What's in Your Transit? Towards Reliably Getting $5\times$ More Science from Exoplanet Transit Data

TL;DR

The paper addresses why transit-depth uncertainties in exoplanet observations systematically exceed photon-noise expectations, introducing the amplification factor $A = \sigma_D/\sigma_{\rm ref}$ and showing that correlations with limb-darkening coefficients (LDCs) largely drive this TDPP. Through injection--retrieval experiments, perturbation analyses, and sensitivity mappings, the authors demonstrate that current LDC priors and limb-darkening parameterizations inflate $A$ (often to $\sim10$), bias $D$, and impede atmospheric inferences. They advocate using multiple, carefully chosen limb-darkening parameterizations (notably a 3rd-order polynomial and a 4th-order nonlinear LDL) and leveraging model-informed priors on LDCs to substantially reduce $A$ (potentially to $\sim2$ with significant model improvements) and biases. This work highlights that advances in stellar atmosphere modeling and LDL orthogonality could yield a 5× higher scientific return from transit data, enabling the same conclusions with up to 25× fewer transits.

Abstract

Exoplanetary science heavily relies on transit depth ($D$) measurements. Yet, as instrumental precision increases, the uncertainty on $D$ appears to increasingly drift from expectations driven solely by photon-noise. Here we characterize this shortfall (the Transit-Depth Precision Problem, TDPP), by defining an amplification factor, $A$, quantifying the discrepancy between the measured transit-depth uncertainty and the measured baseline scatter on a same time bin size. While in theory $A$ should be $\sim\sqrt{3}$, we find that it can reach values $\gtrsim$10 notably due to correlations between $D$ and the limb-darkening coefficients (LDCs). This means that (1) the performance of transit-based exoplanet studies (e.g., atmospheric studies) can be substantially improved with reliable priors on LDCs and (2) low-fidelity priors on the LDCs can yield substantial biases on $D$--potentially affecting atmospheric studies due to the wavelength-dependence of such biases. For the same reason, biases may emerge on stellar-density and planet-shape/limb-asymmetry measurements. With current photometric precisions, we recommend using a 3$^{\rm rd}$-order polynomial law and a 4$^{\rm th}$-order non-linear law, as they provide an optimal compromise between bias and $A$, while testing the fidelity for each parametrization. While their use combined with existing LDC priors (10-20% uncertainty) currently implies $A\sim10$, we show that targeted improvements to limb-darkening models can bring $A$ down to $\sim2$. Improving stellar models and transit-fitting practices is thus essential to fully exploit transit datasets, and reliably increasing their scientific yield by $5\times$, thereby enabling the same science with up to $25\times$ fewer transits.

Figures (4)

• Figure 1: Expectations vs reality in constraining an exoplanet transit depth.Top left: Mock transit generated with the properties listed in \ref{['tab:tab1']}. Black points show the light curve with 600 parts per million (ppm) of injected Gaussian white noise. Blue curves are model realizations drawn from the posterior distribution of $D$ obtained by fitting this dataset. Annotated arrows indicate the quantities relevant to the TDPP. Right: Illustration of the TDPP. Each coloured box plot set corresponds to injection--retrieval tests done with a different limb-darkening prescription. For each set the injection and retrieval models have the same functional form. Box plots summarize 10 MCMC retrievals with identical scatter but different noise realizations. The shaded horizontal line within the boxes marks the distribution's median, while the lower and upper box edges correspond to its first and third quartiles. Whiskers, denoting the minimum and maximum values, are calculated as $1.5\times$ the first and third quartile. The regimes and asymptotic behaviours of these injection--retrieval families are discussed in \ref{['sec:TDPP']}. The injected out-of-transit scatter ($\sigma_{\rm OOT}$) is measured in 120-min bins.
• Figure 2: Parameter perturbations on transit shape. Separate panels highlight the transit ingress/egress, and bottom. Symmetric "bumps" can be clearly seen at ingress and egress, driven by the orbital parameters. In contrast, fluctuations in the bottom are dominated by the LDCs. Perturbations beyond the photometric precision (16 ppm here) are due to underlying correlation, e.g., between $\rho_{\star}$ vs $\sqrt{e}\sin{\omega}$ (top-right). In practice, these correlations link all orbital parameters. The coloured contours indicate the $10, 30, 60$, and $90\%$ equi-probability levels.
• Figure 3: Revealing underlying network of correlations. Results of the sensitivity analysis. Left: Correlation matrix. The lower triangle reports correlation values derived from $\chi^2$ maps, with shading scaled to the absolute values. The upper triangle provides a circular representation of the same matrix, where circle size and shading scale with correlation strength, making the families of correlated parameters easier to distinguish. The transit depth is by far most strongly correlated with $u_1$, $u_2$, and $u_3$. Right: Network representation of the correlations. Parameters are grouped into families using the same colour scheme as in the matrix. Blue lines highlight the strongest links between parameters of different families.
• Figure 4: Towards optimal limb-darkening parametrization. Amplification factor (top) and transit depth bias (bottom) from our injection--retrieval tests. Results are shown as a function of the number of LDCs (x-axis) under different limb-darkening priors assumptions, ranging from uninformative uniform priors to Gaussian priors of varying widths. Polynomial LDLs are used for cases with 2-9 LDCs, while "4NLLD" refers to the fourth-order non-linear LDL. Since this is the law for injection, it yields the best performance. Priors informed by current stellar models are highlighted in blue (typical relative uncertainties between 10 and 20%). As in \ref{['fig:fig1']}, the black dashed horizontal line marks $A$'s theoretical limit, and box plots summarize the results from 10 MCMC retrievals with different noise realizations. Green vertical bands indicate the regimes considered acceptable: for the bias, results within $2\sigma$ of truth and for the amplification factor, $A<10$. As discussed in \ref{['sec:impact_prior_choice']}, given these acceptability criteria we highlight the optimal limb-darkening parameterizations with green x-axis labels, and the sub-optimal ones in red.