Preparing for a first detection of the 21-cm signal during reionization by large-scale interferometer experiments requires rigorous testing of the data analysis and reduction pipelines. Validating that these do not erroneously add/remove features mimicking the signal (e.g. from side lobes or large-scale power leakage) requires simulations extending beyond the primary field of view. However, the Murchison Wide Field Array (MWA) with a field of view of ∼252 deg2 would require simulations spanning several Gpcs, which are currently infeasible. To address this, we developed a simplified version of the seminumerical reionization simulation code 21cmfast, sacrificing some physical accuracy (linear structure formation) in favour of extremely large volumes. We then constructed a 7.5 Gpc comoving volume specifically tailored to the binned spectral resolution of the MWA (∼1.17 cMpc), required for validating the pipeline used in the 2020 MWA 21-cm power spectrum (PS) upper limits. With this large-volume simulation, we then explored: (i) whether smaller volume simulations are biased by missing large-scale modes, (ii) non-Gaussianity in the cosmic variance uncertainty, (iii) biases in the recovered 21-cm PS following foreground wedge avoidance, and (iv) the impact of tiling smaller simulations to achieve large volumes. We found (i) no biases from missing large-scale power, (ii) significant contribution from non-Gaussianity, as expected, (iii) a 10–20 per cent overestimate of the 21-cm PS following wedge mode excision, and (iv) tiling smaller simulations underestimates the large-scale power and cosmic variance.
Observations of the cosmic 21-cm power spectrum (PS) are starting to enable precision Bayesian inference of galaxy properties and physical cosmology, during the first billion years of our Universe. Here we investigate the impact of common approximations about the likelihood used in such inferences, including: (i) assuming a Gaussian functional form; (ii) estimating the mean from a single realization; and (iii) estimating the (co)variance at a single point in parameter space. We compare ‘classical’ inference that uses an explicit likelihood with simulation-based inference (SBI) that estimates the likelihood from a training set. Our forward models include: (i) realizations of the cosmic 21-cm signal computed with 21cmFAST by varying ultraviolet (UV) and X-ray galaxy parameters together with the initial conditions; (ii) realizations of the telescope noise corresponding to a $1000 \, \mathrm{h}$ integration with the low-frequency component of the Square Kilometre Array (SKA1-Low); and (iii) the excision of Fourier modes corresponding to a foreground-dominated horizon ‘wedge’. We find that the 1D PS likelihood is well described by a Gaussian accounting for covariances between wave modes and redshift bins (higher order correlations are small). However, common approaches of estimating the forward-modelled mean and (co)variance from a random realization or at a single point in parameter space result in biased and overconstrained posteriors. Our best results come from using SBI to fit a non-Gaussian likelihood with a Gaussian mixture neural density estimator. Such SBI can be performed with up to an order of magnitude fewer simulations than classical, explicit likelihood inference. Thus SBI provides accurate posteriors at a comparably low computational cost.
more » « less- NSF-PAR ID:
- 10435442
- Publisher / Repository:
- Oxford University Press
- Date Published:
- Journal Name:
- Monthly Notices of the Royal Astronomical Society
- Volume:
- 524
- Issue:
- 3
- ISSN:
- 0035-8711
- Page Range / eLocation ID:
- p. 4239-4255
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
more » « less
ABSTRACT -
Aleatoric uncertainty captures the inherent randomness of the data, such as measurement noise. In Bayesian regression, we often use a Gaussian observation model, where we control the level of aleatoric uncertainty with a noise variance parameter. By contrast, for Bayesian classification we use a categorical distribution with no mechanism to represent our beliefs about aleatoric uncertainty. Our work shows that explicitly accounting for aleatoric uncertainty significantly improves the performance of Bayesian neural networks. We note that many standard benchmarks, such as CIFAR-10, have essentially no aleatoric uncertainty. Moreover, we show that data augmentation in approximate inference softens the likelihood, leading to underconfidence and misrepresenting our beliefs about aleatoric uncertainty. Accordingly, we find that a cold posterior, tempered by a power greater than one, often more honestly reflects our beliefs about aleatoric uncertainty than no tempering --- providing an explicit link between data augmentation and cold posteriors. We further show that we can match or exceed the performance of posterior tempering by using a Dirichlet observation model, where we explicitly control the level of aleatoric uncertainty, without any need for tempering.more » « less
-
Aleatoric uncertainty captures the inherent randomness of the data, such as measurement noise. In Bayesian regression, we often use a Gaussian observation model, where we control the level of aleatoric uncertainty with a noise variance parameter. By contrast, for Bayesian classification we use a categorical distribution with no mechanism to represent our beliefs about aleatoric uncertainty. Our work shows that explicitly accounting for aleatoric uncertainty significantly improves the performance of Bayesian neural networks. We note that many standard benchmarks, such as CIFAR, have essentially no aleatoric uncertainty. Moreover, we show data augmentation in approximate inference has the effect of softening the likelihood, leading to underconfidence and profoundly misrepresenting our honest beliefs about aleatoric uncertainty. Accordingly, we find that a cold posterior, tempered by a power greater than one, often more honestly reflects our beliefs about aleatoric uncertainty than no tempering -- providing an explicit link between data augmentation and cold posteriors. We show that we can match or exceed the performance of posterior tempering by using a Dirichlet observation model, where we explicitly control the level of aleatoric uncertainty, without any need for tempering.more » « less
-
more » « less
ABSTRACT Line-intensity mapping (LIM) is an emerging technique to probe the large-scale structure of the Universe. By targeting the integrated intensity of specific spectral lines, it captures the emission from all sources and is sensitive to the astrophysical processes that drive galaxy evolution. Relating these processes to the underlying distribution of matter introduces observational and theoretical challenges, such as observational contamination and highly non-Gaussian fields, which motivate the use of simulations to better characterize the signal. In this work we present skyline , a computational framework to generate realistic mock LIM observations that include observational features and foreground contamination, as well as a variety of self-consistent tracer catalogues. We apply our framework to generate realizations of LIM maps from the multidark planck 2 simulations coupled to the universemachine galaxy formation model. We showcase the potential of our scheme by exploring the voxel intensity distribution and the power spectrum of emission lines such as 21 cm, CO, [C ii], and Lyman-α, their mutual cross-correlations, and cross-correlations with galaxy clustering. We additionally present cross-correlations between LIM and submillimetre extragalactic tracers of large-scale structure such as the cosmic infrared background and the thermal Sunyaev-Zel’dovich effect, as well as quantify the impact of galactic foregrounds, line interlopers, and instrument noise on LIM observations. These simulated products will be crucial in quantifying the true information content of LIM surveys and their cross-correlations in the coming decade, and to develop strategies to overcome the impact of contaminants and maximize the scientific return from LIM experiments.
-
more » « less
ABSTRACT Most cosmic shear analyses to date have relied on summary statistics (e.g. ξ+ and ξ−). These types of analyses are necessarily suboptimal, as the use of summary statistics is lossy. In this paper, we forward-model the convergence field of the Universe as a lognormal random field conditioned on the observed shear data. This new map-based inference framework enables us to recover the joint posterior of the cosmological parameters and the convergence field of the Universe. Our analysis properly accounts for the covariance in the mass maps across tomographic bins, which significantly improves the fidelity of the maps relative to single-bin reconstructions. We verify that applying our inference pipeline to Gaussian random fields recovers posteriors that are in excellent agreement with their analytical counterparts. At the resolution of our maps – and to the extent that the convergence field can be described by the lognormal model – our map posteriors allow us to reconstruct all summary statistics (including non-Gaussian statistics). We forecast that a map-based inference analysis of LSST-Y10 data can improve cosmological constraints in the σ8–Ωm plane by $\approx\!{30}{{\ \rm per\ cent}}$ relative to the currently standard cosmic shear analysis. This improvement happens almost entirely along the $S_8=\sigma _8\Omega _{\rm m}^{1/2}$ directions, meaning map-based inference fails to significantly improve constraints on S8.
