WorksPhysics and Astronomy
Astrometric microlensing by isolated black holes in the Roman Galactic Bulge Time Domain Survey
Abstract
Most of the hundred million stellar-mass black holes (BHs) expected in the Milky Way are predicted to be isolated. Characterizing this BH population will provide insights into stellar evolution and supernova explosion physics. The Nancy Grace Roman Space Telescope Galactic Bulge Time Domain Survey (GBTDS) will enable the detection of isolated BHs through astrometric microlensing, apparent stellar position shifts caused by deflection of distant starlight. We combine a Galactic model, stellar population synthesis model, BH initial–final mass relation, and kick velocity distribution to simulate events caused by isolated BH lenses. We assess the expected yield, sensitivity, characterization potential, and signal characteristics of these events to inform GBTDS search strategies. We predict isolated-BH microlensing events with detectable astrometric signals, 82% of which have impact parameters beyond 2 Einstein radii and consequently have small or undetectable photometric signals. Standard astrometric fits yield unit weight errors > 1.2 for 18% and > 2 for 2% of these events. Sensitivity peaks for 5-6 BHs and is limited by the five-year GBTDS baseline. Weak microlensing parallax constraints limit BH mass precision, but events with Einstein timescales > 100 days and angular Einstein radii > 1.2 mas have high BH lens probabilities within our Galactic model. Wide-separation astrometric microlensing offers an independent discovery channel for isolated BHs beyond photometrically selected events. Realizing this opportunity requires dedicated multi-season GBTDS searches, joint astrometric and photometric modeling, and accounting for other astrometric variability. Although precise individual masses are challenging for these events, probabilistic lens classification could lead to Galactic abundance constraints on the population of isolated BHs.
Introduction
The Milky Way is expected to harbour about one hundred million black holes (BHs) that formed during the death and collapse of massive stars (e.g., [88]; [8]; [21]; [65]). Characterizing this population will provide insights into stellar evolution mechanisms (e.g., [60]; [39]; [80]), supernova explosion physics (e.g., [93]; [76]), and the means to understand the extragalactic population of massive merging binary BHs detected via gravitational-wave emission (e.g., [5]; Ho et al. 2024).
Despite their expected abundance, little is known about the Galactic BH population. This is due in large part to the difficulty of detecting these dark objects. All but one of the known Galactic BHs have been detected via interactions with a companion in binary systems, e.g., astrometric ([60]; [19 b]; [27]) or X-ray ([75]; [16]) binaries. These detection methods are sensitive only to specific regions of parameter space, e.g., BH and companion masses and orbital parameters, leading to strong selection effects.
This limited sample of BHs leaves a number of open questions about their demographics. A notable example is the “mass gap” observed between known neutron stars (NSs) and BHs ([9]; [67]; [25]). It remains unclear whether the gap is caused by astrophysical effects preventing the formation of remnants within this mass range (e.g., [11]; [26]; Olejok et al. 2022) or by possible observational biases (e.g., [48]; [106]; [49]). Furthermore, in recent years, a new, numerous population of BHs in the mass range of 30–150 has been discovered via gravitational-wave observations ([2, 3, 4, 5 b]; The LIGO Scientific Collaboration et al. 2025; [1]). These results need to be reconciled with the much lower BH masses observed in our Galaxy (e.g., [16]) as well as those predicted by stellar evolution models (e.g., [93]). The recent discovery of a 33 BH in the Milky Way ([27]) has renewed interest in consistent models that could explain the properties of BHs discovered through different observational channels. Importantly, the vast majority of stellar-origin BHs in the Galaxy are expected to be isolated ([54]; [65]), meaning that the sample of known BHs in binary systems is unlikely to be representative of the Galactic population. Therefore, to build a complete and unbiased census of BHs in the Galaxy, the sample of isolated BHs must be increased.
The only currently known isolated stellar-mass black hole was detected recently via a gravitational microlensing event, OB110462 ([85, 53, 63, 50]). During this event, the black hole gravitationally deflected the light of a distant background source, causing a source flux amplification (photometric microlensing; Paczynski [68]) and positional shift (astrometric microlensing; [100, 62, 33]). Using both of these signals, the lens mass was determined to be approximately . This event was initially reported by the Optical Gravitational Lensing Experiment (OGLE; Udalski et al. [97]).
Indeed, there are many examples of isolated black hole candidates that have been found via photometric microlensing events (e.g., [105, 106, 49, 81, 37, 44]). However, such candidates cannot be definitively confirmed unless further follow-up, astrometric or otherwise, is carried out.
Astrometric microlensing allows a measurement of the angular Einstein radius that is crucial for measuring the lens mass ([33, 100, 62, 17]) and is not possible with photometric data alone. Surveys combining astrometric and photometric microlensing observations could be especially powerful in providing precise mass measurements for dark remnants ([82, 43, 86]). Additionally, astrometric microlensing acts at larger lens–source separations than photometric microlensing (Miralda-Escude [61]); astrometric microlensing also has longer and more mass-sensitive timescales (Honma [35]). At the GBTDS’s expected astrometric precision, for some classes of lenses, e.g., massive isolated BHs, there will be events with detectable astrometric signatures but small or even undetectable photometric signatures (e.g., Fardeen et al. [24]). Therefore, including astrometry-only detections would increase the sample sizes of microlensing events and could improve population studies, which is particularly valuable for populations with very few known objects, such as isolated stellar-mass BHs.
Observations of astrometric microlensing show great promise for characterizing stellar remnants ([84, 58]), but remain challenging due to the required precision. Even the most favourable events produce signals of milliarcsecond (mas) and are detectable only with a few observatories. All detections to date have required multi-year astrometric follow-up at this precision, which is difficult to access and highly oversubscribed (see e.g., Kaczmarek et al. [44]). However, we are now entering an era of astrometric surveys providing regular time-series data precise enough to detect shifts caused by microlensing. For example, Gaia (Gaia Collaboration et al. [28]) achieved – mas single-observation precision (Lindegren et al. (2021)), with tentative astrometric microlensing signals already reported (e.g., Jabłońska et al. [40]) and astrometry-only detection and modelling methods proposed for future releases (Chen et al. [15]; ); Janković et al. [41]). However, crowding limits Gaia’s performance in fields where most microlensing events occur (e.g., [57, 14]).
Launched in August 2026, the Roman Space Telescope (Spergel et al. [92], hereafter Roman) is the first space mission with microlensing as one of its key science goals. Roman will monitor the Galactic Bulge with the Wide Field Instrument (WFI) over 5 years of its primary mission performing the Galactic Bulge Time Domain Survey (GBTDS). The survey design is optimized for microlensing science, focusing primarily on cold exoplanet demographics (Zasowski et al. [107]). GBTDS targets a deg field near the Galactic Center observing in a wide near-infrared filter () that minimizes the effects of extinction, while also achieving high photometric precision by maximising collected flux ([69, 59]). Exposures will be obtained every 12.1 minutes with a photometric precision of mmag, in a design tailored for large-scale microlensing exoplanet detection ([69, 102]). Due to WFI’s expected stability and precise tracking, each of these high-cadence exposures will also provide position measurements at a precision of about 1 mas; exposures could be stacked for a further significant improvement in precision (e.g., [59, 51]). However, microlensing exoplanets produce small astrometric signals with angular scales of microarcseconds (Safizadeh et al. [83]); consequently, anomalous astrometry will not be processed in Roman’s microlensing pipeline (Roman Science Support Center (2024)). That means additional efforts must be made to identify astrometric microlensing events in the GBTDS data.
In this study, we simulate and analyse GBTDS observations of the population of isolated BHs that can be detected through astrometric microlensing. We extend previous work on detecting microlensing isolated BHs that considers only small lens–source impact parameter events (typically angular Einstein radii) with both detectable astrometric and photometric signals (Pruett et al. [74]; ); Sajadian & Sahu [86]) to the wider lens–source impact parameter regime (e.g., [98, 99, 24, 41]).
The paper is structured as follows. In Section 2, we cover the astrometric microlensing effect. In § 3, we outline the process of simulating the GBTDS and detecting astrometric microlensing events. In § 4, we report yields of detectable astrometric microlensing events caused by isolated BH lenses and we analyse the constraints that can be placed on physical properties of the lenses. We also investigate characteristics of the astrometric microlensing signals and how they can be selected from GBTDS. Finally, in § 5, we discuss and summarize our findings, including recommendations for finding isolated BHs in GBTDS.
Astrometric microlensing
Consider a lens of mass at a distance and a more distant background source at distance from an observer. If the source, the lens, and the observer are perfectly collinear, an Einstein ring image of the source is formed with angular radius (Einstein [18]);
If the lens–source alignment is imperfect, i.e., the lens and source have an angular separation , pointing towards the source position from the lens position and normalized by , two images of the source are formed. If the images are unresolved and the lens is dark, the apparent astrometric shift of the source-image centroid from the unlensed source position (see e.g., Bramich [13]) is:
The apparent photometric amplification of the source flux is (e.g., Paczynski [68]),
As the lens–source separation changes over the course of an event, it can be parameterised as:
Here, is the angular lens–source impact parameter, pointing towards the source position and normalized by ; hereafter, . is the moment of closest approach, , is the unit vector in the direction of the relative lens–source proper motion, and is the relative lens–source parallax motion.
The parameters that contain information on the lens and source are the Einstein timescale and microlensing parallax, defined as
respectively. Here, , and AU is one astronomical unit. If and can be constrained, the lens mass can be measured (e.g., Gould [30]),
For , assuming a detectability threshold , the region around the lens where a background source would cause a detectable signal can be approximated by a circle of radius (e.g., Dominik & Sahu [17]; Bramich [13]). Assuming a one-day stack of GBTDS exposures corresponding to an astrometric precision of mas for a F146-band magnitude source (e.g., Lam et al. [51]), and mas for a BH, this gives , which is greater than the lens–source separation regimes considered previously (e.g., Sajadian & Sahu [86]). The timescale of the astrometric microlensing effect can be quantified as the time spent by the source, neglecting parallax motion, within an angular radius of of the lens (Honma [35]),
For a given survey duration and assuming sufficiently high cadence, if an event peaks within and , the survey will observe its full amplitude from peak astrometric deviation to baseline. However, for typical lenses in the Galaxy, can be – years (e.g., Belokurov & Evans [12]) and exceed the survey duration, which means that some long-duration events will not be detected solely because of an insufficient temporal baseline. Specifically, we must require some threshold change in the astrometric signal as a detection (e.g., Dominik & Sahu [17]). This point is relevant for the 5-year duration of the GBTDS, and we therefore consider it in our event selection criteria. It is also worth mentioning that blended light from unrelated neighbouring sources changes the form of the centroid shift in Eq. (2) and also reduces the magnitude of the astrometric microlensing signal (Bramich [13]; Lam et al. [52]). To avoid this complication, we limit our sample of events to those with no or low blending.
Simulation
Galactic model, stellar populations, and remnants
To simulate a catalogue of microlensing events, we used the Population Synthesis for Compact Lensing Events code (PopSyCLE;
Notes. The bulge parameters were chosen to be consistent with the “v3” version in Lam et al. [52] (App. A), and match the event rates reported by OGLE (Mróz et al. [64]).
PopSyCLE uses a modified version of Galaxia (see Sharma et al. [89], Appendix of Lam et al. [52]) for the underlying stellar model of the Milky Way that is based on the Besançon model (Robin et al. [77]). BHs, NSs and WDs are injected into the simulation by evolving clusters matching each Galaxia stellar sub-population, assuming they are single-age, single-metallicity populations, via the Stellar Population Interface for Stellar Evolution and Atmospheres code (SPISEA; Hosek et al. [36]). BHs and NSs are given natal kick velocities; initial–final mass relations (IFMRs) are used to determine masses of BHs, NSs and WDs, while white dwarfs (WDs) given the physical properties of their progenitors. The Galactic model parameters and relations adopted for the simulation are detailed in Table 1.
| Parameter | Value |
| Milky Way escape velocity | (Piff et al. 2014) |
| Sun–Galactic Center distance | |
| Mean of initial BH progenitor kick distribution | |
| Mean of initial NS progenitor kick distribution | |
| Initial–Final Mass Relation | Sukhbold N20 (based on models from Sukhbold & Woosley [94]; Sukhbold et al. [93]; Woosley [103]; Woosley et al. [104]) |
| Bar dimensions (radius, major axis, minor axis, height) | |
| Bar angle (Sun–Galactic Center, 2nd, 3rd) | |
| Bulge velocity dispersion (radial, azimuthal, z) | |
| Bar pattern speed |
Table 1. Summary of PopSyCLE simulation parameters.
Galactic Bulge Time Domain Survey
Our simulated survey largely follows the GBTDS simulation of Kaczmarek et al. [45]. The main difference introduced in this work is a much wider permitted range for the impact parameter , allowing for capturing events with no photometric signal detectable only in astrometric time series. We approximate the GBTDS footprint by simulating ten separate fields of within each of the six WFI fields (see Figure 1). Lensing events are selected from the simulated populations as pairs of luminous sources and foreground lenses that reach a separation of within the simulated survey time of 5 years (Roman’s primary mission1). In all the work that follows, predicted event rates are scaled to the full GBTDS area as defined in the Roman Observations Time Allocation Committee Final Report and Recommendations (Zasowski et al. [107], hereafter: ROTAC report).

Figure 1. Layout of the simulation fields (blue circles), overplotted on the extinction map (magenta; Surot et al. [95]; Huston et al. [38], implementation discussed in § 3.2). Shades of blue of the field outlines represent the total number of detectable astrometric microlensing events within each simulation field, while shades of magenta in the background pixel map represent extinction. The event yields are heavily impacted by extinction, as lowest-yield fields overlap with high- regions. The nominal Roman GBTDS footprint has been overplotted with black solid lines using the gbtds_optimizer software1.
Lens and source magnitudes reported by PopSyCLE in the UBV, , , and filters are converted to extinction-free -band magnitudes using the relations of Wilson et al. [102]. Extinction models performing badly near the Galactic Center are a known problem (e.g., [29, 38]). Therefore, we build on SynthPop22 (Klüter et al. [47]; Huston et al. [38]) and its implementation of the Surot et al. [95] extinction map. Surot et al. [95] have created a two-dimensional high-resolution map based on VISTA Variables in the Via Lactea (VVV) observations of red clump and red giant branch stars. Huston et al. [38] have implemented this map as a query-able function by scaling the 2D extinction with the line-of-sight dust distribution; they specifically choose the Surot et al. [95] map for reproducing Roman’s GBTDS fields, optimizing the model for bulge stars. As implemented in SynthPop, we take:
where , . As is a very wide filter, we integrate over its extents. For simplicity, we assume the passband to be a tophat with transmission of 1 for and 0 otherwise. We assume the incoming spectrum to be that of a blackbody with , the median of our simulated source stars (with a 10th–90th percentile range of 3.65–3.80; changing within those ranges only introduces a mag error in the calculated values). With photon-counting weights,
We realise this integration on a 2000-point grid uniform in , and map to . We show the resulting map for a fixed – the Galactic center distance assumed in SynthPop – in Figure 1.
We adopt the nominal GBTDS cadence pattern detailed in the ROTAC report, which follows Galactic bulge visibility seasons centered around the vernal and autumnal equinoxes. Six high-cadence observing seasons are defined, each lasting 70.5 days; within each of the seasons, observations are taken at a 12.1 minute cadence in the bandpass over the full survey area. Seasons three and four are spaced with a “large season gap” of 4 Galactic bulge visibility seasons. We denote this set of 50 346 observing times over the GBTDS .
Event selection
We follow similar criteria as Kaczmarek et al. [45] for selecting detectable astrometric events. We apply an cut for the magnitude at baseline (in the case of luminous lenses, we consider joint source and lens flux) to exclude sources impacted by background noise. We select only events where the closest approach occurs within the bounds of GBTDS: . We also require that there is a detectable change in astrometric signal within the GBTDS time baseline, , where
and for stacking observations taken within 24 hours under a 12.1-minute cadence. We use Eqs. (2), (4) to calculate ; we omit the parallax motion component in Eq. (4), assuming its negligible effect on astrometric signal detectability. In the case of luminous lenses, is calculated including corrections for the blending effect, using relations from Dominik & Sahu [17].
We differ from Kaczmarek et al. [45] in the definition of , as a new detailed study of Roman’s position uncertainties has recently been published by McKinnon & van der Marel [59], hereafter McK26. We use the data and codes published in the repository3 accompanying McK26 to define the single-exposure position uncertainty . We follow McK26 in adding in quadrature a noise floor of 1% of the pixel width; we use simulated data for 66-second exposures in the middle of the 5-field strip in the GBTDS footprint (see Figure 1).
Finally, we require little to no blending – i.e., at least 0.8 of the total (source, lens, and neighbours) blended flux coming from the source, according to the PopSyCLE-reported -band blending parameters. In addition to excluding events with a diluted astrometric signal due to blending, this cut ensures that the event is simple to model astrometrically (see e.g., Bramich [13]). We also add a photometric detectability flag to our dataset, but do not use it as a selection criterion. This flag is defined as in Kaczmarek et al. [45] and analogously to Eq. (10); we require (corresponding to a photometric signal of at peak) and , i.e., a change in magnitude exceeding the magnitude-dependent precision limit determined using Eqs. (6)–(7) from Kaczmarek et al. [45].
Notes. is the uniform distribution with lower and upper bounds of and , respectively. is the normal distribution with mean and standard deviation of mean and std, respectively.
Constraining event parameters
We create simulated observations over for all astrometrically detectable events that have BH lenses. We simulate the ground truth positions and fluxes from Eqs. (2)–(3) and add white Gaussian noise with standard deviations (1-day stacking), consistent with the photometric and astrometric selection criteria described in § 3.3. In addition to the nominal , following Roman User Documentation [79]4 we add gap-filling observations with a cadence of 5 days and added noise of standard deviation (single exposure). The error bars match the standard deviations of the added noise.
The model parameter priors are detailed in Table 2. Parallax factors and standard five-parameter source tracks were calculated using astromet5. We fit the astrometric microlensing signal jointly with the standard five-parameter source astrometric solution for the unlensed source path. The fit is performed by assuming a white Gaussian noise likelihood. We use the nested sampling algorithm Higson et al. [31] implemented by DYNESTY Speagle [90] because of its ability to find multiple modes of the posterior that are common in microlensing event modeling (e.g., Kaczmarek et al. [43]). We use random-walk sampling with multiple bounding ellipsoids and 1000 initial live points. We use a stopping criterion based on the remaining fractional evidence of 0.01 and allocate all weight to posterior computation. For each event, we model two scenarios: with and without photometry.
| Parameter | Prior distribution | Unit | Definition |
| – | impact parameter | ||
| decimal year | time of closest approach (rel. lin. motion) | ||
| years | log (event timescale) | ||
| – | log (microlensing parallax) | ||
| mas | log (Einstein radius) | ||
| truncated to | mas/yr | source proper motion in | |
| truncated to | mas/yr | source proper motion in | |
| mas | log (source parallax) | ||
| truncated to of data | mas | offset to source ref. position in | |
| truncated to of data | mas | offset to source ref. position in | |
| truncated to of data | mag | baseline magnitude (optional) |
Table 2. Prior distributions for modeling the detected events.
Results
Event yields
The total yields of selected events within the GBTDS footprint are shown in Figure 1. We predict a total of 3755 103 astrometric events over the full duration and area of GBTDS, 487 37 of these have BH lenses. Of these events, 1540 65 do not pass photometric detectability ( and ; § 3.3) cuts; 398 33 of which have BH lenses (see Figure 2, left panel). Therefore, we find that including astrometry-only events increases the yield of detectable BH lenses with astrometric signals by a factor of 5.4. This increased BH yield does not necessarily imply a proportional increase in overall population constraints. This is because mass measurements are more challenging and are likely to have higher uncertainties in the absence of a photometric signal (see § 4.3).

Figure 2. Left: Bar chart showing the predicted number of each lens type that will produce detectable astrometric signals during the nominal 5-year GBTDS. Hatched bars represent subsets of astrometric events that are also photometrically detectable. Right: Distributions of astrometric event timescales (; Eq. (7)) for the sample of detected events with non-BH (blue) and BH (magenta) lenses. Each distribution is normalized and therefore contains no information on the relative number of BH and non-BH events. Marked as vertical dashed lines are 70.5 days, 841 days, and 5 years, which correspond to the duration of a single season, large season gap (between seasons 3 and 4), and total survey duration for the nominal GBTDS (Penny et al. [69]). See § 4.2 for discussion.
We also find that astrometry-only events are an efficient channel for BH detection. 25.8% of events with any astrometric signals detectable are caused by BH lenses. This is large compared with the corresponding BH lens proportions of 4.0% and 1.1% of events detectable in photometry and astrometry, and photometry only, respectively. This is due to a combination of the astrometric signals’ dependence on the lens mass, which is typically higher for BHs compared with other lenses, and the astrometric selection criteria that were applied. 97% of the photometrically detectable BH lenses are also astrometrically detectable.
Characteristics of the astrometric signals
The right panel of Figure 2 shows the distribution of astrometric event timescales for BH and non-BH lenses that are detectable by the GBTDS. At this point, it is useful to reiterate our astrometric detection criteria. We require a detectable change in astrometric deflection within the GBTDS survey (Eq. (10)). This criterion excludes events that could have a detectable absolute astrometric signal but do not vary significantly enough within the GBTDS time baseline to be detectable (e.g., Dominik & Sahu [17]).
For events that pass the astrometric detection criteria, Figure 2 shows that the timescale distribution peaks around 1000 days, with many BH events having astrometric timescales comparable to the total primary mission time. For these events, the GBTDS will only detect part of the full astrometric microlensing event, and the source star will be displaced astrometrically for the full GBTDS. This means that for the majority of BH events, multiple seasons, if not all seasons, of the GBTDS will be required to characterize the full astrometric signal. Figure 3 shows example trajectories of different events with BH lenses whose astrometric signals will be detectable during the GBTDS. Most events span multiple GBTDS seasons of data. This also suggests that any gap-filling observations will increase the number of data points on the detectable signal for the majority of BH events.

Figure 3. Example black hole astrometric microlensing event trajectories that will be detectable during the GBTDS (with respect to a reference position at the midpoint of the survey). Grey solid lines show the unlensed background source trajectory. Blue dashed lines show the lensed observed background source trajectory. Magenta circles represent GBTDS observations. The clusters of measurements are the assumed nominal GBTDS 70.5-day observing seasons. The low-cadence gap-filling observations will have significantly higher uncertainties than stacked observations in high-cadence seasons, and therefore are not shown.
Figure 2 also shows that there is a minority of short-timescale events with days, suggesting that gap-filling observations between seasons 3 and 4 are likely to boost yields of events that would otherwise fall completely within this gap. Finally, Figure 2 shows that detectable events with BH lenses that have an astrometric timescale less than the duration of a single high-cadence season, 70.5 days, are extremely rare. This suggests that if an astrometric microlensing signal is completely contained within one GBTDS season, we find that it is unlikely that it is caused by a BH. is often overestimated, resulting in underestimated masses, especially at low signal-to-noise ratios. Figure 4 shows the distribution of these constraints across the detected sample and their accuracy. Similarly to the posterior precision, the posterior accuracy is better for and , compared with and . We note that, for a subset of low signal-to-noise events, multiple or extended posterior modes were found. We defer analysis of these degeneracies to future work. All posterior constraints are reported including all modes.
Figure 4. Posterior precision and accuracy of the microlensing event parameters and the corresponding lens mass for the detected sample of events. Top panels: Fractional posterior constraint is defined as the (84th–16th)/(50th) percentile of the posterior distribution. Including the photometry in the model fitting improves constraints for all parameters. and are typically better constrained than and . Bottom panels: Fitted versus true input values for event parameters. For clarity, data points are coloured by density estimated with a 2D Gaussian KDE, with darker shades (of magenta and green for fits with and without photometry, respectively) signifying higher point density. Similar to the posterior precision in the top panels, and are typically more accurately recovered than and . Including the photometry in the model also improves posterior accuracy.
Lens constraints
For the set of detectable events we find that including (and excluding) the photometry, and can be constrained with a typical (median) precision of (110%) and (135%). and consequently are less well constrained with typical precisions of (520%) and (580%), respectively. In particular, high introduces greater freedom in the modelled trajectory, and the fit may absorb random noise or shifts from linear motion into an artificially inflated parallax signal. Therefore, including the photometry, even when the photometric microlensing signal amplitude is low or undetectable, improves the constraints on all parameters. This is even the case for — a parameter for which only the astrometric signal contains direct information. This is because the amplitude, or lack thereof, of the photometric signal places a lower bound on and/or . This bounding of the normalized lens–source separation allows a tighter constraint on because the astrometric signal is a function of both and (see Eq. (2)). Similar reasoning on bounding can explain the increase in precision for and when including photometry. Additionally, even at low signal to noise, the photometry can contain direct information on and which will also act to increase posterior precision when photometry is included.
Better constraints for and , compared with and consequently , are expected (e.g., [23, 45]. This is because, to first order, and are correlated with the amplitude and duration of the astrometric signal, respectively. These quantities are usually straightforward to constrain. In contrast, typically produces smaller deviations in the astrometric microlensing signal. This is because the astrometric signal is not sensitive to small lens–source parallax-induced deviations at large () separations (Dominik & Sahu [17]).
Figure 4 shows that while there is a heavy left tail of ≈ 20% of events where we can obtain posterior constraints between 1% and 100%, the majority of these events will not yield a tight constraint on . This unfortunate situation suggests that these events will not lead to impactful constraints on the shape of the BH mass function. However, in the context of a Galactic model, these events could still provide constraining information on the number of isolated BHs. In the following Section we turn our attention to exploring this possibility. That is, we ask the question – if we cannot constrain from the event, can we still identify the lens as a BH?
Lens classification
Figure 5 shows where the population of BH lenses is located in the – space relative to the population of non-BH lenses. Analyzing the separation of lens types in microlensing observables is important because it could be used to identify high-confidence BH candidates or, at least, provide a lens class probability that can be marginalized over in population model inference ([72, 87] (Perkins et al. 2024).

Figure 5. Left: The space of high-signal parameters, –, coloured by probabilities of belonging to a given lens class . Non-BH astrophysical lens classes are grouped together, effectively: BH, Star or WD or NS, None. Each point on a 1000 × 1000 grid is coloured by the probability of belonging to either of the astrophysical classes (magenta – BH, blue – non-BH). The inset represents a 2D colorbar; probabilities are normalized to 1, making the space above the diagonal where forbidden. There is a distinct region (which can be conservatively approximated by cuts of days and mas) where the probability of the lens being a BH is . The BH and non-BH regions are well-separated, with a relatively small region belonging to the overlap (purple) where confusion between BH and non-BH classes is high. Center: Distributions of BH and non-BH lenses in the – space, along with inference results for the mock dataset of BH-lens astrometric events (§ 4.3). Contours represent the simulated astrometric events caused by black holes (magenta) vs. all other types of lenses: stars, WDs, and NSs (blue); they were built using a Gaussian kernel density estimate and contain (80, 65, 50, 35, 20, 5)% of the simulation samples. Black data points represent astrometric BH events that are well-constrained, i.e., that have both and fractional posterior constraints < 1; this subset contains 79% of our simulated astrometric BH events. The position in the plot represents median and values from samples of the nested sampling fit to the mock GBTDS observations; error bars represent 16th–84th percentiles from samples. Right: histogram of – the probability of the lens being a black hole based on its posterior and the Galactic model – as returned by popclass for all simulated astrometric BH events (solid filled) and the well-constrained subset (cross-hatched). See § 4.4 for detailed discussion of the popclass classification results.
BH lenses typically have larger and than the other lens types. Although the BH and non-BH lens populations are mostly separated in the – space, there is overlap. Moreover, the BH lens population is situated in the tails of the much more abundant populations of stellar and WD lenses. One notable feature of the – space shown in Figure 5 is that the BH lenses are located in a region of high signal compared with the non-BH population. That is, BH lens events are more likely to have a larger and, therefore, a larger astrometric signal. This suggests that and therefore their location and lens class, in – space can be measured more precisely. This is in contrast to photometric microlensing events monitored from the ground, where and can be measured, but BH lenses are located in a low-signal region of large but small , resulting in uncertain lens classifications (e.g., [52, 44]. The left panel of Figure 5 suggests that there is a region of – space ( days and mas), populated with detectable astrometric events, where the probability of a BH lens is close to 100%.
To predict how well BH lenses can be identified with GBTDS data, we classify our simulated BH events using popclass6 (Sallaberry et al. [87]). The classification framework calculates – the probability of the lens belonging to a given lens class based on its posterior and the Galactic model . We use posteriors from joint modeling of photometric and astrometric time series for classification, as they provide more precise parameter constraints (Figure 4). We construct the astrophysical lens classes (Star, White Dwarf, Neutron Star, Black Hole) on the basis of all astrometrically detectable events from our simulation; we also construct a None class that captures combinations of parameters that are inconsistent with astrophysical lenses in the Galactic model. We refer to Sallaberry et al. [87] for a detailed discussion of the classification procedure, and to Kaczmarek et al. [45] for the construction of the None class over the – space which we reproduce here.
Figure 5 shows the classification results. In the center panel, we show all BH events that are well-constrained with the nested sampling fits to mock GBTDS data. We consider an event well-constrained if it has fractional posterior constraints on both and lower than 1. We find that most of the well-constrained fits are located in the parameter space region occupied by BH lenses. 56% of all events (and 64% of the well-constrained sample) are classified as Black Hole (i.e., is the highest for ). 3.4% (0.7%) are assigned the Star class, 1.1% (1.4%) – the Neutron Star class, and 1.7% (1.2%) – the None class. The most significant source of confusion by far, at 37% (31%), is the White Dwarf class. Kaczmarek et al. [44] have noted that high-probability regions for WD and BH lenses in the observable parameter space are close, giving rise to sharp gradients where ambiguous classifications are likely to occur (including the only known isolated BH, OB110462, yielding a high WD probability). To improve ambiguous WD/BH classifications, parameter constraints must be tightened (e.g., by extending Roman’s observing baseline beyond the nominal mission or incorporating external data).
We also report Spearman’s correlation (Spearman [91]) between event parameters and as a simple gauge of which events are most likely to be recovered. is most correlated with astrometric signal-to-noise ratio () and (); it is moderately () correlated with (which does not meaningfully contribute to astrometric signal and classification, but is a function of ) and UWE; correlation with other parameters (, , ) is weak (). Interestingly, though there are clear linear trends between and or fractional constraints, events with UWE also occur at all ranges and can have well-constrained posteriors. Those events should be identified with other metrics, e.g., taking into account the variability of excess signal in time.
Extracting events from the GBTDS
So far, we have investigated the sample of astrometric microlensing events that can be detected above some detection threshold. This represents an idealized scenario in which we are sure the source in question is undergoing an astrometric microlensing event. In practice, we may not know a priori if this is the case. We must separate astrometric microlensing events from sources with no motion residual relative to the standard proper motion and parallax solution and from sources undergoing accelerations due to different effects (e.g., astrometric binaries; Lam et al. [51]). In this section, we deal with the former and defer the latter issue to future work.
We have performed the 5-parameter (, , , , ) astrometric fits for all our simulated astrometric events and calculated unit weight error (UWE) values of the fits using the astromet package.7 The expected value of UWE for deviations from a 5-parameter track fully explainable by random noise is 1; high excess UWE values point towards extra astrometric signals, such as astrometric microlensing shifts. A similar treatment, using renormalised UWE (RUWE), is commonly applied to extract anomalous astrometric fits in the Gaia mission data, as it provides a
single, easily interpretable value that has been extensively calibrated with real data and simulations (e.g., [70, 14] (Lindegren 2018; Lindegren et al. 2021)) and has been proposed for finding astrometric signals of Gaia microlensing events ([40, 45]). We use UWE as a proxy for RUWE, expecting its renormalisation as a function of magnitude and colour to be performed over the course of Roman’s operations.
Figure 6 (top) shows the UWE distribution with respect to the standard 5-parameter astrometric solution, positions, proper motion, and parallax, for unlensed sources, sources lensed by non-BHs, and sources lensed by BHs. As expected, the population of lensed sources causes a heavy right tail in the UWE distribution that separates from the unlensed source population at UWE . The population of sources lensed by BHs shows an even heavier and flatter tail, with extreme values of UWE . We find no non-BH lenses with UWE . Assuming that other sources of astrometric accelerations, e.g., binaries or systematics, are removed, extreme values of UWE point to a likely BH lens.
Figure 6. Top: Unit weight error (UWE) distributions for a sample of unlensed stars (red, horizontal hatch) and lensing events with non-BH lenses (blue, left-leaning hatch) and BH lenses (magenta, solid filled). Astrometric lensing introduces a tail of high-UWE tracks, which is especially prominent for BH lenses. Bottom: Excess UWE as a function of impact parameter (absolute value, ) and for events with non-BH (left) and BH (right) lenses. As BHs occupy a distinct region of parameter space (see Figure 5), the right plot has smaller extents, marked by the dotted contour within the left plot. Mean values of UWE - 1 are plotted within 2D bins; redder colours signify higher UWE excess, while negligible or negative values are plotted with turquoise. BH lenses have a significantly higher distribution of excess UWE and are responsible for all of the events with extreme (> 0.8) values. While the highest excess (beyond 5-parameter tracks) astrometric signal is observed for events that also have high photometric signal (low ), there is a notable population of events that have a significant excess UWE at high () impact parameters; these events do not have typical microlensing light curves with prominent peaks and are unlikely to be picked by microlensing detection pipelines.
Figure 6 (bottom) shows how UWE varies for different values of and . These event parameters control the amplitude and shape of the astrometric signals. A larger corresponds to a larger astrometric amplitude for an event with a given . Smaller values correspond to more peaked astrometric signals. Figure 6 (bottom, right panel) shows that BH lenses with the smallest values (most peaked signals) and larger values (largest amplitudes) have the highest excess UWE values. These events have the most temporally localized and largest deviations from the standard astrometric solutions.
With held constant, as increases, UWE tends to decrease for both BH and non-BH lenses. This is because larger values lead to less peaked deviations, spreading the signal over a greater timescale. These events are more likely to have a lower UWE because the astrometric signal is more likely to be mistaken for a different set of standard astrometric solution parameters. Notwithstanding the lower UWEs for larger , there is a notable population of events that have a significant excess UWE at high () impact parameters. These events will not have typical microlensing light curves with prominent peaks and therefore are less likely to be identified by the MSOS detection pipeline.
The astrometric sensitivity function
We define the sensitivity to astrometric signal as the fraction ; denotes the number of events with BH lenses that have astrometric signals detectable over the nominal GBTDS in a given sample of simulation results, and – the total number of close () BH lens – background source approaches in the sample. To investigate how the sensitivity varies across event parameter space, we calculate this fraction separately in bins of event parameters (lens mass, event timescale, etc.).
The sensitivity estimates were impacted by significant noise due to the relatively low numbers of detectable events per bin. Astrometric detectability is dependent on parameters that are not physically meaningful, and . Namely, events with lower cause higher astrometric signals (see e.g. Figure 6, Eq. (2)); events with during the large season gap (see e.g. Figure 3) may have no observation epochs covering the high-signal part of the astrometric shift, and are missed more often. To alleviate this effect, we created 50 copies of each simulated BH lens event, for which all parameters except and are unchanged; forrayele the 49 new instances, and are replaced with random values allowed by the simulation bounds: ; . We estimate uncertainties on the detection sensitivity using binomial proportion confidence intervals for selecting detectable events out of total simulated lens–source approaches. To account for the uncertainties decreasing when the impact of nuisance parameters (, ) is minimized, we calculate the effective sample size for each 50-event ensemble using intras-class correlation (ICC; Kish [46]) for binary detection results ( – not detectable or – detectable). The effective sample size of a bin is then (summed over all unique BH lenses within each bin); sensitivity in the bin is , where is the mean for the -th set of 50 ‘copied’ events with the same BH lens. We report Wilson intervals (Wilson [101]) using sensitivity of within a sample size of for each bin. The effective sample sizes and , primarily because a high fraction of events get rejected at the early stages due to magnitude and blending cuts (§ 3.3) which are independent of (, ). Additionally, we find that events with high BH lens masses have low sensitivity to (, ) changes and return particularly high ICC values. However, copying events with (, ) replacement brings a notable decrease of noise on the low- edge, where astrometric signals have shorter timescales and detectability is strongly dependent on .
In Figure 7, we present detection sensitivity as a function of mass and location in the – space. We find that the GBTDS is sensitive to BH masses of . Sensitivity is highest for BHs, whose astrometric signals are rapidly changing and largely contained within the GBTDS timespan. In the – space, astrometric signals can be detectable across a wide range of timescales, with intermediate timescales being most preferred; for each range, the highest astrophysical values are strongly preferred. The high-mass end of the sensitivity function is burdened by high statistical uncertainties, as BHs with masses are rare in the adopted Galactic model (e.g., Rose et al. [80]).

Figure 7. BH lens detection sensitivity as a function of mass (left) and over the joint – observable space (right). The left panel shows that the GBTDS is sensitive through astrometric microlensing to BHs with masses between . BHs on the low-mass end are less likely to be detected due to their typically smaller and therefore smaller astrometric signals. BHs with masses of are most likely to be detected. The sensitivity function increases again for BH masses of , reaching a secondary peak at . This peak marks the transition of median per bin becoming longer than the nominal GBTDS time. Events caused by more massive BHs have high , but a low rate of change of the astrometric shift; to detect those events, additional observations long before or after the nominal surveys would be needed to establish the astrometric baseline for the source. For all bins with , the estimated uncertainty intervals (Wilson intervals for successes) per bin are very large. However, we note that we consistently register no detections of any BH lens with within our simulated dataset. The right panel shows that events with large and intermediate have the highest detection efficiency. These events typically peak and return to baseline within the GBTDS. Detectable events span a wide range of ; within each bin, the highest possible is strongly preferred for detectability.
Discussion and conclusions
We simulated astrometric microlensing in the five-year Roman GBTDS, extending the search to events with impact parameters . Under our adopted Galactic model and selection criteria, we predict 3755 ± 103 events with detectable astrometric signals, 487 ± 37 of which are events caused by isolated BHs. Of these, 1540 ± 66 events, including 398 ± 33 BH events, do not pass our photometric detectability cuts. Therefore, including these astrometry-only events increases the predicted sample of BH lenses with detectable astrometric signals by a factor of 5.5 relative to the photometrically detectable BH sample. This demonstrates the importance of extending GBTDS microlensing searches beyond events identified through their light curves. We find that the GBTDS survey will be most sensitive to BHs of , and not sensitive to BHs with high () masses. Astrometric detection fraction is highest for intermediate timescales and high angular () scales.
Astrometric detection and population constraints
Astrometry-only events are promising for finding isolated BHs. BH lenses account for 25.8% of this sample, compared with 4.0% of events detectable in both photometry and astrometry. This difference reflects both the larger of BH lenses and our selection criteria. Astrometric searches therefore offer not only a larger potential BH sample, but also a sample less dominated by stellar lenses. Our results complement forecasts focused on events with large photometric signals ([74] (Sajadmanesh & Sahu 2023)) and support the importance of the large impact parameter regime originally explored for the GBTDS in Fardeen et al. [24]. However, detecting an astrometric signal, identifying the lens as a BH, and measuring the BH mass are distinct and potentially non-trivial tasks for a given microlensing event. Our simulations show that and are generally better constrained than for these events. This comparatively weak constraint on is the primary obstacle to precise mass measurements. Consequently, this increase in detectable BH events does not translate directly into an equivalent increase in precisely measured BH masses for the GBTDS.
Photometry is valuable even for events that fail our photometric detection criteria. A low-amplitude light curve, or even the absence of measurable amplification, constrains the allowed lens–source separation and reduces degeneracies in the astrometric model. Joint modelling improves constraints even on , despite photometry containing no direct measurement of this parameter.
For this set of events, and are typically well-constrained and provide information about lens class. Within our adopted Galactic model, events with days and mas occupy a region with a high BH lens probability. Most events are correctly classified with the popclass classification software as having BH lenses, based on their mock GBTDS observations. Almost all of the misclassified events are assigned the ‘White Dwarf’ class, owing to the proximity of WD and BH events in the observable parameter space and the high number of WD lenses.
Even events without precise individual mass measurements could constrain the abundance of isolated BHs. This analysis would need to combine event counts and parameter posteriors with a GBTDS astrometric selection function, while marginalizing uncertain classifications and Galactic model assumptions (e.g., [72]). While we expect a minority of these microlensing events to yield precise BH mass measurements, the majority of this sample will not. We therefore expect this sample of microlensing events to be a promising probe of the isolated BH abundance rather than a precise probe of the shape of the isolated BH mass function.
From detectable signals to confirmed BHs
Our predicted yields describe an ideal sample of detectable signals, rather than a predicted yield of recovered and confirmed BH events. Specifically, the selection criterion in Eq. (10) requires a measurable change in the microlensing deflection. It does not require that this change remain significant after fitting the source position, proper motion, and parallax. Slowly varying microlensing signals can be somewhat absorbed into these five astrometric parameters, making them more difficult to extract in a blind search.
The UWE distributions in Figure 6 show this distinction. Only of the selected BH events have UWE , and have UWE . Although we find no non-BH lenses above the latter threshold in our simulated sample, this does not establish a pure BH selection criterion. Astrometric binaries are the main astrophysical source of increased UWE (e.g., [70, 14]); unmodelled blending and instrumental systematics can also produce large residuals (e.g., [22, 34]). On the other hand, a modest UWE does not necessarily imply that a coherent microlensing signal cannot be recovered from the data. In particular, we find events with negligible to no excess UWE that have strong (, ) posterior constraints and high-confidence classification as a BH.
A search strategy with actual GBTDS data could use excess astrometric residuals as an initial diagnostic, followed by joint astrometric and photometric modeling and comparison against alternative explanations. These other explanations should include unlensed stars, astrometric binary systems, and crowded-field systematics, rather than only comparisons among different lens types.
Limitations
The predictions in this study are dependent on the adopted Galactic model, compact-remnant prescription, and observing noise assumptions. The initial–final mass relation determines the underlying BH mass distribution, while natal kicks affect lens kinematics and therefore event timescales. The stellar density, bulge kinematics, extinction prescription, and source luminosity distribution also affect the number of lens–source pairs (e.g., [52, 80]). The uncertainties quoted for our event yields should not be interpreted as including these sources of model uncertainty. Exploring alternative remnant and Galactic models will be required before placing constraints on the Galactic BH population (e.g., [72]).
The mass-dependent GBTDS BH sensitivity requires similar considerations. Within our simulation, sensitivity is highest for BH masses of approximately 5–6 , and detectable BH lenses span approximately 3–19 . We find no detections above 18.3 . This absence is not a model-independent upper mass limit for the GBTDS. Our sensitivity is defined in the context of our event detection criteria using the distances, velocities, and source properties, provided by the Galactic model. The high-mass BH bins in Figure 7 also have substantial statistical uncertainties due to the rarity of high-mass BHs under the adopted initial–final mass relation.
Our simulations and constraint results assume white Gaussian errors and a daily stacked precision obtained from 119 exposures. Realizing this gain requires sufficient independence between the astrometric errors of individual exposures. Temporally correlated errors and crowding residuals could reduce the improvement obtained by stacking, even if the single-exposure uncertainty model is accurate [59, 51]. Their impact should be assessed using the actual incoming GBTDS observations.
The detection criteria that the source star contributes at least 80% of the blended flux selects relatively simple events, but excludes potentially informative events in more crowded environments. The restriction of the event peaking within the GBTDS also excludes some events whose astrometric signatures could be measurable during the survey. Relaxing these detection criteria could lead to a larger yield of BH events.
Finally, our single-lens calculations do not assess the effects of lens or source multiplicity. Binary motion can introduce additional astrometric structure, alter the lensing signal, and complicate lens classification (e.g., Abrams et al. [7]). Quantifying these effects could be particularly important for events without significant photometric microlensing signals.
Implications for the GBTDS
The astrometric timescale distribution peaks near 1000 days, and most detectable BH signals extend across multiple high-cadence observing seasons. Searches restricted to individual seasons will therefore be poorly matched to much of this BH population.
Computation of residuals and summary statistics such as UWE from standard astrometric source fits would enable searches for microlensing BHs that complement the MSOS photometric microlensing pipeline. As the timescales of astrometric signals caused by BH lenses are comparable to the nominal Roman mission duration, the long gaps between observing seasons hinder event characterization. Improving the precision of low-cadence gap-filling GBTDS observations (e.g. by enabling stacking of exposures) could therefore improve constraints on event parameters.
Observations before or after the nominal GBTDS would provide a complementary benefit by extending the astrometric baseline (e.g., pre-GBTDS Gaia data; Gaia Collaboration et al. [28]). Precursor observations (e.g., Beaulieu et al. [10]), an extended Roman mission, or targeted follow-up could help distinguish a long-duration microlensing deflection from the source’s unlensed motion.
In summary, astrometry from the Roman GBTDS can provide an independent discovery channel for isolated BHs, rather than serving only to characterize photometrically selected events. Our simulations predict several hundred detectable BH signals, approximately four-fifths of which will be difficult to detect photometrically. Precise mass measurements will remain challenging for many of these events, but joint photometric and astrometric modelling and probabilistic lens classification could turn this sample into a valuable probe of isolated-BH abundance. Realizing this opportunity will require dedicated multi-season astrometric variability searches, astrometric selection functions, and careful treatment of competing sources of astrometric variability.
Acknowledgements
This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. The document number is LLNL-JRNL-2026695. This work was supported by the LLNL-LDRD Program under Project 27-ERD-017. The authors acknowledge support by the state of Baden-Württemberg through bwHPC.
References
- [1]Abac, A. G., Abouelfettouh, I., Acernese, F., et al. 2025, The Astrophysical Journal Letters, 993, L25
- [2]Abbott, B. P., Abbott, R., Abbott, T. D., et al. 2016, Phys. Rev. Lett., 116, 061102
- [3]Abbott, B. P., Abbott, R., Abbott, T. D., et al. 2017, Phys. Rev. Lett., 119, 141101
- [4]Abbott, R., Abbott, T. D., Abraham, S., et al. 2021, Physical Review X, 11, 021053
- [5]Abbott, R., Abbott, T. D., Acernese, F., et al. 2023a, Physical Review X, 13, 041039
- [7]Abrams, N. S., Lu, J. R., Lam, C. Y., et al. 2025, ApJ, 980, 103
- [8]Agol, E. & Kamionkowski, M. 2002, MNRAS, 334, 553
- [9]Bailyn, C. D., Jain, R. K., Coppi, P., & Orosz, J. A. 1998, ApJ, 499, 367
- [10]Beaulieu, J.-P., Bachelet, E., Gilles, M., et al. 2026, arXiv e-prints, arXiv:2606.25883arxiv.org/abs/2606.25883
- [11]Belczynski, K., Witk...
- [12]Belokurov, V. A. & Evans, N. W. 2002, MNRAS, 331, 649
- [13]Bramich, D. M. 2018, A&A, 618, A44
- [14]Castro-Ginard, A., Penoyre, Z., Casey, A. R., et al. 2024, A&A, 688, A1
- [15]Chen, I. K., Kongsore, M., & Tilburg, K. V. 2023, J. Cosmology Astropart. Phys., 2023, 037
- [16]Corral-Santana, J. M., Casares, J., Muñoz-Darias, T., et al. 2016, A&A, 587, A61
- [17]Dominik, M. & Sahu, K. C. 2000, ApJ, 534, 213
- [18]Einstein, A. 1936, Science, 84, 506
- [19]El-Badry, K., Rix, H.-W., Cendes, Y., et al. 2023a, MNRAS, 521, 4323
- [21]Elbert, O. D., Bullock, J. S., & Kaplinghat, M. 2018, MNRAS, 473, 1186
- [22]Fabricius, C., Luri, X., Arenou, F., et al. 2021, A&A, 649, A5
- [23]Fardeen, J., McGill, P., Perkins, S. E., et al. 2023, arXiv e-prints, arXiv:2312.13249arxiv.org/abs/2312.13249
- [24]Fardeen, J., McGill, P., Perkins, S. E., et al. 2024, ApJ, 965, 138
- [25]Farr, W. M., Sravn, N., Cantrell, A., et al. 2011, ApJ, 741, 103
- [26]Fryer, C. L., Belczynski, K., Witk...
- [27]Gaia Collaboration, Panuzzo, P., Mazeh, T., et al. 2024, A&A, 686, L2
- [28]Gaia Collaboration, Prusti, T., de Bruijne, J. H. J., et al. 2016, A&A, 595, A1
- [29]Gonzalez, O. A., Rejkuba, M., Zoccali, M., et al. 2012, A&A, 543, A13
- [30]Gould, A. 2000, ApJ, 542, 785
- [31]Higson, E., Handley, W., Hobson, M., & Lasenby, A. 2019, Statistics and Computing, 29, 891
- [33]Hog, E., Novikov, I. D., & Polnarev, A. G. 1995, A&A, 294, 287
- [34]Holl, B., Fabricius, C., Portell, J., et al. 2023, A&A, 674, A25
- [35]Honma, M. 2001, PASJ, 53, 233
- [36]Hosek, Matthew W., J., Lu, J. R., Lam, C. Y., et al. 2020, AJ, 160, 143
- [37]Howil, K., Wyrzykowski, Ł., Kruszyńska, K., et al. 2025, A&A, 694, A94
- [38]Huston, M. J., Crisp, A. L., Newman, M., et al. 2026, AJ, 172, 94
- [39]Ivanova, N., Justham, S., Chen, X., et al. 2013, A&A Rev., 21, 59
- [40]Jabłońska, M., Wyrzykowski, L., Rybicki, K. A., et al. 2022, A&A, 666, L16
- [41]Janković, T., Gomboc, A., Wyrzykowski, Ł., et al. 2025, A&A, 699, A156
- [43]Kaczmarek, Z., McGill, P., Evans, N. W., et al. 2022, MNRAS, 514, 4845
- [44]Kaczmarek, Z., McGill, P., Perkins, S. E., et al. 2025, ApJ, 981, 183
- [45]Kaczmarek, Z., Sweeney, D., Wyrzykowski, Ł., & Bastian, U. 2026b, A&A, 714, A374
- [46]Kish, L. 1965, Survey Sampling (Wiley)
- [47]Klüter, J., Huston, M. J., Aronica, A., et al. 2025, AJ, 169, 317
- [48]Kreidberg, L., Bailyn, C. D., Farr, W. M., & Kalogera, V. 2012, ApJ, 757, 36
- [49]Kruszyńska, K., Wyrzykowski, Ł., Rybicki, K. A., et al. 2024, A&A, 692, A28
- [50]Lam, C. Y. & Lu, J. R. 2023, ApJ, 955, 116
- [51]Lam, C. Y., Lu, J. R., El-Badry, K., et al. 2026, arXiv e-prints, arXiv:2608.24998arxiv.org/abs/2608.24998
- [52]Lam, C. Y., Lu, J. R., Hosek, Matthew W., J., Dawson, W. A., & Golovich, N. R. 2020, ApJ, 889, 31
- [53]Lam, C. Y., Lu, J. R., Udalski, A., et al. 2022, ApJ, 933, L23
- [54]Lamberts, A., Garrison-Kimmel, S., Hopkins, P. F., et al. 2018, MNRAS, 480, 2704
- [57]Luna, A., Marchetti, T., Rejkuba, M., & Minniti, D. 2023, A&A, 677, A185
- [58]McGill, P., Anderson, J., Casertano, S., et al. 2023, MNRAS, 520, 259
- [59]McKinnon, K. A. & van der Marel, R. P. 2026, PASP, 138, 044507
- [60]Miller-Jones, J. C. A., Bahramian, A., Orosz, J. A., et al. 2021, Science, 371, 1046
- [61]Miralda-Escude, J. 1996, ApJ, 470, L113
- [62]Miyamoto, M. & Yoshii, Y. 1995, AJ, 110, 1427
- [63]Mroz, P., Udalski, A., & Gould, A. 2022, ApJ, 937, L24
- [64]Mroz, P., Udalski, A., Skowron, J., et al. 2019, ApJS, 244, 29
- [65]Olejak, A., Belczynski, K., Bulik, T., & Sobolewska, M. 2020, A&A, 638, A94
- [67]Özel, F., Psaltis, D., Narayan, R., & McClintock, J. E. 2010, ApJ, 725, 1918
- [68]Paczynski, B. 1986, ApJ, 301, 503
- [69]Penny, M. T., Gaudi, B. S., Kerins, E., et al. 2019, ApJS, 241, 3
- [70]Penoyre, Z., Belokurov, V., & Evans, N. W. 2022, MNRAS, 513, 2437
- [72]Perkins, S. E., McGill, P., Dawson, W. A., et al. 2026, ApJ, 1006, 110
- [74]Pruett, K., Dawson, W., Medford, M. S., et al. 2022, arXiv e-prints, arXiv:2211.06771arxiv.org/abs/2211.06771
- [75]Remillard, R. A. & McClintock, J. E. 2006, ARA&A, 44, 49
- [76]Repetto, S., Igoshev, A. P., & Nelemans, G. 2017, MNRAS, 467, 298
- [77]Robin, A. C., Reylé, C., Derrière, S., & Picaud, S. 2004, A&A, 416, 157
- [79]Roman User Documentation. 2022–, Roman User Documentation (RDoc)
- [80]Rose, S. L., Lam, C. Y., Lu, J. R., et al. 2022, ApJ, 941, 116
- [81]Rybicki, K. A., Shvartzvald, Y., Yee, J. C., et al. 2024, ApJ, 975, 216
- [82]Rybicki, K. A., Wyrzykowski, L., Klencki, J., et al. 2018, MNRAS, 476, 2013
- [83]Safizadeh, N., Dalal, N., & Griest, K. 1999, ApJ, 522, 512
- [84]Sahu, K. C., Anderson, J., Casertano, S., et al. 2017, Science, 356, 1046
- [85]Sahu, K. C., Anderson, J., Casertano, S., et al. 2022, ApJ, 933, 83
- [86]Sajadian, S. & Sahu, K. C. 2023, AJ, 165, 96
- [87]Sallaberry, G., Kaczmarek, Z., McGill, P., et al. 2025, The Journal of Open Source Software, 10, 7769
- [88]Samland, M. 1998, ApJ, 496, 155
- [89]Sharma, S., Bland-Hawthorn, J., Johnston, K. V., & Binney, J. 2011, ApJ, 730, 3
- [90]Speagle, J. S. 2020, MNRAS, 493, 3132
- [91]Spearman, C. 1904, The American Journal of Psychology, 15, 72
- [92]Spergel, D., Gehrels, N., Baltay, C., et al. 2015, arXiv e-prints, arXiv:1503.03757arxiv.org/abs/1503.03757
- [93]Sukhbold, T., Ertl, T., Woosley, S. E., Brown, J. M., & Janka, H. T. 2016, ApJ, 821, 38
- [94]Sukhbold, T. & Woosley, S. E. 2014, ApJ, 783, 10
- [95]Surot, F., Valenti, E., Gonzalez, O. A., et al. 2020, A&A, 644, A140
- [97]Udalski, A., Szymański, M. K., & Szymański, G. 2015, Acta Astron., 65, 1
- [98]Van Tilburg, K., Taki, A.-M., & Weiner, N. 2018, J. Cosmology Astropart. Phys., 2018, 041
- [99]Verma, H. & Rentala, V. 2023, J. Cosmology Astropart. Phys., 2023, 045
- [100]Walker, M. A. 1995, ApJ, 453, 37
- [101]Wilson, E. B. 1927, Journal of the American Statistical Association, 22, 209
- [102]Wilson, R. F., Barclay, T., Powell, B. P., et al. 2023, ApJS, 269, 5
- [103]Woosley, S. E. 2017, ApJ, 836, 244
- [104]Woosley, S. E., Sukhbold, T., & Janka, H. T. 2020, ApJ, 896, 56
- [105]Wyrzykowski, Ł., Kostrzewa-Rutkowska, Z., Skowron, J., et al. 2016, MNRAS, 458, 3012
- [106]Wyrzykowski, Ł. & Mandel, I. 2020, A&A, 636, A20
- [107]Zasowski, G., Jha, S. W., et al. 2025, arXiv:2505.10574

