Rainbow Multi-band Lightcurve Fit

• Rainbow multi-band lightcurve fit is a method that simultaneously models time-domain photometry across multiple bands by coupling a shared timing structure with band-specific parameters.
• It employs diverse algorithmic families such as template fitting, Fourier-based methods, physical SED and transit models, and neural emulators to capture intrinsic lightcurve features.
• This approach improves period recovery and parameter inference under sparse sampling while requiring precise calibration and uncertainty management to control degeneracies.

Searching arXiv for relevant papers on multi-band or “rainbow” light-curve fitting across astronomy domains. Search 1: query "multi-band lightcurve fit" OR "multiband light curve" rainbow astronomy arXiv A rainbow multi-band lightcurve fit is the simultaneous modeling of time-domain photometry across multiple passbands under a shared temporal or physical structure, with band-dependent observables treated as coupled rather than independent. In the literature, this coupling is realized in several distinct ways: a common period and phase with band-specific templates for pulsators, a shared Fourier basis with regularized band residuals for multiband periodograms, a wavelength–time surface defined by bolometric flux and temperature for transients, globally shared orbital geometry with wavelength-dependent depth and limb darkening for transits, and multivariate latent-state models or neural sequence models for stochastic or heterogeneous data (Stringer et al., 2019, VanderPlas et al., 2015, Russeil et al., 2023).

1. Conceptual scope and domain of use

The phrase denotes a methodological class rather than a single algorithm. In periodic-variable work, the central assumption is that all bands trace one underlying clock while amplitudes, mean magnitudes, and sometimes shape corrections vary by passband. This appears in RR Lyrae template fitting for Dark Energy Survey grizY data, in multiband periodograms for sparsely sampled surveys, in seven-band J-VAR RR Lyrae template analysis, and in multi-wavelength Fourier decompositions of RR Lyrae optical-to-infrared light curves (Stringer et al., 2019, Saha et al., 2017, Kulkarni et al., 3 Sep 2025, Das et al., 2018).

In transient work, rainbow fitting is explicitly a wavelength–time reconstruction problem. The “Rainbow” framework for supernova-like transients assumes that the spectral energy distribution is approximated by a blackbody and couples a parametric bolometric light curve to a parametric temperature evolution, yielding a 2-dimensional continuous surface across wavelength and time even when the number of observations in each filter is significantly limited (Russeil et al., 2023). A related but physically different application appears in SALT2-based Type Ia supernova fitting, where a 2D time–wavelength spectral surface is fit across filters simultaneously and returns rest-frame lightcurve parameters used for standardization (Konishi et al., 2011).

In exoplanet transit analysis, the same idea appears as joint inference over all instruments, epochs, and wavelength channels, with wavelength-independent orbital parameters shared globally and wavelength-dependent parameters such as $R_p(\lambda)/R_*$ and limb-darkening coefficients fitted per channel. TransitFit formalizes this structure for JWST, HST, TESS, and ground-based data, with simultaneous detrending and transit inference (Hayes et al., 2021).

In binary-star and stochastic-variability studies, rainbow fitting extends beyond explicit phase models. A neural-network emulator of PHOEBE simultaneously solves multiple-band light curves of contact binaries and infers physical parameters including $q$, $i$, $\Omega$, $f$, third light, and spot parameters across 19 passbands (Li et al., 9 Sep 2025). For stochastic sources such as quasars, a multivariate damped random walk in state-space form models irregularly sampled multi-filter light curves with heteroscedastic errors and evaluates the likelihood by Kalman filtering with worst-case $O(k^3 n)$ complexity (Hu et al., 2020).

2. Shared mathematical structure

Despite large differences in source class, rainbow fitting usually separates into shared parameters and band-dependent parameters. For RRab in DES, the forward model is a single-period, band-dependent template with a common phase across bands,

$$m_b(t) = \mu + M_b(\omega) + a\,\gamma_b(\omega t + \phi),$$

optionally extended by $R_b E(B-V)$ when dust is fitted explicitly. Here $\mu$ is the distance modulus, $\omega=1/P$ is the pulsation frequency, $q$0 is the band-specific mean absolute magnitude, $q$1 is the amplitude referenced to $q$2, $q$3 is the band-dependent template shape, and $q$4 is a global phase offset (Stringer et al., 2019).

The corresponding DES joint objective with pre-corrected dust is

$q$5

with $q$6 mag added in quadrature as a model-error term. This formulation is characteristic of rainbow template fitting: one shared clock, one shared phase reference, and explicit cross-band amplitude scaling embedded in the templates (Stringer et al., 2019).

In the multiband Lomb–Scargle formulation, the shared structure is expressed as a base Fourier model plus regularized band residuals,

$q$7

with the objective

$q$8

Tikhonov regularization drives most variability into the base model common to all bands, while fits for individual bands describe residuals relative to the base model (VanderPlas et al., 2015). The related hybrid algorithm of Lomb–Scargle plus Lafler–Kinman constructs

$q$9

and then sums $i$0 across asynchronously sampled bands to isolate the common fundamental frequency while remaining agnostic to band-dependent shape and phase differences (Saha et al., 2017).

For transient rainbow fitting under a blackbody assumption, the shared structure is an SED:

$i$1

and the passband-averaged flux is

$i$2

Rainbow then parameterizes $i$3 with a Bazin function and $i$4 with a logistic law, so that all filters are linked through a single thermal continuum (Russeil et al., 2023).

Transit fitting uses the same decomposition principle but with transit geometry instead of a pulsation or transient SED. TransitFit models each light curve as

$i$5

where $i$6 is shared and $i$7 varies with wavelength (Hayes et al., 2021).

3. Principal methodological families

The literature contains several recurrent algorithmic families.

Family	Core shared structure	Representative papers
Template/grid search	Common period or phase; band-specific templates, amplitudes, means	(Stringer et al., 2019, Kulkarni et al., 3 Sep 2025)
Fourier/periodogram	Shared Fourier basis or shared frequency across bands	(VanderPlas et al., 2015, Saha et al., 2017, Das et al., 2018)
Physical SED or transit model	Shared bolometric/temperature evolution or orbital geometry	(Russeil et al., 2023, Hayes et al., 2021)
Latent-state stochastic model	Shared multivariate OU process across filters	(Hu et al., 2020)
Neural emulation/embeddings	Shared multiband representation learned from data	(Li et al., 9 Sep 2025, Becker et al., 21 Jan 2025)

Template-fitting methods are particularly prominent for RR Lyrae. The DES pipeline restricts the period search to $i$8, alternates weighted least squares for $i$9 with a Gauss–Newton update for $\Omega$0, and retains the top three residual minima to expose near-degenerate aliases. The period–luminosity term is encoded as a quadratic in $\Omega$1, and the band shapes $\Omega$2 are derived empirically from SDSS Stripe 82 RRab as in Sesar et al. (2010) (Stringer et al., 2019). J-VAR uses the SDSS multiband RR Lyrae template library from Sesar et al. (2010) as well, but explicitly states that each filter was analyzed independently, with per-band period search and best-matching template selection rather than an enforced shared period (Kulkarni et al., 3 Sep 2025).

Fourier-based rainbow fitting ranges from purely periodic decomposition to multiband period detection. The RR Lyrae multi-wavelength Fourier analysis fits

$\Omega$3

with a shared period across bands and derived parameters $\Omega$4 and $\Omega$5 used to study wavelength trends, period–color relations, and amplitude–color relations (Das et al., 2018). The multiband periodogram instead uses a shared low-order base series plus lightly regularized band corrections to improve period recovery under sparse sampling (VanderPlas et al., 2015). The hybrid Lomb–Scargle/Lafler–Kinman method addresses the complementary problems of harmonics, subharmonics, and cadence aliases by combining $\Omega$6 and $\Omega$7 into a bandwise statistic and summing it over passbands (Saha et al., 2017).

Physically constrained forward models dominate transit and transient applications. Rainbow reconstructs a 2D surface across wavelength and time from a blackbody SED, while TransitFit performs nested-sampling inference with batman transit models, per-channel limb darkening, and simultaneous detrending that conserves transit depth at mid-transit (Russeil et al., 2023, Hayes et al., 2021).

Learned multiband models replace explicit templates with emulators or sequence encoders. The PHOEBE-based contact-binary model first pretrains luminosity and light-curve networks on V band and then fine-tunes per band through transfer learning, with inference performed by a two-stage MCMC around the neural forward model (Li et al., 9 Sep 2025). “Multiband Embeddings of Light Curves” builds one LSTM stack per band and a central multiband LSTM that fuses translated band embeddings, colors, and multiband time differences, enabling partial-curve inference without recomputing the entire sequence (Becker et al., 21 Jan 2025).

4. Representative implementations and quantitative performance

Quantitative results vary strongly with source class, cadence, and photometric regime, so reported metrics are method-specific rather than universal.

For sparse DES RRab identification, the joint grizY template fit recovers 89% of RRab periods within 1% of reference on Stripe 82 and, at Random Forest score $\Omega$8, reaches purity $\Omega$9, completeness $f$0, and AUC $f$1. From 8026 candidates with score $f$2, visual vetting yields 5783 RRab candidates, $f$3 of them previously unidentified. The practical runtime is $f$4–$f$5 minutes per light curve and $f$6k CPU hours for $f$7 light curves (Stringer et al., 2019).

For multiband period determination under sparse and irregular sampling, the hybrid $f$8 method identifies the period inferred from the maximum of the co-added $f$9 as the best estimable in $O(k^3 n)$0 of more than 20,000 putative variables, and for $O(k^3 n)$1 RR Lyrae compared to OGLE, more than 98% of periods match within the precision allowed by the 2-year baseline (Saha et al., 2017). The multiband Lomb–Scargle framework, in turn, is reported to find the true period among the top five peaks in $O(k^3 n)$2–$O(k^3 n)$3 of sparse multiband RR Lyrae tests, whereas single-band methods drop to $O(k^3 n)$4 under the same sparse conditions (VanderPlas et al., 2015).

For contact binaries, the neural PHOEBE emulator attains light-curve $O(k^3 n)$5 values across 19 filters of about $O(k^3 n)$6–$O(k^3 n)$7 in mean and about $O(k^3 n)$8–$O(k^3 n)$9 in median on synthetic data. A three-band dataset with 729 points is processed in about 82 s on a workstation, whereas the equivalent PHOEBE analysis took about 4.8 days. Applied to OGLE, the method reported physical parameters for 3,541 systems after quality cuts from an initial 5,332 preselected objects (Li et al., 9 Sep 2025).

For exoplanet transmission spectroscopy, TransitFit recovered a transit depth for WASP-91b to a precision of 111 ppm from 26 TESS transit epochs, fit 180 TESS transits of WASP-126b to test TTV claims, modeled HST observations of WASP-43b with custom detrending, and jointly fit 88 JWST NIRISS and 38 HST WFC3 spectroscopic light curves for WASP-96b in batched mode (Hayes et al., 2021).

For thermal-transient rainbow fitting, the Rainbow framework yields equivalent goodness of fit for SN II and up to 75% better goodness of fit for SN Ibc relative to the Monochromatic approach. In multi-class classification using best-fit values as a parameter space, overall median accuracy increases from 81.9% to 88.4%, and in rising-only classification from 50.9% to 59.5% (Russeil et al., 2023).

For stochastic multiband variability, the state-space multivariate damped random walk offers a worst-case likelihood cost of $$m_b(t) = \mu + M_b(\omega) + a\,\gamma_b(\omega t + \phi),$$0 instead of the $$m_b(t) = \mu + M_b(\omega) + a\,\gamma_b(\omega t + \phi),$$1 maximum cost of stacking all bands into one univariate Gaussian-process vector. In the two-band time-delay application to Q0957+561, the joint fit yields a posterior mean time delay of 414.324 days with standard deviation 2.307 days, whereas single-band fits remain inconsistent and more weakly constrained (Hu et al., 2020).

5. Calibration, uncertainties, and degeneracy control

A defining feature of rainbow fitting is that cross-band coupling is useful only if calibration and uncertainty treatment are explicit. DES combines per-epoch pipeline uncertainties with FGCM zeropoint errors in quadrature, rescales them in each HEALPix region and band through a quadratic fit to $$m_b(t) = \mu + M_b(\omega) + a\,\gamma_b(\omega t + \phi),$$2 versus magnitude, adds a model-error term in the objective, and pre-corrects light curves for extinction using SFD98 maps, the Fitzpatrick law with $$m_b(t) = \mu + M_b(\omega) + a\,\gamma_b(\omega t + \phi),$$3, and Schlafly & Finkbeiner adjustments (Stringer et al., 2019).

Period ambiguity and sparse-phase coverage are treated by astrophysical priors as well as pure statistics. In DES RRab fitting, the period search is restricted to the RRab range, the feature $$m_b(t) = \mu + M_b(\omega) + a\,\gamma_b(\omega t + \phi),$$4 is the von Mises–Fisher concentration parameter of phases in the folded light curve, and distances in $$m_b(t) = \mu + M_b(\omega) + a\,\gamma_b(\omega t + \phi),$$5 from the Oosterhoff I relation and the Oosterhoff I/II boundary are used as classification features (Stringer et al., 2019). The hybrid periodogram further introduces a frequency-dependent threshold $$m_b(t) = \mu + M_b(\omega) + a\,\gamma_b(\omega t + \phi),$$6 built from sampling-window and scrambled-data baselines, while the multiband Lomb–Scargle model uses Tikhonov regularization to concentrate variability in a common base model and to resolve rank deficiencies caused by bandwise offsets (Saha et al., 2017, VanderPlas et al., 2015).

Degeneracy control is equally central in binary and transit problems. For contact binaries, multi-band photometry is specifically used to reduce degeneracies among inclination $$m_b(t) = \mu + M_b(\omega) + a\,\gamma_b(\omega t + \phi),$$7, mass ratio $$m_b(t) = \mu + M_b(\omega) + a\,\gamma_b(\omega t + \phi),$$8, limb darkening, third light $$m_b(t) = \mu + M_b(\omega) + a\,\gamma_b(\omega t + \phi),$$9, and spot parameters. Blue filters emphasize temperature and spot contrast, while red filters stabilize geometry, and the recommended workflow tests phase shifts of 0 and 0.5 with $R_b E(B-V)$0 on and off before selecting the best solution by combined $R_b E(B-V)$1 across bands (Li et al., 9 Sep 2025). In TransitFit, wavelength-dependent limb darkening is regularized by LDTk-conditioned priors, and systematics are modeled per instrument and epoch through depth-conserving polynomials or custom forms such as the HST WFC3 ramp-plus-scan model (Hayes et al., 2021).

For stochastic sources, heteroscedastic errors enter the observation covariance directly. The multivariate OU state-space model treats observations at epoch $R_b E(B-V)$2 as

$R_b E(B-V)$3

and missing bands are handled by subsetting the state and covariance matrices at each epoch. This formulation is one reason the Kalman-filter likelihood scales with the number of bands and epochs rather than with the cube of the full stacked-data length (Hu et al., 2020).

6. Limitations, controversies, and physical non-equivalence across bands

A recurrent misconception is that a rainbow fit implies identical light-curve morphology in every band. The available studies show the opposite. Multi-wavelength RR Lyrae analyses find that Fourier amplitude parameters decrease with wavelength while Fourier phase parameters increase with wavelength at a given period (Das et al., 2018). High-precision observations of SDSS J015450+001501 show that the time of maximum light in H and K lags the maxima at shorter wavelengths by approximately 0.25 in phase, that $R_b E(B-V)$4 varies with amplitude about 0.2 mag while $R_b E(B-V)$5 shows no measurable variation, and that static atmosphere models cannot reproduce the observed color–color hysteresis and phase-dependent inter-band lags (Szabó et al., 2013).

Another misconception is that joint fitting is always present whenever multiple bands are analyzed. J-VAR explicitly derives independent parameter estimates across seven optical bands using mapped SDSS templates, and the paper itself notes that a shared-period, multi-band fit would improve period precision and cross-band phase consistency; this indicates that independent per-band fitting and true rainbow fitting are distinct methodological choices (Kulkarni et al., 3 Sep 2025).

Model validity is source-dependent. The DES RR Lyrae method assumes single-mode RRab with fixed, band-dependent shapes and does not model Blazhko amplitude or phase modulations, multi-mode behavior, or RRc stars (Stringer et al., 2019). The contact-binary neural emulator is designed for W UMa-type contact binaries with common envelopes and synchronous rotation; semi-detached or detached systems lie outside its assumptions, and very large or multiple spot groups may exceed the single-spot parametrization (Li et al., 9 Sep 2025). TransitFit’s core implementation does not natively include Gaussian processes, although a GP likelihood can be supplied through a custom systematics model or likelihood replacement (Hayes et al., 2021). Rainbow for transients assumes a blackbody SED, does not include Milky Way or host extinction or K-corrections in the baseline implementation, and is therefore best matched to cases where that approximation is adequate in the observed passbands (Russeil et al., 2023).

These limitations have a common implication: rainbow fitting is most powerful when the shared structure is physically or statistically justified. A common period, common orbital geometry, common thermal continuum, or common latent stochastic process can sharply reduce effective dimensionality and improve inference under sparse, irregular sampling. When the assumed shared structure is too rigid, however, residuals, host-dependent standardization terms, phase lags, or narrow-band morphology changes become scientifically informative rather than merely nuisances (Konishi et al., 2011, Szabó et al., 2013).