Joint Bayesian calibration and map-making for intensity mapping experiments

Zheng Zhang 1, Philip Bull 1,2, Mario G. Santos 2, and Ainulnabilah Nasirudin 1 1 Jodrell Bank Centre for Astrophysics, University of Manchester, Manchester, M13 9PL, United Kingdom 2 Department of Physics and Astronomy, University of Western Cape, Cape Town 7535, South Africa · 2025

Read Joint Bayesian calibration and map-making for intensity mapping experiments by Zheng Zhang 1, Philip Bull 1,2, Mario G. Santos 2, and Ainulnabilah Nasirudin 1 1 Jodrell Bank Centre for Astrophysics, University of Manchester, Manchester, M13 9PL, United Kingdom 2 Department of Physics and Astronomy, University of Western Cape, Cape Town 7535, South Africa (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-10992.txt.

Introduction

Zheng Zhang and three co-authors build a Bayesian pipeline for turning raw radio telescope data into sky maps in this 2025 paper. Line-intensity mapping charts the universe by the combined glow of unresolved galaxies in a spectral line. Single-dish telescopes fighting for that faint signal suffer wandering gains, thermal drift, and correlated low-frequency noise. Calibration and map-making contaminate each other if done separately.

Their method does both jointly with Gibbs sampling, exposing the full posterior over calibration and sky together. The data model follows the radiometer equation, tying noise to system temperature instead of assuming it fixed. A new noise model for the low frequencies avoids fake periodic patterns, common in older treatments. Efficient numerics, including a Levinson solver and a polynomial emulator, keep it affordable. Simulations modeled on the MeerKLASS survey demonstrate the method, which transfers to other single-dish programs.

Catalog record

CREATORZheng Zhang 1, Philip Bull 1,2, Mario G. Santos 2, and Ainulnabilah Nasirudin 1 1 Jodrell Bank Centre for Astrophysics, University of Manchester, Manchester, M13 9PL, United Kingdom 2 Department of Physics and Astronomy, University of Western Cape, Cape Town 7535, South Africa
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE99 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-10992

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Joint Bayesian calibration and map-making for intensity mapping experiments Zheng Zhang 1, Philip Bull 1,2, Mario G. Santos 2, and Ainulnabilah Nasirudin 1 1 Jodrell Bank Centre for Astrophysics, University of Manchester, Manchester, M13 9PL, United Kingdom 2 Department of Physics and Astronomy, University of Western Cape, Cape Town 7535, South Africa Abstract Line intensity maps have high dynamic range, and demand careful spectral and spatial calibration. We present a Bayesian framework for joint calibration and map-making using Gibbs sampling, which provides access to the full joint posterior of calibration and sky map parameters. Our data model incorporates instrumental noise based on the radiometer equation, capturing the coupling between noise level and system temperature without assuming a fixed noise amplitude. To enable unbiased and fast estimation of gain and system temperature, we develop an iterative generalised least squares (GLS) sampling method. Absolute flux calibration can be achieved either with external sources or internally using known signal injections, such as noise diodes. To handle stochastic gain variations, we introduce a 1 / f 1/f noise model that avoids spurious periodic correlations in the time domain caused by the conventional assumption of diagonal DFT noise covariance. Furthermore, we implement this workflow in an efficient software package, using the Levinson algorithm and a polynomial emulator to reduce the computational complexity of noise parameter sampling, ensuring good scalability. Although demonstrated for auto-correlation measurements, the framework and techniques generalise to cross-correlation and interferometric data. \pagerange Joint Bayesian calibration and map-making for intensity mapping experiments – keywords methods: data analysis – techniques: spectroscopic – radio lines: general 1 Introduction Line-intensity mapping (LIM) is a powerful technique for probing the large-scale structure of the Universe. It involves measuring the integrated emission from unresolved sources across cosmic volumes. By targeting specific spectral lines, such as the 21 cm hyperfine transition of neutral hydrogen ( Chang et al. 2010 ; Pritchard & Loeb 2012 ; Bull et al. 2015 ) or molecular lines like CO ( Righi et al. 2008 ) , LIM can efficiently map the three-dimensional distribution of matter over a wide range of redshifts, with access to fainter populations and larger volumes. Calibration and map-making are foundational steps of data processing in intensity mapping experiments, enabling the translation of raw detector signals into scientifically meaningful sky maps. Traditional approaches often decouple these processes, first calibrating the instrument’s gain and noise properties before reconstructing the sky map using linear or maximum-likelihood solvers ( Tegmark 1997b ; Sutton et al. 2010 ; Doré et al. 2001 ; Keihänen et al. 2005 , e.g.) . However, this sequential treatment can propagate calibration errors into the final map and neglect important interdependencies between gain variations, system temperature fluctuations, and noise statistics. These issues are especially pronounced in low-frequency radio experiments, where 1/f noise, thermal drifts, and other instrumental systematics can significantly impact data quality and bias sky reconstruction ( Bigot-Sazy et al. 2015 ; Harper et al. 2018 ; Chen et al. 2020 ; Hu et al. 2021 ; Li et al. 2021 ; Irfan et al. 2024 ) . Bayesian methods allow for the joint, global analysis of LIM data, rigorously propagating uncertainties and dependencies while incorporating prior knowledge. This approach yields exact posterior distributions and enables straightforward marginalisation over nuisance parameters. Recent advances in hierarchical Bayesian methods, particularly ultra high-dimensional implementations of Gibbs sampling, have enabled joint estimation of calibration parameters and sky signals by iteratively sampling from their conditional posteriors ( Wandelt et al. 2004 , e.g.) . This approach has been successfully applied to Cosmic Microwave Background (CMB) datasets through pipelines such as BeyondPlanck and Cosmoglobe , where calibration, noise modelling, and component separation are treated in a unified framework ( Galloway et al. 2023 ; Eskilt et al. 2023 ) . These developments offer a promising path for extending similar techniques to 21 cm cosmology and other intensity mapping experiments. However, replicating the success of joint Bayesian analysis in intensity mapping experiments is far from straightforward. There are two main reasons for this. First, compared with the CMB, the foreground components in intensity mapping are significantly brighter than the target signal ( Spinelli et al. 2022 , see e.g.) . Within the required dynamic range, these foregrounds cannot be reliably captured by simple empirical models, making standard component separation techniques from the CMB context less directly applicable. Second, intensity mapping experiments often contend with poorly understood or experiment-specific systematics, such as instrument-induced spectral structures ( Thyagarajan et al. 2015 ; Kern et al. 2019 ; Cunnington et al. 2021 ; McCallum et al. 2021 ; Murphy et al. 2024 ; Charles et al. 2024 ) , calibration drifts ( Harper et al. 2018 ; Chen et al. 2020 ; Li et al. 2021 ; Irfan et al. 2024 ) , and environmental effects like groundspill ( Wang et al. 2021 ) and radio-frequency interference ( Offringa et al. 2015 ; Harper & Dickinson 2018 ; Wilensky et al. 2022 ; Engelbrecht et al. 2025 ) . Given the complexity of these challenges, a fully end-to-end Bayesian framework would be rather complex and unwieldy, with many model components. Instead, a modular Bayesian pipeline – dividing the problem into stages – may offer a more tractable solution for the time being. In what follows, we specialise to autocorrelation or ‘single-dish’ intensity mapping experiments, which measure the total power from their receivers and use the resulting time-ordered data to make three-dimensional maps of line brightness temperature. Recent examples of such experiments include GBT ( Chang et al. 2010 ; Masui et al. 2013 ) , Parkes ( Anderson et al. 2018 ) , and FAST ( Li et al. 2023 ) single-dish observations with the 21cm line, the MeerKLASS multi-dish autocorrelation survey ( Santos et al. 2016 ; Wang et al. 2021 ) , also with the 21cm line, and the COMAP survey with CO lines ( Foss et al. 2022 ) . The data reduction and analysis steps for these types of experiment can be split into two major stages: (A) joint calibration and map-making from time-ordered data to a three-dimensional sky map, and (B) component separation and scientific inference from the resulting 3D map. The present work focuses on the implementation and demonstration of the feasibility of a fully Bayesian approach to stage (A). A complementary Bayesian approach to (B) is presented in G. Murphy et al. ( in prep. ) In designing our model and workflow, we identified two common pain points in standard calibration practices. First, noise modelling is often oversimplified: many existing frameworks either assume a fixed noise level during calibration or neglect the coupling between noise and system parameters. This contradicts the radiometer equation, which predicts that the noise variance scales multiplicatively with both the temperature of the system and the gain ( Burke et al. 2019 ) . Such simplifications can lead to biased parameter estimates ( Winkel et al. 2012 ) , especially in high-dynamic-range observations. Second, a major challenge in auto-correlation time-ordered data (TOD) is disentangling stochastic gain fluctuations, commonly referred to as 1/f noise, from the underlying sky signal. Unlike cross-correlation systems (interferometers), where correlated noise between detectors can be mitigated, auto-correlation measurements are inherently more sensitive to

Read the complete text