Impacts of high-contrast image processing on atmospheric retrievals

TL;DR

This study investigates how high-contrast image processing algorithms affect the spectral data of exoplanets and demonstrates that accounting for wavelength correlations improves the accuracy of atmospheric parameter retrievals.

Contribution

The paper systematically analyzes the impact of different image processing algorithms on spectral retrieval accuracy and emphasizes the importance of including wavelength covariance in Bayesian atmospheric retrievals.

Findings

Choice of processing algorithm significantly affects retrieval bias.

Including full covariance matrix improves parameter accuracy.

Full covariance inclusion is statistically favored in retrievals.

Abstract

Many post-processing algorithms have been developed in order to better separate the signal of a companion from the bright light of the host star, but the effect of such algorithms on the shape of exoplanet spectra extracted from integral field spectrograph data is poorly understood. The resulting spectra are affected by noise that is correlated in wavelength space due to both optical and data processing effects. Within the framework of Bayesian atmospheric retrievals, we aim to understand how these correlations and other systematic effects impact the inferred physical parameters. We consider three algorithms (KLIP, PynPoint and ANDROMEDA), optimizing the choice of algorithmic parameters using a series of injection tests into archival SPHERE and GPI data of the HR 8799 system. The wavelength-dependent covariance matrix is calculated to provide a measure of instrumental and algorithmic systematics. We perform atmospheric retrievals using petitRADTRANS on optimally extracted spectra to measure how these data processing systematics influence the retrieved parameter distributions. The choice of data processing algorithm and parameters significantly impact the accuracy of retrieval results, with the mean posterior parameter bias ranging from 1 to 3 $\sigma$ from the true input parameters. Including the full covariance matrix in the likelihood improves the accuracy of inferred parameters, and cannot be accounted for using ad hoc scaling parameters in the retrieval framework. Using the Bayesian information criterion and other statistical measures as a heuristic goodness-of-fit metrics, the retrievals including the full covariance matrix are favoured when compared to using only the diagonal elements.