Fast Sampling of Cosmological Initial Conditions with Gaussian Neural Posterior Estimation
Oleg Savchenko, Guillermo Franco Abellán, Florian List, Noemi Anau Montel, Christoph Weniger · 2025
Read Fast Sampling of Cosmological Initial Conditions with Gaussian Neural Posterior Estimation by Oleg Savchenko, Guillermo Franco Abellán, Florian List, Noemi Anau Montel, Christoph Weniger (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-2502-03139.txt.
Introduction
Oleg Savchenko and co-authors speed up the rebuild of the early universe's matter layout in 2025. All large structure grew from that first layout. Rebuilding it from late data is hard and slow. The space of options has millions of parts. Their method learns from simulations instead of solving the problem head on.
They model the likely range of starts as a bell curve in a wave based space. That choice lets them draw thousands of samples in seconds on one graphics card. Older methods are far slower. A fit for how spread varies with scale turns any best guess into a fast sampler. Checks against true values and a logic test support the samples. The way is open to full scale runs.
Catalog record
| CREATOR | Oleg Savchenko, Guillermo Franco Abellán, Florian List, Noemi Anau Montel, Christoph Weniger |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-03139
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Fast Sampling of Cosmological Initial Conditions with Gaussian Neural Posterior Estimation Oleg Savchenko, Guillermo Franco Abellán, Florian List, Noemi Anau Montel, Christoph Weniger Abstract Knowledge of the primordial matter density field from which the large-scale structure of the Universe emerged over cosmic time is of fundamental importance for cosmology. However, reconstructing these cosmological initial conditions from late-time observations is a notoriously difficult task, which requires advanced cosmological simulators and sophisticated statistical methods to explore a multi-million-dimensional parameter space. We show how simulation-based inference (SBI) can be used to tackle this problem and to obtain data-constrained realisations of the primordial dark matter density field in a simulation-efficient way with general non-differentiable simulators. Our method is applicable to full high-resolution dark matter N-body simulations and is based on modelling the posterior distribution of the constrained initial conditions to be Gaussian with a diagonal covariance matrix in Fourier space. As a result, we can generate thousands of posterior samples within seconds on a single GPU, orders of magnitude faster than existing methods, paving the way for sequential SBI for cosmological fields. Furthermore, we perform an analytical fit of the estimated dependence of the covariance on the wavenumber, effectively transforming any point-estimator of initial conditions into a fast sampler. We test the validity of our obtained samples by comparing them to the true values with summary statistics and performing a Bayesian consistency test. 2026 Fast Sampling of Cosmological Initial Conditions with Gaussian Neural Posterior Estimation – 7 1 Introduction The current era of precise and abundant cosmological data has firmly established the standard model of cosmology, which shows remarkable agreement with a wide variety of observations and is now constrained to percent-level precision (see, e.g., Peebles 2024 ; Efstathiou 2024 ). One of the major probes of this model, which still holds significant untapped potential, is the three-dimensional cosmic web distribution of galaxies, referred to as the large-scale structure (LSS) of the Universe. Next-generation galaxy surveys, such as the Dark Energy Spectroscopic Instrument (DESI, Levi et al. 2019 ), Euclid ( Laureijs et al., 2011 ) , the Nancy Grace Roman Space Telescope ( Spergel et al., 2015 ) , and the Vera Rubin Observatory ( Ivezić et al., 2019 ) , will provide an enormous wealth of new information about the LSS, covering scales from a few Mpc up to the size of the cosmological horizon. However, analysing these vast data volumes often requires computationally expensive simulations and exploration of high-dimensional parameter spaces, posing a serious challenge for traditional statistical inference techniques and, in many cases, rendering them effectively infeasible. The necessity to go beyond standard methods, combined with recent revolutionary progress in deep learning, has led to the emergence of powerful alternative approaches capable of tackling these pressing challenges ( Huertas-Company & Lanusse, 2023 ) . A promising avenue of research in this direction is simulation-based inference (SBI, Cranmer et al. 2020 ). All SBI algorithms perform inference using a generative model of the data, i.e. a stochastic simulator that maps model parameters to data realisations, thus effectively sampling an (implicit) likelihood. SBI has already been successfully applied to a variety of low-dimensional tasks in astrophysics and cosmology ( Makinen et al., 2021 ; Cole et al., 2022 ; Mishra-Sharma & Cranmer, 2022 ; Anau Montel et al., 2022 ; Karchev et al., 2023 ; Lemos et al., 2024 ; Gatti et al., 2024 ; Franco Abellán et al., 2024 ; von Wietersheim-Kramsta et al., 2024 , e.g.,) ), but applying it to high-dimensional problems — such as image or 3D field reconstruction, which often involves inference of \mathrm{\gtrsim}\,10^{6} parameters — still remains an area of active ongoing research ( List et al., 2023 ; Flöss et al., 2024 ; Sabti et al., 2024 , e.g.,) . Table 1: Overview of selected 3D cosmological matter field reconstruction methods from the literature, alongside our work (this table is not exhaustive but highlights some notable works for comparison). We compare the sizes and grid resolutions of the reconstructed simulation boxes, the implemented methods, and whether they allow sampling the posterior or only getting point estimates, together with the computational resources required for convergence or training . Box size, \mathrm{Gpc}/h Grid resolution Reconstruction method Posterior sampling Computational resource BORG ( Jasche & Lavaux 2019 ) 0.6777 256^{3} HMC ✓ \sim 10^{6} CPU hours for \sim 10^{3} samples Schmittfull et al. (2017) 0.5 512^{3} Iterative \times 3 CPU hours Feng et al. (2018) 0.4 128^{3} Gradient-based MAP estimation \times N/A Modi et al. (2021) 0.4 64^{3} Recurrent inference machines \times N/A Shallue & Eisenstein (2023) \simeq 0.44 128^{3} CNN \times 1 Nvidia GeForce GTX 1080 Ti GPU, 4-12 hours Jindal et al. (2023) 0.25 128^{3} V-Net \times N/A Legin et al. (2023) 1 128^{3} Score-based diffusion ✓ 4 Nvidia A100 80GB GPUs, 24 hours Flöss & Meerburg (2024) 1 256^{3} U-Net \times 4 Nvidia A100 40GB GPUs, 4 hours Wang et al. (2024) 0.5 256^{3} U-Net \times N/A Bayer et al. (2023a) 0.256 64^{3} Microcanonical Langevin Monte Carlo ✓ N/A Doeser et al. (2024) 0.25 128^{3} HMC + differentiable emulator ✓ N/A Sabti et al. (2024) \approx 0.5 128^{3} Stochastic interpolants ✓ 1 Nvidia A100 80GB GPU, 3 days This work 1 128^{3} Gaussian NPE ✓ 1 Nvidia A100 40GB GPU, 1.5 hour A prominent example of a high-dimensional inference task in the cosmological setting is the problem of reconstructing the cosmological initial conditions (ICs) from which the observed LSS distribution emerged. While the present-day cosmic web has a highly complex, non-linear and non-Gaussian nature, it originates from a much simpler state that is well characterised by linear Gaussian fluctuations ( Dodelson & Schmidt, 2020 ) . With data-constrained realisations of these ICs at hand, one can then forward-simulate their evolution in time, and reconstruct the local matter distribution, as well as the velocity fields and structure formation histories. This has already found important physical applications, for instance, in the context of primordial magnetic fields ( Hutschenreuter et al., 2018 ) , de-biasing H_{0} inference from standard sirens ( Mukherjee et al., 2021 ) , or the cosmic neutrino background ( Elbers et al., 2023 ) . Most of the standard cosmological analyses rely on using the matter two-point function, which saturates the information content of the LSS only on relatively large scales where the Universe still evolves linearly. In contrast, most of the new information content delivered by the upcoming surveys will come from the small scales where the Universe becomes highly non-linear and non-Gaussian ( Krause et al., 2024 ) . Accessing the full field-level information allows extracting all the information available in higher-order statistics, thereby significantly improving constraints on cosmological parameters ( Leclercq & Heavens, 2021 ; Makinen et al., 2021 ; Andrews et al., 2023 ; Zhou & Dodelson, 2023 ; Lemos et al., 2024 ; Flöss & Meerburg, 2024 ; Nguyen et al., 2024 ; Andrews et al., 2024 , e.g.,) . Early attempts to reconstruct the cosmological ICs using methods such as the least-action principle or optimal mass transportation date back several decades ( Peebles, 1989 ; Frisch et al., 2002 ; Brenier et al., 2003 ) . However, a unique recovery of ICs from a late-time matter density field