An astrometric search for planets in debris disk systems

Abstract

Debris disks are created and sculpted by planetary bodies in the orbital space they share. The properties of these disks, including mass, orbital extent, and morphology, can be indicators of their planetary shepherds. Recently, T. Pearce and collaborators placed limits on the masses and orbits of hypothetical planets around 178 stars with resolved debris disks. We consider 176 of these stars, all the objects that have astrometric data in the Gaia Data Release 3 archive, to assess planet detection from astrometry. Our analysis begins with a set of stellar hosts of known exoplanets, selected to roughly match the parallax, apparent magnitude, and color of the 176 debris disk systems. We confirm that Gaia’s ruwe parameter, a measure of the quality of astrometric fitting to a linear drift model, is sensitive to the presence of massive companions, even planetary ones. Guided by ruwe and a metric derived from a machine-learning algorithm trained on Gaia parameters from the exoplanetary host data set, we identify promising stars with debris disks that may host as-yet-undiscovered planets. These stars will be compelling subjects for time‑series analyses with Gaia Data Release 4.

1 Introduction

Debris disks are powerful tracers of the structure and evolution of planetary systems (Wyatt 2008; Krivov 2010; Hughes et al. 2018; Manara et al. 2023). Composed of dust grains as small as a fraction of a micron, debris disks arise from collisions driven by the gravity of planets or planetesimals (Wyatt & Dent 2002; Kenyon & Bromley 2002a; Kenyon & Bromley 2002b; Dominik & Decin 2003; Wyatt 2008; Mustill & Wyatt 2009; Kenyon & Bromley 2010; Raymond et al. 2011; Najita et al. 2022, e.g.,). Although the more massive bodies may themselves be difficult to detect, copious dust reprocesses and scatters starlight efficiently. The Vega debris disk, the first discovered (Aumann et al. 1984), has a remarkably smooth structure, hosting an asteroid belt, yet no planets as massive as Saturn beyond a few au (Su et al. 2005; Su et al. 2013; Su et al. 2024). But for the dust, this architecture would be missed.

Debris disks with gaps and asymmetries about their orbital axis may indicate the presence of massive bodies that sculpt these features. Prominent examples include $\beta$ Pictoris with debris rings between and beyond the host star’s two planets (Mouillet et al. 1997; Lagrange et al. 2009; Nowak et al. 2020), and HR 8799, with dusty debris between the star’s four giant planets and even extending beyond them (Wyatt et al. 1999; Marois et al. 2008; Su et al. 2009; Booth et al. 2016; Esposito et al. 2020). Both $\beta$ Pic and HR 8799 show clear connections between their debris disks and planets.

Pearce et al. 2022 leverage the connection between debris disks and planets that sculpt them to constrain planetary properties from disk geometry (Faber & Quillen 2007; Mustill & Wyatt 2012; Pearce & Wyatt 2014; Morrison & Malhotra 2015; Nesvold & Kuchner 2015; Shannon et al. 2016; Lazzoni et al. 2018; Regály et al. 2018, see also). They examine 178 debris disk systems and assess gravitational stirring mechanisms for explaining the observed disk structure. Self-stirring models tend to require untenably large disk masses, so Pearce et al. 2022 consider stirring by adjacent planets. With a case-by-case analysis, they predict the orbital distances and masses of planetary candidates that may be responsible for generating the observed dusty debris.

Here we explore an independent approach to inferring the presence of the planetary candidates in the 178 debris disk systems from Pearce et al. 2022. All but two stars in this catalog are also in the Gaia DR3 archive; Gaia’s astronomical measures may be sensitive to the stars’ reflex motion arising from a massive companion planet. A strategy of using Gaia’s astrometric quality indicators to identify companions has proved successful in a range of contexts, including binary stars (Brandt 2021; Belokurov et al. 2020; Fabricius et al. 2021; Kervella et al. 2022; Halbwachs et al. 2023), substellar objects, planets (Holl et al. 2023; Stefánsson et al. 2025; Vioque et al. 2026, e.g) and black holes (El-Badry et al. 2023; Müller-Horn et al. 2025). On the eve of the release of Gaia Data Release 4 (Brown 2025, e.g.,), our hope is to identify the most promising candidates for astrometric confirmation in sources for which there is already compelling evidence for planets (Blakely et al. 2026, for a similar strategy applied to transition disk systems, see).

We begin in §2 with a discussion of the impact of a massive companion on a star’s astrometry, illustrating with stars in Gaia DR3 and the NASA Exoplanet Archive that have known companions (§3). We focus on sources with distances, magnitudes, and color that are similar to stars in the Pearce et al. 2022 catalog, and use both a measure of astrometric quality provided in Gaia DR3 and a machine-learning algorithm based on astrometric errors and their correlations. In §4, we apply our results to the debris systems sample and compare outcomes with the limits on planet mass and location provided by Pearce et al. 2022. We conclude in §5.

2 Astrometric signatures of companions

The astrometry of a star can reveal the presence of an unseen companion due to the star’s reflex motion from the companion’s gravitational force (Lattanzi et al. 2000; Casertano et al. 2008, e.g.,). A star moving on an otherwise straight path in the sky plane will acquire a periodic wobble that may impact the quality of a star’s astrometry if the companion’s influence is not considered. Estimates of position, parallax, and proper motion all fare poorly if a sequence of astrometric measurements of a star is fit with a model of linear drift, depending on the precision of the measurements and their cadence.

This effect is well known. The Gaia DR3 archive includes a parameter ruwe, the renormalized unit weighted error of a constant-drift astrometric model, which turns out to be a robust marker of binary and multiple stars (Brandt 2018; Brandt 2021; Belokurov et al. 2020; Fabricius et al. 2021; Halbwachs et al. 2023; Kervella et al. 2022, e.g.,). Stars with ruwe above a threshold value of $\sim$1.4 are good candidates as members of binary or multiple-star systems, providing the binary orbit and the timing of Gaia’s snapshots of the system are favorable. Castro-Ginard et al. 2024 provide a more detailed description of the connection between ruwe and Gaia’s astrometric measurements.

SK thinks that writing stars with ruwe above 1.4 are ”good candidates” instead of ”excellent candidates” should make it clear there are limits and that we should not belabor the point.

To estimate how a companion impacts astrometry, we compare astrometric sensitivity and the physical motion of a star. When astrometric data are snapshots of sky positions — ignoring radial velocity — orbital motion affects astrometry only when it repositions a star by an amount larger than the positional uncertainty during the total time of the observations, $\Delta t$ (Lattanzi et al. 2000, cf.). We consider two limiting cases. First, at small separation, with binary orbital period $P$ less than $\Delta t$, detectability requires an orbital separation of

$$a \gtrsim \frac{(M+M_{\bullet})}{M_{\bullet}}\frac{\delta\varpi}{\varpi}$$ (1)

$$\gtrsim 0.004\times\frac{(1+q)}{q}\left[\frac{\varpi/\delta\varpi}{250}\right]^{-1}\text{{au}},$$

where $M$ is the mass of the observed partner, which we assume to be the primary, and $M_{\bullet}$ is the companion/secondary mass, $\varpi$ is the true parallax, and $\delta\varpi$ is the typical error for sources at that parallax. In the lower expression, the mass ratio $q=M_{\bullet}/M$ and the fiducial value of parallax-over-error is typical of the sources we consider below. This expression highlights that for planetary masses with $q\lesssim 10^{-2}$, orbital separation for nearby stars must exceed $a\sim$0.4 au. In terms of orbital period $T_{\text{orp}}$, Sun-like, equal-mass binaries vary measurably in their individual positions when $T_{\text{orp}}\gtrsim 1~\text{d}\sim 10^{-3}$ yr (Eq. 1).

Measurements are complicated at small separation when stellar partners are not resolved and only the center of light is observed. With identical twin partners, the position of the center of light does not change at all. When one companion is faint or dark (e.g., Müller-Horn et al. 2025, for black holes), the orbital motion of the brighter star may be detectable. However, rapid orbital motion may well appear as random noise that does not strongly impact an astrometric fit to a linear drift model. For stars in Gaia that we consider here, there are sets of several observations made over the course of a few days, and dozens of such sets (as indicated by Gaia’s visibility_periods_used parameter) spanning several years. We conjecture that even with time-series data, orbital motion more rapid than this cadence may be buried in other noise sources. We thus adopt a lower limit of 0.1 yr for the orbital period of a stellar binary to impact linear-drift astrometry. As in the next section, many binary sources with periods below that value give reasonable fits to a constant-drift astrometric model (${\texttt{ruwe}}<1.4$).

The requirements for astrometric impact for a low-mass companion are even more stringent. From Equation (1),

$$T_{orb}\gtrsim 2.8\times\left[\frac{q}{0.002}\right]^{-3/2}\left[\frac{\delta\varpi/\varpi}{0.004}\right]^{3/2}\left[\frac{M}{1~\text{M${}_{\odot}$}}\right]^{-1/2}\text{yr}.$$ (2)

Shorter-period planets do not impact the observed motion of their host star as the star’s positional changes about the center of mass are too slight to be detected.

For long-period orbits, binary motion may be detected as an acceleration, depending on the resolution of proper motion $\vec{\mu}$ (Brandt 2018; Brandt 2021, e.g.,). By assuming that the orbital acceleration is nearly constant over the observation time $\Delta t$, the velocity deflection $\Delta v$ in this time frame is

$$\frac{\Delta v}{\Delta t}\sim\frac{GM_{\bullet}}{a^{2}}~.$$ (3)

To translate $\Delta v$ into a deflection in proper motion, $\Delta\mu$, we connect physical speed with the proper motion error, $\delta\mu$. Then, a linear drift model could discern a velocity deflection if $\Delta v\approx 2\delta\mu$, so long as $\Delta\mu$ stands out above expected errors in the fitting. From this estimate, we find an upper bound to the orbital separation,

$$a \lesssim \left[\frac{GqM_{\bullet}\,\Delta t}{D\,\delta\mu}\right]^{1/2}$$ (4)

$$\lesssim 120\,q^{1/2}\left[\frac{\varpi}{5~\text{mas}}\right]^{1/2}\left[\frac{\delta\mu}{0.04~\text{mas/yr}}\right]^{-1/2}$$

$$\ \ \ \times\ \left[\frac{M}{1~\text{M${}_{\odot}$}}\right]^{1/2}\left[\frac{\Delta t}{3~\text{yr}}\right]^{1/2}\text{{au},}$$

where $D\sim 1/\varpi$ is the source’s heliocentric distance. Converting this value to the companion’s orbital period, we estimate that for nearby equal-mass, Sun-like stars, binary motion impacts Gaia astrometry when the orbital period is less than about 950 years. When the binary mass ratio drops to $q=0.1$, the bound is about 200 years.

For planetary companions, the upper limit to the orbital period for astrometric impact is roughly

$$P_{pl} \lesssim 13\,\times\left[\frac{q}{0.002}\right]^{3/4}\left[\frac{\delta\mu}{0.04~\text{mas/yr}}\right]^{-3/4}$$ (5)

$$\ \ \ \left[\frac{\varpi}{5~\text{mas}}\right]^{3/4}\left[\frac{M}{1~\text{M${}_{\odot}$}}\right]^{1/4}\left[\frac{\Delta t}{3~\text{yr}}\right]^{3/4}\ \text{yr}.$$

A Sun-like star with a companion of mass $M_{\bullet}=2\text{M${}_{\text{J}}$}$ (where M${}_{\text{J}}$ is the mass of Jupiter) gives an upper bound close to the lower bound in Equation (2). The implication is that for nearby Sun-like stars measured by Gaia, astrometric quality indicators like ruwe are sensitive to planetary companions if the planets are roughly as massive as Jupiter, and located at orbital distances comparable to those of the Sun’s gas giants.

In formulating the results presented in this section, we are guided by Gaia binary star observations. We illustrate those data next. We also explore whether the above expectations for planet hosts are realistic.

3 Astrometric measurement of stars with companions

The quality of astrometric fits can be affected by observed characteristics of individual sources. Therefore, we first consider these characteristics for the stars that are of ultimate interest here, the sources in the Pearce et al. 2022 debris systems catalog that were observed by Gaia. We obtain Gaia DR3 identifiers of sources in the Pearce et al. 2022 catalog from SIMBAD (Wenger et al. 2000). All but two of the 178 stars are in the Gaia DR3 data set. The two stars missing in the archive are the A-type star $\beta$ Leo and the B-type star $\epsilon$ Sgr. Both are too bright for Gaia astrometry, with visual magnitude V $\approx$ 2, as reported in SIMBAD. The remaining stars constitute the 176 debris disk hosts that we study here.

Figures 1 and 2 show a Hertzspring-Russell diagram and a parallax-apparent magnitude plot, respectively, for the 176 debris disk hosts. The gray shaded regions in each plot constitute our “selection zone,” from which all sources used in the analyses provided here are drawn, to keep observational characteristics for all star similar to those of the 176 sources. The selection zone is defined so that sources satisfy these criteria. In terms of Gaia’s data model parameters, the criteria are

$$2\leq\texttt{phot\_g\_mean\_mag}\leq 12,$$ (6)

$$-0.25\leq\texttt{bp\_rp}<3.5,$$

$$\ \ \texttt{parallax}>4,\ \ \text{and}$$

$$-1.5\leq M_{G}-3.5\times\texttt{bp\_rp}<3.5,$$

which imply A-type to M-type main-sequence stars that are nearby (within 250 pc) and bright (brighter than 12${}^{\text{th}}$ magnitude in Gaia G-band).

Figure 1: Hertzsprung-Russel/color-magnitude diagram for 176 debris disk host stars (turquoise) and confirmed exoplanet hosts (goldenrod). The shaded region defines a “selection zone” so that when we compare collections of stars we only consider those that are predominantly on the main sequence, and with a range of color and magnitude similar to the set of 176 debris systems.

These figures include sources drawn from the NASA Exoplanet Archive (Ramirez et al. 2013, “NXA” hereafter), each of which hosts a confirmed exoplanet. We draw a subset of 1,117 NXA stars that meet the selection criteria in Equation (6), which will be part of our planetary companion analysis below.

Figure 2: Distance and apparent magnitude of debris disk host stars (turquoise) and exoplanet hosts from the NASA Exoplanet Archive (“NXA,” goldenrod). Distances for these nearby sources are detemined directly from the inverse of the parallax. A comparison with Bailer-Jones et al. 2021 distance estimates shows that when the latter are available (approximately half of the sources), the difference is small, well within a percent for these nearby sources.

Not shown in Figures 1 and 2 are three other catalogs we consider here. The first set is 1,492 binary stars from the U.S. Naval Observatory Double Star catalog, with derived orbital elements (Hartkopf et al. 2001; Mason et al. 2001, ORB6). The second set consists of 1,617 sources from the Ninth Catalog of Spectroscopic Binaries (Pourbaix et al. 2004, SBC), and the third contains 3,678 randomly selected stars in Gaia DR3 that have $\texttt{non\_single\_star}=0$. These random sources represent single stars, though they are better labeled “not-non-single stars,” in acknowledgment that Gaia’s non-single-star tables have strict admission standards that may miss stellar partners as well as planetary companions (Gaia Collaboration et al. 2023a). The stars identified in all three catalogs are in both SIMBAD and the Gaia DR3 archive, and all meet the criteria in Equation (6), They were further culled to limit their proper motion in Gaia DR3 to within 1,000 mas/yr, so that astrometric quality might be less affected by large changes in position over the course of Gaia’s observations.

We begin assessing the impact of stellar companions on astrometry with Figure 3, which shows the quality metric ruwe as a function of the orbital period. For reference, ${\texttt{ruwe}}\lesssim 1.4$ indicates a good fit to the model of a single star drifting with constant proper motion on a line across the sky (Halbwachs et al. 2023). The effect of the observational time baseline ($\Delta t$ in §2) on the sensitivity of ruwe to binary motion is clear. Binaries with orbital periods less than about a month are so close that their overall motion in the sky is fairly well-fit by a linear drift model about the center of mass. Stars with companions that have orbital periods exceeding a century may also appear to be linearly drifting, since their orbital motion is small over the 3-yr duration of Gaia observations. In between, in a “sweet spot” of orbital periods,

$$0.1~\text{yr}\leq P<100~\text{yr},$$ (7)

binary motion causes a significantly bad fit to the linear drift model.

Figure 3: Gaia’s astrometric quality measure, ruwe, as a function of binary orbital period. In a “sweet spot” of orbital periods ranging from 100 days to 100 years, ruwe is strongly sensitive to binary motion. Outside of the sweet spot, the binaries are unresolved or vary only by small amounts on the plane of the sky compared with the (linearly drifting) binary center of mass. The solid lines indicate the 16th, 50th and 84th percentiles in populations binned by orbital period.

In quantitative terms, stars in Figure 3 with stellar companions on orbits with periods between 0.1 and 100 years have a median ruwe value of 5.4. About 94% of this population have ${\texttt{ruwe}}>1.4$, near the threshold between a good and poor fit to linear drift. For stars in binaries with shorter periods, the fraction of objects with ${\texttt{ruwe}}>1.4$ drops to around 34%, while 54% of stars in longer-period binaries have ruwe above that threshold. Evidently, binarity may well impact the quality of a linear-drift astronomical solution even at short and long orbital periods (Brandt 2021; Castro-Ginard et al. 2024; El-Badry 2025, cf.). However, if we wish to use Gaia DR3 astrometry to detect or confirm binary motion, our sensitivity will be greatest for sources in the sweet spot with $0.1<T_{\text{orb}}/\text{yr}<100$. These conclusions are consistent with the assessments and discussion in the previous section (§2).

To further strengthen the connection between ruwe and binary motion, we identify a population of planet hosts in the NXA (exoplanet archive) that meet our selection criteria (Eq. (6)), obey the proper motion “speed limit” of 1,000 mas/yr (as above), and do not have SIMBAD indicators of stellar partners. Specifically, we admit only sources for which the NXA’s listed number of stars in a host’s planetary system is $\texttt{sy\_snum}=1$. We also select stars with Gaia DR3’s $\texttt{non\_single\_star}=0$. This flag has a high bar for admission, and is generally an indicator of a binary partner, a substellar companion, or even a massive giant planet (Stefánsson et al. 2025). Although we may miss potential discoveries by excluding sources with measured acceleration, our goal here is to focus on planetary systems that are not yet resolved by Gaia.

Furthermore, we require that SIMBAD’s object type list otypes does not contain a double star flag (**, SB*, or EB*). Lastly, we take the maximum mass ratio between planet and star to be less than 0.3 M${}_{\text{J}}$/M_⊙ to avoid the impact of a known giant planet on the star’s astrometry (§2). Our final tally of “solo” NXA stars is 388 sources.

The median ruwe value of solo NXA stars is within a percent of unity, and the fraction of stars with ${\texttt{ruwe}}>1.4$ is just under 3%. The $95^{\text{th}}$ percentile in ruwe is 1.3, and the maximum value is 1.94, corresponding to WASP-131, a $\sim$10 mag G0 star at about 200 pc from us (Ramirez et al. 2013). It hosts a 0.27 M${}_{\text{J}}$ planet (at the high end of the planetary masses for the solo stars) orbiting at about 12 R_⊙(equivalent to 0.3 mas).

We summarize the statistics of these stars and of known binaries in Table 1.

Table 1: Stastics of ruwe in binary stars

		ruwe percentile			ruwe	ruwe$>$1.4
Sample	N	5${}^{\text{th}}$	50${}^{\text{th}}$	95${}^{\text{th}}$	max	(pct.)
short-period	1894	0.89	1.16	7.21	49.19	34.0
long-period	669	0.93	1.55	18.81	37.31	54.1
sweet-spot	3417	1.27	5.39	23.49	65.45	93.6
not-nss	3679	0.82	1.02	4.32	49.30	16.4
solo NXA	388	0.84	1.00	1.30	1.94	2.8

Note. — Short-periods binaries have orbital periods of less than 0.1 yr, long-period binaries have periods exceeding 100 yr, and “sweet-spot” binaries have orbital periods in between. The “not-nss” stars are the random sample from Gaia with $\texttt{non\_single\_star}=0$, as described in the text. The solo NXA stars are selected to have no stellar or giant planetary companions.

The importance of working with sources like the 391 solo stars from the NXA that have been well-studied in terms of companions is underscored when we consider the “not-non-single” stars with $\texttt{non\_single\_star}=0$. In our random download from Gaia DR3, this set has a median ruwe value of 1.02, and over 16% of the sources have ${\texttt{ruwe}}>1.4$. The $99^{\text{th}}$ percentile and maximum ruwe values are at 16.6 and 49.5, respectively. We interpret these results to mean that Gaia’s non_single_star flag is intended to be selective, not all-encompassing (Gaia Collaboration et al. 2023b), so that our subsample is expected to contain some binaries.

These results are summarized in Figure 4, showing ruwe distributions for the NXA single stars, the Gaia non-single stars in the orbital-period sweet spot, and the “not-non-single” stars.

Figure 4: Histograms of Gaia’s astrometric quality measure, ruwe, for several star populations. The black/goldenrod histogram indicates the ruwe distribution for NXA single stars, with no evidence of a stellar or a gas giant companion. The violet/pink histogram corresponds to known binaries with orbital periods in the sweet spot. The power of ruwe as a discriminator of binary motion is evident in these distributions. The value of 1.4 (black dotted line) is the threshold value commonly adopted in the literature. The gray histogram is from Gaia stars with $\texttt{non\_single\_star}=1$, and the significant fraction (16%)of these sources with ${\texttt{ruwe}}>1.4$ is, we conjecture, the result of binaries in this population.

3.1 Comparison with predictions of astrometric quality

As previous work demonstrates, ${\texttt{ruwe}}>1.4$ is a strong indicator of binary motion (Stassun & Torres 2021; Penoyre et al. 2022; Halbwachs et al. 2023; Castro-Ginard et al. 2024). Our own assessment of the orbital configurations of stellar binaries to which ruwe is sensitive implicates a range from $10^{-3}$–$10^{3}$ years, with the caveat that the lower bound would need to be adjusted upward so that Gaia astrometric sampling could track coherent orbital motion, suggesting roughly 100 days. Figure 3 confirms that this lower bound is indeed appropriate, in general, even though approximately half of the binaries with shorter orbital periods have ruwe values above 1.4.

Our upper limit for orbital periods of binaries that may impact astrometry, about 1,000 years, seems validated in Figure 3. Just below $T_{\text{orb}}=100$ yr, ruwe values for binary stars are overwhelmingly above 1.4; for binaries with periods above about $10^{4}$ yr, the distribution is opposite. Our value of $T_{\text{orb}}=100$ yr for the large-period limit of the sweet spot is a conservative estimate.

3.2 Planetary companions

The NASA Exoplanet Archive contains over 6,000 confirmed planets around over 4,200 host stars, including 1,117 sources that match our selection criteria illustrated in Figures 1 and 2. This resource offers opportunity to explore the sensitivity of astrometric quality measures to the presence of a planet. Figure 5, showing ruwe as a function of the planet-to-star mass ratio $q$ for each host’s most massive planet offers a preview.

Figure 5: The astrometric quality parameter ruwe versus companion mass ratio $q$ for NXA planet hosts. The 1,117 sources shown all match our selection criteria (Figures 1 and 2). The value of $q$ corresponds to the most massive planet if a host has multiple planetary companions. The lighter colored points have planets with orbital periods in the “sweet spot”.

As above, we consider our 388 solo NXA sources that show no sign of having a binary stellar or massive planetary companion. We also look at sets of 367 “single” NXA stars that include a massive planetary companion but no stellar companion to identify trends in ruwe statistics. Table 2 highlights our results. When we divide the NXA stars into the4se two groups, solo stars versus single stars with massive planets (mass ratio $q>q_{\text{cut}}=0.3$ in units of M${}_{\text{J}}$/M_⊙), the ruwe distributions are similar. We cannot rule out that the populations are different on the basis of a two-sample Kolmogorov-Smirnov (KS) test. Nor does a test of the proportion of stars with ${\texttt{ruwe}}>1.4$, the proportion-$z$ (“prop-$z$”) test, differ statistically between these two samples. We conclude that planetary hosts with these demographics do not reveal themselves through ruwe statistics.

Table 2: Tests of ruwe distributions for planet hosts with a range of planetary masses.

mass ratio	star	ruwe$>$1.4	KS	prop-$z$
$q_{\text{cut}}$ (M${}_{\text{J}}$/M⊙)	count	count	$p$-value	$p$-value
0.3	367	17	0.2685	0.192
1	253	14	0.0660	0.085
2	180	11	0.0286	0.060
5	93	10	0.0071	0.001
10	43	7	0.0021	$<0.001$

Note. — The first column ($q_{\text{cut}}$) indicates the lower bound to a range of planet-to-star mass ratios ($q$). The second column is the number count of stars whose most massive planetary companion falls in that $q$ range. The third and fourth columns highlight the differences in the ruwe distribution of each set of stars as compared with NXA single stars with no giant planets. The third column contains a KS test $p$-values that measure the differences in the two ruwe distributions, while the fourth column is a prop-$z$ test assessing whether ruwe from the set of stars is different from the NXA singles with no giant planets on the basis of the number counts with ${\texttt{ruwe}}>1.4$.

The situation changes as we increase the minimum mass ratio $q_{\text{cut}}$. When the threshold value is equivalent to 2 M${}_{\text{J}}$ about a solar-mass star, the KS test suggests a significant difference between the ruwe distribution for single stars with a massive planet compared to solo stars ($p\text{-value}=0.029$). The prop-$z$ test gives a marginal $p$-value of about 6%. For a threshold equivalent to 5 M${}_{\text{J}}$, both tests show a significant difference between the two populations. At the highest value of $q_{\text{cut}}=10$ M${}_{\text{J}}$/M_⊙, the KS test $p$-value is about 0.2%, while the prop-$z$ test has a $p$-value less than $10^{-4}$. On the basis of the 43 stars with massive planets in this range (Fig. 6), we conclude that astrometric quality may be a potential signature of planet host candidates.

To check that the ruwe excess among hosts of massive planets might stem from apparent magnitude, parallax or proper motion selection, we binned according to objects above or below the median values in these quantities. The ruwe distributions of these subsets show variations. Closer sources tend to have more ruwe outliers compared with the NXA single stars with no giant planets, while the proper motion and brightness selection do not affect the ruwe distributions as much. We caution that number of sources is small in some of these subsets, particularly for hosts with massive planets.

Figure 6: Histogram of ruwe for groupings of host stars with and without stellar partner. The stars with $q<0.3\text{M${}_{\text{J}}$}/\text{M${}_{\odot}$}$, indicated by the histogram with gold hue, are the same as in Figure 4. Stars with binary periods in the sweet spot are shown in the purple shaded histogram, while the randomly selected population of “not non-single stars” in Gaia appear in the light gray histogram.

3.3 Beyond RUWE: a machine learning approach

While ruwe provides a useful summary diagnostic of excess astrometric scatter, the Gaia DR3 archive has additional information that may help to identify unseen bound companions. For example, orbital motion may increase uncertainties in parallax and proper motion, possibly introducing tell-tale correlations in those errors that are not included in ruwe. We therefore consider a suite of parameters including parallax and proper motions, the errors in each parameter, and their mutual error correlations (e.g., pmra_pmdec_corr). We also add to the analysis error correlations between proper motion components and sky position along with Gaia G-band brightness and BP–RP color.

To explore whether this broader list of features can reveal companions to stars in Gaia, we adopt a machine-learning (ML) approach. We began with Python-based Scikit-Learn’s random forest algorithm applied to a training set of the 755 stars in the NXA for which there is no evidence of stellar binarity. About a third of this group has massive planets in the sweet spot of orbital period and the remaining objects are “solo,” with lower-mass planets. We performed train-test trials with an even random partition of this set, with subsets of the features listed above. The full list of features performed best, with train-test trials yielding about an 80% success rate, nominally better than random. We also tested a neural net classifier on the same features to identify if a neural net would perform better. It did not.

To generate our final ML model, we trained the random forest classifier and a set of about 1,000 stars, with some random duplicates to increase the number of sweet-spot hosts to be equal to the number of solo stars. This model is the basis of our analysis of the Pearce et al. 2022 catalog.

4 178 debris systems (minus two)

Pearce et al. 2022 assess properties of 178 debris disks to place limits on the mass of potential planets. In this section, we explore the 176 sources in Gaia DR3 to evaluate whether there is also evidence of massive planets from Gaia astrometry. Figure 7 shows ruwe and the orbital period of potential massive planets from Pearce et al. 2022. The orbital periods of the candidates are largely within the sweet spot for potential impact on ruwe. This coincidence motivated this work.

Figure 7: The orbital period of potential planets and ruwe of the 176 debris systems from Pearce et al. 2022. The turquoise points correspond to these sources, while for reference, points from the NXA planet hosts (goldenrod) and binary stars (faint light gray) are also shown. The debris systems’s inferred planetary orbital periods lie in the “sweet spot,” although at higher values, typically, than the planet hosts.

To hone in on planetary companions, we check whether sources are associated with known stellar binaries using SIMBAD otypes and Gaia’s non_single_star flag. We identify 55 potential stellar binaries. The remaining 121 stars, “solo” debris systems, have ruwe distributed as in Figure 8. Eleven sources (9.1%) have ruwe above 1.4, which is a significant excess compared with the solo NXA stars (proportion-$z$ test $p$-value of 0.003).

Figure 8: Histogram indicating the distribution of ruwe values in the debris system stars. The blue-green shaded histogram refers to these stars, all of which show no evidence of a stellar companion, as described in the text. The yellowish shade histogram corresponds to solo NXA stars. The excess source count above the ${\texttt{ruwe}}=1.4$ line suggests that the debris systems contain promising candidates for planetary hosts.

Among the Pearce et al. 2022 candidates, 15 were known planet hosts. Since then, two other planetary systems have been discovered in the list of 178: AF Lep (HD 35850), and HD 114082. We also count HN Peg as a planet host though Pearce et al. 2022 do not (HN Peg b is listed in the Exoplanet Archive as a 20 M${}_{\text{J}}$ body, but it is well outside of the debris disk around the host star). Five of the 18 planet hosts have massive planets in the sweet spot of orbital periods. These five, and two others (not in the sweet spot) around “solo” stars, are included in the ML training set described in §3.3.

Table 3 lists the top 25 individual debris systems ranked by high ruwe values. All have bright stellar hosts, with G = 2–6. Because detector saturation or stray light may affect astrometry, we check consistency between brightness in G compared to blue ($B_{P}$) and red ($R_{P}$) bands using the color excess factor $C_{\ast}$, defined in Riello et al. 2021. The top four brightest stars ($\alpha$ CrB, $\beta$ UMa, $\gamma$ Boo and $\tau$ Cet) in the table have anomalous color excess ($|C_{\ast}|>0.05$), perhaps from blending (three of these sources are binaries) or detector saturation ($\texttt{phot\_g\_mean\_mag}\leq 3.3$ mag). All other stars for which we have Gaia astrometry (including those not listed in the table) show no significant color excess.

Sometimes, a high proper motion impacts the overall quality of astrometry. Five sources, including two in Table 3 ($\tau$ Cet and $\epsilon$ Eri), have proper motions close to or over 1,000 mas/yr. Of these five systems, only one (BD-07 4003), with ${\texttt{ruwe}}=1.303$, is not flagged as a possible binary in Gaia (through the non_single_star flag) or SIMBAD (through the object type field, including “**”).

Alongside ruwe, the machine learning classifier from §3.3 provides a separate measure of the probability of a planetary companion. Table 4 lists the top 25 most probable candidate planet hosts. As in Table 3, this table includes columns indicating known companions. Five of the top 25 sources have massive planets in the sweet spot of orbital periods, and they all appear in the table with formal planet host probabilities above 80%. The stars with the highest probability of being planet hosts, according to our ML algorithm (the top three rows in the table), are indeed planet hosts. This result is not an indicator of extraordinary success. We used these five known planet hosts in our ML training set, so that the ML classifier recovered them as it was designed to do.

Table 3: Debris systems with the highest ruwe values

SIMBAD	$\varpi$	$\mu$	$G$	$B_{P}$-$R_{P}$		companions	$M_{p}$	$P_{p}$
main_id	(mas)	(mas/yr)	(mag)	(mag)	ruwe	stars planets	(M${}_{\text{J}}$)	(yr)
V* DE Boo	85.99	491.7	5.766	1.041	16.496	1 0	0.2	750.07
alf CrB	42.24	147.8	2.269	0.530 x	10.338	1 0	0.09	3.20
bet UMa	38.60	86.3	2.399	0.494 x	5.991	0 0	0.2	55.58
zet Lep	44.79	14.2	3.541	0.244	4.317	0 0	0.24	53.14
eta Lep	66.86	145.3	3.636	0.539	4.160	0 0	0.08	48.75
10 Tau	71.84	534.7	4.141	0.758	3.063	0 0	0.03	16.24
eps Pav	31.26	154.9	3.963	-0.017	3.007	0 0	1.2	107.41
gam Oph	33.62	78.2	3.753	0.109	2.879	0 0	0.6	298.56
eps Eri	310.58	975.0	3.466	1.140	2.717	1 1	0.19	422.47
iot Psc	73.24	576.6	3.999	0.705	2.640	1 0	0.07	151.26
tau Cet	273.81	1922.3	3.300	1.051 x	2.634	1 4	0.013	16.86
kap Phe	41.72	111.2	3.919	0.262	2.610	0 0	0.24	119.08
gam Boo	37.91	190.0	3.022	0.480 x	2.589	1 0	0.6	779.57
lam Boo	32.59	246.2	4.169	0.142	2.445	0 0	0.1	11.69
30 Mon	25.34	70.7	3.902	0.011	2.155	0 0	0.4	58.72
gam Dor	48.95	209.8	4.188	0.455	2.008	0 0	0.07	34.17
sig And	23.25	77.7	4.502	0.095	1.971	1 0	0.3	116.19
bet Cir	33.82	167.2	4.061	0.137	1.910	1 0	0.21	53.21
eta Crv	54.81	428.6	4.210	0.526	1.897	1 0	0.34	899.56
d Sco	23.54	106.3	4.782	0.026	1.619	0 0	0.4	85.05
pi.01 Ori	27.01	134.7	4.636	0.128	1.477	1 0	0.5	205.87
HD 218396	24.46	119.3	5.911	0.394	1.474	1 4	2.5	1042.46
nu. Phe	65.53	688.5	4.828	0.728	1.381	0 0	0.08	146.97
HD 146897	7.57	28.2	8.976	0.678	1.379	0 0	2	245.42
mu. Cet	37.59	281.7	4.154	0.441	1.344	1 0	0.3	451.84

Note. — The first column provides the literal main source identifier in SIMBAD, while the second column ($\varpi$ = parallax) through the sixth column (ruwe) are all from Gaia DR3. In the Gaia color index column (bp_rp), an ’x’ marks sources that have a color excess factor $|C_{\ast}|$ that is greater than $0.05$, indicating a potential anomaly (Riello et al. 2021). The two “companions’ columns (“stars” and ”planets”) indicate known companions. We include an “*” in cases where SIMBAD has “**” as an object type, but NXA sy_snum indicates a single star. The second-to-last and last columns are the minimum planet mass and orbital period from Pearce et al. 2022.

Table 4: Debris systems highlighted by machine learning

SIMBAD	$\varpi$	$\mu$	$G$	$B_{P}$-$R_{P}$		companions	$M_{p}$	$P_{p}$
main_id	(mas)	(mas/yr)	(mag)	(mag)	ruwe	stars planets	(M${}_{\text{J}}$)	(yr)	prob.
HD 50554	32.185	103.098	6.715	0.729	1.002	0 1	0.08	248.06	0.99
HD 114082	10.520	44.498	8.106	0.568	0.944	0 1	1.10	85.11	0.99
q01 Eri	57.641	196.720	5.383	0.711	0.875	0 1	0.16	332.59	0.97
V* V1358 Ori	19.353	43.465	7.707	0.724	0.930	0 0	0.20	10.08	0.91
tau01 Gru	30.192	232.513	5.887	0.772	1.004	0 1	0.08	220.19	0.89
V* DK Cet	24.205	115.255	7.914	0.835	0.967	0 0	0.32	40.76	0.84
kap Phe	41.723	111.244	3.919	0.262	2.610	0 0	0.24	119.08	0.83
HD 127821	31.558	177.419	5.998	0.574	0.974	0 0	0.21	379.39	0.83
HD 209253	31.804	30.261	6.505	0.668	1.103	0 0	0.30	439.14	0.82
V* AF Lep	37.254	52.138	6.210	0.736	0.918	0 1	1.10	184.46	0.80
eta Lep	66.857	145.268	3.636	0.539	4.160	0 0	0.08	48.75	0.79
HD 38397	18.659	29.129	7.999	0.737	1.062	0 0	0.60	154.57	0.78
HD 107146	36.404	229.612	6.904	0.793	0.993	0 0	0.44	264.37	0.77
HD 8907	29.941	111.074	6.537	0.660	1.058	0 0	0.28	289.66	0.77
gam Dor	48.949	209.816	4.188	0.455	2.008	0 0	0.07	34.17	0.76
8 Dra	33.967	30.347	5.145	0.429	1.137	0 0	0.22	293.60	0.75
zet Lep	44.794	14.224	3.541	0.244	4.317	0 0	0.24	53.14	0.75
eta Crv	54.813	428.573	4.210	0.526	1.897	1 0	0.34	899.56	0.75
V* NZ Lup	16.627	78.557	7.794	0.834	0.902	0 0	1.30	168.21	0.73
HD 218340	17.815	121.364	8.295	0.787	1.091	0 0	0.40	1259.06	0.73
HD 61005	27.434	92.487	8.017	0.911	0.940	0 0	0.70	273.70	0.71
mu. Cet	37.592	281.673	4.154	0.441	1.344	1 0	0.30	451.84	0.71
V* IS Eri	25.825	144.096	8.300	0.931	0.951	1 0	0.15	37.63	0.69
HD 48370	27.815	72.562	7.746	0.874	1.112	0 0	0.40	91.28	0.69
alf CrB	42.241	147.773	2.269	0.530 x	10.338	1 0	0.09	3.20	0.67

Note. — The columns are the same as in Table 3, with the addition of the machine-learning predicted likelihood that the source hosts a massive planet in the sweet spot of orbital periods.

Just over half of the list of high-ruwe stars in Table 3 are double stars. As a flag for binarity, ruwe is playing its part in the table. The list also contains three known planet hosts. The remainder, including six of the top ten in ruwe, are putatively single stars that are good candidates as substellar or planetary hosts. Table 4, with the top ML probabilities for planet hosts, contains many fewer double stars (4 of 25). It lists five planet hosts, including the top three entries, though only because these sources were in the ML training set. Seven stars are in both top-25 tables, four of which have no known binary partner. We consider these sources below.

In both tables, it is striking that the minimum planet masses from the Pearce et al. 2022 analysis are mostly less than one Jupiter mass. As in §2, the detection of a planetary companion with astrometry typically requires a larger planetary mass. For the few candidates of interest that we consider next, we assume that the putative planets are indeed massive enough to impact the astrometry of their host star. As we mention below, this assumption breaks down at the population level; it is unlikely that most of the 178 debris systems host a planet more massive than Jupiter (Yee & Kenyon 2025)

4.1 Candidates of interest

From Tables 3 and 4, we identify several debris systems as promising potential planet hosts, starting with the four single stars that are in both top-25 tables.

• $\zeta$ Leporis (HD 38678) is a bright (G = 3.5) A2 IV-V(n) star with a ruwe of 4.31. The predicted planet has a mass of 0.24 M${}_{\text{J}}$ and orbital period of 53 years. This planet is well within the sweet spot, but for an astrometric detection, the planet would have to be closer to Jupiter-mass or larger.

• $\eta$ Leporis (HD 40136) is a bright (G = 3.63) F2 V star with a predicted 0.08 M${}_{\text{J}}$ planet in the sweet spot (Pearce et al. 2022, orbital period $\sim$ 50 years;). Although its brightness may be a factor in the large ${\texttt{ruwe}}=4.2$, a planetary companion with a mass of at least an order of magnitude larger than predicted could be sculpting the $\eta$ Lep debris disk.

• $\kappa$ Phe (HD 2262) is an A5 IV(n) star with G = 3.92 and a ruweof 2.6; the nominal planet has a minimum mass of 0.24 M${}_{\text{J}}$ and an orbital period just over a century. The predicted orbital period lies outside the sweet spot; Pearce et al. 2022 note that a larger planet could reside closer to the star, in the sweet spot.

• $\gamma$ Doradus (HD 27290) is an F1 V star with G = 4.19 and ${\texttt{ruwe}}=2.0$. Pearce et al. 2022 predict a minimum mass for a sculpting planet of 0.1 M${}_{\text{J}}$and a maximum orbital period of 34 years. They suggest that there is flexibility in these estimates as the constraints on disk geometry are not tight.

Of these sources, the most promising candidate from an astrometric detection standpoint is $\zeta$ Lep; its predicted planet has the largest minimum mass of about 0.25 M${}_{\text{J}}$ at an orbital period of 53 years. The predictions for the other stars include either a smaller minimum mass or a larger orbital distance. In any case, all four are worthy of a time-series analysis in Gaia DR4 to investigate the source of their unusually high ruwe values.

Other candidates of interests with high ruwe include the following:

• 10 Tauri (HD 22484) is a bright, single F9 IV-V star with G = 4.14 and ruwe= 3.063, but it has a slightly lower ML probability of being a planet host (59%). The Pearce et al. 2022 prediction is a minimum planet mass of roughly 10 Earth masses and a maximum orbital period of 16 years.

• 30 Moncerotis (HD 71155) is a bright (G = 3.90), single A0 Va star with ruwe of 2.155. The predicted minimum planet mass and orbital period are 0.5 M${}_{\text{J}}$ and about 60 years. 30 Mon is one of the higher-mass stars in debris disk systems catalog, which may make it more challenging for astrometric detections of planets.

• d Scorpii (HD 146624) is another single, intrinsically luminous A1 Va star with G = 4.78, ruwe =1.619, a predicted planet mass $\sim$ 0.5 M${}_{\text{J}}$  and a planetary orbital period of 85 years. Like 10 Tau and 30 Mon, this source did not make both top-25 lists because its ML probability score for planet hosting is 0.59, below the list cut-off.

The machine-learning classification in Table 4 highlights sources with other attributes, including demographics and astrometric indicators of companions other than ruwe. Here is a sample of stars with no known companions and predicted planetary orbits from Pearce et al. 2022 in the sweet spot:

• V1358 Orionis (HD 43989) is a G0 V star (G = 7.71) with a low ruwe (0.930). Pearce et al. 2022 infer a 0.3 M${}_{\text{J}}$ (or greater) planet on a 10-year orbit as the source of debris disk sculpting.

• DK Ceti (HD 12039) is somewhat cooler G4 V star with G = 7.91 and ruwe = 0.967. The predicted minimum planet mass is similar to that for V1358 Orionis, although the orbital period is longer (30 years).

• HD 48370 is an even cooler G8–K0 V star with G = 7.75 and ruwe = 1.1. The minimum planetary mass is 0.4 M${}_{\text{J}}$on a 90-year orbit.

In this list, the ML classifier focuses on attributes other than ruwe, since high ruwe values account for a minority of actual planet hosts. A risk with the application of ML in this context is that it more strongly weights features that factor into the demographics and selection of known planet hosts, and less into astrometrical information. However, this aspect of the machine-learning algorithm may be important, for how the astrometry behaves in the presence of a planetary companion will depend on stellar characteristics.

Among these candidates, most have estimated ages ranging from $\sim$200 Myr up to $\sim$5 Gyr (Pearce et al. 2022). However, V1358 Ori ($\sim$40 Myr), DK Cet ($\sim$45 Myr), and HD 48370 ($\sim$60 Myr) are much younger, have large 70–100 $\mu$m excesses, and are members of nearby moving groups (Gagné et al. 2026, see also). Approximately 6% of Spitzer and Herschel debris disk targets belong to a moving group (Kenyon et al. 2026); young systems with such large cold dust luminosities merit special attention once Gaia DR4 is released.

Next, we note two known planet hosts that were not part of the ML training set (they were both listed in SIMBAD as potential double stars) to assess how our analysis fared:

• $\epsilon$ Eri (HD 22049) is a K2 V star (G = 3.47) that hosts a 0.7 M${}_{\text{J}}$ planet in the sweet spot: a period of 7 years (Hatzes et al. 2000) and a semimajor axis of roughly 3 au. The system contains warm dust at 3–20 au (Su et al. 2017; Booth et al. 2023) and cold dust at $\sim$ 70 au (Beichmann 1985). Pearce et al. 2022 propose a second planet with a mass of at least 0.2 M${}_{\text{J}}$ near or within 50 au to account for stirring of the cold disk. SIMBAD lists this star as a double, though the Exoplanet Archive shows no bound stellar companion, leaving the known planet as the potential source of $\epsilon$ Eri’s high ruwe of 2.7.

• HR 8799 (HD 218396), an F-type $\gamma$ Doradus pulsating variable (Sódor et al. 2014, V342 Peg;) depleted in iron-peak elements (Gray & Kaye 1999, $\lambda$ Boo star;), hosts four known giant planets and a broad cold debris disk centered at about 200 au (Marois et al. 2008). The innermost planet with a mass of $\sim$ 10 M${}_{\text{J}}$, has an orbital period ($\sim 50$ years), well inside the sweet spot (Marois et al. 2010). It is reassuring that this system was flagged as a planet host both by its ruwe value (1.474) and the ML classification (59%).

The planet host AF Lep (used in our training set) has a single planet, AF Lep b, that was discovered after Pearce et al. 2022 completed their analysis. Their prediction of a 1.1 M${}_{\text{J}}$ planet at $a\approx$ 36 au with an orbital period of $\sim$ 200 yr differs from the derived parameters (Mesa et al. 2023, mass $\sim$ 3 M${}_{\text{J}}$, projected separation $\sim$ 9 au, and orbital period $\sim$ 25 yr;). The planet is probably not responsible for features in the cold debris disk at 40–50 au.

5 Discussion and conclusion

Our main goal is to find astrometric signatures of planets around stars for which there is other evidence of planetary companions. Because planets can sculpt dusty debris, the debris disk systems discussed in Pearce et al. 2022 offer an excellent opportunity to explore this possibility. We start with criteria for identifying companions in Gaia DR3 based on stellar binaries. We assess how these criteria fare as indicators of exoplanets using stellar hosts drawn from the NASA Exoplanet Archive. In both cases, we consider only sources that are similar to those stars in parallax, brightness, and color. While this process reduces the number counts substantially (only 755 of over 6,000 stellar hosts are similarly bright and close as the debris disk systems), it adds some assurance that measures of astrometric quality are not the result of selection effects.

We confirm a clear connection between the Gaia astrometric quality indicator ruwe and the orbital period of known binary stars. For periods in a sweet spot between days and centuries, approximately spanning the duration of Gaia observations, binary motion impacts astrometric quality as expected (§2 and Fig. 3 above; see also Halbwachs et al. 2023). Astrometric quality is not a perfect identifier of binary systems. Roughly 7% of the binary stars in the sweet spot have ${\texttt{ruwe}}<1.4$; perhaps the geometry and orientation of their orbits fail to impact fitting with a linear drift model. Conversely, putative single stars may have high ruwe values, although these may have unidentified massive companions. Despite these concerns, in a set of well-studied planet hosts with no known companion in the sweet spot (solo NXA stars), fewer than 3% have ${\texttt{ruwe}}>1.4$ (Table 1).

Because stellar reflex motion from planetary companions is typically much smaller than from stellar companions, astrometric quality is less effective for planet detection. Still, on a population level, significantly more stars that host massive planets with orbital periods in the sweet spot have higher ruwe than the “solo” stars with no known massive companions in that orbital regime (Table 2). The fraction of high-ruwe stars with massive planets ($q\geq 10$ M${}_{\text{J}}$/M_⊙) is $\sim$ 16%, compared with 3% for solo stars. The correspondence between astrometry and planets has been successful is identifying several planetary systems (Currie et al. 2020; Currie et al. 2023; Currie et al. 2026; Stefánsson et al. 2025; Vioque et al. 2026, e.g.,). Yet, the majority of planet hosts have low values of ruwe, a reminder that astrometric signatures of planetary companions do not always rise far above Gaia’s detection threshold.

With this background, we consider the 176 debris disk systems listed by Pearce et al. 2022 that are also in the Gaia DR3 archive. Of the 121 sources with no indicator of a stellar companion, 11 have ruwe above the 1.4 threshold. As with the NXA stellar hosts, this distribution suggests a population with massive planets. On a source level, we identify several promising candidates on the basis of their excess ruwe values and the high probability assigned by our machine-learning analysis that they are planet hosts. Since the machine learning algorithm mixed stellar demographics with the astrometry, it provides a separate view of potential hosts than ruwe alone.

With lower limits for the masses of planet candidates in Pearce et al. 2022, it is tempting to consider the impact of larger planet masses closer to the host inside the debris disks in these systems. However, the measured mass functions of planets from direct imaging, microlensing, and radial velocity measurements place strong constraints on the Pearce et al. 2022 planet candidates (Yee & Kenyon 2025). In particular, making all of these candidates significantly more massive creates a strong tension with the microlensing mass function, which in turn roughly agrees with the mass functions of directly imaged and radial velocity planets (Fulton et al. 2021; Vigan et al. 2021; Yee & Kenyon 2025). While several of the debris disk planet candidates could be more massive than the Pearce et al. 2022 estimates, it is unlikely that many are more massive. Thus, the highest probability systems derived here are the most likely planet candidates in the Pearce et al. 2022 sample, based on ruwe and the ML analysis.

To compare these results with giant planet frequencies derived from direct imaging and radial velocity studies, we estimate the implied frequency of giant planets. Some stars in the Pearce et al. 2022 sample are A-type stars, where the frequency of debris disks is $\sim$ 33% (Su et al. 2006, e.g.,). Most are FGK stars, which have a somewhat lower debris disk frequency $\sim$ 20% (Bryden et al. 2009; Carpenter et al. 2009; Eiroa et al. 2013; Sibthorpe et al. 2018, e.g.). If we adopt 25% as the intrinsic debris disk frequency of the mix of stars in Pearce et al. 2022, the 10% frequency of Pearce et al. 2022 single stars with ${\texttt{ruwe}}>1.4$ suggests an occurrence rate of $\sim$ 2%–3% for giant planets in the sweet spot of Gaia orbital periods. This estimate compares well with direct imaging and radial velocity estimates: Fernandes et al. 2019 report a 1–2% occurrence rate of 1–20 M${}_{\text{J}}$] planets at 10–100 au. Including smaller orbital distances ($<10$ au) allows larger rates, 5–10%, for Jupiter mass planets (e.g., Fulton et al. 2021; Vigan et al. 2021, and Table 2, above).

This exploration illustrates the potential and limitations of astrometry for planet discovery. Summary astrometric quality indicators like ruwe are valuable flags of substellar or planetary companions, yet they are not designed to reveal reflex motion, the signature of a star with a massive companion. Adding in additional astrometric information by including measurements from outside of the Gaia observing time frame, as in the Tycho-Gaia Astrometric Solution catalog (Michalik et al. 2015), can extend the sensitivity for long-period planets. Yet Gaia alone has a wealth of data waiting in the wings. With its time-resolved astrometry, Gaia Data Release 4 will provide snapshots of stars, dots along their orbital paths (Brown 2025, e.g.,). With the release of DR4 just around the corner, we are eager to see how those dots connect.