Skip to content

Methods ​

What Exoplanet Science Lab computes and how: the equations as implemented, the parameters and their default priors, the numerical method with its measured accuracy, the assumptions and known limits, and the references. Written to be cited and adapted in the methods section of a paper.

This is the Markdown copy of the site's Methods page (methods/, https://exoplanet.science/methods/). Equations are written in LaTeX between dollar signs.

Each section below describes one technique as the source code implements it; the file is named under the heading. Accuracy statements link to the test or the Validation entry that produces the number. Where something is not modelled, was not tested, or disagrees with a published value, the section says so under "Assumptions and limits".

Notation: R⋆, M⋆ stellar radius and mass; k=Rp/R⋆; a/R⋆ scaled semi-major axis; i inclination; P period; T0 mid-transit time; e, ω eccentricity and argument of periastron; ν, E, M true, eccentric and mean anomaly; μ=1−r2 with r the projected distance from the centre of the stellar disc in stellar radii.

Contents ​

  1. Time standards and conventions
  2. Transit model and limb darkening
  3. Exposure averaging
  4. Kepler solver and orbits
  5. Radial velocities
  6. Rossiter-McLaughlin effect
  7. The likelihood
  8. Nested sampling
  9. Model comparison and its cautions
  10. Ephemerides, timing, periodograms and orbital decay
  11. Several planets and Kepler-tied a/R★
  12. Physical priors: stable orbits and period aliases
  13. Phase curves and eclipse maps
  14. Exomoons
  15. N-body dynamics
  16. Photodynamics
  17. Time-series spectroscopy
  18. Derived quantities and uncertainty propagation
  19. Catalogue sources, merging and provenance
  20. Observability and sky positions
  21. Image photometry
  22. TESS, Kepler and K2 light curves: quality, stitching, transit-aware detrending, K2 systematics
  23. Image inspection, focus and comparison stars
  24. Noise diagnostics
  25. The session format and reproducibility
  26. Validation
  27. Reproducibility table
  28. How to cite and report
  29. References

1. Time standards and conventions ​

Code: src/core/kepler.js, src/core/data.js, src/core/fits.js, src/core/observability.js, src/core/derive.js

Geometry ​

The sky plane is X,Y in units of the stellar radius and +Z points to the observer. With u=ω+ν the planet is at

X=−rcos⁡u,Y=−rsin⁡ucos⁡i,Z=rsin⁡usin⁡i,r=aR⋆(1−ecos⁡E)

so the planet is in front of the star when Z>0. Mid-transit T0 is inferior conjunction, ω+ν=π/2; mid-eclipse is superior conjunction, ω+ν=−π/2. These are not the times of smallest projected separation on an inclined eccentric orbit (see Kepler solver and orbits). Radial velocities are positive when the star recedes. ω is the same number as the ω of radvel and pyaneti (Validation: sign and angle conventions).

Times in the analyzer ​

Times are used as given. The reader (parseLightCurve) recognises the zero point of the time column from a header or from the size of the numbers (JD; MJD = JD − 2400000.5; RJD = JD − 2400000; BTJD = JD − 2457000; BKJD = JD − 2454833) and adds it; it does not convert between time scales (UTC, TT, TDB) or between geocentric, heliocentric and barycentric times. Inside a fit, times are measured from an integer reference (data.tRef, the floor of the first time) so that Julian dates keep their precision in double arithmetic; results are reported as absolute times again. Catalogue epochs are taken as the catalogue lists them and assumed to be BJDTDB; a different stated time system is flagged, not converted.

Times from image headers ​

The photometry page converts the UTC mid-exposure time of each frame to BJDTDB Eastman et al. 2010 when the target coordinates are known (utcToBjdTdb):

BJDTDB=JDUTC+ΔAT+32.184s+0.001657ssin⁡g+0.000014ssin⁡2g86400s+r⊕⋅n^c

with ΔAT = TAI − UTC from a leap-second table (1972 to 2017-01-01, 37 s held afterwards), g=357.53∘+0.9856003∘(JD−2451545.0), n^ the unit vector of the target's ICRS coordinates, and r⊕ the Earth's barycentric position: the low-precision solar theory of Meeus 1998 (chapter 25) for the heliocentric Earth plus the Sun's offset from the barycentre due to Jupiter, Saturn, Uranus and Neptune from the Keplerian elements of Standish & Williams (JPL). The mid-exposure time is the header's start time plus half the exposure, unless the keyword or its comment says it is already a mid or end time; TAI, GPS, TT and TDB header times are first converted to UTC.

The planner uses a second, simpler conversion (jdUtcToBjdTdb in observability.js): the light time to the Sun's centre instead of the barycentre, and TDB = TT.

Constants ​

src/core/derive.js: R⊙ = 6.957e8 m, RJup = 7.1492e7 m, R⊕ = 6.3781e6 m, au = 1.495978707e11 m, GM⊙ = 1.32712440018e20 m³ s⁻², MJup/M⊙ = 9.5459426398e-4, G = 6.6743e-11 (SI), Teff,⊙ = 5772 K, day = 86400 s.

Accuracy and checks ​

  • UTC to BJD_TDB, 35 targets and dates: within 0.045 s. (Validation: From images to a light curve; test/fits.test.mjs)
  • Barycentric time correction (BJD_TDB − JD_UTC): worst 3.92 s. (Validation: Which transits can be observed; test/observability.test.mjs)
  • Sign and ω convention against radvel, pyaneti and a reference obtained by differentiating the orbit in 30-digit arithmetic: node test/validation.test.mjs ("sign", "omega").

Assumptions and limits ​

  • No conversion between time standards in the analyzer, the Timing tab, the catalogue or the planner's input: mixing BJDTDB with JDUTC or HJD moves a mid-transit time by up to about 8 minutes (heliocentric correction) plus 69 s (TDB − UTC in 2017 and later). The user must supply one consistent time standard.
  • The header conversion leaves out the observatory's position on the Earth (up to 21 ms), the Shapiro delay, parallax and proper motion. HJD header times are passed through unconverted. The leap-second table ends in 2017.
  • The project has three separate implementations of Sun position and time conversion (fits.js, astro.js, observability.js) with different accuracy; they have not been merged.

References: Eastman et al. 2010; Meeus 1998; Standish & Williams (JPL); Astropy Collaboration 2022.

2. Transit model and limb darkening ​

Code: src/core/transit.js, src/core/limbdark.js, src/core/ldest.js, src/core/ldspec.js, tools/build_ld_table.py, tools/build_ld_spectral.py

Limb-darkening laws ​

Every law is mapped onto the four-coefficient law of Claret 2000,

I(μ)I(1)=1−∑n=14an(1−μn/2),μ=1−r2

of which the others are exact special cases: linear I=1−u(1−μ) is a2=u; quadratic I=1−u1(1−μ)−u2(1−μ)2 is a2=u1+2u2, a4=−u2; the (q1,q2) form of Kipping 2013 is the quadratic law with u1=2q1q2, u2=q1(1−2q2); the square-root law I=1−c(1−μ)−d(1−μ) is a1=d, a2=c. A profile that is negative anywhere, brightens toward the limb, or has non-positive total flux is outside the prior: the likelihood rejects it (tested on 25 values of μ).

Flux ​

With F(r)=∫0rI(r′)r′dr′, which is in closed form for this law, the vector field F(r)/r2(x,y) has divergence I(r), so the flux behind the planet's disc is a line integral around the boundary of the occulted region (Green's theorem):

ΔFF=12πF(1)∮F(r)r2(xdy−ydx),F(1)=1−∑an2+a12.5+a23+a33.5+a44

The boundary is an arc of the planet's limb, parametrised by the angle φ about the planet's centre, plus, during ingress and egress, an arc of the stellar limb where the integrand is the constant F(1) and the contribution is analytic. With the planet's centre at projected distance z and radius p, r2=z2+p2+2zpcos⁡φ along the planet's limb. When the planet is wholly on the disc, further than 0.02 stellar radii from the limb and not covering the centre, the integrand is smooth and periodic and a 12-node midpoint rule is used. Otherwise the arc is integrated with 20 Gauss-Legendre nodes after the substitution φ=φA+(π−φA)u2, which removes the square-root branch point where the arc meets the stellar limb. Near the stellar centre (r2<10−3) F(r)/r2 is evaluated by a series. A uniform disc uses the closed-form overlap area of two circles. The model flux is 1−ΔF/F when the planet is in front (Z>0) and 1 otherwise.

Where the coefficients come from ​

  • Free parameters, sampled with the other parameters.
  • A model-atmosphere table (src/data/ld_table.js, built by tools/build_ld_table.py): PHOENIX-ACES specific intensities Husser et al. 2013, obtained through the file list and cache of LDTK Parviainen & Aigrain 2015, integrated with photon weighting over 37 passbands for Teff = 2300 to 8000 K, log⁡g = 4.0, 4.5, 5.0 (5.5 at or below 4000 K) and [Fe/H] = −0.5, 0, +0.5. Measured response curves (those shipped with pylightcurve) are used for TESS, Kepler, CHEOPS, Johnson UBV, Cousins RI, SDSS ugriz and 2MASS JHKs; the remaining bands are boxcars. The stellar limb is placed at the steepest gradient of the model intensity and the radius rescaled to it, as in ExoTETHyS Morello et al. 2020; the four-coefficient law is fitted by least squares for μ≥0.1, and refitted under constraints when the unconstrained fit is not monotonic or goes negative. The other laws are fitted to that profile (constrained to physical coefficients). The application interpolates trilinearly in Teff, log⁡g and [Fe/H] and clamps, with a flag, outside the grid.
  • Any wavelength range (src/core/ldspec.js, table built by tools/build_ld_spectral.py): the same models in 357 bins from 200 to 5500 nm, stored as a mean plus 40 principal components. The profile of a band is the flux- and response-weighted mean of the bins' profiles, Iband(μ)=∑bwbFbIb(μ)/∑bwbFb, then constrained to be physical.
  • A grey atmosphere (fallback when no table applies): I(μ)∝∫0∞Bλ[T(μx)]e−xdx with T4=34Teff4[τ+q(τ)], q(τ)=0.7104−0.1331e−3.4488τ, averaged over nine wavelengths of the band with photon weighting; the law is then fitted by least squares on 60 values of μ from 0.02 to 1. It depends on Teff and wavelength only.
  • Following the star in a fit: Teff (and [Fe/H] when a table is in use) are sampled with priors at their measured uncertainties, and the coefficients are computed from each sample (makeLDTie: tabulated once per fit on 17 × 7 knots plus every table node inside the prior, bilinear in between). The coefficients are then confined to the family a star of that kind can produce.

Accuracy and checks ​

Assumptions and limits ​

  • The flux is accurate to about 1e-10, not to rounding error; the closed form of Agol, Luger & Foreman-Mackey 2020 for the quadratic law is more accurate and faster. Every law is integrated numerically here.
  • Not modelled: star spots and faculae, stellar oblateness and gravity darkening, planetary oblateness, rings, refraction, the planet's night-side emission during transit (unless the phase-curve model is on). Dilution by other stars in the aperture is an optional term of the likelihood (The likelihood); the bundled TESS light curve of HD 80606 is diluted by HD 80607, and its stored example result is fitted without that term.
  • One limb-darkening law and band for all light curves in a fit (no multi-band fits).
  • The comparison with ExoTETHyS tests the processing, the grid and the interpolation, not the atmosphere physics: both start from the same PHOENIX files. PHOENIX and ATLAS models differ by 0.0125 rms in I(μ) on average, which is larger than every processing error quoted here. The boxcar and JWST bands have no external reference. Below μ=0.1 the profile is an extrapolation. Stars outside the grid are clamped to its edge. Beyond 5.5 µm there is no table.

References: Claret 2000; Kipping 2013; Husser et al. 2013; Parviainen & Aigrain 2015; Morello et al. 2020; Agol, Luger & Foreman-Mackey 2020; Tsiaras et al. 2016.

3. Exposure averaging ​

Code: src/core/model.js (autoSupersample), src/core/transit.js (transitModel), src/core/session.js

A detector integrates the flux over each exposure. When an exposure time Δt is given, the model at time ti is the mean of n evaluations at the centres of n equal slices of the exposure:

f¯(ti)=1n∑j=0n−1f(ti+(j−n−12)Δtn)

With several planets the product of the planets' fluxes is averaged, not the product of the averages. The number of sub-samples is chosen automatically from the ingress duration τ=(T14−T23)/2 (closed-form durations, at least 1e-4 d) and the depth δ=k2:

n=⌈Δtτδ6×10−6⌉,made odd, at most 41; n=1 when the expression is ≤1

This follows from the error of the midpoint rule, about (curvature) × (Δt/n)2/24 with a curvature of order 4δ/τ2 during ingress, for a target of 1 ppm. Each epoch of a light curve may carry its own exposure time and then gets its own n.

Accuracy and checks ​

  • Against the mean of 4001 evaluations of the exact quadratic-law flux (node test/validation.test.mjs): 30 min exposures of a hot Jupiter (ingress 21 min), automatic n = 41: 5.0e-7; 2 min exposures of the same planet, n = 5: 6.9e-7; 30 min exposures of a small planet on a 1 d orbit (ingress 1.6 min), n = 41: 1.3e-7; with 101 sub-samples the three cases give 7.4e-8, 1.7e-9 and 1.3e-8. See Validation: exposure averaging.

Assumptions and limits ​

  • The count is capped at 41, so the residual grows for exposures much longer than ingress; the analyzer offers no manual setting; a count can be given by hand only through the programming interface (spec.supersample and segments[s].supersample of buildProblem).
  • The exposure is assumed to be a uniform window centred on the time stamp, with no dead time and no read-out weighting. The exposure time is whatever the user or the file header supplies.
  • The count is chosen once, from the parameter values at the start of a fit, not re-evaluated for each sample. Because a session saved after a fit holds the fitted values, re-running a stored session can choose a different count: hd80606-tess.json re-runs to 12,813 likelihood calls instead of the stored 12,837 (0.004σ in the result).

References: Winn 2010.

4. Kepler solver and orbits ​

Code: src/core/kepler.js

The mean anomaly is M(t)=M0+2π(t−T0)/P, with M0 the mean anomaly at inferior conjunction: νc=π/2−ω, Ec=2arctan[(1−e)/(1+e)tan⁡(νc/2)], M0=Ec−esin⁡Ec. Kepler's equation E−esin⁡E=M is solved from the starting value E0=M+0.85esgn(sin⁡M) Danby 1987 with the update

δ1=−ff′,δ2=−ff′+12δ1f″,δ3=−ff′+12δ2f″+16δ22f‴,E←E+δ3

where f=E−esin⁡E−M, iterated until |δ3|<10−15max(1,|E|) (at most 30 iterations; two to four are needed). The true anomaly is ν=2atan2(1+esin⁡E2, 1−ecos⁡E2). Eccentricities are limited to e≤0.999. Circular orbits skip the solver.

Event times follow from the anomaly of the event: periastron ν=0, mid-eclipse ω+ν=−π/2. Durations (T14, T23 and the eclipse equivalents E14, E23) are found numerically: the time at which the projected separation equals 1±k is bracketed from the conjunction and refined by 60 bisections, on each side. The closed-form durations of Winn 2010 (eqs. 14 to 16) are used only to choose the exposure sub-sampling and for catalogue estimates. The impact parameter is b=(a/R⋆)cos⁡i(1−e2)/(1+esin⁡ω).

Accuracy and checks ​

  • node test/validation.test.mjs, against 30-digit arithmetic: true anomaly within 2.4e-15 rad for e = 0.05 to 0.999 (residual of the equation 4.4e-16); periastron and eclipse times of 21 orbits within 2.3e-10 s; T14, T23, E14, E23 of 10 orbits within 1.3e-10 s. The closed-form duration is exact for circular orbits and up to 0.91% short otherwise. See Validation: the orbit and its timing.

Assumptions and limits ​

  • One published number is not reproduced. For HD 80606 b this code gives an eclipse duration E14 = 0.08337 d, equal to root finding on the 30-digit orbit (0.08337 d); Pearson et al. 2022 list 0.0717 d. The transit duration agrees (0.4985 d against 0.499 d). No method tried reproduces the published eclipse duration, and its origin is unknown.
  • For HD 80606 b the time of smallest projected separation is 11.4 min before conjunction. A mid-time measured from the symmetry of a light curve is closer to the midpoint of the contacts than to conjunction; the three should not be mixed for inclined eccentric orbits (Validation).
  • Orbits are fixed Keplerian ellipses: no apsidal or nodal precession, no light-travel-time correction of transit times, no relativistic terms. Mutual perturbations need the N-body option.

References: Danby 1987; Winn 2010; Pearson et al. 2022.

5. Radial velocities ​

Code: src/core/rv.js, src/core/model.js

The velocity of the star in dataset d at time t is

vd(t)=γd+∑pKp[cos⁡(ωp+νp(t))+epcos⁡ωp]+c1(t−tref)+c2(t−tref)2K=(2πGP)1/3Mpsin⁡i(M⋆+Mp)2/31−e2

with the same orbit as the transit model, so the star recedes before mid-transit and approaches after it. tref is the earliest velocity. A planet's amplitude is computed from its mass (the sampled parameter is Mp in Jupiter masses), or K can be given directly. γd is the zero point of instrument d and is integrated out analytically (see The likelihood); each instrument also has a jitter sd added in quadrature to its uncertainties. c1 and c2 (keys rvlin, rvquad) are shared by all instruments. The inverse (mass from K) is solved by fixed-point iteration.

Accuracy and checks ​

Assumptions and limits ​

  • The reflex motion uses the two-body factor Mp/(M⋆+Mp). EXOTIC scales by Mp/M⋆, so for the same velocities the mass from this code is higher by Mp/M⋆ (0.38% for HD 80606 b).
  • Not modelled: stellar activity (no Gaussian process or activity indicators in the likelihood), light-travel time, instrument-specific trends. The jitter is white. The Rossiter-McLaughlin anomaly is optional (Rossiter-McLaughlin effect).
  • Stellar mass (and radius, when a/R⋆ follows from Kepler's law) are held at their input values unless made free; the uncertainty of a planet mass then does not include that of the stellar mass, and the Results tab says so.
  • For the HD 80606 velocities χ2 = 1555 for 264 points with the jitters at zero (stored result public/examples/results/hd80606-rv.json), so formal uncertainties from that fit are too small unless the jitters are freed.

References: Fulton et al. 2018; Pearson et al. 2022; Zellem et al. 2020.

6. Rossiter-McLaughlin effect ​

Code: src/core/rm.js, src/core/model.js (buildProblem)

When vsini is set, every planet in front of the star adds to the velocities the shift of the flux-weighted mean velocity of the light that is not covered. The star rotates as a solid body; in the sky frame of the orbit (planet moving toward +X across the disc, the orbit normal projected onto +Y, the chord at Y=−b) a point of the disc recedes at

v(x,y)=vsin⁡i⋆(xcos⁡λ−ysin⁡λ)

with λ the sky-projected obliquity, the angle from the projected orbit normal to the projected stellar spin (ellc's convention Maxted 2016; λ=0 aligned, 180∘ retrograde). With the Claret-law intensity I of the fit's limb darkening (or another law given for the spectrograph's band through spec.rm) the anomaly is

Δv=−∫occIvdAF⋆−∫occIdA

the quantity Ohta et al. expand for a small planet Ohta, Taruya & Suto 2005. The integral over the occulted area is not approximated: the first moment ∫occIx′dA along the line from the star's centre to the planet's is a line integral of F(r)=∫0rIrdr (the closed form of the transit model) around the occulted region, an arc of the planet's limb with 40 Gauss-Legendre nodes plus, in ingress and egress, an arc of the stellar limb; the moment perpendicular to that line vanishes by symmetry. The blocked flux is the transit model's. Each planet has its own λ (key lambda@c, falling back to lambda); the star's vsini is in km/s.

What a Gaussian fit to the CCF measures (option) ​

With spec.rm.method = 'ccf' the anomaly is instead the shift of the centre of a Gaussian fitted to the cross-correlation function. Every surface element contributes a Gaussian line of dispersion βp (instrumental and thermal, beta; macroturbulence zeta adds ζ2/2 to βp2, a Gaussian approximation), the occulted light is a fraction f with mean velocity vp (both from the integrals above), and to first order in f the linearised least-squares fit moves by

Δv=−f1−fgvp(2β02β02+βp2)3/2exp⁡(−vp22(β02+βp2))

Hirano et al. 2010; Boué et al. 2013, with β0 the dispersion of the out-of-transit CCF (measured, beta0, or that of the Gaussian that best fits the modelled rotation-broadened profile) and g the correction for a profile that is not Gaussian (its area over the fitted Gaussian's, and the curvature term the out-of-transit residual leaves in the Hessian): 1.000 for vsin⁡i/βp = 0.57, 0.974 for 4. Both are tabulated against vsin⁡i/βp once per fit.

Exposures, limb darkening, priors (options) ​

Each velocity dataset may carry its exposure (exposure_min; or spec.rm.exptime for all), over which the anomaly is averaged with supersample sub-samples (default 5); the Keplerian part is not averaged. The limb darkening of the anomaly is the spectrograph's band when spec.rm.ld is given. A catalogue vsin⁡i⋆ enters as a normal prior truncated at zero (vsiniPrior). The true obliquity follows from cos⁡ψ=cos⁡i⋆cos⁡i+sin⁡i⋆sin⁡icos⁡λ when i⋆ is known.

Accuracy and checks ​

  • Against a brute-force pixel sum (4000 × 4000 pixels over the planet's disc, star flux by scipy quadrature; test/make_rm_reference.py): 160 cases (uniform, quadratic and four-coefficient laws, k = 0.01 to 0.2, inside the disc, on the limb, in ingress and egress, four obliquities), largest difference 1.0e-7 vsin⁡i or 8.9 ppm of the anomaly, every case within the pixel sum's own error (node test/rm.test.mjs).
  • Uniform star with the planet on the disc against Δv=−vsin⁡ixsk2/(1−k2): 3.5e-18. The small-planet limit of Ohta et al. Ohta, Taruya & Suto 2005: relative difference 8.0e-4, 8.8e-5 and 7.9e-6 at k = 0.03, 0.01, 0.003, falling as k2.
  • CCF form against a brute-force disc integration (3000 × 3000 pixels of the star, 700 × 700 of the planet, each a Gaussian line, binned in velocity) with Gaussians fitted by scipy to the out-of-transit and in-transit profiles (test/make_rm_ccf_reference.py; 120 cases: vsin⁡i = 1.7 to 12 km/s with β = 3 km/s, and 5 km/s with β = 2.5, ζ = 4 km/s; λ = 0°, 40°; k = 0.05, 0.1; six positions): β0 within 1.7e-4, the anomaly within 0.14% of its largest value at vsin⁡i/β = 0.57, 0.38% at 1, 0.93% at 2, 1.48% at 4. The first moment misses the fitted centre by 2.6%, 7.4%, 21% and 35% at those ratios. Exposure averaging (20 min, 7 sub-samples) is within 0.051 m/s of a 4001-point average where the exposure changes the anomaly by 4.05 m/s (node test/rmccf.test.mjs).
  • HD 80606 b from the bundled velocities (SOPHIE 2009 February 14, 48 points of which 28 in transit; HIRES 2009 June 5, 12 and 8), orbit of the stored joint fit with k = 0.1001 (Winn et al. 2009; the TESS curve is diluted by HD 80607), limb darkening of 390 to 690 nm: vsin⁡i = 1.47 +0.22 −0.17 km/s and λ = 58 +16 −12° with the orbit fixed, 1.49 +0.22 −0.17 km/s and 57 +17 −12° with P, T0, e, ω, Mp free under the joint posterior, 1.47 and 57° with the CCF form (β = 3 km/s) and 10-minute exposures. Hébrard et al. 2010 give λ = 42 ± 8° and vsin⁡i ≈ 1.7 km/s from the full transit of 2010 January, which is not in this file (1.1σ and 0.6σ away); Winn et al. 2009 give 50 +61 −36° from the HIRES night (node test/hd80606rm.test.mjs).
  • Whole curves from the orbit against ellc's flux-weighted velocities Maxted 2016 (test/make_rm_ellc_reference.py; a/R⋆ = 10, k = 0.1, vsin⁡i = 5 km/s, λ = 0°, 30°, −45°, 120° at i = 87° and 89.5°): at most 0.026 m/s of 23 to 39 m/s; the opposite sign of λ misses by at least 5.3 m/s. A fit to 70 simulated in-transit velocities (2 m/s) returns vsin⁡i = 4.97 ± 0.09 km/s (5) and λ = 30.0 ± 1.0° (30).

Assumptions and limits ​

  • By default the anomaly is the first moment of the uncovered light, which is what iodine-cell velocities approximate; the CCF form is an option for the whole fit (not per dataset), first order in the occulted fraction, with Gaussian local lines and a Gaussian macroturbulence (Hirano et al. 2011's radial-tangential profile is not implemented), and its β0 depends on the velocity window of the fit (here ±(vsin⁡i+6βp)) unless the measured one is given.
  • Solid-body rotation, no differential rotation or convective blueshift; the planet is dark (its own light and a moon are not in the anomaly); planets are added independently (no overlap of two planets on the disc).
  • With N-body dynamics the anomaly uses the Keplerian sky positions. Exposure averaging applies to the anomaly only.

References: Ohta, Taruya & Suto 2005; Maxted 2016; Hirano et al. 2010; Boué et al. 2013.

7. The likelihood ​

Code: src/core/model.js (buildProblem), src/core/priors.js, src/main.js (defaultPrior)

There is one likelihood, buildProblem. It is the sum of independent Gaussian terms for the light curve(s), the velocities and the event times, each with its normalisation, so that the evidence is a normalised marginal likelihood.

Light curves ​

A dataset holds one or more epochs s (nights, visits, orbits). The model for point i of epoch s is asmi with

mi=eκsXi∏pfp(ti)

where fp is planet p's (exposure-averaged) transit flux, Xi the airmass and κs the airmass coefficient of the epoch (absent without an airmass column). The normalisation as enters linearly and is integrated out analytically with a uniform prior as∼U(0,2) instead of being sampled:

ln⁡Ls=−12χs,min2+12ln⁡2π∑imi2/σi2−ln⁡2−12∑iln⁡(2πσi2),χs,min2=∑idi2σi2−(∑idimi/σi2)2∑imi2/σi2

The Gaussian integral over as is taken over the whole real line, which is exact to the extent that the conditional posterior of as lies well inside (0, 2); the reader divides fluxes by their median when it is more than 2% from 1 (and always for magnitudes). The conditional best value a^s=∑dimi/σi2/∑mi2/σi2 is the weighted least-squares amplitude and is what the plots show.

buildProblem also accepts linear regressors xk per point (a slope in time, a centroid): the baseline of epoch s becomes eκsX(as+∑kcskxk) and all p linear coefficients are integrated out together, with csk∼U(−1,1):

ln⁡Ls=−12χs,min2+12(pln⁡2π−ln⁡detA)−pln⁡2−12∑iln⁡(2πσi2),A=GTWG

where the columns of G are the model times (1,x1,…). A session stores the detector columns and the detrending choice (docs/FORMAT.md), and specFromSession rebuilds the regressors (or the kernel) from them exactly as the analyzer does, so a session re-run headlessly or from Python is fitted with its systematics model.

Light-curve jitter and dilution ​

An optional white-noise jitter js per epoch (key jitlc, or jitlc@2 for the second epoch; in ppm) is added in quadrature, σi2=ei2+js2, everywhere above: in the weights and in −12∑iln⁡(2πσi2), which is what stops a free jitter from growing without bound. For the information criteria χ2 is replaced by χ2+∑iln⁡(σi2/ei2), as for the velocity jitter.

Dilution (third light) by other stars in the aperture: with d the ratio of their flux to the target's (key dil, per epoch dil@2), the astrophysical model is replaced by

mi→mi+d1+d

before the airmass term, the regressors and the kernel, so it is still 1 out of transit and every depth is divided by 1+d. A contamination fraction D of the light in the aperture (TESS's CROWDSAP is 1−D) is d=D/(1−D). Without a measurement of d (photometry of the neighbour, a second band) it is degenerate with k and should be fixed or given a prior.

Velocities ​

For dataset d with residuals ri=vi−vmodel(ti) (model without γd), variances σi2+sd2, weights wi=1/(σi2+sd2) and γd∼U(−105,105) m/s integrated out:

ln⁡Ld=−12[∑iwiri2−(∑iwiri)2∑iwi]−12∑iln[2π(σi2+sd2)]+12ln⁡2π∑iwi−ln⁡(2×105)

The ln⁡(σ2+s2) term is what keeps a free jitter from growing without bound.

Event times ​

Mid-transit times of a planet are compared with T0+EP and mid-eclipse times with the superior conjunction plus EP, E being the nearest whole orbit; an optional timing noise σt is added in quadrature:

ln⁡Ltim=−12∑i[(ti−T0−EiP)2σi2+σt2+ln⁡2π(σi2+σt2)]

Prior support ​

A point is rejected (ln⁡L=−10100) when e∉[0,0.999), P≤0, (a/R⋆)(1−e)≤1, k≤0, Mp<0, a jitter is negative, the limb-darkening profile is not physical, or an optional model reports an impossible configuration. Priors are uniform or normal (truncated normal and log-uniform exist for programmatic use), applied as transforms of the unit cube.

Default priors in the analyzer ​

These are starting points computed from the current value v of each parameter (defaultPrior in src/main.js); every prior is shown and editable on the Fit tab and is written into the session and the exported script. Report the priors you used.

ParameterDefault prior
Rp/R⋆U(max(0.001,0.5v), min(0.6,1.6v))
iU(max(60∘,v−5∘), 90∘)
Pone light curve: N(v, max(10−6,10−5v)) d; several epochs: U about v wide enough to move the last transit by a quarter of T14 relative to the first
T0U(v±max(0.01, 0.2T14)) d
eU(0, min(0.95,max(0.3,v+0.2)))
ωU(v±30∘)
MpU(0, max(5v,0.02)) MJup
limb darkeningU(v±0.3) per coefficient; q1,q2∼U(0,1)
airmass coefficient κU(−0.2,0.2)
velocity trendc1∼U(−0.1,0.1) m/s/d; c2∼U(−10−3,10−3) m/s/d²
jitterU(0, max(15,5×median σ)) m/s
light-curve jitterU(0, max(500,3×median σ)) ppm
dilution dU(0, max(1,3v)); not ticked when added
vsin⁡i⋆, λU(0, max(20,3v)) km/s; U(−180∘,180∘)
M⋆, R⋆N(v,σ) with the catalogue uncertainty, or 5% of the value
Teff, [Fe/H]N(v,σ) with the catalogue uncertainty, or 100 K and 0.1 dex
planet's lightFp/F⋆∼U(0,max(0.05,3v)); contrast U(0,1); hot-spot offset U(−90∘,90∘); second harmonic U(−0.25,0.25); ellipsoidal U(0,500) ppm; beaming U(−50,50) ppm; map coefficients U(−0.6,0.6) for degree 1, U(−0.3,0.3) above
moonRm/R⋆∼U(0,max(0.05,2.5v)); period U(0.97v,1.03v); phase U(0,1); mass ratio U(0,0.1); tilt U(0∘,60∘)

A look-up in the catalogue suggests wider priors from the published uncertainties (suggestPriors: for example Rp/R⋆ uniform within ±max(10σ,0.5k), P normal with max(5σ,10−5P), T0 uniform within ±max(10σT2+(NσP)2, 0.5T14, 0.02d)). The Fit tab's "EXOTIC's bounds" preset sets Rp/R⋆∈[0,1.25v], i±5∘, T0 within a quarter orbit and the airmass coefficient in [−3,3].

Information criteria ​

BIC=χ2+kln⁡n Schwarz 1978, AIC=χ2+2k and χ2/(n−k) are reported at the best-fitting point, with k the number of sampled parameters plus the analytically integrated ones (one normalisation per epoch, one zero point per velocity dataset) and n the number of measurements of every kind.

Accuracy and checks ​

  • Several nights in one fit: ln L against quadrature over each normalisation: difference 9.5e-8 at ln L = 2468.49. (Validation: Evidence, ephemerides and geometry; test/multiepoch.test.mjs)
  • ln Z against midpoint quadrature, three free parameters: quadrature 1640.001; sampler mean of six seeds within its reported error. (Validation: Evidence, ephemerides and geometry; test/evidence.test.mjs)
  • Light-curve jitter and dilution: ln⁡L of a two-epoch light curve at four values of the jitters and d against the integral over as done numerically (scipy quad; test/make_lcnoise_reference.py): largest difference 1.2e-6; maximum-likelihood jitters 1470.71 and 617.61 ppm against scipy's 1470.70 and 617.57; injected 1500 and 600 ppm recovered as 1489 +107 −101 and 592 +110 −120; every point of a limb-darkened, exposure-averaged transit obeys depth(d) = depth(0)/(1+d) to 1.6e-16; a fit without the correction returns k/1+d (node test/lcnoise.test.mjs).
  • Detrending in sessions: the regressors specFromSession builds equal the centroids standardised by hand plus a Legendre P1(t), the likelihood equals the analyzer's import path exactly, and runSession recovers k = 0.1006 ± 0.0005 (0.1) where the same session without its systematics model gives 0.1124 ± 0.0003 (node test/sessiondetrend.test.mjs; the Python package gives the Node χ2 to 1e-12, python/tests/test_detrend.py).
  • ln⁡L against an independent Python implementation whose models come from exoplanet-core and radvel: largest difference 3.8e-6 over 24 points in four problems (node test/validation.test.mjs).
  • Three-planet toy system: χ² of each light curve at the injected values: 45.11, 88.80 and 190.58 in both codes. (Validation: EXOTIC’s own examples, reproduced; test/exotic.test.mjs)

Assumptions and limits ​

  • The noise model is white and Gaussian: the uncertainties as given, plus the optional light-curve jitter. There is no correlated-noise term in the likelihood and no outlier model. Underestimated error bars without a jitter give posteriors that are too narrow and exaggerate every evidence difference.
  • Detrending is a multiplicative eκX airmass term per epoch plus, when the light curve carries detector columns, linear regressors or a kernel model. Space-based systematics without such columns must be removed before the fit; the phase-curve examples use data detrended elsewhere.
  • All epochs share one T0 and P (no per-epoch mid-times in a joint fit), one exposure model per epoch and one limb-darkening band.
  • With several planets the light curve is the product of the planets' fluxes: planet-planet occultations during simultaneous transits are not modelled, and when planets have their own light the star's light is counted twice at the 1e-6 level.

References: Schwarz 1978; Zellem et al. 2020.

8. Nested sampling ​

Code: src/core/nested.js (MiniNest), src/core/fitRunner.js, src/core/stats.js

Posteriors and the evidence Z=∫L(θ)π(θ)dθ are computed by nested sampling Skilling 2006 with a sampler written for this project, MiniNest, which follows the strategy of UltraNest Buchner 2021. Parameters are sampled in the unit cube and mapped through the inverse cumulative distribution of each prior.

Algorithm ​

  • Live points. N=max(400,2d+10) by default for d free parameters, drawn uniformly in the unit cube.
  • Shrinkage. At each iteration the live point of lowest likelihood Li is retired with weight wi=Xi−1(1−e−1/N) and the prior volume shrinks as ln⁡Xi=ln⁡Xi−1−1/N; then Z=∑iLiwi plus the remaining live points, each with weight X/N. Live points with exactly equal likelihood (plateaus) are retired together with the volume rule of Fowlie, Handley & Su 2021.
  • Region. New points are drawn uniformly from the MLFriends region Buchner 2016; Buchner 2019: live points are whitened, t=A−1(u−u¯) with AAT the covariance of the live points about the mean of their own cluster; the region is the union of balls of radius r around the live points, intersected with an ellipsoid |t|2≤R2 and the unit cube. r is the largest distance from a left-out point to its nearest selected neighbour over 20 bootstrap rounds; the ellipsoid is enlarged by the largest ratio found when it is re-fitted to each bootstrap sample (UltraNest keeps the shape fixed, which biased ln⁡Z upward with dimension in our tests). Clusters are found by linking points closer than r. The region is rebuilt after every max(20,N/5) replacements and its volume is not allowed to grow.
  • Slice sampling. When drawing from the region costs more than three times a slice-sampled point (measured in likelihood calls), the sampler switches to slice sampling Neal 2003 from a random live point: max(10,2d) steps, each along either the difference of two live points of the same cluster or a random direction in the whitened space, with a bracket from the ellipsoid chord or from stepping out, and shrinkage on rejection. It switches back when the region costs less than 1.5 times a slice point.
  • Self-test. The rank at which each new point enters the live set must be uniform Buchner 2016; Buchner 2023. A Kolmogorov-Smirnov test of the ranks of region-drawn points is made after each rebuild (threshold p<10−5, at least 600 ranks) and at termination (p<10−3). On failure the region is no longer trusted: the sampler continues in slice mode and, if less than 30% of the budget is spent, restarts from fresh prior draws.
  • Termination. The run stops as converged when the live points can hold less than a fraction f of the evidence, LmaxX/(Z+LmaxX)<f, with f = 0.01 by default. It also stops at a budget of likelihood calls (default 10⁶) or seconds; such a run returns what it has with converged: false.

Uncertainty of the evidence ​

The volumes are random: each shrinkage factor is distributed as Beta(ni,1), ni being the number of live points at step i. ln⁡Z is recomputed for 200 simulated sequences Xi+1=XiU1/ni (fixed seed) and its standard deviation taken; terms for plateaus, for the scatter of the final live points, and, in an unconverged run, half the range the remaining live points could add, are added in quadrature. The information H and the effective sample size (∑w)2/∑w2 are reported with it.

Posterior summaries ​

The result is a weighted sample. Reported values are weighted medians with the 15.87% and 84.13% quantiles (no interpolation between sample points). Equal-weight draws for plots, derived quantities and stored posteriors are made by systematic resampling with a fixed seed.

Warnings ​

The sampler writes warnings in plain words: not converged (with the fraction of evidence still in the live points); fewer than 150 effective samples; sampling efficiency below 1%; a failed insertion-order test and whether the run restarted; more than half of the points from slice sampling; a posterior piled against an edge of its prior, or equal to its prior; separate peaks with their weights; pairs of parameters correlated above 0.95, with advice. They are stored in the session and belong in a paper's description of the fit when present.

Reproducibility ​

The random-number generator is sfc32 seeded from the session (seed 42 by default). For a given seed a run is the same every time on the same JavaScript engine; only the optional time limit depends on the clock.

What the evidence can and cannot be used for ​

ln⁡Z is a normalised marginal likelihood: the Gaussian normalisations are included, the priors are proper, and the analytically integrated parameters carry their prior widths. It can therefore be compared between models fitted to the same data with the same error bars, when both runs converged. It cannot be used: across different data or error bars; from an unconverged run (it is then a lower limit); between an N-body fit with the profiled straight line and a Keplerian fit; or without regard to the priors, since widening a prior on a constrained parameter lowers ln⁡Z by the logarithm of the factor. Its quoted uncertainty is the sampling error only.

Accuracy and checks ​

  • Correlated Gaussian, 10 parameters, ρ = 0.99: ln Z = −7.729 ± 0.335, exact −6.931 (2.4σ low on this seed; over 32 seeds the mean offset is −0.078 ± 0.054). (Validation: The nested sampler on hard problems; test/stress.test.mjs)
  • Transit with a/R★ and inclination both free (0.99 correlated): ln Z = 744.121 ± 0.196, quadrature 744.090; means within 0.025 sd. (Validation: The nested sampler on hard problems; test/stress.test.mjs)
  • Period aliases: three separate peaks: ln Z = 2.166 ± 0.154, exact 2.121; the peaks are reported with their weights. (Validation: The nested sampler on hard problems; test/stress.test.mjs)
  • Neal’s funnel, 4 parameters: ln Z = −16.872 ± 0.158, exact −16.588 (1.8σ low on this seed); over 48 seeds −0.045 ± 0.026 since the ellipsoid check, −0.24 ± 0.05 before. (Validation: The nested sampler on hard problems; test/stress.test.mjs)
  • Correlated Gaussian, 15 parameters, ρ = 0.99: ln Z = −10.245 ± 0.359, exact −10.397. (Validation: The nested sampler on hard problems; test/stress.test.mjs)
  • Banana, 5 parameters: ln Z = −17.033 ± 0.204, exact −16.732 (1.5σ); posterior widths off by 23% on this seed, beyond the 20% tolerance. (Validation: The nested sampler on hard problems; test/stress.test.mjs)
  • ln Z against midpoint quadrature, three free parameters: quadrature 1640.001; sampler mean of six seeds within its reported error. (Validation: Evidence, ephemerides and geometry; test/evidence.test.mjs)
  • Four simulated problems against dynesty (3000 live points), UltraNest and emcee on an independent Python likelihood (node test/validation.test.mjs): posterior means within 0.04σ and widths within 3.5%; ln⁡Z within 0.31 (1.5 combined σ) of dynesty and UltraNest; over eight seeds the mean ln⁡Z is within 0.09 of the reference and its scatter between seeds (0.07 to 0.21) is no larger than about the reported error (0.12 to 0.20). Validation: fits and their error bars.
  • Coverage, 300 simulated data sets each: the truth lies inside the 68% interval in 67.0 to 68.0% of transit fits and 69.0 to 73.3% of velocity fits (expected 68 ± 2.7%; the velocity amplitude and ω are 2σ high), and inside the 95% interval in 94.0 to 96.7% (same test). The data are simulated with this code's own model and white noise, so this tests the sampler and the likelihood, not the model.

Assumptions and limits ​

  • Funnel-shaped posteriors bias the evidence low by a few tenths, sometimes without a warning; the project's notes record 0.5 to 0.9 on some seeds.
  • In slice mode the scatter of ln⁡Z between runs exceeds the quoted error (rms pull up to 1.9 in the stress suite).
  • In the stored run of test/nested.test.mjs an 8-dimensional correlated Gaussian gives ln⁡Z = −23.357 ± 0.274 against the exact −23.966, that is 0.61 (2.2 reported σ) high; the 2- and 5-dimensional cases are within 1σ.
  • A mode much narrower than another can be lost. Twenty parameters with correlations of 0.999 do not converge within the default budget (reported as such).
  • A run stopped early has an effective sample size near 1; its intervals mean nothing, and the Results tab says so instead of showing them.
  • Not implemented: a variable number of live points, a separate metric per mode, warm starts, resuming from disk.
  • Coverage has been verified only on data simulated by this code's model with white noise. Correlated noise and model misspecification are untested.

References: Skilling 2006; Buchner 2016; Buchner 2019; Buchner 2021; Buchner 2023; Neal 2003; Fowlie, Handley & Su 2021; Speagle 2020; Foreman-Mackey et al. 2013.

9. Model comparison and its cautions ​

Code: src/main.js (readBayes, runAlarms), src/core/model.js (infoCriteria)

The Results tab keeps every run on the same data and compares them by Δln⁡Z, the logarithm of the Bayes factor Kass & Raftery 1995, against the run with the highest evidence. The thresholds 2.5 and 5 (odds of about 12 and 150 to 1) are the scale of Trotta 2008, an adaptation of Jeffreys' scale; the requirement that a difference also exceed 1 and twice its uncertainty is the software's own, deliberately conservative. With σΔ the quadrature sum of the two evidence uncertainties,

ConditionReading
either run not convergednot comparable
|Δln⁡Z|<max(1,2σΔ)indistinguishable
up to 2.5weakly disfavoured
2.5 to 5 (odds of about 12 to 150)moderately disfavoured
above 5strongly disfavoured

Runs on other data are hidden from the comparison. Each run also lists reasons not to trust its evidence: not converged; fewer than 200 effective samples; insertion-order test with p<0.01; χ2/N above 1.6 (error bars too small or model incomplete, which exaggerates every Δln⁡Z) or below 0.5; sampler warnings; priors on shared parameters that differ from the reference run (part of Δln⁡Z is then the change of prior volume); different kinds of data in the fit.

The Timing tab compares ephemeris models by BIC=χ2+kln⁡n Schwarz 1978 and describes a difference above 2, 6 and 10 as positive, strong and very strong, the scale of Kass & Raftery 1995 for twice the log Bayes factor.

Accuracy and checks ​

  • test/evidence.test.mjs checks that a transit model is preferred over no transit on simulated data (Δln⁡Z = 6164.6 in the stored output); there is no test of model comparison between two close models.

Assumptions and limits ​

  • The evidence depends on the prior widths chosen. A model comparison in a paper should state the priors of both models and, ideally, show that the conclusion survives a reasonable change of them.
  • The uncertainty of ln⁡Z is sampling error. Systematic bias of the sampler (see Nested sampling) is of the same size as the "indistinguishable" threshold; differences below about 1 should not be interpreted.
  • The BIC is an approximation that ignores the priors. On signal-free timing lists the BIC alone prefers a sinusoid in 58% of cases (project notes), which is why the Timing tab also requires a significant periodogram peak.
  • Both criteria assume the error bars are right. They are not rescaled.

References: Kass & Raftery 1995; Trotta 2008; Schwarz 1978.

10. Ephemerides, timing, periodograms and orbital decay ​

Code: src/core/ephemeris.js, src/core/periodogram.js, src/ui/timingTab.js

Ephemerides ​

Mid-transit times ti±σi at whole-orbit epochs Ei are fitted by weighted linear least squares with

T(E)=T0+PEorT(E)=T0+PE+12dPdEE2,dPdt=1PdPdE

The fit is done in a centred, scaled epoch variable with a time offset removed, then mapped back to epoch 0, so Julian dates keep their precision. The covariance is (GTWG)−1 from the error bars as given; it is not rescaled by χ2, and the tab says so when χ2/dof>2. The uncertainty of a predicted time is σT2(E)=gTCg with g=(1,E,E2) in the scaled variable, and σ(dP/dt) includes the covariance with P. When a list has no epoch column, Ei=round[(ti−tref)/P] with the system period; a time more than 0.2 periods from a whole orbit raises a warning. The quadratic term is the only decay model: it describes orbital decay, or apsidal precession to first order; there is no separate precession model. Linear and quadratic fits are compared by BIC.

Periodic signals in the residuals ​

The residuals from the ephemeris are searched with the generalised Lomb-Scargle periodogram Zechmeister & Kürster 2009 (weights 1/σi2, floating mean), in the normalisation p(f)=1−χ2(f)/χ02, where χ2(f) is that of the best sinusoid plus constant at frequency f. Epochs are a lattice, so the shortest period searched is two lattice steps (a true Nyquist limit) and the longest twice the time span, on a grid oversampled 10 times. The false-alarm probability of the highest peak is the analytic bound of Baluev 2008, FAP≈1−(1−FAPsingle)e−τ(z), as implemented in astropy VanderPlas 2018; a bootstrap over the residuals (1000 resamples) is available on request. A sinusoid is then fitted to the times themselves together with the ephemeris (linear least squares with a Gauss-Newton refinement of the period), the periodogram of what is left is computed, and a second sinusoid is tried. A sinusoid is recommended only if its peak has FAP below 1% and it lowers the BIC.

Accuracy and checks ​

Assumptions and limits ​

  • No published list of mid-transit times has been fitted end to end and compared with its paper; only the conversion from dP/dE to dP/dt was (WASP-12 b, Leonardi et al. 2024).
  • The analytic false-alarm probability trusts the error bars: with errors underestimated by 45 s of unmodelled scatter it raises false alarms in 68% of signal-free lists, the bootstrap in 1% (stored test output). Use the bootstrap when χ2/dof of the ephemeris is well above 1.
  • Times are taken as given, with no conversion between time standards; the calendar column of predicted transits reads the listed time as UTC.
  • A periodic signal in transit times has many physical explanations; the periodogram gives a period and amplitude, not a companion. See N-body dynamics.

References: Zechmeister & Kürster 2009; Baluev 2008; VanderPlas 2018; Astropy Collaboration 2022; Leonardi et al. 2024; Schwarz 1978.

11. Several planets and Kepler-tied a/R★ ​

Code: src/core/model.js, src/core/rv.js (arsFromKepler)

A fit can hold several planets. The first planet's parameters have plain keys (per, rprs); the others carry their letter (per@c). Limb darkening belongs to the star and is shared. The light curve is the product of the planets' fluxes, the velocity the sum of their reflex terms, and each planet's event times are compared with its own ephemeris.

a/R⋆ is never a sampled parameter. With the inclination free the two trade off through the impact parameter b=(a/R⋆)cos⁡i: a transit of fixed duration and depth can be reproduced along a ridge in the (a/R⋆, i) plane, which costs the sampler several times more likelihood calls and inflates the uncertainty of i without adding information that the stellar density does not already give. a/R⋆ is therefore either held at its input value, or, with "a/R★ from Kepler's law", computed for every sample from

aR⋆=1R⋆[G(M⋆+Mp)P24π2]1/3

so that the photometry, the velocities and the timing describe one orbit. M⋆ and R⋆ can then be free parameters with normal priors at their measured uncertainties, which propagates the stellar density uncertainty into a/R⋆, b and the durations.

Accuracy and checks ​

  • Transit with a/R★ and inclination both free (0.99 correlated): ln Z = 744.121 ± 0.196, quadrature 744.090; means within 0.025 sd. (Validation: The nested sampler on hard problems; test/stress.test.mjs)
  • Three-planet toy system: χ² of each light curve at the injected values: 45.11, 88.80 and 190.58 in both codes. (Validation: EXOTIC’s own examples, reproduced; test/exotic.test.mjs)
  • HD 80606 b, light curve + velocities + timing, against Table 7 of Pearson et al. 2022: P within 5e-4 d, i within 0.03°, e within 0.001, K within 3 m/s. (Validation: EXOTIC’s own examples, reproduced; test/exotic.test.mjs)
  • Project notes (README, "The sampler"): a transit fit with both a/R⋆ and i free needed 47,000 likelihood calls and gave σ(i) = 0.84°, against 11,600 calls and 0.076° with a/R⋆ fixed. These two numbers are from a development run and are not asserted by a test.

Assumptions and limits ​

  • Holding a/R⋆ fixed without freeing the star understates the uncertainty of i, b and the durations: the quoted intervals are conditional on that value. For publication, use Kepler's law with M⋆ and R⋆ free under their measured uncertainties, or state the value a/R⋆ was fixed at and its source.
  • The stellar density implied by a transit is not fitted independently of M⋆ and R⋆; an inconsistent stellar density (blend, wrong star) shows only as a poor fit.
  • Seventeen free parameters for three planets converge in about 950,000 likelihood calls, close to the default budget.

References: Winn 2010; Seager & Mallén-Ornelas 2003.

12. Physical priors: stable orbits and period aliases ​

Code: src/core/physical.js, src/core/aliases.js, src/core/model.js (spec.physical, nalias), src/core/priors.js (categorical)

Optional prior constraints, each off unless spec.physical switches it on, applied in the likelihood before any model is computed (a point that fails has zero prior mass). Semi-major axes come from Kepler's law with each planet's period and mass, masses from mplanet or from K. crossing: the radial ranges of adjacent orbits do not overlap, a2(1−e2)>a1(1+e1). hill: Gladman's Hill stability of each adjacent pair on coplanar eccentric orbits Gladman 1993,

μ−3(μ1+μ2ℓ2)(μ1γ1+μ2γ2ℓ)2>1+34/3μ1μ2μ4/3,ℓ=a2/a1, γi=1−ei2

which for circular orbits is a separation of 23 mutual Hill radii RH=[(m1+m2)/3M⋆]1/3(a1+a2)/2. hillSpacing: every adjacent pair at least that many RH apart (about 10 for three or more planets to survive 108 orbits Chambers et al. 1996). amd: AMD stability Laskar & Petit 2017 with coplanar orbits, C/Λout<Cc(α,γ) for each pair and C/Λ1<1. roche: periastron outside the star and outside fRp(M⋆/Mp)1/3, f = 2.16 (Rappaport et al. 2013). megno: a short integration of the variational equations Cincotta & Simó 2000 for each candidate (expensive).

A truncated prior changes the evidence. The sampler integrates Lπ over the allowed region, so its Z is that of "this prior and a viable system"; the evidence of the renormalised prior π/V is Z/V, with V the prior mass that passes, estimated from 20,000 prior draws and stored by runSession as result.prior_volume (ln⁡V, and ln⁡Z−ln⁡V). Use ln⁡Z to compare viable models with each other; ln⁡Z−ln⁡V always gains about −ln⁡V when the data already sit in the allowed region.

Period aliases ​

A planet seen in transits a span S apart has P=S/N. aliasScan predicts every transit of each N and drops an N that puts a transit inside covered data (more than half of its T14 window observed) more than 0.1 d from every one seen, the rule of the owner's EXOTIC analysis of TOI-1873. The survivors go to the sampler as one categorical parameter nalias with palias = S, P = palias/nalias, uniform or geometric (∝N2/3, the transit probability) weights; or one run per N, compared afterwards.

Accuracy and checks ​

  • Gladman's criterion against a numpy transcription on 200 random eccentric pairs: 2.2e-16; its circular limit is 3.4640, 3.4637, 3.4624, 3.4563, 3.4283 mutual Hill radii at μ = 1e-7 to 1e-3, tending to 23 = 3.4641. AMD, critical AMD and β against numpy/brentq on 100 systems of two to four planets: 1.9e-15. Two Jupiters around a Sun 3.0 RH apart are rejected and 4.0 accepted (test/make_viability_reference.py, node test/viability.test.mjs).
  • With a flat likelihood the sampler's ln⁡Z with the crossing test is −2.197 against the analytic ln⁡V = −2.282 (expected scatter 0.15). Two planets in 30 simulated velocities (25 m/s, e = 0.05 and 0.10): ln⁡Z = −161.69 ± 0.22 without tests, −161.45 (crossing, ln⁡V = −1.44), −162.00 (hill, −2.54), −161.99 (amd, −1.82): the data sit in the allowed region; a spacing of 8 RH, which excludes the true pair (6.4 RH), costs 49 in ln⁡Z against ln⁡V = −3.32.
  • TOI-1873 d (three transits 900.191 and 900.186 d apart, TESS coverage of 16 sectors read from the owner's time series, test/make_alias_reference.py): allowed N = 1, 2, 3, 4, 7, as in the owner's analysis, with the covered and missed transits of every N up to 25 equal to EXOTIC's stored scan. A simulated planet with three transits 900 d apart (truth N = 3): the categorical run gives p(N=3) = 0.895 and p(N=2) = 0.105 against 0.882 ± 0.028 and 0.117 ± 0.028 from one run per N, and its ln⁡Z (3295.99 ± 0.18) equals the mean of theirs (3295.76 ± 0.16) (node test/aliases.test.mjs).

Assumptions and limits ​

  • Mutual inclinations are not known from transits and velocities, so Hill and AMD tests assume coplanar orbits. Neither criterion knows about resonances, which can protect crossing orbits (and the AMD test ignores the mean-motion-resonance overlap correction of later work).
  • Masses from K are minimum masses when the inclination is unknown. A planet without a mass counts as massless (only crossing is tested for it).
  • The MEGNO screen integrates a few hundred orbits per likelihood call; it is for a final check of a narrow posterior, not for a broad search.

References: Gladman 1993; Chambers et al. 1996; Laskar & Petit 2017; Cincotta & Simó 2000.

13. Phase curves and eclipse maps ​

Code: src/core/phasecurve.js

When a planet has the parameter Fp/F⋆ (fpfs), its own light is added to the model. The flux relative to the star alone is

F(t)=S(t)T(t)+Fp(t)V(t)

with T the transit, S optional stellar terms, Fp the planet's flux and V the visible fraction of the planet's disc. Fp/F⋆ is the dayside surface brightness relative to the star's, so the eclipse depth is D=(Fp/F⋆)k2. The orbital phase is ϕ=ω+ν+90∘, zero at mid-eclipse and 180∘ at mid-transit for any eccentricity.

Harmonic phase curve ​

Fp(ϕ)=D(2−C)+Ccos⁡(ϕ+ϕh)2−C+Ccos⁡ϕh+D[h2c(cos⁡2ϕ−1)+h2ssin⁡2ϕ]

C∈[0,1] is the contrast, (maximum − minimum)/maximum of the first-order curve; ϕh is the hot-spot offset in degrees, positive east, so that the curve peaks ϕh before mid-eclipse; h2c, h2s are second-harmonic amplitudes as fractions of the depth. By construction Fp(0)=D and Fp≥0. The flux is floored at zero.

Eclipse ​

In this mode the planet is a uniform disc and V=1−Alens/(πk2), with Alens the area common to the stellar and planetary discs (closed form), when the planet is behind the star. The planet's own disc is occulted, with its true radius.

Stellar terms ​

S=1−Aellip(ar)3cos⁡2ϕ−Abeam[cos⁡(ω+ν)+ecos⁡ω]

with the ellipsoidal and beaming amplitudes in ppm; both default to zero.

Light-travel time ​

Optionally (it needs R⋆) the model is evaluated at the retarded time t′=t−(R⋆/c)[rtrsin⁡i−Z(t)], one fixed-point step, which is zero at mid-transit and delays mid-eclipse by (R⋆/c)(rtr+recl)sin⁡i, that is 2a/c for a circular orbit.

Maps ​

In map mode the planet's surface brightness is I(θ,λ)=I0[1+∑l,mylmYlm(θ,λ)] in real spherical harmonics up to degree 3, normalised as in starry Luger et al. 2019 with Y00=1, in a frame whose z axis points to the star (longitude measured east from the substellar point) and whose y axis is the orbit normal; rotation is synchronous. Out of eclipse the disc-integrated flux is analytic, Fp=Dn[1+23v⋅o+14oTQo], with o the direction to the observer and v, Q the degree-1 and degree-2 parts of the map; degree 3 contributes nothing out of eclipse. Dn is set so that the unocculted flux at ϕ=0 equals D. During eclipse ingress and egress the flux of the visible part of the disc is obtained from ten moments 1π∫xaybzcdA, each a Green's-theorem line integral: closed form along the planet's limb and 12 Gauss-Legendre nodes along the stellar limb. A fit rejects maps whose disc-integrated flux is negative; rejecting maps with negative surface brightness is an option, off by default.

Positive planets ​

Always: the planet's unocculted flux must be non-negative at every phase (so the nightside too); the smallest value over the orbit is found from 72 phases and a golden-section search. With phase.positive_map the surface brightness must also be non-negative everywhere on the sphere, zero prior mass otherwise. Its minimum is exact for degree 1 (1−|v|) and degree 2 (the minimum of 1+v⋅b+bTQb on |b|=1, a trust-region problem: the multiplier is the root of a secular equation in the eigenbasis of Q, plus the "hard case"), and from a 5° grid and a pattern search for degree 3. In harmonic mode the option tests the degree-1 map of the first-order curve, which is stricter than physics requires: a contrast above 0.8 needs degree 2 or more, not a negative map. phaseSummary(requirePositive) gives no temperatures, albedo or hot spot for a model that is not physical, so posterior summaries can be taken over physical draws only.

Temperatures ​

A brightness temperature is found by solving ∫Bλ(Tb)dλ=(Fp/F⋆)∫Bλ(Teff)dλ over the band by bisection, the star being a blackbody at Teff. Day and night temperatures are those of the unocculted flux at ϕ=0 and 180∘. Bond albedo and heat redistribution follow Cowan & Agol 2011: T0=Teff/a/R⋆, AB=1−12(3x+5y), ε=8y/(3x+5y) with x=(Td/T0)4, y=(Tn/T0)4.

Accuracy and checks ​

  • Map basis curves to degree 3, through ingress and egress: max 4.6e-13 over 2256 values in 5 systems. (Validation: Phase curves and eclipse mapping; test/phasecurve.test.mjs)
  • Uniform-disc eclipse against batman 2.5.1: 4.9e-13 for circular orbits, 2.5e-7 at e = 0.93. (Validation: Phase curves and eclipse mapping; test/phasecurve.test.mjs)
  • Map curves against starry 1.2.0: max 5.6e-7 of the planet’s flux over 48 curves. (Validation: Phase curves and eclipse mapping; test/phasecurve.test.mjs)
  • Harmonic phase curve against EXOTIC: 5.0e-13 outside transit and eclipse. (Validation: Phase curves and eclipse mapping; test/phasecurve.test.mjs)
  • Brightness temperature against scipy quadrature and root finding: relative 1.6e-9 over 48 cases. (Validation: Phase curves and eclipse mapping; test/phasecurve.test.mjs)
  • The same model inside the fitting likelihood: χ² identical to the model used directly. (Validation: Phase curves and eclipse mapping; test/phasemodel.test.mjs)
  • Smallest brightness of 120 random maps of degree 1 to 3 (37 negative somewhere) against a 0.25° grid polished by Nelder-Mead in numpy (test/make_mappos_reference.py): 2.2e-16 (degree 1), 1.6e-15 (degree 2), 2.4e-12 (degree 3), none classified differently. The smallest flux over the orbit is within 5.6e-11 of the depth of 100,000 phases (node test/phasepositive.test.mjs).
  • WASP-43 b eclipse map refitted with positive_map (PHASEPOS_FULL=1): degree 1 holds y10 at the edge, 0.568 +0.005 −0.008 (stored, unconstrained: 0.703 ± 0.037, a map negative on 8.4% of the sphere), y1,−1 is bounded at −0.03 ± 0.06 (stored: 0.11 +0.22 −0.24, its prior), and ln⁡Z falls by 11.6 (ln⁡V = −0.63 of prior mass removed). Degree 2 positive keeps y10 = 0.705 ± 0.034 with y20 = 0.17 ± 0.05, y1,−1 = 0.00 ± 0.06, and ln⁡Z 1.2 below the unconstrained degree-1 fit (ln⁡V = −1.22).

Assumptions and limits ​

  • The examples fit Spitzer light curves that were already detrended elsewhere, jointly with that pipeline's own phase-curve model; instrument systematics are not modelled here, and those light curves are not independent of a degree-1 map.
  • The north-south coefficient y1,−1 is unconstrained in both map examples and runs into its prior unless the map is required to be positive, which bounds it (above).
  • With several planets the fluxes are multiplied, which counts the star's light twice at the 1e-6 level.
  • The star is a blackbody for brightness temperatures. Reflected light and thermal emission are not distinguished. A contrast above 0.8 corresponds to a degree-1 map that is negative near the antistellar point.
  • EXOTIC's phase-curve model differs on purpose: its eclipse ingress is that of a body of radius kFp/F⋆ and starts late (up to 29% of the depth during ingress).

References: Luger et al. 2019; Cowan & Agol 2011; Stevenson et al. 2017; Kreidberg 2015; Zellem et al. 2020.

14. Exomoons ​

Code: src/core/exomoon.js

A planet has a moon when it has the parameter Rm/R⋆ (rmoon). Planet and moon orbit their common barycentre, which follows the planet's Keplerian orbit; the moon's orbit is circular. Parameters: Rm/R⋆; the moon's period Pm; its phase at the planet's mid-transit (0 = between planet and observer); the mass ratio q=Mm/Mp (default 0, in which case the planet does not move about the barycentre); and one tilt angle about the transit chord (0 = coplanar and prograde). The separation follows from Kepler's law,

amR⋆=aR⋆(PmP)2/3(μ1+μ)1/3,μ=Mp(1+q)M⋆

and the positions are barycentre −q1+qd for the planet and barycentre +11+qd for the moon, d being the moon-planet vector. The blocked flux is that of the union of the two discs: the sum of the two single-disc terms of the transit model minus the flux behind the region common to planet, moon and star, a Green's-theorem sum over arcs of the three circles (20 Gauss-Legendre nodes per half-arc). A moon wholly behind or in front of the planet's disc adds nothing.

Stability is reported, not enforced: the Hill radius rH=(a/R⋆)(1−e)(μ/3)1/3, the limits of Domingos et al. 2006 (am<0.4895rH(1−1.0305e) prograde, 0.9309rH(1−1.0764e) retrograde) and a Roche limit. The likelihood only requires that the two bodies do not overlap.

"Search for a moon" is a matched filter on the residuals of the planet-only fit: templates of a reference moon on a grid of period (uniform in frequency) and phase, with the planet's own depth and mid-time projected out; the statistic is the signal-to-noise ratio ∑wrm/∑wm2. The false-alarm probability is the fraction of searches of white-noise realisations, drawn from the error bars, that reach the same value.

Timing and duration variations are compared with the amplitudes of Kipping 2009: TTV=aw/vB and TDV=T¯(2πaw/Pm)/vB with aw=amq/(1+q).

Accuracy and checks ​

  • Two overlapping discs on the star: max 6.5e-14 over 205 cases and 5 limb-darkening profiles. (Validation: Exomoons; test/exomoon.test.mjs)
  • exomoon2.py, step by step with this kernel: planet + moon within 1.3e-8. (Validation: Exomoons; test/exomoon.test.mjs)
  • Pandora, whole light curves: max difference 5.2e-6. (Validation: Exomoons; test/exomoon.test.mjs)
  • Timing and duration variations against Kipping’s formulas: within 2.4% and 0.5% of the amplitude. (Validation: Exomoons; test/exomoon.test.mjs)
  • Search for an injected moon: found at S/N 26; 100 noise-only searches reach 5.0. (Validation: Exomoons; test/exomoon.test.mjs)
  • Moonless light curves: 1 of 40 above the 95% level of noise (2 expected). (Validation: Exomoons; test/exomoon.test.mjs)

Assumptions and limits ​

  • The moon's orbit is circular; the star's tide on the moon is ignored (two bodies about a Keplerian barycentre); one orientation angle only. The moon's own light is not modelled.
  • No real detection has been reproduced: there is none. The example is simulated.
  • The configuration of the reference script this was ported from is unbound (2.4 Hill radii), which the page flags.
  • The search's false-alarm rate assumes white noise with the stated error bars. The moon is not drawn in the three-dimensional view.

References: Kipping 2009; Domingos et al. 2006; Heller 2014; Hippke & Heller 2022.

15. N-body dynamics ​

Code: src/core/nbody.js, src/core/model.js

With the N-body option, mid-transit times and stellar velocities come from an integration of the star and planets instead of independent Keplerian orbits. The planets' P, T0, e, ω, i (and an optional node angle) are read as osculating elements at a stated epoch, in Jacobi coordinates with planets ordered by period, in the frame of the conventions. Units are au, days and solar masses. With all planet masses zero the result is the Keplerian model.

  • Wisdom-Holman map (default): the splitting of Wisdom & Holman 1991 in Jacobi coordinates as in WHFast Rein & Tamayo 2015, drift-kick-drift, with the Kepler drift solved in universal variables and symplectic correctors Wisdom 2006 of order 5 by default. The step is fixed at minp[Pp(1−ep)3/2/1+ep] divided by 30 (20 in the example and the companion search).
  • IAS15: the 15th-order Gauss-Radau integrator of Rein & Spiegel 2015 (after Everhart 1985) with adaptive steps, precision parameter 10−9. It is used only when selected; a close encounter under the Wisdom-Holman map (separation below one mutual Hill radius) is reported with advice to rerun, not handled automatically.
  • Transit times: the root of the star-centred sky-plane condition, by default the conjunction Xcos⁡Ω+Ysin⁡Ω=0 in the likelihood (or the minimum separation Rsky⋅Vsky=0 as in TTVFast Deck et al. 2014), with the planet in front; refined by Newton iteration on a two-body arc.
  • Velocity: −Z˙⋆ of the star in the barycentric frame, positive when receding.
  • In the likelihood: observed mid-transit times are matched to the nearest simulated transit. With nbody.linear a weighted straight line in epoch is fitted to each planet's residuals inside the likelihood, so the transiting planet's own T0 and P need not be sampled.

Stability is reported from the mutual Hill separation (two planets are Hill-stable beyond 23 mutual Hill radii, Gladman 1993; with three or more planets a separation below 10 is flagged, Chambers et al. 1996), orbit crossing, the AMD criterion of Laskar & Petit 2017 and, on request, a short integration with a shadow particle. "Search for an unseen companion" scores a grid of perturber mass and period (eight starting phases each) by the χ2 of the observed times against the integration with a straight line profiled out, then polishes the best cell by Nelder-Mead. The analytic amplitudes of Lithwick, Xie & Wu 2012 and Agol et al. 2005 are available for comparison.

Accuracy and checks ​

  • Transit times against REBOUND 5.2.2 (IAS15), EXOTIC’s example system: 1.3e-6 s over 610 transits. (Validation: N-body dynamics; test/nbody.test.mjs)
  • Same map as REBOUND’s WHFast: positions within 2.0e-12 AU after 8000 steps. (Validation: N-body dynamics; test/nbody.test.mjs)
  • One planet: identical to the Keplerian model: timing 3.4e-6 s over 400 orbits, velocity 1.6e-9 m/s. (Validation: N-body dynamics; test/nbody.test.mjs)
  • Critical angular-momentum deficit against a brute-force minimum: relative 3.8e-12 over 25 cases. (Validation: N-body dynamics; test/nbody.test.mjs)
  • Unseen companion recovered from 72 transit times: 0.1040 M Jup (injected 0.1), mean period 5.3336 d (5.3322). (Validation: N-body dynamics; test/nbody.test.mjs)
  • The same dynamics inside the fitting likelihood: χ² 82.9 against 82.9 from the other integrator directly. (Validation: N-body dynamics; test/nbodymodel.test.mjs)

Assumptions and limits ​

  • Light curves and mid-eclipse times from the integration are described in the next section (Photodynamics). Light-travel time is optional (nbody.ltt).
  • A fit that samples the transiting planet's own period and T0 as osculating elements did not converge in 150,000 likelihood calls, which is why the straight line is profiled by default; mean-ephemeris elements (next section) make that fit converge at about six integrations per call. A profiled parameter is not marginalised: the evidence of such a fit is not comparable with that of a Keplerian fit, and the uncertainties are conditional on the best line.
  • The χ2 surface of a perturber has period aliases every few thousandths of a day. The search returns one solution that fits, not the only one.
  • TTVFast, given the same Jacobi elements, differs by hundreds of seconds because it defines them differently (stored test output); compare codes through state vectors or transit times, not elements.

References: Wisdom & Holman 1991; Wisdom 2006; Rein & Tamayo 2015; Rein & Spiegel 2015; Everhart 1985; Rein & Liu 2012; Deck et al. 2014; Lithwick, Xie & Wu 2012; Agol et al. 2005; Holman & Murray 2005; Gladman 1993; Chambers et al. 1996; Laskar & Petit 2017; Pearson 2019.

16. Photodynamics ​

Code: src/core/nbody.js (dynamicalModel, simulate, osculatingFromMean, megno), src/core/model.js (spec.nbody)

With spec.nbody set and a light curve in the fit, the light curve is computed from the same integration as the transit times and velocities. The integrator writes every body's position and velocity at evenly spaced nodes (150 per orbit of the innermost planet by default) only over the spans covered by data; each planet's sky-plane position relative to the star at a sample time is interpolated by cubic Hermite polynomials from the positions and velocities at the neighbouring nodes, and the separation is passed to the transit kernel of the transit-model section, through the same exposure averaging as Keplerian curves.

Light-travel time (nbody.ltt): each body is taken at its retarded time, with the barycentre's sky plane as reference, so a transit is seen about a/c early and an eclipse about a/c late.

Mid-eclipse times in the timing likelihood come from the same integration (roots of the superior-conjunction condition).

Mean-ephemeris elements (nbody.elements: "mean"): a planet's sampled period and T0 are those of the straight line through its simulated transits over the span of the data; the osculating elements at the reference epoch are found by fixed-point iteration (each step one integration). This removes the ridge between a companion's mass and the transiting planet's osculating period that defeats sampling in osculating elements.

MEGNO Cincotta & Simó 2000 is integrated alongside a short run (variational equations, Gauss–Radau quadrature on the IAS15 nodes): about 2 for a regular orbit, growing linearly for a chaotic one; it enters the stability findings.

Accuracy and checks ​

  • node test/nbodyphoto.test.mjs (references from REBOUND 5.2.2 and batman, test/make_nbodyphoto_reference.py): with all masses zero the photodynamical curve equals the Keplerian one to 1.0e-9 of the flux (3.6e-14 at 600 nodes per orbit); a strongly perturbed pair (1 and 2 MJ at 10 and 20.3 d, 2819 points over 125 d, where the Keplerian curve is wrong by 12,049 ppm) agrees with REBOUND IAS15 positions through batman's kernel to 0.008 ppm with IAS15, 57 ppm with the Wisdom–Holman map at its defaults and 14 ppm at 60 steps per orbit; 30-minute exposures 0.004 ppm; light-travel time 0.02 ppm against REBOUND at retarded times. Mid-eclipse times: 3.8e-8 s from REBOUND for 18 eclipses. MEGNO 2.0004, 1.9955, 1.9987 on the example system after 2000 orbits (REBOUND 2.0008, 2.0020).
  • Mean elements: periods recovered to 1.6e-11 d after a round trip; a five-parameter fit of the timing example (both planets' ephemerides and the companion's mass) converged in 19,032 calls, companion 0.1004±0.0063MJ (injected 0.100), with NBODY_FIT=1.

Assumptions and limits ​

  • Planets are treated one at a time in the light curve: the area two planets share on the disc during a mutual event is counted twice. A planet with its own light or a moon keeps a Keplerian curve.
  • The Wisdom–Holman map's timing error grows with time and with the square of the masses (7 s over 125 d for the pair above); pairs of giant planets with precise photometry need IAS15 or more steps per orbit.
  • The straight-line profile (nbody.linear) corrects times only and is refused together with a light curve. Mean elements cost about six integrations per call and are not found inside a mean-motion resonance (the point is rejected).
  • The value of MEGNO for a chaotic orbit is not reproducible between codes, only "about 2" against "much more"; the Lyapunov time from a run stopped at MEGNO = 20 is an order of magnitude.

References: Rein & Spiegel 2015; Rein & Tamayo 2015; Wisdom & Holman 1991; Cincotta & Simó 2000; Kreidberg 2015.

17. Time-series spectroscopy ​

Code: src/core/spectroscopy.js, src/core/missionbands.js, src/core/model.js (data.regressors), src/spectra/

Input is extracted one-dimensional spectra over time: JWST x1dints.fits files (both pipeline layouts, one row per integration or one extension per integration, segments joined in time and detectors in wavelength) or a table of flux against time and wavelength (CSV, ECSV). There is no spectral extraction from images.

Cleaning and binning ​

A point more than 5 robust standard deviations from an 11-point running median in time is replaced. A column is dropped when more than 20% of its points are bad, when its median flux is not positive or differs by more than 50% from its neighbours', or when its error bars are five times the typical ones. Channels are sums over columns (constant resolving power, constant width, N channels, or a fixed number of columns), with errors added in quadrature, normalised by their median. Error bars are widened to the point-to-point scatter when that is larger (the default; pipeline errors or scatter alone can be chosen).

Baseline ​

Every light curve, white or spectroscopic, goes through the one likelihood (buildProblem) with optional regressors xk per epoch (scaled time, its square, step functions at given times):

F(t)=m(t)ea2X(t)(a1+∑kckxk(t))

with a1 and the ck integrated out analytically together (a Gaussian integral with the determinant of the weighted normal matrix), each ck under a flat prior of width 2.

White light, then channels ​

The white light curve is fitted by nested sampling for k, T0 and i; a/R⋆ is fixed or follows from Kepler's law with R⋆ free under a normal prior. Each channel is then fitted for its depth (k, or Fp/F⋆ for an eclipse) with the geometry fixed at the white-light solution and limb-darkening coefficients from the PHOENIX table Husser et al. 2013 for the channel's wavelength range and the star. With one free parameter the posterior is integrated on a 241-point grid of the same likelihood. Limb darkening with a normal prior, or the geometry with normal priors, sends the channel fit to the nested sampler instead.

Products ​

The spectrum with 16th–84th percentile errors; per-channel residual rms against the error bars; the binned-rms test of the noise section; the depth in atmospheric scale heights H=kBTeq/(μmHg) for interpretation (with μ chosen by the user, 2.3 by default), the change of depth per scale height being 2kH/R⋆.

Simulator ​

Simulated observations use photon noise from a black-body star seen through a telescope of collecting area A (Ariel 0.64 m² Tinetti et al. 2018, JWST 25.4 m²) and a throughput τ, plus a noise floor f:

Ne=AτΔtπ(R⋆d)2∫2cλ4dλehc/λkBT−1,σ2=(1+x)2Ne+f2T14Δt

τ (0.3) and f (20 ppm Ariel, 10 ppm JWST) are editable inputs and are not taken from mission documents. This is a simple radiometric estimate, not ArielRad or PandExo.

Accuracy and checks ​

  • node test/spectroscopy.test.mjs (44 checks): readers against astropy, exact (real G395H files to 2.4e-15); the marginal likelihood with regressors against a scipy.integrate.tplquad integral to 9.4e-5 in ln⁡L, coefficients against numpy.linalg.solve to 1.4e-9; photon rate against scipy.integrate.quad to 8.4e-9; Ariel-to-JWST noise ratio 6.300 =25.4/0.64.
  • Injection and recovery: pulls (recovered − injected)/σ of mean 0.084 and standard deviation 0.980 over 228 simulated Ariel channels (70% within 1σ, 96% within 2σ); 0.296 and 0.981 over 36 channels made with batman Kreidberg 2015 with the geometry at the truth, 0.471 and 0.978 with the geometry from the white-light fit.
  • Real data, WASP-39 b with NIRSpec G395H (JWST ERS program 1366, standard-pipeline x1dints): the mirror-tilt step is found automatically (−3275 ppm, 75σ); white-light depth 21324 ± 13 ppm, residual rms 136 ppm. At R=100 (56 channels) the spectrum follows the published one Alderson et al. 2023 with correlation 0.974 and a scatter of the differences of 91 ppm (1.10 combined σ), but lies 202 ppm lower on average. The offset is not explained (geometry, fixed limb darkening and the standard-pipeline extraction against the paper's custom reductions are candidates; none was isolated). This test needs 207 MB of public files and is skipped without them.

Assumptions and limits ​

  • Channel error bars do not include the uncertainty of the shared geometry; a change of the white-light inclination moves all channels together.
  • All bands are top-hats; no measured throughput curves are used. The JWST and Ariel wavelength limits are those of the mission documentation and should be checked against the instrument's current values before quoting.
  • Limb darkening beyond 5.5 µm (MIRI LRS, the red end of Ariel AIRS-Ch1) lies outside the PHOENIX table and is extrapolated; the page flags it.
  • Not read: HST/WFC3 spatial scans and any two-dimensional product (calints, rateints). NIRSpec PRISM data and the published PRISM spectrum were not compared. Sessions hold the binned channel light curves, not the cube.
  • Exposure averaging uses midpoint sub-samples; batman's end-to-end spacing differs by 3.7 ppm in the test.

References: Alderson et al. 2023; Tinetti et al. 2018; Husser et al. 2013; Kreidberg 2015; Winn et al. 2008; Mancini et al. 2018.

18. Derived quantities and uncertainty propagation ​

Code: src/core/derive.js, src/core/session.js (summarizeFit), src/main.js, src/core/stellar.js, src/core/hz.js

The system table (modelled on Table 7 of Pearson et al. 2022) is computed from the fitted parameters and the stellar inputs:

QuantityExpression
transit depthk2 (geometric, no limb darkening)
planet radiuskR⋆
semi-major axis(a/R⋆)R⋆
impact parameterb=(a/R⋆)cos⁡i(1−e2)/(1+esin⁡ω); at eclipse with 1−esin⁡ω
durations, eclipse and periastron timesnumerical, see Kepler solver and orbits
stellar densityρ⋆=3π(a/R⋆)3/(GP2) (planet mass neglected)
equilibrium temperatureTeq=TeffR⋆/2a (zero albedo, full redistribution, at the semi-major axis)
insolationR⋆2(Teff/5772K)4/a2 in Earth units, a in au
velocity semi-amplitudeK of Radial velocities, when the planet has a mass
planet densityfrom Mp and kR⋆

Propagation ​

Derived quantities are computed on equal-weight posterior draws (600), one evaluation per draw, and summarised by the median and the 15.87% and 84.13% quantiles, so correlations between parameters are carried through. In the analyzer, R⋆, M⋆ and Teff are drawn for each sample from normal distributions with their input uncertainties, unless they were fitted or are tied through Kepler's law or the velocities, in which case the fitted or held values are used. In the headless path (runSession, the Python package) the stellar values are used as given, without their uncertainties, unless they were free parameters.

Stellar mass and radius ​

When only Teff, log⁡g and [Fe/H] are known, mass and radius come from the eclipsing-binary calibration of Torres, Andersen & Giménez 2010, with its scatter (0.027 dex in log⁡M, 0.014 dex in log⁡R) added in quadrature to the propagated input errors; the result is flagged as extrapolated below 0.6 M⊙ or outside 3800 to 10000 K.

Habitable zone ​

The limits are the effective fluxes of Kopparapu et al. 2013; Kopparapu et al. 2014, Seff=S⊙+aT+bT2+cT3+dT4 with T=Teff−5780 K (clamped to 2600 to 7200 K), converted to d/R⋆=215.032(Teff/5780)2/Seff. The band drawn is averaged over the Teff uncertainty and an albedo range; the albedo scaling is a heuristic of this application, not part of the published limits.

Accuracy and checks ​

Assumptions and limits ​

  • A planet radius or mass in physical units is only as good as the stellar radius and mass. If the star was held fixed, the quoted uncertainty of Rp, a and Mp excludes the stellar uncertainty in the headless path and includes it only through the resampling described above in the analyzer. State which applies.
  • The equilibrium temperature assumes zero albedo and full redistribution. The transit depth is k2, not the limb-darkened minimum of the light curve.
  • The Torres calibration returns 1.05 M⊙ and 1.02 R⊙ for solar inputs, the known offset of that relation.

References: Pearson et al. 2022; Torres, Andersen & Giménez 2010; Kopparapu et al. 2013; Kopparapu et al. 2014; Winn 2010.

19. Catalogue sources, merging and provenance ​

Code: src/core/catalog.js, src/core/sources/, src/core/relations.js, tools/build_catalog.py, tools/build_catalog_table.py, tools/build_catalog_eu.py, tools/build_esl_catalog.py

Sources ​

The site works from dated snapshots, because the archives cannot be queried from a browser page. (1) The NASA Exoplanet Archive Akeson et al. 2013 Planetary Systems table ps, every published reference of every planet, and the composite table pscomppars, through its TAP service. (2) The Extrasolar Planets Encyclopaedia, exoplanet.eu Schneider et al. 2011, from its CSV export. (3) The TESS Objects of Interest table Guerrero et al. 2021 as served by the NASA archive. Each snapshot records its date, query and row counts; the merged offline catalogue (public/esl-catalog/) carries checksums of its inputs.

Choosing the parameters of a planet ​

For a look-up in the analyzer, every reference of the planet is scored and the highest score is used (chooseParameters): 100 × completeness (weighted by what the chosen kind of analysis needs; a derivable quantity counts in part) + consistency (−25 per failed check, −60 per fatal one, −8 per warning, +1.5 per passed check up to 6) + recency (up to 3) + small terms (the archive's default flag, stellar parameters from the same paper, candidates and controversial planets marked down). The consistency checks compare a reported quantity with the value implied by the reference's other numbers and count the difference in σ with a fractional floor: a/R⋆ against Kepler's third law; stellar density from a/R⋆ and P against M⋆, R⋆; log⁡g against M⋆, R⋆; a against (a/R⋆)R⋆; the impact parameter; the transit duration (numerical); the depth against k2; k against the radii; K against the mass; periastron outside the star; the transit geometry. A difference up to 3σ passes, up to 5σ warns, beyond fails.

Whatever the chosen reference lacks is filled in a fixed order and labelled: reported by the chosen reference; derived from its own numbers; taken from another reference (named); derived from a mix; assumed (for example e=0 when only an upper limit exists, or [Fe/H] = 0). The page shows every value with its label and source, and the alternatives with their scores.

Merging catalogues ​

For the Search and Plan pages the two planet catalogues are matched in stages: identical normalised names; aliases; host and planet letter; then position within 1 arcminute with periods within 3%. A match from the later stages is rejected when the positions differ by more than 0.25° (unless the periods agree to 0.1%) or the periods by more than 20%. Period and epoch are always taken from the same catalogue. Values that differ by more than 3σ and by more than a per-quantity floor are marked as disagreeing. For the planner, the ephemeris is the published (P, T0) pair that predicts the smallest timing uncertainty today, σT02+(NσP)2.

Estimates ​

Missing quantities are estimated from named relations, never overwriting a catalogue value, and marked as estimates (with a flag when an estimate is built on another estimate): the mass-radius relation of Chen & Kipping 2017 (no mass above the radius where it stops being single-valued); the dwarf sequence of Pecaut & Mamajek 2013; Torres, Andersen & Giménez 2010; and exact geometry for depth Seager & Mallén-Ornelas 2003, duration and impact parameter Winn 2010, equilibrium temperature (zero albedo, full redistribution) and K. The relation of Otegi, Bouchy & Helled 2020 is implemented as an alternative.

Accuracy and checks ​

  • Chen & Kipping mass-radius relation against forecaster: largest difference 1.37%. (Validation: Estimates in the catalogue; test/relations.test.mjs)
  • Pecaut & Mamajek dwarf sequence against the published table: 63 rows, B9V to M9.5V, identical. (Validation: Estimates in the catalogue; test/relations.test.mjs)
  • Dwarf masses against Torres et al. 2010: within 0.022 dex (that calibration’s own scatter is 0.027). (Validation: Estimates in the catalogue; test/relations.test.mjs)
  • Snapshot of 2026-11-15 (the index files under public/catalog*/ and public/esl-catalog/manifest.json): 40,414 rows of ps for 6,470 planets around 4,862 hosts; 8,479 exoplanet.eu planets; 8,175 TOIs; 8,561 planets in the merged catalogue, 6,388 matched in both (5,882 by name, 254 by alias, 31 by host and letter, 221 by position).
  • Selection logic on constructed references with planted inconsistencies and on real systems: node test/catalog.test.mjs, node test/sources.test.mjs, node test/eslcatalog.test.mjs.

Assumptions and limits ​

  • The scoring weights and thresholds are judgement calls, not calibrated against anything. The chosen reference is a starting point; a paper should cite the original publications of the values it uses, which the provenance labels identify.
  • Catalogue values are not re-derived from data. Time systems are flagged but not converted. Choosing an ephemeris by its quoted errors favours publications with optimistic errors.
  • Population relations have large scatter and are not measurements. Estimated values must not be reported as measured.
  • TOIs are not matched to confirmed planets. Only the offline path of the merged-catalogue builder has been run.

References: Akeson et al. 2013; Schneider et al. 2011; Guerrero et al. 2021; Chen & Kipping 2017; Otegi, Bouchy & Helled 2020; Pecaut & Mamajek 2013; Torres, Andersen & Giménez 2010; Seager & Mallén-Ornelas 2003; Winn 2010.

20. Observability and sky positions ​

Code: src/core/observability.js, src/core/astro.js, src/core/horizon.js

Two modules compute positions; both take UT1 = UTC and use the algorithms of Meeus 1998: IAU 1976 precession, the mean sidereal time of 1982, a four-term nutation, the low-precision Sun (chapter 25) and a truncated lunar theory (chapter 47). The sky view (astro.js) adds the planets from the Keplerian elements of Standish & Williams (JPL) with light-time iteration, apparent sidereal time, and refraction for display by the formulas of Bennett and Sæmundsson as given by Meeus 1998 (see also Bennett 1982). The planner (observability.js) uses mean sidereal time for the target, topocentric parallax for the Moon, and the airmass of Kasten & Young 1989.

Which transits are observable ​

Transit n is at Tn=T0+nP, converted from BJDTDB to UTC, with ingress and egress at Tn∓T14/2. Its mid-time uncertainty is

σn=σT02+(nσP)2

which assumes T0 and P uncorrelated. A transit is listed as observable in full when the whole region of interest, from one hour before ingress to one hour after egress, has the target above 30° altitude with the Sun below −18°; complete transits with a shorter baseline, and partial ones, are listed as such. All three limits are settings. Rise, set and twilight times are found by sampling every 10 minutes and bisecting.

Terrain horizon ​

The horizon profile is traced through elevation tiles (Mapzen/Tilezen Terrarium) along great circles out to 150 km on a spherical Earth, with terrestrial refraction through an effective radius R/(1−k), k = 0.13.

Accuracy and checks ​

Assumptions and limits ​

  • Positions are good enough to find and frame a target and to decide observability, not for astrometry: annual aberration of stars (20″), proper motion, parallax and polar motion are not modelled, and Jupiter and Saturn are limited to about 10′ by the simple elements.
  • Planner altitudes are geometric (no refraction). The planner's barycentric correction uses the Sun's centre and is good to about 4 s.
  • The ephemeris uncertainty ignores the covariance of T0 and P and trusts the published errors; an ephemeris that is stale or optimistic gives a window that is too narrow.
  • The comparison with the Swarthmore transit finder covers one site and one week. Terrain data have 30 m to 300 m pixels and contain no buildings or trees; summits come out 0.08° to 0.14° low in the three places checked.

References: Meeus 1998; Standish & Williams (JPL); Bennett 1982; Kasten & Young 1989; Jensen 2013; Astropy Collaboration 2022; Anderson & Francis 2012.

21. Image photometry ​

Code: src/core/photometry.js, src/core/fits.js, src/photometry/pipeline.js, src/photometry/transitfit.js

The photometry page turns a stack of FITS frames, or a TESS, Kepler or K2 target pixel file, into a differential light curve. The method follows EXOTIC Zellem et al. 2020, with the differences listed below. Pixel centres are at integer coordinates.

  • Sky. The median of the pixels whose centres lie in an annulus around the star, after iterative clipping at 3σ about the median (5 iterations). The annulus starts 2 px outside the aperture.
  • Centroid. A flux-weighted centroid above a threshold of 3σ over the local sky in a box, re-centred until it moves by less than 0.01 px, followed by a Gaussian-windowed centroid. Stars are tracked from frame to frame; when the field jumps, rotates or flips (meridian flip), sources are re-detected and matched by triangle similarity Groth 1986; Valdes et al. 1995; Beroiz et al. 2020 and a similarity transform is fitted.
  • Aperture sum. The exact overlap area of a circle (or an ellipse of the same area) with each pixel, the algorithm of photutils Bradley et al. (Photutils). Flux = sum − area × sky.
  • PSF photometry. An elliptical Gaussian, B+Aexp⁡[−(x′2/2σx2+y′2/2σy2)] with flux 2πAσxσy, or a Moffat profile, B+A(1+r2/γ2)−β with flux πAγ2/(β−1), fitted to a box by unweighted Levenberg-Marquardt with analytic derivatives; the model is averaged over each pixel when the star is undersampled.
  • Optimal extraction. The sky-limited estimator of Naylor 1998 with a Gaussian profile P taken from the brightest well-fitted star of the frame: F=∑P(d−B)/∑P2. The pixel variance is assumed uniform.
  • Comparison stars. The target's flux is divided by one comparison star or by the unweighted sum of all of them; the alternative "no comparison star" is always tried.
  • Choice of aperture, annulus and comparison. Every combination of 12 aperture radii from 1.5 to 6 times the PSF σ of the first frame, 4 annulus widths from 6 to 15σ, each comparison choice, and the PSF and optimal fluxes, is reduced to a light curve; the one with the smallest scatter of residuals is kept. The residuals are from a least-squares fit of a baseline linear in airmass (or time) and a trapezoidal transit, located once and then held to the same window for every candidate, with one pass of clipping at 4 robust σ. With a known planet the candidates can instead be scored with the transit model of this document through the same likelihood, with EXOTIC's bounds.
  • Calibration. Bias, dark and flat frames found among the files (IMAGETYP, OBSTYPE or the file name) are combined pixel by pixel with the median. The master dark and master flat have the master bias subtracted; the flat is divided by its median. Each light frame becomes (frame − bias − dark) / flat; pixels where the flat is below 0.05 are blanked.
  • Airmass is read from the header or computed as sec⁡z.
  • Time. Mid-exposure UTC converted to BJDTDB as described under Time standards.

Uncertainties ​

The uncertainty attached to every point of the exported light curve is one number: the robust scatter (1.4826 × median absolute deviation) of the residuals of the fit just described, relative to the median flux. A photon-noise estimate is computed for diagnosis only,

σF2=Fg+npixσsky2(1+npixnsky),σRR=σT2T2+σC2C2

in ADU, with g the gain in e⁻/ADU (from the header keywords GAIN or EGAIN, or typed; 1 when unknown), σsky the scatter in the annulus (which contains sky, read and dark noise), npix the aperture area, nsky the annulus area, and R=T/C the ratio of target to comparison.

Detrending ​

For pointing-driven systematics (K2, TESS) two detrenders are offered: the self-flat-field of Vanderburg & Johnson 2014 (flux against arclength along the principal axis of the centroids, in windows, with a running-median trend removed) and a Gaussian process with a squared-exponential kernel on the centroid coordinates (optionally time), whose hyperparameters maximise the marginal likelihood on a subset of points; the light curve is divided by 1 + trend.

Accuracy and checks ​

Assumptions and limits ​

  • The exported error bars are a single empirical scatter, the same for every point; they do not follow airmass, clouds or the star's count rate, and they include any real signal that the trapezoid does not describe. There is no read-noise or scintillation term in the per-point estimate, and the gain defaults to 1 e⁻/ADU.
  • Choosing the aperture and comparison star that minimise the residual scatter of a transit fit is a selection on the data: it can bias the scatter low. On EXOTIC's sample the choice of comparison star moves Rp/R⋆ by more than its formal uncertainty.
  • Darks are not scaled by exposure time: dark frames must have the light frames' exposure (or be left out). The flat is not dark-subtracted. There is no saturation or non-linearity correction beyond flagging saturated stars and leaving them out of the comparison candidates.
  • Checked on synthetic frames, two archived TESS and Kepler pixel files, and one set of real ground-based frames compared with EXOTIC's reduction of the same frames; no reduction here has been compared with a published light curve. At EXOTIC's own aperture this code's scatter is higher (0.599% against 0.573%) and the reason was not found.
  • HCOMPRESS-compressed FITS files are not read.

References: Zellem et al. 2020; Bradley et al. (Photutils); Naylor 1998; Groth 1986; Valdes et al. 1995; Beroiz et al. 2020; Vanderburg & Johnson 2014; Merline & Howell 1995; Ricker et al. 2014; Borucki et al. 2010.

22. TESS, Kepler and K2 light curves: quality, stitching, transit-aware detrending, K2 systematics ​

Code: src/core/lcprep.js, src/core/k2.js, src/core/zip.js, src/photometry/prep.js, src/photometry/mast.js

The photometry page lists every product the public MAST bucket holds for a star (SPOC 2-minute and 20-second postage stamps, TESS-SPOC and QLP products of the full-frame images, Kepler and K2 long and short cadence; light curves and target pixel files) and loads one product per sector, as the owner's TOI extraction does: 20 s and 2 min of the same sector are the same photons and must not both enter a likelihood. A target pixel file is reduced with the aperture search of the previous section.

  • Quality. lightkurve's bitmasks Lightkurve Collaboration 2018: 'default' (TESS 17087, Kepler and K2 1130799) unless another preset is chosen. Light-curve files: PDCSAP (or SAP) with the same mask.
  • Stitching. Each sector (quarter, campaign) is divided by its own median; a cadence less than 1 s after another is dropped.
  • Transit mask. Points with |((t−T0+P/2)modP)−P/2|<1.5T14/2 for any known planet, T14=(P/π)arcsin((1+k)2−b2/(a/R⋆sin⁡i)) (with the eccentric factor); planets from the catalogue and the TOI list Guerrero et al. 2021, or typed.
  • Trend. wotan's time-windowed biweight Hippke et al. 2019: for every point, Tukey's biweight location (c = 5, from the median and the MAD, iterated to 10−6 MAD) of the unmasked points within ±w/2, segments split at gaps longer than 1 h. With "auto" the window w is chosen per sector by injecting box transits of the longest duration (1500 ppm) away from the real ones, unmasked, and maximising recovered depth × T14/Δt / noise (EXOTIC's optimize_window).
  • Clip and trim. 4σ about a Savitzky–Golay smooth of 30 min, never in transit; 30 min after every gap longer than an hour.
  • Windows. ±2.5T14 around every predicted transit with points inside it; another planet's transit in a window is reported and, when its geometry is known, divided out.
  • K2. Pixel-level decorrelation Deming et al. 2015 as EVEREST Luger et al. 2016; Luger et al. 2018 and lightkurve's PLDCorrector set it up for K2: the 3σ threshold-mask pixels over the aperture flux, their products to third order (16 principal components per order), the principal components of the background pixels (below the median image, normalised), a degree-5 B-spline in time with n/50 columns for the star, fitted jointly by least squares with Gaussian priors (10× the flux scatter) and 5σ clipping, five times; the pixel and background parts are removed. Thruster firings are flagged from QUALITY bits 20 and 21 and from 5σ jumps of the centroid along the roll arc Vanderburg & Johnson 2014.
  • CDPP is lightkurve's estimate_cdpp (Savitzky–Golay flatten, 5σ clip, the standard deviation of a 13-cadence running mean), with both lengths scaled from Kepler's 29.4-minute cadence when sectors of other cadences are compared.
  • Zips. Zip archives are read from their central directory (ZIP64 included); a deflated entry becomes the slice of the archive between a 10-byte gzip header and a trailer made from the directory's CRC-32 and size, so frames are inflated one at a time as the pipeline reads them.

Accuracy and checks ​

  • Against lightkurve 2.6 on the WASP-18 (TESS), TrES-2 (Kepler) and K2-3 (K2) pixel files: the cadences of every quality preset, times to 0 µs, aperture flux 3e-7 and centroids 3e-7 of the CCD coordinate (lightkurve sums float32), SAP_FLUX of the pipeline's light-curve files to 1.2e-7, threshold masks identical, CDPP to 4e-5: node test/tpf_lightkurve.test.mjs.
  • Biweight trend against wotan to 1.9e-9 with and without transit masks, EXOTIC's T14, mask, clip, trim, binning, window injection (7e-7) and alias scan exactly, lightkurve's flatten (5e-12) and CDPP (1e-8), the chain on TrES-2's pixels (trend 2.4e-7, depth 13184 ppm both): node test/lcprep.test.mjs.
  • K2-3, campaign 1, CDPP at 6.5 h: raw SAP 5822 ppm; self-flat-field 326 ppm (lightkurve SFFCorrector 158); PLD order 3 40.2 ppm (lightkurve PLDCorrector 48.0; the pipeline's PDCSAP 186); thruster firings from the centroid match 154 of 172 flagged cadences: node test/k2.test.mjs.
  • Zips written by Python's zipfile (stored, deflated, ZIP64, nested, .fits.gz inside) and EXOTIC's sample-data zip: node test/zip.test.mjs.

Assumptions and limits ​

  • TESScut cutouts, TGLC, K2SFF and EVEREST light curves cannot be fetched from a page (no CORS, or not in the bucket). Sectors of different cadence go to the analyzer as one light curve with one exposure time (the median); the analyzer does not yet read the window index the hand-off carries ("Split by orbit" does the same).
  • The self-flat-field here is worse than lightkurve's on the whole K2-3 campaign (326 against 158 ppm) and better on its first 24 days; the reason was not found. Dropping the unflagged thruster firings did not lower the CDPP of K2-3.
  • The window injection is EXOTIC's procedure; its point counts are scaled for cadences coarser than 2 min, and on a quiet star (TrES-2) every window recovers the injected depth equally, so the choice is then arbitrary.

References: Hippke et al. 2019; Lightkurve Collaboration 2018; Guerrero et al. 2021; Deming et al. 2015; Luger et al. 2016; Luger et al. 2018; Vanderburg & Johnson 2014.

23. Image inspection, focus and comparison stars ​

Code: src/core/imexam.js, src/photometry/inspect.js, src/core/lcfits.js

The photometry page has the inspection tools of IRAF's imexam Tody 1986 and SAOImage DS9 Joye & Mandel 2003: a radial profile, a line cut, region statistics, a pixel table, the FITS header, and a map of the focus across the field. They act on the frame shown, calibrated as above.

  • Radial profile. The centre is refined as for photometry. The sky is the 3σ-clipped median of the pixels between rmax+2 and rmax+8 px. A Gaussian Ae−r2/2σ2 (FWHM 22ln⁡2σ) and a Moffat profile A(1+r2/γ2)−β Moffat 1969 (FWHM 2γ21/β−1) are fitted by least squares to every pixel inside rmax, sky subtracted. The binned profile drawn is the mean in 1 px annuli with exact pixel overlaps, as photutils' RadialProfile Bradley et al. (Photutils). rmax is set to about 3 FWHM after a first pass.
  • Curve of growth and half-flux diameter. Sky-subtracted sums in circles every 0.25 px with exact overlaps. The half-flux diameter (HFD) is twice the radius holding half the flux inside rmax, found by linear interpolation; it equals the FWHM for a Gaussian and is the usual focusing number because it does not depend on the profile's shape.
  • Signal to noise. F/F/g+npixσsky2 for the circle of radius rmax.
  • Saturation. SATURATE, SATLEVEL, DATAMAX or MAXLIN from the header, else the largest value of the integer format (with BZERO and BSCALE).
  • Elongation. A Gaussian with free position angle fitted to a box of about 2 FWHM (the PSF fit of the photometry): ellipticity 1−b/a and the direction of the major axis.
  • Focus across the field. Stars are detected, and saturated stars, stars with a neighbour brighter than 5% of their flux within 2.5 box half-widths, and failed fits are dropped. Each remaining star gets the rotated Gaussian fit (twice: the second with a box of 4σ of the median star). The field is cut into n×n cells, each with the median FWHM and ellipticity and the ellipticity-weighted mean of e2iθ. A quadratic surface in x,y is fitted to the FWHM with one 3σ rejection; its minimum, when inside the frame, is marked as the sharpest point. The summary compares the corner cells with the centre cell (curvature), the corners with each other (tilt), and reports |⟨ee2iθ⟩|/⟨e⟩, near 1 when every star is stretched the same way (tracking, guiding, wind) and near 0 for optical patterns, and ⟨ecos⁡2(θ−ϕ)⟩/⟨e⟩ over stars away from the centre (ϕ the direction from the centre: +1 radial, −1 tangential).
  • Ranking comparison stars. For each detected star: |ln⁡(F/Ftarget)|, plus 3 when the flux ratio is outside 1/10 to 10, plus 10 when saturated, plus 1+2(4−d)/4 when a neighbour brighter than 10% of it lies within d<4 FWHM, plus 2 within 4 FWHM of the edge, plus 1.5 when its FWHM exceeds the median by 30% (a blend or a galaxy), plus 1 when its ellipticity exceeds 0.3. Lower is better; the four best usable stars are proposed when only the target is chosen. The light-curve search still tries every comparison choice and keeps the one with the least scatter.
  • FITS light-curve tables. The analyzer reads TESS, Kepler and K2 light-curve files: the first binary table, TIME plus BJDREFI + BJDREFF (BTJD = BJDTDB − 2457000, BKJD = BJDTDB − 2454833), the first of PDCSAP_FLUX, KSPSAP_FLUX, DET_FLUX, FCOR, FLUX, SAP_FLUX (or a column chosen by name) with its error column, rows with a non-zero QUALITY flag or a blank time or flux left out, and flux and errors divided by the median.

Accuracy and checks ​

  • Against photutils and scipy on six stars of a real frame (EXOTIC's HAT-P-32 b sample), and on synthetic stars and fields: node test/imexam.test.mjs. FITS tables written by astropy in the SPOC and Kepler layouts: node test/lcfits.test.mjs.

Assumptions and limits ​

  • The FWHM fits assume a round profile; the elongation comes from a separate elliptical fit. Undersampled stars (FWHM below about 2 px) are fitted to pixel values without averaging the model over the pixel, which widens the FWHM by about σ2+1/12/σ (2% at σ = 1.5 px).
  • The focus verdicts use fixed thresholds (15% for curvature and tilt, ellipticity 0.12 to 0.15, coherence 0.6) chosen on synthetic fields; they describe the frame, not its cause. Seeing and wind change from frame to frame, so a single frame can mislead.
  • Comparison stars are ranked on one frame. Variability, colour and the star's position in later frames are not considered, and no catalogue of variable stars is queried.
  • Saturation is read from the header or the data format only; a camera that saturates below its format limit (common for 16-bit CMOS) is not recognised without a SATURATE keyword.

24. Noise diagnostics ​

Code: src/core/psd.js, src/core/gp.js, src/photometry/diagnostics.js

Residuals are examined in three ways. None of them enters the likelihood; they tell the user whether the white-noise assumption of the fit holds.

Power spectrum ​

The one-sided power spectral density is normalised so that its integral from 0 to the Nyquist frequency is the variance: for evenly sampled data Pk=2Δt|∑jwjxje−2πijk/n|2/∑jwj2 (not doubled at zero and Nyquist), by a fast Fourier transform, or averaged over half-overlapping Hann-windowed segments Welch 1967; for uneven sampling the Lomb-Scargle periodogram is scaled to the same density. White noise of variance σ2 then sits at

Pwhite=2σ2Δt

and the level expected from the error bars is 2⟨σi2⟩Δt. A model P(f)=W+Af−γ can be fitted to the logarithm of the (bias-corrected) spectrum; red noise is flagged when it improves the fit significantly and the knee lies inside the sampled band.

RMS against bin size ​

The residuals are averaged in non-overlapping bins of N points (M bins) and the rms of the bin means is compared with the white-noise expectation Pont, Zucker & Queloz 2006; Winn et al. 2008, in the convention of Cubillos et al. 2017:

σNexpected=σ1NMM−1,β=maxNσNσNexpected

with an uncertainty σN/2M and the maximum taken over bin sizes with at least 10 bins. β>1 indicates correlated noise.

Photon and scintillation noise ​

For reference the page computes the CCD equation of Merline & Howell 1995, σ2=S+npix(1+npix/nsky)(gB+D+R2)+(σscintS)2 in electrons, and the scintillation estimate σscint=0.09D−2/3X7/4(2t)−1/2e−h/8000m Young 1967; Dravins et al. 1998 for a telescope of diameter D (cm), airmass X, exposure t (s) and altitude h.

Pixel map ​

Residuals binned against the centroid position, with a polynomial or Gaussian-process surface, show intrapixel sensitivity.

Accuracy and checks ​

  • node test/psd.test.mjs (references: scipy 1.13.1, astropy 6.0.1, MC3 3.2.1): periodogram density against scipy.signal.periodogram within 5.2e-14 of the highest peak; Welch against scipy.signal.welch within 1.0e-15; Lomb-Scargle density against astropy within 7.7e-15; rms against bin size, expectation and β against mc3.stats.time_avg within 7.1e-15; red-noise fit against scipy.optimize.least_squares within 1.8e-8.
  • Gaussian-process detrending: likelihood against scikit-learn: 1.1e-13. (Validation: From images to a light curve; test/gp.test.mjs)

Assumptions and limits ​

  • β is a maximum over bin sizes and is biased high for white noise (1.14 on a 20,000-point white series in the test). MC3's asymmetric uncertainty for fewer than 36 bins is not reproduced.
  • These are diagnostics. β is not applied to the error bars and no correlated-noise model enters the fit; if the diagnostics show red noise, the posterior widths from the white-noise likelihood are too small, and the paper should say what was done about it.

References: Welch 1967; Pont, Zucker & Queloz 2006; Winn et al. 2008; Cubillos et al. 2017; Merline & Howell 1995; Young 1967; Dravins et al. 1998; Virtanen et al. 2020.

25. The session format and reproducibility ​

Code: docs/FORMAT.md, src/core/session.js, src/core/exportScript.js, python/, tools/build_example_results.py

An analysis is one JSON document (docs/FORMAT.md, format esl-session version 1): the star, the limb-darkening law and coefficients, the planets' parameters, the data (light curve with epochs and exposure, velocities per instrument, event times), which kinds of data enter the fit, every free parameter with its prior, every fixed value, the sampler settings including the seed, the software version, and, after a run, the result: evidence and its uncertainty, information, effective sample size, number of likelihood calls, convergence and stop reason, warnings, medians and quantiles, derived quantities, the best-fitting point with its χ2 by kind of data, and optionally equal-weight posterior draws. All times in the file are absolute.

The same document runs in the browser, headless in Node (runSession), in the Python package, which executes the same JavaScript core rather than a translation, and through the MCP server. The Results tab also writes a Python script that rebuilds the fit with every prior as an editable line, a Markdown report and LaTeX tables.

To make a published result reproducible, archive the session file with posterior draws (for example as supplementary material or in a data repository) and cite the software version it records. Re-running the file with the same version reproduces the run.

Accuracy and checks ​

  • node test/session.test.mjs: a session turned into a likelihood reproduces EXOTIC's χ2 on the three-planet example (light curves 324.49 against 324.49, velocities 37.58 against 37.58); a run recovers the injected masses and survives a JSON round trip.
  • Python against Node on stored sessions (python/README.md, "What was reproduced"): five of six reproduce exactly (same number of likelihood calls, identical medians and ln⁡Z); hd80606-tess.json takes 12,813 calls against 12,837 and agrees to 0.004σ. An exported script run under the package reproduces the JavaScript result to 4e-14 relative (test/exportScript.test.mjs).
  • Every bundled example opens with a stored posterior (public/examples/results/), rebuilt by tools/build_example_results.py; the reproducibility table below quotes those files.

Assumptions and limits ​

  • Identical results need the same software version and the same family of JavaScript engine: the sampler's path depends on the last digit of the engine's mathematical functions. Between Node 22 and the Python package's default runtime the call counts are identical and medians agree to 1e-15 relative; another engine may differ within sampling noise.
  • The session does not record the time standard of the data, the source of fixed limb-darkening coefficients, or whether star-tied limb darkening used the table or the grey atmosphere. The paragraph generator below lists these as items to state by hand.
  • The software version in a session is a string (0.2.0); no commit hash is recorded yet, and there is no archived release with a DOI. The Python package was run on macOS arm64 with Python 3.12 only and is not on PyPI.
  • Star-tied limb darkening is set up by the application, not by specFromSession; outside the browser use prepareSession of the Python entry points.

26. Validation ​

How the models compare with other codes, with independent references and with published results. On the site this is the Validation tab of the Methods page (https://exoplanet.science/methods/#validation), which recomputes this code's side of the first table in the browser, draws the comparisons and lists where this code is weaker; below are the numbers as the tests print them.

Comparisons with other codes ​

Numbers as printed by node test/validation.test.mjs on 2026-11-15; the page recomputes this code's side in the browser.

ComparisonReference dataResult
Transit fluxpublic/validation/transit.json, generated by tools/validation/make_transit.py (inputs chosen by this project; reference: scipy quadrature on a 30-digit orbit)41 light curves: worst error 3.1e-11 (tolerance 1e-10); branch points 1.4e-10; automatic exposure averaging 1.3e-7 to 6.9e-7
Orbit and timingpublic/validation/orbit.json, tools/validation/make_orbit.py (reference: 30-digit arithmetic)true anomaly 2.4e-15 rad up to e = 0.999; event times 2.3e-10 s; durations 1.3e-10 s. HD 80606 b eclipse duration: 0.08337 d here and in the reference, 0.07169 d published: NOT REPRODUCED
Radial velocitypublic/validation/rv.json, tools/validation/make_rv.py5 orbits up to e = 0.93: worst 1.3e-14 K; K from masses 3.1e-8 relative
Posteriors, evidence, coveragepublic/validation/fits.json and esl.json (four simulated data sets; tools/validation/make_fits.py, esl_reference.mjs)means within 0.04σ and widths within 3.5% of dynesty or the exact posterior; ln Z within 0.31; 68% coverage 67.0 to 73.3% over 300 trials (expected 68 ± 2.7%)

Checks elsewhere in the project ​

Each line is printed by the test named and stored in public/validation/project.json; node test/validation.test.mjs fails when a number quoted here is no longer in that output.

TopicReference data and codesWhat was reproducedRegenerate
EXOTIC’s own examples, reproducedAgainst EXOTIC, pylightcurve. Stored references: test/exotic_reference.json, test/exotic_fits.json; generated by test/make_exotic_reference.py, test/make_exotic_fits.py.Transit model against pylightcurve, five planets up to e = 0.93: largest flux difference 1.0e-7
Single transit: posterior against EXOTIC’s lc_fitter: means within 0.25σ, widths within 15% (the test’s limits)
Five nights with airmass trends: against glc_fitter: means within 0.11σ, widths within 6%
Three-planet toy system: χ² of each light curve at the injected values: 45.11, 88.80 and 190.58 in both codes
HD 80606 b, velocities from seven instruments: K = 469.6 m/s (published 469.2, EXOTIC 469.9)
HD 80606 b, light curve + velocities + timing, against Table 7 of Pearson et al. 2022: P within 5e-4 d, i within 0.03°, e within 0.001, K within 3 m/s
Central values are comparable in the velocity and joint fits; error bars are not, because EXOTIC averages a reduced χ² over instruments and counts the light curve twice. This code’s planet mass is 0.38% higher for the same velocities (two-body factor).
node test/exotic.test.mjs
node tools/validation/project_reference.mjs exotic
The nested sampler on hard problemsAgainst UltraNest. Stored references: test/stress_reference.json, test/stress_ultranest.json; generated by test/make_stress_reference.mjs, test/stress_ultranest.py.Correlated Gaussian, 10 parameters, ρ = 0.99: ln Z = −7.729 ± 0.335, exact −6.931 (2.4σ low on this seed; over 32 seeds the mean offset is −0.078 ± 0.054)
Transit with a/R★ and inclination both free (0.99 correlated): ln Z = 744.121 ± 0.196, quadrature 744.090; means within 0.025 sd
Period aliases: three separate peaks: ln Z = 2.166 ± 0.154, exact 2.121; the peaks are reported with their weights
Neal’s funnel, 4 parameters: ln Z = −16.872 ± 0.158, exact −16.588 (1.8σ low on this seed); over 48 seeds −0.045 ± 0.026 since the ellipsoid check, −0.24 ± 0.05 before
Correlated Gaussian, 15 parameters, ρ = 0.99: ln Z = −10.245 ± 0.359, exact −10.397
Banana, 5 parameters: ln Z = −17.033 ± 0.204, exact −16.732 (1.5σ); posterior widths off by 23% on this seed, beyond the 20% tolerance (NOT PASSED)
One narrow peak, 8 parameters (information 63 nats): ln Z = 0.803 ± 0.373, exact 0 (2.2σ, just past the 0.8 ceiling of the check, and the run raised no warning); over 32 seeds the mean offset is 0.085 ± 0.076, so this seed is a fluctuation, not a bias (NOT PASSED)
Correlated Gaussian, 20 parameters, ρ = 0.999: not converged at the one-million-call budget (reported as such) (NOT PASSED)
Funnel shapes are the weak spot: ln Z can come out low (see “Where this code is weaker”). In slice-sampling mode the scatter of ln Z between runs exceeds the quoted error, and a peak much narrower than another can be lost.
npm run test:stress
node tools/validation/project_reference.mjs sampler
Limb darkening from stellar parametersAgainst ExoTETHyS, pylightcurve. Stored references: test/limbdark_reference.json, test/ldspec_reference.json; generated by test/make_ld_reference.py, test/make_ldspec_reference.py, tools/build_ld_table.py, tools/build_ld_spectral.py.HD 80606, TESS band, against the coefficients EXOTIC’s examples use: 0.56 ppm in flux
Four-coefficient law, worst spectral type (K): I(μ) rms 0.0009; 12.9 ppm in a transit
Quadratic law, worst spectral type (M): I(μ) rms 0.0029; 20.8 ppm
Same table against ATLAS models (different atmosphere physics): worst I(μ) rms 0.0327: the model-to-model spread dominates
Any wavelength range (357 bins, 200 to 5500 nm): 39 HST/WFC3 channels: I(μ) rms median 4.8e-4, worst 9.2e-3; transit depth within 0.72%
ExoTETHyS starts from the same PHOENIX files, so this tests the processing, the grid and the interpolation, not the physics. The boxcar and JWST bands have no external reference.
node test/limbdark.test.mjs && node test/ldspec.test.mjs
node tools/validation/project_reference.mjs limbdark
Periodic signals in transit timingAgainst astropy, EXOTIC. Stored references: test/periodogram_reference.json; generated by test/make_periodogram_reference.py.Power, four normalisations: 20,788 values within 4.0e-11 of the peak
False-alarm probabilities (Baluev and three others): relative difference 1.4e-14
A 2-minute signal in 30 of 500 epochs: detected in 93% of 300 trials
Lists with no signal, 1000 trials: flagged at the 1% level in 1.00%
Error bars too small by 45 s of unmodelled scatter: the analytic test raises false alarms in 68% of lists, the bootstrap in 1%
The analytic false-alarm probability is only as good as the error bars; with underestimated errors it is far too optimistic (last row), which is why the page offers a bootstrap.
node test/periodogram.test.mjs
node tools/validation/project_reference.mjs periodogram
Positions in the skyAgainst astropy. Stored references: test/astro_reference.json; generated by test/make_astro_reference.py.Stars, altitude and azimuth: max 25.2″ (20″ of it is annual aberration, not modelled)
Sun: max 33.8″
Moon: max 12.2″
Saturn (the worst planet): max 10.0′
Sidereal time: max 0.30″
Refraction above 10°: max 6.30″
Enough to find and frame an object, not for astrometry: Jupiter and Saturn are limited by the simple orbital elements used.
node test/astro.test.mjs
node tools/validation/project_reference.mjs sky
Which transits can be observedAgainst astropy, swarthmore. Stored references: test/observability_reference.json, test/swarthmore_reference.csv; generated by test/make_observability_reference.py, tools/compare_swarthmore.mjs.Target altitude, 400 positions: worst 0.0077°
Sunset, sunrise and the three twilights, 274 crossings: worst 6.4 s
Barycentric time correction (BJD_TDB − JD_UTC): worst 3.92 s
Swarthmore finder: mid-transit times: 32 events, worst 0.70 min
Swarthmore finder: elevations: worst 0.52°
Altitudes ignore refraction, time systems are not converted, and the Swarthmore comparison covers one site and one week.
node test/observability.test.mjs
node tools/validation/project_reference.mjs planner
From images to a light curveAgainst photutils, astropy, EXOTIC. Stored references: test/photometry_reference.json; generated by test/make_photometry_reference.py.Aperture sums with exact pixel overlap, 4 radii × 6 stars: 4.4e-16 relative
Flux-weighted centroid: 1.4e-14 px
Star detection: the same six stars as DAOStarFinder, within 0.03 px
FITS pixel values, six data types: 3.7e-16
UTC to BJD_TDB, 35 targets and dates: within 0.045 s
Gaussian-process detrending: likelihood against scikit-learn: 1.1e-13
End to end on synthetic frames with an injected transit: scatter 0.372%, at the photon-noise level
Synthetic frames and two archived TESS and Kepler pixel files. One set of real ground-based frames (EXOTIC’s HAT-P-32 b sample, 142 frames) is reduced in a separate test, test/exotic_photometry.test.mjs, and compared with EXOTIC’s reduction of the same frames: the two choose different apertures and comparison stars and give Rp/R★ = 0.1452 ± 0.0032 here against 0.1562 ± 0.0034. No reduction has been compared with a published light curve.
node test/photometry.test.mjs && node test/fits.test.mjs && node test/gp.test.mjs
node tools/validation/project_reference.mjs photometry
Estimates in the catalogueAgainst forecaster, astropy. Stored references: test/relations_reference.json; generated by test/make_relations_reference.py.Chen & Kipping mass-radius relation against forecaster: largest difference 1.37%
Pecaut & Mamajek dwarf sequence against the published table: 63 rows, B9V to M9.5V, identical
Dwarf masses against Torres et al. 2010: within 0.022 dex (that calibration’s own scatter is 0.027)
Velocity semi-amplitude of Jupiter and Earth against astropy constants: 12.474 and 0.08947 m/s in both
These are population averages with large scatter (0.25 dex in mass at 2.5 Earth radii); no mass is estimated above 11.3 Earth radii, where the relation is not single-valued.
node test/relations.test.mjs
node tools/validation/project_reference.mjs relations
Evidence, ephemerides and geometryStored references: test/ephemeris_reference.json, test/reference.json; generated by test/make_ephemeris_reference.py, test/make_reference.py.Flux dips exactly when the discs overlap: 160,000 samples, 0 mismatches
ln Z against midpoint quadrature, three free parameters: quadrature 1640.001; sampler mean of six seeds within its reported error
Ephemeris coefficients against exact rational arithmetic: within 1.8e-6 σ
Orbital decay of WASP-12 b from the published dP/dE: −30.74 ± 2.69 ms/yr here, −30.72 ± 2.67 published
Several nights in one fit: ln L against quadrature over each normalisation: difference 9.5e-8 at ln L = 2468.49
No published list of transit times has been fitted end to end and compared with its paper; only the unit conversion was.
node test/physics.test.mjs && node test/evidence.test.mjs && node test/ephemeris.test.mjs && node test/multiepoch.test.mjs && node test/kernel.test.mjs && node test/nested.test.mjs
node tools/validation/project_reference.mjs basics
Phase curves and eclipse mappingAgainst batman, EXOTIC, scipy. Analytic or computed in the test.Map basis curves to degree 3, through ingress and egress: max 4.6e-13 over 2256 values in 5 systems
Uniform-disc eclipse against batman 2.5.1: 4.9e-13 for circular orbits, 2.5e-7 at e = 0.93
Map curves against starry 1.2.0: max 5.6e-7 of the planet’s flux over 48 curves
Harmonic phase curve against EXOTIC: 5.0e-13 outside transit and eclipse
Brightness temperature against scipy quadrature and root finding: relative 1.6e-9 over 48 cases
The same model inside the fitting likelihood: χ² identical to the model used directly
EXOTIC’s eclipse ingress is that of a smaller body and starts late (up to 29% of the depth during ingress); here the planet’s own disc is occulted. A map of degree 3 is invisible in the phase curve and constrained only by ingress and egress. The examples are fits to already-detrended Spitzer data; instrument systematics are not modelled.
node test/phasecurve.test.mjs && node test/phasemodel.test.mjs
node tools/validation/project_reference.mjs phasecurve
ExomoonsAgainst scipy. Analytic or computed in the test.Two overlapping discs on the star: max 6.5e-14 over 205 cases and 5 limb-darkening profiles
exomoon2.py, step by step with this kernel: planet + moon within 1.3e-8
Pandora, whole light curves: max difference 5.2e-6
Timing and duration variations against Kipping’s formulas: within 2.4% and 0.5% of the amplitude
Search for an injected moon: found at S/N 26; 100 noise-only searches reach 5.0
Moonless light curves: 1 of 40 above the 95% level of noise (2 expected)
The moon’s orbit is circular and the planet-moon pair is treated as two bodies about a Keplerian barycentre; the star’s tide on the moon is ignored. The exomoon2.py configuration itself is unbound (2.4 Hill radii), which the page flags. No real detection has been reproduced: there is none.
node test/exomoon.test.mjs
node tools/validation/project_reference.mjs exomoon
N-body dynamicsAgainst EXOTIC, scipy. Stored references: test/nbody_reference.json; generated by test/make_nbody_reference.py.Transit times against REBOUND 5.2.2 (IAS15), EXOTIC’s example system: 1.3e-6 s over 610 transits
Same map as REBOUND’s WHFast: positions within 2.0e-12 AU after 8000 steps
One planet: identical to the Keplerian model: timing 3.4e-6 s over 400 orbits, velocity 1.6e-9 m/s
Critical angular-momentum deficit against a brute-force minimum: relative 3.8e-12 over 25 cases
Unseen companion recovered from 72 transit times: 0.1040 M Jup (injected 0.1), mean period 5.3336 d (5.3322)
The same dynamics inside the fitting likelihood: χ² 82.9 against 82.9 from the other integrator directly
Light curves stay Keplerian (no photodynamics), light-travel time is not applied, and there is no MEGNO: stability is judged by the Hill and AMD criteria and a short integration. The chi-square surface of a perturber has period aliases every few thousandths of a day; the search returns one solution that fits, not the only one. TTVFast, given the same Jacobi elements, differs by hundreds of seconds because it defines them differently.
node test/nbody.test.mjs && node test/nbodymodel.test.mjs
node tools/validation/project_reference.mjs nbody

27. Reproducibility table ​

For each bundled example: where the data come from and under what terms, what was reproduced, and the command that regenerates it. Values of this code are read from the stored result of each example (public/examples/results/, extracted from files dated up to 2026-11-16). Rows marked "NOT REPRODUCED" are results this code does not match. Published values are quoted as this repository records them; where it holds no uncertainty for a published value, none is shown. The validation topics, with the command that reruns each, are in the Validation section.

Bundled examples ​

HAT-P-32 b, one simulated night (single.json) ​

  • Data: Simulated by this code (EXOTIC example single_transit: a ground-based night with an airmass trend).
  • Terms: Generated here; no third-party data.
  • Cite: Zellem et al. 2020
  • Stored run: 4 free, 19,978 likelihood calls, ln Z = 1254.00 ± 0.20, converged, 1 sampler warning; stored 2026-11-16 with version 0.2.0
  • Outcome: EXOTIC’s posterior on the same simulated data is reproduced within the limits shown
  • Regenerate: node test/exotic.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
Single transit: posterior against EXOTIC’s lc_fittermeans within 0.25σ, widths within 15% (the test’s limits)test/exotic.test.mjs (Validation: EXOTIC’s own examples, reproduced)

HAT-P-32 b, five simulated nights (multi.json) ​

  • Data: Simulated by this code (EXOTIC example multiple_transit: five nights, each with its own airmass trend; the period is free).
  • Terms: Generated here; no third-party data.
  • Cite: Zellem et al. 2020
  • Stored run: 9 free, 175,905 likelihood calls, ln Z = 6244.55 ± 0.29, converged; stored 2026-11-16 with version 0.2.0
  • Outcome: EXOTIC’s posterior on the same simulated data is reproduced within the limits shown
  • Regenerate: node test/exotic.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
Five nights with airmass trends: against glc_fittermeans within 0.11σ, widths within 6%test/exotic.test.mjs (Validation: EXOTIC’s own examples, reproduced)

HD 80606 b, TESS transit (hd80606-tess.json) ​

  • Data: TESS 2-minute light curve of the February 2020 transit, the file of EXOTIC’s examples/joint_rv_transit (public/examples/hd80606/2458888_08_lightcurve.csv).
  • Terms: No licence stated in this repository for the file. TESS data are public (MAST); the reduced light curve is from the EXOTIC repository.
  • Cite: Pearson et al. 2022; Ricker et al. 2014; Zellem et al. 2020
  • Stored run: 3 free, 14,221 likelihood calls, ln Z = 4801.32 ± 0.24, converged; stored 2026-11-16 with version 0.2.0
  • Outcome: partly: the mid-time agrees; the inclination is 0.03° from the published value, 3.6 times this fit’s formal uncertainty (a/R★ is held fixed); the radius ratio is NOT reproduced (dilution)
  • Regenerate: .venv/bin/python tools/build_example_results.py # or open the example in the analyzer and press Run
QuantityThis codePublished or referenceSource of the comparison value
i [°]89.2718 +0.0089 −0.008389.24Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
T mid [BJD]2458888.0765 +0.0033 −0.00352458888.07466Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
Rp/R★0.0720 ± 0.0012 NOT REPRODUCEDnot comparablethe published radius uses a depth corrected for the light of HD 80607; this light curve is diluted and no correction is applied (README, "System table": the fitted radius is about 30% too small)

HD 80606 b, velocities from seven instruments (hd80606-rv.json) ​

  • Data: 264 radial velocities (ELODIE, HIRES, HRS, SOPHIE, APF) as compiled in EXOTIC’s examples/radial_velocity (public/examples/hd80606/HD80606_RV.csv).
  • Terms: No licence stated in this repository; the measurements are from the publications compiled by Pearson et al. 2022.
  • Cite: Pearson et al. 2022; Zellem et al. 2020
  • Stored run: 4 free, 24,161 likelihood calls, ln Z = -1465.92 ± 0.23, converged, 1 sampler warning; stored 2026-11-16 with version 0.2.0
  • Outcome: reproduced within the test’s limits (K within 5 m/s, e within 0.002, ω within 0.5°, P within 0.002 d); formal errors are too small because χ² = 1555 for 264 points without jitter
  • Regenerate: node test/exotic.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
K [m/s]469.60 +0.42 −0.40469.22Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
e0.93156 +0.00013 −0.000120.93183Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
ω [°]-58.832 ± 0.042−58.887Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
P [d]111.436747 +0.000082 −0.000086111.436765Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs

HD 80606 b, transit + velocities + timing (hd80606-joint.json) ​

  • Data: The TESS light curve and velocities above, four mid-transit times and one mid-eclipse time (sources named in each row of the session file).
  • Terms: As the two rows above.
  • Cite: Pearson et al. 2022; Zellem et al. 2020
  • Stored run: 7 free, 57,440 likelihood calls, ln Z = 3359.94 ± 0.22, converged, 1 sampler warning; stored 2026-11-16 with version 0.2.0
  • Outcome: orbit and mass agree within the test’s limits (P within 5e-4 d, i within 0.03°, e within 0.001, K within 3 m/s), though several differ from the published values by 1.5 to 2 times this fit’s formal uncertainties, which are too small (χ² = 1675 for 1168 measurements; the star is held fixed). The transit duration of this fit and the eclipse duration are NOT reproduced
  • Regenerate: node test/exotic.test.mjs && node test/validation.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
P [d]111.436866 +0.000065 −0.000063111.436765Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
i [°]89.2412 +0.0086 −0.007889.24Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
e0.93160 ± 0.000110.93183Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
ω [°]-58.823 ± 0.033−58.887Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
T mid [BJD]2458888.0711 +0.0021 −0.00192458888.07466Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
K [m/s]469.84 +0.38 −0.39469.22Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
a/R★94.452697 +0.000096 −0.00009894.452Pearson et al. 2022, Table 7, as recorded in test/exotic.test.mjs
T14 [d]0.4699 +0.0063 −0.0062 NOT REPRODUCED0.499Pearson et al. 2022, as recorded in public/validation/orbit.json. With the paper’s own parameters this code gives 0.4985 d (test/validation.test.mjs); the stored fit differs because its radius ratio and impact parameter come from the diluted TESS light curve
eclipse duration E14 [d]0.08135 +0.00012 −0.00013 NOT REPRODUCED0.0717Pearson et al. 2022, as recorded in public/validation/orbit.json; the 30-digit orbit gives 0.08337 d

HD 80606 b, the Rossiter–McLaughlin effect (hd80606-rm.json) ​

  • Data: The 264 velocities above and the five mid-times; the two partial transits in the file, SOPHIE 2009 February 14 (Moutou et al. 2009) and HIRES 2009 June 5 (Winn et al. 2009), are datasets of their own with free jitters (public/examples/hd80606/README.md). Rp/R★ held at 0.1001 (Winn et al. 2009) because the TESS radius is diluted by HD 80607; the anomaly uses quadratic limb darkening of the 390–690 nm band (the session’s rm block); no light curve in the fit.
  • Terms: As the rows above.
  • Cite: Pearson et al. 2022
  • Stored run: 9 free, 313,913 likelihood calls, ln Z = -1259.00 ± 0.26, converged, 3 sampler warnings; stored 2026-11-16 with version 0.2.0
  • Outcome: λ and v sin i★ agree (about 1σ and 0.6σ) from two partial transits; the same data with the orbit fixed give 1.47 km/s and 58° (test/hd80606rm.test.mjs). The model is the first moment of the uncovered light (a CCF form exists, not used here), without exposure averaging
  • Regenerate: node test/examples_real.test.mjs && node test/hd80606rm.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
λ [°]57 +17 −1242 ± 8Hébrard et al. 2010 (A&A 516, A95), from the full SOPHIE transit of 2010 January 14, which is not in this file; as quoted in test/examples_real.test.mjs
v sin i★ [km/s]1.47 +0.24 −0.17about 1.7as given in the task notes of test/hd80606rm.test.mjs (taken as ± 0.3); not re-read from a paper here

TOI-1873, TESS + NEID (real data; planet d a candidate) (toi1873.json) ​

  • Data: TESS SPOC light curves of 13 sectors as extracted and detrended in EXOTIC’s examples/toi-3planet, windows around every transit of b, c and d binned to 10 min (2876 points, 53 epochs); 37 NEID velocities of the published pyaneti analysis (public/examples/toi1873/README.md).
  • Terms: TESS data are public (MAST). The NEID velocities and the pyaneti results were shared with the owner by collaborators; no licence is stated and they may be unpublished: decide before the repository goes public.
  • Cite: Ricker et al. 2014; Barragán et al. 2019; Zellem et al. 2020
  • Stored run: 19 free, 1,500,000 likelihood calls, ln Z = 16179.51 ± 0.27, NOT converged, 7 sampler warnings; stored 2026-11-16 with version 0.2.0
  • Outcome: ephemerides, inclinations and both masses agree with the published two-planet fit within 0.6σ (c’s mass 0.03σ, b’s −0.4σ); the radius ratios differ by −1.5σ (b) and +1.6σ (c). Planet d: N = 7 (128.6 d) holds 85% of the posterior, 900 d 0%; its mass is 13 +8 −7 M⊕. The run stopped at its budget of 1.5 million calls with 2.6% of the evidence possibly left (NOT converged): read ln Z as a lower limit
  • Regenerate: node test/examples_real.test.mjs # ESL_FULL=1 reruns the sampler (about 2 hours)
QuantityThis codePublished or referenceSource of the comparison value
P b [d]7.005718 ± 0.0000137.0057048 +0.0000265 −0.0000208EXOTIC/examples/toi-3planet/2p_circ_prod_params.dat (pyaneti, Barragán et al. 2019), as typed in test/examples_real.test.mjs
Rp/R★ b0.02539 +0.00062 −0.000660.02683 ± 0.00074same
i b [°]88.66 +0.16 −0.1588.91 +0.45 −0.43same
Mp b [M Jup]0.0243 +0.0078 −0.00760.0285 ± 0.0075 (9.07 ± 2.39 M⊕)same
P c [d]24.933989 +0.000059 −0.00006324.9339464 ± 0.000055same
Rp/R★ c0.04167 +0.00088 −0.000830.03979 +0.00085 −0.00082same
i c [°]89.66 +0.16 −0.1289.72 +0.19 −0.24same
Mp c [M Jup]0.088 ± 0.0140.0871 ± 0.0158 (27.7 ± 5.0 M⊕)same
Mp d [M Jup]0.042 +0.026 −0.022none (candidate)EXOTIC’s own three-planet fit gives 17.6 ± 7.1 M⊕ (examples/toi-3planet/README.md); not a publication

toy-1873: three planets, light curves and velocities (synthetic) (toy1873.json) ​

  • Data: Synthetic system of EXOTIC’s examples/multi_planet (modelled on TOI-1873; the real data are the row above): three light-curve files and 40 velocities with injected truth.
  • Terms: Synthetic; from the EXOTIC repository.
  • Cite: Zellem et al. 2020
  • Stored run: 5 free, 18,641 likelihood calls, ln Z = 1725.88 ± 0.14, converged; stored 2026-11-16 with version 0.2.0
  • Outcome: reproduced (identical χ² at the injected values)
  • Regenerate: node test/exotic.test.mjs && node test/session.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
Three-planet toy system: χ² of each light curve at the injected values45.11, 88.80 and 190.58 in both codestest/exotic.test.mjs (Validation: EXOTIC’s own examples, reproduced)

WASP-43 b, full-orbit phase curve at 4.5 µm (phasecurve.json) ​

  • Data: One Spitzer IRAC channel 2 visit (BJD 2458738.3 to 2458739.4), already detrended, from github.com/pearsonkyle/Exoplanet-Mapping (commit 3af871c); binned to 120 s here.
  • Terms: That repository has no licence file; it is the site owner’s repository and its README asks for its white paper to be cited (public/examples/phasecurve/README.md). Spitzer data are public (IRSA).
  • Cite: Stevenson et al. 2017; Hellier et al. 2011
  • Stored run: 6 free, 80,241 likelihood calls, ln Z = 4433.49 ± 0.20, converged; stored 2026-11-16 with version 0.2.0
  • Outcome: reproduces the source repository’s own fit; the light curve was detrended jointly with that fit, so this is not an independent check, and no value from a journal article is compared
  • Regenerate: node test/make_phasecurve_examples.mjs <Exoplanet-Mapping checkout> && node test/phasemodel.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
T mid [BJD]2458738.834829 +0.000072 −0.0000812458738.834783 ± 0.000070header of the Exoplanet-Mapping file, as recorded in public/examples/phasecurve/README.md
eclipse depth [ppm]3849 ± 0.000000 (from the medians)3810 ± 100same
hot-spot offset [°]1.1 ± 1.31.6 ± 1.4same

WASP-43 b, eclipse map of degree 1 (eclipsemap.json) ​

  • Data: The visit above plus windows around four more eclipses and two transits from two other Spitzer visits (same repository).
  • Terms: As the row above.
  • Cite: Stevenson et al. 2017; Hellier et al. 2011; Luger et al. 2019
  • Stored run: 8 free, 64,864 likelihood calls, ln Z = 9635.36 ± 0.22, converged, 1 sampler warning; stored 2026-11-16 with version 0.2.0
  • Outcome: no published value to compare; internal check only (χ² through the likelihood equals the model used directly)
  • Regenerate: node test/phasemodel.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
Fp/F★0.1513 +0.0029 −0.0031noneno published map of these data is recorded in this repository
y1,−1 (north-south)0.11 +0.22 −0.24noneunconstrained: the posterior fills its prior

Simulated eclipse map of degree 2 (WASP-18 b orbit) (eclipsemap-sim.json) ​

  • Data: Simulated by this code: three eclipses and a transit, 60 ppm per 60 s, seed 20261115; orbit of WASP-18 b.
  • Terms: Generated here; no third-party data.
  • Cite: Coulombe et al. 2023
  • Stored run: 9 free, 120,675 likelihood calls, ln Z = 9749.56 ± 0.26, converged, 1 sampler warning; stored 2026-11-16 with version 0.2.0
  • Outcome: injected values recovered within the posterior (a self-consistency test: the data were made with this model)
  • Regenerate: node test/make_phasecurve_examples.mjs <Exoplanet-Mapping checkout>
QuantityThis codePublished or referenceSource of the comparison value
Fp/F★0.35977 +0.00101 −0.000980.36 (injected)notes of the session file
y1,00.417 +0.054 −0.0460.42 (injected)same
y1,10.0711 +0.0034 −0.00350.07 (injected)same
y1,−10.231 +0.045 −0.0620.15 (injected)same
y2,0-0.093 +0.062 −0.064−0.06 (injected)same
y2,2-0.005 ± 0.0960.04 (injected)same

Simulated planet with a moon, 12 transits (exomoon.json) ​

  • Data: Simulated by this code (public/examples/exomoon/make_example.mjs, seed 20261115): a Jupiter-size planet at 45 d with a moon, 250 ppm.
  • Terms: Generated here; no third-party data.
  • Stored run: 6 free, 62,950 likelihood calls, ln Z = 28515.61 ± 0.25, converged; stored 2026-11-16 with version 0.2.0
  • Outcome: injected values recovered (self-consistency; no real exomoon exists to reproduce)
  • Regenerate: node public/examples/exomoon/make_example.mjs && node test/exomoon.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
Rm/R★0.02017 ± 0.000270.02 (injected)notes of the session file
moon period [d]3.69996 +0.00012 −0.000133.7 (injected)same

Simulated hot Jupiter with a companion near 2:1, 72 mid-transit times (nbody.json) ​

  • Data: Simulated with this project’s IAS15 integrator (public/examples/nbody/make_example.mjs).
  • Terms: Generated here; no third-party data.
  • Stored run: 3 free, 8,804 likelihood calls, ln Z = 414.68 ± 0.12, converged; stored 2026-11-16 with version 0.2.0
  • Outcome: injected mass recovered; the transiting planet’s T0 and P are profiled, not sampled
  • Regenerate: node public/examples/nbody/make_example.mjs && node test/nbody.test.mjs && node test/nbodymodel.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
companion mass [M Jup]0.1034 +0.0069 −0.00650.100 (injected)public/examples/nbody/truth.json

TOI-1873, linear ephemerides and O − C of 50 TESS mid-times (real data) (toi1873-oc.json) ​

  • Data: The mid-transit times of EXOTIC’s per-transit fits of the TESS light curves (examples/toi-3planet, toi1873_transit_fits.json): 39 of b, 8 of c, 3 of d (public/examples/toi1873/transit_times.csv).
  • Terms: TESS data are public (MAST); the times are EXOTIC’s measurements, no licence stated.
  • Cite: Ricker et al. 2014; Zellem et al. 2020
  • Stored run: 6 free, 21,265 likelihood calls, ln Z = 151.00 ± 0.18, converged; stored 2026-11-16 with version 0.2.0
  • Outcome: the periods equal EXOTIC’s on the same times (0.00σ and 0.01σ; b is 1.6σ from pyaneti’s, c 0.4σ). No periodicity in the O − C: strongest peaks at false-alarm probabilities of 25% (b) and 80% (c), as EXOTIC also finds
  • Regenerate: node test/examples_real.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
P b [d]7.005755 +0.000018 −0.0000197.0057545 ± 0.0000182 (EXOTIC, same times)toi1873_transit_fits.json, weighted least squares; pyaneti from the light curves: 7.0057048
P c [d]24.933974 +0.000055 −0.00005224.9339738 ± 0.0000528 (EXOTIC, same times)same; pyaneti: 24.9339464

Kepler-88 (KOI-142), the non-transiting planet c weighed by the transit timing of b (real data) (kepler88.json) ​

  • Data: 119 mid-transit times of Kepler-88 b (two flagged outliers left out) from the Kepler TTV catalogue of Holczer et al. 2016 (ApJS 225, 9; VizieR J/ApJS/225/9, table3, rows of KOI-142.01; public/examples/kepler88/README.md). N-body model, b’s T0 and P profiled, c at i = 87°, both eccentricity vectors free; 8.66 min of timing noise added in quadrature (the catalogue errors, given to one figure, are too small for this model).
  • Terms: VizieR catalogue data (CDS terms, https://cds.unistra.fr/vizier-org/licences_vizier.html); Kepler data are public (MAST).
  • Cite: Borucki et al. 2010
  • Stored run: 7 free, 943,766 likelihood calls, ln Z = 404.60 ± 0.35, converged, 4 sampler warnings; stored 2026-11-16 with version 0.2.0
  • Outcome: c’s mass, 0.671 ± 0.006 M Jup from b’s timing alone, agrees with Weiss et al. 2020 (−0.2σ) and Nesvorný et al. 2013 (+1.4σ); its osculating period is 22.3349 ± 0.0007 d against Nesvorný’s osculating 22.3395 (−2.4σ; Weiss’s 22.265 is defined differently and not comparable); both eccentricities (0.0558, 0.0572) match the published 0.0559 and 0.0563 to 0.0572. The uncertainties rest on the 8.66 min of timing noise added to the catalogue errors
  • Regenerate: node public/examples/kepler88/make_example.mjs && node tools/build_example_results_node.mjs kepler88 && node test/examples_real.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
Mp c [M Jup]0.6706 +0.0062 −0.00580.674 ± 0.016 (Weiss et al. 2020); 0.626 +0.029 −0.033 (Nesvorný et al. 2013)NASA Exoplanet Archive rows in this repository’s snapshot (public/catalog), as typed in test/examples_real.test.mjs
P c [d]22.33488 +0.00070 −0.0007222.26492 ± 0.00067 (Weiss et al. 2020); 22.3395 +0.0019 −0.0017 (Nesvorný et al. 2013)same; this fit’s period is osculating at BJD 2454954.5, so it compares with Nesvorný’s osculating value

Simulated pair near 2:1, masses from timing alone (ttvpair.json) ​

  • Data: Simulated with this project’s IAS15 integrator (public/examples/ttvpair/make_example.mjs, seed 20261116): every transit of two warm Saturns over 1460 d, 113 times.
  • Terms: Generated here; no third-party data.
  • Cite: Lithwick, Xie & Wu 2012
  • Stored run: 4 free, 17,777 likelihood calls, ln Z = 551.81 ± 0.19, converged; stored 2026-11-16 with version 0.2.0
  • Outcome: both injected masses recovered (−0.3σ and +0.1σ) with circular orbits held; no velocities in the fit
  • Regenerate: node public/examples/ttvpair/make_example.mjs && node test/examples_real.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
Mp b [M Jup]0.1392 +0.0026 −0.00280.14 (injected)public/examples/ttvpair/truth.json
Mp c [M Jup]0.0901 ± 0.00100.09 (injected)same

Mid-transit times: EXOTIC’s 40-time list and its WASP-12 b list ​

  • Data: public/examples/ephemeris/transit_times.csv (40 simulated times, written from the arrays of EXOTIC’s examples/ephemeris; they run to 2028 and their observatory labels are made up) and wasp12_decay_times.csv (457 real times of WASP-12 b compiled in EXOTIC’s examples/ephemeris/fit_decay.py from Exoplanet Watch, ExoClock and the Exoplanet Transit Database; the file labels every row “literature”).
  • Terms: No licence stated in this repository; compiled in the EXOTIC repository from the sources labelled in each row.
  • Cite: Zellem et al. 2020; Leonardi et al. 2024
  • Outcome: the conversion dP/dE to dP/dt is reproduced on published numbers; a fit of the 457-time list is NOT compared with a published decay rate
  • Regenerate: node test/ephemeris.test.mjs && node test/periodogram.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
Orbital decay of WASP-12 b from the published dP/dE−30.74 ± 2.69 ms/yr here, −30.72 ± 2.67 publishedtest/ephemeris.test.mjs (Validation: Evidence, ephemerides and geometry)

HAT-P-32 b, 142 MicroObservatory frames (image photometry) ​

  • Data: EXOTIC’s sample data set (EXOTIC_sampledata-1.0.0, HatP32Dec202017), kept in test/data as cut-outs and not served by the site.
  • Terms: The sample data state no licence (README); for that reason the frames are not in public/.
  • Cite: Zellem et al. 2020
  • Outcome: timing agrees; Rp/R★ differs by 2.3σ, mostly through the choice of comparison star. Compared with EXOTIC’s reduction, not with a published light curve
  • Regenerate: node test/exotic_photometry.test.mjs
QuantityThis codePublished or referenceSource of the comparison value
Rp/R★0.1452 ± 0.0032 NOT REPRODUCED0.1562 ± 0.0034 (EXOTIC 4.3.1 on the same frames)README, "Photometry"; printed by test/exotic_photometry.test.mjs
T mid − 24581070.71311 ± 0.001060.71359 ± 0.00108 (EXOTIC 4.3.1)same
residual scatter0.496%0.573% (EXOTIC 4.3.1)same

Third-party data in the site and their terms ​

As far as the files in this repository state them. The project itself has no licence file yet.

DataWhereTerms as stated in the repositoryCite
NASA Exoplanet Archive (ps, pscomppars, TOI tables)public/catalog/, public/catalog-table/No licence text stored; the archive’s acknowledgement sentence is stored and shown.Akeson et al. 2013; Guerrero et al. 2021
exoplanet.eupublic/catalog-eu/CC BY 4.0, as stated in the site’s footer and stored in the snapshot; acknowledgement shown under the table.Schneider et al. 2011
PHOENIX-ACES specific intensities (derived coefficient tables only)src/data/ld_table.js, src/data/ld_spectral.jsNo licence stated in this repository.Husser et al. 2013
Star catalogue for the sky view (XHIP through d3-celestial)public/sky/d3-celestial data: BSD 3-Clause, © 2015 Olaf Frohn (stored in each file). No licence stated for XHIP itself.Anderson & Francis 2012
TESS and Kepler target pixel files (test fixtures; one TESS cut-out served as a demo)public/photometry/, test/data/No licence stated in this repository; public mission data from MAST.Ricker et al. 2014; Borucki et al. 2010
Land outlinepublic/observatories/land-110m.jsonNatural Earth, public domain (stored in the file).

28. How to cite and report ​

A result from this software can be reported like any other fit, provided the paper states: the software and its version; the model components and the limb-darkening law with the source of its coefficients; every free parameter with its prior, and what was held fixed (in particular a/R⋆ and the stellar mass and radius); the sampler settings and the outcome (convergence, likelihood calls, effective samples, ln⁡Z with its uncertainty, any sampler warning); the time standard of the data; and where the session file with the posterior draws is archived.

The function methodsParagraph(session) in src/core/methodsText.js writes a draft of that paragraph from a session file, with the BibTeX of exactly the references its model components need:

js
import { methodsParagraph } from './src/core/methodsText.js';
const { text, bibtex, keys, notes } = methodsParagraph(session, { ldSource, sliceUsed });

Its output for the stored example hd80606-joint.json (a draft to adapt; citation keys in square brackets match the BibTeX below):

We fitted the light curve (899 points; TESS), 264 radial velocities from 7 instruments (ELODIE, HIRES_k, HRS_1, HRS_2, HIRES_j, SOPHIE, APF), 4 mid-transit times and 1 mid-eclipse time of HD 80606 jointly with Exoplanet Science Lab 0.2.0 [esl]. The transit model is the fraction of stellar flux behind the planet's disc for the four-coefficient limb-darkening law of [claret2000], evaluated as a line integral around the boundary of the occulted region (Green's theorem), with mid-transit defined as inferior conjunction; the software's test suite holds this flux within 1e-10 of a quadrature reference. The limb-darkening coefficients were held fixed at (0.49387, -0.072946, 0.45785, -0.21582) for the band recorded as "TESS (approx.)". The flux normalisation was marginalised analytically over a uniform prior U(0, 2) instead of being sampled. The 2 min exposures are short compared with ingress, so the model was evaluated at the mid-exposure times. Orbits are Keplerian and eccentric; Kepler's equation is solved to machine precision from the starting value of [danby1987]. The scaled semi-major axis a/R_star was not a free parameter: for each planet it follows from Kepler's third law with the stellar mass and radius, the period and the planet mass, so that all data constrain one orbit. The stellar radial velocity is the sum over planets of K [cos(omega + nu) + e cos(omega)], with K from the planet mass, the stellar mass, the period, the eccentricity and the inclination by the two-body relation. Each instrument has its own zero point, marginalised analytically over a uniform prior of width 200000 m/s, and a jitter term added in quadrature to its uncertainties (held at zero); the likelihood includes the ln(sigma) normalisation. Mid-transit and mid-eclipse times were compared with the linear ephemeris of the same orbit. The 7 free parameters and their priors were: Rp/R★ ~ U(0.01, 0.2); i [°] ~ U(87, 90); P [d] ~ U(111.43633, 111.43717); T mid [d] ~ U(2458887.8757, 2458888.2757); e ~ U(0.93096, 0.93216); ω [°] ~ U(-59.04, -58.62); Mp [M Jup] ~ U(4.172, 4.202). Posterior samples and the Bayesian evidence were obtained by nested sampling [skilling2006] with the software's sampler, which follows UltraNest [buchner2021]: 400 live points, new points drawn from an MLFriends region [buchner2016, buchner2019] in whitened coordinates, stopping when the live points hold less than 0.01 of the evidence (seed 42; budget 1000000 likelihood evaluations). The run converged after 57440 likelihood evaluations with 2804 effective posterior samples and ln Z = 3359.94 +/- 0.22 (chi-square 1675.1 for 1168 measurements at the best fit). We report posterior medians with the 15.87 and 84.13 per cent quantiles as uncertainties; derived quantities were computed on posterior draws. Times were used as given in the data; the software does not convert between time standards.

To state or check by hand:

  • State where the fixed limb-darkening coefficients came from (the session records the band and the values, not their source).
  • The sampler raised 1 warning; report or resolve it: e and ω are -0.98 correlated: fit √e cos ω and √e sin ω instead, or fix one.
  • State the time standard of your data (for example BJD_TDB): the session does not record it.
  • Data in the session: light curve "TESS", 899 points, median uncertainty 0.00177 relative flux; radial velocity "ELODIE", 74 points, median uncertainty 13 m/s; radial velocity "HIRES_k", 39 points, median uncertainty 1.8 m/s; radial velocity "HRS_1", 7 points, median uncertainty 2.5 m/s; radial velocity "HRS_2", 16 points, median uncertainty 8.1 m/s; radial velocity "HIRES_j", 75 points, median uncertainty 1.07 m/s; radial velocity "SOPHIE", 48 points, median uncertainty 4.4 m/s; radial velocity "APF", 5 points, median uncertainty 2.33 m/s; mid-transit times "planet b", 4 points, median uncertainty 0.0049 d; mid-eclipse times "planet b", 1 points, median uncertainty 0.00025 d.
bibtex
@software{esl,
  author = {Pearson, Kyle A.},
  title = {{Exoplanet Science Lab}: transit, radial-velocity and timing analysis in the browser},
  version = {0.2.0},
  year = {2026},
  url = {https://exoplanet.science}
}

@article{claret2000,
  author = {Claret, A.},
  year = {2000},
  journal = {Astronomy \& Astrophysics},
  volume = {363},
  pages = {1081},
  url = {https://ui.adsabs.harvard.edu/abs/2000A%26A...363.1081C}
}

@article{danby1987,
  author = {Danby, J. M. A.},
  title = {{The solution of Kepler's equation, III}},
  year = {1987},
  journal = {Celestial Mechanics},
  volume = {40},
  pages = {303},
  doi = {10.1007/BF01235847}
}

@article{skilling2006,
  author = {Skilling, J.},
  title = {{Nested sampling for general Bayesian computation}},
  year = {2006},
  journal = {Bayesian Analysis},
  volume = {1},
  pages = {833},
  doi = {10.1214/06-BA127}
}

@article{buchner2021,
  author = {Buchner, J.},
  title = {{UltraNest - a robust, general purpose Bayesian inference engine}},
  year = {2021},
  journal = {Journal of Open Source Software},
  volume = {6},
  pages = {3001},
  doi = {10.21105/joss.03001},
  eprint = {2101.09604},
  archivePrefix = {arXiv}
}

@article{buchner2016,
  author = {Buchner, J.},
  title = {{A statistical test for Nested Sampling algorithms}},
  year = {2016},
  journal = {Statistics and Computing},
  volume = {26},
  pages = {383},
  doi = {10.1007/s11222-014-9512-y},
  eprint = {1407.5459},
  archivePrefix = {arXiv}
}

@article{buchner2019,
  author = {Buchner, J.},
  title = {{Collaborative Nested Sampling: Big Data versus Complex Physical Models}},
  year = {2019},
  journal = {Publications of the Astronomical Society of the Pacific},
  volume = {131},
  pages = {108005},
  doi = {10.1088/1538-3873/aae7fc},
  eprint = {1707.04476},
  archivePrefix = {arXiv}
}

Citing the software ​

Cite the software with its version (CITATION.cff at the repository root; BibTeX key esl), and the papers of the methods you used: the reference list below gives them, and docs/references.bib holds all of them. If the catalogue supplied your starting values, cite the original publications named by the provenance labels and the archive's own acknowledgement.

29. References ​

109 entries: every work cited by the methods and by the validation (src/validation/refs.js). Each was checked against the Crossref record of its DOI, arXiv, or the Astrophysics Source Code Library on 15 November 2026; entries that could not be checked in full say so. The key in brackets is the BibTeX key in docs/references.bib.

MIT licensed · Kyle A. Pearson