On the computational feasibility of Bayesian end-to-end analysis of LiteBIRD simulations within Cosmoglobe

R. Aurvik, M. Galloway, E. Gjerløw, U Fuskeland, A. Basyrov, M. Bortolami, M. Brilenkov, P. Campeti, H. K. Eriksen, L. T. Hergt, D. Herman, M. Monelli, L. Pagano, Giuseppe Puglisi, N. Raffuzzi, Nils-Ole Stutzer, R. M. Sullivan, H. Thommesen, Dj Watts, I. K. Wehus, D Adak, Erwan Allys, Ashish Anand, J. Aumont, C Baccigalupi, M. Ballardini, A. J. Banday, R. B. Barreiro, N. Bartolo, S. Basak, M. Bersanelli, A Besnard, T Brinckmann, Erminia Calabrese, E. Carinos, F. J. Casas, K. Cheung, M Citran, Lionel Clermont, F Columbro, Gabriele Coppi, A Coppolecchia, Bo Peng, P de Bernardis, E. de la Hoz, Mario de Lucia, S Della Torre, P. Diego-Palazuelos, Thomas Essinger-Hileman, C Franceschet, G. Galloni, M. Gerbino, M. Gervasi, R.t Genova-Santos, T. Ghigna, S. Giardiello, C. Gimeno-Amo, A. Gruppuso, M. Hazumi, S. Henrot–Versillé, Kazunori Kohri, L. Lamagna, T. Lari, M. Lattanzi, C. Leloup, F. Levrier, A.i. Lonappan, M. López-Caniego, G. Luzzi, J. F. Macías–Pérez, B. Maffei, E. Martínez-González, S. Masi, S. Matarrese, T. Matsumura, S. Micheli, L. Montier, G. Morgante, L. Mousset, Ryo Nagata, Andrea Novelli, Ippei Obata, Antonio Occhiuzzi, A Paiella, D. Paoletti, G Pascual-Cisneros, F Piacentini, Michele Pinchera, G Polenta, L. Porcelli, M Remazeilles, A. Ritacco, A. Rizzieri, M. Ruiz-Granda, J. Sanghavi, V. Sauvage, Maresuke Shiraishi, S. L. Stever, Y Takase, Konstantinos Tassis

Abstract

We assess the computational feasibility of end-to-end Bayesian analysis of the JAXA-led LiteBIRD experiment by analysing simulated time ordered data (TOD) for a subset of detectors through the Cosmoglobe and Commander3 framework. The data volume for the simulated TOD is 1.55 TB, or 470 GB after Huffman compression. From this we estimate a total data volume of 238 TB for the full three year mission, or 70 TB after Huffman compression. We further estimate the running time for one Gibbs sample, from TOD to cosmological parameters, to be approximately 3000 CPU hours. The current simulations are based on an ideal instrument model, only including correlated 1/f noise. Future work will consider realistic systematics with full end-to-end error propagation. We conclude that these requirements are well within capabilities of future high-performance computing systems.

Introduction

The standard model of cosmology predicts that the universe underwent a period of rapid expansion called inflation, which should have left faint "B-mode" polarization patterns in the Cosmic Microwave Background (CMB). The JAXA-led LiteBIRD satellite, launching in the 2030s, aims to detect these elusive signals. To achieve the required sensitivity, scientists must disentangle the true CMB signal from instrumental noise and astrophysical foregrounds. The BeyondPlanck and Cosmoglobe projects recently demonstrated that a global Bayesian end-to-end analysis—where instrument calibration, map-making, and component separation are modeled jointly—is the most robust approach. However, LiteBIRD will feature 4508 bolometers observing the sky for 3 years, generating an unprecedented volume of Time-Ordered Data (TOD). This paper evaluates whether such a computationally heavy Bayesian framework is feasible for a dataset of this magnitude.

Methods

The authors tested computational limits using Commander3, a Bayesian analysis framework employing Gibbs sampling to iteratively sample the parameter space of the CMB sky and instrument model. Given the prohibitive cost of processing the full 3-year, 4508-detector dataset for a test run, the team simulated a scaled-down subset: 88 detectors and 1 year of observations. By measuring the RAM, storage, and CPU requirements for this subset, they linearly extrapolated the TOD processing costs to estimate the burden of the full LiteBIRD mission.

Instrument and Data

The study utilizes simulated LiteBIRD TOD generated by the LiteBIRD Simulation Framework (LBS). The instrument model covers a wide range of frequencies across three telescopes: the Low Frequency Telescope (LFT), Medium Frequency Telescope (MFT), and High Frequency Telescope (HFT).

Table 1: LiteBIRD instrument specifications, including frequency bands, beam sizes, and bolometer counts for the LFT, MFT, and HFT telescopes.

Foregrounds were modeled using the Python Sky Model (PySM), incorporating galactic synchrotron, thermal dust, free-free, anomalous microwave emission (AME), and CO emission. Extragalactic sources were modeled via WebSky simulations, while noise simulations included white noise and correlated $1/f$ noise based on expected detector specifications.

Results and Computational Insights

The analysis reveals that as the dataset scales, the computational cost of component separation in the map domain becomes negligible compared to processing the 1D Time-Ordered Data. The most expensive step is correlated noise sampling, which accounts for approximately 24% of the total CPU time.

Memory footprint is a significant constraint, as Commander3 stores data in RAM. However, the application of Huffman compression optimized for CMB data reduces RAM requirements by a factor of ~3.4.

Full Mission Estimates (4508 detectors, 3 years): * Data Volume: 238 TB uncompressed, reducible to 70 TB via compression. * RAM Required: 70 TB. * Compute Cost: ~3000 CPU hours per Gibbs sample.

A full production run of 1000 Gibbs samples would require approximately 3 million CPU hours. This is well within the capabilities of modern high-performance computing (HPC) clusters.

Conclusion

The study demonstrates that Bayesian end-to-end analysis is a viable path for next-generation CMB experiments like LiteBIRD. While the current simulations assume an idealized instrument model, future work will incorporate realistic systematic effects (e.g., non-ideal beams, gain fluctuations, and half-wave plate non-idealities), which may increase CPU costs by an order of magnitude. To address memory bottlenecks, the development of Commander4 aims to provide a massively parallelized framework capable of distributing computation across standard computing nodes.

Summary updated

Abstract

We assess the computational feasibility of end-to-end Bayesian analysis of the JAXA-led LiteBIRD experiment by analysing simulated time ordered data (TOD) for a subset of detectors through the Cosmoglobe and Commander3 framework. The data volume for the simulated TOD is 1.55 TB, or 470 GB after Huffman compression. From this we estimate a total data volume of 238 TB for the full three year mission, or 70 TB after Huffman compression. We further estimate the running time for one Gibbs sample, from TOD to cosmological parameters, to be approximately 3000 CPU hours. The current simulations are based on an ideal instrument model, only including correlated 1/f noise. Future work will consider realistic systematics with full end-to-end error propagation. We conclude that these requirements are well within capabilities of future high-performance computing systems.
BibTeX
@article{aurvik2025computational,
  author = {R. Aurvik and M. Galloway and E. Gjerløw and U Fuskeland and A. Basyrov and M. Bortolami and M. Brilenkov and P. Campeti and H. K. Eriksen and L. T. Hergt and D. Herman and M. Monelli and L. Pagano and Giuseppe Puglisi and N. Raffuzzi and Nils-Ole Stutzer and R. M. Sullivan and H. Thommesen and DJ Watts and I. K. Wehus and D Adak and Erwan Allys and Ashish Anand and J. Aumont and C Baccigalupi and M. Ballardini and A. J. Banday and R. B. Barreiro and N. Bartolo and S. Basak and M. Bersanelli and A Besnard and T Brinckmann and Erminia Calabrese and E. Carinos and F. J. Casas and K. Cheung and M Citran and Lionel Clermont and F Columbro and Gabriele Coppi and A Coppolecchia and Bo Peng and P de Bernardis and E. de la Hoz and Mario de Lucia and S Della Torre and P. Diego-Palazuelos and Thomas Essinger-Hileman and C Franceschet and G. Galloni and M. Gerbino and M. Gervasi and R.T Genova-Santos and T. Ghigna and S. Giardiello and C. Gimeno-Amo and A. Gruppuso and M. Hazumi and S. Henrot–Versillé and Kazunori Kohri and L. Lamagna and T. Lari and M. Lattanzi and C. Leloup and F. Levrier and A.I. Lonappan and M. López-Caniego and G. Luzzi and J. F. Macías–Pérez and B. Maffei and E. Martínez-González and S. Masi and S. Matarrese and T. Matsumura and S. Micheli and L. Montier and G. Morgante and L. Mousset and Ryo Nagata and Andrea Novelli and Ippei Obata and Antonio Occhiuzzi and A Paiella and D. Paoletti and G Pascual-Cisneros and F Piacentini and Michele Pinchera and G Polenta and L. Porcelli and M Remazeilles and A. Ritacco and A. Rizzieri and M. Ruiz-Granda and J. Sanghavi and V. Sauvage and Maresuke Shiraishi and S. L. Stever and Y Takase and Konstantinos Tassis},
  title = {{On the computational feasibility of Bayesian end-to-end analysis of LiteBIRD simulations within Cosmoglobe}},
  journal = {Journal of Cosmology and Astroparticle Physics},
  year = {2025},
  doi = {10.1088/1475-7516/2025/11/041},
}
Figure