling parameters within their uncertainties.
- Degeneracies: The method correctly identified the known degeneracy between $\Omega_m$ and $\sigma_8$ . It also revealed a correlation between $H_0$ and $\Omega_m$ driven by the survey volume vs. mass function trade-off, which becomes significant in large survey areas (1000 deg²).
Goodness-of-Fit: A modified $\chi^2$ test (Poissonian log-likelihood) confirmed that the observed mock data is statistically consistent with the posterior predictive distribution (p-value $\approx$ 0.5).
Systematic Uncertainty Analysis:
- Mass Function Choice: Comparing four different HMF models (Tinker, Despali, Watson, Bocquet) showed that for a 50 deg² survey, the choice of HMF is a subdominant source of error compared to statistical noise.
- Sample Variance: Variations across different 50 deg² regions of the simulation were found to be negligible compared to statistical errors.
- Mass Calibration (Critical Finding): The absolute mass scale normalization is the leading source of systematic error. A $\pm 10\%$ shift in the Mass-Temperature normalization caused a $\sim 17-28\%$ shift in $\Omega_m$ and a $\sim 5-11\%$ shift in $\sigma_8$ .
- Robustness of $S_8$ : Crucially, the quantity $S_8 = \sigma_8(\Omega_m/0.3)^{0.3}$ remained nearly constant (varying by only $\sim 4\%$ ) despite large shifts in $\Omega_m$ and $\sigma_8$ . This indicates that $S_8$ is robust against mass calibration uncertainties.
Survey Size: Increasing the survey area from 50 deg² to 1000 deg² significantly reduced parameter uncertainties, demonstrating the method's suitability for future large-scale surveys like eROSITA and Euclid.

5. Significance

This paper establishes Simulation-Based Inference as a promising and robust alternative to traditional analytic methods for galaxy cluster cosmology.

Flexibility: It allows for the inclusion of all available observables and complex selection effects without simplifying assumptions.
Precision: It demonstrates that with accurate mass calibration (better than 10%), cluster counts can provide competitive constraints on $\Omega_m$ and $\sigma_8$ .
Future Outlook: The method is particularly well-suited for upcoming large-area X-ray (eROSITA) and optical (Euclid) surveys, where handling complex systematics and marginalizing over scaling relations is critical for resolving tensions in the $\Lambda$ CDM model (such as the $S_8$ tension). The finding that $S_8$ is robust to mass calibration errors suggests it may be a more reliable cosmological probe in the near term than individual $\Omega_m$ or $\sigma_8$ constraints.