The Universal Bayesian Imaging Kit

Bayesian imaging of astrophysical measurement data shares universal properties across the electromagnetic spectrum: it requires probabilistic descriptions of possible images and spectra, and instrument responses. To unify Bayesian imaging, we present the Universal Bayesian Imaging Kit (UBIK). Currently, UBIK images data from Chandra, eROSITA, JWST, and ALMA. UBIK is based on information field theory (IFT), the mathematical theory of field inference, and on NIFTy, a package for numerical IFT. UBIK provides sky models that are instrument independent and instrument interfaces that share common parts of their response representations. It is open source, can provide spatio-spectral image cubes, jointly analyses data from several instruments, and separates diffuse emission, point sources, and extended emission regions.

💡 Research Summary

The paper introduces the Universal Bayesian Imaging Kit (UBIK), an open-source software framework designed to unify the Bayesian imaging of astrophysical data across the entire electromagnetic spectrum. The authors argue that while current astronomical imaging relies on instrument-specific software, the fundamental task—reconstructing an image of the sky from imperfect measurement data while properly quantifying uncertainties—is universal and should be addressed within a consistent probabilistic framework.

UBIK is built upon Information Field Theory (IFT), the mathematical theory of field inference, and its numerical implementation package, NIFTy. It allows users to construct a generative forward model of the measurement process. This model starts with latent Gaussian variables, which are non-linearly transformed into sky signal components (like diffuse emission fields and point source properties) with desired prior statistics. The final layer is a likelihood function that compares simulated detector data—incorporating effects like the point spread function, exposure variations, and noise (Poisson/Gaussian)—with the actual observed data.

To perform the high-dimensional inference required for imaging, UBIK employs Variational Inference (VI) methods, specifically Metric Gaussian VI (MGVI) and geometric VI (geoVI). MGVI approximates the posterior distribution with a Gaussian whose covariance is informed by the Fisher information metric. GeoVI improves upon this by constructing approximate Riemann normal coordinates, offering a more accurate approximation for non-Gaussian posteriors on curved manifolds. Both methods generate samples from the approximate posterior, enabling the calculation of summary statistics like mean maps and uncertainty estimates.

A key architectural feature of UBIK is the separation between instrument-independent sky models and instrument-specific interfaces. This modular design promotes reusability and enables the joint analysis of data from multiple telescopes. Currently, UBIK provides ready-to-use interfaces for several major facilities: the Chandra and eROSITA X-ray observatories, the James Webb Space Telescope (JWST) in the infrared, and the Atacama Large Millimeter/submillimeter Array (ALMA) in the radio regime. This allows UBIK to produce spatio-spectral image cubes and separate different emission components (diffuse, point-like, extended).

In summary, UBIK represents a significant step towards standardizing Bayesian imaging in astrophysics. By providing a unified, open-source toolkit based on a rigorous mathematical foundation, it aims to improve reproducibility, enable sophisticated multi-instrument analyses, and offer a coherent framework for uncertainty quantification in astronomical image reconstruction.