betaFeedback

WorksPhysics and Astronomy

Gravitational-Wave Microlensing as a Probe of Compact Object Population

Abstract

The abundance of compact objects ranging from stellar to intermediate masses (∼1−105M⊙\sim 1 - 10^{5}M_{\odot}) remains poorly constrained, particularly for populations such as primordial black holes and dark matter subhalos that emit little or no electromagnetic radiation. Gravitational-wave (GW) microlensing provides a complementary way to probe these populations through wave-optics effects imprinted on the GW signals, with implications for astrophysical and cosmological inference. We develop a hierarchical Bayesian framework to infer the population properties of microlenses from an ensemble of microlensed GW events, constraining the slope of the power-law lens mass function (αMLα_{\rm ML}), the lens number density (n0n_0), and the fraction of dark matter in compact lenses (fDMf_{\rm DM}). Using simulated microlensed GW signals drawn from a realistic binary black hole population, we find that ∼100\sim 100 detected microlensed events are able to constrain both αMLα_{\rm ML} and n0n_0, although a degeneracy between the two remains. For a low-mass-dominated lens population (αML=2α_{\rm ML}=2) with n0=106 Mpc−3n_0=10^6~\mathrm{Mpc}^{-3}, a three-year observation at the O4 sensitivity yields ∼130\sim 130 detectable microlensed events and recovers the injected dark-matter fraction, fDM≃1.2×10−4f_{\rm DM}\simeq 1.2 \times 10^{-4}, with a 90% credible interval of 7.3×10−5≲fDM≲1.8×10−47.3 \times 10^{-5} \lesssim f_{\rm DM} \lesssim 1.8 \times 10^{-4}.

Introduction

The abundance and mass distribution of dark compact objects are difficult to determine as they may produce little or no electromagnetic (EM) emission [26]. Such objects may include astrophysical black holes [18, 40, 79] (The LIGO Scientific Collaboration et al. 2026), primordial black holes (PBHs) [27, 86, 62, 30], dark-matter (DM) clumps [93, 46, 21], and more exotic compact objects [42, 35, 85, 105, 20].

Gravitational microlensing provides a powerful way to probe such otherwise elusive populations that could constitute a fraction of DM [11, 53, 106, 81, 80, 8, 102, 19, 51, 82, 84, 61] (P. Mróz et al. 2019). Rather than requiring the lens to be directly observable, microlensing relies on the gravitational imprint of the lens on the source signal. The resulting lensing effect depends on the mass distribution of the lens, the relative geometry between the lens and source, and the evolution of this geometry due to their relative motion [87].

However, conventional EM microlensing has intrinsic limitations because photometric observations primarily constrain the Einstein-radius crossing timescale, which depends on a combination of the lens mass, lens-source geometry, and relative transverse velocity rather than determining these quantities individually [74]. Additional observables, such as microlensing parallax, finite-source effects, and astrometric deflections, can help break these degeneracies [56, 57, 47, 58] (J. H. An et al. 2002), but are available only for a limited subset of events. Moreover, photometric microlensing surveys are primarily sensitive to sufficiently bright and well-monitored sources, predominantly accessible in the nearby Universe, limiting the range of sources and lensing configurations that can be probed [55]. These limitations motivate complementary probes of compact dark objects.

Gravitational-wave (GW) microlensing provides such a complementary probe. Compact lenses can imprint wave-optics effects on GW signals, producing frequency-dependent diffraction and interference features ([43, 78, 95, 24, 68, 39, 67]). For ground-based GW detectors, these effects are particularly relevant for lens masses in the range ∼1\sim1–105 M⊙10^{5}\,M_{\odot}, depending on the GW frequency and lens–source geometry.

In this regime of wave optics, the lensing signature is encoded directly in the waveform rather than solely in the magnification of an EM source, providing information that is complementary to conventional photometric microlensing. Moreover, GWs can be observed from sources at cosmological distances and are unaffected by dust extinction, allowing compact lenses along distant lines of sight to be probed ([3]).

However, individual GW microlensing events provide only limited information, typically constraining only the detector-frame (redshifted) lens mass and leaving a degeneracy between the intrinsic lens mass and lens redshift ([1]). Recovering the underlying lens population, therefore, requires combining information from multiple detections. Hierarchical Bayesian inference offers a framework for doing this by jointly modeling the lens-mass and redshift distributions across an ensemble of events, thereby constraining the hyperparameters that govern the lens population (e.g., [54, 69]).

To date, no compelling evidence for GW lensing has been reported by the LIGO–Virgo–KAGRA (LVK) Collaboration ([63, 5, 7, 1]), although intense scrutiny has been given to the lensed candidate GW231123 ([92, 33, 32, 59, 101] (Q. Hu et al. 2026; M. H.-Y. Cheung et al. 2026)). In the absence of confirmed detections, existing studies have instead used GW observations to place constraints on the abundance of compact DM through the non-observation of lensing signatures ([17, 16]). A complementary question is then how much can be learned once a population of microlensed GW events becomes available? Although intermediate-mass black holes (IMBHs) and globular clusters are unlikely to produce GW microlensing ([100]), if compact objects are linked to the DM, future GW observatories are expected to detect such events in sufficient numbers, making population-level inference possible ([50, 99]).

In this Letter, we investigate the extent to which compact-lens populations can be recovered from GW microlensing observations. Using astrophysically motivated simulations of microlensed GW events, we perform hierarchical Bayesian inference to jointly constrain the lens-mass distribution and number density while accounting for selection effects. We then translate these constraints into measurements of the compact-lens abundance, expressed as their contribution to the DM density.

Throughout this work, we adopt the Planck15 cosmology ([10]) as our fiducial model.

Data Setup

Point-Mass Microlensing

When a propagating GW signal passes near a compact object with a Schwarzschild radius that is comparable to the GW wavelength, the waveform undergoes frequency-dependent amplification, a phenomenon referred to as GW microlensing. The waveform of a microlensed GW (hMLh_{\mathrm{ML}}) in the frequency domain can be obtained by multiplying the amplification factor (FF) by the corresponding unlensed GW waveform (huh_{\mathrm{u}}}) ([95]):

hML(w,y)=F(w,y)hu(w),(1)h_{\mathrm{ML}}(w,y)=F(w,y)h_{\mathrm{u}}(w), \tag*{(1)}

where w≡8πGMlzf/c3w \equiv8\pi GM_{\mathrm{l}}^{z}f/c^{3} is the dimensionless frequency of the waveform. Mlz≡Ml(1+zl)M_{\mathrm{l}}^{z}\equiv M_{\mathrm{l}}(1+z_{\mathrm{l}}) denotes the redshifted lens mass, with MlM_{\mathrm{l}} and zlz_{\mathrm{l}} the lens-frame mass and redshift, respectively, and yy is the dimensionless impact parameter. These parameters fully specify an isolated point-mass (PM) lens. The corresponding amplification factor is written analytically as follows ([78, 95]):

F(w,y)=exp⁡{πw4+iw2[ln⁡(w2)−2ϕm(y)]}×Γ(1−i2w)1F1(i2w;1;i2wy2),(2)F(w,y)=\exp\left\{\frac{\pi w}{4}+i\frac{w}{2}\left[\ln\left(\frac{w}{2}\right)-2\phi_{\mathrm{m}}(y)\right]\right\} \times\Gamma\left(1-\frac{i}{2}w\right){}_{1}F_{1}\left(\frac{i}{2}w;1;\frac{i}{2}wy^{2}\right), \tag*{(2)}

where 1F1{}_{1}F_{1} is the confluent hypergeometric function and ϕm(y)=(xm−y)2/2−ln⁡xm\phi_{\mathrm{m}}(y)=(x_{\mathrm{m}}-y)^{2}/2-\ln x_{\mathrm{m}}, where xm=(y+y2+4)/2x_{\mathrm{m}}=(y+\sqrt{y^{2}+4})/2.

Simulated Microlensed GW Events

To simulate microlensed GW signals using Eqs. (1) and (2), we consider only binary black holes (BBHs), whose detection horizon for current ground-based detectors extends to the largest distances among compact-binary sources (The LIGO Scientific Collaboration et al. 2026). Assuming a 3-year observing period (Tobs=3 yrsT_{\mathrm{obs}}=3\,\mathrm{yrs}), we draw source parameters from the GWTC-3 population models [6], fixing the hyperparameters to their posterior median values. Specifically, we adopt the PowerLaw+Peak model for the primary mass distribution and a Madau–Dickinson-like redshift distribution [72] over zs∈[0.001,2.3]z_s \in[0.001, 2.3], yielding a total of NUL≃7.9×105N_{\mathrm{UL}} \simeq7.9 \times10^{5} BBHs (see Appendices § 7.§ 1 and § 7.§ 2). GW signals are then generated using the IMRPhenomXPHM-SpinTaylor waveform model [83, 41].

The resulting synthetic BBH population remains broadly consistent with the population properties inferred from the most recent GWTC-5 catalog, which retains the main features of the BBH mass distribution identified in earlier catalogs [6, 2] (The LIGO Scientific Collaboration et al. 2026). We further find that our results are largely insensitive to the choice between the GWTC-3 and GWTC-5 population models. Details of the adopted models and hyperparameters, together with the comparison between the GWTC-3 and GWTC-5 populations, are given in Appendix § 7.§ 3.

We draw microlens parameters based on a population described by two hyperparameters: (1) n0n_0, the constant comoving number density of microlenses, and (2) αML\alpha_{\mathrm{ML}}, the mass-function slope. The intrinsic lens mass MlM_l is distributed as:

p(Ml∣αML)∝Ml−αML,(3)p(M_l \mid\alpha_{\mathrm{ML}}) \propto M_l^{-\alpha_{\mathrm{ML}}}, \tag*{(3)}

motivated by stellar lenses following the initial mass function [31], of PBHs, for which a power-law mass function arises naturally from scale-invariant density fluctuations or the collapse of cosmic string loops [28, 66], and by compact objects observed by the LVK network (The LIGO Scientific Collaboration et al. 2026). We consider lens masses up to Ml=105 M⊙M_l = 10^{5}\,M_{\odot}, above which lensing effects enter the geometrical-optics regime across the LVK frequency band.

Motivated by J. Gais et al. [54], which showed that an injected IMBH number density of n0=106 Mpc−3n_0 = 10^{6}\,\mathrm{Mpc}^{-3} yields a well-constrained posterior on the number density, we adopt this comoving number density. For the mass-function slope, we inject αML=[1,2]\alpha_{\mathrm{ML}} = [1, 2] to explore two representative population scenarios.

The choice αML=1\alpha_{\mathrm{ML}} = 1 corresponds to a log-uniform mass distribution, providing an agnostic baseline model (see, for example, B. Carr et al. [29] for PBH formation via spectator-field perturbations during an early matter-dominated era). In contrast, αML=2\alpha_{\mathrm{ML}} = 2 describes a steeper distribution in which low-mass lenses dominate the population more strongly, representative of the general class of power-law PBH mass functions [28, 66]. Comparing these two cases allows us to assess how constraints on the underlying lens population depend on αML\alpha_{\mathrm{ML}}.

Since GW microlensing measurements constrain the detector-frame lens mass, MlzM_l^z, we draw MlzM_l^z directly rather than MlM_l. The microlensing optical depth for a BBH at source redshift zsz_s is derived in Appendix § 8. Combining this optical depth with the assumed MlM_l distribution in Eq. (3), the probability that a BBH at zsz_s is microlensed by a lens with MlzM_l^z can be written as:

p(Mlz∣αML,zs)p(M_l^z \mid\alpha_{\mathrm{ML}}, z_s)
∝∫0zsdzl(Mlz)1−αML(1+zl)αML−2DlsDsDldVcdzl,(4)\propto\int_0^{z_s} \mathrm{d}z_l (M_l^z)^{1-\alpha_{\mathrm{ML}}}(1+z_l)^{\alpha_{\mathrm{ML}}-2}\frac{D_{ls}}{D_sD_l}\frac{\mathrm{d}V_c}{\mathrm{d}z_l}, \tag*{(4)}

where VcV_c is the comoving volume, and DlD_l, DsD_s, and DlsD_{ls} are the angular diameter distances from the observer to the lens, from the observer to the source, and from the lens to the source, respectively.

Using Eq. (4), we sample MlzM_l^z within the range [1,105]M⊙[1,10^{5}]M_{\odot}. Note that the source and lens parameters are linked through zsz_s, which determines the lens redshift and MlzM_l^z distribution. We then sample yy from the nearest-effective-lens distribution [54], which incorporates the geometric factor (p(y)∝y)(p(y) \propto y) arising from the differential lensing area for a spatially uniform source population in the source plane. Restricting yy to the range [0.01,3.0][0.01,3.0]2, we adopt

p(y∣n0,αML,zs)=2πΣ(n0,αML,zs)ye−πΣ(n0,αML,zs)y2,(5)\begin{aligned} p(y \mid n_0, \alpha_{\mathrm{ML}}, z_s) \\ &= 2\pi\Sigma(n_0,\alpha_{\mathrm{ML}},z_s)y e^{-\pi\Sigma(n_0,\alpha_{\mathrm{ML}},z_s)y^2}, \tag*{(5)} \end{aligned}

where Σ\Sigma is the effective lens number density within the comoving volume out to zsz_s, projected onto the sky, which is given by

Σ(n0,αML,zs)=NL(zs)4π⟨θE2⟩αML,zs,(6)\Sigma(n_0,\alpha_{\mathrm{ML}},z_s) = \frac{N_L(z_s)}{4\pi}\left\langle\theta_E^2 \right\rangle_{\alpha_{\mathrm{ML}},z_s}, \tag*{(6)}

where NL(zs)≡∫0zsn0dVc(zl)N_L(z_s) \equiv\int_0^{z_s} n_0 \mathrm{d}V_c(z_l) is the total number of lenses within the comoving volume out to zsz_s, and ⟨θE2⟩αML,zs\left\langle\theta_E^2 \right\rangle_{\alpha_{\mathrm{ML}},z_s} is the mean squared angular Einstein radius averaged over the lens population according to their mass and redshift distributions.

For the αML=1\alpha_{\mathrm{ML}} = 1 and 22 scenarios, the expected numbers of microlensing events given by Eq. (23) are NML=74,933N_{\mathrm{ML}} = 74,933 and 16,27516,275, respectively. Each microlensed event is then assigned a source parameter set and a corresponding set of lens parameters drawn from the distributions above.

Microlensing Detectability

Although the optical depth determines the expected number of microlensed GW events, only a subset of these events produces detectable wave-optics signatures in GW observations. The detectability of GW microlensing depends not only on the network SNR, ρnet\rho_{\rm net}, but also on the waveform mismatch ϵ\epsilon between microlensed and unlensed signals, which quantifies the degree of waveform distortion induced by wave-optics effects.

Following J. M. Ezquiaga et al. [48], we adopt a fiducial distinguishability criterion based on the statistic Δχ2\Delta\chi^2, which is interpreted as a measure of how distinguishable a microlensed waveform is from its unlensed counterpart. In the limit of small waveform mismatches, Δχ2\Delta\chi^2 can be approximated as:

Δχ2≃2ρnet2ϵnet,(7)\Delta\chi^2 \simeq2\rho_{\rm net}^2\epsilon_{\rm net}, \tag*{(7)}

where ϵnet\epsilon_{\rm net} denotes the network-averaged mismatch computed from the individual detector signals. We impose the conventional GW detection threshold, ρnet>8\rho_{\rm net}>8, and require an additional distinguishability criterion to identify a detected signal as microlensed. Using Eq. (7), a fiducial mismatch threshold of ϵnet>0.03\epsilon_{\rm net}>0.03 at the detection threshold ρnet=8\rho_{\rm net}=8 corresponds to Δχ2>3.84\Delta\chi^2>3.84 (or log⁡Δχ2≳1.35\log\Delta\chi^2\gtrsim1.35). Signals satisfying both criteria are therefore considered identifiable as microlensed.

Figure 1 shows the resulting Δχ2\Delta\chi^2 values between the simulated microlensed waveforms and their corresponding unlensed counterparts for microlensed events that satisfy the SNR detection threshold ρnet>8\rho_{\rm net}>8, obtained over three years of simulated observations, as a function of MlM_l and zl/zsz_l/z_s for both the αML=1\alpha_{\rm ML}=1 and αML=2\alpha_{\rm ML}=2 populations. In both cases, Δχ2\Delta\chi^2 shows a decreasing trend with increasing yy, reflecting the progressive suppression of wave-optics distortions at larger yy. We estimate ymax⁡y_{\max} used in Eq. (21) based on the adopted Δχ2\Delta\chi^2 threshold.

Figure 1. Distributions of Δχ2\Delta\chi^2 between microlensed and intrinsic unlensed waveforms, for signals with ρnet>8\rho_{\rm net}>8, shown as a function of the intrinsic lens mass (MlM_l) and the ratio of lens to source redshifts (zl/zsz_l/z_s). The color scheme indicates the dimensionless impact parameter (yy).

In addition, a positive correlation between MlM_l and Δχ2\Delta\chi^2 is evident, as expected from the stronger and more distinguishable waveform distortions produced by more massive lenses. Since the αML=1\alpha_{\mathrm{ML}}=1 population contains a larger fraction of high MlM_l lenses than the steeper αML=2\alpha_{\mathrm{ML}}=2 distribution, a correspondingly larger fraction of simulated events satisfy the final distinguishability criterion in the αML=1\alpha_{\mathrm{ML}}=1 case. The minimum lens mass capable of producing detectable microlensing signatures is O(2) M⊙\mathcal{O}(2)\ M_\odot for αML=1\alpha_{\mathrm{ML}}=1, whereas detectable signals in the αML=2\alpha_{\mathrm{ML}}=2 case can extend down to O(1) M⊙\mathcal{O}(1)\ M_\odot. Moreover, no clear correlation is observed between Δχ2\Delta\chi^2 and zl/zsz_l/z_s. While the lensing probability is maximized for lenses located roughly midway between the source and observer due to geometric effects, the strength of observable wave-optics distortions depends more strongly on MlzM_l^z and yy.

It is important to note that some simulated microlensed signals may satisfy the distinguishability criterion while still failing the network SNR requirement (ρnet>8\rho_{\mathrm{net}}>8). Such events would not be initially detectable. We therefore retain only events satisfying both detectability and distinguishability criteria and classify them as detectable microlensed GW signals.

Population Inference

Selection of Microlensed Events

From the resulting detectable populations, containing 1727 and 291 events for αML=1\alpha_{\mathrm{ML}}=1 and 2, respectively, we randomly select 200 events from each population for computationally tractable parameter estimation. Random sampling ensures that no additional selection is imposed beyond the detectability criterion.

We then inject these events into simulated data from the LIGO Livingston and Hanford detectors (E. Capote et al. [25]) and Virgo (F. Acernese et al. [9]), assuming the expected O4 sensitivities and Gaussian noise realizations (B. P. Abbott et al. [4]).

To estimate the parameters of microlensed waveforms, we adopt Gravelamps (M. Wright & M. Hendry [103]), which is based on bilby (G. Ashton et al. [13]) and dynesty (J. S. Speagle [94]). The pipeline yields posterior distributions for MlzM_l^z and yy, together with the Bayesian evidence for the microlensing hypothesis, BML\mathcal{B}_{\mathrm{ML}}. Performing a parameter estimation for the same signal under the unlensed hypothesis additionally yields the corresponding evidence BU\mathcal{B}_{\mathrm{U}}, allowing the Bayes factor between microlensing and unlensed hypotheses, BUML\mathcal{B}_{\mathrm{U}}^{\mathrm{ML}}, to be evaluated. A microlensed signal is expected to favor the lensing hypothesis with BUML>1\mathcal{B}_{\mathrm{U}}^{\mathrm{ML}}>1. In realistic searches, confident microlensing detections would additionally require a robust false-positive probability (FPP) assessment based on a large population of unlensed background events. For simplicity, however, we adopt a pragmatic criterion and regard microlensed events with

BUML>1\mathcal{B}_{\mathrm{U}}^{\mathrm{ML}}>1 and well-constrained posteriors for MlzM_l^z and yy as confidently identified microlensing events, yielding totals of 158 and 132 events for the αML=1\alpha_{\mathrm{ML}}=1 and αML=2\alpha_{\mathrm{ML}}=2 cases, respectively. Appendix § 9 compares the lens parameters of identified microlensed events with the underlying population models, illustrating how detectable lens properties depend on the population parameters.

Hierarchical Bayesian Framework

Given the set of detected microlensing events (dd), the posterior of the hyperparameters (λ\lambda) is obtained through hierarchical Bayesian inference:

p(λ∣d)∝p(λ)p(NML,det∣λ)×∏iNML,detp(di∣θi,Ml,iz,yi)ppop(θi,Ml,iz,yi∣λ)β(λ)NML,det,(8)\begin{aligned} p(\lambda\mid d) \propto p(\lambda)p(N_{\mathrm{ML,det}}\mid\lambda) \\ &\times\frac{\prod_i^{N_{\mathrm{ML,det}}}p(d_i\mid\boldsymbol{\theta}_i,M_{l,i}^z,y_i)p_{\mathrm{pop}}(\boldsymbol{\theta}_i,M_{l,i}^z,y_i\mid\lambda)}{\beta(\lambda)^{N_{\mathrm{ML,det}}},} \tag*{(8)} \end{aligned}

where NML,detN_{\mathrm{ML,det}} is the number of detected microlensed events and β(λ)\beta(\lambda) indicates the detection efficiency for a given hyperparameter set, corresponding to a selection-bias correction (I. Mandel et al. [73]; W. M. Farr [52]). The hyperparameter set consists of source-population, cosmological, and lens-population parameters, denoted by λ=(λs,λc,λl)\lambda=(\lambda_s,\lambda_c,\lambda_l). Here, λs\lambda_s describes the BBH population as introduced in Eqs. (16)–(20), while λc\lambda_c denotes the cosmological parameters, such as the Hubble constant H0H_0 and the density parameters Ωm\Omega_m and ΩΛ\Omega_\Lambda. In this work, the lens-population parameters are λl=(αML,n0)\lambda_l=(\alpha_{\mathrm{ML}},n_0).

The first term in Eq. (8) is the prior on the hyperparameters. In this work, we fix λs\lambda_s and λc\lambda_c to equal their injected values to focus exclusively on the inference of λl\lambda_l. We adopt a broad uniform prior for the lens mass-function slope, αML∈U[−4,12]\alpha_{\mathrm{ML}}\in\mathrm{U}[-4,12]. For the lens number density, we adopt log⁡10n0∈U[0,10]\log_{10}n_0\in\mathrm{U}[0,10], following the prior range adopted in J. Gais et al. [54].

The second term describes the probability of detecting NML,detN_{\mathrm{ML,det}} microlensing events for a given set of λ\lambda, assuming a Poisson distribution with an expected number given by:

⟨NML,det(λ)⟩=NML(λ)×∫ppop(θs,Mlz,y∣λ)Pdet(θs,Mlz,y) dθsdMlzdy(9)\begin{aligned} \langle N_{\mathrm{ML,det}}(\lambda)\rangle=N_{\mathrm{ML}}(\lambda) \\ &\times\int p_{\mathrm{pop}}(\boldsymbol{\theta}_s,M_l^z,y\mid\lambda)P_{\mathrm{det}}(\boldsymbol{\theta}_s,M_l^z,y)\,\mathrm{d}\boldsymbol{\theta}_s\mathrm{d}M_l^z\mathrm{d}y \tag*{(9)} \end{aligned}

where NML(λ)N_{\mathrm{ML}}(\lambda) is the total number of microlensed events defined in Eq. (23), while ppop(θs,Mlz,y∣λ)p_{\mathrm{pop}}(\boldsymbol{\theta}_s,M_l^z,y\mid\lambda) encodes the underlying astrophysical population model for the source and lens parameters. In this work, the lens population is described by MlzM_l^z distribution given in Eq. (4), together with yy distribution given in Eq. (5). The term PdetP_{\mathrm{det}} denotes the probability that a microlensed signal with the given source and lens parameters satisfies the adopted detection and distinguishability criteria.

The third term in Eq. (8) represents the joint likelihood of all detected microlensed events, constructed as a product of individual event contributions, assuming that each event is statistically independent given λ\lambda. This assumption allows the joint likelihood to be factorized into a product of event-level terms. In the numerator, p(di∣θs,i,Ml,iz,yi)p(d_i \mid\theta_{s,i}, M_{l,i}^{z}, y_i) denotes the event-level likelihood of each set of observed microlensed GW data, evaluated using posterior samples obtained from parameter estimation. In the denominator, β(λ)\beta(\lambda) denotes the detection efficiency, defined as the probability that a source drawn from the underlying population is identified as a microlensed event satisfying both the adopted SNR detection threshold and distinguishability criterion. It therefore accounts for the selection effects associated with both requirements. The detection efficiency for microlensed signals can be written as:

β(λ)=⟨NML,det(λ)⟩NML=∫ppop(θs,Mlz,y∣λ)Pdet(θs,Mlz,y).(10)\beta(\lambda) = \frac{\langle N_{\mathrm{ML,det}}(\lambda)\rangle}{N_{\mathrm{ML}} = \int p_{\mathrm{pop}}(\theta_s, M_l^z, y \mid\lambda) P_{\mathrm{det}}(\theta_s, M_l^z, y).} \tag*{(10)}

In practice, Eq. (10) can be estimated using importance sampling over a large set of simulated microlensed injections:

β(λ)≃1Ninj∑iNinj,detppop(θs,i,Ml,iz,yi∣λ)π(θs,i,Ml,iz,yi),(11)\beta(\lambda) \simeq\frac{1}{N_{\mathrm{inj}}} \sum_{i}^{N_{\mathrm{inj,det}}} \frac{p_{\mathrm{pop}}(\theta_{s,i}, M_{l,i}^z, y_i \mid\lambda)}{\pi(\theta_{s,i}, M_{l,i}^z, y_i)}, \tag*{(11)}

where NinjN_{\mathrm{inj}} and Ninj,detN_{\mathrm{inj,det}} denote the total number of injected and successfully detected microlensed signals, respectively, and π\pi is the joint prior distribution used in the injection campaign.

Results

Inference of the Compact-Lens Distribution

Using Eq. (8), we infer the joint posterior distributions for αML\alpha_{\mathrm{ML}} and n0n_0. To examine how the constraints on these population parameters improve as the mock microlensed-event sample accumulates, we consider cumulative samples containing NML,det=50N_{\mathrm{ML,det}}=50, 100, and 132 events for both the αML=1\alpha_{\mathrm{ML}}=1 and 2 scenarios. The αML=2\alpha_{\mathrm{ML}}=2 scenario has a lower detectable-event rate than the αML=1\alpha_{\mathrm{ML}}=1 scenario, and therefore requires a longer observing time to accumulate the same number of events. We compare the two scenarios at fixed NML,detN_{\mathrm{ML,det}} to control for differences in sample size and isolate the impact of the underlying lens-mass distribution in Eq. (3) on the population constraints.

As shown in Figure 2, the widths of the posteriors for both hyperparameters become progressively narrower as the number of detections increases, while the posterior peaks converge toward their respective true values. Notably, n0n_0 is less well constrained than αML\alpha_{\mathrm{ML}}. This difference arises from its strong degeneracy with yy, which more directly governs the microlensing probability and magnification than MlzM_l^z.

Figure 2. Corner plots showing the posteriors of αML\alpha_{\mathrm{ML}} and log⁡10n0\log_{10} n_0 inferred from the NML,det=50N_{\mathrm{ML,det}}=50, 100, and 132 cases for the two fiducial scenarios αML=1\alpha_{\mathrm{ML}}=1 (left) and αML=2\alpha_{\mathrm{ML}}=2 (right). The black dotted lines denote the corresponding true values.

Furthermore, across both the αML=1\alpha_{\mathrm{ML}} = 1 and αML=2\alpha_{\mathrm{ML}} = 2 scenarios, the posteriors show a positive correlation between αML\alpha_{\mathrm{ML}} and n0n_0, regardless of NML,detN_{\mathrm{ML,det}}. Physically, a larger αML\alpha_{\mathrm{ML}} shifts the lens population toward lower-mass compact objects, which generally produce weaker microlensing signatures and are therefore less likely to be detected. A higher n0n_0 is then required to reproduce the same number of detections, giving rise to the observed αML\alpha_{\mathrm{ML}}–n0n_0 degeneracy. Increasing the catalog size tightens the constraints on both parameters, but the correlation persists, since it arises from the structure of the detection rate rather than from limited statistics alone.

In the αML=2\alpha_{\mathrm{ML}} = 2 scenario, the intrinsic lens-mass distribution is strongly weighted towards the lower-mass end. Even for NML,det=50N_{\mathrm{ML,det}} = 50, the detected population is therefore dominated by low-mass lenses, and the absence of high-mass lenses provides a strong constraint on αML\alpha_{\mathrm{ML}}. Because of the αML\alpha_{\mathrm{ML}}–n0n_0 correlation, this in turn tightens the constraint on n0n_0.

For each dataset simulated under αML=1\alpha_{\mathrm{ML}} = 1 or αML=2\alpha_{\mathrm{ML}} = 2, we calculate the Bayes factor between the two hypotheses by analyzing them separately with αML\alpha_{\mathrm{ML}} fixed to each value. We denote the log Bayes factor in favor of the αML=1\alpha_{\mathrm{ML}} = 1 and αML=2\alpha_{\mathrm{ML}} = 2 hypotheses by ln⁡B21\ln B^{1}_{2} and ln⁡B12\ln B^{2}_{1}, respectively. The resulting Bayes factors are reported in Table 1. As expected, the Bayes factor in favor of the injected hypothesis increases with NML,detN_{\mathrm{ML,det}}.

Injected αML\alpha_{\mathrm{ML}} (Bayes factor)50100132
1 (ln⁡B12\ln B^{2}_{1})69127161
2 (ln⁡B21\ln B^{1}_{2})76159226

Table 1. Log Bayes factors in favor of the injected αML\alpha_{\mathrm{ML}} hypothesis for different NML,detN_{\mathrm{ML,det}}. Each entry is defined as ln⁡Baltinj=ln⁡Zinj−ln⁡Zalt\ln B^{\mathrm{inj}}_{\mathrm{alt}} = \ln Z_{\mathrm{inj}}-\ln Z_{\mathrm{alt}}, where ZinjZ_{\mathrm{inj}} and ZaltZ_{\mathrm{alt}} are the Bayesian evidences for the injected and alternative fixed-αML\alpha_{\mathrm{ML}} hypotheses, respectively.

Constraining the Dark Matter Fraction in Microlenses

According to E. L. Turner et al. [98], the comoving number density n0n_0 appearing in Eq. (21) can be written as:

n0=ρl⟨Ml⟩=ρcΩl⟨Ml⟩=3H02Ωl8πG⟨Ml⟩,(12)\begin{aligned} n_0 &= \frac{\rho_l}{\langle M_l\rangle}\\ &= \frac{\rho_c\Omega_l}{\langle M_l\rangle}\\ &= \frac{3H_0^2\Omega_l}{8\pi G\langle M_l\rangle}, \tag*{(12)} \end{aligned}

where ρl\rho_l and ρc\rho_c denote the present-day compact-lens density and the critical density of the Universe, respectively, and Ωl≡ρl/ρc\Omega_l \equiv\rho_l/\rho_c is the corresponding density parameter. ⟨Ml⟩\langle M_l\rangle denotes the expectation value of the intrinsic lens mass, analytically derived from the assumed mass distribution p(Ml∣αML)p(M_l \mid\alpha_{\mathrm{ML}}) truncated to a specific mass range. For a given αML\alpha_{\mathrm{ML}}, the theoretical ⟨Ml⟩\langle M_l\rangle is given by:

⟨Ml⟩={(1−αML2−αML)Mmax⁡2−αML−Mmin⁡2−αMLMmax⁡1−αML−Mmin⁡1−αML,αML≠1,2,Mmax⁡−Mmin⁡ln⁡(Mmax⁡/Mmin⁡),αML=1,(Mmin⁡Mmax⁡Mmax⁡−Mmin⁡)ln⁡(Mmax⁡Mmin⁡),αML=2.(13)\langle M_l\rangle= \begin{cases} \left(\frac{1-\alpha_{\mathrm{ML}}}{2-\alpha_{\mathrm{ML}}}\right)\frac{M_{\max}^{2-\alpha_{\mathrm{ML}}}-M_{\min}^{2-\alpha_{\mathrm{ML}}}}{M_{\max}^{1-\alpha_{\mathrm{ML}}}-M_{\min}^{1-\alpha_{\mathrm{ML}}}}, & \alpha_{\mathrm{ML}}\ne1,2,\\ \frac{M_{\max}-M_{\min}}{\ln(M_{\max}/M_{\min})}, & \alpha_{\mathrm{ML}}=1,\\ \left(\frac{M_{\min}M_{\max}}{M_{\max}-M_{\min}}\right)\ln\left(\frac{M_{\max}}{M_{\min}}\right), & \alpha_{\mathrm{ML}}=2. \end{cases} \tag*{(13)}

where Mmin⁡M_{\min} and Mmax⁡M_{\max} are the lower and upper bounds of the intrinsic lens-mass distribution, respectively. Since we consider redshifted lens masses Mlz∈[1,105]M⊙M_l^{z}\in[1,10^5]M_\odot and the maximum source redshift is zs,max⁡=2.3z_{s,\max}=2.3, we adopt the corresponding global bounds on the intrinsic lens mass, Mmin⁡=1/(1+zs,max⁡)M⊙≃0.303M⊙M_{\min}=1/(1+z_{s,\max})M_\odot\simeq0.303M_\odot and Mmax⁡=105M⊙M_{\max}=10^5M_\odot. If the compact lenses are interpreted as DM objects, such as PBHs or DM clumps, their density parameter can be expressed as Ωl≡fDMΩDM\Omega_l\equiv f_{\mathrm{DM}}\Omega_{\mathrm{DM}}, where fDMf_{\mathrm{DM}} denotes the fraction of total DM density contributed by compact lenses within the mass range [Mmin⁡,Mmax⁡][M_{\min},M_{\max}]. Note that fDM=1f_{\mathrm{DM}}=1 corresponds to the case in which all DM is contained in compact lenses within this mass range.

Consequently, the posterior constraints on n0n_0 and αML\alpha_{\mathrm{ML}} obtained from the hierarchical Bayesian analysis in § 4.1 can be translated into constraints on fDMf_{\mathrm{DM}}, assuming a fiducial DM density parameter ΩDM=0.26\Omega_{\mathrm{DM}}=0.26 ([10]). Under these assumptions, the DM fraction associated with compact lenses can be written as:

fDM=8πGn0⟨Ml⟩3H02ΩDM.(14)f_{\mathrm{DM}}=\frac{8\pi G n_0\langle M_l\rangle}{3H_0^2\Omega_{\mathrm{DM}}}. \tag*{(14)}

Unlike monochromatic-mass constraints, our analysis accounts for an extended lens-mass range and its power-law spectrum, allowing the inferred fDMf_{\mathrm{DM}} to reflect both the lens abundance and its mass distribution.

Figure 3 shows the posteriors of fDMf_{\mathrm{DM}}, derived from the posterior samples of n0n_0 and αML\alpha_{\mathrm{ML}} for each NML,detN_{\mathrm{ML,det}} case. In the αML=1\alpha_{\mathrm{ML}}=1 scenario, the relatively high mean lens mass leads to a comparatively large true DM fraction of fDM≃0.23f_{\mathrm{DM}}\simeq0.23. For NML,det=50N_{\mathrm{ML,det}}=50, the posterior is weakly constrained, with substantial support over a broad range and its peak noticeably offset from the true value. As shown similarly in Figure 2, the posterior becomes progressively narrower and more closely centered on the true value as NML,detN_{\mathrm{ML,det}} increases. In particular, the lower bound on fDMf_{\rm DM} becomes progressively more stringent with increasing NML,detN_{\rm ML,det}.

Figure 3. Posteriors of fDMf_{\rm DM} inferred for the αML=1\alpha_{\rm ML}=1 (left) and αML=2\alpha_{\rm ML}=2 (right) scenarios using catalogs with NML,det=50N_{\rm ML,det}=50, 100, and 132 detected microlensing events. The hatched gray region marks the unphysical range with fDM>1f_{\rm DM}>1. The black dotted lines denote the corresponding true values.

In contrast, for the αML=2\alpha_{\rm ML}=2 scenario, the substantially lower mean lens mass yields a much smaller true DM fraction of fDM≃1.2×10−4f_{\rm DM}\simeq1.2\times10^{-4}. The posterior is already peaked near the true value for NML,det=50N_{\rm ML,det}=50, reflecting the relatively strong constraint on αML\alpha_{\rm ML}.

It is worth noting that while the physical boundary strictly requires fDM≤1f_{\rm DM}\leq1, our posterior samples for the αML=1\alpha_{\rm ML}=1 scenario extend into the unphysical regime (fDM>1f_{\rm DM}>1). This issue arises because the analysis was performed using broad hyperparameter priors that do not explicitly impose the physical constraint on fDMf_{\rm DM}. In our simplified population model, the physically allowed region in the (n0,αML)(n_0,\alpha_{\rm ML}) parameter space can be derived analytically from Eq. (14). In more realistic compact-object population models, however, the mapping between hyperparameters and the implied value of fDMf_{\rm DM} is expected to become substantially more complex. For this reason, adopting broad priors at the hyperparameter level and subsequently excluding posterior samples with fDM>1f_{\rm DM}>1 provides a more flexible and model-independent approach than imposing hard physical boundaries a priori.

In addition to fDMf_{\rm DM}, which quantifies the total contribution of compact lenses across the given lens-mass range, the differential dark-matter fraction as a function of lens mass, dfDM/dMldf_{\rm DM}/dM_l, provides a mass-resolved view of this contribution. In our power-law lens mass distribution model, this quantity can be derived analytically.

In particular,

dfDMdlog⁡10Ml∝Ml2−αML,(15)\frac{df_{\rm DM}}{d\log_{10} M_l}\propto M_l^{2-\alpha_{\rm ML}}, \tag*{(15)}

where αML=2\alpha_{\rm ML}=2 corresponds to an equal contribution to the DM density from each logarithmic mass interval. For αML<2\alpha_{\rm ML}<2, the DM contribution is weighted towards higher lens masses, whereas αML>2\alpha_{\rm ML}>2 shifts the contribution towards lower masses. Figure 7 in Appendix § 10 illustrates this dependence.

Figure 7. The differential fraction of dark matter, dfDM/dlog⁡10Ml\mathrm{d}f_{\mathrm{DM}}/\mathrm{d}\log_{10} M_l, for the αML=1\alpha_{\mathrm{ML}}=1 (left) and αML=2\alpha_{\mathrm{ML}}=2 (right) scenarios, from catalogs with NML,det=50N_{\mathrm{ML,det}}=50, 100, and 132 detected events. The black dashed lines denote the corresponding true spectra, while the colored shaded regions indicate the 22-σ\sigma credible regions after imposing fDM≤1f_{\mathrm{DM}}\leq1.

Summary and Discussion

We have investigated the feasibility of constraining compact-object populations through hierarchical Bayesian inference applied to a catalog of identified GW microlensing events, while consistently accounting for selection effects. Using simulated microlensed BBH signals drawn from two representative power-law lens-mass distributions, with αML=1\alpha_{\rm ML}=1 and αML=2\alpha_{\rm ML}=2, we demonstrate that the accumulation of microlensed GW detections can place meaningful constraints on both the mass-function slope αML\alpha_{\rm ML} and the comoving lens number density n0n_0, while simultaneously breaking the degeneracy between these hyperparameters.

We further translate the inferred population parameters into constraints on the compact-DM fraction, fDMf_{\rm DM}, and investigate its dependence on both microlens mass and redshift. For a fixed number of detections, steeper microlens mass functions yield tighter constraints on fDMf_{\rm DM}, implying that compact-object populations dominated by lower-mass lenses are intrinsically more informative for measuring it. Accordingly, both the differential compact-DM fraction and the redshift-binned compact-DM fraction are recovered with smaller uncertainties for the αML=2\alpha_{\mathrm{ML}} = 2 scenario than for the αML=1\alpha_{\mathrm{ML}} = 1 scenario.

An additional advantage of our hierarchical framework is that it does not require explicit hard priors derived from the physical bounds on fDMf_{\mathrm{DM}}. Constructing such priors becomes increasingly difficult as the relationship between the population parameters and the physically allowed fDMf_{\mathrm{DM}} grows more complicated in realistic population models. Instead, the physically allowed and disallowed regions can be identified directly from the inferred fDMf_{\mathrm{DM}} posterior, making the framework both flexible and readily extensible to more general compact-object population models. Moreover, the inferred fDMf_{\mathrm{DM}} as a function of lens mass and redshift can reveal how the abundance of compact objects in different mass ranges evolves across cosmic time. Therefore, if observations can probe sufficiently distant lenses, the redshift evolution of fDM(zl)f_{\mathrm{DM}}(z_l) beyond the peak of the cosmic star-formation rate can provide a test of whether the inferred compact-object population is consistent with conventional stellar remnants alone or requires additional populations such as PBHs.

Although our framework provides a robust foundation for population inference, several simplifying assumptions adopted in this work motivate important directions for future refinement. We employed a simplified population configuration in which the source-population hyperparameters are fixed, thereby focusing exclusively on the lens population. Incorporating uncertainties in the source population would weaken the constraints on the population properties. However, such an extension would also enable a direct investigation of potential degeneracies between the source and lens populations, particularly in scenarios where both originate from dark compact objects. This possibility becomes especially relevant if a fraction of the BBH population detected by LVK is itself primordial in origin [36, 70], since the source and lens populations could then share common formation channels and population characteristics.

In addition, we adopted a simple power-law model for the intrinsic lens-mass distribution. More physically motivated population distributions should be considered for PBH and other DM-related object populations when their theoretical predictions are sufficiently well established (see e.g., A. M. Green 2016; M. Fairbairn et al. 2023; E. Bagui et al. 2025; L. Vujeva et al. 2026a; M. Ho-Yeuk Cheung et al. 2026). We also assumed a constant n0n_0, despite the possibility that the compact-lens population evolves over cosmic time. A more realistic treatment could incorporate redshift evolution through a parameterization, such as nl(z)=n0(1+z)κn_l(z) = n_0(1+z)^\kappa, where κ\kappa characterizes the redshift evolution of the lens abundance. Such models generally require additional hyperparameters and would introduce further correlations between the source and lens population properties.

Regarding detectability, two aspects of the selection effects should be further refined. First, current GW search pipelines may fail to detect highly distorted microlensed signals, even when their optimal SNRs are sufficiently large, because signal-consistency tests can downrank waveforms that deviate substantially from the unlensed template bank [34]. A realistic selection function should therefore account for the additional detection criteria used by GW search pipelines, such as the ξ2\xi^2 signal-consistency test [14]. Second, among the events recovered by the search pipeline, identifying a signal as microlensed requires a sufficiently low FPP. Rather than relying solely on a Bayes-factor threshold, this probability should be calibrated through dedicated background injections, allowing the microlensing identification criterion and its associated selection effects to be modeled consistently.

Lastly, our injection-and-recovery framework relies exclusively on the isolated point-mass lens model. However, microlensed GW signals may also originate from galaxies in which strong lensing by their gravitational potential is possible, with stellar or compact-object populations embedded within the lens galaxy introducing additional microlensing perturbations. Such systems can produce waveform distortions substantially more complex than those generated by an isolated point-mass lens alone [45, 44, 37, 76, 75, 88, 104, 90, 91, 89, 59, 100]. Similar diffraction is also expected near caustics [49], the regions of highest magnification, leading to unique observational signatures (R. K. L. Lo et al. 2025) and boosting the effects of DM subhalos [99]. Addressing these limitations will require both more robust statistical detection frameworks and the development of more sophisticated waveform models in future analyses.

Acknowledgments

We thank Srashti Goyal for her valuable comments and feedback. E. S. is supported by grants from the College of Science and Engineering of the University of Glasgow. J.M.E. is supported by Research Grant No. VIL53101 from Villum Fonden and the Marie Sklodowska-Curie Grant Agreement No. 847523 INTERACTIONS. The Center of Gravity is a Center of Excellence funded by the Danish National Research Foundation under grant no. DNRF184.. M. H. is supported by the Science and Technology Facilities Council (Grant Ref. ST/V005634/1). The authors are grateful for computational resources provided by the LIGO Laboratory and supported by the National Science Foundation Grants PHY-0757058 and PHY-0823459. This material is based upon work supported by NSF’s LIGO Laboratory which is a major facility fully funded by the National Science Foundation.

Mock BBH Population

We generate the population of simulated unlensed source BBHs from the PowerLaw+Peak primary mass model and a Madau–Dickinson-like (P. Madau & M. Dickinson [72]) redshift distribution, adopting the posterior median values of the population hyperparameters inferred from GWTC-3 (R. Abbott et al. [6]), as summarized in Table 2.

ParameterDescriptionInjected value
α\alphaSpectral index for the power law of the primary mass distribution3.78
β\betaSpectral index for the power law of the mass ratio distribution0.81
mmin⁡m_{\min}Minimum mass of the power-law component of the black hole mass distribution4.98M⊙4.98M_{\odot}
mmax⁡m_{\max}Maximum mass of the power-law component of the black hole mass distribution112.5M⊙112.5M_{\odot}
λg\lambda_{\mathrm{g}}Fraction of BBH systems in the Gaussian component0.03
μg\mu_{\mathrm{g}}Mean of the Gaussian component in the primary mass distribution32.27M⊙32.27M_{\odot}
σg\sigma_{\mathrm{g}}Standard deviation of the Gaussian component in the primary mass distribution3.88M⊙3.88M_{\odot}
δm\delta_mRange of mass tapering at the lower end of the mass distribution4.8M⊙4.8M_{\odot}
zpz_{\mathrm{p}}Redshift at which the distribution peaks2.7
γ\gammaPower-law index of the redshift distribution for z<zpz < z_{\mathrm{p}}4.59
κ\kappaPower-law index of the redshift distribution for z>zpz > z_{\mathrm{p}}2.86

Table 2. List of hyperparameters of the BBH population model and their injected values used in our GWTC-3-based simulations.

Mass distribution

The primary source-frame mass m1m_1 is drawn from a truncated power law defined over mmin⁡=5M⊙m_{\min}=5M_{\odot} and mmax⁡=112M⊙m_{\max}=112M_{\odot}, with a Gaussian peak and a smooth low-mass taper:

ppop(m1∣α,λg,μg,σg,mmin⁡,δm)=[(1−λg)m1−α+λg2πσgexp⁡(−(m1−μg)22σg2)]S(m1∣mmin⁡,δm),(16)p_{\mathrm{pop}}(m_1 \mid\alpha,\lambda_{\mathrm{g}},\mu_{\mathrm{g}},\sigma_{\mathrm{g}},m_{\min},\delta_m)=\left[(1-\lambda_{\mathrm{g}})m_1^{-\alpha}+\frac{\lambda_{\mathrm{g}}}{\sqrt{2\pi}\sigma_{\mathrm{g}}}\exp\left(-\frac{(m_1-\mu_{\mathrm{g}})^2}{2\sigma_{\mathrm{g}}^2}\right)\right]S(m_1\mid m_{\min},\delta_m), \tag*{(16)}

where α\alpha denotes the power-law index, and λg\lambda_{\mathrm{g}}, μg\mu_{\mathrm{g}}, and σg\sigma_{\mathrm{g}} correspond to the mixture fraction, mean, and standard deviation of the Gaussian component, respectively. S(m1∣mmin⁡,δm)S(m_1\mid m_{\min},\delta_m) represents a smoothing function, which is given by

S(m∣mmin⁡,δm)={0(m<mmin⁡),[exp⁡(δmm−mmin⁡+δmm−mmin⁡−δm)+1]−1(mmin⁡≤m<mmin⁡+δm),1(m≥mmin⁡+δm),(17)S(m\mid m_{\min},\delta_m)= \begin{cases} 0 & (m<m_{\min}),\\ \left[\exp\left(\frac{\delta_m}{m-m_{\min}}+\frac{\delta_m}{m-m_{\min}-\delta_m}\right)+1\right]^{-1} & (m_{\min}\leq m<m_{\min}+\delta_m),\\ 1 & (m\geq m_{\min}+\delta_m), \end{cases} \tag*{(17)}

where δm\delta_m sets the width of the low-mass taper applied near mmin⁡m_{\min}.

The secondary mass m2m_2 is determined by the mass ratio q≡m2/m1q\equiv m_2/m_1, which is drawn from a truncated power-law distribution with the low-mass smoothing function defined in Eq. (17):

ppop(q∣β,m1,mmin⁡,δm)∝qβS(qm1∣mmin⁡,δm),mmin⁡<m2≤m1,(18)p_{\mathrm{pop}}(q\mid\beta,m_1,m_{\min},\delta_m)\propto q^{\beta}S(qm_1\mid m_{\min},\delta_m),\qquad m_{\min}<m_2\leq m_1, \tag*{(18)}

where β\beta denotes the power-law index of the qq distribution.

Redshift distribution

For a source-frame merger-rate density R(z)\mathcal{R}(z) at redshift zz, the number of mergers per unit detector-frame time and redshift is given by:

d2Ndt dz=R(z)1+zdVcdz,(19)\frac{d^2N}{dt\,dz}=\frac{\mathcal{R}(z)}{1+z}\frac{dV_{\mathrm{c}}}{dz}, \tag*{(19)}

where VcV_c is the comoving volume at zz. Following the parametrization of T. Callister et al. [23], which builds on the star-formation-rate form of P. Madau & M. Dickinson [72], the merger rate density can be written as:

R(z)=R0(1+z)γ1+(1+zp)−(γ+κ)1+(1+z1+zp)γ+κ(20)\mathcal{R}(z) = \mathcal{R}_0(1+z)^\gamma\frac{1+(1+z_p)^{-(\gamma+\kappa)}}{1+\left(\frac{1+z}{1+z_p}\right)^{\gamma+\kappa}} \tag*{(20)}

where R0=R(0)\mathcal{R}_0=\mathcal{R}(0), zpz_p is the redshift at which R(z)\mathcal{R}(z) peaks, and γ\gamma and κ\kappa denote the asymptotic power-law indices z≪zpz \ll z_p and z≫zpz \gg z_p, respectively. In our simulations, we adopt the local merger rate density R0=20 Gpc−3yr−1\mathcal{R}_0 = 20\ \mathrm{Gpc}^{-3}\mathrm{yr}^{-1}.

Consistency with GWTC-5

The most recent BBH population model is based on GWTC-5 (A. Abac et al. [3]). Since repeating the full parameter-estimation analysis with the updated population is computationally expensive, we instead repeat the mock-catalog generation using the posterior-median hyperparameters from the GWTC-5 population analysis (The LIGO Scientific Collaboration et al. 2026), which adopts a Broken Power Law + 2 Peaks model for the primary-mass distribution (C. Talbot & E. Thrane [96]; The LIGO Scientific Collaboration et al. 2026) and a simple power law for the redshift distribution, while retaining the same mass-ratio distribution as that given in Eq. (18).

Relative to the GWTC-3 population, the GWTC-5 population yields a larger total number of simulated events satisfying ρnet>8\rho_{\mathrm{net}} > 8 owing to its higher detector sensitivity, and therefore a larger absolute number of microlensed events.

To compare the microlensing distinguishability on an equal footing, we instead consider the fraction of detectable microlensed events that satisfy Δχ2>3.84\Delta\chi^2>3.84. For αML=1\alpha_{\mathrm{ML}}=1 (22), this fraction changes from 1727/1914=90.2%1727/1914=90.2\% (291/446=65.2%291/446=65.2\%) for GWTC-3 to 2223/2543=87.4%2223/2543=87.4\% (342/522=65.5%342/522=65.5\%) for GWTC-5. We further compare the distributions of the detector-frame chirp mass (Mc,det\mathcal{M}_{\mathrm{c,det}}) and Δχ2\Delta\chi^2, which are directly related to the microlensing distinguishability. As shown in Figure 4, the GWTC-3 and GWTC-5 populations predict nearly identical distributions of these observable quantities. The choice of BBH population model is therefore unlikely to have a significant impact on our inference of λl\lambda_l.

Figure 4. Comparison of the detectable microlensed BBH populations for αML=1\alpha_{\mathrm{ML}} = 1 (top) and αML=2\alpha_{\mathrm{ML}} = 2 (bottom) generated from the GWTC-3 (blue) and GWTC-5 (orange) population models. The left and right panels show the distributions of Mc,det\mathcal{M}_{\mathrm{c,det}} and Δχ2\Delta\chi^2, respectively.

Microlensing Optical Depths

Only a fraction of GW signals are expected to undergo detectable microlensing, with the microlensing probability characterized by the optical depth, τ(zs)\tau(z_s), for a GW source at redshift zsz_s. Assuming that the lenses are uniformly distributed, for a given n0n_0 value, the optical depth can be written as:

τ(zs)=∫0zsdzl∫dMl n0p(Ml) πymax⁡2RE214πDl2dVcdzl(21)\tau(z_s)=\int_0^{z_s} dz_l \int dM_l\,n_0p(M_l)\,\pi y_{\max}^2R_E^2\frac{1}{4\pi D_l^2}\frac{dV_c}{dz_l} \tag*{(21)}

where p(Ml)p(M_l) is the distribution of MlM_l given by Eq. (3), assuming a fixed αML\alpha_{\mathrm{ML}} over the mass range [Mmin⁡,Mmax⁡][M_{\min},M_{\max}]. The factor ymax⁡y_{\max} is introduced because, in the wave-optics regime, diffraction effects allow detectable microlensing signatures to extend beyond RER_E commonly associated with EM microlensing (ymax⁡=1y_{\max}=1) (see e.g. M. Çalışkan et al. [22]). Consequently, the effective lensing cross-section expands to σ=πymax⁡2RE2\sigma=\pi y_{\max}^2R_E^2 (R. Takahashi & T. Nakamura [95]). As discussed in § 2.3 (see Figure 1), we adopt ymax⁡=2.75y_{\max}=2.75 and ymax⁡=2.07y_{\max}=2.07 for the αML=1\alpha_{\mathrm{ML}}=1 the αML=2\alpha_{\mathrm{ML}}=2 populations, respectively.

The population-averaged optical depth is obtained by averaging Eq. (21) over the simulated BBH redshift distribution:

τˉ=∫0zmax⁡τ(zs)psrc(zs) dzs(22)\bar{\tau}=\int_0^{z_{\max}}\tau(z_s)p_{\mathrm{src}}(z_s)\,dz_s \tag*{(22)}

where psrc(zs)p_{\mathrm{src}}(z_{\mathrm{s}}) is the normalized source-redshift distribution and zmax⁡z_{\max} denotes the maximum source redshift considered in the simulation. For a survey containing NULN_{\mathrm{UL}} GW sources, the expected number of microlensed events is given by:

NML=NUL τˉ.(23)N_{\mathrm{ML}} = N_{\mathrm{UL}}\,\bar{\tau}. \tag*{(23)}

Lens parameters of identified microlensed events

The lens properties of identified microlensed events can provide an early indication of the underlying lens-mass distribution. Figure 5 compares the posterior medians of MlzM_l^z and yy, together with the network SNRs, of events identified with BUML>1B_{\mathrm{U}}^{\mathrm{ML}} > 1. For each αML\alpha_{\mathrm{ML}} scenario, we also show the corresponding distributions for 1000 randomly selected events from the initial NMLN_{\mathrm{ML}} simulated population.

Figure 5. Histograms of the posterior medians of MlzM_l^z and yy, and ρnet\rho_{\mathrm{net}} of identified microlensed events with BMLU>1\mathcal{B}_{\mathrm{ML}}^{\mathrm{U}}>1. The upper and lower rows correspond to αML=1\alpha_{\mathrm{ML}}=1 and 22, respectively, with n0=106 Mpc−3n_0=10^6\,\mathrm{Mpc}^{-3} fixed. For MlzM_l^z and yy, the histograms are compared with those of 1000 samples drawn from the corresponding population models.

The two scenarios differ most notably in their MlzM_l^z distributions. For αML=1\alpha_{\mathrm{ML}} = 1, events with Mlz<103 M⊙M_l^z < 10^3\,M_\odot are rare in the initial population and are consequently unlikely to be identified. By contrast, the αML=2\alpha_{\mathrm{ML}} = 2 scenario yields a broader distribution with a substantially larger fraction of low-mass lenses, making such events more likely to be identified. Repeated detections of microlensed events with low MlzM_l^z could therefore provide an early indication in favor of αML=2\alpha_{\mathrm{ML}} = 2. The hierarchical Bayesian analysis quantifies this distinction by incorporating the information carried by individual events into the population-level inference.

Mass-dependent dark matter fraction

fDMf_{\mathrm{DM}} in Eq. (14) compresses the lens population into a single number, and in our hierarchical model it is constrained only through the correlated pair (αML,n0)(\alpha_{\mathrm{ML}}, n_0). We therefore map the joint posterior onto the differential DM fraction that (i) provides a direct view of which lens-mass scales contribute to the inferred compact-object mass density, and (ii) allows direct comparison with mass-dependent constraints on compact-object scenarios whose abundance is expected to be concentrated within specific mass ranges, such as PBHs [27]. The differential DM fraction per log⁡10Ml\log_{10} M_l is given by:

dfDMdlog⁡10Ml=8πGn03H02ΩDMMl2−αMLln⁡(10).(24)\frac{\mathrm{d}f_{\mathrm{DM}}}{\mathrm{d}\log_{10} M_l} = \frac{8\pi G n_0}{3H_0^2\Omega_{\mathrm{DM}}} M_l^{2-\alpha_{\mathrm{ML}}}\ln(10). \tag*{(24)}

We evaluate Eq. (24) for each joint posterior sample of (αML,n0)(\alpha_{\mathrm{ML}}, n_0). As discussed in § 4.2, in the αML=1\alpha_{\mathrm{ML}} = 1 scenario, a fraction of the joint posterior lies in the unphysical region fDM>1f_{\mathrm{DM}} > 1. Figure 6 illustrates the corresponding unphysical posterior samples, which will be discarded in evaluating Eq. (24).

Figure 6. Joint posterior samples in the (αML,n0)(\alpha_{\mathrm{ML}}, n_0) plane for NML,det=50N_{\mathrm{ML,det}} = 50, 100, and 132. The gray dots indicate the unphysical domain fDM>1f_{\mathrm{DM}} > 1 and are discarded in evaluating the differential fDMf_{\mathrm{DM}}.

Figure 7 shows the differential spectra obtained from the remaining physical samples. For αML=1\alpha_{\mathrm{ML}} = 1, the differential DM fraction increases approximately as Ml2−αML∼MlM_l^{2-\alpha_{\mathrm{ML}}} \sim M_l, so uncertainties in αML\alpha_{\mathrm{ML}} increasingly broaden the spectrum toward higher MlM_l. Note that the fDM≤1f_{\mathrm{DM}} \leq1 truncation preferentially removes the high-n0n_0 tail, particularly for NML,det=100N_{\mathrm{ML,det}} = 100, shifting the median and narrowing the credible interval so that its constraint can appear nearly as tight as that for NML,det=132N_{\mathrm{ML,det}}=132 despite the smaller number of detected events. This apparent tightening is caused by the physical truncation rather than by additional constraining power from the data.

For αML=2\alpha_{\mathrm{ML}}=2, by contrast, the differential spectrum is substantially flatter. In this case, dfDM/dlog⁡10Ml\mathrm{d}f_{\mathrm{DM}}/\mathrm{d}\log_{10} M_l is approximately independent of MlM_l, reducing the propagation of uncertainty in αML\alpha_{\mathrm{ML}} toward higher MlM_l. Reflecting the improved constraining power of the larger event sample, NML,det=132N_{\mathrm{ML,det}}=132 yields the narrowest credible interval and the closest agreement with the true spectrum among the three cases.

References

  1. [1]Abac, A., Abouelfettouh, I., Acernese, F., et al. 2025a, arXiv preprint arXiv:2512.16347arxiv.org/abs/2512.16347
  2. [2]Abac, A., Abouelfettouh, I., Acernese, F., et al. 2025b, arXiv preprint arXiv:2508.18083DOI
  3. [3]Abac, A., Abe, A., Abouelfettouh, I., et al. 2026, arXiv preprint arXiv:2605.27225arxiv.org/abs/2605.27225
  4. [4]Abbott, B. P., Abbott, R., Abbott, T., et al. 2020, Living reviews in relativity, 23, 1
  5. [5]Abbott, R., Abbott, T. D., Abraham, S., et al. 2021, The Astrophysical Journal, 923, 14
  6. [6]Abbott, R., Abbott, T., Acernese, F., et al. 2023, Physical Review X, 13, 011048
  7. [7]Abbott, R., Abe, H., Acernese, F., et al. 2024, The Astrophysical Journal, 970, 191
  8. [8]Abrams, N. S., & Takada, M. 2020, The Astrophysical Journal, 905, 121
  9. [9]Acernese, F., Agapito, A., Agarwal, D., et al. 2026, arXiv preprint arXiv:2607.26872arxiv.org/abs/2607.26872
  10. [10]Ade, P. A. A., Aghanim, N., Arnaud, M., et al. 2016, Astronomy & Astrophysics, 594, A13
  11. [11]Alcock, C., Allsman, R., Alves, D. R., et al. 2001, The Astrophysical Journal Letters, 550, L169
  12. [13]Ashton, G., Hübner, M., Lasky, P. D., et al. 2019, The Astrophysical Journal Supplement Series, 241, 27
  13. [14]Babak, S., Biswas, R., Brady, P., et al. 2013, Physical Review D, 87, 024033
  14. [16]Barsode, A., Kapadia, S., & Ajith, P. 2024, The Astrophysical Journal, 975, 48
  15. [17]Basak, S., Ganguly, A., Haris, K., et al. 2022, The Astrophysical Journal Letters, 926, L28
  16. [18]Belczynski, K., Sadowski, A., & Rasio, F. A. 2004, The Astrophysical Journal, 611, 1068
  17. [19]Blaineau, T., Moniez, M., Afonso, C., et al. 2022, Astronomy & Astrophysics, 664, A106
  18. [20]Bramante, J., Diamond, M., & Kim, J. L. 2024, Journal of Cosmology and Astroparticle Physics, 2024, 002
  19. [21]Buckley, M. R., & DiFranzo, A. 2018, Physical review letters, 120, 051102
  20. [22]Çalışkan, M., Anıl Kumar, N., Ji, L., et al. 2023, Phys. Rev. D, 108, 123543, doi: 10.1103/PhysRevD.108.123543
  21. [23]Callister, T., Fishbach, M., Holz, D. E., & Farr, W. M. 2020, The Astrophysical Journal Letters, 896, L32
  22. [24]Cao, Z., Li, L.-F., & Wang, Y. 2014, Phys. Rev. D, 90, 062003, doi: 10.1103/PhysRevD.90.062003
  23. [25]Capote, E., Jia, W., Aritomi, N., et al. 2025, Physical Review D, 111, 062002
  24. [26]Cardoso, V., & Pani, P. 2019, Living Reviews in Relativity, 22, 4
  25. [27]Carr, B., Kühnel, F., & Sandstad, M. 2016, Physical Review D, 94, 083504
  26. [28]Carr, B., Raidal, M., Tenkanen, T., Vaskonen, V., & Veermäe, H. 2017a, Physical Review D, 96, 023514
  27. [29]Carr, B., Tenkanen, T., & Vaskonen, V. 2017b, Physical Review D, 96, 063507
  28. [30]Carr, B. J., & Green, A. M. 2025, in Primordial black holes (Springer), 3–33
  29. [31]Chabrier, G. 2003, Publications of the Astronomical Society of the Pacific, 115, 763
  30. [32]Chakraborty, A., & Mukherjee, S. 2026, The Astrophysical Journal, 1003, 20
  31. [33]Chan, J. C., Ezquiaga, J. M., Lo, R. K., et al. 2026, Physical Review D, 114, L201502
  32. [34]Chan, J. C., Seo, E., Li, A. K., Fong, H., & Ezquiaga, J. M. 2024, arXiv preprint arXiv:2411.13058DOI
  33. [35]Chang, J. H., Egana-Ugrinovic, D., Essig, R., & Kouvaris, C. 2019, Journal of Cosmology and Astroparticle Physics, 2019, 036
  34. [36]Chen, Z.-C., Yuan, C., & Huang, Q.-G. 2022, Physics Letters B, 829, 137040
  35. [37]Cheung, M. H., Gais, J., Hannuksela, O. A., & Li, T. G. 2021, Monthly Notices of the Royal Astronomical Society, 503, 3326
  36. [39]Christian, P., Vitale, S., & Loeb, A. 2018, Phys. Rev. D, 98, 103022, doi: 10.1103/PhysRevD.98.103022DOI
  37. [40]Coleman Miller, M., & Colbert, E. J. 2004, International Journal of Modern Physics D, 13, 1
  38. [41]Colleoni, M., Ramis Vidal, F. A., García-Quirós, C., Akçay, S., & Bera, S. 2025, Phys. Rev. D, 111, 104019, doi: 10.1103/PhysRevD.111.104019DOI
  39. [42]d’Amico, G., Panci, P., Lupi, A., Bovino, S., & Silk, J. 2018, Monthly Notices of the Royal Astronomical Society, 473, 328
  40. [43]Deguchi, S., & Watson, W. D. 1986, Astrophysical Journal, Part 1 (ISSN 0004-637X), vol. 307, Aug. 1, 1986, p. 30-37., 307, 30
  41. [44]Diego, J. M. 2020, Phys. Rev. D, 101, 123512, doi: 10.1103/PhysRevD.101.123512DOI
  42. [45]Diego, J. M., Hannuksela, O. A., Kelly, P. L., et al. 2019, Astronomy & Astrophysics, 627, A130
  43. [46]Diemand, J., Kuhlen, M., Madau, P., et al. 2008, Nature, 454, 735
  44. [47]Dominik, M., & Sahu, K. C. 2000, The Astrophysical Journal, 534, 213
  45. [48]Ezquiaga, J. M., Hu, W., Lagos, M., Lin, M.-X., & Xu, F. 2022, Journal of Cosmology and Astroparticle Physics, 2022, 016
  46. [49]Ezquiaga, J. M., Lo, R. K. L., & Vujeva, L. 2025, Phys. Rev. D, 112, 043544, doi: 10.1103/tkk2-x9stDOI
  47. [50]Fairbairn, M., Urrutia, J., & Vaskonen, V. 2023, Journal of Cosmology and Astroparticle Physics, 2023, 007
  48. [51]Fardeen, J., McGill, P., Perkins, S. E., et al. 2024, The Astrophysical Journal, 965, 138
  49. [52]Farr, W. M. 2019, Research Notes of the AAS, 3, 66
  50. [53]Fedorova, E., Slusar, V., Zhdanov, V., et al. 2016, Monthly Notices of the Royal Astronomical Society, 457, 4147
  51. [54]Gais, J., Ng, K. K., Seo, E., Wong, K. W., & Li, T. G. 2022, The Astrophysical Journal Letters, 932, L4
  52. [55]Gaudi, B. S. 2012, Annual Review of Astronomy and Astrophysics, 50, 411
  53. [56]Gould, A. 1992, Astrophysical Journal, Part 1, 392, 442
  54. [57]Gould, A. 2000, The Astrophysical Journal, 542, 785
  55. [58]Gould, A. 2004, The Astrophysical Journal, 606, 319
  56. [59]Goyal, S., Villarrubia-Rojo, H., & Zumalacárregui, M. 2026, The Astrophysical Journal Letters, 1008, L12
  57. [61]Green, A. M. 2026, arXiv preprint arXiv:2602.15974arxiv.org/abs/2602.15974
  58. [62]Green, A. M., & Kavanagh, B. J. 2021, Journal of Physics G: Nuclear and Particle Physics, 48, 043001
  59. [63]Hannuksela, O., Haris, K., Ng, K., et al. 2019, The Astrophysical Journal Letters, 874, L2
  60. [66]James-Turner, C., Weil, D. P., Green, A. M., & Copeland, E. J. 2020, Physical Review D, 101, 123526
  61. [67]Jung, S., & Shin, C. S. 2019, Phys. Rev. Lett., 122, 041103, doi: 10.1103/PhysRevLett.122.041103DOI
  62. [68]Lai, K.-H., Hannuksela, O. A., Herrera-Martín, A., et al. 2018, Phys. Rev. D, 98, 083005, doi: 10.1103/PhysRevD.98.083005DOI
  63. [69]Lin, X.-Y., Wang, X.-J., Zhou, H., et al. 2025, arXiv preprint arXiv:2508.13577arxiv.org/abs/2508.13577
  64. [70]Liu, L., You, Z.-Q., Wu, Y., & Chen, Z.-C. 2023, Physical Review D, 107, 063035
  65. [72]Madau, P., & Dickinson, M. 2014, Annual Review of Astronomy and Astrophysics, 52, 415
  66. [73]Mandel, I., Farr, W. M., & Gair, J. R. 2019, Monthly Notices of the Royal Astronomical Society, 486, 1086
  67. [74]Mao, S. 2012, Research in Astronomy and Astrophysics, 12, 947
  68. [75]Meena, A. K., Mishra, A., More, A., Bose, S., & Bagla, J. S. 2022, Monthly Notices of the Royal Astronomical Society, 517, 872
  69. [76]Mishra, A., Meena, A. K., More, A., Bose, S., & Bagla, J. S. 2021, Monthly Notices of the Royal Astronomical Society, 508, 4869
  70. [78]Nakamura, T. T. 1998, Physical review letters, 80, 1138
  71. [79]Narayan, R. 2005, New Journal of Physics, 7, 199
  72. [80]Niikura, H., Takada, M., Yasuda, N., et al. 2019, Nature Astronomy, 3, 524
  73. [81]Oguri, M., Diego, J. M., Kaiser, N., Kelly, P. L., & Broadhurst, T. 2018, Physical Review D, 97, 023518
  74. [82]Perkins, S. E., McGill, P., Dawson, W., et al. 2024, The Astrophysical Journal, 961, 179
  75. [83]Pratten, G., García-Quirós, C., Colleoni, M., et al. 2021, Physical Review D, 103, 104056
  76. [84]Romao, M. C., Croon, D., Crossey, B., & Godines, D. 2025, Journal of Cosmology and Astroparticle Physics, 2025, 066
  77. [85]Ryan, M., & Radice, D. 2022, Physical Review D, 105, 115034
  78. [86]Sasaki, M., Suyama, T., Tanaka, T., & Yokoyama, S. 2018, Classical and Quantum Gravity, 35, 063001
  79. [87]Schneider, P., Kochanek, C., & Wambsganss, J. 2006, Gravitational lensing: strong, weak and micro: Saas-Fee advanced course 33, Vol. 33 (Springer Science & Business Media)
  80. [88]Seo, E., Hannuksela, O. A., & Li, T. G. 2022, The Astrophysical Journal, 932, 50
  81. [89]Seo, E., Shan, X., Janquart, J., et al. 2025, The Astrophysical Journal, 988, 159
  82. [90]Shan, X., Chen, X., Hu, B., & Cai, R.-G. 2023, arXiv preprint arXiv:2301.06117DOI
  83. [91]Shan, X., Chen, X., Hu, B., & Li, G. 2024, Sci. China Phys. Mech. Astron., 67, 269511, doi: 10.1007/s11433-023-2334-9DOI
  84. [92]Shan, X., Yang, H., & Mao, S. 2025, arXiv preprint arXiv:2512.19118arxiv.org/abs/2512.19118
  85. [93]Silk, J., & Stebbins, A. 1993, Astrophysical Journal-Part 1, 411, 439
  86. [94]Speagle, J. S. 2020, Monthly Notices of the Royal Astronomical Society, 493, 3132
  87. [95]Takahashi, R., & Nakamura, T. 2003, The Astrophysical Journal, 595, 1039
  88. [96]Talbot, C., & Thrane, E. 2018, The Astrophysical Journal, 856, 173
  89. [98]Turner, E. L., Ostriker, J. P., & Gott III, J. R. 1984, Astrophysical Journal, Part 1 (ISSN 0004-637X), vol. 284, Sept. 1, 1984, p. 1-22., 284, 1
  90. [99]Vujeva, L., Ezquiaga, J. M., Gilman, D., Goyal, S., & Zumalacárregui, M. 2026a, Phys. Rev. D, 114, 083024, doi: 10.1103/kp86-q1h2
  91. [100]Vujeva, L., Ezquiaga, J. M., Lo, R. K. L., & Zwick, L. 2026b, https://arxiv.org/abs/2608.07667
  92. [101]Wang, B., & Yang, T. 2026, arXiv preprint arXiv:2604.08179arxiv.org/abs/2604.08179
  93. [102]Winch, H., Setford, J., Bovy, J., & Curtin, D. 2022, The Astrophysical Journal, 933, 177
  94. [103]Wright, M., & Hendry, M. 2022, The Astrophysical Journal, 935, 68
  95. [104]Yeung, S. M., Cheung, M. H., Seo, E., et al. 2023, Monthly Notices of the Royal Astronomical Society, 526, 2230
  96. [105]Zhang, K., Luo, L.-W., Tsao, J.-S., Chen, C.-S., & Lin, F.-L. 2023, Results in Physics, 53, 106967
  97. [106]Zumalacarregui, M., & Seljak, U. 2018, Physical review letters, 121, 141101

Paper details

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

views · 0

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents