WorksPhysics and Astronomy
Deconstructing the cosmic baryon distribution with fast radio bursts
Abstract
The precise distribution of baryons in the Universe remains an open question in cosmology, as their highly diffuse nature hinders direct observation. However, mapping it is integral for precision cosmology in studies of gravitational lensing, galaxy evolution and the cosmic microwave background with current and upcoming surveys. We use fast radio bursts (FRBs) and their dispersion measures (DMs) to probe and localise this baryon distribution in the large-scale structure (LSS). Using IllustrisTNG simulations, we classify the LSS into eight bins of increasing matter density, analogous to voids, filaments, and haloes. We model the DM contribution from each substructure and compare it with the simulated DMs to infer the density estimator, the ratio of electron number to total matter density. The resulting overdetermined system is solved in a Bayesian framework with a simulation-physics-informed model. To demonstrate observational applicability, we create a Euclid-like mock galaxy catalogue and build a density estimator based on luminous matter along FRB sightlines. We apply a bias correction for the underlying dark matter distribution and compare the two density estimators to recover the electron number density across the eight bins. Electrons are predominantly localised in low-density regions, while the ensemble mean across all bins aligns with the global theoretical estimate. Combining the FRB and galaxy density estimators, we recover the density of ionised baryons as (simulation: 0.0462), corresponding to for . Our bin-wise electron density estimate reproduces the simulation ground truth to within 10 per cent and gives gas fractions of , and in voids, filaments and haloes, against ground truths of 0.167, 0.173 and 0.067.
Introduction
Fast radio bursts (FRBs) are bright and typically millisecond transient signals detected in the radio band of 100 MHz to 10 GHz with observed flux in the range of 50 mJy to 100 Jy [140, 110]. After being discovered by Lorimer et al. [93], the number of FRB detections has increased by orders of magnitude due to surveys such as the Australian Square Kilometre Array Pathfinder (ASKAP) [99], the Canadian Hydrogen Intensity Mapping Experiment Fast Radio Burst project (CHIME/FRB) [58], Apertif [108], Five-Hundred-Meter Aperture Spherical Radio Telescope (FAST) [106], MeerKAT [19], FLIMFLAM (Khrykin et al. 2024) and DSA-110 [87]. An updated compilation of FRB sources is available in Xu et al. (2023). These discoveries have revealed that FRBs are predominantly of extragalactic origin, with distances spanning from 3.6 Mpc [13] to a redshift of [29], or more recently discovered at by Caleb et al. [20].
One characteristic observational property of an FRB is its dispersion measure (DM), which arises from the frequency-dependent phase velocity of electromagnetic waves propagating through the plasma in the intergalactic medium (IGM) [125]. In the local Universe (), where the baryonic component is largely ionised, it is proportional to the electron column density along the line of sight (LoS) and hence traces the underlying baryon distribution and scales with redshift.
This DM- or Macquart relation has been used to estimate the total baryon content of the Universe with localised FRBs [95], establishing it as a pivotal probe of the missing baryon problem, where the total baryon estimate from Big Bang nucleosynthesis (BBN) and the cosmic microwave background (CMB) [37, 31, 113] do not coincide with the late-time observations [59, 130]. A recent review of the baryon problem and possible solutions can be found in [117].
Further studies suggest that the majority of baryons are expected to be in a diffuse form, occupying the low-density regions outside dark matter haloes or the warm-hot intergalactic medium (WHIM). Traditionally, techniques like the kinetic (kSZ) and thermal Sunyaev-Zel’dovich (tSZ) effect (Bonjean et al. 2018; Tanimura et al. (2020), X-ray emission and absorption (Eckert et al. [45]), OVII absorption line (Nicastro et al. [105]) and Lyman- forest ([109] (); Hu et al. 2024)) are employed to probe the diffuse gas in the WHIM. The critical parameter dictating the effectiveness of different probes is the temperature. The SZ effect is sensitive to higher average temperatures of K, whereas Lyman- detects regions at K. Moreover, the WHIM region is very faint in electromagnetic radiation due to its diffuse nature. The absorption spectra also rely on bright background sources, such as quasars, or on stacking multiple observations to increase the signal.
The DM of FRBs, particularly the scatter in the DM- relation, serves as a complementary probe for localising the inhomogeneous baryon distribution ([72, 147, 29]). As the DM arises from interactions with electrons along the LoS, it can effectively trace diffuse baryons regardless of temperature or density. While Macquart et al. [95] used the homogeneous DM- relation to yield the total baryon estimate consistent with estimates from BBN and the CMB, its distribution in the Universe remains largely unknown. This mapping is critical for studies of galaxy evolution ([8] (Khrykin et al. 2024)). Particularly, FRBs can constrain feedback mechanisms like active galactic nuclei (AGN) and supernovae ([119, 129]), which in turn constrain the star formation rate ([17]). Similarly, weak lensing (WL) is dependent on comprehensive baryon distribution maps for unbiased inference, particularly at non-linear, smaller spatial scales, where matter distribution differs significantly from the collisionless dark matter distribution due to baryonic feedback effects ([26, 128]). The kSZ effect, similar to the DM, is the integrated effect of the electron number density and the velocity field, where a robust knowledge of the baryon distribution can isolate the velocity field to measure the growth parameter ([96, 137, 23]).
Since the baryon distribution in the Universe is inhomogeneous, models focus on decomposing the total integrated DM into contributions from different components of the cosmic web. Prior studies with cosmological observations and simulations partition the total DM into structures like voids, filaments and haloes, where results indicate that the dominant contribution arises from filaments (see e.g. [76, 157, 146]). While they provide a qualitative or categorical estimate of the DM contribution, a quantitative description remains a significant challenge. Furthermore, as the DM is an integrated signal, for real data, it alone cannot localise the baryons in the cosmic web. We require additional tracers of the spatial density distribution between the FRB source and the observer to recover the density-dependent baryon information. Therefore, a framework utilising the observational data alongside simulation-based insights is essential.
In this paper, we propose such a method, using galaxy surveys as tracers of the density distribution, and explicitly linking the FRB DM distribution and the density distribution in the large-scale structure (LSS) via a simple linear density estimator. Here, density estimation is the process of inferring a continuous underlying field from discrete points. The classification of that continuous field into eight equidistant logarithmic bins of the underlying density, analogous to voids, filaments, and haloes, but without any ambiguous boundaries, disentangles the integrated DM value into individual contributions. This framework not only provides a global estimate but also describes the baryon distribution across different sub-regions of the LSS.
To validate this framework, we simulate data using snapshot information from the IllustrisTNG project, a large-scale cosmological hydrodynamical simulation. It consists of various tracers and sub-grid physical models, calibrated against observational data, to capture small-scale effects. For multiple FRB sightlines, the density estimator is defined as the ratio of electron number density to the total mass density of the Universe.
Our analysis integrates observational estimates with the simulated density field by comparing the FRB density estimator with matter density probed via (upcoming) spectroscopic galaxy surveys mimicking observations with e.g. the Dark Energy Spectroscopic Instrument (DESI; Abareshi et al. [1]), Vera Rubin Observatory Legacy Survey of Space and Time (LSST; Ivezić et al. [75]), Euclid (Euclid Collaboration et al. [53]), Nancy Grace Roman Space Telescope (Roman; [131, 56]) and SPHEREx (Crill et al. [35]). Specifically, in this work, we create a Euclid-like mock catalogue as a subset of all the subhaloes in the IllustrisTNG simulations.
For real observational data, we require localised or host-identified FRBs that fall within the footprint of galaxy surveys. In particular, with upcoming surveys such as the Canadian Hydrogen Observatory and Radio-transient Detector (CHORD; Vanderlinde et al. [142]) and the Deep Synoptic Array (DSA)-2000 (Hallinan et al. [67]) detecting FRBs each year, the localised number is expected to increase by several orders of magnitude. Similarly, galaxy surveys with large sky coverage, such as Euclid and LSST, will enable a precise comparison of the density estimators presented in our work.
The paper is structured as follows. In § 2, we describe the observable properties of FRBs. § 3 presents the simulation setup used to calculate the physical properties, followed by the density estimation in § 4. In § 5, we create a density estimator using a mock galaxy catalogue. § 6 contains the principal results from this work. Finally, we discuss our findings in § 7 and conclude in § 8.
The fiducial cosmology in IllustrisTNG is set from the Planck-2015 cosmological results (Planck Collaboration et al. [112]) and in this work we use the same to maintain consistency, i.e., , , , , and km s Mpc.
Fast radio bursts
The radio signal emanating from an FRB disperses due to interactions with ionised plasma, which is characterised by a frequency-dependent time delay in the arrival of the signal, i.e. , where is the frequency, and DM is called the dispersion measure. This DM is a proxy for the integrated electron density between the FRB source and the observer, generally expressed as
where is the physical free electron number density at a redshift of , is the speed of light and is the Hubble function. The total dispersion is divided into three primary components based on the sky position and redshift via
where is the contribution from the Milky Way, is from the LSS of the Universe in the LoS towards the FRB, and finally is the contribution from the host galaxy of the FRB progenitor.
Note that the Milky Way contribution is local and depends solely on the direction. The host contributions can vary based on the position of the FRB progenitor inside the host galaxy [120], and it is redshifted with a term for the cosmological redshift. As a result, the positional dependence of DM inside the host galaxy is suppressed in the observed DM value. There exist models like YMW16 [152], NE2001 [32], or the updated NE2025 [107] based on pulsar observations to calculate the Milky Way contribution. However, with a lack of a global FRB population, the host contribution is the least constrained, and studies focus on a statistical estimate with a log-normal distribution based on analytical and hydrodynamical insights (see e.g. [76, 118, 100, 120, 83]). Despite these contributions, for an FRB at cosmological distances, most of the signal is encoded in the LSS part.
To study the electron distribution in the LSS, therefore, we focus on the part and substitute the electron number density in Eq. (1) as an integral over redshift [73, 72]:
is the Hubble constant, is the baryon density in the present Universe, is the gravitational constant, is the proton mass, is the unitless electron density contrast and is the expansion function. Additionally, is the fraction of baryons in the diffuse ionised phase, i.e. excluding baryons in stars, compact objects, and the dense neutral interstellar medium, and is the electron fraction calculated via [39, 147, 95]:
where and are the hydrogen and helium abundances [113, 3]. Note that we are considering a redshift of , where the IGM is effectively fully ionised [7].
The mean of Eq. (3) is called the DM- or Macquart relation, which probes the total baryon content of the Universe [95]. The inhomogeneous part, on the other hand, leads to sample variance between FRBs observed at the same redshift and encodes the distribution of baryons in the Universe along with baryon feedback effects via AGN and supernovae [119, 129].
Simulation
Classifying the LSS and studying its contribution to the DM observations requires density contrast, which cannot be probed directly via observations. However, if we have a proxy for the density along the LoS together with a relation between total density and electron density, we can use an ensemble of FRBs to infer the relation statistically. To validate such a relation, we require the ground truth. Hence, we use cosmological hydrodynamical simulations, where the underlying relation is directly accessible. Generally, such simulations evolve a large number () of particles in time in a three-dimensional box of cosmological dimensions. The evolution is tracked via snapshots at multiple redshifts. Various simulations differ in their scale, box size, resolution, particle mass, and implementation of physical laws.
Specifically, simulations such as Magneticum [41], MICE [114], EAGLE [6], RAMSES [158], and CROCODILE (Zhang et al. 2025) exist and have been used in cosmological and astrophysical studies of FRBs.
IllustrisTNG
Here, we use TNG300-1 simulations from the IllustrisTNG project [104, 111, 133, 103, 97], which is a large-scale, cosmological, magnetohydrodynamical (MHD) simulation that includes gas, dark matter, tracers, stars, and black holes. In addition, it contains observationally calibrated sub-grid models for star formation, gas cooling, stellar and AGN feedback [148, 111]. The simulation evolves the particles from an initial redshift of to a final redshift of while capturing the evolution with 100 snapshots. There are two types of snapshots, full and mini, where mini contains a subset of the full information3.
The choice of IllustrisTNG is motivated by the advantage of a higher resolution over Magneticum, better gas dynamics than MICE, larger box sizes compared to EAGLE, more pre-computed fields than RAMSES and different combinations of the kinetic and thermal feedback models based on AGN accretion, which is lacking in CROCODILE (Zhang et al. 2025). Moreover, since we are concerned with the LSS contribution to the DM in Eq. (2), we adopt the largest-volume simulation, i.e., TNG300-1. The redshift range is set to to match the current host-identified FRB sample at a redshift of [29]. Note that we choose to exclude the highest entry at discovered by Caleb et al. [20] because of the lack of localised FRB samples in between. Choosing to do so would allow the extrapolation to a single data point, which is not ideal because we would lose information about scatter. With that in mind, we use full snapshots at redshifts 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.7, 1.0 and 1.5. With a box size of 205 Mpc, there are empty regions or gaps in comoving distance, where we reuse sections of the snapshots to interpolate (more in § 3.3). The following sections describe the calculation of all the required physical quantities in these snapshots.
Electron density
Simulating the DM values requires the electron density in Eq. (1), or equivalently, the electron density contrast in Eq. (3). However, IllustrisTNG does not natively provide electron density as a field. Hence, we follow the formalism laid out in [76, 157, 146]. The “ElectronAbundance” field, defined as the fractional electron number density with respect to the total hydrogen number density, can be utilised. Moreover, IllustrisTNG includes both star-forming and non-star-forming Voronoi cells, which are the spatial discretisation units on which the MHD fluid dynamics are run. The star-forming cells, probed by “StarFormationRate”, consist of gases in both warm and cold phases, for which the simulation output needs to be inverted to extract the free electrons from the subgrid model (Springel & Hernquist [132]). However, Konietzka et al. [84] found that neglecting the contribution from star-forming cells yields only a sub-per cent deviation from the total DM estimate and is therefore omitted here. Then, the free electron density of each cell is
where is the electron abundance given by the “ElectronAbundance” field, is the gas cell comoving mass density and is proton mass. Here, electron abundance represents the number of free electrons per hydrogen nucleus, which implicitly accounts for the contribution from ionised helium, i.e., in a fully ionised region, . As we compute the electron density using the “ElectronAbundance” field, no additional ionised baryon fraction is applied to Eq. (5).
Dispersion measure
With the electron density for each Voronoi cell, we can approximate the integral in Eq. (1) as a sum of incremental distance elements. Specifically, the discretised version of the line-of-sight integral is
where and are the physical electron number density and differential physical path length on the LoS. The physical path length can be calculated by multiplying the comoving distance in simulation units by the snapshot’s scale factor. Note that in each IllustrisTNG snapshot, the redshift or cosmic time is constant, i.e., there is no evolution along the line of sight inside one box. Hence, the redshift dependence of the DM is not written explicitly. As for the spatial vector (i.e., direction), we define it implicitly and drop the notation for brevity.
The full TNG300-1 snapshots are only available at a specific redshift with a length of . Hence, the discrete sum in Eq. (6) is not continuous in redshift, i.e., the full snapshots used in § 3.1 do not overlap in comoving distance. One of the solutions is the trapezoidal rule with interpolation as presented in Zhang et al. [157]; Walker et al. [146]. This interpolation or stretching, however, distorts the DM distribution, specifically affecting the higher-order moments, leading to a biased estimate of the cosmic variance of the DM, i.e., the interpolation produces a lower scatter in the DM– relation than otherwise expected from different LoS through the LSS (Konietzka et al. [84]). Improvements include placing the periodic boxes parallel to each other, where a LoS traces the same structure multiple times, or randomly rotating the boxes before placing them. The latter avoids some level of repetition but introduces incorrect transitions at the simulation boundary due to random rotations. However, we choose this method because it is computationally feasible and yields a better estimate of cosmic variance than the trapezoidal method.
Note that because the primary goal of this work is to demonstrate that the electron-to-total mass density profile can be successfully reconstructed from a set of DM observations, the spatial correlation introduced by reusing simulation boxes is not critical. This approach, however, results in an underestimation of the error bars, since some LoS in the simulations might not be fully independent. Nevertheless, it does not affect the core results. For any missing region between two snapshots, we use a slice of the snapshot parallel to the axis that has a lower redshift than the next available snapshot. As the underlying cosmology is known and the LoS vector is parallel to the axis, this slicing is exact and implicitly avoids overlap.
Having established a continuous LoS in the simulation, we focus on the mock FRB population. In our work, we study the precise cosmological contribution of the LSS to the DM. Therefore, the redshifts of the FRBs are required. While the true redshift distribution of FRBs is unknown at this point, we create the mock redshift distribution following the works by Chen et al. [24]; Zhang et al. (2025). The unnormalised probability distribution function for the redshifts is parametrised by
where is the comoving distance, and we combine the number density evolution scale of with a cosmological factor of , along with the differential comoving volume element. We restrict the final sampled redshifts to . This is a catalogue selection motivated by the currently available localised FRB sample rather than a physical claim that the intrinsic FRB rate vanishes above this redshift.
The sky positions are randomly sampled from and , leading to a random sky distribution, which is then converted to Cartesian coordinates using the underlying cosmology. We apply periodic boundary conditions to ensure the FRBs remain within the simulation box.
Starting with a snapshot at a given redshift, we use the coordinates of the full box to select a LoS mask, with the vector parallel to the -axis and spanning the whole box. The centre of each LoS is the coordinates of the FRB sample. Afterwards, we use all necessary fields inside this mask to calculate the electron density in Eq. (5). The distance element, , is simply the derivative of the increasing -coordinates due to the parallel LoS vector. Finally, the DM is the sum over segments of the physical electron density times the physical path length, divided by of the corresponding snapshot. We derive the expression in the following section.
Decomposition
In each LoS, the measured DM contains contributions of ionised electrons from many over- and under-dense regions of the LSS. Here, we decompose this integrated signal into contributions from specific parts of the cosmic structure. Generally, in the works by Haider et al. [66]; Martizzi et al. [98]; Artale et al. [2], the LSS is classified based on the ratio of dark matter density to critical density, . In this scheme, typically three distinct regions are defined, namely voids (), filaments () and haloes ().
However, these boundaries are a little arbitrary and ambiguous, with significant overlap, and the definition itself is method-dependent (Libeskind et al. 2018). Hence, in this work, we define 8 logarithmic bins on the underlying density without labelling them as any morphological structures. These 8 bins are equidistant in log space, where the minima and maxima are the two extrema of the snapshot density. This ensures we account for the full range of densities within the snapshot.
Next, we derive the general expression of the discretised DM for an arbitrary number of LoS and density bins. Let there be sightlines that are indexed by (). The differential path segment on a single LoS is indexed via () and finally, we have logarithmic density bins indexed by (). For the -th sightline, the observer frame DM is written as
where is the redshift of the -th path segment, which is constant within the individual IllustrisTNG snapshots. Both electron density and the distance increment are in physical units here. Note that for real observations, we cannot directly probe the electron density; instead, we measure the mass overdensity traced by galaxies. Hence, we introduce the ratio
where is the electron density and is the total mass density. In a later section, we discuss how we derive from the galaxy population.
An analytical expression for the mean of this ratio across all density bins can be derived. Starting with the definitions of electron number density and matter density, one finds
where () is the baryon density, () is the matter density and is the mean molecular weight per free electron. Note that the DM arises due to contributions from diffuse ionised baryons expressed via in Eq. (3).
Here, we adopt , the value as given for TNG300-1 at ([136, 84]). Further, we keep this value constant and assume no redshift evolution. Substituting the values yields , where the uncertainty propagates the errors on and (neglecting their positive correlation). Within the simulation, these parameters are exact inputs, so this uncertainty is only relevant for observational data. Since is a global ratio, it corresponds to the mass-weighted mean of the bin values, i.e.,
where is the total matter mass in bin . We also note that this value is used as a reference, and not imposed in the inference. It enters neither the mock DMs (Eq. (5)) nor the configuration matrix that we derive in the following.
As stated earlier, with this definition of the ratio and 8 log bins, no classification of the LSS into voids, filaments, and haloes is necessary. First, we convert the total matter densities, accessed via the “SubfindDensity” field, from simulation units to physical units. Then, the minimum and maximum bin edges are the extrema of the matter density on a logarithmic scale. Finally, there are equidistant bins.
Substituting Eq. (9) into Eq. (8), the DM for a single LoS is
where is the realisation of the electron-to-matter ratio in cell of the simulation snapshot. All the quantities here are in physical units, related to the comoving quantities via and , with being the redshift. To pivot the calculation on the density bins, we partition the set of cells into bins ,
As the ratio is modelled as a function of density, we approximate the cell-specific value by a common bin-averaged value for all cells belonging to density bin . Therefore, we can write
where we have defined the configuration matrix as
Finally, for all sightlines and logarithmic density bins, we can stack them in matrices with the compact notation
where is the observer-frame DM, is the configuration matrix, and is the density estimator as the ratio between electron number density and total matter density. The last term, , is the measurement error. As the total DM has three distinct contributions in Eq. (2), the Milky Way and host parts are simply subtracted from the total DM. Here, we simulate the LSS contribution and omit the remaining components because the density estimator is designed to detect contributions from various regions of the LSS. Adding and then subtracting the Milky Way and host DM values results in a net-zero effect.
Assuming the noise is independent and uncorrelated, we model the measurement error as a zero-mean Gaussian with standard deviation 20. Then the entries are random draws from such a distribution, i.e., , where . The choice is based off the average instrumental noise of in the second CHIME catalogue ([58]), a larger coming from the galactic DM models like NE2001, NE2025 and YMW16, and finally, from the host model. In comparison, the sightline-to-sightline noise dominates at and contains the cosmological information of the inhomogeneous LSS ([119]).
Note that Eq. (15) is a linear system in , and one can estimate the least-squares or the maximum likelihood solution when a trivial inversion is ill-posed due to over-determinism. That is the principal approach of our work, where given the observed values of and for an FRB dataset, the solution yields a density estimator .
Density estimation with FRB
To ensure that the sightlines are largely independent, we quantify how similar the density structure traversed by each pair of LoS is. For the -th LoS, we compute the mean gas density in slices along the LoS vector, averaging over the Voronoi cells within a transverse column of width ; is the length of the snapshot box. Then, the correlation is calculated as
where and are the mean and standard deviation of along the LoS. Two sightlines that cross through the same structures have close to unity, whereas statistically independent LoS yield . We compute in the snapshot and retain a LoS only if . This choice preserves per cent of the simulated data while minimising redundant, closely clustered LoS. Based on this threshold, 470 LoS are selected for the final analysis. We discuss the effects of this choice on our results in § 7.1. Physically, a higher correlation implies two sightlines probe the same cosmic structure and therefore bias the expected variance. Note that in real observations, the angular separation between FRBs is typically much larger than in our simulation snapshot, which is 205 Mpc.
In the general case, if , i.e., we have more FRB DM values than logarithmic bins, Eq. (15) is over-determined. Here, we have and . Assuming the signal and noise are Gaussian, the maximum likelihood estimate (MLE) is
where is the covariance matrix and the term is the Fisher information matrix. The full covariance matrix has four distinct contributions: three corresponding to the quantities in Eq. (2) and a noise covariance. The covariance for the Milky Way can be obtained from models like NE2001, NE2025, or YMW16, while the host component comes from the variance of a log-normal distribution ([83]). Note that the Milky Way and the measurement uncertainties are subdominant compared to the variance in the LSS, and the host contribution gets suppressed with the factor. Consequently, to a good approximation, the final covariance is dominated by the variance coming from the LSS part, i.e.,
We therefore retain only the diagonal terms, approximating the LoS as statistically independent after imposing the selection.
For a real FRB dataset, the observed DM already contains all the components in Eq. (2). The full covariance, in that case, is the summation of the individual covariances, which can be estimated following Reischke & Hagstotz [119] for a Gaussian likelihood, or one can implement a likelihood-free inference using forward simulations ([83]). Note that the covariance depends on the angular separation between FRBs on the sky, with sightlines converging on the observer. However, in our case, the LoS vectors are parallel to the axis, so we adopt a diagonal covariance after imposing the selection. Even though it is not critical with the current observational dataset, as the number of localised FRBs increases, precise cosmological inference would rely on accurate subtraction of the Milky Way and host contributions from the total DM.
The form of Eq. (15) and its inversion for an over-determined system is a well-documented problem in statistics and regression modelling (see e.g. Cox [34]; Ruymaart (1993); Donoho [42]; Bissantz et al. [14]; Stuart [135]). As there may be multiple solutions for that reproduce the data for a given matrix, the inversion can be ill-posed, especially for low-density bins that contribute less information to the measured DMs. The MLE solver in Eq. (17) for such a case is noisy. Therefore, a maximum a posteriori (MAP) solution is preferred that includes a regularisation scheme like Tikhonov regularisation to the Fisher matrix in Eq. (17) ([63]). In such a regularisation scheme, the solution reads
where is the regularisation operator, which in its simplest form can be an identity matrix (i.e., Tikhonov regularisation). Alternatively, a finite-difference matrix can be used to obtain a smoother solution than Tikhonov. The term , a scalar, is called the regularisation scale or ‘ridge’ and essentially controls the amplitude of Tikhonov regularisation. This is equivalent to providing a prior probability distribution next to the likelihood function. The choice of can be optimised from the condition number of the Fisher matrix ([28, 11]). Similar to MLE, we find that the solution is also prone to noise, particularly at low-density bins and produces unphysical negative values.
Therefore, we implement a Bayesian prior with a parametric form on (see e.g. Dries et al. 2016, addressing a similar problem with a hierarchical approach). Recalling that we define 8 logarithmic bins on the density, let us define the bin centres as , i.e., . Then, the elements in are modelled using a scaled skewed Gaussian function as
with as the scale relative to the theoretical mean value , and () as the number of density bins, a fixed constant included only so that is of order unity. It is fully degenerate with and does not affect the reconstructed . In other words, is a function of density with a shape function that is defined as4
where is the location parameter of the profile, which approximately marks the transition from at low densities to at high densities, due to feedback. Similarly, is the width of the profile in , which dictates how rapidly the value of changes. Finally, the skewness parameter, , being positive or negative, determines whether the resulting distribution is right- or left-skewed, respectively, while corresponds to a Gaussian shape. Here, is normalised to unit area in . No normalisation condition is imposed on , so is a free parameter of the fit. Moreover, since the profile extends beyond the sampled density range, is neither the arithmetic nor the mass-weighted mean of over the bins. Therefore, these are the four free parameters in the inference.
This functional form is an empirical choice and represents a higher electron-to-matter density ratio in the LSS regions below , scaled by . At very low densities, we expect the baryons to follow the underlying dark matter distribution more closely. Alternatively, at densities above , it drops to lower values following the tail of a Gaussian distribution. Such a choice is motivated by the feedback effect via AGN and supernovae, which expels hot gas from the central part of the halo to the outskirts and is expected to be dominant in high-density regions, as shown in several hydrodynamical simulations (see e.g. [25] (Ayromlou et al. 2023; Khrykin et al. 2024)).
However, note that treating the four parameters independently implies that only the data dictate the amplitude and overall shape. As such, we impose wide uniform priors and list the prior ranges in Table 1 with a brief description of how each parameter alters the shape of the distribution in Eq. (20). As scales the theoretical value, an upper limit of four is extreme enough to describe all plausible DMs. In principle, one could impose narrower prior boundaries on from previous observational values of from CMB studies, as the mass-weighted mean of must reproduce the cosmic ionised baryon fraction. We choose to keep a wider boundary as an implicit validation test, avoiding a circular argument. A narrow prior can be implemented when observational data is used with unknown ground truth. The prior for the logarithm of the transition density (or the distribution’s location) is also set to be uniform over the two extrema. Similarly, the width of the distribution of densities controls the curvature of the profile, with the prior range allowing for a sharp drop or a flat profile. The skew parameter has a prior distribution over both positive and negative values, allowing flexibility for left- or right-skewed distributions. Combined, the prior choices provide a general profile to characterise the baryon distribution across the density bins. Alternatively, in § 7.2, we assess how a non-parametric model on performs with our dataset.
| Parameter | Prior | Description |
| Amplitude relative to | ||
| Location of the bin where feedback starts affecting | ||
| Determines the steepness of the drop from the flat profile | ||
| Determines the skew |
Table 1. Parameters and the prior ranges used in the inversion step.
As the functional form of the density profile is chosen in Eq. (20) based on empirical knowledge, we constrain the four free parameters that best describe the DM vector for a fixed matrix. Given an instance of the prior , the forward prediction is then
where is the parameter in Eq. (20). Then, the Gaussian log-likelihood can be written as
Together with the uniform priors, the final log posterior is
We utilise an affine-invariant ensemble sampler in emcee (Foreman-Mackey et al. [57]) to sample the posterior in Eq. (24). Specifically, we use 64 walkers, with a burn-in of 2000 steps, a reset, and a production phase of 4000 steps. We use the least-squares solution as the starting position for the walkers to achieve faster convergence.
Mock catalogue
The inversion presented in § 4 is implemented using simulated data, in which we used the total matter density directly from the simulations. This is inaccessible in a real scenario and requires a proxy. Hence, in this section, we create a mock galaxy catalogue using the subhaloes from IllustrisTNG.
In IllustrisTNG, dark matter haloes are first identified by the Friends-of-Friends (FOF) algorithm using a linking length parameter. Subsequently, the collapsed subhaloes are located using the Subfind algorithm (Springel et al. [133]). The subhalo catalogue in each snapshot contains all such subhaloes and their baryonic properties. Next, we treat these subhaloes or galaxies as luminous tracers of the LSS density field and compute a more realistic observations-driven density estimator.
Selection
IllustrisTNG natively provides the absolute magnitudes in photometric bands, namely U (300-400 nm), B (350-540 nm), V (380-455 nm), K(1800 - 2580 nm), r (570-590 nm), i (620-740 nm) and z (750 - 1400 nm) via the “SubhaloStellarPhotometrics” field in the subhalo catalogues. We remove the IllustrisTNG samples flagged with “SubhaloFlag”, which implies a non-cosmological origin.5
Next, we create a mock Euclid-like survey catalogue from these samples. Euclid is a stage-IV galaxy survey with a wide deg field (Euclid Collaboration et al. [53, 50]). It consists of four photometric bands, including the Visible Camera (VIS), which covers the 550-900 nm range, and the Near IR Spectrometer and Photometer (NISP). It has three bands: Y (950 - 1192 nm), J (1192 - 1544 nm), and H (1544 - 2000 nm).
As the Euclid filters do not exactly match the IllustrisTNG bands, we need to convert the photometric values into specific Euclid bands. In that regard, we calculate the flux and the corresponding K correction of each galaxy using Kcorrect6 (Blanton & Roweis [15]). Then, for a given flux and the Euclid filter response curves, as presented in Euclid Collaboration et al. [54], the AB magnitudes in Euclid bands are obtained using S\textsc{pecLite}.7 We select the galaxies based on Euclid selection criteria in the VIS band, i.e., and signal-to-noise (SNR) (Euclid Collaboration et al. [53]). Finally, we randomly select 60 per cent of the galaxies to imitate the sample incompleteness (Hamaus et al. [68]; Euclid Collaboration et al. [49]).
We provide a more detailed description of each step in converting magnitudes from IllustrisTNG simulation output to the final selection based on the Euclid specifications in Appendix A.§ 1. Essentially, this selection creates a subset of galaxies from the full subhalo catalogue.
The last required ingredient for density estimation is the halo mass, which is provided by the “SubhaloMass” field. However, in observations, the halo mass is a derived quantity. There exist multiple techniques to infer mass via scaling relations and indirect methods, e.g. velocity dispersion (Bernardi et al. [12]; Zahid et al. [154]), gravitational lensing (Hoekstra & Jain [69]), mass-to-light ratio (Taylor et al. [139]), and photometric information (Lower et al. [94]; Leja et al. [90]).
For Euclid data, stellar mass can be estimated, for example, from spectral energy distribution (SED) fitting or machine learning ([47]). Here, we use the colour information and luminosities calculated using the apparent magnitudes in the four bands. For a given luminosity, the mass can be estimated using the mass-to-light ratio . In Appendix A.§ 2, we present more specific details. Next, with a catalogue of mock galaxy positions and masses, we build a density estimator that coincides with the FRB LoS.
Density estimation using galaxies
Density estimation attempts to reconstruct a continuous underlying field using discrete data points or tracers of that field. Specifically, density estimation using galaxies assumes that luminous matter is a biased tracer of the underlying dark matter distribution ([143, 40]). Mathematically, we can write
where and are the galaxy and matter overdensities, respectively, related by the bias parameter .
A large body of work exists on various methods of density estimation and is broadly divided into three categories. Firstly, in the grid-based method, galaxies are distributed on a 3D Cartesian grid and weighted by their mass before applying a smoothing kernel ([38]). Next, in tessellation techniques, the kernel is made self-adaptive and has two major algorithms: Voronoi Tessellation Field Estimator (VTFE; [141, 71]) and Delaunay Tessellation Field Estimator (DTFE; [126, 122, 55]). Finally, the Bayesian Origin Reconstruction from Galaxies (BORG; [77]) method applies Hamiltonian Monte Carlo (HMC) sampling to create an initial field and then evolves it to reproduce the observed galaxy distribution. An extensive list of available tools implementing these methods, along with a broader overview and application, is presented in Contarini et al. [30].
In our work, we use Python libraries for the analysis of numerical simulations (Pylians8, [144]). It creates a continuous field on a 3D grid and progressively smooths the field from large to small spatial scales. This preserves the natural hierarchical structure of the LSS. We use Cartesian coordinates and the galaxies’ masses as inputs to the code. The values are the sky positions of a galaxy, while the coordinate is the comoving distance calculated from the redshift of the galaxy for the Planck-2015 cosmological model. In the fiducial mock catalogue, we use the galaxy positions in real space and do not include the peculiar-velocity contributions, i.e., redshift-space distortion (RSD). The effect of including RSD is discussed in § 7.3.1.
One of the largest sources of error here stems from uncertainties in the galaxies’ redshifts. We assume that we have spectroscopic redshifts for all the galaxy samples and include the upper limit of Euclid’s spectroscopic redshift error of ([50]). As the LoS vectors are parallel to the axis, this results in uncertainties in the galaxies’ positions, which are calculated using the underlying cosmology. The assumption of a full spectroscopic sample is not realistic, as most wide-field surveys capture photometric information of the galaxies, where the uncertainties are larger. Therefore, we systematically evaluate the effect of spectroscopic and photometric redshift uncertainties on the galactic density estimator in § 7.3.2. Similarly, uncertainties in the mass estimates derived from luminosities are included in the final error of the density estimator.
We choose as the grid resolution, deposit the galaxies on that grid, and add the mass via a mass assignment scheme like cloud-in-cell (CIC), triangular-shaped cloud (TSC), nearest grid point (NGP), and piecewise cubic spline (PCS) (see e.g. Cui et al. [36], for comparisons). We test the effect of the algorithm via cross-spectra and coherence for all possible pair combinations and find no noticeable influence. Finally, we use CIC. To test the sensitivity to grid resolution, we used two additional grids at and , while keeping the galaxy catalogue, CIC assignment and density-bin definition fixed. We discuss the findings in § 7.3.1.
Furthermore, we compare the quality of the Pylians cube with estimates from the Void Analysis Software Toolkit (VAST9, [43]) package, implementing both VoidFinder and V2 algorithms. For the same galaxy sample, the voids identified by VAST occupy per cent of the total simulation box volume, where the choice of algorithm has no noticeable effect. For the Pylians’ estimate, we defined a void as following [66, 98, 2] and the fraction is 83 per cent. Additionally, the centres of the voids in the VAST catalogue coincide with the centre of the voxels in the Pylians cube. This comparison holds as we use the same grid resolution in both cases.
For a given LoS vector for an FRB, the contributing voxels are divided into the same density bin definition used for FRBs. This creates a density estimator of the visible matter along the LoS. As galaxies are biased tracers of the underlying dark matter distribution, the density obtained from the galaxy survey is also a biased proxy for it. We can translate the bias in Eq. (25) as
where is the mean matter density, is the galaxy overdensity, and is the bias in the -th bin. We define as the density estimator for galaxies.
The bias correction term is the last required quantity and remains an active area of research, with various pre-existing methods available. Traditionally, the galaxy power spectrum is used to estimate the linear bias. The non-linear bias estimation requires independent probes. For example, the weak lensing signal, particularly galaxy-galaxy lensing, can be used, as gravitational lensing is sensitive to both visible and dark matter ([5, 116, 61]). Similarly, the velocity field is affected by the total matter, and the effect, redshift-space distortion (RSD), quantifies the bias ([80, 27]). Beyond the two-point statistics, galaxy bispectra and higher-order statistics can infer the bias, as shown in works by Gaztañaga & Frieman [60]; Pollack et al. [115]; Ivanov et al. [74]. Cross-correlation of galaxies with other tracers, such as cosmic microwave background (CMB) lensing maps ([127]) and the tSZ effect ([151]), also constrains the bias parameter.
Here, we calibrate the bias using values taken from the simulation itself. We calculate both the total matter and galaxy number overdensities, as in Eq. (25), on the same 512-grid and apply logarithmic binning. The result is the bias parameter in each bin. This calibration ensures the estimator is unbiased. However, for real observational data, when the galaxy distribution is converted to a density distribution, the noise is larger compared to our simulation-derived values.
Results
This section contains the primary findings of this work. We start with a density estimator based on FRBs and investigate how its shape encodes both gravitational and non-gravitational information that affects the baryons. Further, we estimate and compare electron densities across the eight binned regions by incorporating density estimators from FRBs and galaxies.
Reconstructed density estimator
First, we present the results from the density estimation using FRBs, described in § 4. Within the Bayesian framework, we used a parametric model in Eq. (20) with four free parameters, i.e., the scale , transition density , the width , and the skew parameter .
In Figure 1, we show the marginalised contour plots of the posterior distributions. The optimal parameter values, estimated as the means of the respective posterior distributions, are , where the quoted uncertainties represent the credible intervals.
Figure 1. Marginalised contour plots for all the parameters in our inference process. The posterior mean values are shown as grey dotted lines at .
Given , we reconstruct the density profile with the definition in Eq. (20) and denote it by . This is the density-ratio estimator obtained via FRBs. In principle, one could choose a functional form of any arbitrary complexity or even a fully non-parametric form for , provided the data can constrain such a model robustly. In § 7.2, we shortly compare a density estimator with a non-parametric prior on .
The profile of is shown in blue points in Figure 2, alongside the ground truth from the simulations in grey (dashed line). While the ground truth is inaccessible in real observational data, including it in the figure serves as a validation of our result. Crucially, our inversion for density estimation functions independently of the true values and requires only the configuration matrix and the DM values. We show the values in units of . The horizontal solid black line at unity represents , whereas the grey and blue lines denote the means of and . The mass-weighted mean of the simulated data (grey line) is , which confirms that the binned ground truth recovers the electron budget of the snapshot. The analogous value from the reconstruction yields , which we indicate with the dotted blue line.
Figure 2. The reconstructed values from the MCMC analysis calculated using Eq. (20) are shown in blue. The vertical dotted lines represent the bin edges, and the grey data points indicate the ground truth from the simulation. The horizontal black line at unity marks . The mass-weighted means of the ground truth and the reconstructed profiles are shown by dotted grey and blue horizontal lines at 1.0007 and 0.9860, respectively.
Recalling that summarises the ratio of the electron and total matter density per density bin, we observe that the values are greater (smaller) than unity for low (high) density bins. Such behaviour is caused by the interplay of gravitational interactions within the LSS, combined with non-gravitational baryonic feedback via AGNs and supernovae, which drives electrons from the central part of the halo or high-density bins to the outskirts with lower density. Additionally, gas inflows induce shock heating that persists since reionisation due to very long cooling and recombination times in the ionised IGM. Note that this rise at low densities is largely compensated by a fall at higher densities, so that the mass-weighted mean recovers the theoretical value to 1.4 per cent (). The theoretical mean is .
Figure 2 shows that electrons are mostly localised in low- to mid-density regions, consistent with the global theoretical estimate. The location of the highest value hints at the effects of gravitational interaction, suggesting most baryons are localised in low-density filamentary regions. The relative decrement of electrons in high-density environments is expected to be largely due to feedback, and, in principle, more bins can describe the baryon redistribution more robustly. Increasing the number of bins requires finer sensitivity to the contribution of individual bins to the total DM, which in turn requires more FRB sightlines. However, the number of LoS within the simulation box is not arbitrary and is limited by the correlation threshold (see § 7.1). Therefore, for this study, we restrict ourselves to eight density bins. Conversely, real observational data of localised FRBs can avoid this limitation due to larger angular separation.
In the next section, we anchor this density estimator to a more data-driven case using galaxy surveys and provide a quantitative estimate of the electron density in the eight bins.
Electron density estimation
The density estimator using FRBs is the ratio of electron number density and the total matter density, i.e., . Similarly, the density estimator from the galaxy survey, , is defined in Eq. (26). Then, the estimated electron number density can be expressed in terms of the two density estimators via
where is the reconstructed electron number density, and are the estimators from FRBs and galaxies, respectively, and finally, is the bias.
In the top panel of Figure 3, we show the result. The grey scattered points represent the electron density calculated from the full snapshot data, which serves as the ground truth in this case. The vertical lines are the bin edges used to classify the LSS. In blue, we show the electron densities, where the connecting line aligns well with the snapshot data across all the density regions.
Figure 3. Top: electron density inferred by multiplying from FRBs and from galaxies. In grey, we show the electron density obtained from the snapshot data, with blue dots indicating the reconstructed values. The associated error bars are shown with blue lines. The vertical dotted lines are the bin edges used in FRBs, which are the same as those in Figure 2. Bottom: the cumulative distribution function for Eq. (28) is shown. This describes the fraction of total electrons in each density bin.
To quantify the contribution of each density bin to the global electron budget, we can calculate the electron fraction for the -th bin via
where and are the electron density and the volume of the -th density bin, respectively and is the total number of electrons across all bins. Physically, this shows the fraction of the total electron content in each bin.
We show the equivalent cumulative distribution function (CDF) in the bottom panel of Figure 3. The distribution indicates that most electrons in the LSS are concentrated in low-density regions. Specifically, the first three to four bins contain approximately 75 per cent and 92 per cent of electrons, respectively, consistent with the findings in Martizzi et al. [98].
Further, we calculate the ionised gas fraction or , for the eight bins using
where the last equality uses . Therefore, the gas fractions depend only on the FRB-based estimator and not on the galaxy-based matter density, which enters only the electron densities and . The ground truth follows in the same way from (grey points in Figure 2).
Now, following Artale et al. [2], we categorise the eight bins into three regions (see Eq. (12) in Walker et al. [146]). Specifically, void, filament and halo regions are roughly covered by bin 1, bins 2-5 and bins 6-8, respectively. Note that the density boundaries separating these environments do not, in general, coincide with the nine bin edges of our eight logarithmic density bins. Therefore, this mapping is an approximate assignment of the discrete density bins to the three environments. To minimise the effect of their different volume occupancies and focus on their characteristic physical states, we assign equal weight to the density bins and use the arithmetic mean. Therefore, these values should be interpreted as comparative, equal-bin-weighted estimates rather than volume- or mass-weighted integrated gas fractions. The ground truth is grouped and weighted similarly, and the uncertainties are propagated from the per-bin errors as , where is the number of bins of environment . Note that this neglects the correlation between bins induced by the parametric profile in Eq. (20), so these uncertainties are approximate.
We obtain , and , compared with the simulation ground truths of 0.167, 0.173 and 0.067, i.e., within , and , respectively. As we use equal bin weights, the appropriate global check is the mass-weighted mean over all bins, , compared with in the simulation, consistent with .
In Figure 4, we show how the SNR of changes based on the number of FRB LoS. The SNR for each of the environments is calculated as the ratio of the mean to the uncertainty, i.e., , where . Since depends only on (Eq. (29)), the variation is coming from changes in the estimates at varying numbers of LoS. We note that with reduced LoS, the fidelity of the density estimator, particularly in the high-density bins, worsens (see § 7.1). However, averaging estimates from multiple bins into three results in a more stable value, with reduced uncertainty, as seen in the bottom panel of Figure 4. The halo estimate lies systematically below the ground truth. As the high-density bins occupy a small volume fraction, their columns of A carry little weight in the DMs, and after the correlation cut, only a few sightlines through haloes remain. The likelihood is therefore dominated by the low- and intermediate-density bins, and r at high densities is set largely by the tail of the parametric profile in Eq. (20). Moreover, the sample variance is very high, which implies that a precise measurement of the halo contribution would require a larger number of FRBs and hence a larger simulation volume.
Figure 4. Top: SNR values of for three regions are shown as a function of the number of LoS. The horizontal line represents the threshold. Bottom: The parameter for the three regions is shown in the respective colours, where the dotted lines represent the corresponding ground-truth values from the simulation, computed identically from .
Finally, we compute the density of ionised baryons as the discrete volume integral of the electron densities over all bins via , where is the bin index. We obtain , which agrees with the simulation value of 0.0462 () at the level. This is consistent with the mass-weighted mean , since . To calculate , we break the degeneracy using , and find , compared with the true value of 0.0486 ().
Discussion
In this section, we review the effect of various correlation thresholds for the LoS, the choice of a non-parametric prior model for density estimation, and the impact of RSD, grid resolution, and redshift uncertainties on the galactic density estimator. In summary, we assess the sensitivity of the fiducial results for the considered modelling choices.
Effect of correlation
To quantify the relative independence of the LoS, we compute correlations between all pairs of LoS vectors using Eq. (16), with values of 0 and 1 representing independent and perfectly correlated LoS vectors, respectively. Moreover, in our fiducial calculation, we adopt a correlation requirement of < 0.3. Here, we quantify its impact on the final reconstruction using nine correlation-threshold cuts ranging from 0.1 to 0.9. Physically, a lower threshold indicates that the LoS are more sparsely distributed, whereas a higher threshold implies greater clustering.
For the 0.1 threshold, the sampling preferably selects contributions from the low-to-mid-density regions (bins 1-4), as they occupy a larger volume fraction of the simulation box. At the same time, we undersample the high-density regions (bins 7 and 8). Combined, we are left with a limited number of LoS, resulting in a very low SNR () across all bins.
Alternatively, for a very high correlation threshold of 0.9, the sampling and thus the SNR increase in low-density regions. On the other hand, highly correlated LoS vectors may reflect contributions from the same haloes, where the redundancy does not necessarily translate into higher SNR. The corresponding DMs become highly correlated while the diagonal covariance treats them as independent, which leads to multiple counting of the same structure. The estimates in the high-density regions become dominated by a few repeatedly sampled structures. As a result, although the average SNR across all bins is higher, the diagonal-covariance inversion overestimates the amount of independent information and can assign excessive weight to repeatedly sampled structures.
Our test indicates that any threshold below 0.5 avoids overweighting repeatedly sampled structures in the high-density bins without appreciably reducing the average SNR. However, we select 0.3 as a conservative threshold. Increasing the threshold to 0.5 reduces the uncertainties in the high-density bins by only 3 per cent, which does not justify the additional correlation between sightlines.
The threshold also affects the final reconstructed profile of the density estimator. As pointed out by Walker et al. [146] using the same IllustrisTNG simulation boxes, the void contribution remains almost constant at per cent across redshifts. Our results suggest the same, as the model provides a smooth profile, partly due to the parametric modelling choice, rather than picking out the curvature as seen in the albeit low-signal-to-noise grey dashed line in Figure 2. The immediate solution is to increase the number of LoS of localised FRBs. However, because the currently used simulation box size is finite, this inevitably increases correlations among different LoS. Alternatively, one can adopt two distinct LoS schemes to sample the low and high densities independently. This approach ensures that the DM contribution remains sensitive to subtle changes in the low density, while also preventing redundant sampling of the same halo. For a real observation, the angular separations between FRBs are larger, which allows for a larger number of LoS that are mutually independent. Since we aim to apply this method to observational data, we do not further explore the two-way sampling scheme in the present work.
Effect of non-parametric prior
In this section, we contrast our results obtained using a parametric prior with those derived from a non-parametric specification. Specifically, we employ hierarchical Bayesian inference with a Gaussian prior centred on the theoretical mean value , whose covariance is hyperparameterised by the prior strength, smoothness, and correlation length scale. By analytically marginalising over , the posterior distribution can be expressed as
where denotes the hyperparameterised covariance matrix. For each posterior sample of , we then compute the conditional Gaussian posterior for and generate corresponding conditional realisations. The final posterior distribution for is obtained by aggregating these conditional samples and marginalising over the hyperparameters.
The results indicate an average correlation length of 2.63 bins (i.e., dex), which enforces relatively smooth variations across the eight bins. Nonetheless, within the non-parametric framework, the overall profile of exhibits a markedly stronger contribution from low-density regions than from high-density regions. As a consequence, the mean estimator is systematically biased towards values larger than .
Such behaviour in a non-parametric model with wide priors is due to model degeneracies. The total DM of an FRB is degenerate among the contributions from the Milky Way, the LSS, and the host galaxy, as seen in Eq. (2). In our approach, we simulate only the LSS contribution and then decompose it into contributions from eight LSS regions. This signal is then distributed among the bins. Recalling that the number of LoS is comparatively small for high-density regions, most of the information in the operator is concentrated in the low-density bins. Moreover, the contribution to the total DM from individual haloes is much larger than in void regions, leading to a log-normal distribution of observed DM values (Walker et al. [146]). This translates to a sharp increase in the respective column for a few LoS. The linear model and its inversion using and with a non-parametric prior, therefore, dampens the overall high-density contribution in favour of a higher low-density contribution. The parametric prior model in Eq. (20), on the other hand, imposes a mean density estimator or along with for the onset and steepness of the turnover and avoids the collapse to smaller values at high-density bins.
Furthermore, the conditional Gaussian posterior can produce unphysical negative values of . One solution is to use a truncated Gaussian or a log-normal distribution, which makes the marginalisation in the first step non-trivial. Therefore, a more complex non-parametric approach is not warranted over our simulation-driven parametric modelling approach with the current data.
Effects on the galaxy-based density estimation
RSD and grid resolution
To test the effect of RSD, we included the peculiar velocities of the galaxies along the LoS and repeated the density reconstruction at a fixed grid resolution. Compared to the fiducial real-space case, the galaxy density estimate increases by – per cent in low-density bins as RSD partly fills in the voids, and decreases by up to 20 per cent in the high-density bins because of structure smearing along the LoS. The mean absolute change over the eight bins is 14 per cent. As for the result, the density estimator based on FRB () and the gas fractions in Eq. (29) are not affected by RSD. It changes the reconstructed electron density in Eq. (27) by the same factors, and raises the recovered by per cent. Since real observations are in redshift space, our real-space estimator is optimistic in this case. RSD is the dominant systematic effect, and dedicated redshift-space calibration is beyond the scope of the work presented here.
The grid resolution of is another fiducial choice, and we repeat the real-space density-field reconstruction at and . This also sets the scales over which the density is averaged, which are 0.80, 0.40 and 0.20 Mpc. Therefore, the matter-density contrast also increases with increasing resolution, with , 25.14 and 35.73, respectively. At the same time, the galaxy field becomes sparser, and the voxel-level correspondence between galaxy and matter overdensities weakens at higher resolution, indicating increasing departures from the simple linear-bias approximation on smaller grid scales. For each resolution, the bias is recalibrated on the same grid. At , applying the bias calibration of systematically underestimates the mean matter density in bins 2-8 by 3-18 per cent. Moreover, in the lowest-density bin, most cells at this high resolution contain no particles, which makes the binned matter density dominated by discreteness. At , the estimator underestimates the matter density by 16 per cent on average. With the bin-wise changes, the recovered decreases by 0.4 and 0.7 per cent at and , respectively. This is primarily because the changes in the different bins compensate, leaving the overall shift in small compared with its statistical uncertainty. Finally, we adopt as a compromise between resolving the density field, maintaining adequate sampling of the galaxy distribution, and preserving the linear-bias approximation. The fiducial bias calibration uses the grid.
Redshift uncertainty
For the fiducial results, we assume a spectroscopic redshift for all galaxies with an upper limit of uncertainty of . In realistic surveys, it is unlikely that spectroscopic redshifts can be obtained for all galaxies. Hence, we include photometric redshift information to quantify its effect on density estimation. Euclid expects a photometric redshift uncertainty of [86]. However, such an error margin has a particularly significant impact on distance measurements at higher redshifts. For instance, as the simulation box has a side length of 205 Mpc, or the associated error in comoving distance, Mpc, is comparable to the simulation box, which implies the density estimates are wrong. For Euclid, the spectroscopic redshifts for the wide and deep fields are in the range and [50], where the effect of photometric uncertainty can be minimised with more accurate spectroscopic redshift values.
To test the impact of less complete galaxy surveys, due to the requirement for spectroscopic follow-up, we reduce the number of galaxies by reducing the final subhalo sample by 10 per cent, 50 per cent, and 80 per cent. This reduces the number of galaxies along the LoS and introduces errors in the estimation of galactic density. The noticeable effects are in the two extreme bins. A pairwise comparison shows that the low-density estimate for the third sample is 31 per cent and 27 per cent lower than those for the first and second samples, respectively. The same applies to the high-density yields of 53 per cent and 46 per cent. This reveals that galactic density estimation relies on both the number density and precise redshift measurements. Moreover, as there is no significant difference between the 10 per cent and 50 per cent selection, we expect a robust galactic density estimator for real observations with spectroscopic redshift on a subset of the full catalogue.
The electron density calculation relies on the mean matter density, which yields accurate mean values even with small galaxy samples. However, because structure growth is non-Gaussian, density estimates in regions of low galaxy density can yield erroneous estimates of higher-order moments due to interpolation in mass-assignment schemes. This constitutes a loss of cosmological information and can lead to potentially biased estimates.
Summary and conclusion
In this work, we decomposed the cosmic web into eight density regimes and estimated their (ionised) baryon content relative to the underlying total density field to quantitatively address the missing baryon problem. Below, we summarise our findings:
We employed fast radio bursts (FRBs) and their dispersion measures (DMs) to quantify the ionised baryon density along the line of sight. To simulate the DM, we used snapshot data from the IllustrisTNG project, a magnetohydrodynamical simulation that incorporates subgrid baryonic feedback, including AGN feedback. Decomposing the cosmic web into eight density bins and calculating the individual contributions to the total DM yielded a linear equation that maps the DM to the density in Eq. (15). We defined a density estimator in such a system as the ratio between the electron number density and the total matter density, so that it measures the relative abundance of electrons in each bin.
For the simulated DM values and the linear operator, we infer the density estimator using an empirical prior model in Eq. (20) within a Bayesian framework. In Figure 2, we show a recovery of 98.60 per cent of the global ionised baryon content, while also accounting for the bin-wise distribution.
Furthermore, to bridge the gap between simulation and observation, we integrated this FRB-based estimator with a
mock galaxy catalogue. To this end, we constructed a density estimator along FRB sightlines using galaxies as luminous tracers of the underlying dark matter density field, based on the subhalo catalogue from the IllustrisTNG simulations. The resulting catalogue is a Euclid-like mock subset that includes Euclid-specific detection criteria and instrumental effects, as described in § 5.1. This estimator provides a biased proxy for the total matter density. Therefore, we applied bias correction using the snapshot data.
iv. We obtained the electron density in each bin by combining these two density estimators from FRBs and galaxies. A comparison with the ground truth in Figure 3 reveals an accurate reconstruction. Furthermore, we evaluated the electron fraction and concluded that approximately 75 per cent of the free electrons are located in the first three bins with . We find the gas fractions in three substructures of the cosmic web defined in § 3.4 as , , and , against ground truths of 0.167, 0.173 and 0.067. Finally, we calculate the ionised baryon density as , which agrees with the ground truth of the simulation at the level, corresponding to for .
In conclusion, we demonstrate that the linear framework based on FRB dispersion measures, coupled with density estimators, can effectively probe diffuse baryons in the large-scale structure (LSS) of the Universe. By integrating mock observational data, we provide a real-world anchor for our simulations, establishing the groundwork for incorporating future datasets as they become available. The fiducial results not only agree with global estimates of the baryon distribution but also quantify the baryon content in different subregions of the LSS. Furthermore, our modelling approach allows us to trace the effects of baryonic feedback across density regions. This provides an alternative to the more traditional two-point correlation function or power spectrum.
While our framework is robust, it currently relies on simulation-calibrated bias corrections. For real datasets, this could represent the most significant systematic uncertainty. Furthermore, although IllustrisTNG has high resolution, its subgrid feedback modelling is fixed. Applying this methodology to suites that provide variable feedback prescriptions, such as FLAMINGO, would be a critical next step. Such an analysis could determine whether the density estimator itself can be used to directly constrain feedback strength by mapping the flow from high-density regions. The density estimator, however, could also be directly calibrated against weak-lensing observations, where available.
Looking ahead, we aim to apply the methods developed here to real localised FRBs from CHORD and DSA-2000, combined with galaxy surveys with Euclid and LSST, where we expect orders-of-magnitude improvement in number density and redshift information. At larger angular separations, this would also allow for more independent line-of-sight traversal, leading to better constraints with a larger number of density bins.
Acknowledgements. The authors thank Steven Murray for his helpful comments during the preparation of the manuscript and the funding support from Stellenbosch University. We thank the Center for Information Technology of the University of Groningen for their support and for providing access to the Hábrók high-performance computing cluster. Y.-Z. Ma acknowledges South Africa Research Chair Initiative (NRF) Grant No. RCA250325306198) from South Africa’s Department of Science, Technology and Innovation and National Research Foundation (NRF), and other NRF Grants with No. 150580, No. CHN2211069370, No. ERC250324306141, No. AUPP250310302114.
Details on mock catalogue
Catalogue using photometric data
Here we present a detailed overview of the mock catalogue created in § 5.1. We focus on the VIS band, and all the instrument specifications are obtained from the SurveyCodex10 package. The selection depends on Euclid’s detection limit of 24.5 (10) in the VIS and 24.0 (5) in the NISP band ([52, 48]).
We begin with the IllustrisTNG subhalo catalogue and the absolute magnitudes () in eight bands. First, subhaloes with floored absolute magnitudes () are removed from the full IllustrisTNG catalogue.
For a given absolute magnitude, the apparent magnitude () is calculated using the luminosity distance, which is known from the underlying cosmology and the redshift of the subhalo catalogue.
The SED fitting is carried out using Kcorrect, which expects input in linearised magnitudes or maggies, calculated from the apparent magnitudes () as .
The SED fitting step returns the rest-frame SED and the K correction values via coefficient fitting as shown in Fig. A.1.
Figure A.1. The SEDs of the first 50 subhaloes are shown. The colour bar shows the average apparent magnitudes of all the IllustrisTNG photometric bands. The throughputs of the four Euclid filters, which act as convolution kernels, are presented in the inset. The vertical dotted lines are the effective wavelengths of bands VIS, Y, J and H, respectively.
We then calculate the rest-frame flux and redshift it along with the wavelength.
We convolve with Euclid’s filter response curves from Speclite and calculate the AB magnitudes in the four Euclid bands. We show the apparent magnitudes from IllustrisTNG and their equivalent in Euclid bands in Fig. A.2.
Instead of a hard selection cut at Euclid’s magnitude limit, we base our selection criteria on the number of electrons () via
where is the apparent magnitude, is the VIS zero-point value and s is the exposure time. With this step, we can include instrument-specific noise.
Figure A.2. Top: Histogram of apparent magnitudes in IllustrisTNG photometric bands. Bottom: Equivalent bands in a Euclid-like survey obtained via SED fitting. The two vertical lines represent the magnitude cuts in Euclid at 24.0 (dot-dash) for NISP and 24.5 (solid) for the VIS band.
The sky brightness is similarly given as
where is Euclid’s sky brightness. Note that we multiply by the area, , so that is in electrons pixel.
To simulate the galaxy brightness profile, we utilise GALSIM11 (Rowe et al. [123]). Specifically, a Sérsic profile is used with an index of 1, and we set the half-light radius equal to the half-mass stellar radius from IllustrisTNG, accessed via the “SubhaloHalfmassRadType” field. The total flux is anchored to the number of electrons in the exposure to maintain uniformity.
Next, we add two convolution kernels
PSF: The point spread function (PSF) for Euclid is Gaussian with a full width at half maximum (FWHM) of 0.17”.
Optical: The optical effects are included as the Airy disc model in GALSIM using the effective wavelength of 750.637 nm for the VIS filter, a diameter of 1.2 m, and an obscuration of 0.13.
The final true image is the convolution of the Sérsic profile with the PSF and the optical effects.
Noise: We add a mean-zero Gaussian read noise of 4.4 electrons12 along with Poisson noise.
We choose a fixed aperture radius of 0′.6 and the corresponding flux is
where is the measured source electrons inside the aperture, is the sky electrons/pixel, is the read noise and is the number of pixels in the aperture, i.e.,
where arcsec/pixel is Euclid’s pixel scale.
The signal-to-noise ratio (SNR) is then
In the penultimate step, we impose the VIS detection criteria of a magnitude cut at and SNR on the full subhalo catalogue ([53]). A subhalo that satisfies both criteria simultaneously is considered a true detection and included in the final catalogue. See Fig. A.3 for a visual representation of the pipeline, where red/green indicates if the sample is rejected/selected.
Figure A.3. A visual representation of the Euclid detection pipeline. The top row is the full simulated image for a given apparent magnitude and stellar half-mass radius. The bottom row shows the masked selection of subhaloes, with the corresponding SNR values displayed. The green and red colours in the bottom panel indicate whether the sample is selected. In the first two columns, the magnitude is increased while the stellar half-mass radius is held fixed, thereby raising the SNR above the detection threshold. The third and fourth columns show the effect of this radius for a fixed magnitude. As the flux is concentrated in a smaller area, the SNR exceeds the detection limit.
As the sample selection in real Euclid observations is not complete, we remove 40 per cent of the selected samples following Hamaus et al. [68]; Euclid Collaboration et al. [49]. The remaining 60 per cent is our final mock catalogue.
For SED fitting, other alternatives are available, such as Bagpipes (Carnall et al. [22]), CIGALE (Boquien et al. [18]), and Prospector (Johnson et al. [78]).13 Moreover, note that we have not included any complex post-processing of the images before computing the flux. A more realistic and physically accurate approach is to use a radiative transfer code such as SKIRT14 (Camps & Baes [21]) and apply dust models ([134, 79]). For Euclid-specific data, another alternative is to use the dedicated package Phosphoros that calculates the photometric values and the bandpass effects (Euclid Collaboration et al. [51]). However, as we eventually create a 3D density field with the galaxies, the required level of accuracy does not warrant a more complex detection pipeline. Such detailed applications can be found in [121, 46, 65, 64].
Stellar and halo mass estimation
For our sample catalogue, IllustrisTNG provides the stellar and halo mass as ground truths. For a real galaxy survey, these are not accessible trivially and need to be inferred from observable properties. To showcase a more realistic application, we use the luminosities and the apparent magnitudes in the four Euclid bands in Appendix A.§ 1 to first estimate the stellar mass. Conversion from the apparent magnitudes to solar luminosities requires the absolute solar magnitudes, which are taken from Willmer [149].15 One of the most conventional ways to obtain the stellar mass is to employ SED fitting (see e.g. [101, 85]; and for a review Mobasher et al. [102]). However, such a dedicated application is not the focus of our study. With access to the true stellar mass from the simulation, we instead employ a random-forest regression to estimate the mass-to-light ratio, . Typically, one assumes - for galaxies with a non-zero star-formation rate (SFR) and - for quenched galaxies. In our model, we include the specific star formation rate and colour information via , where denote different filters, as features that allow a dynamic .
Further, we split the full dataset into 7:3 for training and testing. As the final target, we use 60 per cent of the true stellar mass, i.e., we assume that the true stellar mass for 60 per cent of the galaxies is known. Post-training, we obtain the stellar mass for the full sample, and the result is shown in the top panel of Fig. A.4.
Figure A.4. Shown are the ratios of the reconstructed and true stellar (top panel) and halo (bottom panel) masses against the true values from IllustrisTNG. The horizontal dotted lines are at unity and represent a perfect reconstruction. The grey scatter points in both panels show the calculated values, where the means are indicated by solid lines. The shaded areas are the 68 per cent error margins around the mean.
To go from stellar mass to the halo mass, we need the stellar-to-halo mass relation (SHMR, Girelli et al. [62]), which can be inferred, e.g. via subhalo abundance matching (SHAM, Legrand et al. [89]), weak lensing ([153] (Li et al. 2025)), neural networks (Villanueva-Domingo et al. [145]) or a halo occupation distribution prescription (HOD, [88, 33]). Here, we can use the relation from Behroozi et al. [9]16 and interpolate to estimate the halo mass for a given stellar mass. Note that this relation is non-trivially redshift-dependent (Legrand et al. [89]). However, with Euclid’s stringent spectroscopic and photometric uncertainty limits, the errors only from redshifts are negligible. The largest source of error in this conversion (0.15-0.25 dex) is the intrinsic scatter of the stellar mass at a fixed halo mass. The reconstructed halo masses are shown in the bottom panel of Fig. A.4.
References
- [1]Abareshi, B., Aguilar, J., Ahlen, S., et al. 2022, AJ, 164, 207
- [2]Artale, M. C., Haider, M., Montero-Dorta, A. D., et al. 2022, MNRAS, 510, 399
- [3]Aver, E., Skillman, E. D., Pogge, R. W., et al. 2026, ApJ, 1008, 240
- [5]Bartelmann, M. & Schneider, P. 2001, Phys. Rep., 340, 291
- [6]Batten, A. J., Duffy, A. R., Wijers, N. A., et al. 2021, MNRAS, 505, 5356
- [7]Becker, G. D., Bolton, J. S., Haehnelt, M. G., & Sargent, W. L. W. 2011, MNRAS, 410, 1096
- [8]Behroozi, P., Wechsler, R. H., Hearin, A. P., & Conroy, C. 2019, MNRAS, 488, 3143
- [9]Behroozi, P. S., Wechsler, R. H., & Conroy, C. 2013a, ApJ, 762, L31
- [11]Belsley, D. A., Kuh, E., & Welsch, R. E. 1980, Regression diagnostics: Identifying influential data and sources of collinearity (Wiley), 100–104
- [12]Bernardi, M., Shankar, F., Hyde, J. B., et al. 2010, MNRAS, 404, 2087
- [13]Bhardwaj, M., Gaensler, B. M., Kaspi, V. M., et al. 2021, ApJL, 910, L18
- [14]Bissantz, N., Hohage, T., Munk, A., & Ruymgaart, F. 2007, SIAM J. Numer. Anal., 45, 2610
- [15]Blanton, M. R. & Roweis, S. 2007, AJ, 133, 734
- [17]Booth, C. M. & Schaye, J. 2013, Sci. Rep., 3, 1738
- [18]Boquien, M., Burgarella, D., Roehlly, Y., et al. 2019, A&A, 622, A103
- [19]Caleb, M., Driessen, L. N., Gordon, A. C., et al. 2023, MNRAS, 524, 2064
- [20]Caleb, M., Nanayakkara, T., Snappers, B., et al. 2026, Science, eadz2675, in press
- [21]Camps, P. & Baes, M. 2020, Astron. Comput., 31, 100381
- [22]Carnall, A. C., McLure, R. J., Dunlop, J. S., & Davé, R. 2018, MNRAS, 480, 4379
- [23]Chaussidon, E., Hotinli, S. C., Ferraro, S., et al. 2026, Phys. Rev. D [arXiv:2604.04867], in pressarxiv.org/abs/2604.04867
- [24]Chen, J. H., Jia, X. D., Dong, X. F., & Wang, F. Y. 2024, ApJL, 973, L54
- [25]Chisari, N. E., Mead, A. J., Joudaki, S., et al. 2019, OJAp, 2, 4
- [26]Chisari, N. E., Richardson, M. L. A., Devriendt, J., et al. 2018, MNRAS, 480, 3962
- [27]Chuang, C.-H., Prada, F., Pellejero-Ibanez, M., et al. 2016, MNRAS, 461, 3781
- [28]Cline, A. K., Moler, C. B., Stewart, G. W., & Wilkinson, J. H. 1979, SIAM J. Numer. Anal., 16, 368
- [29]Connor, L., Ravi, V., Sharma, K., et al. 2025, Nat. Astron., 9, 1226
- [30]Contarini, S., Verza, G., & Pisani, A. 2026, A&A, 34, 1
- [31]Cooke, R. J., Pettini, M., & Steidel, C. C. 2018, ApJ, 855, 102
- [32]Cordes, J. M. & Lazio, T. J. W. 2002, arXiv e-prints, astro
- [33]Coupon, J., Arnouts, S., van Waerbeke, L., et al. 2015, MNRAS, 449, 1352
- [34]Cox, D. D. 1993, Ann. Stat., 21, 903
- [35]Crill, B. P., Werner, M., Akeson, R., et al. 2020, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 11443, Space Telescopes and Instrumentation 2020: Optical, Infrared, and Millimeter Wave, ed. M. Lystrup & M. D. Perrin, 114430I
- [36]Cui, W., Liu, L., Yang, X., et al. 2008, ApJ, 687, 738
- [37]Cyburt, R. H., Fields, B. D., Olive, K. A., & Yeh, T.-H. 2016, Rev. Mod. Phys., 88, 015004
- [38]Darvish, B., Mobasher, B., Sobral, D., Scoville, N., & Aragon-Calvo, M. 2015, ApJ, 805, 121
- [39]Deng, W. & Zhang, B. 2014, ApJL, 783, L35
- [40]Desjacques, V., Jeong, D., & Schmidt, F. 2018, Phys. Rep., 733, 1
- [41]Dolag, K., Gaensler, B. M., Beck, A. M., & Beck, M. C. 2015, MNRAS, 451, 4277
- [42]Donoho, D. L. 1994, Ann. Stat., 22, 238
- [43]Douglass, K., Veyrat, D., O’Neill, S., et al. 2022, J. Open Source Softw., 7, 4033
- [45]Eckert, D., Jauzac, M., Shan, H., et al. 2015, Nature, 528, 105
- [46]Eisert, L., Bottrell, C., Pillepich, A., et al. 2024, MNRAS, 528, 7411
- [47]Euclid Collaboration, Bisigello, L., Conselice, C. J., et al. 2023, MNRAS, 520, 3529
- [48]Euclid Collaboration, Cropper, M. S., Al-Bahlawan, A., et al. 2025a, A&A, 697, A2
- [49]Euclid Collaboration, Kraljic, K., Laigle, C., et al. 2026a, A&A, 708, A164
- [50]Euclid Collaboration, Mellier, Y., Abdurro’uf, et al. 2025b, A&A, 697, A1
- [51]Euclid Collaboration, Palatini, S., Coupon, J., et al. 2024, A&A, 681, A66
- [52]Euclid Collaboration, Romelli, E., Kümmel, M., et al. 2026b, A&A, 711, A4
- [53]Euclid Collaboration, Scaramella, R., Amiaux, J., et al. 2022a, A&A, 662, A112
- [54]Euclid Collaboration, Schirmer, M., Jahnke, K., et al. 2022b, A&A, 662, A92
- [55]Feldbrugge, J. 2025, MNRAS, 536, 807
- [56]Foley, R. J., Kosec, P., M. Spergel, D. N., et al. 2018, arXiv e-prints, arXiv:1812.00514arxiv.org/abs/1812.00514
- [57]Foreman-Mackey, D., Hogg, D. W., Lang, D., & Goodman, J. 2013, PASP, 125, 306
- [58]FRB Collaboration, Abbott, T., Andersen, B. C., et al. 2026, ApJS, 283, 34
- [59]Fukugita, M. & Peebles, P. J. E. 2004, ApJ, 616, 643
- [60]Gaztanaga, E. & Frieman, J. A. 1994, ApJL, 437, L13
- [61]Giannini, G., Camacho-Curcio, G., Whyley, A., et al. 2026, Phys. Rev. D, 113, 103503
- [62]Girelli, G., Pozzetti, L., Bolzonella, M., et al. 2020, A&A, 634, A135
- [63]Golub, G. H., Hansen, P. C., & O’Leary, D. P. 1999, SIAM J. Matrix Anal. Appl., 21, 185
- [64]Gonçalves, G. F., Machado, R. E. G., Valença, R. R., et al. 2026, MNRAS, 546, stag263
- [65]Gong, J.-Y., Lin, W., Tang, L., & Lan, Y. 2025, ApJS, 279, 16
- [66]Haider, M., Steinhauser, D., Vogelsberger, M., et al. 2016, MNRAS, 457, 3024
- [67]Hallinan, G., Ravi, V., Weinreb, S., et al. 2019, in Bulletin of the American Astronomical Society, Vol. 51, 255
- [68]Hamaus, N., Aubert, M., Pisani, A., et al. 2022, A&A, 658, A20
- [69]Hoekstra, H. & Jain, B. 2008, Annu. Rev. Nucl. Part. Sci., 58, 99
- [71]Hung, D., Lemáeux, B. C., Gal, R. R., et al. 2020, MNRAS, 491, 5524
- [72]Inoue, S. 2004, MNRAS, 348, 999
- [73]Ioka, K. 2003, ApJL, 598, L79
- [74]Ivanov, M. M., Philcox, O. H. E., Nishimichi, T., et al. 2022, Phys. Rev. D, 105, 063512
- [75]Ivezić, Ž., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111
- [76]Jaroszynski, M. 2019, MNRAS, 484, 1637
- [77]Jasche, J. & Wandelt, B. D. 2013, MNRAS, 432, 894
- [78]Johnson, B. D., Leja, J., Conroy, C., & Speagle, J. S. 2021, ApJS, 254, 22
- [79]Jones, A. P., Köhler, M., Ysard, N., Bocchio, M., & Verstraete, L. 2017, A&A, 602, A46
- [80]Kaiser, N. 1987, MNRAS, 227, 1
- [83]Konar, K., Reischke, R., Hagstotz, S., Nicola, A., & Hildebrandt, H. 2025, OJAp, 8
- [84]Konietzka, R. M., Connor, L., Semenov, V. A., et al. 2026, ApJ, 1010, 122
- [85]Kouroumpatzakis, K., Zezas, A., Kyritsis, E., Salim, S., & Svoboda, J. 2023, A&A, 673, A16
- [86]Laureijs, R., Amiaux, J., Arduini, S., et al. 2011, arXiv e-prints, arXiv:1110.3193arxiv.org/abs/1110.3193
- [87]Law, C. J., Sharma, K., Ravi, V., et al. 2024, ApJ, 967, 29
- [88]Leauthaud, A., Tinker, J., Behroozi, P. S., Busha, M. T., & Wechsler, R. H. 2011, ApJ, 738, 45
- [89]Legrand, L., McCracken, H. J., Davidzon, I., et al. 2019, MNRAS, 486, 5468
- [90]Leja, J., Speagle, J. S., Johnson, B. D., et al. 2020, ApJ, 893, 111
- [93]Lorimer, D. R., Bailes, M., McLaughlin, M. A., Narkevic, D. J., & Crawford, F. 2007, Science, 318, 777
- [94]Lower, S., Narayanan, D., Leja, J., et al. 2020, ApJ, 904, 33
- [95]Macquart, J.-P., Prochaska, J. X., McQuinn, M., et al. 2020, Nature, 581, 391
- [96]Madhavacheril, M. S., Battaglia, N., Smith, K. M., & Sievers, J. L. 2019, Phys. Rev. D, 100, 103532
- [97]Marinacci, F., Vogelsberger, M., Pakmor, R., et al. 2018, MNRAS, 480, 5113
- [98]Martizzi, D., Vogelsberger, M., Artale, M. C., et al. 2019, MNRAS, 486, 3766
- [99]McConnell, D., Allison, J. R., Bannister, K., et al. 2016, PASA, 33, e042
- [100]Medlock, K., Nagai, D., Singh, P., et al. 2024, ApJ, 967, 32
- [101]Mitchell, P. D., Lacey, C. G., Baugh, C. M., & Cole, S. 2013, MNRAS, 435, 87
- [102]Mobasher, B., Dahlen, T., Ferguson, H. C., et al. 2015, ApJ, 808, 101
- [103]Naiman, J. P., Pillepich, A., Springel, V., et al. 2018, MNRAS, 477, 1206
- [104]Nelson, D., Pillepich, A., Springel, V., et al. 2018, MNRAS, 475, 624
- [105]Nicastro, F., Kaastra, J., Krongold, Y., et al. 2018, Nature, 558, 406
- [106]Niu, C.-H., Aggarwal, K., Li, D., et al. 2022, Nature, 606, 873
- [107]Ocker, S. K. & Cordes, J. M. 2026, ApJ, 1002, 3
- [108]Pastor-Marazuela, I., van Leeuwen, J., Bilous, A., et al. 2025, A&A, 693, A279
- [109]Penton, S. V., Stocke, J. T., & Shull, J. M. 2004, ApJS, 152, 29
- [110]Petroff, E., Hessels, J. W. T., & Lorimer, D. R. 2022, A&ARv, 30, 2
- [111]Pillepich, A., Springel, V., Nelson, D., et al. 2018, MNRAS, 473, 4077
- [112]Planck Collaboration, Ade, P. A. R., Aghanim, N., et al. 2016, A&A, 594, A13
- [113]Planck Collaboration, Aghanim, N., Akrami, Y., et al. 2020, A&A, 641, A6
- [114]Pol, N., Lam, M. T., McLaughlin, M. A., Lazio, T. J. W., & Cordes, J. M. 2019, ApJ, 886, 135
- [115]Pollack, J. E., Smith, R. E., & Porciani, C. 2014, MNRAS, 440, 555
- [116]Prat, J., Sánchez, C., Miquel, R., et al. 2018, MNRAS, 473, 1667
- [117]Priyadarshini, V., Vijikumar, S., & Bhuvaneshwari, V. 2025, Theor. Math. Phys., 223, 879
- [118]Reischke, R. & Hagstotz, S. 2023a, MNRAS, 523, 6264
- [119]Reischke, R. & Hagstotz, S. 2023b, MNRAS, 524, 2237
- [120]Reischke, R., Kovač, M., Nicola, A., Hagstotz, S., & Schneider, A. 2025, OJAp, 8
- [121]Rodriguez-Gomez, V., Snyder, G. F., Lotz, J. M., et al. 2019, MNRAS, 483, 4140
- [122]Romano-Díaz, E. & van de Weygaert, R. 2007, MNRAS, 382, 2
- [123]Rowe, B. T. P., Jarvis, M., Mandelbaum, R., et al. 2015, Astron. Comput., 10, 121
- [125]Rybicki, G. B. & Lightman, A. P. 1979, Radiative processes in astrophysics (New York: Wiley)
- [126]Schaap, W. E. & van de Weygaert, R. 2000, A&A, 363, L29
- [127]Schmittfull, M. & Seljak, U. 2018, Phys. Rev. D, 97, 123540
- [128]Schneider, A., Stoira, N., Refregier, A., et al. 2020, JCAP, 2020, 019
- [129]Sharma, K., Krause, E., Ravi, V., et al. 2025, ApJ, 989, 81
- [130]Shull, J. M., Smith, B. D., & Danforth, C. W. 2012, ApJ, 759, 23
- [131]Spergel, D., Gehrels, N., Breckinridge, J., et al. 2013, arXiv e-prints, arXiv:1305.5425arxiv.org/abs/1305.5425
- [132]Springel, V. & Hernquist, L. 2003, MNRAS, 339, 289
- [133]Springel, V., Pakmor, R., Pillepich, A., et al. 2018, MNRAS, 475, 676
- [134]Stevens, J. A., Amure, M., & Gear, W. K. 2005, MNRAS, 357, 361
- [135]Stuart, A. M. 2010, Acta Numer., 19, 451
- [136]Takahashi, R., Ioka, K., Mori, A., & Funahashi, K. 2021, MNRAS, 502, 2615
- [137]Takahashi, R., Ioka, K., Shirasaki, M., & Osato, K. 2026, PASJ, 78, 1829
- [139]Taylor, E. N., Hopkins, A. M., Baldry, I. K., et al. 2011, MNRAS, 418, 1587
- [140]Thornton, D., Stappers, B., Bailes, M., et al. 2013, Science, 341, 53
- [141]van de Weygaert, R. 1994, A&A, 283, 361
- [142]Vanderlinde, K., Liu, A., Geach, J. L., et al. 2019, in Canadian Long Range Plan for Astronomy and Astrophysics White Papers, Vol. 2020, 28
- [143]Verde, L., Heavens, A. F., Percival, W. J., et al. 2002, MNRAS, 335, 432
- [145]Villanueva-Domingo, P., Villaescusa-Navarro, F., Anglés-Alcázar, D., et al. 2022, ApJ, 935, 30
- [146]Walker, C. R. H., Spitler, L. G., Ma, Y.-Z., et al. 2024, A&A, 683, A71
- [147]Walters, A., Ma, Y.-Z., Sievers, J., & Weltman, A. 2019, Phys. Rev. D, 100, 103519
- [148]Weinberger, R., Springel, V., Hernquist, L., et al. 2017, MNRAS, 465, 3291
- [149]Willmer, C. N. A. 2018, ApJS, 236, 47
- [151]Yan, Z., van Waerbeke, L., Tröster, T., et al. 2021, A&A, 651, A76
- [152]Yao, J. M., Manchester, R. N., & Wang, N. 2017, ApJ, 835, 29
- [153]Zacharegkas, G., Chang, C., Prat, J., et al. 2026, OJAp, 9
- [154]Zahid, H. J., Geller, M. J., Fabricant, D. G., & Hwang, H. S. 2016, ApJ, 832, 203
- [157]Zhang, Z. J., Yan, K., Li, C. M., Zhang, G. Q., & Wang, F. Y. 2021, ApJ, 906, 49
- [158]Zhu, W. & Feng, L.-L. 2021, ApJ, 906, 95







