Inverse-problem versus principal component analysis methods for angular differential imaging of circumstellar disks

Abstract

Context. Circumstellar disk images have highlighted a wide variety of morphological features. Recovering disk images from high-contrast angular differential imaging (ADI) sequences is, however, generally affected by geometrical biases, leading to unreliable inferences of the morphology of extended disk features. Recently, two types of approaches have been proposed to recover more robust disk images from ADI sequences: iterative principal component analysis (I-PCA) and inverse problem (IP) approaches. Aims. We introduce mustard, a new IP-based algorithm specifically designed to address the problem of the flux invariant to rotation in ADI sequences – a limitation inherent to the ADI observing strategy – and discuss the advantages of IP approaches with respect to PCA-based algorithms. Methods. The mustard model relies on the addition of morphological priors on the disk and speckle field to a standard IP approach to tackle rotation-invariant signals in circumstellar disk images. We compared the performance of mustard, I-PCA, and standard PCA on a sample of high-contrast imaging data sets acquired in different observing conditions, after injecting a variety of synthetic disk models at different contrast levels. Results. Mustard significantly improves the recovery of rotation-invariant signals in disk images, especially for data sets obtained in good observing conditions. However, the mustard model inadequately handles unstable ADI data sets and provides shallower detection limits than PCA-based approaches. Conclusions. Mustard has the potential to deliver more robust disk images by introducing a prior to address the inherent ambiguity of ADI observations. However, the effectiveness of the prior is partly hindered by our limited knowledge of the morphological and temporal properties of the stellar speckle halo. In light of this limitation, we suggest that the algorithm could be improved by enforcing a data-driven prior based on a library of reference stars.