betaFeedback

WorksPhysics and Astronomy

Geodesic-informed neural networks for testing spacetimes with S2

Abstract

Stellar orbits around Sagittarius A*, in particular that of S2, test the spacetime of the Galactic Center. Such analyses usually integrate the equations of motion for every likelihood evaluation and sample the posterior. We develop a differentiable orbit model for static, spherically symmetric spacetimes that obtains the parameters, their uncertainties and the Bayesian evidence by optimization rather than sampling. In the Geodesic-Informed Neural Network (GINN), the conserved quantities, the shape of the orbit and the redshift are computed exactly from the metric, and a neural network constrained by the geodesic equation represents the time evolution of the radial phase. The physical parameters and the network weights minimize a single loss with Adam and Gauss-Newton steps, and the curvature at the minimum gives the uncertainties and the evidence. We test this Gaussian approximation against a second-order expansion and an exact geodesic solver, and apply GINN to the S2 data in the Schwarzschild, Reissner-Nordström, hairy Horndeski, Bardeen, Janis-Newman-Winicour and Yukawa spacetimes. GINN reproduces the estimates and uncertainties of the exact solver. For Schwarzschild we find M=(4.32±0.22)×106 M⊙M=(4.32\pm0.22)\times10^6\,M_\odot and D=8.33±0.21D=8.33\pm0.21 kpc. No spacetime deviates significantly from Schwarzschild; a Yukawa correction, for example, is limited to ∣δ∣<0.011|δ|<0.011 (95%) at a range of about 180 AU. No additional parameter improves the fit enough to pay its Occam penalty, so the Bayes factors range from inconclusive to moderate in favor of Schwarzschild. The Gaussian approximation is accurate when the hair enters the metric linearly and conservative when it enters only through its square. When the posterior is close to Gaussian, GINN gives the estimates, their uncertainties and the evidence from a single optimization, and the same derivatives quantify departures from it.

Introduction

Scientific progress depends on the ability to extract physical information from observations. Ideally, one would infer physical laws directly from the data, with few or no assumptions about the underlying model. In practice, statistical and systematic uncertainties, instrumental limitations and sparse sampling make such a direct reconstruction difficult, and the usual task is more modest. One adopts a physical model and infers the parameter values that best describe the observations, along with their uncertainties. When the observables depend non-analytically on the parameters, the posterior is commonly explored by sampling with Markov chain Monte Carlo (MCMC) or nested sampling. Sampling is robust, but it requires a large number of evaluations of the forward model and does not use its derivatives. When the forward model is differentiable, gradient-based optimizers and samplers, and uncertainty estimates from the local curvature of the posterior, become available. In exoplanet studies, for example, differentiable orbit models have made gradient-based inference routine [37, 77].

Machine learning offers flexible, differentiable models of this kind for modelling. In particular, artificial neural networks (ANNs) can approximate complex relations between physical quantities and observables, and automatic differentiation [9] provides the derivatives of the network outputs with respect to all inputs and weights. In astrophysics and cosmology, however, data sets are often small, while ANNs typically need large training sets to converge. More importantly, an unconstrained ANN has no knowledge of the physical laws governing the system and can return unphysical predictions. Physics-informed neural networks (PINNs) address both problems by including the governing differential equations in the training objective, so that the network is constrained to physically allowed solutions (Raissi et al. 2017a,b, 2019). In cosmology, PINNs have been trained to solve, for example, the background (Friedmann) equations over a range of parameters, so that they replace the numerical solver in fits to supernova, cosmic-chronometer and BAO data [16], to infer the dark-energy equation of state [79], and to reconstruct the Hubble function from supernova data [71]. In astrophysics, they have been used to solve the pulsar equation for axisymmetric, force-free pulsar magnetospheres [76].

Another application of PINNs, which is the focus of this work, is to infer the spacetime around the supermassive black hole at the center of our Galaxy, which can be directly probed by the orbits of the S-stars. Three decades of high-precision observations of the S-stars near Sagittarius A∗A^{*} (Sgr A∗A^{*}) with the ESO New Technology Telescope, the Very Large Telescope (VLT) and the Keck Observatory have provided robust evidence for a compact object of about 4×106M⊙4 \times10^{6} M_{\odot} [34, 40, 72, 41, 42, 39, 43], consistent with the horizon-scale images of the Event Horizon Telescope [35]. Among the S-stars, S2 plays a central role, owing to its short orbital period (∼16\sim16 yr) and high eccentricity (e∼0.88e \sim0.88). Around its 2018 pericenter passage, observations with the interferometer GRAVITY [45] and with Keck detected the combined gravitational redshift and transverse Doppler shift (GRAVITY Collaboration 2018; Do et al. 2019) and measured the distance to the Galactic Center to 0.3% [47]. In 2020 the GRAVITY collaboration detected the Schwarzschild precession of about 12′ per orbit, with fSP=1.10±0.19f_{\mathrm{SP}} = 1.10 \pm0.19 (fSP=1f_{\mathrm{SP}} = 1 in GR), consistent with General Relativity (GR) [2], and later multi-star fits limited any extended mass inside the orbit of S2 [3, 48].

The astrometric and radial-velocity data of S2 have been used to test a wide range of modified-gravity theories (Della Monica et al. 2022a; Monica et al. 2022; Benisty & Davis 2022; de Martino et al. 2021; Della Monica et al. 2022b; Zhang et al. 2024; Abd El Dayem et al. 2025; De Martino et al. 2024; Yan et al. 2024), alternative compact objects [28, 7, 4, 17, 85], and dark-matter distributions (Della Monica & de Martino 2023b,a; Shen et al. 2023; Abuter et al. 2022; Heißel et al. 2022; Becerra-Vergara et al. 2020) (see De Laurentis et al. 2023 for a review). Most recently, several static black hole and horizonless spacetimes were examined in a Bayesian framework, with no significant preference for any of them over Schwarzschild [60, 6]. In these analyses the orbit is obtained by integrating the equations of motion numerically and the posterior is explored by sampling, as in GC-OrbitFit [64], PyGRO [27] and SNOPy [49]. Each likelihood evaluation then requires a new numerical integration of the orbit, and the derivatives of the observables with respect to the parameters are not available.

In this work, we develop a differentiable forward model for the same problem, the Geodesic-Informed Neural Network (GINN). The conserved quantities, the shape of the orbit along with its precession, and the redshift are derived exactly from the metric at each point of the orbit, and the Rømer delay is included in the projection onto the sky. What remains is when the star reaches each point, that is, its radial motion in time, r(t)r(t). This is the relativistic generalization of Kepler’s problem and has no closed-form solution. GINN represents it by a neural network built on the Keplerian solution and trained jointly on the data and on the geodesic equation of the metric under study. Every predicted position and velocity therefore depends on the network output, and the whole model is differentiable with respect to all orbital, reference-frame and metric parameters. The posterior can then be located by gradient-based minimization of a single loss and approximated by a Gaussian (Laplace approximation) with the network weights eliminated through their response to the parameters, instead of being sampled. GINN can be easily applied to any static, spherically symmetric metric, and here we apply it to six spacetimes.

The paper is organized as follows. Section 2 lists the spacetimes and their parameters. The data are described in Sec. 3. Section 4 describes the orbit model, the network, the optimization and the uncertainty estimate. Section 5 presents the results, and Sec. 6 discusses their validation, and Sec. 7 concludes. The appendices give details of the optimization, the evidence, the higher-order checks and the validation tests.

Spacetimes

We consider static, spherically symmetric spacetimes,

ds2=−f(r) dt2+g(r) dr2+h(r)(dϑ2+sin⁡2ϑ dφ2),(1)ds^2=-f(r)\,dt^2+g(r)\,dr^2+h(r)(d\vartheta^2+\sin^2\vartheta\,d\varphi^2), \tag*{(1)}

in geometric units (G=c=1G = c = 1). The spin of Sgr A* affects the orbit of S2 at higher order than the Schwarzschild precession and is below the sensitivity of the available data [21]. Each spacetime is characterized by the mass MM and, except for Schwarzschild, one additional parameter ξ\xi, which we call the hair parameter. All of them reduce to Schwarzschild for vanishing hair (the Yukawa metric to its weak-field form). We follow a selection of recent MCMC studies [60, 6] so that the results can be compared directly. A spacetime enters the analysis only through the functions ff, gg and hh in the forward model of Sec. 4.

Since S2 never comes closer than its pericenter rp≃2800 Mr_p \simeq2800\,M, the observations probe only the exterior spacetime. Whether a horizon, a regular core or a singularity lies inside is not tested. We also treat the region outside the central object as empty, apart from the fields of the theory itself. This is supported by the upper limit of 1200 M⊙≃3×10−4M1200\,M_\odot\simeq3 \times10^{-4}M (at 1σ1\sigma confidence) on any extended mass, such as a stellar cusp or a dark-matter spike, inside the orbit of S2 [48]. In GR, Birkhoff’s theorem and its electrovacuum extension make the Schwarzschild and RN metrics the exterior of any spherical distribution of mass or charge. For the other theories, no such uniqueness result is available, so our analysis tests the specific metrics rather than every object these theories allow.

Schwarzschild spacetime

The unique spherically symmetric vacuum solution of GR [73],

f(r)=1−2Mr,g(r)=f(r)−1,h(r)=r2,(2)f(r)=1-\frac{2M}{r},\qquad g(r)=f(r)^{-1},\qquad h(r)=r^2, \tag*{(2)}

is the reference model. It has a horizon at r=2Mr = 2M, an unstable circular photon orbit at r=3Mr = 3M and an innermost stable circular orbit at r=6Mr = 6M. The pericenter of S2, rp≃2800 Mr_p \simeq2800\,M, lies far outside this strong-field region, so S2 probes the weak field, where M/rp≃4×10−4M/r_p \simeq4 \times10^{-4} sets the size of the relativistic effects in the orbit.

Reissner-Nordström spacetime

The exterior of a spherical mass with electric charge QQ [70, 63] is

f(r)=1−2Mr+Q2r2,g(r)=f(r)−1,h(r)=r2,(3)f(r)=1-\frac{2M}{r}+\frac{Q^2}{r^2},\qquad g(r)=f(r)^{-1},\qquad h(r)=r^2, \tag*{(3)}

with hair parameter qRN=Q/Mq_{\mathrm{RN}} = Q/M. It contains a black hole for ∣qRN∣≤1|q_{\mathrm{RN}}| \le1 and a naked singularity for ∣qRN∣>1|q_{\mathrm{RN}}| > 1. The charge enters only through qRN2q_{\mathrm{RN}}^2, so its sign cannot be measured. In the radial equation (see (11)), the Q2/r2Q^2/r^2 term in ff adds to the centrifugal term L2/r2L^2/r^2, so for given turning points its leading effect on the radial motion is absorbed by LL (Fig. 2). The electric charge of Sgr A* is expected to be negligible [84], so a bound on qRNq_{\mathrm{RN}} is best read as a bound on any positive $1/r^2$ term in ff. The same metric with a coefficient β\beta of either sign in place of Q2Q^2 describes braneworld black holes with tidal charge [19], whereas our parametrization covers only $\beta = Q^2 \ge 0.

Hairy Horndeski spacetime

Horndeski gravity is the most general scalar-tensor gravity with second-order field equations [53]. Its quartic sector admits the exact hairy solution [12]

f(r)=1−2Mr+QHrln⁡(r2M),g(r)=f(r)−1,h(r)=r2,(4)f(r)=1-\frac{2M}{r}+\frac{Q_H}{r}\ln\left(\frac{r}{2M}\right),\qquad g(r)=f(r)^{-1},\qquad h(r)=r^2, \tag*{(4)}

with hair parameter qH=QH/Mq_H=Q_H/M. For qH>−2q_H>-2, which includes all values considered here, the horizon stays at r=2Mr=2M. The hair modifies the Newtonian potential itself, with m(r)=M−(QH/2)ln⁡(r/2M)m(r)=M-(Q_H/2)\ln(r/2M). Since m(r)m(r) has no finite limit as r→∞r\to\infty, MM is the parameter that fixes the horizon rather than an asymptotic mass. Over the orbit of S2 the logarithm grows from about 7 at pericenter to about 10 at apocenter, so the star feels a different mass at the two ends of its orbit. The constant part of this shift is indistinguishable from a change of MM, which makes qHq_H partly degenerate with the mass (and sign of qHq_H matters).

Bardeen spacetime

The Bardeen spacetime [8] replaces the central singularity with a regular core and can be interpreted as a magnetic monopole of charge QBQ_B in nonlinear electrodynamics [5]. In that case, the metric coefficients read as

f(r)=1−2Mr2(r2+QB2)3/2,g(r)=f(r)−1,h(r)=r2,(5)f(r)=1-\frac{2Mr^2}{(r^2+Q_B^2)^{3/2}},\qquad g(r)=f(r)^{-1},\qquad h(r)=r^2, \tag*{(5)}

with hair parameter qB=QB/Mq_B=Q_B/M. Its mass function m(r)=Mr3/(r2+QB2)3/2m(r)=Mr^3/(r^2+Q_B^2)^{3/2} vanishes at the center, which is a regular de Sitter core, and approaches MM for r≫QBr\gg Q_B. It has horizons for ∣qB∣≤4/(33)≃0.77|q_B|\le4/(3\sqrt{3})\simeq0.77 and is a horizonless regular object otherwise. At large radii m(r)≃M[1−3QB2/(2r2)]m(r)\simeq M[1-3Q_B^2/(2r^2)], so the charge enters only through qB2q_B^2, and at the pericenter of S2 the relative change of the mass is about 2×10−7qB22\times10^{-7}q_B^2. Only ∣qB∣≫1|q_B|\gg1 can therefore affect the orbit, which lies deep in the horizonless regime.

Janis-Newman-Winicour (JNW) spacetime

The JNW solution describes a mass with a minimally coupled1, massless scalar field of charge QJNWQ_{\rm JNW} [54]. In the notation of Virbhadra [80],

f(r)=g(r)−1=(1−br)γ,h(r)=r2(1−br)1−γ,(6)f(r)=g(r)^{-1}=\left(1-\frac{b}{r}\right)^\gamma,\qquad h(r)=r^2\left(1-\frac{b}{r}\right)^{1-\gamma}, \tag*{(6)}

with b=2M2+QJNW2b=2\sqrt{M^2+Q_{\rm JNW}^2}, γ=2M/b\gamma=2M/b and hair parameter qJNW=QJNW/Mq_{\rm JNW}=Q_{\rm JNW}/M. For qJNW≠0q_{\rm JNW}\ne0, the surface r=br=b is a naked curvature singularity, and a photon sphere exists only for ∣qJNW∣<3|q_{\rm JNW}|<\sqrt{3} [81]. Since rr is not the areal radius R=hR=\sqrt{h}, we expand in RR: f=1−2M/R+MQJNW2/(3R3)+…f=1-2M/R+MQ_{\rm JNW}^2/(3R^3)+\ldots and gRR≡g(r)(dr/dR)2=1+2M/R+(4M2−QJNW2)/R2+…g_{RR}\equiv g(r)(dr/dR)^2=1+2M/R+(4M^2-Q_{\rm JNW}^2)/R^2+\ldots. The charge enters only through qJNWq_{\rm JNW}, with corrections to Schwarzschild suppressed by QJNW2/R2∼10−7qJNW2Q_{\rm JNW}^2/R^2\sim10^{-7}q_{\rm JNW}^2 at the radii of S2. Only ∣qJNW∣≫1|q_{\rm JNW}|\gg1 can therefore affect the orbit, lying far outside the photon sphere regime. Thus, S2 and horizon-scale imaging probe different regions of the JNW parameter space.

Yukawa-type correction

In f(R)f(R) gravity and other theories with a massive scalar degree of freedom, the weak-field potential acquires a Yukawa term (De Martino et al. 2018; De Laurentis et al. 2018; De Martino et al. 2021),

f(r)=1+Φ(r),g(r)=1−Φ(r),h(r)=r2,f(r)=1+\Phi(r),\qquad g(r)=1-\Phi(r),\qquad h(r)=r^2,
Φ(r)=−2M(1+δ)r(1+δe−r/λ),(7)\Phi(r)=-\frac{2M}{(1+\delta)r}\left(1+\delta e^{-r/\lambda}\right), \tag*{(7)}

with strength δ\delta and range λ\lambda, the Compton wavelength of the extra degree of freedom. The mass function m(r)=M(1+δe−r/λ)/(1+δ)m(r)=M(1+\delta e^{-r/\lambda})/(1+\delta) equals MM for r≪λr\ll\lambda and M/(1+δ)M/(1+\delta) for r≫λr\gg\lambda. For λ\lambda much larger than the orbit,

Φ(r)≃−2Mr+2δM(1+δ)λ−δMr(1+δ)λ2.(8)\Phi(r)\simeq-\frac{2M}{r}+\frac{2\delta M}{(1+\delta)\lambda}-\frac{\delta Mr}{(1+\delta)\lambda^2}. \tag*{(8)}

The constant term does not affect the Newtonian motion and only shifts the redshift by a constant, which is absorbed by the radial-velocity offset (Sec. 4). The last term is a uniform radial force whose strength depends only on δ/[(1+δ)λ2]\delta/[(1+\delta)\lambda^2]. Since δ\delta and λ\lambda cannot be estimated separately in this regime, we fix λ\lambda and take ξ≡δ\xi\equiv\delta (Sec. 5). Yukawa-type corrections have previously been constrained with S-stars by Borka et al. [14]; Hees et al. [51]; de Martino et al. (2021); Abd El Dayem et al. [1]; Jovanović et al. [55].

Data

We use the public astrometric and spectroscopic data of S2 from Gillessen et al. [43], obtained from the catalog of Gillessen et al. [44]. The dataset consists of 145 astrometric positions and 44 line-of-sight radial velocity measurements between 1992 and 2016, that is, N=2×145+44=334N=2\times145+44=334 data points. The astrometry includes speckle imaging with the SHARP camera at the ESO New Technology Telescope (NTT) before 2002 with adaptive-optics imaging with NACO at the Very Large Telescope (VLT) from 2002 onward. The radial velocities come mostly from the integral-field spectrograph SINFONI at the VLT, complemented by earlier measurements (before 2003) with NIRC2 at the Keck Observatory in slit-spectroscopy mode. We take all uncertainties as published and treat them as independent and Gaussian.

The positions of S2 are measured in near-infrared images relative to an astrometric reference frame tied to the radio frame through SiO maser stars (evolved stars detected at both infrared and radio wavelengths) [65]. The orbit of S2 is centered on the dynamical center, whose position and motion relative to the frame origin are known only to the accuracy of this calibration. We therefore fit five offset parameters together with the orbit: the position (x0,y0)(x_0,y_0) and proper motion (vx,0,vy,0)(v_{x,0},v_{y,0}) of the dynamical center, and a constant offset vz,0v_{z,0} of the radial velocities. The first four have Gaussian priors of zero mean and widths 0.2 mas (at epoch 2009.0) and 0.1 mas yr−1^{-1} from this calibration, as in Gillessen et al. [43]; vz,0v_{z,0} has no prior.

Method

Figure 1 summarizes the method, whose steps are described in the following subsections.

Workflow of GINN

Figure 1. Workflow of GINN. Everything that can be computed exactly from the metric (blue) is computed exactly; the network (orange) represents only the radial phase ϕ(t)\phi(t) and is constrained by the geodesic equation through the physics residual. All parameters minimize one joint loss in three stages (green: Keplerian pre-fit, Adam, Gauss–Newton); the curvature of the data and frame-prior residuals at the minimum, with the network following the parameters, gives the Laplace uncertainties.

Orbit modeling

The orbit of S2 is described by the six orbital elements: semi-major axis aa, eccentricity ee, inclination ii, longitude of the ascending node Ω\Omega, argument of pericenter ω\omega and time of pericenter passage tpt_p. The spacetime is fixed by the mass MM and the hair parameter of Sec. 2,

ξ∈{qRN,qH,qB,qJNW,δ}.(9)\xi\in\{q_{\mathrm{RN}}, q_{\mathrm{H}}, q_{\mathrm{B}}, q_{\mathrm{JNW}}, \delta\}. \tag*{(9)}

Distances are converted to angles with the distance DD to the Galactic Center. The parameters MM, aa and DD must be positive, and a bound orbit requires 0<e<10 < e < 1. We therefore work with the transformed variables ln⁡(M/M⊙)\ln(M/M_\odot), ln⁡(a/AU)\ln(a/\mathrm{AU}), ln⁡(D/kpc)\ln(D/\mathrm{kpc}) and logit⁡e=ln⁡[e/(1−e)]\operatorname{logit} e = \ln[e/(1-e)]. The constraints are then satisfied automatically, without bounds in the optimization, and all parameters have a similar numerical range, which improves the conditioning of the optimization. Together with the offset parameters (Sec. 3), with (x0,y0)(x_0,y_0) given at the reference epoch treft_{\mathrm{ref}}, the complete parameter vector is

Θ={ln⁡M,ln⁡a,logit⁡e,i,Ω,ω,tp,ln⁡D,x0,y0,vx,0,vy,0,vz,0,ξ}.(10)\Theta= \{\ln M, \ln a, \operatorname{logit} e, i, \Omega, \omega, t_p, \ln D, x_0, y_0, v_{x,0}, v_{y,0}, v_{z,0}, \xi\}. \tag*{(10)}

Spherical symmetry confines a geodesic to a plane, which we take as ϑ=π/2\vartheta= \pi/2. Time-translation and rotational invariance of Eq. (1) give two conserved quantities per unit rest mass, the specific energy E=f(r)t˙E = f(r)\dot{t} and angular momentum L=h(r)ϕ˙L = h(r)\dot{\phi}, where dots denote derivatives with respect to proper time τ\tau. The normalization gμνx˙μx˙ν=−1g_{\mu\nu}\dot{x}^{\mu}\dot{x}^{\nu} = -1 leads to the radial equation of motion,

g(r)(drdτ)2=E2f(r)−1−L2h(r)≡V(r).(11)g(r)\left(\frac{dr}{d\tau}\right)^2 = \frac{E^2}{f(r)} - 1 - \frac{L^2}{h(r)} \equiv V(r). \tag*{(11)}

A bound orbit oscillates between the pericenter rp=a(1−e)r_p = a(1-e) and the apocenter2{}^{2} ra=a(1+e)r_a = a(1+e), where V(rp)=V(ra)=0V(r_p) = V(r_a) = 0. These two conditions are linear in E2E^2 and L2L^2,

E2f(ra)−L2h(ra)=1,E2f(rp)−L2h(rp)=1,(12)\frac{E^2}{f(r_a)}-\frac{L^2}{h(r_a)}=1,\qquad\frac{E^2}{f(r_p)}-\frac{L^2}{h(r_p)}=1, \tag*{(12)}

and determine EE and LL uniquely. The elements aa and ee are thus defined through the turning points of the areal radius, which coincides with rr for h=r2h=r^2.

Figure 2 compares the radial potential V(r)V(r) across the spacetimes of Sec. 2, with EE and LL fixed by the common turning points. The shape of VV between the turning points sets the radial speed, and gg rescales it in Eq. (11) without explicitly entering V(r)V(r).

Figure 2. Radial function V(r)V(r) of Eq. (11) for the S2 orbit in each spacetime, with the turning points rpr_p and rar_a marked by the vertical lines. The hair parameters are chosen for illustrative purposes to distinguish the curves.

Instead of solving for r(τ)r(\tau) directly, we introduce the relativistic anomaly ϕ\phi [20, 18],

r(ϕ)=p1+ecos⁡ϕ,p≡a(1−e2).(13)r(\phi) = \frac{p}{1+e\cos\phi},\qquad p \equiv a(1-e^2). \tag*{(13)}

This is an exact relabeling of the radial coordinate between the two turning points, valid for any metric, and in Newtonian gravity ϕ\phi is the true anomaly. With dt/dτ=E/fdt/d\tau= E/f, Eq. (11) gives the rate of ϕ\phi in coordinate time,

dϕdt=f(r)EV(r)g(r)∣dr(ϕ)dϕ∣−1≡F(ϕ;Θ),(14)\frac{d\phi}{dt} = \frac{f(r)}{E}\sqrt{\frac{V(r)}{g(r)}}\left|\frac{dr(\phi)}{d\phi}\right|^{-1} \equiv F(\phi;\Theta), \tag*{(14)}

which, for given parameters, is a closed-form function of ϕ\phi. Both VV and dr/dϕdr/d\phi vanish at the turning points while their ratio remains finite, so we evaluate VV relative to the nearest turning point where the leading constant terms cancel exactly. The radial period and the radial frequency follow from the rate,

Pr=∫02πdϕF(ϕ;Θ),nr=2πPr.(15)P_r = \int_{0}^{2\pi} \frac{d\phi}{F(\phi;\boldsymbol{\Theta})}, \qquad n_r = \frac{2\pi}{P_r}. \tag*{(15)}

Similarly, conservation of specific angular momentum dϕ/dτ=L/h(r)d\phi/d\tau= L/h(r), gives the shape of the orbit,

φ(ϕ)=∫0ϕL/h(r(ϕ′))dϕ′/dτ dϕ′.(16)\varphi(\phi) = \int_{0}^{\phi} \frac{L/h(r(\phi'))}{d\phi'/d\tau}\,d\phi'. \tag*{(16)}

The pericenter advance per radial period is then Δϕ=φ(2π)−2π\Delta\phi= \varphi(2\pi) - 2\pi. In Newtonian gravity (where ϕ=φ\phi= \varphi), Δϕ=0\Delta\phi= 0, and in the Schwarzschild case Δϕ≃6πM/p\Delta\phi\simeq6\pi M/p to leading order. We evaluate both integrals (15) and (16) with Gauss-Legendre quadrature.

The shape of the orbit, including its precession, is therefore known exactly from the metric. What remains is its time dependence ϕ(t)\phi(t), i.e. the inverse of t(ϕ)=∫F−1dϕt(\phi) = \int F^{-1}d\phi. Even in Newtonian gravity, where this inversion is Kepler’s equation, E−esin⁡E=n(t−tp)E-e\sin E=n(t-t_p) for the eccentric anomaly EE, has no closed-form solution and is strongly non-linear at the eccentricity of S2. This is the part of the problem that we represent with a neural network (Sec. 4.3).

From (t) to the observables

Once ϕ(t)\phi(t) is known, the observables follow from the following steps.

  1. At each epoch we compute the areal radius R=h(r)R = \sqrt{h(r)} and the azimuth φ(ϕ)\varphi(\phi), and hence the position coordinates X=Rcos⁡φX = R\cos\varphi, Y=Rsin⁡φY = R\sin\varphi in the orbital plane, with XX pointing to pericenter. The velocities follow from the conserved quantities,

vR=±dRdrV(r)/g(r)dt/dτ,vϕ=RL/h(r)dt/dτ,(17)v_R = \pm\frac{dR}{dr}\frac{\sqrt{V(r)/g(r)}}{dt/d\tau}, \qquad v_\phi= R\frac{L/h(r)}{dt/d\tau}, \tag*{(17)}

with dR/dr=h′/(2h)dR/dr = h'/(2\sqrt{h}), which equals one for h=r2h = r^2, and vR>0v_R > 0 while the star moves outward (sin⁡ϕ>0\sin\phi> 0). In Cartesian coordinates, vX=vRcos⁡φ−vϕsin⁡φv_X = v_R\cos\varphi- v_\phi\sin\varphi and vY=vRsin⁡φ+vϕcos⁡φv_Y = v_R\sin\varphi+ v_\phi\cos\varphi.

  1. The orbital plane is projected on the sky with the unit vectors of the XX and YY axes,

eX=(cos⁡Ωcos⁡ω−sin⁡Ωsin⁡ωcos⁡isin⁡Ωcos⁡ω+cos⁡Ωsin⁡ωcos⁡isin⁡ωsin⁡i),eY=(−cos⁡Ωsin⁡ω−sin⁡Ωcos⁡ωcos⁡i−sin⁡Ωsin⁡ω+cos⁡Ωcos⁡ωcos⁡icos⁡ωsin⁡i).(18)\mathbf{e}_X = \begin{pmatrix} \cos\Omega\cos\omega-\sin\Omega\sin\omega\cos i\\ \sin\Omega\cos\omega+\cos\Omega\sin\omega\cos i\\ \sin\omega\sin i \end{pmatrix}, \qquad \mathbf{e}_Y = \begin{pmatrix} -\cos\Omega\sin\omega-\sin\Omega\cos\omega\cos i\\ -\sin\Omega\sin\omega+\cos\Omega\cos\omega\cos i\\ \cos\omega\sin i \end{pmatrix}. \tag*{(18)}

whose components are the Thiele-Innes constants, so that

(xDec,xRA,Z)T=XeX+YeY,vz=vX(eX)3+vY(eY)3,(19)(x_{\mathrm{Dec}},x_{\mathrm{RA}},Z)^{\mathsf{T}} = X\mathbf{e}_X + Y\mathbf{e}_Y, \qquad v_z = v_X(\mathbf{e}_X)_3 + v_Y(\mathbf{e}_Y)_3, \tag*{(19)}

where ZZ and vzv_z are positive away from the observer.

  1. Light received at time tobst_{\mathrm{obs}} was emitted at tem=tobs−Z(tem)/ct_{\mathrm{em}} = t_{\mathrm{obs}}-Z(t_{\mathrm{em}})/c. We account for the Rømer delay between emission and reception. Identifying tt with the distant-observer time, up to a constant absorbed into tpt_p, we solve this relation with one fixed-point iteration, tem≃tobs−Z(tobs)/ct_{\mathrm{em}} \simeq t_{\mathrm{obs}} - Z(t_{\mathrm{obs}})/c whose timing error is of order ∣Zv∣/c2|Z_v|/c^2, below one hour for S2. We evaluate the orbital quantities at ϕ(tem)\phi(t_{\mathrm{em}}).

  1. The positions are converted to angles with DD, and the frame offsets are added,

RAmod(tobs)=xRAD+x0+vx,0(tobs−tref),\mathrm{RA}^{\mathrm{mod}}(t_{\mathrm{obs}}) = \frac{x_{\mathrm{RA}}}{D} + x_0 + v_{x,0}(t_{\mathrm{obs}}-t_{\mathrm{ref}}),
Decmod(tobs)=xDecD+y0+vy,0(tobs−tref).(20)\mathrm{Dec}^{\mathrm{mod}}(t_{\mathrm{obs}}) = \frac{x_{\mathrm{Dec}}}{D} + y_0 + v_{y,0}(t_{\mathrm{obs}}-t_{\mathrm{ref}}). \tag*{(20)}

where treft_{\mathrm{ref}} is the mean astrometric epoch. The line-of-sight velocity follows from the frequency shift seen by a static observer at infinity,

1+z=dtdτ(1+vz)=Ef(r)(1+vz),vLOSmod=cz+vz,0.(21)1+z = \frac{dt}{d\tau}(1+v_z) = \frac{E}{f(r)}(1+v_z), \qquad v_{\mathrm{LOS}}^{\mathrm{mod}} = cz+v_{z,0}. \tag*{(21)}

The factor E/fE/f contains the gravitational redshift and the transverse Doppler shift, and (1+vz)(1+v_z) the longitudinal Doppler shift. Since both are evaluated with the metric under study, the redshift is consistent with each spacetime. We neglect the Shapiro delay, which is below the precision of these data.

Geodesic-informed neural network (GINN)

We write the relativistic anomaly as a Keplerian part plus a neural-network correction,

ϕ(t)=ϕK(nr(t−tp);e)+Nθ(ϕK)−Nθ(0).(22)\phi(t) = \phi_K(n_r(t-t_p);e) + \mathcal{N}_{\theta}(\phi_K) - \mathcal{N}_{\theta}(0). \tag*{(22)}

The Keplerian part ϕK\phi_K is the true anomaly obtained from Kepler’s equation with the mean anomaly nr(t−tp)n_r(t-t_p), where nrn_r is the exact radial frequency of the metric, (15), so that ϕK\phi_K already has the correct period. The network Nθ\mathcal{N}_{\theta}, with weights θ\theta, is fully connected, takes (cos⁡ϕK,sin⁡ϕK)(\cos\phi_K,\sin\phi_K) as input, has two hidden layers of 32 tanh units and a linear output scaled by 10−310^{-3} rad, about 1200 weights in total. Because its input is periodic, the correction repeats every radial period and cannot produce a secular drift. Subtracting Nθ(0)\mathcal{N}_{\theta}(0) makes the correction vanish at pericenter, so that tpt_p keeps its meaning and is not degenerate with a constant network output. Since the precession is already carried by φ(ϕ)\varphi(\phi), the network represents the small, periodic, non-Keplerian part of the radial motion, ϕ(t)−ϕK(t)\phi(t)-\phi_K(t).

The network is tied to the equation of motion (14) through the relative residual

Rc=ϕ˙NN−F(ϕ;Θ)F(ϕ;Θ),R_c = \frac{\dot{\phi}_{\mathrm{NN}}-F(\phi;\boldsymbol{\Theta})}{F(\phi;\boldsymbol{\Theta})},

with,

ϕ˙NN=[1+Nθ′(ϕK)]nr(1+ecos⁡ϕK)2(1−e2)3/2,(23)\dot{\phi}_{\mathrm{NN}} = \left[1+\mathcal{N}'_{\theta}(\phi_K)\right]\frac{n_r(1+e\cos\phi_K)^2}{(1-e^2)^{3/2}}, \tag*{(23)}

where ϕ\phi is the anomaly of (22) and the last factor is the Keplerian rate ϕ˙K\dot{\phi}_K. The derivative Nθ′\mathcal{N}'_{\theta} is computed exactly by automatic differentiation [9]. We evaluate RcR_c at Nc=256N_c=256 collocation phases, spaced uniformly in the Keplerian eccentric anomaly over one radial period and shifted together by a random amount at each training step. The random shift enforces the equation of motion at all phases rather than at a fixed set of points. This stochastic resampling is related to sampling strategies used in PINNs (Wu et al. 2023). Because $R_c is a relative error of the rate, it has the same weight near pericenter, where ϕ\phi changes fast, and near apocenter, where it changes slowly.

Unlike a standard PINN, which represents the full solution, GINN computes exactly everything that follows from the metric (EE, LL, nrn_r, φ(ϕ)\varphi(\phi) and the redshift) and leaves only ϕ(t)−ϕK(t)\phi(t)-\phi_K(t) to the network. For S2, this correction is about 10−410^{-4} rad in Schwarzschild, or ≲0.02\lesssim0.02 mas on the sky, below the present astrometric precision. The network therefore adds no information beyond an exact solution of the equation of motion. Its role is to make the forward model differentiable with respect to all parameters, so that they can be estimated jointly by gradient-based optimization. The same construction applies to metrics for which inverting t(ϕ)t(\phi) is inconvenient, and to more precise data, for which the correction becomes measurable.

Loss and priors

All parameters are obtained by minimizing a single loss,

L(Θ,θ)=χdata2+χframe2+∑c=1Nc(Rcϵ)2+∥θ∥2σw2,(24)\mathcal{L}(\Theta,\theta)=\chi^2_{\mathrm{data}}+\chi^2_{\mathrm{frame}}+\sum_{c=1}^{N_c}\left(\frac{R_c}{\epsilon}\right)^2+\frac{\lVert\theta\rVert^2}{\sigma_w^2}, \tag*{(24)}

with

χdata2=∑i[(RAimod−RAiσRA,i)2+(Decimod−DeciσDec,i)2]+∑j(vLOS,jmod−vLOS,jσv,j)2,(25)\chi^2_{\mathrm{data}}=\sum_i\left[\left(\frac{\mathrm{RA}^{\mathrm{mod}}_i-\mathrm{RA}_i}{\sigma_{\mathrm{RA},i}}\right)^2+\left(\frac{\mathrm{Dec}^{\mathrm{mod}}_i-\mathrm{Dec}_i}{\sigma_{\mathrm{Dec},i}}\right)^2\right]+\sum_j\left(\frac{v^{\mathrm{mod}}_{\mathrm{LOS},j}-v_{\mathrm{LOS},j}}{\sigma_{v,j}}\right)^2, \tag*{(25)}

where ii runs over the astrometric epochs and jj over the radial-velocity epochs. The Gaussian frame priors of § 3 enter as χframe2\chi^2_{\mathrm{frame}}, the sum of four squared residuals: the position of the dynamical center at epoch 2009.0, x0+vx,0(2009.0−tref)x_0+v_{x,0}(2009.0-t_{\mathrm{ref}}) and likewise in yy, divided by 0.2 mas; and vx,0v_{x,0} and vy,0v_{y,0} divided by 0.1 mas yr−1^{-1}. They are treated as four additional data residuals throughout, in the optimization, the uncertainties and the evidence. The physics term has weight 1/ϵ21/\epsilon^2. It enforces the geodesic equation up to a relative tolerance ϵ\epsilon, analogous to the tolerance set in an ODE solver. We set ϵ=10−4\epsilon=10^{-4} by a convergence test: for 10−4≲ϵ≲10−310^{-4}\lesssim\epsilon\lesssim10^{-3} all results agree with an exact geodesic solver and do not depend on ϵ\epsilon. A loose tolerance lets the network adapt to noise, and a tighter one makes the problem too stiff to optimize (Appendix § 11). The last term is a weak Gaussian prior on the network weights (σw=10\sigma_w=10) [58]. It has no noticeable effect on the fit, but makes the curvature matrix invertible: the network has more weights than there are residuals (334+4+256=594<1200334+4+256=594<1200), so without it the curvature of the loss (the Gauss–Newton matrix) is singular in the weight directions, and the prior adds σw−2\sigma_w^{-2} to its diagonal (see § 4.5).

The hair parameters have the uniform priors of Table 1. A uniform prior is constant inside its interval and adds nothing to the loss or its curvature. We enforce the bounds by projection; that is, an update that leaves the interval is set back to the boundary. The two kinds of prior therefore act differently. A uniform prior is a hard wall that is felt only when a step reaches it, whereas a Gaussian frame prior is soft: it adds to the loss and its curvature everywhere and pulls the offsets towards zero with a strength set by its width, so that values beyond the width are allowed but penalized. Apart from the frame priors, the mass, the distance, the orbital elements and vz,0v_{z,0} have no prior term in the loss; that is, their priors are flat in the fitted variables of Θ\Theta. Since these parameters are common to all spacetimes, their priors cancel in the Bayes factors.

SpacetimeHair parameter ξ\xiPrior
Reissner–NordströmqRN=Q/Mq_{\mathrm{RN}}=Q/MU[−50,50]\mathcal{U}[-50,50]
Hairy HorndeskiqH=QH/Mq_H=Q_H/MU[−0.05,0.05]\mathcal{U}[-0.05,0.05]
BardeenqB=QB/Mq_B=Q_B/MU[−200,200]\mathcal{U}[-200,200]
JNWqJNW=QJNW/Mq_{\mathrm{JNW}}=Q_{\mathrm{JNW}}/MU[−600,600]\mathcal{U}[-600,600]
Yukawaδ\delta (at fixed λ\lambda)U[−0.5,0.5]\mathcal{U}[-0.5,0.5]

Table 1. Hair parameters and their uniform priors. For qRNq_{\mathrm{RN}}, qBq_B and qJNWq_{\mathrm{JNW}} the likelihood depends only on ξ2\xi^2, so ξ\xi and −ξ-\xi describe the same spacetime. We nevertheless use a prior symmetric about zero rather than, e.g., U[0,50]\mathcal{U}[0,50]: both give the same posterior for ∣ξ∣|\xi| and the same evidence, but the symmetric prior keeps the maximum at ξ=0\xi=0 inside the prior rather than on its boundary, where the Laplace approximation would not apply. For qHq_H and δ\delta the sign is physical.

Optimization

The optimization has to deliver two things: the minimum of the loss, which is the maximum of the posterior, and the width of the posterior around it, which gives the uncertainties. Although the loss is defined over the full parameter space, its gradient and curvature describe its local behaviour at a given point. While the former indicates the direction in which the loss decreases, the latter describes how rapidly it rises away from the minimum and, hence, how tightly the data constrain each combination of parameters. An estimate of the curvature therefore serves three purposes: it sets the length and direction of each step, it tells when the minimum has been reached, and it gives the error bars (see § 4.6). Implementation details are given in Appendix § 8.

A first-order optimizer such as Adam [56] uses only the gradient, rescaled separately for each parameter, and the length of its steps is set by a learning rate. This works well when the parameters are roughly independent, but in the orbit problem the physical parameters are strongly correlated, most notably MM, DD and aa, and a hair parameter, where present, is correlated with MM. Along such correlated directions the appropriate step length differs by orders of magnitude, according to how tightly the data constrain each direction, and a single learning rate cannot capture that. The physics term in Eq. (24) also makes the problem stiff. With ϵ=10−4\epsilon=10^{-4}, a change of the hair that alters the rate FF by 10−310^{-3} raises each physics residual Rc/ϵR_c/\epsilon by about 10, unless the network changes at the same time. The minimum therefore lies in a narrow, curved valley in the joint space of physical parameters and network weights, and moving along it requires a coordinated change of the hair, the other physical parameters and the network. Adam progresses only slowly along this valley.

The curvature is available at little extra cost because the loss is a sum of squares. Let r\mathbf{r} be the vector of the normalized data residuals, frame-prior terms and the physics residuals Rc/ϵR_c/\epsilon in Eq. (24), so that L=rTr+∥θ∥2/σw2\mathcal{L}=\mathbf{r}^{\mathsf{T}}\mathbf{r}+\lVert\theta\rVert^2/\sigma_w^2. With JJ, the Jacobian of r\mathbf{r} with respect to the physical parameters u\mathbf{u} in the whitened coordinates (of Eq. (A.1)), and to the weights θ\boldsymbol{\theta}, the Hessian (matrix of second derivatives) of the loss, ∇2L\nabla^2\mathcal{L}, is,

∇2L=2(JTJ+∑krk∇2rk+Dw),(26)\nabla^2\mathcal{L}=2\left(J^{\mathsf{T}}J+\sum_k r_k\nabla^2r_k+D_w\right), \tag*{(26)}

where DwD_w is zero on u\mathbf{u} and the identity divided by σw2\sigma_w^2 on the weights. The frame-prior terms are linear in physical parame­ters. ters uu so their second derivatives vanish. Near the minimum, the physics residuals are nearly zero, and the data residuals scatter with random signs, so their contributions to the second-derivative term largely cancel and it can be dropped (we quantify its effect in Appendix § 10). This gives the Gauss–Newton (GN) approximation of the curvature, together with the gradient [62],

H=JTJ+Dw=(HuuHuθHθuHθθ),g=JTr+1σw2(0θ)=(gugθ).(27)H = J^{\mathsf T}J + D_w = \begin{pmatrix} H_{uu} & H_{u\theta} \\ H_{\theta u} & H_{\theta\theta} \end{pmatrix}, \qquad\mathbf{g} = J^{\mathsf T}\mathbf{r} + \frac{1}{\sigma_w^2}\begin{pmatrix} 0 \\ \theta\end{pmatrix} = \begin{pmatrix} g_u \\ g_\theta\end{pmatrix}. \tag*{(27)}

Notice that HH needs only first derivatives, which automatic differentiation provides, and it is positive semi-definite. For Gaussian data, its χdata2\chi^2_{\mathrm{data}} part is the Fisher information matrix (with the network following the parameters), whose inverse is the covariance of the parameters.

With HH and g\mathbf{g}, a small change of the parameters and weights, Δ≡(Δu,Δθ)\Delta\equiv(\Delta u,\Delta\theta), changes the loss to approximately

L(Δ)≃L(0)+2gTΔ+ΔTHΔ.(28)\mathcal{L}(\Delta) \simeq\mathcal{L}(0) + 2\mathbf{g}^{\mathsf T}\Delta+ \Delta^{\mathsf T}H\Delta. \tag*{(28)}

Setting its derivative to zero defines the step, Δ=−H−1g\Delta=-H^{-1}\mathbf{g}, at which the quadratic reaches its minimum, L(Δ)=L(0)−gTH−1g\mathcal{L}(\Delta)=\mathcal{L}(0)-\mathbf{g}^{\mathsf T}H^{-1}\mathbf{g}. The step is long along directions that the data constrain weakly and short along those they constrain tightly; it follows correlated directions, and it needs no learning rate. Its predicted decrease, gTH−1g\mathbf{g}^{\mathsf T}H^{-1}\mathbf{g}, is a change of χ2\chi^2. Since a change of Δχ2=1\Delta\chi^2=1 corresponds to one standard deviation, a predicted decrease d2d^2 means that the minimum is still about dσd\sigma away, which gives a natural stopping criterion.

The quadratic model, Eq. (28), follows from linearizing the residuals, r(Δ)≃r+JΔ\mathbf{r}(\Delta)\simeq\mathbf{r}+J\Delta, and is exact only if the residuals depend linearly on the parameters, that is, if JJ is the same everywhere. Far from the minimum, it can be poor, which is why the GN steps are taken only after an Adam stage (see below). At the minimum, terms of higher order would make the posterior skewed or bend the degeneracies. We tested this by carrying the expansion of the residuals to second order (Appendix § 10). The Gaussian reproduces the means and is slightly conservative in the widths; only for a hair that enters the metric through its square is the posterior noticeably non-Gaussian (§ 6).

We therefore proceed in two stages. Starting from a Keplerian orbit determined from the astrometry, Adam first trains all parameters and weights jointly and brings them close to the minimum; it is robust to a random initialization of the network and tolerates the random shift of the collocation points. GN steps then complete the minimization in the physical parameters. Minimizing Eq. (28) over the weights first (Appendix § 8.§ 2) gives the step

Δu=−Hm−1gm,\Delta u=-H_m^{-1}\mathbf{g}_m,
Hm=Huu−HuθHθθ−1Hθu,gm=gu−HuθHθθ−1gθ.(29)H_m=H_{uu}-H_{u\theta}H_{\theta\theta}^{-1}H_{\theta u}, \qquad\mathbf{g}_m=\mathbf{g}_u-H_{u\theta}H_{\theta\theta}^{-1}\mathbf{g}_\theta. \tag*{(29)}

Here, HmH_m is the curvature of the loss in the physical parameters after the network has adjusted to them. The step therefore already accounts for the change of the network needed to keep satisfying the geodesic equation, and it moves the hair, the mass and the orbital elements together along the valley. This is the coordinated move that Adam does not make.

After the step, the old weights no longer solve the geodesic equation for the new orbit. Instead of using the linear estimate of their change from the quadratic model, which is accurate only for small steps, we hold the hair at its new value and re-train the network and the other physical parameters with Adam in a short run, after which the next step is computed3. We stop when the predicted decrease gmTHm−1gm\mathbf{g}_m^{\mathsf T}H_m^{-1}\mathbf{g}_m falls below 0.020.02. That is, when the minimum is located to about 0.14σ0.14\sigma. Both stages minimize the same loss of the same model and differ only in how the parameters are updated. Each optimization is repeated from different starting values of the hair and, for Schwarzschild, from three initializations of the network. We accept an estimate only if the final values agree to better than 0.2σ0.2\sigma (Appendix § 11).

Uncertainties

In the Laplace approximation [58], the posterior is a Gaussian centered on the minimum of the loss, u∗u^\ast. The network weights are not physical parameters. For a given uu, they are set by minimizing the loss, that is, mainly by the geodesic equation, and when uu changes, they follow through their linear response A≡∂θ/∂uA\equiv\partial\theta/\partial u (Appendix § 8.§ 2). Along this response, the data and frame-prior residuals rd\mathbf{r}_d change by J^Δu\hat{J}\Delta u, with J^=drd/du=Ju+JθA\hat{J}=\mathrm{d}\mathbf{r}_d/\mathrm{d}u=J_u+J_\theta A, where JuJ_u and JθJ_\theta are the rows of JJ for rd\mathbf{r}_d. The posterior of the physical parameters is therefore

p(u)∝exp⁡[−12(χdata2+χframe2)]≃exp⁡(−12ΔuTJ^TJ^Δu).(30)p(u)\propto\exp\left[-\frac{1}{2}\left(\chi^2_{\mathrm{data}}+\chi^2_{\mathrm{frame}}\right)\right]\simeq\exp\left(-\frac{1}{2}\Delta u^{\mathsf T}\hat{J}^{\mathsf T}\hat{J}\Delta u\right). \tag*{(30)}

with covariance matrix Σu=(J^TJ^)−1\Sigma_u=(\hat{J}^{\mathsf T}\hat{J})^{-1} and Δu=u−u∗\Delta u=u-u^\ast.

Because the network follows the parameters, any trade-off between the network and the physical parameters is included in the error bars. Moreover, only the data and the frame priors measure uu: the physics residuals and the weight prior determine how the network follows, but they do not enter Σu\Sigma_u. Freezing the network, or integrating over the weights as if they were parameters, would violate one of these properties and underestimate the uncertainties (Appendix § 8.§ 2). With Eq. (30), GINN reproduces the uncertainties of the exact solver, which has no network (Appendix § 11).

We draw samples from N(u∗,Σu)\mathcal{N}(u^\ast,\Sigma_u) and map them to the physical parameters through Eq. (A.1), which preserves nonlinear correlations such as the MM-DD degeneracy. The same Gaussian gives the Bayesian evidence of each spacetime (Appendix § 9). Departures from it are examined in Appendix § 10.

A hair parameter that enters the metric only through ξ2\xi^2 (qRNq_{\mathrm{RN}}, qBq_{\mathrm{B}} and qJNWq_{\mathrm{JNW}}) needs a slightly different treatment, because at ξ=0\xi=0 the residuals do not change to first order and the GN curvature in ξ\xi vanishes. To leading order χdata2\chi^2_{\mathrm{data}} is instead linear in ξ2\xi^2, χ2≃χ02+sξξ2\chi^2\simeq\chi_0^2+s_\xi\xi^2, so the posterior is even in ξ\xi and approximately Gaussian with width σξ=sξ−1/2\sigma_\xi=s_\xi^{-1/2}. We compute the slope sξ=∂χdata2/∂(ξ2)s_\xi=\partial\chi^2_{\mathrm{data}}/\partial(\xi^2) from the data residuals, with the network following the equation of motion (Appendix § 8.§ 3).

For the Yukawa correction, δ\delta and λ\lambda are degenerate (§ 2). We therefore fix λ\lambda, obtain the posterior of δ\delta with all other parameters marginalized, and repeat this on a grid of λ\lambda to map the constraint on δ\delta as a function of the range.

Results

Inferred parameters

Table 2 lists the estimates and their 1σ1\sigma uncertainties, and Figure 4 shows the best-fit orbits. The posteriors are shown in Figure 5 for Schwarzschild and in Appendix § 12 for the other spacetimes.

Figure 4. Best-fit orbits of S2 in all spacetimes. Left: positions on the sky relative to Sgr A* (denoted by a star), with the astrometric data. Right: RA, Dec, and line-of-sight velocity as functions of time, each with the residuals of every spacetime below it. The orbits are almost indistinguishable at the precision of the data.

Posterior of all 13 parameters for Schwarzschild

Figure 5. Posterior of all 13 parameters for Schwarzschild, from the Laplace approximation, with the network following the parameters. The contours enclose 39% and 86% of the two-dimensional probability (1σ and 2σ). In the titles we show the median and the 16th and 84th percentiles.

For Schwarzschild, χdata2=302.9\chi^2_{\mathrm{data}} = 302.9 for N=334N = 334 data points and χframe2=4.4\chi^2_{\mathrm{frame}} = 4.4, so that χmin⁡2=307.3\chi^2_{\min} = 307.3 for 338 residuals and 13 parameters (χ2/ν=0.95\chi^2/\nu= 0.95), with M=(4.32±0.22)×106 M⊙M = (4.32 \pm0.22) \times10^6\,M_\odot and D=8.33±0.21D = 8.33 \pm0.21 kpc. These values are consistent with Gillessen et al. [43] and with more recent determinations (e.g., Abuter et al. [2, 3]; ); GRAVITY Collaboration [48]). In all other spacetimes, the orbital elements agree with the Schwarzschild ones to within 1σ1\sigma. The largest differences occur for Horndeski and Yukawa, which are the two spacetimes where the hair changes the Newtonian potential, with their masses larger by 0.2σ0.2\sigma.

None of the hair parameters differs significantly from its GR value. The Horndeski hair, qH=(1.9±3.5)×10−3q_H = (1.9 \pm3.5) \times10^{-3}, is consistent with zero. The Reissner-Nordström charge, qRN=0.3±2.5q_{\mathrm{RN}} = 0.3 \pm2.5, does not separate the black-hole range ∣qRN∣≤1|q_{\mathrm{RN}}| \le1 from a naked singularity, and is best read as a bound on a positive 1/r21/r^2 term in the metric function f(r)f(r). For Bardeen and JNW, whose corrections at the radii of S2 are suppressed by ∼10−7ξ2\sim10^{-7}\xi^2, the data exclude only ∣ξ∣≳102|\xi| \gtrsim10^2 with qB=3±70q_B = 3 \pm70 and qJNW=1±128q_{\mathrm{JNW}} = 1 \pm128. These values lie far outside the horizon and photon-sphere ranges of both spacetimes, so S2 and horizon-scale imaging probe different parts of their parameter space. Our conclusions agree with the MCMC analyses of the same spacetimes by Navarrete et al. [60] and Bambhaniya et al. [6], who also find no preference over Schwarzschild.

For the three hair parameters that enter only through ξ2\xi^2, the posterior is even in ξ\xi. The best-fit values differ from zero only within the stopping tolerance of the optimization (0.14σ0.14\sigma). We therefore summarize these posteriors by upper limits on ∣ξ∣|\xi|. Since the posterior of these hairs is not Gaussian, we take the limits from the higher-order posterior (§ 6): ∣qRN∣<3.9|q_{\mathrm{RN}}| < 3.9, ∣qB∣<92|q_B| < 92 and ∣qJNW∣<175|q_{\mathrm{JNW}}| < 175 (95%), in agreement with the exact solver to within 2%. The Gaussian of Table 2 gives the weaker limits 5.0, 138 and 251.

At fixed λ=103\lambda= 10^3 AU, comparable to the semi-major axis of S2 (a≃1030a \simeq1030 AU), we find δ=0.005±0.012\delta= 0.005 \pm0.012, consistent with zero at 0.4σ0.4\sigma. To map the constraint as a function of the range, we use 15 values of λ\lambda between 10 and 3×1043 \times10^4 AU, logarithmically spaced. At each value, all other parameters and the network are re-optimized with GN steps, starting from the solution at the neighbouring λ\lambda. The posterior of δ\delta, a Gaussian N(δ∗,σδ2)\mathcal{N}(\delta^*,\sigma_\delta^2) restricted to the prior, gives the 95% upper limit δ95\delta_{95} on ∣δ∣|\delta| through P(∣δ∣<δ95)=0.95P(|\delta| < \delta_{95}) = 0.95.

Figure 3. Yukawa correction as a function of its range λ\lambda. Blue (left axis): 95% upper limit on ∣δ∣|\delta|; the dashed lines mark the tightest limit. Purple (right axis): estimate of δ∗\delta^* and its 1σ1\sigma uncertainty. The grey band marks the radial range of S2, from pericenter to apocenter. Red and green points are the tightest limits of Hees et al. [51] and Abd El Dayem et al. [1].

Figure 3 shows the result. The limit is tightest at λ≃180\lambda\simeq180 AU (2.7×10132.7 \times10^{13} m), with ∣δ∣<0.0105|\delta| < 0.0105 and δ∗=0.0027±0.0047\delta^* = 0.0027 \pm0.0047. For λ\lambda much smaller than the pericenter distance, the Yukawa term has decayed along the whole orbit and only rescales the mass. For λ\lambda much larger than the orbit, it reduces to a constant and a uniform force. The limit therefore weakens on both sides, to ∣δ∣<0.11|\delta| < 0.11 at λ≃32\lambda\simeq32 AU and ∣δ∣<0.14|\delta| < 0.14 at λ≃3×103\lambda\simeq3 \times10^3 AU, and reaches the prior boundary for λ≲18\lambda\lesssim18 AU and λ≳104\lambda\gtrsim10^4 AU. The estimate δ∗\delta^* is consistent with zero at all λ\lambda. Our tightest limit is comparable to that of Hees et al. [51], ∣α∣<0.016|\alpha| < 0.016 at λ=150\lambda= 150 AU4^{4}, and is about 3.5 times weaker than that of Abd El Dayem et al. [1], ∣α∣<0.003|\alpha| < 0.003 at λ≃200\lambda\simeq200 AU, which includes GRAVITY astrometry up to 2022. Adding four more stars and data up to 2025, Foschi et al. [38] recently tightened this to ∣α∣<6×10−4|\alpha| < 6 \times10^{-4} at the same range.

Model comparison

The Laplace approximation also gives the integral of the posterior, so the evidence of each spacetime follows from quantities already computed at the minimum. Relative to Schwarzschild, the logarithm of the Bayes factor is (Appendix § 9)

ln⁡B=−Δχmin⁡22+ln⁡2π σξmξΔξ+12Δln⁡det⁡Σrest.(31)\ln B = -\frac{\Delta\chi^2_{\min}}{2} + \ln\frac{\sqrt{2\pi}\,\sigma_\xi m_\xi}{\Delta_\xi} + \frac{1}{2}\Delta\ln\det\Sigma_{\mathrm{rest}}. \tag*{(31)}

Here Δχmin⁡2\Delta\chi^2_{\min} is the change of the minimum, Δξ\Delta_\xi the width of the hair prior, mξm_\xi the part of the Gaussian N(ξ∗,σξ2)\mathcal{N}(\xi^*,\sigma_\xi^2) inside it, and Σrest\Sigma_{\mathrm{rest}} the covariance of the other parameters at fixed hair. The second term is the Occam factor, the fraction of the prior range of the hair allowed by the data. It penalizes a hair that the data do not require, and it depends on the chosen prior. The last term is below 0.02 for all spacetimes.

Table 3 lists the values. All the Bayes factors are negative, so none of the spacetimes is preferred over Schwarzschild. On the scale of Trotta [78] (∣ln⁡B∣<1|\ln B| < 1 inconclusive, 1-2.5 weak, 2.5-5 moderate), the evidence for Schwarzschild ranges from inconclusive to moderate. In every case, the Bayes factor is set by the Occam penalty of the hair rather than by the fit. Horndeski and Yukawa lower χmin⁡2\chi^2_{\min} by 0.3 and 0.2, less than needed to pay for the prior range of their hair, and the other three spacetimes do not improve the fit at all. For Reissner-Nordström, ln⁡B\ln B is therefore entirely the penalty of the prior range ∣qRN∣≤50|q_{\mathrm{RN}}| \le50. Restricted to the black-hole range ∣qRN∣≤1|q_{\mathrm{RN}}| \le1, narrower than the posterior, the prior would give ln⁡B≃0\ln B \simeq0. For hair that enters the metric through its square, the posterior is narrower than its Gaussian, and using the higher-order posterior changes ln⁡B\ln B by less than 0.1, which does not change these conclusions (Appendix § 10).

Discussion

The results rest on three considerations: the network replaces the exact solution of the geodesic equation, the optimizer finds the minimum of the loss, and the posterior is Gaussian. We test each of them separately (Appendices § 10 and § 11).

SpacetimeΔχmin⁡2\Delta\chi^{2}_{\min}ln⁡O\ln Oln⁡B\ln B
Hairy Horndeski−0.31-0.31−2.45-2.45−2.29-2.29
Reissner-Nordström0.010.01−2.76-2.76−2.77-2.77
Bardeen0.000.00−0.82-0.82−0.82-0.82
JNW0.000.00−1.32-1.32−1.32-1.32
Yukawa (λ=103\lambda= 10^{3} AU)−0.18-0.18−3.51-3.51−3.41-3.41

Table 3. Laplace estimate of the Bayes factor relative to Schwarzschild, Eq. (31), with the Occam factor ln⁡O=ln⁡[2π σξmξ/Δξ]\ln O = \ln[\sqrt{2\pi}\,\sigma_{\xi}m_{\xi}/\Delta_{\xi}]. ln⁡B>0\ln B > 0 would favour the spacetime over Schwarzschild.

  • Accuracy of the network: The decisive test is a comparison with an exact geodesic solver, which inverts t(ϕ)t(\phi) numerically and contains neither the network nor the physics term. For Schwarzschild, the estimates of the two models differ by less than 0.01σ0.01\sigma and their uncertainties by less than 1%. The comparison also holds away from the minimum; over the whole posterior of every spacetime, the exact solver and GINN assign nearly the same probability to the same parameters. The network is therefore as accurate as the exact solver wherever the posterior has weight. Letting the network follow the parameters nevertheless matters for the uncertainties: with the weights held fixed, they can be up to 20% smaller, because the response of the network to the parameters partly compensates the change in the Keplerian timing.

  • Convergence: Repeating each optimization from different starting values of the hair, and for Schwarzschild from different initializations of the network, gives final values of the hair, MM and DD that agree to within 0.15σ0.15\sigma in all spacetimes. The Adam stage alone does not achieve this, and the GN stage is what makes the result independent of the starting point.

  • Beyond the Gaussian approximation: A Gaussian built from the local curvature is fast, but it can misrepresent both the size and the orientation of the posterior when the observables depend nonlinearly on the parameters. We test it in two steps (Appendix § 10). First, we expand the residuals to second order in the parameters, as in the DALI method of Sellentin et al. [74], with the network eliminated to the same order. The resulting χ2\chi^2 remains a sum of squares and can describe

skewed and curved posteriors. Second, we reweigh samples of this expansion with the exact solver, which gives the exact posterior.

The six spacetimes fall into two groups. For Schwarzschild, Horndeski and Yukawa, the Gaussian describes the posterior well. It reproduces the means of all parameters to within about 0.1σ0.1\sigma of the exact posterior and the widths of qHq_{\mathrm{H}} and δ\delta to within about 1%, and it overestimates the widths of the orbital elements by 6-12%. The reason is the second-derivative term ∑krk∇2rk\sum_k r_k\nabla^2r_k dropped in Eq. (26); for these data it does not cancel along the direction in which all orbital elements are correlated, and steepens χ2\chi^2 along it by 25% (Appendix § 10). For Reissner-Nordström, Bardeen and JNW, whose hair enters only through ξ2\xi^2, the posterior of the hair is not Gaussian. Its χ2\chi^2 also contains a term in ξ4\xi^4, and the other parameters compensate it along a curved degeneracy, so that χ2\chi^2 rises faster than the parabola at large ∣ξ∣|\xi| (Figure 6). The exact posterior can then be up to 30% narrower than the Gaussian, and for these three spacetimes we take the limits on ∣ξ∣|\xi| from the higher-order posterior (§ 5). In neither group does the Gaussian underestimate the uncertainties by more than a per cent, and in both the higher-order posterior follows from derivatives of the same model.

Figure 6. Profile Δχ2\Delta\chi^2 for each hair parameter (minimized over all others), comparing the Gaussian (Laplace) approximation (dashed), the second-order expansion (solid), and the exact solver (points). For linear hairs (qHq_{\rm H}, δ\delta) all three agree closely; for squared hairs (qRNq_{\rm RN}, qBq_{\rm B}, qJNWq_{\rm JNW}), χ2\chi^2 rises faster than a parabola, yielding a narrower posterior than the Gaussian. For these, the parameter that the data constrain is ξ2\xi^2 (Appendix § 8.§ 3).

Conclusions

We have presented the Geodesic-Informed Neural Network (GINN), a forward model for stellar orbits that is exact in the metric and differentiable in all parameters. A small network, constrained by the geodesic equation, represents only the radial phase ϕ(t)\phi(t). A single optimization that includes Gauss-Newton steps gives the estimates, and the curvature at the minimum, with the network following the parameters, gives their uncertainties and the Bayesian evidence.

Applied to the 1992-2016 data of S2, GINN reproduces an exact geodesic solver and finds no significant deviation from Schwarzschild in any of the six spacetimes considered. The tightest constraints are on hair that modifies the Newtonian potential, and the Bayes factors favour Schwarzschild mainly through the Occam penalty of the additional parameter. For Schwarzschild, Horndeski and Yukawa, whose hair enters the metric linearly, the Gaussian describes the posterior well. For Reissner-Nordström, Bardeen and JNW, whose hair enters only through its square, the Gaussian is conservative, and the limits on the hair follow from a second-order expansion of the same model, checked against the exact solver.

The method has limitations. The Laplace approximation and its higher-order extension describe the posterior around one maximum and would not reveal a separate mode; the multi-start test makes such a mode unlikely but does not exclude it. The forward model assumes a static, spherically symmetric metric and test-particle motion, so a rotating black hole would require a different parametrization of the orbit. Our data do not include GRAVITY astrometry, and we do not model astrometric confusion [66].[66]

Because a spacetime enters only through ff, gg and hh, GINN applies to any static, spherically symmetric metric without modification. It can be extended directly to several stars sharing one metric, each with its own network, and to GRAVITY data, for which the non-Keplerian correction becomes measurable. Finally, the same construction can represent the metric functions themselves by networks, turning the test of given spacetimes into a reconstruction of the spacetime probed by the orbits, which will be discussed in a future work.

Acknowledgements. PHP is funded by Fundação de Amparo à Pesquisa e Inovação do Espírito Santo (FAPES, Brazil) and Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES, Brazil). We acknowledge the use of computational resources from the Sci-Com Lab of the Department of Physics at UFES, supported by FAPES, CAPES, and CNPq. DF, AB and DFM thank the Research Council of Norway for their support and the resources provided by UNINETT Sigma2 – the National Infrastructure for High Performance Computing and Data Storage in Norway.

Optimization details

Optimization steps

  1. Initialization: We initialize the parameters with a Keplerian fit to the astrometry alone. For a fixed period, eccentricity and time of pericenter (PP, ee, tpt_p), the projected orbit is linear in the Thiele-Innes constants, which follow from weighted least squares. (PP, ee, tpt_p) are found by a grid search refined with the Nelder-Mead method (Nelder & Mead [61]), and the constants are converted to (aa, ii, Ω\Omega, ω\omega). The mass follows from Kepler’s third law at a nominal distance. Astrometry alone cannot separate MM from DD, so this value is only a starting guess, and DD is determined together with MM afterwards. The network output is scaled by 10−310^{-3} rad, so the initial correction to the Keplerian orbit is small. The hair starts slightly away from its GR value; for a hair entering only through ξ2\xi^2, the gradient would otherwise vanish.

  1. Whitened coordinates: The physical parameters differ in scale by orders of magnitude and are correlated. Adam works in whitened coordinates u\mathbf{u},

Θ=Θ0+SWu,WWT=H0−1,(32)\boldsymbol{\Theta} = \boldsymbol{\Theta}_0 + S W\mathbf{u}, \qquad WW^{\mathsf{T}} = H_0^{-1}, \tag*{(32)}

where Θ0\boldsymbol{\Theta}_0 is the current point and SS a diagonal matrix of rough parameter scales. H0H_0 is the GN matrix of χdata2+χframe2\chi^2_{\mathrm{data}}+\chi^2_{\mathrm{frame}} in the scaled variables S−1(Θ−Θ0)S^{-1}(\boldsymbol{\Theta}-\boldsymbol{\Theta}_0), computed with the network switched off and with a small ridge to keep it invertible. WW is the lower-triangular Cholesky factor of H0−1H_0^{-1}. Since WTH0W=IW^{\mathsf{T}}H_0W=I, the components of u\mathbf{u} are uncorrelated with unit variance, and a step of given size corresponds to about the same number of standard deviations in every direction. The whitening is recomputed at the start of every Adam run, including the re-training between GN steps. It is only a change of variables: the loss, its minimum, the GN step and the covariance mapped back to Θ\boldsymbol{\Theta} do not depend on it.

  1. Adam optimization: In the first stage Adam runs for NA=3000N_A=3000 steps with learning rates η0=0.1\eta_0=0.1 for u\mathbf{u} and 5×10−45\times10^{-4} for the weights. A larger rate for the weights makes Adam oscillate, because the physics term is stiff. Both rates follow a cosine schedule (Loshchilov & Hutter [57]),

ηk=ηmin⁡+η0−ηmin⁡2[1+cos⁡(πkK)],ηmin⁡=0.01 η0.(33)\eta_k=\eta_{\min}+\frac{\eta_0-\eta_{\min}}{2}\left[1+\cos\left(\frac{\pi k}{K}\right)\right],\qquad\eta_{\min}=0.01\,\eta_0. \tag*{(33)}

Because the collocation points are shifted randomly at each step, the physics term fluctuates, and with a constant rate, Adam keeps moving around the minimum by an amount proportional to the rate. The decreasing rate lets it settle. We keep the state with the lowest loss on a fixed collocation grid, so a rate that decays before convergence cannot leave a worse state behind. Final convergence in the physical parameters is in any case reached in the GN stage. Each re-training between GN steps is a short Adam run of 400 steps with the same schedule, with the hair held fixed and a learning rate of 2×10−32\times10^{-3} for the weights, since the network then only has to adapt to a small change of the orbit. The GN iteration also stops after 30 steps or after more than five halvings of the step, but in all our fits the criterion on the predicted decrease was met first.

Reduced step and Laplace posterior

In block form, Eq. (28) reads

L(Δ)≃L(0)+2(guTΔu+gθTΔθ)+ΔuTHuuΔu+2ΔuTHuθΔθ+ΔθTHθθΔθ,(34)\mathcal{L}(\Delta)\simeq\mathcal{L}(0)+2(\mathbf{g}_{u}^{\mathsf{T}}\Delta\mathbf{u}+\mathbf{g}_{\theta}^{\mathsf{T}}\Delta\boldsymbol{\theta})+\Delta\mathbf{u}^{\mathsf{T}}H_{uu}\Delta\mathbf{u}+2\Delta\mathbf{u}^{\mathsf{T}}H_{u\theta}\Delta\boldsymbol{\theta}+\Delta\boldsymbol{\theta}^{\mathsf{T}}H_{\theta\theta}\Delta\boldsymbol{\theta}, \tag*{(34)}

where the weight prior DwD_w makes HθθH_{\theta\theta} invertible. Completing the square in Δθ\Delta\boldsymbol{\theta} gives

L(Δ)≃L(0)−gθTHθθ−1gθ+(Δθ−Δθ∗)THθθ(Δθ−Δθ∗)+2gmTΔu+ΔuTHmΔu,(35)\mathcal{L}(\Delta)\simeq\mathcal{L}(0)-\mathbf{g}_{\theta}^{\mathsf{T}}H_{\theta\theta}^{-1}\mathbf{g}_{\theta}+(\Delta\boldsymbol{\theta}-\Delta\boldsymbol{\theta}^{*})^{\mathsf{T}}H_{\theta\theta}(\Delta\boldsymbol{\theta}-\Delta\boldsymbol{\theta}^{*})+2\mathbf{g}_{m}^{\mathsf{T}}\Delta\mathbf{u}+\Delta\mathbf{u}^{\mathsf{T}}H_m\Delta\mathbf{u}, \tag*{(35)}

with Δθ∗=−Hθθ−1(gθ+HθuΔu)\Delta\boldsymbol{\theta}^{*}=-H_{\theta\theta}^{-1}(\mathbf{g}_{\theta}+H_{\theta u}\Delta\mathbf{u}), and with HmH_m and gm\mathbf{g}_m as in Eq. (29). This one identity gives both the step and the response of the network that enters the covariance.

  1. Step: For any Δu\Delta\mathbf{u}, the loss is lowest when the network takes its linear response Δθ=Δθ∗\Delta\boldsymbol{\theta}=\Delta\boldsymbol{\theta}^{*}. What remains is a quadratic in Δu\Delta\mathbf{u} with curvature HmH_m and gradient gm\mathbf{g}_m, which is minimized by the step of Eq. (29). The total predicted decrease is gθTHθθ−1gθ+gmTHm−1gm\mathbf{g}_{\theta}^{\mathsf{T}}H_{\theta\theta}^{-1}\mathbf{g}_{\theta}+\mathbf{g}_m^{\mathsf{T}}H_m^{-1}\mathbf{g}_m. The first term is the gain from adjusting the network alone. It is nearly zero after re-training, since then gθ≃0\mathbf{g}_{\theta}\simeq0, so only the second term is used to stop.

  1. Laplace posterior: At the minimum gu=gθ=0\mathbf{g}_u=\mathbf{g}_{\theta}=0, so the response of the network to a displacement Δu\Delta\mathbf{u} is Δθ∗=AΔu\Delta\boldsymbol{\theta}^{*}=A\Delta\mathbf{u} with A=−Hθθ−1HθuA=-H_{\theta\theta}^{-1}H_{\theta u}. We take this response as the solution of the equation of motion for the new parameters. Along it, the data and frame-prior residuals change by J^Δu\hat{J}\Delta\mathbf{u}, with J^=Ju+JθA\hat{J}=J_u+J_{\theta}A restricted to these rows, and the posterior of the physical parameters is p(Δu)∝exp⁡(−12∣J^Δu∣2)p(\Delta\mathbf{u})\propto\exp(-\frac{1}{2}\lvert\hat{J}\Delta\mathbf{u}\rvert^2), a Gaussian with covariance Σu=(J^TJ^)−1\Sigma_u=(\hat{J}^{\mathsf{T}}\hat{J})^{-1}, Eq. (30). The two other ways of treating the network both underestimate the uncertainties:

    • Network fixed, Δθ=0\Delta\boldsymbol{\theta}=0, gives Σu=(JuTJu)−1\Sigma_u=(J_u^{\mathsf{T}}J_u)^{-1}. This ignores that a change of the orbit changes the solution of the geodesic equation, so the trade-off between the network and the physical parameters is missing. For example, the uncertainty of MM for Schwarzschild is then about 20% too small (0.175 instead of 0.22).

    • Network integrated over: integrating the Gaussian of Eq. (A.4) over Δθ\Delta\boldsymbol{\theta} gives Σu=Hm−1\Sigma_u=H_m^{-1}, the u\mathbf{u} block of H−1H^{-1}. This treats the physics residuals and the weight prior as measurements of u\mathbf{u}, although they only fix how the network represents the solution. For example, the uncertainty of the Yukawa parameter δ\delta is then about 15% too small.

The GN step uses HmH_m, because it minimizes the full loss, including the physics term; the covariance uses J^TJ^\hat{J}^{\mathsf{T}}\hat{J}, because only the data and the frame priors are measurements.

Hair parameters entering as ξ2\xi^2

Let the residuals depend on the hair only through ξ2\xi^2. By the chain rule, ∂rk/∂ξ=2ξ ∂rk/∂(ξ2)\partial r_k/\partial\xi=2\xi\,\partial r_k/\partial(\xi^2), which vanishes at ξ=0\xi=0, and so does the GN curvature 2∑k(∂rk/∂ξ)22\sum_k(\partial r_k/\partial\xi)^2. The curvature of χdata2=∑krk2\chi^2_{\mathrm{data}}=\sum_k r_k^2 comes entirely from the term dropped in Eq. (26). With ∂2rk/∂ξ2=2 ∂rk/∂(ξ2)\partial^2r_k/\partial\xi^2=2\,\partial r_k/\partial(\xi^2) at ξ=0\xi=0,

∂2χ2∂ξ2∣ξ=0=2∑krk∂2rk∂ξ2=2∂χ2∂(ξ2)=2sξ.(36)\left.\frac{\partial^2\chi^2}{\partial\xi^2}\right|_{\xi=0}=2\sum_k r_k\frac{\partial^2r_k}{\partial\xi^2}=2\frac{\partial\chi^2}{\partial(\xi^2)}=2s_{\xi}. \tag*{(36)}

To leading order in the hair, the residuals are linear in ξ2\xi^2, so χ2≃χ02+sξξ2\chi^2\simeq\chi_0^2+s_{\xi}\xi^2. Since the posterior is ∝e−χ2/2\propto e^{-\chi^2/2}, it is approximately a Gaussian of width σξ=sξ−1/2\sigma_{\xi}=s_{\xi}^{-1/2}, provided sξ>0s_{\xi}>0. It is even in ξ\xi, so at ξ=0\xi= 0 the hair is uncorrelated with the other parameters, which keep the covariance Σu\Sigma_u of Eq. (30).

The slope cannot be read off a Jacobian at ξ=0\xi= 0, where the Jacobian vanishes. We therefore evaluate it at a finite value ξe\xi_e, where χ2\chi^2 is still linear in ξ2\xi^2. The hair is set to ξe\xi_e and the network is re-trained. The data residuals are then linearized in all free parameters, with the network following through the linear response that keeps the physics residuals fixed. Minimizing the resulting quadratic model over the other parameters gives the slope ∂χ2/∂ξ\partial\chi^2/\partial\xi at ξe\xi_e, and sξs_\xi is this slope divided by 2ξe2\xi_e.

This is the same rule as for a hair that enters linearly. In both cases σξ\sigma_\xi is the shift of the hair from its best-fit value at which the χ2\chi^2, minimized over all other parameters, has risen by one. For a linear hair, this χ2\chi^2 is quadratic in ξ\xi, which gives the Laplace width; for a squared hair it is linear in ξ2\xi^2, which gives sξ−1/2s_\xi^{-1/2}. The next-order terms make the squared-hair posterior somewhat narrower than this Gaussian (Appendix § 10). The parameter that the data constrain is therefore η=ξ2\eta= \xi^2, in which the residuals are linear to first order. Fitting η\eta instead of ξ\xi would make the Gaussian more accurate, but a prior uniform in ξ\xi is not uniform in η\eta, and the limits in the literature are quoted for ∣ξ∣|\xi|, so we keep ξ\xi.

Bayesian evidence

The evidence of a model with parameters Θ\Theta and prior π(Θ)\pi(\Theta) is Z=∫P(y∣Θ) π(Θ) dΘZ = \int\mathcal{P}(y\mid\Theta)\,\pi(\Theta)\,\mathrm{d}\Theta. The Laplace approximation replaces the integrand by a Gaussian around the maximum, which gives

Z≃P(y∣Θ∗)π(Θ∗)(2π)n/2det⁡Σ1/2,(37)Z \simeq\mathcal{P}(y\mid\Theta^*)\pi(\Theta^*)(2\pi)^{n/2}\det\Sigma^{1/2}, \tag*{(37)}

where P∝e−χ2/2\mathcal{P} \propto e^{-\chi^2/2}, Σ\Sigma is the Laplace covariance of Sec. 4.6 expressed in Θ\Theta, and nn is the number of parameters. All spacetimes share the same 13 orbital and frame parameters. The determinant factorizes as det⁡Σ=σξ2det⁡Σrest\det\Sigma= \sigma_\xi^2\det\Sigma_{\mathrm{rest}}, with σξ2\sigma_\xi^2 the marginal variance of the hair and Σrest\Sigma_{\mathrm{rest}} the covariance of the other parameters at fixed hair. A flat hair prior of width Δξ\Delta_\xi contributes 1/Δξ1/\Delta_\xi, and restricting the Gaussian to the prior contributes the fraction mξm_\xi inside it. Together these give Eq. (31).

We make three assumptions here: (i) only χdata2\chi^2_{\mathrm{data}} and χframe2\chi^2_{\mathrm{frame}} enter Δχmin⁡2\Delta\chi^2_{\min} and Σ\Sigma; the normalization of the frame priors is the same in all spacetimes and cancels. The physics residuals are left out because at the minimum they vanish to within the tolerance, and since the network follows the parameters rather than being integrated over (§ 4.6), it contributes no factor of its own to the evidence. (ii) the Occam term depends on the prior through −ln⁡Δξ-\ln\Delta_\xi, so the Bayes factors are meaningful only together with Table 1. (iii) for squared hair the widths of Appendix A.3 are used, and the remaining non-Gaussianity changes ln⁡B\ln B by less than 0.1, without changing the conclusions. We compute this change by importance sampling the higher-order posterior of Appendix C, with a proposal built from its samples.

Beyond the quadratic approximation

In coordinates zz in which the Laplace posterior is N(0,I)\mathcal{N}(0,I), u=u∗+Czu = u^* + Cz with CCT=ΣuCC^{\mathsf T} = \Sigma_u, we expand the data residuals to second order,

r(z)≃r0+Jz+12K(z,z),(38)\mathbf{r}(z) \simeq\mathbf{r}_0 + Jz + \frac{1}{2}K(z,z), \tag*{(38)}

with Jk,i=∂rk/∂ziJ_{k,i} = \partial r_k/\partial z_i and Kk,i,j=∂2rk/∂zi∂zjK_{k,i,j} = \partial^2 r_k/\partial z_i\partial z_j. Squaring gives

χ2(z)=∣r0∣2+2r0TJz+zT(JTJ+r0⋅K)z+(Jz)TK(z,z)+14∣K(z,z)∣2,(39)\chi^2(z) = |\mathbf{r}_0|^2 + 2\mathbf{r}_0^{\mathsf T}Jz + z^{\mathsf T}(J^{\mathsf T}J+\mathbf{r}_0\cdot K)z + (Jz)^{\mathsf T}K(z,z) + \frac{1}{4}|K(z,z)|^2, \tag*{(39)}

with r0⋅K=∑kr0,kKk,i,j\mathbf{r}_0\cdot K = \sum_k r_{0,k}K_{k,i,j}. Keeping only JJ gives back the Laplace Gaussian. The term r0⋅K\mathbf{r}_0\cdot K is the curvature dropped in Eq. (26), the cubic term skews the posterior, and the quartic term bends its degeneracies. The gradient r0TJ\mathbf{r}_0^{\mathsf T}J vanishes at the minimum, up to the stopping tolerance, so the expansion changes the shape of the posterior but not its maximum. For r0=0\mathbf{r}_0 = 0 and without the network, Eq. (C.2) is the DALI doublet of Sellentin et al. [74] (also see de Souza & Quartin 2026); the terms in r0\mathbf{r}_0 arise when we expand at the best fit to real data rather than at a fiducial model. As in DALI, the truncation is in the order of the derivatives of the residuals, so that χ2\chi^2 remains a sum of squares and e−χ2/2e^{-\chi^2/2} a proper distribution.

When the physical parameters change, the network must keep solving the geodesic equation. We eliminate its weights to the same order, θ(z)=θ∗+Az+12B(z,z)\theta(z) = \theta^* + Az + \frac{1}{2}B(z,z). Let J~θ\tilde{J}_\theta be the Jacobian of all residuals (data, physics and weight prior) with respect to the weights, and P=J~θ(J~θTJ~θ)−1J~θTP = \tilde{J}_\theta(\tilde{J}_\theta^{\mathsf T}\tilde{J}_\theta)^{-1}\tilde{J}_\theta^{\mathsf T} the projector onto the changes that the network can produce. The first-order response is that of Appendix A.2, A=−Hθθ−1HθzA = -H_{\theta\theta}^{-1}H_{\theta z}, and turns the Jacobian into (I−P)Jz(I-P)J_z, whose data and frame-prior rows give JTJ=IJ^{\mathsf T}J = I in these coordinates. At second order, minimizing ∣K(z,z)∣2|K(z,z)|^2 over BB gives Bij=−(J~θTJ~θ)−1J~θTKijlinB_{i j} = -(\tilde{J}_\theta^{\mathsf T}\tilde{J}_\theta)^{-1}\tilde{J}_{\theta}^{\mathsf T}K^{\mathrm{lin}}_{i j}, where KlinK^{\mathrm{lin}} is the second derivative along θ∗+Az\theta^* + Az, so that K=(I−P)KlinK = (I-P)K^{\mathrm{lin}}. Only the data and frame-prior rows enter Eq. (C.1), since the physics residuals are the constraint that the network follows rather than measurements, as for the covariance, Eq. (30). JJ and KlinK^{\mathrm{lin}} are computed by automatic differentiation.

Since the expansion leaves the maximum in place, its effect is on integrals over the posterior, which we obtain by sampling e−χ2(z)/2e^{-\chi^2(z)/2}, restricted to the hair prior, with emcee [36]. The polynomial χ2\chi^2 is cheap to evaluate, and we use 4nn walkers for nn parameters and run each chain until it is longer than 50 autocorrelation times and gives at least 10410^4 independent samples, discarding the first five autocorrelation times. The chains are longer than 180 autocorrelation times, so that their Monte Carlo errors are well below the effects we report. To obtain the exact posterior, we reweight 6000 of these samples by w∝exp⁡[−12(Δχexact2−Δχ2)]w \propto\exp[-\frac{1}{2}(\Delta\chi^2_{\mathrm{exact}}-\Delta\chi^2)], where Δ\Delta denotes the change from the minimum and χexact2\chi^2_{\mathrm{exact}} is computed with the exact solver. The effective fraction of samples, (∑iwi)2/(ns∑iwi2)(\sum_i w_i)^2/(n_s\sum_i w_i^2) for nsn_s samples, measures the agreement: it equals 1 when the two posteriors coincide.

The exact solver keeps 86–100% of the samples effective in all spacetimes, so the expansion, and with it the network, describes the posterior well. The term r0⋅K\mathbf{r}_0\cdot K explains the narrowing of the orbital elements. For Schwarzschild, the eigenvalues of JTJ+r0⋅KJ^{\mathsf T}J+\mathbf{r}_0\cdot K lie within 2% of one, except one of 1.25. Its eigenvector is correlated by 0.8–0.99 with each orbital element, so it is the direction that sets their uncertainties, and narrowing it by 1.25−1/2≃0.901.25^{-1/2}\simeq0.90 reduces their widths by 6–10%, as found. The orbital elements narrow by the same amount in all spacetimes. Over noise realizations r0⋅K\mathbf{r}_0\cdot K averages to zero, but for one data set it need not be small. For a hair entering as ξ2\xi^2, Eq. (C.1) gives, along ξ\xi, with c=∂r/∂(ξ2)c = \partial\mathbf{r}/\partial(\xi^2), χ2=∣r0+cξ2∣2=χ02+ξ2r0⋅c+ξ4∣c∣2\chi^2 = |\mathbf{r}_0+c\xi^2|^2 = \chi_0^2+\xi^2\mathbf{r}_0\cdot c+\xi^4|c|^2. Here sξ=2r0⋅cs_\xi= 2\mathbf{r}_0\cdot c gives the width of Appendix A.3, and the quartic term, together with a curved degeneracy with the other parameters, makes the posterior narrower.

Validation

  • Exact solver: The exact solver obtains ϕ(t)\phi(t) by Newton inversion of t(ϕ)=∫F−1 dϕt(\phi)=\int F^{-1}\,\mathrm{d}\phi and is otherwise identical to GINN, with the same observables and optimizer but without the net-

work and the physics term. For Schwarzschild, the estimates of both models differ by less than 0.01σ\sigma, and their uncertainties by less than 1%. The network reproduces the exact non-Keplerian correction to 0.4% (rms), and the improvement of the fit over a Keplerian orbit is the same for both, Δχdata2=0.127\Delta\chi^2_{\mathrm{data}} = 0.127 with the network and 0.126 with the exact solver.

– Physics tolerance: Repeating the Schwarzschild optimization for ϵ\epsilon between 3×10−23 \times10^{-2} and 10−510^{-5} gives three regimes. For ϵ≳3×10−3\epsilon\gtrsim3 \times10^{-3} the network is only loosely bound to the equation of motion and absorbs noise, so that the minimum χ2\chi^2 falls below the exact-solver value and MM and DD are biased. For 10−4≲ϵ≲10−310^{-4} \lesssim\epsilon\lesssim10^{-3} the results agree with the exact solver and do not depend on ϵ\epsilon. For ϵ≲3×10−5\epsilon\lesssim3 \times10^{-5} the problem becomes too stiff to converge in the same number of steps. We use ϵ=10−4\epsilon= 10^{-4}.

– Multi-start: Each optimization is repeated from different starting values of the hair spanning the prior or several orders of magnitude, and for Schwarzschild from different initializations of the network. The final values are stable: across all spacetimes, the largest spread in the final values of the hair, MM, and DD over three starts is below 0.04σ0.04\sigma (except for Hairy Horndeski spacetime where they agree within 0.15σ0.15\sigma).

– Exact posterior: Furthermore, replacing the Gaussian approximation with the exact non-Gaussian posterior shifts the means by at most 0.1σ0.1\sigma in all cases. The non-Gaussianity mainly affects the widths of the squared-hair parameters: the exact posteriors for qRNq_{\mathrm{RN}}, qBq_{\mathrm{B}}, and qJNWq_{\mathrm{JNW}} have standard deviations 0.82, 0.70, and 0.72 times those of the Gaussian, the orbital elements 0.89–0.94 times, and qHq_{\mathrm{H}} and δ\delta 1.01 times. The exact posterior reweights between 86% (Bardeen, JNW) and 100% (Schwarzschild, Horndeski, Yukawa) of the higher-order samples effectively.

Posteriors of the other spacetimes

Figures E.1–E.5 show the posteriors of the mass, the orbital elements and the hair for the other spacetimes, from the Laplace approximation, with the network following the parameters. The reference-frame parameters are omitted for brevity; their values are listed in Table 2. The contours enclose 39% and 86% of the two-dimensional probability (1σ\sigma and 2σ\sigma), and the titles give the median and the 16th and 84th percentiles.

ParameterSchwarzschildHairy HorndeskiReissner-NordströmJNWBardeenYukawa
MM [106 M⊙10^{6}\,M_{\odot}]4.32±0.224.32 \pm0.224.36±0.244.36 \pm0.244.32±0.224.32 \pm0.224.32±0.224.32 \pm0.224.32±0.224.32 \pm0.224.36±0.234.36 \pm0.23
DD [kpc]8.33±0.218.33 \pm0.218.32±0.218.32 \pm0.218.32±0.218.32 \pm0.218.33±0.218.33 \pm0.218.32±0.218.32 \pm0.218.35±0.218.35 \pm0.21
Hair ξ\xi–qHq_{\mathrm{H}}qRNq_{\mathrm{RN}}qJNWq_{\mathrm{JNW}}qBq_{\mathrm{B}}δ\delta
–0.0019±0.00350.0019 \pm0.00350.33±2.530.33 \pm2.530.9±1280.9 \pm1283.0±70.43.0 \pm70.40.0047±0.01190.0047 \pm0.0119
aa [mas]124.58±0.96124.58 \pm0.96124.61±0.96124.61 \pm0.96124.60±0.96124.60 \pm0.96124.58±0.97124.58 \pm0.97124.58±0.96124.58 \pm0.96124.51±0.97124.51 \pm0.97
ee0.8821±0.00220.8821 \pm0.00220.8817±0.00240.8817 \pm0.00240.8822±0.00220.8822 \pm0.00220.8821±0.00220.8821 \pm0.00220.8822±0.00220.8822 \pm0.00220.8818±0.00250.8818 \pm0.0025
ii [∘^\circ]134.80±0.45134.80 \pm0.45134.78±0.45134.78 \pm0.45134.79±0.45134.79 \pm0.45134.80±0.45134.80 \pm0.45134.80±0.45134.80 \pm0.45134.83±0.45134.83 \pm0.45
Ω\Omega [∘^\circ]227.13±0.71227.13 \pm0.71227.08±0.72227.08 \pm0.72227.14±0.71227.14 \pm0.71227.13±0.71227.13 \pm0.71227.13±0.71227.13 \pm0.71227.06±0.74227.06 \pm0.74
ω\omega [∘^\circ]65.25±0.6965.25 \pm0.6965.11±0.7565.11 \pm0.7565.26±0.6965.26 \pm0.6965.25±0.6965.25 \pm0.6965.25±0.6965.25 \pm0.6965.10±0.7965.10 \pm0.79
tp−2002t_{\mathrm{p}} - 2002 [yr]0.3210±0.00600.3210 \pm0.00600.3202±0.00660.3202 \pm0.00660.3211±0.00600.3211 \pm0.00600.3210±0.00600.3210 \pm0.00600.3210±0.00600.3210 \pm0.00600.3200±0.00680.3200 \pm0.0068
x0x_0 [mas]0.130±0.1360.130 \pm0.1360.114±0.1390.114 \pm0.1390.130±0.1370.130 \pm0.1370.129±0.1360.129 \pm0.1360.130±0.1370.130 \pm0.1370.120±0.1400.120 \pm0.140
y0y_0 [mas]−0.278±0.191-0.278 \pm0.191−0.284±0.191-0.284 \pm0.191−0.279±0.191-0.279 \pm0.191−0.279±0.192-0.279 \pm0.192−0.279±0.192-0.279 \pm0.192−0.286±0.192-0.286 \pm0.192
vx,0v_{x,0} [mas yr−1^{-1}]0.108±0.0410.108 \pm0.0410.132±0.0600.132 \pm0.0600.109±0.0410.109 \pm0.0410.108±0.0410.108 \pm0.0410.108±0.0410.108 \pm0.0410.124±0.0600.124 \pm0.060
vy,0v_{y,0} [mas yr−1^{-1}]0.105±0.0540.105 \pm0.0540.113±0.0560.113 \pm0.0560.105±0.0540.105 \pm0.0540.105±0.0540.105 \pm0.0540.105±0.0540.105 \pm0.0540.109±0.0560.109 \pm0.056
vz,0v_{z,0} [km s−1^{-1}]5.9±7.55.9 \pm7.56.7±7.76.7 \pm7.75.8±7.55.8 \pm7.55.9±7.55.9 \pm7.55.9±7.55.9 \pm7.56.7±7.86.7 \pm7.8

Table 2. Estimates and 1σ1\sigma uncertainties from the Gauss-Newton minimum and the Laplace approximation, with the network following the parameters. The hair parameters and their priors are listed in Table 1. For qRNq_{\mathrm{RN}}, qBq_{\mathrm{B}} and qJNWq_{\mathrm{JNW}}, which enter the metric only through their square, the sign of the estimate has no meaning, and the width is computed as in § 4.6; their 95% upper limits on ∣ξ∣|\xi| are given in § 5. The Yukawa results are for λ=103\lambda= 10^{3} AU.

Posterior for the hairy Horndeski spacetime

Figure E.1. Posterior for the hairy Horndeski spacetime.

Posterior for the Yukawa correction at fixed $\lambda = 10^3$ AU

Figure E.2. Posterior for the Yukawa correction at fixed λ=103\lambda= 10^3 AU.

Posterior plot for the Reissner-Nordström spacetime

Figure E.3. Posterior for the Reissner-Nordström spacetime.

Posterior plot for the Bardeen spacetime

Figure E.5. Posterior for the Bardeen spacetime.

Posterior plot for the JNW spacetime

Figure E.4. Posterior for the JNW spacetime.

References

  1. [1]Abd El Dayem, K. et al. 2025, Astron. Astrophys., 698, L15
  2. [2]Abuter, R. et al. 2020, Astron. Astrophys., 636, L5
  3. [3]Abuter, R. et al. 2022, Astron. Astrophys., 657, L12
  4. [4]Argüelles, C. R., Mestre, M. F., Becerra-Vergara, E. A., et al. 2022, MNRAS, 511, L35
  5. [5]Ayón-Beato, E. & García, A. 2000, Phys. Lett. B, 493, 149
  6. [6]Bambhaniya, P., Dalal, P., Vicentini, G. H., et al. 2026, Phys. Dark Univ., 53, 102356
  7. [7]Bambhaniya, P., Joshi, A. B., Dey, D., et al. 2024, Eur. Phys. J. C, 84, 124
  8. [8]Bardeen, J. M. 1968, in Proceedings of the 5th International Conference on Gravitation and the Theory of Relativity (GR5), Tbilisi, 174
  9. [9]Baydin, A. G., Pearlmutter, B. A., Radul, A. A., & Siskind, J. M. 2018, J. Mach. Learn. Res., 18, 1
  10. [12]Bergliaffa, S. E. P., Maier, R., & de Oliveira Silvano, N. 2021 [arXiv:2107.07839]arxiv.org/abs/2107.07839
  11. [13]Bhadra, A., Sarkar, K., Datta, D. P., & Nandi, K. K. 2007, Mod. Phys. Lett. A, 22, 367
  12. [14]Borka, D., Jovanović, P., Borka Jovanović, V., & Zakharov, A. F. 2013, JCAP, 11, 050
  13. [15]Brans, C. H. 1962, Phys. Rev., 125, 2194
  14. [16]Chantada, A. T., Landau, S. J., Protopapas, P., Scóccola, C. G., & Garraffo, C. 2023, Phys. Rev. D, 107, 063523
  15. [17]Crespi, V., Argüelles, C. R., Becerra-Vergara, E. A., et al. 2026, MNRAS, 546, staf1854
  16. [18]Cutler, C., Kennefick, D., & Poisson, E. 1994, Phys. Rev. D, 50, 3816
  17. [19]Dadhich, N., Maartens, R., Papadopoulos, P., & Rezania, V. 2000, Phys. Lett. B, 487, 1
  18. [20]Darwin, C. G. 1959, Proc. R. Soc. Lond. A, 249, 180
  19. [21]Dayem, K. A. E., Vincent, F. H., Heissel, G., Paumard, T., & Perrin, G. 2026, Astron. Astrophys., 707, A207
  20. [27]Della Monica, R. 2025, Astron. Astrophys., 698, A193
  21. [28]Della Monica, R. & de Martino, I. 2022, JCAP, 03, 007
  22. [34]Eckart, A. & Genzel, R. 1997, Mon. Not. Roy. Astron. Soc., 284, 576
  23. [35]Event Horizon Telescope Collaboration. 2022, Astrophys. J. Lett., 930, L12
  24. [36]Foreman-Mackey, D., Hogg, D. W., Lang, D., & Goodman, J. 2013, Publ. Astron. Soc. Pac., 125, 306
  25. [37]Foreman-Mackey, D. et al. 2021, J. Open Source Softw., 6, 3285
  26. [38]Foschi, A. et al. 2026, Improving constraints on the Yukawa correction at the Galactic Center with multiple stellar orbits
  27. [39]Genzel, R., Eisenhauer, F., & Gillessen, S. 2010, Rev. Mod. Phys., 82, 3121
  28. [40]Ghez, A. M., Klein, B. L., Morris, M., & Becklin, E. E. 1998, Astrophys. J., 509, 678
  29. [41]Ghez, A. M. et al. 2008, Astrophys. J., 689, 1044
  30. [42]Gillessen, S., Eisenhauer, F., Trippe, S., et al. 2009, Astrophys. J., 692, 1075
  31. [43]Gillessen, S., Plewa, P. M., Eisenhauer, F., et al. 2017, Astrophys. J., 837, 30
  32. [44]Gillessen, S., Plewa, P. M., Eisenhauer, F., et al. 2016, Centre de Donnees astronomique de Strasbourg (CDS)
  33. [45]GRAVITY Collaboration. 2017, Astron. Astrophys., 602, A94
  34. [47]GRAVITY Collaboration. 2019, Astron. Astrophys., 625, L10
  35. [48]GRAVITY Collaboration. 2024, Astron. Astrophys., 692, A242
  36. [49]Hassan Puttasiddappa, P., Mushtaq, M., Ramirez, W., & Mota, D. F. 2026, Bounds on A at the Galactic Center
  37. [50]Hassan Puttasiddappa, P., Rodrigues, D. C., & Mota, D. F. 2025, Eur. Phys. J. C, 85, 974
  38. [51]Hees, A. et al. 2017, Phys. Rev. Lett., 118, 211101
  39. [53]Horndeski, G. W. 1974, International Journal of Theoretical Physics, 10, 363
  40. [54]Janis, A. I., Newman, E. T., & Winicour, J. 1968, Phys. Rev. Lett., 20, 878
  41. [55]Jovanović, P., Borka, D., & Jovanović, V. B. 2026, Symmetry, 18, 557
  42. [56]Kingma, D. P. & Ba, J. 2015, in International Conference on Learning Representations (ICLR)
  43. [57]Loshchilov, I. & Hutter, F. 2017, in International Conference on Learning Representations (ICLR)
  44. [58]MacKay, D. J. C. 1992, Neural Comput., 4, 448
  45. [60]Navarrete, C., Vázquez-Chávez, F., Cruz-Osorio, A., & Ortiz, N. 2026, Mon. Not. Roy. Astron. Soc., 546, stag059
  46. [61]Nelder, J. A. & Mead, R. 1965, Comput. J., 7, 308
  47. [62]Nocedal, J. & Wright, S. J. 2006, Numerical Optimization, 2nd edn. (New York: Springer)
  48. [63]Nordström, G. 1918, Koninklijke Nederlandse Akademie van Wetenschappen Proceedings, 20, 1238
  49. [64]Plewa, P. M. 2026, GC-OrbitFit: Fit the orbit of the star S2, https://github.com/pmplewa/GC-OrbitFit, gitHub repository, accessed October 2026
  50. [65]Plewa, P. M., Gillessen, S., Eisenhauer, F., et al. 2015, MNRAS, 453, 3234
  51. [66]Plewa, P. M. & Sari, R. 2018, Mon. Not. Roy. Astron. Soc., 476, 4372
  52. [70]Reissner, H. 1916, Annalen der Physik, 355, 106
  53. [71]Röver, L., Schäfer, B. M., & Plehn, T. 2024 [arXiv:2403.13899]arxiv.org/abs/2403.13899
  54. [72]Schödel, R. et al. 2002, Nature, 419, 694
  55. [73]Schwarzschild, K. 1916, Sitzungsber. Preuss. Akad. Wiss. Berlin (Math. Phys.), 189
  56. [74]Sellentin, E., Quartin, M., & Amendola, L. 2014, Mon. Not. Roy. Astron. Soc., 441, 1831
  57. [76]Stefanou, P., Urbán, J. F., & Pons, J. A. 2023, Mon. Not. Roy. Astron. Soc., 526, 1504
  58. [77]Thompson, W. et al. 2023, Astron. J., 166, 164
  59. [78]Trotta, R. 2008, Contemp. Phys., 49, 71
  60. [79]Verma, A., Sourav, S., Aluri, P. K., & Mota, D. F. 2025 [arXiv:2508.12032]arxiv.org/abs/2508.12032
  61. [80]Virbhadra, K. S. 1997, Int. J. Mod. Phys. A, 12, 4831
  62. [81]Virbhadra, K. S. & Ellis, G. F. R. 2002, Phys. Rev. D, 65, 103004
  63. [84]Zajaček, M., Tursunov, A., Eckart, A., & Britzen, S. 2018, Mon. Not. Roy. Astron. Soc., 480, 4408
  64. [85]Zakharov, A. F. 2022, MNRAS, 513, L6

Paper details

vlnc:W_c9chmnv5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents