In this paper
AbstractIntroductionExisting toolsModelAstrometric orbitProjection through the single-star fitRadial velocitiesRelative astrometryLikelihood and priorsAstrometric jitterAcceleration and scan-angle termsInferenceSignal screenPeel and joint fitSamplersModel comparisonDiagnosticsValidation on simulated dataA commensurate two-planet systemAstrometric jitterInjection–recovery campaignDetection threshold and false positivesOrientation priorPeriods near one year and high eccentricityTwo companionsOrbit or accelerationUpper limitsApplication to the Gaia DR4 pre-release epoch astrometryDiscussion and conclusionsSummary of resultsLimitations and scope of the validationOutlookReferencesCorner plotsWorksPhysics and Astronomy
ExoPlanetLab: Bayesian orbit fitting of Gaia-DR4 epoch astrometry, radial velocities and relative astrometry
Abstract
Gaia Data Release 4 (DR4) provides epoch astrometry from which thousands of giant planets and many substellar and stellar companions can be characterised. We present ExoPlanetLab, an open-source package for Bayesian orbit inference from Gaia epoch astrometry, radial velocities (RVs) and relative astrometry, validated on simulated and real data. The model orbit is compared with the epoch residuals after removing the part absorbed by the single-star fit. A Student-t likelihood describes heavy-tailed data. A robust Keplerian periodogram with bootstrap false-alarm probabilities gives starting values for a joint fit. Five samplers are available: tempered ensemble MCMC, two nested samplers and two reversible-jump samplers. Evidences for a fixed number of companions select that number. Acceleration and scan-angle models test whether an orbit is warranted. The samplers agree within 0.1 sigma on a simulated two-planet system. In 180 injections all fits converge, and all signals with signal-to-noise ratio S >= 10 are detected. The posteriors show no significant bias, but the period coverage is below nominal: 90% of the true periods lie within their 95% credible intervals. None of 30 companion-free data sets gives a false detection. On DR4 pre-release data the orbit of Gaia BH3 is recovered within 1.4 sigma of published solutions. The Student-t model brings the astrometric period and eccentricity of HD 114762 onto its RV orbit. Eight simulated relative positions narrow the period's credible interval to 0.27 of its width, and Octofitter agrees within 0.2 sigma on the same data. The code is validated for single sources with DR4-like epoch astrometry, with or without RVs. Relative astrometry is tested on simulated data only. The main limits are an uncorrelated noise model and the computing cost, which grows linearly with the number of measurements and steeply with the number of companions.
Introduction
A star with an unseen companion moves on the sky around the system barycentre. Gaia measures that motion one dimension at a time, along the instantaneous scan direction, with a per-transit precision of tens of micro-arcseconds for bright stars [15]. Gaia DR3 published orbital solutions for about stars [13, 19, 22]. The epoch data behind them was not released. DR4 publishes the epoch astrometry of all sources. These data are predicted to yield about astrometric orbits, most of them of binary stars [9], and of the order of giant planets. Lammers & Winn [27] predict 7500 2100 giant planets in DR4 and 120 000 22 000 in DR5. Perryman et al. [33] predicted 21 000 6000 and 70 000 20 000, but their criteria also count planets with a poorly constrained period. With those criteria the DR4 estimate of Lammers & Winn [27] is about 16 000. Tens of thousands of brown dwarfs come in addition [21].
Fitting an orbit to Gaia epoch astrometry differs from the RV and direct-imaging problems for which most orbit codes were written. The data are one-dimensional projections along a scan angle that changes from transit to transit. The published epoch values are residuals of a single-star fit (position, parallax, proper motion), which absorbs part of any long-period orbit. When the period exceeds the time span of the data, only an arc of the orbit is seen, and many orbits fit it equally well. The posterior is then far from Gaussian, with long tails in period and separate modes in orbital phase. Combining the astrometry with ground-based RVs, which may extend decades beyond the Gaia baseline, removes the degeneracy. It adds instrument offsets and jitters, and a sign ambiguity between and . Several codes now address parts of this problem (Sect. 2). ExoPlanetLab was written to study it with one likelihood and several samplers, so that results that depend on the sampler can be separated from properties of the data.
Sections 3 and 4 describe the model and the inference pipeline. Section 5 validates them on simulations, and Sect. 6 on the Gaia DR4 pre-release epoch astrometry. Section 7 summarises the results and discusses the limitations.
Existing tools
Table 1 summarises the main publicly available codes that fit orbits to absolute astrometry, based on their papers and documentation. kima (Faria et al. [10]) has recently been extended with models for epoch astrometry alone and jointly with RVs, sampled with diffusive nested sampling with a free number of Keplerians (Baycroft et al. [2]); it fits the five single-star parameters (optionally seven or nine, with acceleration terms), an astrometric jitter, an optional Student- likelihood, and scan-angle harmonic terms to discriminate spurious signals, and has been applied to Gaia BH3. kepmodel (Delisle & Ségransan [8]) models astrometric and RV time series with Keplerians, linear predictors and Gaussian-process noise, and provides a linear periodogram with analytical false-alarm probabilities. Octofitter (Thompson et al. [37]) fits relative astrometry, RVs, Hipparcos intermediate astrometric data (IAD), proper-motion anomalies, images and interferometric visibilities, with Hamiltonian Monte Carlo or parallel-tempered sampling, a variable number of planets, optional -body integration, and preliminary support for DR4 epoch astrometry. orvara (Brandt et al. [7]) combines RVs, relative astrometry and the Hipparcos–Gaia Catalog of Accelerations (Brandt [6]); orbitize! (Blunt et al. [5, 4]) targets directly imaged companions and adds Hipparcos intermediate data and -body modelling; BINARYS (Leclerc et al. [29]) combines Hipparcos and Gaia data with relative astrometry and RVs for binaries. gaiamock (El-Badry et al. [9]) simulates Gaia epoch astrometry from the scanning law and fits it with the DR3 hierarchy of single-star, acceleration and orbital solutions, and the DR3 orbital pipeline itself is described by Holl et al. [22].
The main methodological differences between ExoPlanetLab and these packages are threefold. (i) Before comparing a model orbit with the published residuals, it removes from the orbit’s along-scan signal the part that the single-star fit would have absorbed, by applying the same linear fit, with the same weights, that produced the residuals (Sect. 3.2). (ii) The same likelihood can be sampled with five different algorithms, which exposes sampler-dependent failures and has been essential for the multi-planet case (Sect. 5). (iii) It includes a Keplerian periodogram with bootstrap false-alarm probabilities and a cross-refinement step for commensurate periods (in a ratio of small integers, such as 2:1). The code provides an opt-in Hipparcos IAD mode (with warnings), validated against the Hipparcos catalogue, injections and Octofitter on HD 114762 (described in the user guide that comes with the code; see Code availability). By default, however, the pipeline operates on epoch astrometry, optionally joint with RVs and relative (imaging) astrometry (Sect. 3.4): it avoids combining multi-catalogue proper-motion anomalies (such as the HGCA), whose cross-mission systematics are complex, so that all orbital constraints derive from observational data whose error models can be evaluated directly. Acceleration solutions are included (Sect. 3.7): they decide whether an orbit is warranted when the period exceeds the time span of the data.
Model
Astrometric orbit
Each orbiting companion is described by seven parameters . With , and , the photocentre displacement is
in the standard Thiele–Innes convention (orbital-plane axis to the north), and the along-scan signal at scan angle is
and with several companions the signals of the individual companions add. The Campbell elements follow from the constants (Green [18]). Because the constants enter linearly, no trigonometry of the orientation angles is evaluated in the likelihood, and the periodogram of Sect. 4.1 can solve for them in closed form.
Projection through the single-star fit
The DR4 epoch values are not the measured along-scan positions themselves but the residuals, observed minus computed (O−C), of the single-star astrometric solution. Let be the measured along-scan positions, the design matrix of the single-star model (position, parallax, proper motion and a pseudo-colour term with prior precision ; it is usually called in papers, a symbol we keep for a Thiele–Innes constant), and the weights. The pipeline estimates the single-star parameters by weighted least squares,
and publishes the residuals
If the star has a companion, , with the true single-star parameters, the along-scan orbit signal of Sect. 3.1 and the noise. The single-star solution knows nothing of the orbit and absorbs into whatever part of it resembles a change of position, parallax or proper motion. To see what is left, substitute the solution for into :
Applied to any vector of along-scan values, returns the best-fitting single-star curve (it is the hat matrix of the fit; Hoaglin & Welch 1978), so returns the O−C of a single-star fit to that vector. A pure single-star signal is fitted exactly, (for ; the prior is negligible in practice, Sect. 6), and therefore
The model to compare with the published residuals is not the orbit but , i.e., the part of the orbit that a single star can mimic has been removed from the data and must be removed from the model in the same way. In terms, is what one obtains by running the single-star solution, with the pipeline’s weights, on the noise-free Notes. Preliminary support. Simulation and DR3-style fitting. Removed from the likelihood by the projection and recovered for each posterior draw (Sect. 3.2). Through htof. In separate models (RVs with the Hipparcos–Gaia proper-motion anomaly; Gaussian processes for RVs), not in the Gaia epoch-astrometry models. Maximum likelihood with errors from the Hessian; no sampler. As user-defined linear terms.
| ExoPlanetLab | kima | kepmodel | Octofitter | orvara | orbitize! | BINARYS | gaiamock | |
| Language | Python | C++/Python | Python | Julia | Python | Python | R/C++ | Python |
| Gaia epoch astrometry | • | • | • | •^a | – | – | – | •^b |
| Single-star terms | projected^c | fitted | fitted | fitted | – | – | – | fitted |
| Hipparcos IAD | • | – | – | • | •^d | • | • | – |
| HGCA / PM anomaly | – | •^e | – | • | • | • | – | – |
| Relative astrometry | • | – | – | • | • | • | • | – |
| Multi-instrument RVs | • | • | • | • | • | • | • | – |
| Variable number of planets | • | • | – | • | – | – | – | – |
| Evidence (ln ) | • | • | – | • | – | • | – | – |
| Several samplers | • (5) | – | –^f | • | – | • | – | – |
| Normalising-flow proposals | • | – | – | – | – | – | – | – |
| Periodogram + FAP screen | • | – | • | – | – | – | – | – |
| Astrometric jitter | • | • | • | • | – | – | – | – |
| Acceleration models (7/9 par.) | • | • | •^g | – | – | – | – | • |
| Scan-angle harmonic terms | • | • | •^g | – | – | – | – | – |
| Student- likelihood | • | • | – | – | – | – | – | – |
| Correlated noise (GP) | – | •^e | • | • | – | – | – | – |
| -body | – | – | – | • | – | • | – | – |
Table 1. Capabilities of codes that fit orbits to absolute astrometry, from their papers and documentation (•: yes; –: no).
orbit signal and taking its O–C. For a period long compared with the time span of the data this removes most of the orbit. With the weights absorbed into the design matrix, , the corresponding symmetric projector is ; its diagonal gives the fraction of each measurement absorbed by the single-star fit, and its trace is . ExoPlanetLab rebuilds from the processing, verifies that it reproduces the delivered residuals to mas, and applies it to every model compared with the data. The projection is required. Without it, a published orbit fits the published residuals worse than no orbit at all (Sect. 6).
Projected versus fitted single-star terms. Other codes (e.g. kima) instead fit the single-star parameters together with the orbit, modelling . With Gaussian noise and a flat prior on the two approaches give the same orbital posterior. Writing for the of a residual vector,
the left side is our , the right side that of a joint fit at the best single-star parameters, and both equal , because the minimising is the single-star solution for the data . This is the Frisch–Waugh–Lovell theorem of linear regression [12, 31]. Removing the linear parameters of a nonlinear fit in this way is also known as variable projection [16]. Our is therefore the fitted one profiled over , and marginalising over instead adds only to the log-likelihood, a constant because does not depend on .
The equivalence is exact for Gaussian errors without a prior on . The weak pseudo-colour prior makes it approximate, negligibly so in practice (Sect. 6), and so does a fitted jitter (Sect. 3.6), at the level of in the jitter variance, because keeps the pipeline’s weights. The differences between the two approaches are practical. Projection samples seven parameters per companion rather than , reuses the pipeline’s own weights and prior, and needs no priors on . It does not report , the orbit-corrected parallax and proper motion, but these can be recovered for each posterior sample as ; and it cannot take an informative prior on the parallax. Finally, the projection assumes Gaussian errors: a least-squares fit is pulled by outliers, and spreads their effect over the other epochs. For heavy-tailed noise ExoPlanetLab therefore refits robustly at every likelihood evaluation (Sect. 3.6).
Radial velocities
The RV model is , with the true anomaly of companion and derived from the constants, so the RVs share with the astrometry and add only and one offset and jitter per instrument . The constants determine only up to ; is taken on the branch with and is allowed to be negative, which selects the other node. Reported values are converted to , moving both angles by when , so that with RVs the node is determined over the full circle. The bounds on , and the jitter scale with the range of the loaded RVs, so that absolute velocities of stellar-mass companions are handled without clipping. RV times must declare their zero-point, and are shifted onto the astrometric one.
For a dark companion the photocentre orbit is the orbit of the star, so follows from the astrometric orbit and the parallax. can optionally be tied to the orbit in this way, either in the fit or afterwards by importance sampling of the posterior draws of a fit with free; the two give the same result.
Relative astrometry
A companion resolved by imaging, interferometry or speckle observations is measured relative to its host. With the mass fraction and the flux fraction, the host moves about the barycentre by and the unresolved photocentre by , where is the companion’s position relative to the host, so that
with the photocentre displacement of Sect. 3.1 and and the angular semi-major axes of the relative and photocentre orbits. The relative positions thus share all orbital elements with the Gaia data and add one parameter per imaged companion, , with a log-uniform prior. Their likelihood is a two-dimensional Gaussian per epoch with the quoted errors and correlation, to which a fitted jitter can be added in quadrature; the reflex motion of the host due to other companions is neglected.
The scale separates the two masses: Eq. (8) is the photocentre relation used for the Gaia DR3 astrometric binaries (Eq. 14 of Halbwachs et al. [19]; Eq. 2 of Gaia Collaboration et al. [13]), written with and . With the parallax, the relative semi-major axis in au is , and Kepler’s third law gives the total mass (au, yr, ) without a stellar model. Solving Eq. (8) for the mass fraction gives , and the individual masses follow as and . For a dark companion , so ; for a luminous one must come from photometry, from the measured flux ratio in the band, and it enters only this mass calculation in which is free.
Likelihood and priors
For Gaussian noise the astrometric and RV terms are independent:
The first sum runs over the along-scan measurements: is the published residual and the projected orbit model (Sect. 3.2), the formal uncertainty of measurement , and the astrometric jitter (Sect. 3.6), added in quadrature. The second sum runs over the radial velocities: is the measured velocity with uncertainty , the model of Sect. 3.3, and the instrument that made measurement , with jitter . The constant is . The logarithmic terms normalise each Gaussian and matter only when a jitter is fitted: with all jitters fixed they are constant, and Eq. (9) reduces to , the weighted least-squares of Sect. 3.2 plus its RV counterpart. Priors are log-uniform on between 1 d and twice the nominal 10-yr mission (7305 d, about four DR4 baselines; a larger upper bound let samplers settle on spurious long-period, high-eccentricity fits), uniform on , on over the baseline extended by one maximal period on either side, isotropic in orientation with a log-uniform (below), uniform on and , log-uniform on the jitters (an RV instrument with fewer than three points has no free jitter: with one point its offset reproduces the measurement exactly, so the data say nothing about the jitter, and the posterior in offset and jitter becomes a funnel, the offset pinned at small jitter and free at large jitter, which nested samplers cannot explore; the jitter is then held at 0.01 m s), and Gaussian on the stellar mass and parallax when their external uncertainties are given (these enter only the mass estimates). Multi-planet fits impose . Uniform priors on the Thiele–Innes constants imply , the Jacobian of the transformation from to , which favours large, edge-on orbits; Baycroft et al. [2] find that such priors can distort Bayes factors for low-amplitude signals. By default the prior is therefore isotropic (, , uniform) with a log-uniform ( mas), i.e. . Divided by the Jacobian, this is a factor in Thiele–Innes space, within mas; is floored at to keep the factor finite for face-on orbits. The uniform Thiele–Innes prior remains an option; it is the only one for the RJ samplers, and outside evidence runs (Sect. 4.4), for dynesty, whose prior transforms assume it. It should not be used for weak signals (Sect. 5.5). For the period of an orbit longer than the baseline the posterior depends on the upper bound of the period prior (Sect. 5.3).
Astrometric jitter
Real epoch astrometry is not fully described by its formal errors (Sect. 6). An astrometric jitter , one parameter per source, is therefore added in quadrature to every along-scan error, , and fitted with the orbit (log-uniform prior from mas to the larger of 1 mas and three times the scatter of the data); it is identifiable because the likelihood keeps its term. It measures how much noise the formal errors miss, whatever its origin.
For the DR4 pre-release data a single white jitter is the wrong shape. With the formal errors, the orbit fits give , 3.1 and 9.9 for Gaia-4, Gaia BH3 and HD 114762. The scatter of the CCD measurements of one transit about their mean gives almost the same values: 1.4, 3.1 and 9.8. The whole excess is therefore already present within a transit, and it is not a per-transit (attitude) term. The core of the distribution follows the formal errors. The excess comes from a few per cent of outliers, up to . The default model for epoch data is therefore a Student- distribution per CCD measurement [28], , of scale and with two parameters per source: the number of degrees of freedom , which sets the weight of the tails (small gives heavy tails, and the Gaussian), and a factor that scales the formal errors of the core. Both are fitted, with log-uniform priors on and . Because outliers also corrupt a linear least-squares single-star fit, the robust likelihood avoids using the published residuals and instead uses the raw along-scan positions, , with the unprojected orbit signal of Sect. 3.1 and noise distributed as . The single-star terms are refitted for every likelihood evaluation, by iteratively reweighted least squares (IRLS): starting from the least-squares solution, each meameasurement is weighted by , which down-weights large residuals, and the weighted fit is repeated until the best fit converges. The weights follow from the likelihood. Setting the derivative of with respect to to zero gives the normal equations of a weighted least-squares fit, , with the -th row of . The weights depend on the residuals, hence the iteration [23]. This is the maximum-likelihood fit of a regression with -distributed errors [28]. Here are the residuals of this robust fit. Marginalising over with a Laplace approximation [38] about , in which the Fisher information of the -distribution [28], per measurement, takes the place of the Gaussian weights , gives
with , and the pseudo-colour prior of Sect. 3.2. With RVs, the RV sum of (9) is added to (10) unchanged; and enter only the astrometric term. As and , (10) strictly converges to the exact Gaussian marginal likelihood over single-star terms of Sect. 3.2. The fitted tails are , and with , and ; on simulated Gaussian data goes to large values, and the orbit is unchanged, the widths growing by 2%. The Gaussian model with jitter remains available.
Acceleration and scan-angle terms
A companion with a period much longer than the data shows only the curvature of its orbit: a constant acceleration of the proper motion, or an acceleration and its rate of change (jerk). Such terms are part of the core astrometric solution (AGIS) used for the DR4 epoch astrometry, and in DR3 they gave the seven- and nine-parameter acceleration solutions of the non-single-star processing [19]. We add them as linear terms of the single-star model,
with the time in years from the reference epoch of the epoch data, the acceleration (mas yr) and the jerk (mas yr) in the two coordinates; the seven-parameter model has the acceleration only. In the along-scan direction these displacements are projected as in (2), and the fitted parameters are the acceleration and jerk themselves. Two further models test a different explanation: an along-scan time independent shift that depends on the direction of the scan, ,
with harmonic amplitudes in mas, fitted directly, and or . The harmonic, , is a fixed offset in and and is already in the single-star model. Such a signal arises from source structure or calibration errors, which the scanning law can turn into an apparent orbit [22]. It is an alternative to an orbit or an acceleration and not a reparameterisation of them. Being linear, all these terms are fitted with the single-star terms at every likelihood evaluation and cost no sampled parameter. The evidence of these models needs a proper prior on the terms, which is built in two levels. Each pair, , or , has a Gaussian prior with zero mean and standard deviation on each component, ; sets the typical size of the terms and is itself not fixed, but log-uniform over five decades, to in the unit of the terms (mas yr, mas yr or mas). The mixture of Gaussians over [1] is isotropic in direction and close to log-uniform in amplitude, like the prior of an orbit. The first level can be integrated without sampling. For a fixed both the likelihood and the prior are Gaussian in the linear terms (the likelihood exactly for Gaussian noise, and in the Laplace approximation of Sect. 3.6 for Student- noise), so the integral of their product over the terms has the closed form [32]
where is the likelihood with the new terms added to the single-star model and given a flat prior (Sects. 3.2 and 3.6), and are the best fit of the new terms and their covariance under that flat prior, and is the covariance of the Gaussian prior. With normalised as in (10), the factors of of the two Gaussian integrals cancel. The last two terms are the price of the prior. The quadratic term penalises fitted terms that are large compared with . The determinant term is an Occam penalty: it grows with how much more precisely the data fix a term than the prior does, and it vanishes for a term that the data do not constrain. The second level, the average of over the log-uniform prior on (one scale per pair), is computed numerically on a grid. The models are used in two ways: alone, their evidences are direct integrals (a few seconds per source) that an orbit has to beat; inside an orbit fit, they absorb the curvature of a wider companion, or a scan-angle signal, without sampling it. With RVs a linear velocity trend common to all instruments can be integrated in the same way.
Inference
Signal screen
A time-only periodogram is invalid for scan-projected data, and a circular-orbit approximation places the maximum at the wrong period for eccentric orbits. The screen therefore profiles and on a grid at each trial period and, since the signal is linear in the Thiele–Innes constants, solves for them by weighted least squares, with the same projection as the orbit model (Sect. 3.2). The significance of a peak is tested with a permutation bootstrap: the residuals are shuffled among the epochs, keeping the times and scan angles, which destroys any orbital signal but keeps the noise and the sampling, and the highest peak of each shuffled periodogram gives its distribution for noise alone. The false-alarm probability (FAP) is the fraction of 200 shuffles whose highest peak reaches the observed one, and a peak is accepted if it is below 1%. Peaks are further vetted by a bounded refinement with a failure sentinel and a rejection of fits pinned to the period boundary. A Gauss–Seidel cross-refinement then re-fits each candidate against the data with only the other candidates subtracted, removing the duplicate and harmonic peaks that sequential subtraction produces at commensurate periods. The screen provides the signal count and the starting point; no fit is initialised from a published orbit.
Least squares is the right statistic only for Gaussian noise; in the heavy-tailed noise of the epoch data a few outliers dominate . With the Student- noise model the screen therefore first fits to the residuals and works on their score, with , the locally most powerful statistic for a small linear signal in noise; the periodogram and the permutation bootstrap are unchanged, and for Gaussian residuals (Sect. 5.4).
Peel and joint fit
For planets, single-planet “peel” fits on progressively cleaned residuals provide warm starts. At commensurate periods the peels are biased, because each one absorbs part of the signal of the others (Sect. 5). Every backend therefore depends on a joint fit of all planets. This is an affine-invariant ensemble MCMC ([17, 11]), parallel-tempered with four temperatures. With RVs, tempering is needed because RVs taken years after the astrometry make the posterior multimodal in phase. Without RVs, it is what makes weak signals, near-circular orbits and periods beyond the baseline converge (Sect. 5.3).
Tempering does not free the fit from its starting point when the modes are separated in period, so the start is searched for (Sect. 5.7). For one orbit the candidates are the screen’s detections and the five highest distinct peaks of the first periodogram; each is refitted as a single orbit to the full data and they are ranked by the likelihood of the noise model of the fit. For two or more orbits a beam search builds the combinations up orbit by orbit: every kept combination is extended by the three highest distinct peaks of the robust periodogram of its residual, its orbits are refitted in turn against each other, and the four combinations with the highest likelihood are kept. The fit starts from the best-ranked candidate, or the peel start if that is better, and can be repeated from the others, the highest evidence being kept.
Samplers
emcee. The joint fit of Sect. 4.2 is itself the emcee result.
dynesty ([35]). A cold nested run over the full prior it took calls or converged on a wrong mode. The joint stage then runs with emcee.
rjmcmc and rjflow. These two reversible-jump samplers use ern (Karnesis et al. 2023). They sample the number of planets together with the orbits. Both start from an emcee joint fit at the detected number of planets. Births are proposed from its posterior and from the strongest peak of the residual periodogram. Within a model both samplers mix differential-evolution and adaptive Gaussian moves. rjmcmc adds a KDE move and rjflow a RealNVP flow, both built from the emcee posterior. Both run with parallel tempering. Neither supports RVs.
Model comparison
The reversible-jump samplers accept almost no trans-dimensional moves on these problems, so the number of companions is chosen from the evidences of fixed- fits: stepping up from , a companion is added when the Bayes factor satisfies (odds of about 150:1). For and for the acceleration and scan-angle models only the noise parameters remain, and the evidence is integrated directly. For orbits it comes from the tempered fit with 16 temperatures down to the prior, each rung’s evidence ratio being estimated by bridge sampling ([3]) from the samples of both its temperatures; the error is the scatter of consecutive blocks of the chain.
A strong signal switches on over a narrow range of temperature, and the chains there equilibrate far more slowly than the cold chain, whose autocorrelation time normally ends the run. A temperature ladder frozen after the burn-in can then leave a gap at the transition, and keeps changing with the length of the run (Sect. 6). An evidence run therefore stops only when the two halves of its production chain give the same within 0.5; when they do not, the ladder adapts again and the evidence restarts, the cold chain keeping every step. A run still unstable at its step limit is flagged.
The nested samplers give an independent value. They search a box around the posterior; in evidence runs they integrate the likelihood times the full normalised prior over that box (including the isotropic orientation prior and, for , the period ordering), with the periastron time sampled as an orbital phase, because the likelihood repeats every period in and a box around one copy would miss the others. With RVs the nested values are not yet validated, and the tempered evidence is used (Sect. 6).
Diagnostics
Every run records its random seed, settings and package versions, so that it can be repeated exactly, and reports the integrated autocorrelation times against a criterion (at which the joint fit stops, except in evidence runs), the joint residual RMS in units of the formal errors, the folded to the first observed cycle together with its raw value, and the companion mass from the orbit and, with RVs, from and the astrometric inclination.
Validation on simulated data
A commensurate two-planet system
Two planets with and d, and and and mas (SNR 53 and 19) were injected into 140 along-scan epochs over 9.9 yr with 0.1 mas white noise; the RV runs add 50 epochs with 10 m s noise and , m s. The 2:1 ratio makes single-planet fits biased, and the outer planet is observed for only 1.2 orbits.
Without RVs, all five samplers (emcee, nessai, dynesty and the two RJ samplers, whose parameters come from the joint MCMC) agree to within 0.1 in every parameter, with a joint residual of 1.04–1.05 times the noise. All eight runs were made side by side with the jitter tests of Sect. 5.2 and the DR4 runs of Sect. 6, one CPU core each and up to nine at a time, on a 2021 MacBook Pro with an Apple M1 Max (10 cores: 8 performance and 2 efficiency; 64 GB), in 45 min in total. Each run took 8–19 min for emcee, nessai and dynesty and 29–31 min for the RJ samplers without RVs, and 9–13 min with RVs; the periodogram screen takes 0.5–1.8 min, and the tempered joint sampling 5–13 min without RVs and 3–9 min with them, since it stops once the chain exceeds 100. The outer planet’s and lie along a nearly degenerate ridge, and its median sits 2–3 from the injected values in every sampler. With RVs, emcee, dynesty and nessai again agree, every offset falls to ≤ 1.1 and the width halves (Table 2). The RJ samplers return but accepted no trans-dimensional move. Figure A.1 shows the ridge and its removal by the RVs.
| true | astrometry | + RV | |
| (d) | 2900 | 2898.6 ± 5.2 −0.3 | 2901.7 ± 4.4 +0.4 |
| (d) | 800 | 911 ± 47 +2.3 | 828 ± 28 +1.0 |
| (mas) | 1.8 | 1.54 ± 0.14 −1.9 | 1.75 ± 0.07 −0.7 |
| (mas) | 3.5 | 3.66 ± 0.04 +3.1 | 3.56 ± 0.06 +0.8 |
| (mas) | −1.0 | −1.41 ± 0.16 −2.4 | −1.12 ± 0.11 −1.1 |
| (mas) | 1.2 | 0.36 ± 0.38 −2.2 | 1.00 ± 0.21 −0.9 |
Table 2. Outer planet of the simulated 2:1 system: posterior median and offset from the injected value, without and with RVs.
Astrometric jitter
The same benchmark with RVs was re-run with extra white noise added to the astrometry but not to its quoted errors (0.10 mas). With none added the fitted jitter is consistent with zero ( mas) and every parameter is within 0.05 of a run without the jitter term; with 0.10 and 0.30 mas added it returns and mas, and all 16 parameters remain within 2 of the truth while the widths grow. The RV amplitudes are recovered as and m s (true 500 and 200).
Injection–recovery campaign
To test detection, convergence and the calibration of the posteriors we injected 180 single companions into the real observing geometry – epoch times, scan angles, formal errors and single-star projection – of five quiet sources of the DR4 pre-release (–14.3, 576–781 epochs, baselines 1733–1902 d; their own measurements are not used). The data are with Gaussian noise at the formal errors, and are fitted as real data with the default settings (fitted jitter, isotropic prior, tempered joint fit). The orbits were drawn by Latin hypercube sampling, in three campaigns of 60 with different seeds, from the fit’s prior restricted to –2 baselines (log-uniform), (uniform), isotropic orientation, uniform phase, and a total signal-to-noise ratio of the projected signal log-uniform from 5 to 100 (which uses ). The posteriors are restricted to the same ranges before comparison, so that the prior equals the injection distribution and the coverage of the credible intervals must be exact if the fit is correct. A campaign takes 72–91 min on 9 cores (median 9.3 min per run).
All 180 runs converged. Tempering is essential here: an earlier campaign with the untempered sampler converged in only 41 of 60 runs, failing for weak signals, near-circular orbits – where a shift of is a rotation that absorb, leaving a ring in the posterior – and periods beyond the baseline. We re-fitted 23 problem runs (the 19 that failed and 4 that converged with other problems) and 4 controls with the tempered sampler. All 27 converged, whereas the untempered sampler had converged only 8 of them, and the tempered fits took a fifth of the total sampling time. All signals with are detected, and 83% of those with –10; 64%, 95%, 91% and 100% of the runs recover the period to 20% for –10, 10–20, 20–50 and 50–100.
Table 3 and Figure 1 give the calibration over the 173 converged runs with a detection. All parameters are unbiased, and all but the period are calibrated ( slightly conservatively: its intervals are a little too wide). For the period, 90% of the truths are inside the 95% intervals (17 outliers against 8.7 expected), but the mean posterior quantile of the truth, , and a Kolmogorov–Smirnov test of its uniformity () show no bias. We verified that the prior matches the injections (reweighting the posteriors to the injection distribution changes nothing), that the likelihood is exact (the stored single-star gain uses the same weights ), and that the sampler returns the correct posterior (an independent grid integration, with marginalised analytically, reproduces the chain’s period quantiles for eight of the worst runs, and for every outlier the best posterior sample fits better than the truth). The excess is therefore statistical: it was found in the first campaign () and is weaker in the two independent repeats (10 of 115 outside the 95% intervals against 5.8 expected, ).
Figure 1. Injection–recovery calibration (173 runs). Left: fraction of truths inside the central credible interval against its level, per parameter; calibrated posteriors follow the diagonal, within the grey 95% binomial range. Right: posterior quantile of the true period against the true period in units of the baseline (1733–1902 d), coloured by the total signal-to-noise ratio ; dashed lines bound the 95% interval. For true periods just below one baseline the posterior tends to lie above the truth (points near ): such periods tend to be overestimated, while near 1 means it is underestimated.
Calibration holds on average over the injections, not at each true period. For true periods within 10% of the baseline the posterior tends to lie above the truth (in 6 of 12 runs the truth is at its 0.2–7th percentile, mostly for ), balanced by true periods of 1–2 baselines (mean quantile 0.54 against 0.39 for 0.5–1 baselines). We attribute this, without having verified it, to orbits slightly longer than the baseline reproducing the same projected signal with more prior volume. A period close to the baseline may therefore be overestimated, and its credible interval extends to longer periods.
Relative astrometry (Sect. 3.4) was tested with two further campaigns of 60 injections, with Student- noise () and the Student- likelihood: one on the Gaia data alone, and one that adds to the same data sets eight simulated relative positions with (Eq. (8)), Gaussian errors of 2 mas per coordinate and epochs spread over three Gaia baselines. All 120 fits converged, and the screen, which uses the Gaia data only, detects the same 53 signals in both. With relative astrometry the posteriors remain calibrated: averaged over the nine orbital parameters, 71% and 96% of the truths lie inside the 68% and 95% intervals (67% and 94% without relative astrometry), and 70% and 98% for . Every value is inside the range expected from chance for 53 runs: 55–79% for the 68% intervals and 89–100% for the 95% intervals. Run by run the credible intervals are narrower, with median width ratios of 0.27 for , 0.38 for , 0.67 for and 0.64 for , and the period is within 20% of the truth in all 53 detected runs, against 45 without. The gain is largest for the 12 orbits longer than the Gaia baseline, which the relative positions cover in full (0.02 for , 0.10 for ). Figure A.4 shows a pair with the typical gain. The scale is measured to a median 8%. On two such data sets with Gaussian noise ( and 3000 d) Octofitter, fitting the same along-scan points and relative positions, agrees with ExoPlanetLab within 0.16 in all orbital elements, and the parallax, and the widths of the 68% intervals agree to 4%. The only difference is a convention: Octofitter gives the argument of periastron of the companion’s orbit about its host, and ExoPlanetLab that of the photocentre orbit, which differs by 180°.
| in 68% | in 95% | median | rms | |
| 64% | 90% | +0.09 | 1.12 | |
| 65% | 95% | +0.02 | 1.05 | |
| 78% | 97% | −0.18 | 0.87 | |
| 64% | 91% | +0.11 | 1.08 | |
| 68% | 95% | +0.13 | 1.02 | |
| 70% | 94% | −0.05 | 1.00 | |
| 68% | 95% | −0.12 | 1.02 | |
| 63% | 96% | −0.05 | 1.06 | |
| 69% | 95% | +0.13 | 1.01 | |
| calibrated | 68% | 95% | 0 | 1 |
| range | 61–75% | 91–98% |
Table 3. Calibration of the posteriors in the injection–recovery campaign (173 converged runs with a detection): fraction of truths inside the central 68% and 95% credible intervals, and the median and rms of the normal score of the truth’s posterior quantile , where is the standard normal cumulative distribution function ( when the posterior lies above the truth). The last two rows give the calibrated values and the binomial 95% ranges of the coverage for 173 runs.
Detection threshold and false positives
The tests of this and the following subsections use Student- noise with , like the pre-release data, in the geometry of the same quiet sources, and single-orbit evidence runs. Thirty data sets without a companion give (Sect. 4.4) between −1.44 and −0.24: none is above zero. An orbit with d and injected at with six noise realisations each is detected () in (1, 1, 5, 5, 6) of them (Figure 2a), so half of the companions are found between and 8, and every detected period is within 5% of the truth. A threshold of 3 would add three detections at without admitting a null, however, 30 nulls do not exclude a false-positive rate of several per cent.
Figure 2. Detection in Student- () noise. (a): evidence for one companion against none for an injected orbit ( d, ), six noise realisations per signal-to-noise ratio ; the band is the range of 30 data sets without a companion and the dashed line the adopted threshold. (b): fraction of injected signals found by the periodogram screen with least squares and with the robust statistic (50 screens per point; 68% binomial intervals).
The screen is the less sensitive stage in such noise. With least squares it finds 12, 36, 56 and 80% of the signals at and 12 (50 screens each); on the Student- score (Sect. 4.1) it finds 24, 74, 86 and 100% (Figure 2b). The gain costs neither false alarms nor sensitivity in Gaussian noise. On 300 data sets of pure Student- noise the robust screen reports a signal once (the least-squares screen twice), where the 1% false-alarm threshold allows about three; on 200 data sets of pure Gaussian noise it reports none; and with a signal at in Gaussian noise both screens find 49 of 50. For this kind of data one companion is always fitted, so a signal the screen misses is still recovered by the joint fit from a blind start; the screen decides how many signals are reported.
Orientation prior
For strong signals the orientation prior does not matter; for weak ones it decides the outcome. At the isotropic prior gives and +1.7 in four tests, and the uniform Thiele–Innes prior −5.4, −7.7, −3.2 and −18.6 for the same data, with posteriors that run to and an from twice to thousands of times the true one: its volume outweighs a weak likelihood, as Baycroft et al. [2] warned. At the two priors give the same orbit. The effect on the real sources is given in Sect. 6.
Periods near one year and high eccentricity
Orbits with , 365.25 and 385 d, and a 300-d control, were injected at and 15 with four noise realisations each. At a given a period of one year is no harder than 300 d: all 16 runs at are detections and 12 of 16 at (the four missed are at 300, 345 and 385 d, none at one year), the periods are within 0.2–1.7% of the truth, and the 28 detected runs are calibrated (the true period inside the 68% interval in 68% of them). The cost of the parallax degeneracy is the amplitude needed to reach that .
Orbits with and 0.95 are detected in seven of eight runs each (one missed at $S = 8) and their periods recovered to 1–2%, but and come out too small: for the posterior median is 0.60–0.76 of the true and half the true , the truth lying outside the 68% interval in all seven detected runs. The eccentricity prior is uniform, so only 5% of it lies above , and at –15 the data do not rule out rounder orbits, which then dominate the posterior. The posteriors are calibrated on average over truths drawn from the prior (Sect. 5.3), not for every fixed truth, and a truth at the edge of the prior is pulled inwards. A weak, highly eccentric orbit is therefore reported rounder and smaller than it is.
Two companions
To test the choice of the number of companions, a second companion was added to an inner one ( d, ). It was placed at period ratios of 1.3, 2 and 3, with its own signal , 10 and 15, in two noise realisations. The 18 cases were fitted with one and two orbits, and four of them also with three (Table 4). All fits start from the best of several ranked starting configurations (Sect. 4.2).
| (d) | recovered (d) | ||||
| 390 | 6 | +0.9 | +0.3 | – | – |
| 10 | +14.3 | +5.7 | |||
| 15 | +20.1 | +20.8 | |||
| 600 | 6 | −0.1 | +2.4 | – | – |
| 10 | +13.4 | +14.4 | |||
| 15 | +16.4 | +26.2 | |||
| 900 | 6 | −0.3 | +2.8 | – | – |
| 10 | +11.0 | +17.7 | |||
| 15 | +28.9 | +45.5 |
Table 4. Two injected companions (inner: d, ): evidence for the second orbit, , for the two noise realisations, and the period of the second orbit in the two-orbit fit.
The number of companions is correctly recovered in all 18 cases. At the second companion is not accepted ( from to ). At and 15 it always is ( to ). A third orbit never is ( to ). In all 12 detected cases the two-orbit posterior has a single period pair, the injected one, with at least 98% of the samples.
The ranked starts are needed. Started from orbits fitted one after another to the residuals, the two-orbit fit recovered the second period in only 9 of the 12 detected cases. In the other three, two of them at a period ratio of 1.3, the posterior held several period pairs, and only 12–48% of the samples were on the injected one. In these 16-parameter fits the sampler moves slowly and stays close to where it starts, which here was a wrong pair.
The one-orbit fit has a similar problem. In one case with two equally strong companions, at 300 and 600 d, it found a period of 16.8 d. This alias fits the combined signal about as well as either true period. Fits started from different periods give evidences within 1.6 of each other, and we use the highest. A second-companion solution is considered reliable when its posterior has a single period mode or fits from different starts agree. The test is small as it has one inner orbit, one orientation of the second companion and only two noise realisations per case. It therefore shows that the number of companions can be chosen correctly. It does not show often this succeeds for other orbits.
Orbit or acceleration
We examine the period range over which an orbital model is preferred to an acceleration model. One companion () was injected with periods of 1 to 4 times the 1733-d time span of the data and and 5 mas. Up to twice the time span the orbit is strongly preferred over the better of the two acceleration models (with and without jerk; Figure 3). The preference is lost at about 2.5 time spans for 1 mas and 3 for 5 mas, beyond which the acceleration models have equal or higher evidence and the period posterior Notes. One- and two-orbit fits started from the best of several ranked starting configurations (Sect. 4.2).
Figure 3. Evidence of the one-orbit fit over the better of the acceleration models (with and without jerk), against the injected period in units of the time span of the data. The vertical scale is linear below 10 and logarithmic above.
is the same whatever the true period. All twelve are overwhelming detections against the single star. The comparison separates an orbit from mere curvature. For strong signals the jerk term is needed: at 5 mas and three time spans the orbit has against the model with acceleration only, which alone would wrongly make the orbit look strongly warranted, but −2.5 against the one with acceleration and jerk (the value shown in Figure 3), so the curvature is described as well without an orbit. Both acceleration models must therefore be compared with the orbit.
An inner orbit ( d, ) with a real outer companion ( mas) shows the use of the terms inside a fit. With one orbit and no acceleration terms the fit takes the stronger outer signal and loses the inner companion. For an outer period of 3.5 time spans, acceleration and jerk describe the outer companion as well as a second orbit does ( against 201.6) and recover the inner period equally well (302.5 2.2 d). The acceleration alone is not enough. For an outer period of 1.7 time spans the terms are not an adequate description and two orbits are needed (). When the same inner orbit is injected without an outer companion and fitted with the acceleration terms, the evidence is only 0.76 lower than without them. The fitted orbit is unchanged so including the terms costs little when there is nothing for them to describe.
A pure scan-angle signal of 0.04 mas, injected without an orbit, is detected by the scan-angle model of Eq. (12) () and not by the orbit fit (+0.4). A real orbit keeps its evidence when the terms of Eq. (12) with are in the fit: is only 1.2 lower. An orbit does leak into them so a significant scan-angle term by itself does not show a signal to be spurious, however, the comparison with the orbit does.
Upper limits
For a source without a detection, we compute an upper limit on the semi-major axis of a companion as a function of period. This needs no sampling. At a fixed period, eccentricity and periastron time, the orbit is linear in its four Thiele–Innes constants. These can therefore be fitted together with the single-star terms, like the terms of Sect. 3.7. The limit is the 95% point of the posterior of . It assumes
an isotropic orientation and a prior uniform in , and is marginalised over eccentricity and phase. We tested the limits on 400 injected companions for each of two sets. For circular orbits the true is above the 95% limit in 2.2% of the cases. For eccentric orbits () it is above it in 3.5%. The limits are therefore slightly conservative.
Application to the Gaia DR4 pre-release epoch astrometry
The DR4 pre-release (June 2026) contains CCD-level epoch astrometry for 12 sources. We fitted the three with published orbits, with and without RVs (Table 5).
| Gaia BH3 published | Gaia BH3 this work | Gaia-4 published | Gaia-4 this work | HD 114762 published | HD 114762 this work | |
| , time span (d) | 558, 1861 | 824, 1804 | 558, 1798 | |||
| (d) | ||||||
| (mas) | ||||||
| (deg) | ||||||
| – | ||||||
| (mas) | ||||||
| – | – | – |
Table 5. The three pre-release sources with published orbits: joint fits of the epoch astrometry and the radial velocities (Student- noise, posterior medians and 68% intervals) next to the published values. is the number of CCD measurements, the companion mass from , and the stellar mass, and the parallax corrected for the orbit.
For Gaia BH3 (Gaia Collaboration et al. [14]) all parameters are within 1.4 of both published solutions. The samplers emcee, dynesty and nessai agree to 0.14 in every parameter (Figure 4 and Figure 5). The companion mass agrees with the published one. The RV amplitude can also be predicted from the astrometric orbit and the parallax. The measured agrees with this prediction for mas, the parallax of the published combined solution. For our orbit-corrected parallax it is 2.0 below the prediction, and the purely astrometric mass is correspondingly 6% higher. No parallax zero-point is applied in the fit. Subtracting the Gaia DR3 zero-point of Lindgren et al. (2021) for this source, −0.035 mas, from our parallax gives 1.696 0.008 mas. Our results are therefore consistent with a zero-point of that size in the pre-release astrometry.
Figure 4. Gaia BH3, joint fit of the epoch astrometry and the radial velocities. Left: orbit of the photocentre on the sky (north up, east to the left) for 60 posterior draws and the posterior median; the part covered by the 1861 days of epoch astrometry is drawn thicker. Right: radial velocities of Gaia Collaboration et al. [14] with each instrument’s fitted offset removed, and the same posterior draws.
Figure 5. Gaia BH3: posterior of the period, eccentricity, photocentre semi-major axis and inclination from the astrometry alone (solid) and with the radial velocities (dashed). Contours enclose 39% and 86% of the probability.
In the joint fit is free. Tying it to the astrometric orbit (Sect. 3.3) tests whether the data agree with a dark companion, and gives one mass instead of an astrometric and an RV-based one. For Gaia BH3 the orbit then moves by 0.6. The parallax becomes 1.668 0.007 mas and the single mass 33.9 0.5 . For none of the three sources do the data contradict a dark companion.
Fitted to the epoch data of Gaia Collaboration et al. [14] (their Table B.1), the method reproduces their astrometric solution to better than 0.7 in every parameter. The remaining differences on the pre-release data come Notes. Combined astrometric and spectroscopic solution of Gaia Collaboration et al. [14]. Adopted Gaia + RV solution of Stefánsson et al. [36]; the parallax is the value they adopt. , and : RV orbit of Kane et al. [24], their adopted solution without a linear trend; and parallax: Gaia DR3 orbital solution [13]; and : Winn [40], from that solution and the RVs. in km s and in for Gaia BH3; in m s and for the other two.
from the differences between the two reductions of the same CCD measurements. After removing the single-star terms, their positions differ by 0.11 mas rms per CCD, comparable to the formal errors. Adding these measured differences, reassigned to the data of Gaia Collaboration et al. [14], reassigned at random among transits, moves the solution by 1.2–1.8. This contains every difference we find.
For HD 114762 the astrometry gives a nearly face-on orbit. The data have a large excess noise from heavy-tailed CCD outliers: with the formal errors, (Sect. 3.6). With a single white jitter, the astrometry-only period and eccentricity (83.850 0.038 d, 0.311 0.029) lie 1.8 and 1.1 from the RV orbit. The inclination is then bounded only from above ( deg, , 95%). With the Student- model the astrometry alone gives
d and , within 0.15 of the RV orbit. The joint fit then bounds the inclination: deg and (95%, isotropic orientation prior). 90% of the posterior is above the hydrogen-burning limit.
HD 114762 is the only one of the three sources whose posterior depends on the orientation prior (Sect. 5.5). With RVs its inclination is deg under the isotropic prior and deg under the uniform Thiele–Innes prior. Gaia-4 and Gaia BH3 have the same posteriors under both. The companion is most likely an M dwarf. Kiefer [26] first concluded this from the DR1 excess noise. Winn [40] combined the Gaia DR3 orbital solution with the RVs and found deg and (225 14 ). Our mass in Table 5, 133, is lower and much wider. With free, the inclination comes from the astrometric orbit alone. For a nearly face-on orbit this bounds it only from above, and follows the prior. With tied to the astrometric orbit through the parallax (Pourbaix & Jorissen [34], Eq. 12), which the data allow, the RVs fix the inclination as well. We then find deg and , close to the values of Winn [40].
For Gaia-4 (Stefánsson et al. [36]) the eccentricity, inclination and agree with the published solution, but the period and orientation differ by –. The published solution rests on DR3 data. The period differs by of the published width, or of the two widths combined. The errors of Gaia-4 are nearly Gaussian, and the differences are the same with the formal errors, a Gaussian jitter and the Student- model. We also held the published orbit fixed and computed against the epoch data, for measurements. With no orbit, is 2869. With the orbit projected through the single-star fit (Sect. 3.2), it is 1227. Without the projection it is 3087, worse than no orbit at all. For Gaia BH3 (), is with no orbit, 2016 with the projected orbit and without the projection. The projection is therefore essential when a published orbit is compared with the epoch data.
With the formal errors, the orbit fits give for Gaia-4, 3.1 for Gaia BH3 and 9.9 for HD 114762. The excess sits within transits and is heavy-tailed (Sect. 3.6). It is not an error common to the CCDs of a transit, and part of it belongs to particular gates (Sect. 7.2). Its origin could be calibration, the source (e.g. the photocentre of a partly resolved companion) or other unmodelled effects; we cannot identify it. Refitting Gaia BH3 with an error factor per gate moves its orbit by less than .
With the Student- model, the orbital values of Gaia BH3 change by less than their uncertainties, compared with a Gaussian model with a fitted jitter (0.116 mas). The uncertainties shrink by about a quarter: against d in , and against in . They are now somewhat narrower than the published ones ( d, ), which use Gaussian errors.
The evidences select one companion for each of the three sources. Against no companion, , and (Gaia-4, HD 114762, Gaia BH3, in that order throughout). For a second companion, is for Gaia-4 and for Gaia BH3. For HD 114762 the tempered emcee fit with two orbits did not reach a stable evidence in 200 000 steps. Its second orbit collapses to the floor of the amplitude prior. The nested samplers give (nessai) and (dynesty).
The one-orbit evidence of HD 114762 needed the stability test of Sect. 4.4. A run stopped by the autocorrelation time of the cold chain gave , too low by 14. Longer runs give , and so do nessai and dynesty. For Gaia BH3 with RVs the two nested samplers disagree by 4, so we use the tempered evidence there.
The orbit is strongly preferred over the simpler models. Against the better of the two acceleration models, , and . Against the better of the two scan-angle models, , and . We also added the acceleration terms or the scan-angle terms (Sect. 3.7) to the orbit fit. The evidence then drops by 1–4. The fitted accelerations and scan-angle amplitudes are all consistent with zero, and the orbits are unchanged. None of the three sources shows curvature beyond its orbit or a scan-angle signal. The nine other pre-release sources show no companion, acceleration or scan-angle term above the threshold. For them we computed 95% upper limits on (Sect. 5.9). The limit varies with period. Its median over periods up to the time span of the data is 0.03–0.06 mas at –14 and 0.6–1.7 mas at –20.5.
Three other codes were run on the same measurements. Octofitter (Thompson et al. [37]) reduces each transit to the median of its CCD measurements. Fitted to the astrometry alone, it agrees with our fits within on every parameter of the three sources. kepmodel (maximum likelihood, Gaussian noise) agrees within in every parameter for Gaia-4 and Gaia BH3. It also reproduces our Gaussian-model fit of HD 114762. kima (Student- likelihood, priors matched to ours) agrees within for Gaia-4, including the tail parameter . Its sampler did not converge for the two strong signals.
Recovering the single-star terms for each posterior draw (Sect. 3.2) gives the orbit-corrected parallaxes of Table 5. For Gaia BH3 the parallax is below that of the published astrometric solution ( mas). It is below that of the combined solution before a zero-point correction, and above it with the DR3 zero-point of Lindegren et al. [30]. For Gaia-4 it equals the value adopted by Stefánsson et al. [36]. For HD 114762 it is below the parallax of the DR3 orbital solution, mas. It is 3.4% below the DR3 single-star parallax (26.198 mas, RUWE 3.2), which ignores the 84-d orbit. The pseudo-colour prior has a negligible effect on the projection (Sect. 3.2). With the fitted orbits, the projected and that of a joint fit without the prior differ by less than for all three sources.
Discussion and conclusions
Summary of results
ExoPlanetLab fits Keplerian orbits to Gaia epoch astrometry, alone or with radial velocities and relative astrometry, with one likelihood and five samplers. The principal results are the following.
The model orbit must be compared with the epoch residuals only after removing the part that the pipeline’s single-star fit absorbs. Without this projection, a published orbit fits the DR4 epoch data worse than no orbit at all. The single-star parameters are recovered for every posterior draw.
The excess noise of the DR4 pre-release data is heavy-tailed at the CCD level. A Student- likelihood with robustly refitted single-star terms describes it. For HD 114762 it brings the astrometric period and eccentricity onto the RV orbit. A robust version of the periodogram screen detects at – what least squares detects at –.
Three pre-release sources have published orbits. For Gaia BH3 all parameters are within of both published solutions. For Gaia-4 the period and orientation differ by – from the published solution, which rests on DR3 data. For HD 114762 the astrometry bounds the inclination, deg (95%). The companion is therefore not a planet but most likely an M dwarf, as Winn [40] found from the DR3 orbit and the RVs. With free its mass follows the orientation prior and lies below theirs. With tied to the astrometric orbit through the parallax we find deg and , close to their values. The orbit-corrected parallaxes of the three sources are within of the published ones.
The samplers agree with each other, and with Octofitter, kepmodel and (for Gaia-4) kima on the same data. In 180 injections, the posteriors of the 173 runs with a detection are unbiased. All parameters but the period are calibrated; 90% of the true periods lie within their 95% credible intervals. Highly eccentric orbits at low signal-to-noise ratio are reported rounder and smaller than they are. A uniform Thiele–Innes prior loses weak signals.
Evidences of fits with a fixed number of companions select that number. There is no false detection in 30 data sets without a companion, and half of the companions are detected at –. With two injected companions the count is right in all 18 cases. The orbit of the second companion depends on where the fit starts. The fit therefore starts from the best of several ranked candidates. The evidence of a strong signal needs a temperature ladder adapted to equilibrated chains and a stability test. Without them it was low by 10–14 for HD 114762.
Acceleration and scan-angle models show three things. An orbit is warranted up to periods of 2–3 times the time span of the data. An inner orbit survives a wide companion when acceleration and jerk are fitted with it. None of the three pre-release companions is a scan-angle artefact or carries a second signal.
Relative astrometry of a resolved companion enters with one parameter, the ratio of its relative orbit to the photocentre orbit. With the parallax, it gives the total mass without a stellar model. In paired injections, eight relative positions leave the posteriors calibrated. They narrow the credible intervals to 0.27 (), 0.38 (), 0.67 () and 0.64 () of their widths, and far more for periods beyond the time span of the data. Octofitter agrees within on the same simulated data.
Limitations and scope of the validation
The excess noise of the pre-release data is heavy-tailed, and its origin is not identified. Part of it depends on the gate. An error factor per gate, however, explains little of the heavy tail. The number of degrees of freedom of the Student- likelihood (Sect. 3.6), which is small for heavy tails, rises only from 2.9 to 4.0 for Gaia BH3 and from 2.2 to 3.0 for HD 114762. The Student- likelihood treats the CCD measurements as independent. No term common to the CCDs of a transit is detected.
The log-uniform prior extends far below the detection limit. An undetectable extra companion, acceleration or scan-angle term therefore lowers by only 1–4. A term is added only when it raises by more than 5, and in our tests no spurious term was added. A negative of −1 to −4, however, only says that the data do not need the term. It does not rule the term out.
With two or more orbits the tempered sampler does not move between period modes. The result therefore depends on the ranked starts (Sect. 4.2). The posterior can have several modes, for example the true pair of periods and an alias pair. A single run then finds only one of them, and the evidences of separate runs are needed to weigh them.
A fit of one companion with its evidence takes about one core-hour for a source with 600 measurements. The time grows in proportion to the number of measurements, and faster with the number of companions.
The validation has limits of its own. The injections use the geometry of five sources at –14. Fainter sources were only screened and given upper limits. The real orbits are those of three sources with strong signals and published solutions. The tests with two companions and of the transition from orbit to acceleration are small, with one orbit shape and one or two noise realisations per case. Relative astrometry was tested on simulated data only.
Outlook
ExoPlanetLab can be applied to the DR4 epoch astrometry as released, alone or with radial velocities and relative astrometry. The first test after the release is a comparison with the orbital solutions of a larger sample, such as the Gaia DR3 non-single stars. Each source is fitted independently, so a large sample is spread over as many cores as are available. At about one core-hour per source, sources take about four days on a cluster with 1000 cores. For the full DR4 sample the main open question is therefore the origin of the excess noise.
Acknowledgements. This work has made use of data from the European Space Agency (ESA) mission Gaia (https://www.cosmos.esa.int/gaia), processed by the Gaia Data Processing and Analysis Consortium (DPAC, https://www.cosmos.esa.int/web/gaia/dpac/consortium). This work was financially supported directly by the Swedish National Space Agency (SNSA/Rymdstyrelsen). The code and its documentation have been developed by the author working with Claude AI (Opus 5.5 and earlier Sonnet 5.5) using a Human Intelligence (HI)–AI iterative procedure. Many hundreds of HI-AI iterations have been necessary during development and testing and the author takes full responsibility for any remaining errors.
Code availability. ExoPlanetLab is available at https://github.com/hobbsSe/ExoPlanetLab and archived at https://doi.org/10.5281/zenodo.24101622 (version 1.0.0), under the GPL-3.0 licence. The code comes with a user guide (ExoPlanetLab.pdf). It describes the model, the priors, the samplers and every setting, and gives worked examples and exercises.
References
Corner plots
Figures A.1–A.4 show posteriors of the period, eccentricity, photocentre semi-major axis and inclination in pairs of fits to the same astrometry, in the style of Figure 5: contours enclose 39% and 86% of the probability, and the histograms are normalised.
Figure A.1. Outer planet of the simulated system (Sect. 5, Table 2), observed for 1.2 orbits, from the astrometry alone (solid) and with the radial velocities (dashed); thin dark lines and squares mark the injected values (true values). Without RVs the periastron time lies along a ridge that leaves its median 2–3 from the truth; the RVs remove it.
Figure A.2. Gaia-4, from the astrometry alone (solid) and with the radial velocities of Stefánsson et al. [36] (dashed).
Figure A.3. HD 114762, from the astrometry alone (solid) and with the radial velocities (dashed), which fix the period and eccentricity; the orbit is close to face-on.
Figure A.4. An injected companion ( d, , ; Sect. 5.3) from the astrometry alone (solid) and with eight simulated relative positions (dashed); thin dark lines and squares mark the injected values (true values). This pair has the typical gain of the campaign: the 68% intervals narrow to 0.31 (), 0.39 (), 0.71 () and 0.68 () of their widths. For periods beyond the time span of the astrometry the gain is far larger (0.01–0.03 in ).
References
- [1]Andrews, D. F. & Mallows, C. L. 1974, Journal of the Royal Statistical Society. Series B (Methodological), 36, 99
- [2]Baycroft, T. A., Faria, J. P., & Delisle, J.-B. 2026, A&A, in press; arXiv:2606.24132DOI
- [3]Bennett, C. H. 1976, Journal of Computational Physics, 22, 245
- [4]Blunt, S., Wang, J. J., Hirsch, L., et al. 2024, JOSS, 9, 6756
- [5]Blunt, S., Wang, J. J., Hirsch, L., et al. 2020, AJ, 159, 89
- [6]Brandt, T. D. 2021, ApJS, 254, 42
- [7]Brandt, T. D., Dupuy, T. J., Li, Y., et al. 2021, AJ, 162, 186
- [8]Delisle, J.-B. & Ségransan, D. 2022, A&A, 667, A172
- [9]El-Badry, K., Lam, C., Holl, B., et al. 2024, The Open Journal of Astrophysics, 7, 100
- [10]Faria, J. P., Santos, N. C., Figueira, P., & Brewer, B. J. 2018, JOSS, 3, 487
- [11]Foreman-Mackey, D., Hogg, D. W., Lang, D., & Goodman, J. 2013, PASP, 125, 306
- [12]Frisch, R. & Waugh, F. V. 1933, Econometrica, 1, 387
- [13]Gaia Collaboration, Arenou, F., Babusiaux, C., et al. 2023a, A&A, 674, A34
- [14]Gaia Collaboration, Panuzzo, P., Mazeh, T., Arenou, F., et al. 2024, Astronomy & Astrophysics, 686, L2
- [15]Gaia Collaboration, Vallenari, A., Brown, A. G. A., Prusti, T., et al. 2023b, A&A, 674, A1
- [16]Golub, G. H. & Pereyra, V. 1973, SIAM Journal on Numerical Analysis, 10, 413
- [17]Goodman, J. & Weare, J. 2010, Communications in Applied Mathematics and Computational Science, 5, 65
- [18]Green, R. M. 1985, Spherical Astronomy (Cambridge: Cambridge University Press)
- [19]Halbwachs, J.-L., Pourbaix, D., Arenou, F., et al. 2023, A&A, 674, A9
- [21]Holl, B., Perryman, M., Lindegren, L., Ségransan, D., & Raimbault, M. 2022, A&A, 661, A151
- [22]Holl, B., Sozzetti, A., Sahlmann, J., et al. 2023, A&A, 674, A10
- [23]Holland, P. W. & Welsch, R. E. 1977, Communications in Statistics - Theory and Methods, 6, 813
- [24]Kane, S. R., Henry, G. W., Dragomir, D., et al. 2011, The Astrophysical Journal Letters, 735, L41
- [26]Kiefer, F. 2019, Astronomy & Astrophysics, 632, L9
- [27]Lammers, C. & Winn, J. N. 2026, AJ, 171, 18
- [28]Lange, K. L., Little, R. J. A., & Taylor, J. M. G. 1989, Journal of the American Statistical Association, 84, 881
- [29]Leclerc, A., Babusiaux, C., Arenou, F., et al. 2023, A&A, 672, A82
- [30]Lindegren, L., Bastian, U., Biermann, M., et al. 2021, A&A, 649, A4
- [31]Lovell, M. C. 1963, Journal of the American Statistical Association, 58, 993
- [32]MacKay, D. J. C. 1992, Neural Computation, 4, 415
- [33]Perryman, M., Hartman, J., Bakos, G. Á., & Lindegren, L. 2014, ApJ, 797, 14
- [34]Pourbaix, D. & Jorissen, A. 2000, A&AS, 145, 161
- [35]Speagle, J. S. 2020, MNRAS, 493, 3132
- [36]Stefánsson, G., Mahadevan, S., Winn, J. N., et al. 2025, The Astronomical Journal, 169, 107
- [37]Thompson, W., Lawrence, J., Blakely, D., et al. 2023, AJ, 166, 164
- [38]Tierney, L. & Kadane, J. B. 1986, Journal of the American Statistical Association, 81, 82
- [40]Winn, J. N. 2022, The Astronomical Journal, 164, 196








