WorksPhysics and Astronomy
Localization of gravitational-wave source behind galaxy clusters by gravitational-wave lensing
Abstract
Gravitational-wave observations have opened new opportunities to study massive object events like merging binary black holes. When combined with electromagnetic counterparts, accurate localization of gravitational-wave sources could allow us to identify their host galaxies and carry out follow-up studies, such as study the environment of merging black holes and determine the Hubble constant. One of the possible methods for precise localization is strong gravitational-wave lensing. Previous studies have considered strong gravitational-wave lensing with galaxy lenses. However, galaxy clusters have yet to be investigated due to the complexity of their mass structures. Here we take the list of known cluster lenses from the Hubble Frontier Fields, simulate lensing properties of gravitational waves from their reconstructions as they would be observed by the ground-based gravitational-wave detectors. When same model family is used in simulation and localization, we find that the correct lens is identified for around 90% of the simulated events. We also show that the source positions within the cluster can be pinpointed at sub-arcsecond precision with an estimate on source redshift. When we marginalize over all possible lens models for the clusters, the lens identification and localizations remain valid as long as the correct lens model is contained in the available cluster lens models. However, most localization attempts fail at initialization if the correct cluster model is excluded. These tests establish a proof of concept of localization under idealized conditions, while more work is needed to understand the cluster lens modeling discrepancies and reconstruction bias to validate the approach further.
Introduction
Observations of gravitational waves (GWs) from binary compact objects provide a powerful tool for gravitational research. Analysis of GW signals allows us to test gravitational theories and directly detect electromagnetically invisible sources, such as binary black holes (BBHs) [6].
For cosmological measurements such as the Hubble constant [2], GW observations can provide strong constraints if the sources are well localized. Identifying their host galaxies further enables multi-messenger observations of GW and electromagnetic (EM) signals from local environments [33]. Such studies support independent distance measurements without relying on indirect methods such as the distance ladder [2], highlighting the essential role of GW localization.
However, localizing GW sources remains a major challenge. Hundreds of GW events have been identified by the LIGO–Virgo–KAGRA (LVK) detector networks. The number of detected GW events is expected to rise substantially in future LVK observing runs (B. P. Abbott et al. 2020) and increase by orders of magnitude with planned third-generation ground-based GW observatories, such as the Einstein Telescope and Cosmic Explorer [26, 10]. However, most current LVK network localizations still span tens to hundreds of square degrees, enclosing large numbers of galaxies [13], even though some well-measured events can be localized much more tightly (for example, GW190814 was localized to 18.5 deg at 90% credibility (R. Abbott et al. 2020)). Without an EM counterpart, associating GWs with their host galaxies is extremely difficult, particularly for BBHs, which do not emit light. One of the few exceptions is GW170817, the binary neutron star merger that produced a number of EM counterparts including a gamma-ray burst, allowing better estimation of its sky location [2]. By contrast, GW events by BBHs still face the challenge of poor localization.
Strong gravitational lensing of GWs may be a solution to the problem. Like EM waves, GWs are bent and magnified when passing near massive objects such as galaxies or galaxy clusters, producing effects such as delaying and amplifying GW signals [44]. Strong lensing, in particular, can create multiple copies of the same signal with different arrival times and amplitudes, and potentially distorting the signals in a frequency-dependent manner [41]. Previous studies have shown that if a GW is strongly lensed, the possible sky location of the source can be greatly reduced [18], and the candidate host galaxies can be narrowed to one or a few using the ratio between time delays and magnifications [13, 45].
Previous GW lensing studies have focused on galaxy lenses [13, 42], while almost all such studies exclude detailed analyses of cluster lensing. The reason is that complex and irregular mass distributions make cluster lenses challenging to model analytically [30, 38].
However, cluster lensing has unique advantages over galaxy lensing in both GW-only and multi-messenger studies. Complex cluster mass structures allow greater image multiplicity than typical smooth lens models [28]. While common galaxy-lens models, which are usually smooth mass models, produce a maximum of four images, cluster lenses can produce more. A lensed GW signal with larger number of images has more constraints to cross-match the GW image properties against the EM lens reconstructions. Also, the images can impose stronger constraints on the sky locations of GW sources in parameter estimation (PE). The association of lensed GW signals with cluster lenses is also easier. Galaxy clusters are much larger than galaxies and are therefore easier to observe. Moreover, there are fewer candidate lenses in a sky patch where the lensed signals are localized using GW data only. The distinctive shapes of clusters can also limit the possible candidates for a detected lensing signal. With a multi-messenger association, it may also be possible to estimate the source redshift. This quantity is closely tied to lensed GW signals, as described in the next section. Therefore, cluster lenses should be included for a comprehensively study on how strong lensing affects GW source localization.
Although no lensing signals have been confirmed in current observations [15, 5, 1] (A. G. Hannuksela et al. 2023), the rate of lensing events in O5 is expected to be around 1 yr [46, 39, 34]. Cluster lensed events may be more difficult to detect because of the larger time delay allowing higher fraction of false coincidences [14]. However, the third-generation GW detectors are expected to detect orders of magnitude more signals [35, 26, 10, 34] and hence more lensed signals, making cluster lensed events detections increasingly promising in future observing runs. In this study, we ask whether the detection of multiple images of a cluster-lensed GW signal can identify the cluster lens and constrain the position and redshift of the source. This extends previous work on localization with galaxy lenses to galaxy clusters.
In this paper, we first describe the detailed method to study localization of cluster-lensed GW signals by simulations in § 2, then present the localization results of the simulated events in § 3. Finally, we conclude in § 4 with discussions on the results, design and potential extensions of the study.
Method
Instead of analytical models, we adopt published numerical galaxy cluster lens reconstructions available in the literature and build the GW localization framework on top of these models. In particular, using selected numerical lens models (Section 2.1), we simulate GW events with a given position in the source plane and a source redshift and obtain the lensing observables: the effective luminosity distances and the relative time delays between image and the first image (Section 2.2). Next, we input these observables into the localization algorithm for one of the clusters (Section 2.3), which maximizes the likelihood of producing images with the observed delays and distances. We repeat this process for each cluster lens candidate, then determine the most probable lens using an approximation of posterior odds.1
We assume throughout that the true cluster is among the candidate lenses. We use the same lens reconstruction for simulation and recovery as an idealized test of the localization framework. We also perform localization tests marginalizing over all available models of the clusters to verify the framework when there are multiple models of the same cluster with discrepancies. Finally, we test recovery with the generating reconstruction excluded for robustness tests when available models may not adequately recover the detail lensing properties of the cluster.
Numerical Galaxy Cluster Models
Current analytical models of galaxy clusters, such as Navarro–Frenk–White profiles [31] and singular isothermal sphere (SIS)-like models, can fail to capture realistic cluster mass structures due to the complicated lensing properties in clusters, such as substructures and asymmetries [29].
Therefore, we use existing numerical models for known cluster lenses to probe our method while retaining realistic cluster lens complexity. The numerical cluster lens models used in this study are from the Hubble Space Telescope Frontier Fields (HSTFF) program (J. M. Lotz et al. [25]). We apply the numerical lens models from the Clusters As Telescopes (CATS) team (M. Jauzac et al. [20, 21]; J. Richard et al. [36]), Keeton (C. McCully et al. [27]) and Sharon (T. L. Johnson et al. [22]). The idealized test adopts CATS reconstruction for each cluster, while other tests make use of all the reconstructions. Details of such tests are shown in § 2.4 and § 2.5.
Figure 1 shows a full-depth image of the galaxy cluster MACS J0416.1-2403 and its numerical model of magnification map as an illustration of the cluster structure.

Figure 1. Full-depth image of the galaxy cluster MACS J0416.1-2403 (left, imaged with HST (J. M. Lotz et al. [25])) and its magnification map (right). The full-depth image shows the distribution of galaxies within the cluster and images produced by the cluster, and the magnification map shows the model caustic indicated by white curves. The irregular magnification pattern illustrates the complex mass distribution of the cluster.
The numerical lens models (M. Jauzac et al. [21]; J. Richard et al. [36]) consist of deflection-angle maps and lens-potential maps, which allow the calculation of image positions given a source position. The details of the calculation for model deflection angles and the uncertainty of the model are presented in Appendix § 7.
Simulation of Lensed GW Event Parameters
In order to test the localization approach with expected observables of cluster-lensed signals, we simulate lensed GW events using the numerical lens models and a set of simulated source position and redshift parameters , then extract the required simulated data for injection. GW sources are introduced as point sources in this study because the source size is negligible compared with the lensing scale. Given the source positions, the image positions of a lensed GW event can be obtained directly using the lens equation (P. Schneider et al. [37]):
where is the image position, is the source position and is the deflection angle at the image position. Therefore, we can calculate the image positions with the source position and deflection angle map by ray tracing.
The numerical models include maps that contain the lens potential at each pixel. The time delays between an unlensed GW signal and its lensed images can then be calculated as
where is the arrival-time function in terms of Fermat potential and is given by
where , , are the angular diameter distance of the lens, the source, and between the lens and source respectively.
For actual observations, the relative time delays between images 2 are considered instead. Note that scales with , thus changing will significantly change the value of .
In addition to the relative time delays, PE of the GW signals can also recover the effective luminosity distance of the source. The luminosity distance is related to by:
where we fix in this study. For lensed events, the observed luminosity distance of image is equivalent to the effective luminosity distance due to the signal amplification by lensing:
where is the magnification of corresponding image. Therefore, the observed luminosity distances can act as extra constraints for the estimation of in localization, and we include these observables in the localization algorithm.
For each source, we use the Python package Lenstronomy (S. Birrer et al. [7]) to compute the relative arrival time delays and effective luminosity distances of its images. We use these values as zero-noise injection.
We use the Lenstronomy (S. Birrer et al. [7]) Python package to calculate lensed images with numerical models, source position and . Relative time delays and effective luminosity distances can then be calculated by Lenstronomy and used as injections for localization.
Here, only and are simulated as lensed GW observables. After simulation, we treat and as the results of GW signal PE and then inject them into the localization algorithm to construct a mock-data study. The uncertainties of observables due to the lens model uncertainties are accounted in the localization. We also assume that has already been measured by EM observations and is included in the numerical models.
We simulate lensed GW events from the numerical lens models of 6 galaxy clusters and assume that the given clusters are all the possible candidate lenses of the lensed GW events. In the main analysis, we generate parameter sets that produce 5-image lensed GW events (yielding 4 and 5 from the observations). We simulated a total of 60 5-image lensed GW events, with 10 events behind each galaxy cluster. We consider five-image GW configurations as a high-information scenario to investigate the constraining power of cluster lensing specifically. A large number of images is possible in cluster lensing due to the complex lensing environment of galaxy clusters. An example of a 5-image lensed GW event is shown in Figure 2.
Figure 2. Example of a five-image lensing event lensed by the galaxy cluster MACS J0416.1-2403 with its critical curves shown. The white curves represent critical curves. The red plus and the crosses mark the simulated source position in source plane and its corresponding images in image plane, respectively. The numbers next to the crosses indicate the absolute magnification. When the source is located inside the inner caustic and near the complex mass distribution, more images can be produced beyond the typical maximum number of images of analytical galaxy lens models which is 4.
The high-multiplicity lensing configurations are not restricted to idealized mass distributions and have been observed electromagnetically. For example, the galaxy cluster SDSS J2222+2745 produces six observed images H. Dahle et al. [8]. These systems demonstrate that galaxy clusters can generate image multiplicities of five or more. Note that the detectable frequency of such systems depends on the image magnifications, source population and detector sensitivity, which are not evaluated in this work.
For the marginalized localization analysis, we select 18 of the 60 simulated five-image events. For the model-mismatch test, we simulated five 4-image events (yielding 3 and 4 from the observations) for each model used for each of the 6 clusters. The simulation of 4-image events is to avoid larger differences between reconstructions for high-multiplicity configurations.
Localization and Evaluation
We perform the localizations by sampling the source positions and redshifts under each detected cluster lens candidate so as to obtain the localization accuracy in general. We do so using the same numerical lens model as in the simulation and assume that the CATS model is adequate to describe the lensing structure for a constrained study in the main analysis. For a candidate cluster , we denote the source position and redshift parameters by , . The GW data includes the relative arrival-time delays and the observed effective luminosity-distances of the associated images , while
denotes the EM lens model of candidate and the full EM candidate catalog is , where is the number of detected EM candidates. The posterior for candidate and its source parameters, assuming the GW is lensed and the lens is in the detected EM candidate set, is then
We assign equal prior probabilities to the candidate clusters, . Since the evidence is independent of the positional parameters and cluster candidates, the denominator of posterior can be treated as a constant in this study and does not affect the localization results.
The prior is adopted to be uniform for and uniform spatial prior is assumed within the cluster model.
For a fixed candidate , we define the localization log-likelihood as:
Assuming Gaussian likelihood, we can express the posterior as simple log-likelihood (see Appendix § 8):
where is the number of images; the hypothesis ; is the observed ; is the predicted under the hypothesis. The uncertainties of and are denoted as and respectively. A detailed explanation of these uncertainties are shown in Appendix § 7 and Appendix § 8.
We first employ the global optimization method, the differential evolution (DE) algorithm (R. Storn & K. Price [40]), to obtain an initial estimate of with relatively high likelihood, preventing the sampling procedure from trapping in local maxima. We then perform MCMC by the Python package emcee (J. Goodman & J. Weare [12]; D. Foreman-Mackey et al. [11]) to explore the posterior distribution and determine the credible intervals of .
To evaluate the determination of the correct galaxy cluster lens, we compare each candidate with the union of the other candidates in the catalog. The posterior odds assuming true cluster in the catalog and using the evidence of each candidate is:
where , and is hypothesis that a lens other than cluster is the lens. The derivation is given in Appendix § 6. The appendix further approximate the odds by maximum likelihood estimators:
where is the best-fit log-likelihood under candidate :
and we can neglect the Occam factors if the detected candidates have comparable prior volumes and comparable posterior widths. We thus use to estimate the confidence of lens identification. We determine whether a cluster lens is identified by a threshold. The integrated-evidence calculation used for model marginalization is described separately below.
Model marginalization
A cluster can have several published reconstructions that predict different lensing observables. We study this by marginalizing the evidence of each model. Let denote the index of one of the available reconstructions of cluster , and , be the normalized likelihood and source prior for that reconstruction. Using the normalized likelihood of Appendix § 8, the individual and model-marginalized evidences are
where is the prior model probability and superscript denotes the marginalized evidence. The adopted prior weights are again uniform over each model.
The cluster-identification odds become
Conditional on cluster , the source posterior is
with posterior model weights
The evidences of each reconstructions are evaluated using the same DE and MCMC stages as in § 2.3. The reconstruction of simulation is retained in the set of reconstructions.
Reconstruction-mismatch test and evaluation
For each event simulated with reconstruction A, we attempt the localization with each alternative reconstruction B A of the same cluster. Here, we perform the tests with four-image events due to the reason described in § 2.2. We use three parametric models in HSTFF which are CATS, Keeton and Sharon for the tests. Five events for each cluster for each family of models are localized and we require the predicted image count to match the simulated event as a pilot test. We record whether each attempt can obtain a valid initial guess on a cluster and the posterior if the correct cluster lens is identified. We then assess the discrepancies between the injected and localized source position and the coverage of the 90% credible intervals.
Results
The 60 source positions for the idealized analysis and caustics of the clusters are shown in Figure 3.
Figure 3. Caustics of the 6 galaxy clusters used in this study with simulated source positions, and the surface mass density. The green curves indicate the cluster caustics, red pluses represent simulated source positions in the source plane. The gray shading represents surface mass density in the image plane for reference. The sources are located inside the inner caustic, often in regions associated with complex critical curves, which permit the formation of multiple lensed images.
We first show the results of matched-model cluster identification (§ 3.1), followed by the source localization results with sampled (§ 3.2) and the results with fixed to its injected value (§ 3.3). In § 3.4, we study the possible correlations between lens identification and properties of both the events and lenses. § 3.5 and § 3.6 present the model-marginalized and reconstruction-mismatch results.
Identification of Galaxy Cluster Lens
We determine the successful identification of the galaxy cluster lens by the approximation of posterior odds defined in the method section. We show the identification confidence of galaxy cluster lens of the simulated events with and without sampling by their posterior odds displaying in Figure 4.
Figure 4. Cumulative frequency distribution of posterior odds of lens identification of events with successful localization, with sampling (upper panel) and without sampling (lower panel) in double log-scale for clarity. The posterior odds are between most (i) and other hypotheses that the corresponding cluster is the lens. For both distributions, are all , and the 10th percentiles are larger than 5 in both sampling cases, indicating strong confidence of the cluster identification.
We find that 88.3% of the simulated events pass the threshold of . The failed events are due to the computational failure of DE to find reasonable initial guesses, while one event has wrong lens identification, making the successful events slightly less than 90% of simulated events. For the simulations with sampled, the posterior odds values are smaller than simulations with known , which is expected due to larger parameter space volume when sampling in higher dimensions. For both types of simulations, successful localized events all have . Also, and 10 for around 10th percentile of the localized event with and without sampled respectively, denoted by the shaded region in the figures. This indicates strong confidence of determining the correct cluster lens for most of the simulated events.
Localization Results with unknown redshift
We jointly infer where the source redshift is not known before identifying the host galaxy of the GW source. The lens identification and localization performance is summarized in Table 1.
| No. events | Identification | |||
| 60 | 53 | 0.33″ | 0.33″ | 0.27 |
Table 1. Summary of the localization results with sampled. The number of simulated events and the number of events with successful lens identification and localization are shown. The corresponding median widths of the 90% intervals for each sampled parameter are also shown.
The table demonstrates that the localization algorithm achieves high precision. We identify the correct cluster in 53 of the 60 events. Four of these 53 events fail the numerical
acceptance criterion for the MCMC fit4, leaving 49 events for source localization. For these 49 events, the median widths of the marginal 90% intervals are 0.33″ in each source coordinate and 0.27 in redshift. Thus, the position precision are in sub-arcsec precision. Besides, the table 1 shows the median uncertainty of , which is relatively large comparing to the injected value of . Although this constraint is less precise than a spectroscopic redshift measurement, it provides useful information when searching for the host galaxy, especially when the inferred redshift is combined with the source-plane localization to narrow the possible host galaxy candidates.
Figure 5 shows the corner plot of one event behind MACS J0717.5+3745 as an example. The sampled marginal posteriors are single-peaked, and the injected value lie within 90% interval for each parameter.

Figure 5. An example of corner plot of a simulated event with MACS J0717.5+3745 as lens and sampling. The blue cross is the true value of the source position, and the dotted lines mark the 90% credible intervals. The true parameter values are within the credible intervals of the localization, with uncertainties of below arcsec, and for . The sources are localized to sub-arcsec precision, while is less constrained.
For the localization uncertainties, we also show the distribution of the localized region and uncertainties for each simulated event in Figure 6. We define as the product of the marginal 90% interval widths of and ; this is a convenient axis-aligned scale and is not a joint 90% credible area in the source plane. The histogram bars show and the marginal 90% interval widths of for each event, excluding failed localizations. Most events have below 0.4 arcsec, and most of the 90% interval widths of span from around 0.1 to 0.5, which is large considering the values of . Therefore, while the source-plane constraints reach arcsec scales in this proxy, the results carry larger uncertainties and serve as rough estimations of their true values.
Figure 6. Distribution of the localization-scale proxy and of the marginal 90% interval width of . The upper panel shows the product of the marginal 90% widths of and (not a joint 90% credible area). Most events have below 0.4 arcsec. The lower panel shows the corresponding interval widths from lensed GW data alone.
In summary, we can identify the galaxy cluster lens of the lensed GW event, constrain the source position with sub-arcsec precision and estimate without a prior identification of its host galaxy. The inferred source position and redshift can then be compared with EM reconstructions of galaxy members in the cluster to narrow down the possible hosts.
Localization results with known redshift
We next investigate a follow-up scenario where has been measured independently by EM observations and infer only . To quantify the potential improvement on source localization and lens identification, we fix to its injected value and infer only the source coordinates . Table 2 summarizes the results.
| No. events | Identification | ||
| 60 | 54 | 0.26″ | 0.26″ |
Table 2. Summary of the localization results with known . Same types of data are shown as in Table 1. The median widths of 90% intervals are approximately 0.26 arcsec in and , indicating higher precision in sky localization than the case with sampling.
Again, only localization results of 50 successful events are shown due to MCMC numerical convergence issues. The success rates of both lens identification and localization are slightly higher than that in the unknown- analysis. Compared with the unknown-redshift scenario, the median marginal 90% interval width is reduced to 0.26″ in each source coordinate. Fixing the source redshift removes one degree of freedom and restricts the range of source configurations compatible with the observed lensing properties. The posterior odds of lens identification also show substantial increase. These improvements demonstrate the potential of EM counterpart redshift measurement to further constrain the source position.
A corner plot for a typical localized event is shown in Figure 7 as an example similar to the previous section. The posteriors of all parameters are single-peaked, indicating no distinct degeneracies in this sampling, which is also the case for most of the performed localizations sampling only .

Figure 7. An example of corner plot of a simulated event with MACS J0717.5+3745 as lens and known . The blue cross is the true value of the source position, and the dotted lines mark the 90% credible intervals. The true parameter values are within the credible intervals of the localization, with uncertainties of below arcsec. The source position is more tightly constrained than in the case where is sampled, showing even higher localization precision.
We again report the localization-scale proxy constructed from the marginal 90% widths of the source-position parameters, and display the distribution in Figure 8. Most events have below 0.2 arcsec, smaller than in the free- analysis because one degree of freedom is removed.
Figure 8. Distribution of the localization-scale proxy when is fixed. As in Figure 6, this is the product of marginal 90% widths rather than a joint 90% credible area. Most events have below 0.2–0.3 arcsec.
An independent redshift constraint could therefore improve positional precision within this framework. We emphasize that fixing to its true value represents an idealized scenario in which is independently determined. In practice, an EM redshift measurement would be associated with the GW observations. It could introduce finite range to help improving the localization which is more realistic than locating the exact .
Correlation of Lens identification Confidence and Event Properties
We analyze the possible correlations between the lens identification confidence represented by and the properties of events and their respective lens models here. For the properties of events, we inspect the relationship between and maximum relative time delay among image pairs and the largest image magnification of each simulated event. There is no clear dependence of posterior odds with and , suggesting single summary statistics of the observables may not predict identification confidence. also has no clear correlation with when it is also sampled.
We then analyze the possible correlations between the cluster lens model properties and the posterior odds. One of the important properties is the relative time delay uncertainty of image positions , which can represent the total uncertainties in estimations. We show the spatial distribution of it in Appendix § 7. From the plot, the region with higher are mostly in the center of the clusters as those are the regions with more complex lensing structures.
We hence display the correlation plots between the lens model properties and in Figure 9. Shown in the figure, has no clear correlations with the largest among the images for each event. However, for the apparent cluster

Figure 9. Correlation plot of with the cluster properties: the largest and the size of the cluster. The upper panel shows the plots of localization results with sampling, and lower panel shows the results with known . There is positive correlation (Spearman correlation 0.3-0.5 with p<0.05) between apparent cluster size and lens-identification confidence.
size5, both lens identifications with or without sampling show clear correlation () with the posterior odds. with Spearman correlation coefficient 0.57 (sampling with ) and 0.3 (sampling without ). Therefore, there is a moderately strong correlation between the apparent size of a cluster lens and its identification confidence.
Model-marginalized identification and localization
We perform tests of lens identification and localization where multiple models of the same cluster is marginalized to account for the model discrepancies. We select 18 five-image events that were successfully localized in the matched-reconstruction analysis. For each event, we average over the available reconstructions of each cluster, including the reconstruction used to generate the event. All 18 events retain the correct unique cluster identification and the source localization. The posterior odds of lens identification are shown in Figure 10.

Figure 10. Probability density of of select events in marginalized-model test, comparing with the values in idealized test. The median of of marginalized-model test is slightly lower than that of the idealized test, suggesting lower confidence in multiple-model situation.
Note that the median of log posterior odds (11.1) is slightly lower than that of the idealized test (11.3). The possible reason is due to the inclusion of evidence of wrong model in the calculation of , diluting the overall confidence of lens identification of one cluster lens. The localization results are the same as in the idealized test due to the correct cluster lens and models included in this test. This test shows that the lens identification and localization remains possible with several available models when the set includes an adequate reconstruction, while the confidence might be slightly lowered.
Reconstruction-mismatch results
We finally test the localization when all reconstructions does not reliably identify the correct cluster by simulating events with one cluster model and attempt to perform localization by other available models of that same cluster. We report per pair of models and cluster, the fraction of events where the correct cluster is identified. The results are presented in Figure 11.
Figure 11. Distribution of successful and failed events in lens identification when there are mismatches of reconstruction used in simulation and localization. Distribution of such events per cluster (upper panel) and per model pair (lower panel) is shown. Around 90% of localization attempts fail to identify the correct cluster lens. Therefore, the localization generally fails in the reconstruction-mismatch setting.
As shown in the figure, only 10.5% of tests can identify the correct cluster lens. Most attempts failed as no cluster candidate yields an initial fit; the remaining failed events have wrong cluster lens identified or invalid fit after MCMC when an initial fit exists. Due to the low proportion of successful tests even in lens identification, the successful localization fraction is also low. Therefore, we conclude that the localization fails in general when there is no adequate reconstruction that describes the cluster lens structure well enough.
Discussion and Conclusion
Gravitational-wave lensing is forecasted to become detectable in the coming years, with current forecasts based on population-I/II stars and the known population of galaxies yielding between one in a few hundred to one in a thousand events being strongly lensed. Lensing is expected to become commonplace with the third-generation detectors S.-S. Li et al. (2018); K. K. Y. Ng et al. (2018); H. Phurailatpam et al. [34]. Although galaxy cluster lensing is predicted to occur less frequently than galaxy-scale lensing S. Hilbert et al. [16, 17]; G. P. Smith et al. [39]; L. Vujeva et al. [43], the optical depth may be more dominant at higher magnification owing to the long caustic curves in galaxy clusters S. Hilbert et al. [16, 17]. With around – total mergers in the third-generation detectors, galaxy cluster lensing should be observed in the future with third-generation detectors. The increasing number of expected cluster-lensed GW signals observed in future motivates investigation of cluster lenses as a complementary population of strong gravitational lenses.
In this study, we demonstrate the feasibility of identifying galaxy cluster lenses and precisely localizing the sources of strongly lensed GW events using numerical lens models of galaxy cluster lens candidates. When the same cluster models are used for simulation and localization, correct cluster lens is identified in 53 out of 60 simulated events and the precision of localization can reach sub-arcsec scale. When is sampled, the injected values lie in the 90% credible intervals with narrow width of 0.33″ in each source coordinate. Therefore, multi-image GW observations, combined with EM cluster lens reconstructions, may provide precise source-plane localization without requiring prior knowledge of the source redshift. An independent measurement of the source redshift can further strengthen the lens association and tighten the localization constraints.
The five-image configurations considered in this study are motivated by the high image multiplicities possible in galaxy cluster lenses. The lensing rate of different image multiplicities in galaxy cluster lenses has not been studied extensively. Nevertheless, such events are plausible when the GW source lies near specific regions of a galaxy cluster, as illustrated in Figure 2. Such configuration is also observed by EM H. Dahle et al. [8]. We adopt five-image systems to explore the additional information available from high-multiplicity cluster lenses. Future work can extend this analysis to events with fewer detected images, as the information provided by effective luminosity distances can compensate for the reduced number of measured time delays.
All 18 previously localized events retain the correct cluster identification and source localization when we marginalize over multiple lens reconstructions that include the generating model. However, if the alternative models have negligible posterior weight, the resulting posterior may be nearly same as the posterior of the model used in simulation. The discrepancies between models are also not quantitatively analyzed. Therefore, the results in this test suggest the feasibility of localization in multi-model scenario instead of verifying it in general.
Lens identifications and localizations are less successful in the four-image test with the generating reconstruction excluded. Most attempts fail during initialization. Also, this pilot study which adopts four-image events may not be aligned with the situation of five-image events; we also only use parametric models and free-form models are not tested yet. In realistic lensed GW detections, we are not able to guarantee an adequate lens reconstruction is available. Therefore, whether the localization framework can be applied with current status of EM lens observations and reconstructions remains unresolved and further studies needed to be done to address the robustness of this localization method.
Observational developments motivate applications of this localization framework beyond the six HSTFF clusters investigated here. Euclid’s first strong-lensing cluster catalog demonstrates the potential of wide-field imaging to expand the sample of candidate cluster lenses ([9]). We test the localization method on six HSTFF clusters, while the localization method in general requires a complete all-sky catalog of cluster lenses. Future observational missions such as Euclid may hence fulfill the data required for this localization. More well-observed clusters may also improve the quality of lens reconstruction, which can be beneficial to the localization method.
With the successful localization of GW events lensed by galaxy clusters, and current observations of such clusters, existing and future GW events can be searched retrospectively to investigate the possibility of lensing by galaxy clusters. If such signals are detected, we can precisely constrain their source positions and estimate the source redshift, which can then locate the host galaxy. With further studies such as a full cluster-lensed GW detection analysis, the localization framework may enable detailed studies of galaxy cluster lenses, host galaxies and support multi-messenger studies of such lensed GW events.
Acknowledgments
This work utilizes gravitational lensing models produced by PIs Bradač, Natarajan & Kneib (CATS), Merten & Zitrin, Sharon, Williams, Keeton, Bernstein and Diego, and the GLAFIC group. This lens modeling was partially funded by the HST Frontier Fields program conducted by STScI. STScI is operated by the Association of Universities for Research in Astronomy, Inc. under NASA contract NAS 5-26555. The lens models were obtained from the Mikulski Archive for Space Telescopes (MAST). H.Y.C., L.E.U., S.J., S.L., and O.A.H. acknowledge support by grants from the Research Grants Council of Hong Kong (Project Nos. CUHK 14304622, 14307923, and 14307724), the start-up grant from the Chinese University of Hong Kong, and the Direct Grant for Research from the Research Committee of The Chinese University of Hong Kong. L.E.U. is supported by the Hong Kong PhD Fellowship Scheme (HKPFS) from the Hong Kong Research Grants Council (RGC).
Posterior odds derivation
In order for the GW to be lensed by galaxy cluster in our catalog of electromagnetic data, the GW event must be established as a strong-lensing candidate, the lens that produced the GW must then be identified, and the GW must originate from cluster . The probability that this occurs
where is the hypothesis that cluster lensed the gravitational wave, is the hypothesis that the lens is detected by EM observations, and is the hypothesis that the gravitational wave is lensed. The data denotes all the available information from both EM and GW observations, including the GW strain data and the EM data of all the galaxy clusters in the catalog.6 This scenario should be favored over the alternative, that either (a) the gravitational-wave is not lensed, or (b) it is lensed but its lens system is not detected in the electromagnetic band, or (c) it is lensed and its lens system is detected in the electromagnetic band, but it is not cluster that lensed the gravitational wave, but instead any other cluster in the catalog. The probability of this alternative scenario
where is the hypothesis that the lens is not detected by EM observations, and is the hypothesis that the GW is not lensed, and is the hypothesis that the lens is detected by EM observations and the GW is lensed, but it is not cluster that produced the GW, but instead any other cluster in the catalog:
The posterior odds in favor of cluster having produced the GW is then given by the ratio of these two probabilities,
We then assume that the GW lensing has been confirmed and use the chain rule, so the posterior odds
We consider five-image lensed gravitational-wave signals, which typically require massive cluster lenses [28, 16]. Such massive systems are comparatively rare and are well covered by existing deep and wide surveys at the relevant masses and redshifts [25, 36]. Therefore, we treat as negligible for the events considered here.
We then define the detected-candidate odds
where , and . Therefore, we can approximate full odds with .
We apply Laplace approximation to simplify the estimation of . For the posterior which can be well-approximated with Gaussian likelihood in parameters,
where is the mode and is the local covariance. For a uniform prior of volume , the factor multiplying the maximum likelihood is . We set prior volumes for each sampling runs on different cluster lens to be the same. Therefore,
which reduces to the main-text expression for equal cluster priors.
Calculation of Deflection Angle and Uncertainty from the Numerical Models
The values of and depend on the source redshift and the Hubble constant [24]:
where is the angular diameter distance, and and are the matter and dark energy density parameters respectively. Flat CDM cosmology is assumed in this study, with and for simplicity.
Also, the data in 2D deflection maps are reduced deflection vector:
Thus, the actual deflection angle can be obtained as
The corresponding image positions can then be obtained using .
The angular diameter distances depend on . For , however, the dependence cancels out in the ratio . Therefore, the deflection maps can be scaled directly without requiring the precise value of .
The uncertainty of is evaluated using a set of simulated lens potential maps from simulations of numerical lens models. To obtain a rough estimate of the uncertainty in the deflection angle, we first obtain the set of deflection angle maps by simple differentiation (). We then compute the standard deviation of the deflection angle at each pixel () across the simulated maps and adopt the upper percentile as the scalar uncertainty value . The uncertainty of lens potential is also obtained by the similar approach with the lens potential map directly.
We then calculate the uncertainty of time delays by error propagation:
As relative time-delay is used in the study, the uncertainty is
where is the uncertainty of time delay of image i and is the uncertainty of time delay of the first image detected. For simplicity, the upper bound of the possible values are used. The resulting maps of the lens models used are shown in Figure 12. We then obtain for each pixel of the numerical lens model and obtain the value of a given image position by linear interpolation.
Figure 12. maps of the 6 galaxy clusters used in this study. The color bar represents the relative uncertainty on time delay measurements for each pixel, and red curves are critical curves for each cluster lens. In the core of the galaxy clusters, the relative uncertainties are higher and can reach up to 12%. Therefore, sources in this study which are behind core region of cluster lens are subjected to higher uncertainties of time delays of images.
Approximation of Likelihood for Localization
For this study, we assume Gaussian noise in the detection of signals, which is a standard approach. By modeling the observed data as , where is the observation, is the noise, and is the true signal, the likelihood of the data given the signal is:
Therefore, set and as GW signals, the likelihood becomes:
and
Note that we assume and
since is much smaller than .
References
- [1]Abac, A. G., et al. 2025, arXiv. https://arxiv.org/abs/2512.16347
- [2]Abbott, B. P., Abbott, R., Abbott, T. D., et al. 2017, Nature (London), 551, 85, doi: 10.1038/nature24471DOI
- [5]Abbott, R., et al. 2021, Astrophys. J., 923, 14, doi: 10.3847/1538-4357/ac23dbDOI
- [6]Abbott, R., Abe, H., Acernese, F., et al. 2025, Phys. Rev. D, 112, 084080, doi: 10.1103/PhysRevD.112.084080DOI
- [7]Birrer, S., Amara, A., & Refregier, A. 2015, The Astrophysical Journal, 813, 102, doi: 10.1088/0004-637X/813/2/102DOI
- [8]Dahle, H., Gladders, M. D., Sharon, K., et al. 2013, The Astrophysical Journal, 773, 146, doi: 10.1088/0004-637X/773/2/146DOI
- [9]Euclid Collaboration, Bergamini, P., Meneghetti, M., Acebron, A., et al. 2026, Astronomy & Astrophysics, 711, A33, doi: 10.1051/0004-6361/202554577
- [10]Evans, M., Adhikari, R. X., Afle, C., et al. 2021, A Horizon Study for Cosmic Explorer: Science, Observatories, and Community, https://arxiv.org/abs/2109.09882
- [11]Foreman-Mackey, D., Hogg, D. W., Lang, D., & Goodman, J. 2013, PASP, 125, 306, doi: 10.1086/670067
- [12]Goodman, J., & Weare, J. 2010, Communications in Applied Mathematics and Computational Science, 5, 65, doi: 10.2140/camcos.2010.5.65DOI
- [13]Hannuksela, O. A., Collett, T. E., Çalışkan, M., & Li, T. G. F. 2020, Mon. Not. Roy. Astron. Soc., 498, 3395, doi: 10.1093/mnras/staa2577DOI
- [14]Hannuksela, O. A., Haris, K., Janquart, J., et al. 2026, The Astrophysical Journal, 1002, 42, doi: 10.3847/1538-4357/ae5816DOI
- [15]Hannuksela, O. A., Haris, K., Ng, K. K. Y., et al. 2019, Astrophys. J. Lett., 874, L2, doi: 10.3847/2041-8213/ab0c0fDOI
- [16]Hilbert, S., White, S. D. M., Hartlap, J., & Schneider, P. 2007, Mon. Not. Roy. Astron. Soc., 382, 121, doi: 10.1111/j.1365-2966.2007.12931.x
- [17]Hilbert, S., White, S. D. M., Hartlap, J., & Schneider, P. 2008, Mon. Not. Roy. Astron. Soc., 386, 1845, doi: 10.1111/j.1365-2966.2008.13190.xDOI
- [18]Janquart, J., Hannuksela, O. A., K., H., & Van Den Broeck, C. 2021, Mon. Not. Roy. Astron. Soc., 506, 5430, doi: 10.1093/mnras/stab1991DOI
- [20]Jauzac, M., Clément, B., Limousin, M., et al. 2014, Monthly Notices of the Royal Astronomical Society, 443, 1549, doi: 10.1093/mnras/stu1355DOI
- [21]Jauzac, M., Richard, J., Limousin, M., et al. 2016, Monthly Notices of the Royal Astronomical Society, 457, 2029–2042, doi: 10.1093/mnras/stw069DOI
- [22]Johnson, T. L., Sharon, K., Bayliss, M. B., et al. 2014, The Astrophysical Journal, 797, 48, doi: 10.1088/0004-637x/797/1/48DOI
- [24]Liddle, A. R. 2015, An introduction to modern cosmology, third edition. edn. (Chichester, England: Wiley)
- [25]Lotz, J. M., Koekemoer, A., Coe, D., et al. 2017, The Astrophysical Journal, 837, 97, doi: 10.3847/1538-4357/837/1/97DOI
- [26]Maggiore, M., Broeck, C. V. D., Bartolo, N., et al. 2020, Journal of Cosmology and Astroparticle Physics, 2020, 050, doi: 10.1088/1475-7516/2020/03/050DOI
- [27]McCully, C., Keeton, C. R., Wong, K. C., & Zabludoff, A. I. 2014, MNRAS, 443, 3631, doi: 10.1093/mnras/stu1316DOI
- [28]Meneghetti, M., Argazzi, R., Pace, F., et al. 2007, A & A, 461, 25, doi: 10.1051/0004-6361:20065722DOI
- [29]Meneghetti, M., Bartelmann, M., & Moscardini, L. 2003, Monthly Notices of the Royal Astronomical Society, 340, 105–114, doi: 10.1046/j.1365-8711.2003.06276.xDOI
- [30]Meneghetti, M., Natarajan, P., Coe, D., et al. 2017, MNRAS, 472, 3177, doi: 10.1093/mnras/stx2064DOI
- [33]Nuttall, L. K., & Sutton, P. J. 2010, PhRvD, 82, 102002, doi: 10.1103/PhysRevD.82.102002DOI
- [34]Phurailatpam, H., More, A., Narola, H., et al. 2024, arXiv e-prints, arXiv:2407.07526, doi: 10.48550/arXiv.2407.07526DOI
- [35]Reitze, D., Adhikari, R. X., Ballmer, S., et al. 2019, Cosmic Explorer: The U.S. Contribution to Gravitational-Wave Astronomy beyond LIGO, https://arxiv.org/abs/1907.04833
- [36]Richard, J., Jauzac, M., Limousin, M., et al. 2014, Monthly Notices of the Royal Astronomical Society, 444, 268, doi: 10.1093/mnras/stu1395DOI
- [37]Schneider, P., Ehlers, J., & Falco, E. E. 1992, Gravitational lenses as astrophysical tools (Berlin, Heidelberg: Springer Berlin Heidelberg), 467–515, doi: 10.1007/978-3-662-03758-4_13DOI
- [38]Şengül, A. Ç., Birrer, S., Natarajan, P., & Dvorkin, C. 2023, Monthly Notices of the Royal Astronomical Society, 526, 2525, doi: 10.1093/mnras/stad2784DOI
- [39]Smith, G. P., Robertson, A., Mahler, G., et al. 2023, Mon. Not. Roy. Astron. Soc., 520, 702, doi: 10.1093/mnras/stad1140
- [40]Storn, R., & Price, K. 1997, Journal of Global Optimization, 11, 341, doi: 10.1023/A:1008202821328DOI
- [41]Takahashi, R., & Nakamura, T. 2003, Astrophys. J., 595, 1039, doi: 10.1086/377430DOI
- [42]Uronen, L. E., Li, T., Janquart, J., et al. 2025, Phil. Trans. Roy. Soc. Lond. A, 383, 20240152, doi: 10.1098/rsta.2024.0152
- [43]Vujeva, L., Ezquiaga, J. M., Lo, R. K. L., & Chan, J. C. L. 2025, Phys. Rev. D, 112, 063044, doi: 10.1103/1hmj-pxjrDOI
- [44]Wambsganss, J. 1998, Living Reviews in Relativity, 1, 12, doi: 10.12942/lrr-1998-12DOI
- [45]Wempe, E., Koopmans, L. V. E., Wierda, A. R. A. C., Hannuksela, O. A., & van den Broeck, C. 2024, Mon. Not. Roy. Astron. Soc., 530, 3368, doi: 10.1093/mnras/stae1023DOI
- [46]Wierda, A. R. A. C., Wempe, E., Hannuksela, O. A., Koopmans, L. é. V. E., & Van Den Broeck, C. 2021, Astrophys. J., 921, 154, doi: 10.3847/1538-4357/ac1bb4DOI






