In this paper
AbstractIntroductionMagnification mapsMacroscopic ModelSubhaloesMagnification mapsLensed star detection and spatial distributionDiscussionConstraints on SIDM cross-section strengthPreference in Comparison with wave DM predictionObservations required to distinguish between modelsEffect of choice of model parametersLens model dependencySubstructure mass fractionCore sizeTruncation radiusSlope of luminosity functionAbundance of microlensesGlobular clusters as extra perturbersConclusionAnalytical understanding of formation of milliCCsSpatial distribution for different sReferencesWorksPhysics and Astronomy
Self-interacting Dark Matter signatures from stars extremely magnified by galaxy clusters
Abstract
The spatial distribution of cluster-lensed star transients with respect to the macroscopic critical curve (CC) is sensitive to underlying substructures predicted by different classes of dark matter (DM) models. \citet{Broadhurst_2025} have shown that wave-like DM composed of ultra-light bosons better explains both the width and negative skewness (towards the interior of the CC) of the lensed star spatial distribution discovered in the ``Dragon'' lensed galaxy () in the Abell 370 (A370, ) cluster field, compared with the case of regular cold dark matter (CDM) subhaloes. In this work, we consider the case of self-interacting DM (SIDM) and investigate how the lensed star spatial distribution depends on the velocity-dependent self-interaction cross-section of SIDM particles, and whether constraints can be placed on the cross-section strength at relevant velocities. Establishing a cross-section strength-dependent (sub)halo gravothermal core-collapse probability, we found that the SIDM appears to be the only known class of models thus far in which the sign of skewness is set by its own particle parameters. The observed lensed star distribution in the Dragon galaxy can be well explained, in terms of both skewness and width, when a significant fraction of subhaloes have undergone core collapse and exhibit a steep power-law density slope. Given the cluster redshift, this qualitatively favours an (effective) cross-section strength of for halos with mass (velocity dispersion of ), when the role of environmental effects on the core-collapse (de)acceleration is not considered, and limits the three-dimensional inner density slope of the collapsed halo to (Bayes factor of ).
Introduction
The regular detection of lensed stars through deep, repeated Hubble Space Telescope (HST) and James-Webb Space Telescope (JWST) observations of galaxy clusters (e.g., [48, 47, 98, 32, 75]) has stimulated studies of Dark Matter (DM) substructures based on lensed star spatial distributions ([96, 8]) with respect to the cluster critical curve (CC), where the lensing magnification is at its maximum. Broadhurst et al. [8] showed that both the broadness and the negative-parity skewness (towards the cluster core, interior of CC) of the spatial distribution of lensed stars in the “Dragon” lensed galaxy (, Fudamoto et al. [32]) in cluster Abell 370 (A370, ) aligns with the prediction made adopting the wave-like interference assuming ultra-light axions, whereas standard cold and collisionless DM subhaloes predict a narrower spatial distribution, and more importantly, an opposite, positive-parity (away from the cluster core, exterior of CC) skewness of lensed star distribution (see also Williams et al. [96]). Recent observations and calculations pinned down the CC position to a higher confidence level with different methodologies (e.g., Zhou et al. [102]; Diego et al. (2025)) and revealed more lensed stars (Palencia et al. [75]) that reinforce both the broadness and the negative lensed star skewness stressed in Broadhurst et al. [8].
In this paper, we explore for the first time the effect of another class of DM model known as self-interacting DM (SIDM) in the context of lensed stars, especially to see if SIDM can qualitatively explain the observed lensed star spatial distribution – and if so, what can we say about the nature of SIDM particles. SIDM was first proposed to solve the small-scale problems of the cold and collisionless DM paradigm [86]. It provides solutions to the cusp-core problem, the too-big-to-fail problem, and the diversity problem (see, e.g., Tulin & Yu 2018, for a review). Self-interactions also arise naturally when DM belongs to a dark sector, i.e., a collection of hidden particles that interact feebly with the Standard Model. New hidden particles, such as dark photons, can mediate non-gravitational interactions among DM particles. A heavy mediator gives rise to a constant and isotropic self-interaction cross-section, while a light mediator can give rise to a velocity-dependent cross-section1, as in Yukawa SIDM, resonant SIDM, and many others [30, 9, 7, 44, 35]. Therefore, astrophysical probes of SIDM, which rely only on the gravitational effects of DM, offer a new way to study the nature of DM and dark sectors. For example, recent lensing observations through image reconstruction point towards overly concentrated subhaloes in both galaxy- and cluster-scale lenses, favouring the SIDM scenario [e.g. 28, 26, 88, 94, 66] after which subhaloes have undergone core collapse.
This paper is structured as follows. In § 2, we describe the methodology that computes the magnification distribution when a fraction of the lensing cluster is constituted of SIDM subhaloes under different conditions. In § 3, we outline the way in which the lensed star spatial distribution is computed based on the result in § 2. We then discuss the possible application of our result in constraining the interaction cross-section of SIDM particles and properties of SIDM subhaloes, followed by a brief estimation of how many events in total are required to distinguish between DM models with lensed star spatial distribution in § 4. We then discuss briefly how different choices of hyperparameters and model uncertainties affect our results in § 5. Throughout this work, we adopt the AB magnitude system [71], along with Planck18 cosmological parameters: , , and [78].
Magnification maps
To estimate how the lensed star spatial distribution is affected by substructures, we have to first compute the magnification maps by considering a macroscopic DM halo (representing, for example, a galaxy cluster) being perturbed by DM substructures, and then compute the relative lensed star detection rate for different magnifications.
Macroscopic Model
For simplicity, we consider a galaxy cluster with following a Navarro-Frenk-White (NFW, Navarro et al. 1997) profile:
as the macroscopic model. Such mass is common for massive clusters, including A370 [e.g. 91], which we shall compare against our simulations later on. For a given halo, its mass, (the enclosed mass within such that its enclosed density is 200 times the critical density), and its concentration, , follows the concentration-mass-redshift relation [e.g. 25] with a scatter of dex in log space, such that the NFW scale radius and density can be written as:
given the critical density at redshift , , and
For our cluster halo, , which is typical for clusters as found from lensing observations [e.g. 58, 61]. Throughout this work, we adopt a lens redshift of 0.375 and source redshift of 0.7251.
Subhaloes
The gravothermal evolution of the SIDM (sub)haloes is based on an initial NFW profile (Eqn. (1)) where their time evolution is normalized by the collapse time scale:
with the effective cross-section strength of SIDM particles [29, 99] and the long-mean-free-path conduction factor is calibrated according to Koda & Shapiro [51] (i.e., 0.75). Note that Eq. (4) is only valid when the halo stays at the long-mean-free-path regime at its formation [79, 29, 92], or more concretely, . For to exceed the limit, the assumption of an initial NFW profile is not valid. We estimate the (sub)halo evolution time , i.e., the age of halo, as , where and are cosmic times at the halo observation redshift and its formation redshift, , respectively. If , we consider that the halo has completely core-collapsed. Otherwise, the halo is in the core expansion stage or begins to enter the core collapse phase.
Using CDM cosmology and Eqs. (2) and (3), we can rewrite Eq. (4) of a halo as a function of , , and at its formation redshift. The concentration can be further fixed by using a concentration-mass-redshift relation. For a halo with mass at , its core-collapsing probability, with formation concentration and , is given by
where is the Heaviside step function.
To apply the probability to a population of similar halos, we replace by their average formation redshift and update Eq. (5) to
where and the scatter are functions of halo mass and lens redshift, determined by fitting Eqn. (6) with a set of isolated halo samples generated by SatGen Jiang et al. [45] merging tree, where we trace back the half-mass redshift and identify it as the formation redshift. The is also take value at . One can see from Figure 1 that, for the subhalo masses of interest between , the core collapse probability is either 0 or 1 except in the range of for a formation redshift of (solid curves), which is expected from high-resolution cluster simulation (e.g., Xie & Gao [97]). In other words, under such a setup, this method is expected to have sensitivity only to the SIDM interaction cross-section within this window, as any cross-sections below or higher than this range would be identical to the boundary cases.
Figure 1. Core collapse probability as a function of SIDM interacting cross-section, for different subhalo masses thus velocities (colour), and formation redshifts (linestyle).
Note that environmental effects such as tidal stripping (Nishikawa et al. [69]) and the central baryonic component could have led to accelerated core-collapse (Zhong et al. [101]), and the self-interactions could have velocity dependence such that small haloes, with larger cross-sections, will go through core-collapse quicker, still leading to over-concentrated subhaloes, as motivated by the claims of overly-dense subhaloes found through galaxy-scale strong lensing analysis mentioned earlier in § 1 (Enzi et al. [28]; Vegetti et al. [94]). We shall discuss the possible role of tidal stripping in further detail later in § 4.1.
To model the density profiles of core-collapsed subhaloes, we adopt a cored power-law profile from Gilman et al. [35]; ); He et al. (2026):
which has a concentrated, cuspy profile outside the core radius with a power-law slope characterized by a constant, (Turner et al. [90]; Fischer et al. [31]). Similarly steep three-dimensional inner slopes, , have also been inferred from cluster strong-lensing analyses (Natarajan et al. [66]), and we shall discuss this later in Sec. 4.2.
The core radius, , changes as the halo evolves (Zhong et al. [101]). It first increases to during the core expansion stage and then decreases during the core-collapse stage in a self-similar manner, during which the shrinkage accelerates near the onset of gravothermal catastrophe. The end state of the core-collapse leads to a secondary core with radius , which departs from the single-power-law profile () of its outskirts. To be concrete, we fix for subhaloes that have not yet core-collapsed (e.g., Outmezguine et al. [72]; Yang et al. [99]), and for those that have core-collapsed. This is because the core-collapse probability is nearly a sigmoid function for most interaction cross-sections, and only at very specific cross-sections would one have a thorough mix of subhaloes at different degrees of core-collapse. Moreover, as we shall show in § 5.3, the exact choice of in either case is not the main driver of our results, as the primary driver of lensed star spatial distribution is , as well as .
of Eq. (7) is computed by matching the total integrated mass of Eqn. (7) to the total integrated mass Eqn. (1) to within (Jiang et al. [46]; Zhong et al. [101]). Since the projected mass of the NFW and power-law profiles diverges at large , we have to truncate at some characteristic radius. For the NFW profile, the truncation would naturally be at . Effects such as tidal stripping could have truncated the subhaloes to a radius smaller than , where we shall discuss this further in § 4. Nevertheless, to conserve the number density of core-collapsed SIDM subhaloes with respect to the number density of NFW subhaloes (as SIDM core-collapsed subhaloes "evolve" from the initial NFW subhaloes), we truncate the power-law NFW profiles at some radius, , such that the total projected mass within is equal to that of the NFW profile within .
A special scenario is . A prior matching between the initial NFW and SIDM subhalo mass within will lead to a failure in finding a matching . This is because integrating the SIDM halo's cored power-law profile with at will only yield at most of the mass in the initial NFW subhalo. To preserve the mass and number density of subhaloes, instead of matching to an initial NFW profile, we match the cored power-law profile with to a King [50] profile:
truncated at the calculated for an NFW subhalo with the same initial mass and concentration. We show a suite of the adopted profiles in Figure 2 alongside an NFW profile for reference.
Figure 2. An example of profiles adopted in this work, for (orange), (red) and (purple). Core-collapsed subhalos with are shown in dotted line style, and cored subhalos with are in dashed line. The solid black curve refers to an NFW subhalo for reference. All these haloes have , and a therefore concentration of provided the Diemer & Kravtsov [25] mass-concentration relation.
Given that self-interactions sufficiently randomize the velocity distributions of the scattered DM particles, it is reasonable to assume the subhaloes are perfectly spherically symmetric. We project the three-dimensional mass profile along the line-of-sight into surface mass density, . Within the thin lens approximation, the reduced deflection field can then be expressed as:
If subhaloes are to take up a fraction of the DM halo, we can replace some fraction of the mass of the macroscopic NFW halo with subhaloes. We consider a case where these subhaloes take up 1% of the DM mass (same as previous lensed star spatial distribution studies, e.g., Williams et al. [96]; Broadhurst et al. [8]), following a typical subhalo mass function (Dai et al. [17]; Dai et al. [16]):
Instead of simply summing up their mass, we consider their total mass integrated within the field-of-view (FOV) of interest (i.e., near the cluster CC, as it is computationally impossible to do a fine-scale simulation for the whole cluster) to ensure the conservation of projected surface mass density. We adopt an image plane resolution of pc ( arcsec) to maximize the resolution of microscopic CCs formed from subhaloes while retaining a reasonable computational cost. Sampling all subhaloes within the FOV, we can make use of the linearity of the deflection field to co-add the deflection field contributed by the cluster NFW profile, as well as those from subhaloes, to compute the lensing magnification:
with being the image plane position.
Magnification maps
We show a suite of magnification maps in Figure 3, where the top row features the scenario where one adopts regular NFW subhaloes following the standard CDM convention. Other than that, each row shows simulations adopting SIDM core-collapsed subhaloes for different , whereas each column is a case with different characteristic interaction cross-sections.4 For consistency’s sake

Figure 3. Macro magnification maps of CDM subhaloes (top row), SIDM core-collapsed subhaloes with (second row), (third row), and (bottom row) when (left column), (middle column), and (right column) respectively. Subhaloes take up 1% of the DM mass in all these maps. The white, almost vertical dashed line shows the smooth CC when there is only the cluster lensing effect without substructures. The region towards the right of the smooth CC is known as the negative parity, whereas the region towards the left is the positive parity.
sake and to best compare the effect of the choice of and , we use the same seed to generate these magnification maps such that they have the same subhalo mass function sampled, the same concentration for each subhalo despite the scatter in the mass-concentration relation, as well as the same position within the FOV. Notice that the magnification map of the case of CDM subhaloes follows closely the distribution in Williams et al. [96]; Broadhurst et al. [8] where the originally smooth cluster CC (white dashed line) is extended towards the positive parity (left) in most regions, with extra milliCCs forming exclusively at the positive parity side due to subhaloes.
For shallow SIDM profiles with , and at a smaller cross-section strength of (left column), none of the subhaloes have undergone core collapse (as indicated in Figure 1), and have retained a large core of . With increasing , these subhaloes create stronger perturbations towards the macroscopic CC, bringing the CC further out towards the positive parity, like the case of NFW subhaloes. At an intermediate cross-section strength of (middle column), some of the subhaloes have already gone through core collapse, as shown in Figure 1. While the case of remains vastly similar to the case of NFW subhaloes, smaller CCs begin to form in the negative parity when . Since CCs form when the projected surface mass density transitions between sub- and supercritical, this explains why milliCCs are only found where , as these steeper profiles can generate supercriticality in the negative parity (see Sec A for an analytical understanding). The steeper the density profile, the more dramatic the CCs are. Lastly, at a larger cross-section strength of (right column), where all subhaloes have their core collapsed, as shown in Figure 1, the CCs are most dramatically perturbed. For the steepest case, one can see that not only are more milliCCs formed in the negative parity, but the macroscopic CC is also “deformed”, similar to what is expected for wave-like DM ([1, 8]) when they perturb macroscopic CC through the under- and over-density created by wave-like interference.
In Figure 4, we also show the macroscopic magnification probability density function in each of the maps in Figure 3, which helps interpret the result. In the positive parity, for a low interaction cross-section, where no subhaloes have undergone core collapse, all three values give rise to very similar PDFs. When we increase the interaction cross-section, the collapsed core makes the profile much steeper depending on , such that the steeper the power-law slope, the more demagnification. For the higher magnification tail, although all the cases demonstrate the slope, the case, when subhaloes have collapsed, has a lower probability of producing higher magnification than the cases of and . In the negative parity, a similar trend is observed, except that now cases have a higher probability of gaining higher magnification.
Figure 4. Probability density function of macroscopic lensing magnification near the CC (from the magnification maps in Figure 3), for different SIDM interaction cross-section strengths (linestyles) and different power-law index, (colour). We show the case for the cluster positive parity and the negative parity in the left and right panels, respectively.
Significant demagnification with subhaloes (or point mass lenses) only occurs in the negative parity5 (e.g., Diego [18]; Palencia et al. [74]) – the reason why demagnification is also observed in the positive parity, as shown in Figure 4, is that the “main” CC has been dragged towards the positive parity by massive subhaloes (see Figure 3). In this case, part of the original positive parity in the macroscopic model is now actually negative parity, where demagnification occurs.
Lensed star detection and spatial distribution
To estimate the lensed star detection rate under the effect of microlensing on top of strong lensing, we have to first assume a stellar luminosity function, , of the background galaxy and estimate the lensed star detection probability given some magnification. Instead of computing from star formation history and stellar isochrones, we assume a power-law stellar luminosity function where the relative abundance of stars with luminosity is:
which in apparent magnitude, , space, is:
truncated at an absolute magnitude of , roughly corresponding to the brightest stars known in the local universe and the prediction made by isochrones (e.g., Choi et al. [14]). Recent lensed star observations deduced (Diego et al. [21]; Broadhurst et al. [8]; Palencia et al. [75]), and we adopt as a fiducial value. We shall show later in § 5.5 that or ( lower and upper limits in Palencia et al. [75]) only changes the statistical significance of our result but not the qualitative conclusion. Based on the underlying macroscopic magnification6, (perturbed by substructures or not) and the abundance of microlenses (in terms of surface mass density, ), background stars have some probability of attaining some total magnification, , under the influence of microlensing as approximated by Palencia et al. [74]. Here we adopt a fiducial , which is typical for cluster-lensed galaxies with lensed star detections, including the Dragon Arc, against which we will compare our simulations (e.g., Diego et al. [22]; Li et al. (2024); Meena et al. [60]; Li et al. (2025a)). Such a value is deduced from analysis of intra-cluster light (e.g., Morishita et al. [64]), and assuming there are little to no dark compact objects such as primordial black holes at the relevant stellar mass range of , which is constrained by different kinds of observations like galactic microlensing and gravitational waves (e.g., Green & Kavanagh [38]). We shall inspect what would happen to our result if the inferred surface mass density of microlenses differs significantly in § 5.
If we assume that the surface brightness of the lensed galaxy is more or less constant (which is true for most lensed galaxies, as the brightness variation is of order ), then we can follow the methodology in Palencia et al. [73]; ); Li et al. (2025a) and estimate the detection rate as:
with describing the action of lensing magnifying the apparent magnitude of background stars. We adopt a fiducial yet typical JWST observation depth of 29 mag throughout our calculation, whereas is the minimum magnification required for a background star to be detected at this detection limit. By applying to the macroscopic magnification map [8] generated following the methodology stated in Sec. 2, we can simulate the expected spatial distribution of lensed stars with respect to the macroscopic cluster CC (i.e., when there is only the main NFW halo).
We simulate the expected spatial distribution of lensed stars under microlensing for all the magnification maps shown in Figure 3 as well as for the case of a smooth CC with no perturbation (which roughly resembles the prediction made with warm dark matter, Lovell et al. [56]) in Figure 5. Owing to the fact that microlensing can demagnify images in the negative parity (Diego [18]; Palencia et al. [74]), which is compensated by a higher probability of getting higher magnification, detection of a lensed star is more likely in the negative macroscopic magnification (Broadhurst et al. [8]; Li et al. 2025a; Palencia et al. [73]). Based on this effect, a smooth model where the magnification follows with the separation from the CC (Schneider et al. [83]; Diego [18]) should show a slight negative skewness of the lensed star under the effect of microlensing, as shown in Broadhurst et al. [8] and the black histogram in Figure 5. We found a positive skewness of the lensed star spatial distribution in the case of CDM subhaloes (green histogram), aligning with the earlier findings in Williams et al. [96]; Broadhurst et al. [8].
Figure 5. Expected strong+micro lensed star spatial distribution of smooth (black), CDM subhalo (green), SIDM core-collapsed subhalo with (top row, orange), (middle row, red), and (bottom row, purple) respectively, with respect to the cluster CC, given (left column, dotted line), (middle column, dashed line), and (right column, dash-dot line). Positive projected distance refers to the negative parity, whereas negative projected distance refers to the positive parity. Each bin has a width of 0.18 arcsec, symmetric around the smooth CC (). We also show the observed spatial distribution of lensed stars extracted from Kelly et al. [47]; Fudamoto et al. [32]; Palencia et al. [75] anchored to the Diego et al. (2025) lens model at each corresponding bin as black data points, with the error bars showing the Poisson uncertainties at the corresponding bins. The reduced , KS-test statistics, , and p-value for each simulation run are shown at the top of each panel.
In the case of core-collapsed SIDM subhaloes, the skewness of the lensed star spatial distribution depends on both and : for , all models demonstrate a positive skewness, as the CDM model does, as few to no subhaloes are core-collapsed such that they do not have sufficiently steep profiles to induce super-criticality at the negative parity. For , their profiles are not steep enough such that even when all the haloes are core-collapsed, the lensed star spatial distribution is still towards the positive parity. In fact, they resemble a similar spatial distribution to that of CDM subhaloes (green) for all cases of . In the case of , low also resembles the case of CDM subhaloes and . However, when grows to , all subhaloes have undergone core collapse, and a slight negative skewness is observed. At , if no subhaloes have undergone core collapse (), the large core still prohibits milliCCs from forming in the negative parity despite the steep profile, thus resulting in a positive skewness. When some subhaloes have core-collapsed (), milliCCs begin to form at the negative parity and are stronger than in the case of , see Sec. A) and begin to create the negative skewness.
Discussion
Constraints on SIDM cross-section strength
Here, we compare our simulation results with observations to determine whether we can place constraints on the SIDM cross-section to the first order. Three lensed star catalogs published so far have detections in the “Dragon” galaxy (Kelly et al. [47]; Fudamoto et al. [32]; Palencia et al. [75]), and we adopt the Diego et al. (2025) lens model as the baseline to compute the separation of lensed stars from the cluster CC. Since the position of the CC is highly uncertain at the location of the cluster members near the head of the Dragon galaxy, we only adopt lensed stars detected away from that region, the same as was done in Broadhurst et al. [8], yielding a total of 67 lensed stars in our sample. For all our simulations, we compare the prediction against the observed lensed star spatial distribution and compute the reduced (). Since our prediction is not in absolute value, we normalize all our predictions to have a total of 67 events.
Figure 5 shows that simulations with a larger interacting cross-section and larger are preferred in general. The best-performing models are the ones with , and , as they give a reduced of 1.15 and 1.37, respectively. The third best-performing model would be the one with and , which has a reduced of 1.91. These models all predict a skewness towards the negative parity, which is visualized in Figure 3, where an abundant amount of milli-CCs is generated in the cluster negative parity when the subhaloes have undergone core-collapse with a steep power-law density profile. As a reference, is and for the case of the smooth model (black histogram) and regular NFW subhaloes in the case of standard CDM (green histogram). We can evaluate the significance of statistical preference between the models with the Asimov estimate (Cowan et al. [15]):
Here, is the predicted count of models and to be compared upon at the -th data point (spatial bin) with observed count in the corresponding data bin. The best-fitting SIDM model has compared with the smooth model, meaning that it is better than the smooth model. Compared with the CDM model, , so the SIDM model is preferred over the CDM model by . The qualitative difference between these models will eventually allow the distinction when sufficient observations have been accumulated. We shall address this aspect in Sec. 4.4.
As a binning-independent check, we adopt the one-sample Kolmogorov-Smirnov (KS) test on the 67 event positions. The KS statistic, , is given as the maximum difference between the observed and model cumulative distribution of the 67 independent observed positions. Since no parameters are fitted, we also compute the -values from simulated samples for each model, with 67 events sampled for each realization. The KS test confirms the above test – the model rankings are the same, rejecting both the smooth and the CDM model at a similar confidence level ( for both models, with and , respectively). Both the models with and are fully consistent with the data ( for both models, with and , respectively), where is typical for a correct model. A paired KS test with the best-fitting SIDM model and the smooth (CDM) model gives () preference, lower than the confidence level inferred from the Asimov estimate. This is attributed to the KS test only using the maximum cumulative deviation, whereas the Asimov estimate combines the difference between the models in every bin. Nevertheless, this demonstrates that the quantitative preference among models is broadly consistent between the two statistical tests.
Since best reproduces the observations among all our models, we rerun our simulation at a finer grid of for and show the reduced as a function of within the range of in Figure 6. Instead of fixing to one seed in generating ten solutions with different , we run the suite of simulations ten times with ten different seeds to sample the uncertainty in the randomly sampled subhalo mass function and mass-concentration relation that affect the result. One can see that there are two main sets of reduced – one group with () and another with (). This trend is apparent when one refers to Figure 1, as – is the interaction cross-section where the core collapse probability of subhalo quickly transitions from zero to unity. Since these haloes are most abundant, this transition in core collapse probability propagates into the magnification distribution; thus, into the spatial distribution of lensed stars and the reduced . Estimated from the scatter in Figure 6, the first group has a seed noise of , and the second group has a seed noise of . Propagating the seed noise would then yield uncertainty in relevant statistics.
Figure 6. Reduced of simulations adopting with different . Black data points refer to the mean value and the errorbar standard deviation among ten different realizations for each cases. The reduced of the individual realizations are shown in red data points.
While the current scattering is large due to the Poissonian statistics, at face value, this shows that the lensed star spatial distribution provides a pathway to constrain the interacting cross-section of SIDM particles for subhaloes in cluster environments. However, the constraining power from the current calculation is incomplete, as the effect of on is not explored, and we have simply taken as a hyperparameter that describes the SIDM density profile. Furthermore, we have neglected the fact that tidal forces could accelerate (or decelerate) the core collapse (Nishikawa et al. [70]; Zeng et al. [100]; Mace et al. [57]) and make the truncation radii smaller, meaning that we could be underestimating (overestimating) the core-collapse probability as calculated in Eqn. (6), thus overestimating (underestimating) the cross-section strength that is required to match the observed lensed star spatial distribution.
In principle, the environmental effects can be considered as a nuisance parameter that affects our result through Eqn. (6). To date, there are no tailored simulations that estimate the effect of tidal stripping upon subhalo core-collapse on cluster scales. One can expect that tidal disruption is more severe for slow encounters, so the high relative velocities in clusters make individual encounters between subhaloes inefficient in tidal stripping (e.g., Gnedin et al. [36]). Instead, it would be the cumulative effect of many such fast encounters, known as harassment (Moore et al. [62]), on top of the repeated passage through the cluster potential that dominates the stripping. Another environmental effect comes from ram pressure. The self-interacting nature of SIDM makes ram-pressure stripping possible, especially when the interaction cross-section is larger (Shirasaki et al. [84]). In such a case, the mass loss from ram-pressure stripping will make the (sub)haloes more susceptible to tidal stripping [65], as well as making the truncation radius much smaller (see discussion in Chiang et al. 2026) and yielding even more compact subhaloes. We shall discuss briefly the potential effect later in Sec. 5.4
Because of the complication in accessing the environmental effect in cluster subhaloes, even though Mace et al. [57] infer that a stronger tidal effect does not appear to speed up the core-collapse process significantly in smaller-scale host haloes at a larger interaction cross-section (what is required to explain the observed lensed star spatial distribution according to our calculation), we hereby do not simply extrapolate their result to infer the constraining power of the lensed star spatial distribution upon the interacting cross-section of SIDM particles.
Although lensing observations have constrained the cluster-scale interaction cross-section down to level [e.g. 81, 37, 40, 82, 49, 10, 19], there are few constraints on the interaction cross-section of subhaloes embedded in cluster-scale halos. The lower velocity of these subhaloes, with masses of , could allow for a larger interaction cross-section for velocity-dependent SIDM. One of the most relevant constraints comes from Gilman et al. [34], where the flux ratios from 29 quadruply-lensed quasars [33] constrain the core-collapse timescale of low-mass subhaloes in galaxy lenses that reflect the interaction cross-section of SIDM particles. Their result disfavours for velocities of (, the upper mass end of our subhalo mass function), where two sets of more conservative and aggressive assumptions upon structure of collapsed-haloes and tidal effects leads to peak of likelihood at and , respectively. This aligns with our inference at face value, especially since they have adopted and , also roughly the that best reproduces the observed lensed star distribution. Since we in
served lensed star distribution. Since we include subhaloes with even lower mass (thus velocity) than their benchmark velocity of , meaning that our preferred interaction cross-section could have been larger than what they have inferred. Another relevant constraint comes from Sánchez Almeida [87], where they used the stellar cores of ultra-faint dwarfs with masses of to infer that lies between , assuming that the cores are the end result of DM self-interactions. This is at the more massive end of our subhalo mass function, where one would expect the less massive end of the mass function would have also significantly contributed to our inference. Neither inference is for the cluster environment, as we have done here; the different environmental effects mean that such a comparison only serves as a first-order sanity check of our result, and our methodology could be a complementary method to provide insights into the SIDM interaction cross-section at the relevant mass range in different environments.
Preference in
Following the calculation in the previous section, we can also deduce that a steeper profile with better explains the observations, regardless of the interaction cross-section. We can marginalize over and compute the preference of through the Poissonian deviance for all given sets of and with Eqn. 15, with the likelihood defined as:
if we set one of the models to be the observed value for Eqn. 15. Marginalizing over choices of with a flat prior:
would then yield posterior weights of , and for , and , respectively. is moderately preferred over with a Bayes factor of , whereas the hypothesis of is definitively preferred over with a Bayes factor of .
Recent findings claim that the observational cross-section of galaxy-galaxy strong lensing is significantly larger than the prediction of the standard CDM model, an issue that can be alleviated when (cluster) subhaloes adopt a SIDM-consistent core with [26, 66]. Although aligning with our inference, such subhaloes are in the range of , significantly larger than the subhaloes considered in our work. For our mass range, the most relevant constraints come from galaxy strong lens modelling, which prefer for a subhalo [80]. That is, however, the case for subhaloes inside a field galaxy instead of a cluster environment. Given the different ambient environment, we cannot directly compare our inference with such a constraint. The different preference in at the two regimes could reflect different degrees of, for example, tidal disruption in SIDM subhaloes, and detailed simulation is required to better understand the root cause behind the different preferred s.
In either weighting scheme, they prefer a density profile of
Comparison with wave DM prediction
Broadhurst et al. [8] showed that wave DM with a characteristic wavelength of , which corresponds to a typical cluster halo mass of and an axion mass of , best fits the observed lensed star spatial distribution in the Dragon arc with the Kelly et al. [47] and Fudamoto et al. [32] data. Since we have additional events from Palencia et al. [75], we also predict and compare the expected lensed star spatial distribution with the wave DM simulations in Broadhurst et al. [8] as both a sanity check of our method and a comparison to the SIDM predictions.
As in Broadhurst et al. [8], we run the simulation for wave DM with the characteristic de Broglie wavelengths of pc, pc, and pc and compare with the observed lensed star spatial distribution. We found the reduced (KS-test -value) of the three models to be (), (), and (), respectively. The quality of fit for the first two models is comparable with SIDM simulations adopting . Our best-fitting model, and cm/g, performs similarly to the and pc wave DM simulations. This specific SIDM case corresponds to a de Broglie wavelength of pc if we make use of the relation in Broadhurst et al. [8] to fit for the lensed stars distribution. The current difference in between the best-fitting SIDM ( and cm/g) and waveDM ( pc) models is only with ten bins, which is dominated by Poissonian statistical fluctuations, as the statistical uncertainty is larger than the difference itself. Future observations increasing the statistical confidence of the lensed star spatial distribution is required to distinguish between the two classes of models in the context of lensed stars.
Observations required to distinguish between models
At this stage, an interesting question would be to ask how many lensed stars (and epochs of observations) are required to distinguish between different DM models. We begin by comparing the best-fitting SIDM (, cm/g) and wave DM (de Broglie wavelength of pc) models and see how many observations are required to distinguish between the two fixed models to confidence again through the Asimov estimate. The current observation yields , meaning that the two models are not significantly discriminated, with only . We can take this one step further and assume one model is true, such that its prediction would then serve as the typical data – comparing the other model against it will then inform us of the goodness-of-fit. The current two models give . grows linearly with the increasing number of events, , such that the significance scales as . To achieve , we would need times the current , meaning that a total of events are required to distinguish between the two classes of DM models based on the lensed star spatial distribution. Fudamoto et al. [32] and Palencia et al. [75] together gave events (since a fraction falls on regions where the position of the CC is ambiguous) over three epochs of observation. In other words, about total observations are required to distinguish the two models at the current rate of lensed star discovery. Right now, the JWST cycle 4-6 “Dragon” survey (GO-7345, PI: Fudamoto) is being executed and is expected to return five extra epochs of data ( extra events). Expanding the calculation to other lensed galaxies (e.g., those observed by Vast Exploration for Nascent, Unexplored Sources, “VENUS”, GO-6882, PI: Fujimoto) with proper marginalization would boost the amount of data available and help with the comparison.
We can also consider the case for the same best-fitting SIDM model, but instead, compare with the case of the smooth cluster model () and the CDM model (). Since the smooth model is already at significance from the best-fitting SIDM model, and the CDM model is currently at with this Asimov estimate, instead of asking for the number of events required to distinguish between CDM and our best-fitting SIDM at , we compute the number of events required for significance. Repeating the same calculation but putting as , we require a total of events to distinguish between SIDM model of , cm/g, and the CDM model, which is expected to be achievable with the upcoming data from the Dragon survey.
As an important note, none of these calculations have taken into account intrinsic scatter in the fitting quality arising from the stochastic sampling of the subhalo mass function and the placement of subhaloes, as shown in Figure 8. Also, the actual number of observations that is required to separate between models is very likely to be larger than what we estimated here, which is the median expectation. Thus, this exercise serves no more than a rough estimation that, so at least, that the lensed star spatial distribution is a viable way of distinguishing between DM models.
Figure 8. Mean absolute macroscopic magnification across the CC when integrating along the CC. The smooth model is shown in black as a reference curve, where one can see that the magnification drops inversely with the distance from the peak. Blue curves show the case for and (case of minimum CC perturbation), whereas orange curves show and (case of maximum CC perturbation). The left and right panels are the cases when subhaloes take up 1% and 0.5% of the total mass of the cluster, respectively.
Effect of choice of model parameters
Lens model dependency
One of the uncertainties is the dependency on the lensed model, from which we infer the spatial distribution. While we have already avoided using the region with the highest lens model uncertainty near the head of the Dragon, models behave slightly differently at other parts of the Dragon owing to different modelling methods and different numbers of constraints used. Apart from Diego et al. (2025), we test if adopting two other recent models (after 2025, Eid & Keeton [27]; Cha et al. [12]) will affect our inference. Eid & Keeton [27] is a tailored parametric model for the Dragon using the full pixel brightness and allows for extra degrees of freedom for the masses and truncation radii of the cluster members nearby; we adopt an updated version of Cha et al. [12] for the Dragon, constructed by the hybrid code MrMartist (Cha & Jee [11]) that includes ten extra knots as constraints compared with the published version, which only used one image of the Dragon as a constraint. Image reconstruction of these models resembles the morphology of the Dragon galaxy satisfactorily and thus is expected to be reliable for our purpose. Also, at the region with most lensed stars (\alpha\approx2^{\mathrm{h}}39^{\mathrm{m}}52\fs9, ), all models align within of the inference through velocity dispersion mapping in Zhou et al. [102].
We show in Figure 7 the different inferred lensed star spatial distributions with respect to the model critical curves. All three models exhibit a similar skewness towards the negative parity, and agree to within . We recompute the for all our nine models compared with the Eid & Keeton [27] and Cha et al. [12] models, finding that the statistics behave similarly among the models. The , cm/g model remains the best-fitting under the three models, with , , and for Diego et al. (2025), Cha et al. [12], and Eid & Keeton [27], respectively. The KS test gives the exact same conclusion, with , , and for the three models. Such variation is within the Poissonian uncertainty; thus, the choice of lens model does not appear to be critical to our results.

Figure 7. Dependency of lensed star spatial distribution upon choice of recent lens models (Diego et al. 2025; Cha et al. [12]; Eid & Keeton [27]). Again, negative parity is to the right (positive x-axis), and positive parity to the left (negative x-axis).
Substructure mass fraction
Another unknown is the fraction of cluster mass that is taken up by subhaloes. Recent inference from asymmetries in cluster
Perera et al. [77] places a constraint on the subhalo mass fraction of and (Perera in prep.) under the same subhalo mass function. Our adopted 1% is slightly above the former with , but lies within of the latter, as well as the inference from galaxy-lens inferences (e.g., Vegetti et al. [93]; Hsueh et al. [42]). Since increasing the subhalo fraction in our simulation will deviate further from the previous constraints and significantly increase the computational cost, here we alternatively consider qualitatively what would happen if subhaloes take up more or less fraction of the cluster mass by looking at the magnification distribution across the CC given different subhalo mass fractions. We rerun our simulations with 0.5% cluster mass (closer to the inference in Perera et al. [77]) as subhalo instead of 1%, then show the magnification distribution across the CC (by averaging along the CC) for two representative cases that represent the least perturbation (, , blue curve) and the most perturbation (, , orange curve) in Figure 8.
One of the outstanding trends is that the fluctuations in the CC (with respect to the smooth CC) vanish at the same distance of pixels from the CC for the negative parity in the case of , regardless of the mass fraction. For the opposite parity, the fluctuation extends further to pixels when subhaloes take up more of the cluster mass. This trend is not observed in the other case (, ), as milliCCs are forming everywhere in both parities. This could be understood through our derivation in Sec. A – shallow subhaloes can never form milli-CCs far beyond the macroscopic CC in the negative parity, so increasing the fraction of subhaloes does not help at all; on the other hand, adding more subhaloes will nucleate more milli-CCs that are further away from the macroscopic CC in the positive parity by both increasing the probability of having more massive subhaloes and the number density of subhaloes. This aligns with the fact that breaking the criticality in (milli)CC formation depends strongly on the slope of the substructure profile, especially when the substructure is lower in mass. Under this circumstance, the less mass fraction is taken up by subhaloes, the narrower the lensed star spatial distribution would be when and , as the magnification distribution resembles the smooth model. Such models would be less preferred by the observations – as one could observe from Figure 5, a narrow, strongly peaked lensed star spatial distribution like the smooth case (black histogram) will produce an excess of events close to the macroscopic CC instead of a plateaued distribution. In other words, the lensed star spatial distribution would also simultaneously place a limit on the subhalo mass fraction.
The amplitude, on the other hand, depends on the fraction of cluster mass in DM subhaloes: in both cases, , and , , the fluctuation in magnification is more dramatic with more subhaloes, while bringing down the peak magnification. This could be easily understood as a more abundant number of subhaloes increasing the perturbation towards the CC, redistributing the peak magnification to regions further away from the CC, as shown also in Figure 3.
Qualitatively, the subhalo mass fraction should not shift the skewness of the lensed star spatial distribution: milliCCs – the prerequisite of lensed stars' negative skewness – can barely form in the negative parity as long as the subhaloes are of lower mass as considered in this work and have a shallow profile, regardless of the mass fraction. In fact, increasing the amount of shallow subhaloes will only form more milliCCs in the positive parity, leading to even more lensed stars towards the positive parity. This is also true for the case of steep subhaloes – increasing the mass fraction does not significantly alter the magnification distribution at the two parities, and therefore should not affect the skewness of lensed stars.
Core size
The size of the core for core-collapsed subhaloes is also unknown in a cluster environment. It was shown in Williams et al. [96] that a smaller core (for a different profile) will lead to stronger CC perturbation in the negative parity. From Figure 3, one can see that the core size is certainly affecting our case – magnification distributions are drastically different when uncollapsed subhaloes (with large cores, , left column) and collapsed subhaloes (with small cores, , right column) are respectively dominating the subhalo population, regardless of . Here, we would like to test the sensitivity of our method to the collapsed core size by rerunning our simulations while replacing the core size of the core-collapsed subhaloes with different values. For the sake of computational power, we only compare the magnification maps for , where the perturbation of CC is most dramatic across all models, as all subhaloes are core-collapsed.
We show the magnification maps in Figure 9 for (left), (middle) and (right). For a shallower slope of , the choice of is insignificant. While the formation of milliCCs at the positive parity decreases with increasing , no milliCCs are formed at the negative parity at any . This is because the profile is too shallow and could never have achieved supercriticality in the negative region, even with a small core, let alone with a larger core now. Qualitatively, does not affect the lensed star spatial distribution when is small enough. At , a small fraction of milliCCs are formed at negative parity when . Increasing makes it more and more difficult to achieve supercriticality in the negative parity region, generating fewer and fewer milliCCs. At , no milliCC could be formed at the negative parity. A similar effect is observed in the – however, since is extremely steep, even if we increase to , milliCCs are still formed at the negative parity, and the macro-CC is still being perturbed in both parities following the location of subhaloes. As shown earlier in Figure 3, the dramatic perturbation for is only absent when , where none of the subhaloes have undergone core collapse and thus remain at . Since the formation of milliCC is the main driver behind the negative skewness of lensed stars’ spatial distribution, the choice of core size of core-collapsed SIDM subhaloes when is less significant for the magnification map and therefore the expected lensed star spatial distribution, unless is increased to a value at which essentially the core size is the same as when it is not collapsed.

Figure 9. Same as Figure 3, except we adopt for all the simulations, and each column now instead represents a different core radius, , as indicated in the title of each panel.
Truncation radius
Also, we assumed that subhaloes follow a truncation radius that matches the same total enclosed mass for an initial NFW profile truncated at . As briefly mentioned in § 4, tidal effects in the cluster are expected to not only speed up the core-collapse process, but also strip down the subhaloes to become smaller in radius. Since the stripping will return some of the subhalo mass back to the ambient environment, and the number of subhaloes is conserved upon tidal stripping8, this would naturally lower the fraction of cluster mass that is taken up by subhaloes. To inspect what happens if the subhaloes are tidally stripped to smaller truncation radii, we repeat our calculation following the same suite of sampled NFW subhaloes truncated at , then strip off their mass at and redistribute back to the cluster halo. The SIDM subhaloes are then matched to the total mass of each of the corresponding subhaloes enclosed within , the same way as described in § 2. This way, the setup is identical to our original simulation but having subhaloes with small truncation radius. The choice of is motivated by extrapolating the inference in Niemiec et al. [68]; Sirks et al. [85]; Chiang et al. [13], which return roughly of the mass back to the cluster. Nevertheless, these works study the mass loss and stripping of more massive haloes () in simulated clusters, which are at least an order of magnitude more massive than our considered subhaloes. Our less massive, more compact (given the mass-concentration relation, Diemer & Kravtsov 2015) subhaloes could have lost less mass [68] and sink more slowly than the massive subhaloes [85]. We hereby stress that this is no more than a qualitative assessment of how the tidal stripping is expected to affect our result.
We show all the revised magnification maps with a smaller truncation radius of in Figure 10. Although the number density of the subhaloes is identical, the smaller truncation radii made them more compact, and formation of milli-CC is more vigorous even at a smaller in the negative parity. The magnification distribution behaves no differently than the original simulation when is small, and no subhaloes have undergone core-collapse; however, the steeper profile thus yields more abundant milli-CC in the negative parity, creating a negatively skewed and broader lensed star spatial distribution even at . The quality-of-fit is nearly unaffected for the case of , but improved significantly for the case of and under to and , comparable (or even slightly better) than the original best-fitting SIDM model of and (). For the case of , despite the tidally truncated profiles being more compact, they failed to create sufficient milli-CCs on both sides of the CC that can redistribute the magnification and explain the observed width of the lensed star spatial distribution. The improvement in for SIDM subhaloes under tidal truncation is within the seed noise.
Figure 10. Same as Figure 3, except we adopt a smaller truncation radius for the matched NFW subhaloes, instead of , to mimic a scenario where our subhaloes are significantly stripped through tidal effects.
One extra contribution of tidal truncation we have neglected in our calculation is that it would simultaneously speed up the core collapse. Nishikawa et al. [70] demonstrated that NFW haloes truncated at will core-collapse times faster than an untruncated halo. For our given range of subhalo mass, would roughly correspond to , the same truncation radius we tested in this section. In other words, the preferred interaction cross-section could be read as up to times smaller when we consider tidal truncation at , i.e., preferring to be instead of .
This demonstrates that the truncation radius does affect our simulation and will affect the preferred interaction cross-section, , at least for the considered alternative truncation radius. Since this tested truncation radius is not a stringent result from tailored simulation and calculation, the tidal truncation of SIDM subhaloes in cluster environments has to be thoroughly explored (as much as how is coupled with ) before we can refine our simulation to better understand the constraining power of the interaction cross-section of subhaloes in clusters.
Slope of luminosity function
We adopted for the slope of our adopted luminosity function (Eqn. (13)) – as the slope of the luminosity function affects the degree of magnification bias (Broadhurst et al. [8]) and therefore the lensed star distribution, we follow the uncertainty inferred in Palencia et al. [75] and inspect how a different changes our results by repeating our calculations with (lower from Palencia et al. [75]) and (value deduced in Diego et al. 2024; Broadhurst et al. [8]). The predicted spatial distributions are shown in Fig. B.1 and Fig. B.2, respectively.
The main difference with different s is that a steeper luminosity function creates more events in the central region and makes all the fits worse. This is because of the magnification bias – the increase in the number of fainter stars outperforms the reduction in source plane area at higher macroscopic magnification. While , remains the best fit for all the s considered, the statistical preference among models increases (decreases) for a steeper (shallower) luminosity function. As a steeper luminosity function makes all fits worse, the best-fitting SIDM model is now preferred over the smooth model, and preferred over the CDM model. One would just need extra events to distinguish between CDM and the best-fitting SIDM to confidence. With , the preference of over grows further to a Bayes factor of , and the model is definitively preferred over with Bayes factor of . Vice versa, swapping to weakens the preference of the best-fitting SIDM model over the smooth model to , and the CDM model is no longer significantly distinguished from the SIDM model (). One would need a total of events to distinguish between CDM and SIDM, and events to distinguish between the smooth and SIDM model, to confidence. While is still strongly preferred over with a Bayes factor of , and are no longer distinguished with a Bayes factor of only . Although the results of our work remain unchanged qualitatively, should be treated as a nuisance parameter for future work.
As a side note, the effect of is weak for the case of waveDM simulations in Broadhurst et al. [8], like our case of SIDM model with and . Swapping from to barely moves the reduced – the reduced s remain as , , and for the three de Broglie wavelengths of , and , respectively. This is very likely arising from the fact that the magnification has been drastically redistributed across the CC in these cases (see, e.g., the orange curve in Figure 8), further departing from the smooth CC case. The slope of the luminosity function controls the level of magnification bias (Broadhurst et al. [8]), which determines the sensitivity of the lensed star detection rate towards higher magnification. A steeper slope, as we have shown earlier, gives rise to more events in the central region where the magnification is high. Once the high magnification is redistributed across the CC, the sensitivity becomes less relevant to the lensed star spatial distribution, as there is high magnification everywhere away from the original peak.
Abundance of microlenses
While the abundance of stellar microlenses of is well motivated by intracluster light (ICL) analysis, portions of the Dragon arc are submerged by the stellar populations in cluster members. Although our subsample of lensed stars is far from such regions and the estimated ICL density should be robust, we briefly explore what would happen to our result if the abundance of microlenses increases by a factor of two, to , by rerunning our simulations with this revised abundance. This would also correspond to a possible case in which dark compact objects have taken up more fractions of DM than what was previously constrained (Green & Kavanagh [38]).
The difference between the reduced of the two most distinctive models, , and , , dropped from 2.89 and 1.15 (difference of 1.74) to 1.64 and 1.23 (difference of 0.4). Such a difference would then fall within the sampling uncertainty as shown earlier in § 4. While the qualitative preference persists, the statistical significance in distinguishing the models dropped significantly. This is unsurprising – an abundant amount of stellar microlenses will act as extra substructures and wash out the difference driven dominantly by DM substructures. Alternatively, we can bring down the abundance of stellar microlenses to better understand the situation. Half the original abundance of stellar microlenses, at , would lead to a more significant difference between the two aforementioned models, with and 1.39 respectively. The lensed star spatial distribution would peak more strongly at the center; thus, those models with little redistribution of macroscopic magnification (i.e., small , and low core-collapse fraction, as shown in Figure 3) will be more disfavoured by the data. This is similar to the case where we have a steeper for the background luminosity function, as discussed previously in § 5.5.
Nevertheless, we stress that this is purely a sensitivity test, as the uncertainty in the abundance of stellar microlenses should be much lower than the other unknowns given that it is purely photometric. Stacking repeated JWST observations would help increase the signal-to-noise in the ICL light, allowing for an even more precise determination of the abundance of stellar microlenses. Also, studying light curves of individual lensed stars could help one to distinguish the underlying low-mass initial mass function of the lensing population (e.g., Kelly et al. [48]), further reducing the uncertainty in the abundance of stellar microlenses. Based on our results, lensed galaxies with a lower abundance of surface mass density would be beneficial in testing the nature of DM, given their relatively higher sensitivity to the underlying macroscopic magnification perturbed by DM substructures. Notice that background stars would be able to attain even higher maximum magnification given the lower abundance of stellar microlenses (Li et al. 2026), beneficial in increasing the number of lensed star detections, which is critical in distinguishing DM models.
Globular clusters as extra perturbers
We have neglected the existence of globular clusters (GCs) in our simulation. GCs are known to be well described by the very compact King [50] profile, the same as SIDM core-collapsed subhaloes. Although GCs typically have masses in a range lower () than the subhaloes considered in our simulations, they can also perturb the magnification distribution locally (e.g., Diego et al. [21]; Palencia et al. [73]) and lead to local over-density of lensed stars. The number density of GCs traces the ICL distribution closely, with both contributing, to the first order, the same surface mass density in galaxy clusters (Diego, J. M. et al. [24]).
The latest deep observation of cluster AS1063 at mag with GLIMPSE (Atek et al. [5]) revealed GCs, yielding a GC surface density scaling as with the lensing convergence (Diego et al. [20]). Our subsample of lensed stars spans an area of , with according to the Diego et al. (2025) lens model. If we assume the relation holds also at A370, this yields GCs that can act as potential perturbers. The GC mass function follows a log-normal distribution, centered at with a width of dex (e.g., McLaughlin & Fall [59]; Baumgardt & Hilker [6]; Moreno-Hilarío et al. [63]), meaning that we would expect only GC with within our region of interest. The majority of the GCs would have mass significantly lower than our considered mass range of subhaloes, meaning that they would have limited perturbation towards the macroscopic magnification, and are rather unlikely to lead to extra bias in our result. Nevertheless, one strategy would be to avoid lensed stars discovered near massive GCs revealed by deep observations to minimize the effect of GCs on the inferred lensed star spatial distribution.
Conclusion
In this paper, we study for the first time the role of SIDM core-collapsed subhaloes in the spatial distribution of lensed stars now regularly found in deep, repeated HST and JWST cluster observations. Adopting core-collapsed power-law profiles from Gilman et al. [35], we explore how the steepness, core size of core-collapsed subhaloes, as well as the SIDM interaction cross-section affect the magnification distribution across the macroscopic CC and therefore the lensed star spatial distribution.
We found that the power-law slope, , is determinant of the skewness of the lensed star spatial distribution – only when is large enough () can milliCCs be formed in the negative parity of the macro CC when the subhaloes are core-collapsed, therefore leading to negative skewness of lensed stars as observed in clusters. A larger interaction cross-section that leads to a larger fraction of subhaloes that have undergone core-collapse is a prerequisite for milliCCs to form in the negative parity with density slopes that are steep enough to explain the negative skewness. At face value, the observed lensed star spatial distribution qualitatively favours and , which fits the observed lensed star spatial distribution in the Dragon galaxy (, Kelly et al. [47]; Fudamoto et al. [32]; Palencia et al. [75]) compared with the Diego et al. (2025) lens model down to a reduced of (p-value of 0.45 from KS test). Moreover, models marginalizing over indicate a definitive preference for over with a Bayes factor of , and moderately prefer over with a Bayes factor of .
The Poissonian statistics, combined with the fact that the effect of tidal disruption upon the subhalo core collapse probability is unclear in a cluster environment, as well as the unknown correlation between and , made it impossible to draw solid constraints on the SIDM interaction cross-section in this work. Nevertheless, our methodology demonstrated that the lensed star spatial distribution can be used to constrain the SIDM interaction cross-section when sufficient data has been accumulated, and when the tidal effects are better understood. We also show that while both the best-fit claimed in Broadhurst et al. [8] – wave DM with a de Broglie wavelength of pc – and our best-fitting SIDM can reproduce the current observations down to , the two models can be distinguished at level with a total of (extra ) cluster lensed star events, assuming that one model is the ground truth. To distinguish the best-fitting SIDM model and the standard CDM model, we would require a total of (extra ) events to reach confidence level.
We further explored the parameter space, including the subhalo mass fraction, the core size and truncation radius of the core-collapsed SIDM subhaloes, as well as the luminosity function of the background galaxy and the abundance of stellar microlenses. We found that the choice of these parameters does not affect our main result qualitatively, but affects the statistical significance of our results. We also deduce that the GC population does not carry a significant effect on our inference, but a careful treatment of excluding lensed stars near bright GCs (if any) would be beneficial. The ongoing monitoring campaign on the Dragon, combined with scheduled observations at many different clusters, is expected to return a suite of events and narrow down the uncertainty of some of the parameters (e.g., the slope of the luminosity function). This could eventually allow for distinguishing DM models based on lensed star observations, upon a careful treatment of lens model systematics and marginalization over parameter space.
Acknowledgements. S.K.L. acknowledges the generous support from the Croucher Foundation through the Croucher Postdoctoral Fellowship. S.K.L., J.L. and A.A. acknowledge support from the Research Grants Council (RGC) of Hong Kong through the General Research Fund (GRF) 17312122 and 17302236. Y.H. and Y.Z. are supported by the GRF Grants No. 11302824 and 11310925, and PROCORE F-CityU/06/25 from RGC, and the Grants No. 9610645 and No. 7020130 from the City University of Hong Kong. J.M.D. acknowledges the support of projects PID2022-138896NB-C51 (MICIN/AEI/10.13039/501100011033 and FEDER, UE) Ministerio de Ciencia, Investigación y Universidades and SA101P24. A.A. acknowledges the support from the RGC Junior Research Fellow Scheme. T.B. and P.M. are supported by the Spanish Grant PID2023-149016NB-I00 (MINECO/AEI/FEDER, UE). PM acknowledges financial support from fellowship PIF22/177 (UPV/EHU). T.B. is supported by the Basque Government Grant No. IT1977-26. S.C. acknowledges support from the EACOA Fellowship awarded by the East Asia Core Observatories Association. L.W.H.F. and R.M. acknowledge funding from the UK Science and Technology Facilities Council via grant ST/X001075/1. The Instituto de Física de Cantabria (IFCA) has been accredited as Unidad de Excelencia María de Maeztu, a distinction awarded by the Spanish State Research Agency for the period 01/06/2026 - 31/05/2030. This publication is part of the grant «Excelencia María de Maeztu» CEX2025-001620-M, funded by MICIU/AEI/10.13039/501100011033. We acknowledge the use of the following programs: Astropy ([4, 3, 2]), Numpy ([39]), Matplotlib ([43]), and Scipy ([95]). Data will be shared upon reasonable request.
Analytical understanding of formation of milliCCs
Our numerical simulations have shown that milliCCs always form at positive parity, and only form at negative parity when is large enough and the core size is small after core collapse. Here, we perform an analytical derivation to better understand why this would happen.
The formation of (milli)CC happens when the sign of the magnification changes. In other words, there exists a criticality for which this happens. We can rewrite Eqn. 11 in terms of lensing convergence, , and shear, , such that for a tangential arc we are interested in this work, the criticality happens when the tangential eigenvalue, , given that the radial eigenvalue, . When a subhalo (subscript, ) is perturbing the macroscopic cluster environment (subscript, ), we can rewrite the convergence and shear as:
The Abel projection of Eqn. 7, assuming , gives:
The shear produced by the subhalo as a function of distance from the center is:
with the mean enclosed convergence at :
Substituting and rearranging these gives:
One can see that when , the ratio is one. But then when increases, this ratio grows steeply and eventually diverges at . This aligns with one’s expectation, as a steeper profile has more mass () towards the center. In other words, for the range of considered in this work, the shear component of subhaloes at any given radius will become strong and stronger than the convergence component when we increase .
With these ingredients, we can ask when a subhalo perturbing the cluster field nucleates a new (milli)CC. The combination in Eqn. A.1, adding directly onto , is only valid along one specific direction: the line joining the subhalo to the cluster center, where the subhalo’s own shear (circular about its own center) happens to point the same way as the macro shear (circular about the cluster center). This is the direction that controls the tangential eigenvalue, , and since can only push this quantity down, it can only be driven through zero from the positive-parity side, explaining why regardless of the subhalo concentration, CCs are always nucleating in the positive parity.
Perpendicular to that line, the subhalo’s shear instead opposes rather than reinforcing it, and it is the radial eigenvalue, , that gets pushed toward zero there. Because everywhere in the negative-parity region by definition, this channel is available throughout that region, not just at its boundary.
Substituting (Eqn. A.1) into each of these two conditions in turn gives two thresholds on where CC can form in the two respective parities:
Since and are different numbers whenever —a subhalo at fixed can satisfy one comfortably while falling short of the other. Since for any subhalo embedded in the macro-model shear field, , so the positive-parity threshold is always lower than its negative-parity counterpart the positive-parity condition is the easier of the two to satisfy, for any . As , both thresholds approach , so only the steepest subhaloes can nucleate a milliCC at all; as grows relative to , both thresholds drop well below , admitting progressively shallower profiles. Because the positive-parity threshold has the smaller denominator, it falls faster as increases, so a subhalo of intermediate steepness can cross it while the negative-parity threshold remains out of reach. This is why steep, core-collapsed subhalos nucleate milliCCs on both sides of the macroscopic CC, while shallower ones are confined to the positive parity side alone. This also explains why the orientation of the hourglass-shaped milliCCs are orthogonal in the two parities, as shown in the literature [e.g. 18, 96], as well as in Figure 3.
Spatial distribution for different s
Figure B.1. Same as Figure 5, but with . A shallower slope washes down the central concentration and leads to broader distributions, diminishing the statistical significance among different models.
Figure B.2. Same as Figure 5, but with . One can see that most distributions are strongly peaked in the center, promoting the statistical significance betwen different models.
References
- [1]Amruth, A., Broadhurst, T., Lim, J., et al. 2023, Nature Astronomy, 7, 736
- [2]Astropy Collaboration, Price-Whelan, A. M., Lim, P. L., et al. 2022, ApJ, 935, 167
- [3]Astropy Collaboration, Price-Whelan, A. M., Śipőcz, B. M., et al. 2018, AJ, 156, 123
- [4]Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33
- [5]Atek, H., Chisholm, J., Kokorev, V., et al. 2025, JWST’s GLIMPSE: an overview of the deepest probe of early galaxy formation and cosmic reionizationarxiv.org/abs/2511.07542
- [6]Baumgardt, H. & Hilker, M. 2018, MNRAS, 478, 1520
- [7]Boddy, K. K., Kaplinghat, M., Kwa, A., & Peter, A. H. G. 2016, Phys. Rev. D, 94, 123017
- [8]Broadhurst, T. Li, S. K., Alfred, A., et al. 2025, ApJ, 978, L5
- [9]Buckley, M. R. & Fox, P. J. 2010, Phys. Rev. D, 81, 083522
- [10]Cha, S., Cho, B. Y., Joo, H., et al. 2025, The Astrophysical Journal Letters, 987, L15
- [11]Cha, S. & Jee, M. J. 2026, The Astrophysical Journal, 997, 18
- [12]Cha, S., Limousin, M., & Jee, M. J. 2026, Testing Offsets Between Cluster-Scale Halos and BCGs in Strong Lensing Models Using the Jackknife Methodarxiv.org/abs/2609.38310
- [13]Chiang, B. T., Dutra, I., & Natarajan, P. 2026, The Astrophysical Journal, 997, 106
- [14]Choi, J., Dotter, A., Conroy, C., et al. 2016, The Astrophysical Journal, 823, 102
- [15]Cowan, G., Cranmer, K., Gross, E., & Vitells, O. 2011, The European Physical Journal C, 71
- [16]Dai, L., Kaurov, A. A., Sharon, K., et al. 2020, MNRAS, 495, 3192
- [17]Dai, L., Venumadhav, T., Kaurov, A. A., & Miralda-Escud, J. 2018, The Astrophysical Journal, 867, 24
- [18]Diego, J. M. 2019, Astronomy & Astrophysics, 625, A84
- [19]Diego, J. M. 2026, The large cores of dark matter and globular clusters in AS1063. Possible evidence of self-interacting dark matter. Or not
- [20]Diego, J. M., Goolsby, C., Conelisce, C. J., & Palencia, J. M. 2026, Astronomy & Astrophysics, 709, A248
- [21]Diego, J. M., Li, S. K., Amruth, A., et al. 2024a, A&A, 689, A167
- [22]Diego, J. M., Li, S. K., Meena, A. K., et al. 2024b, A&A, 681, A124
- [24]Diego, J. M., Pascale, M., Frye, B., et al. 2023, A&A, 679, A159
- [25]Diemer, B. & Kravtsov, A. V. 2015, The Astrophysical Journal, 799, 108
- [26]Dutra, I., Natarajan, P., & Gilman, D. 2024, The Astrophysical Journal, 978, 38
- [27]Eid, L. & Keeton, C. R. 2025, The Astrophysical Journal, 990, 196
- [28]Enzi, W. J. R., Krawczyk, C. M., Ballard, D. J., & Collett, T. E. 2024, The over-concentrated dark halo in the strong lens SDSS J0946+1006 is a subhalo: evidence for self interacting dark matter?arxiv.org/abs/2411.08565
- [29]Essig, R., McDermott, S. D., Yu, H.-B., & Zhong, Y.-M. 2019, Phys. Rev. Lett., 123, 121102
- [30]Feng, J. L., Kaplinghat, M., Tu, H., & Yu, H.-B. 2009, JCAP, 07, 004
- [31]Fischer, M. S., Yu, H.-B., & Dolag, K. 2025, Astron. Astrophys., 703, A234
- [32]Fudamoto, Y., Sun, F., Diego, J. M., et al. 2025, Nature Astronomy, 9, 428
- [33]Gilman, D., Nierenberg, A. M., Gurian, J., et al. 2026a, JWST lensed quasar dark matter survey V: Hints of self-interacting dark matter from 29 quadruply imaged quasars
- [34]Gilman, D., Petraki, K., Pignard, T., et al. 2026b, Hints of self-interacting dark matter from strong gravitational lensingarxiv.org/abs/2610.40766
- [35]Gilman, D., Zhong, Y.-M., & Bovy, J. 2023, Physical Review D, 107
- [36]Gnedin, O. Y., Hernquist, L., & Ostriker, J. P. 1999, ApJ, 514, 109
- [37]Gómez, P. L., Valkonen, L. E., Romer, A. K., et al. 2012, AJ, 144, 79
- [38]Green, A. M. & Kavanagh, B. J. 2021, Journal of Physics G: Nuclear and Particle Physics, 48, 043001
- [39]Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357
- [40]Harvey, D., Tittley, E., Massey, R., et al. 2014, Monthly Notices of the Royal Astronomical Society, 441, 4046
- [42]Hsueh, J.-W., Enzi, W., Vegetti, S., et al. 2020, Monthly Notices of the Royal Astronomical Society, 492, 3047
- [43]Hunter, J. D. 2007, Computing in Science & Engineering, 9, 90
- [44]Huo, R., Yu, H.-B., & Zhong, Y.-M. 2020, JCAP, 06, 051
- [45]Jiang, F., Dekel, A., Freundlich, J., et al. 2021, Monthly Notices of the Royal Astronomical Society, 502, 621–641
- [46]Jiang, F. et al. 2023, Mon. Not. Roy. Astron. Soc., 521, 4630
- [47]Kelly, P. L., Chen, W., Alfred, A., et al. 2022, arXiv e-prints, arXiv:2211.02670arxiv.org/abs/2211.02670
- [48]Kelly, P. L., Diego, J. M., Rodney, S., et al. 2018, Nature Astronomy, 2, 334
- [49]Kim, S. Y., Peter, A. H. G., & Wittman, D. 2017, Monthly Notices of the Royal Astronomical Society, 469, 1414
- [50]King, I. 1962, AJ, 67, 471
- [51]Koda, J. & Shapiro, P. R. 2011, Mon. Not. Roy. Astron. Soc., 415, 1125
- [56]Lovell, M. R., Frenk, C. S., Eke, V. R., et al. 2014, Monthly Notices of the Royal Astronomical Society, 439, 301–317
- [57]Mace, C., Dhanasingham, B., Zeng, Z. C., et al. 2026, The Sensitivity of Substructure Lensing to SIDM Core-collapse Model Variationarxiv.org/abs/2605.24174
- [58]Mandelbaum, R., Seljak, U., & Hirata, C. M. 2008, J. Cosmology Astropart. Phys., 2008, 006
- [59]McLaughlin, D. E. & Fall, S. M. 2008, The Astrophysical Journal, 679, 1272–1287
- [60]Meena, A. K., Li, S. K., Zitrin, A., et al. 2025, A&A, 699, A299
- [61]Merten, J., Meneghetti, M., Postman, M., et al. 2015, ApJ, 806, 4
- [62]Moore, B., Katz, N., Lake, G., Dressler, A., & Oemler, A. 1996, Nature, 379, 613
- [63]Moreno-Hilario, E., Li, H., & Martinez-Medina, L. A. 2026, Reconstructing the Globular Cluster Initial Mass Function from Present-Day Globular Cluster Systems
- [64]Morishita, T., Abramson, L. E., Treu, T., et al. 2017, The Astrophysical Journal, 846, 139
- [65]Nadler, E. O., Banerjee, A., Adhikari, S., Mao, Y.-Y., & Wechsler, R. H. 2020, Astrophys. J., 896, 112
- [66]Natarajan, P., Chiang, B. T., & Dutra, I. 2026, The Astrophysical Journal Letters, 1001, L12
- [68]Niemiec, A., Jullo, E., Giocoli, C., Limousin, M., & Jauzac, M. 2019, Monthly Notices of the Royal Astronomical Society, 487, 653–666
- [69]Nishikawa, H., Boddy, K. K., & Kaplinghat, M. 2020a, Phys. Rev. D, 101, 063009
- [70]Nishikawa, H., Boddy, K. K., & Kaplinghat, M. 2020b, Physical Review D, 101
- [71]Oke, J. B. & Gunn, J. E. 1983, ApJ, 266, 713
- [72]Outmezguine, N. J., Boddy, K. K., Gad-Nasr, S., Kaplinghat, M., & Sagunski, L. 2023, Monthly Notices of the Royal Astronomical Society, 523, 4786–4800
- [73]Palencia, J. M., Diego, J. M., Dai, L., et al. 2025, A&A, 699, A295
- [74]Palencia, J. M., Diego, J. M., Kavanagh, B. J., & Martínez-Arrizabalaga, J. 2024, A&A, 687, A81
- [75]Palencia, J. M., Sun, F., Diego, J. M., et al. 2026, arXiv e-prints, arXiv:2604.22702DOI
- [77]Perera, D., Gilman, D., Williams, L. L. R., et al. 2026, Monthly Notices of the Royal Astronomical Society, 547, stag486
- [78]Planck Collaboration, Aghanim, N., Akrami, Y., et al. 2020, A&A, 641, A6
- [79]Pollack, J., Spergel, D. N., & Steinhardt, P. J. 2015, Astrophys. J., 804, 131
- [80]Powell, D. M., McKean, J. P., Vegetti, S., et al. 2025, Nature Astronomy, 9, 1714–1722
- [81]Randall, S. W., Markevitch, M., Clowe, D., Gonzalez, A. H., & Bradač, M. 2008, ApJ, 679, 1173
- [82]Robertson, A., Massey, R., & Eke, V. 2016, Monthly Notices of the Royal Astronomical Society, 465, 569–587
- [83]Schneider, P., Ehlers, J., & Falco, E. E. 1992, Gravitational Lenses
- [84]Shirasaki, M., Okamoto, T., & Ando, S. 2022, Monthly Notices of the Royal Astronomical Society, 516, 4594–4611
- [85]Sirks, E. L., Oman, K. A., Robertson, A., Massey, R., & Frenk, C. 2022, Monthly Notices of the Royal Astronomical Society, 511, 5927–5935
- [86]Spergel, D. N. & Steinhardt, P. J. 2000, Phys. Rev. Lett., 84, 3760
- [87]Sánchez Almeida, J. 2025, Astronomy & Astrophysics, 704, A210
- [88]Tajalli, M., Vegetti, S., O’Riordan, C. M., et al. 2025, Monthly Notices of the Royal Astronomical Society, 543, 540–564
- [90]Turner, H. C., Lowell, M. R., Zavala, J., & Vogelsberger, M. 2021, Mon. Not. Roy. Astron. Soc., 505, 5327
- [91]Umetsu, K., Ueda, S., Hsieh, B.-C., et al. 2022, The Astrophysical Journal, 934, 169
- [92]van den Bosch, F. C. & Tokayer, Y. M. 2026 [arXiv:2610.28890]arxiv.org/abs/2610.28890
- [93]Vegetti, S., Koopmans, L. V. E., Auger, M. W., Treu, T., & Bolton, A. S. 2014, Monthly Notices of the Royal Astronomical Society, 442, 2017–2035
- [94]Vegetti, S., White, S. D. M., McKean, J. P., et al. 2026, Nature Astronomy, 10, 440–447
- [95]Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, Nature Methods, 17, 261
- [96]Williams, L. L. R., Kelly, P. L., Treu, T., et al. 2024, ApJ, 961, 200
- [97]Xie, L. & Gao, L. 2015, Monthly Notices of the Royal Astronomical Society, 454, 1697–1703
- [98]Yan, H., Ma, Z., Sun, B., et al. 2023, ApJS, 269, 43
- [99]Yang, D., Nadler, E. O., Yu, H.-B., & Zhong, Y.-M. 2024, Journal of Cosmology and Astroparticle Physics, 2024, 032
- [100]Zeng, Z. C., Peter, A. H. G., Du, X., et al. 2025, Phys. Rev. D, 111, 063001
- [101]Zhong, Y.-M., Yang, D., & Yu, H.-B. 2023, Mon. Not. Roy. Astron. Soc., 526, 758
- [102]Zhou, R., Dai, L., Ji, L., et al. 2025, arXiv e-prints, arXiv:2511.18479DOI








