Gaussian Process Regression for foreground removal in HI intensity mapping experiments

We apply for the first time Gaussian Process Regression (GPR) as a foreground removal technique in the context of single-dish, low redshift H i intensity mapping, and present an open-source python toolkit for doing so. We use MeerKAT and SKA1-MID-like simulations of 21cm foregrounds (including polarisation leakage), H i cosmological signal and instrumental noise. We find that it is possible to use GPR as a foreground removal technique in this context, and that it is better suited in some cases to recover the H i power spectrum than Principal Component Analysis (PCA), especially on small scales. GPR is especially good at recovering the radial power spectrum, outperforming PCA when considering the full bandwidth of our data. Both methods are worse at recovering the transverse power spectrum, since they rely on frequency-only covariance information. When halving our data along frequency, we find that GPR performs better in the low frequency range, where foregrounds are brighter. It performs worse than PCA when frequency channels are missing, to emulate RFI flagging. We conclude that GPR is an excellent foreground removal option for the case of single-dish, low redshift H i intensity mapping in the absence of missing frequency channels. Our python toolkit gpr4im and the data used in this analysis are publicly available on GitHub.

Comparing foreground removal techniques for recovery of the LOFAR-EoR 21 cm power spectrum

ABSTRACT We compare various foreground removal techniques that are being utilized to remove bright foregrounds in various experiments aiming to detect the redshifted 21 cm signal of neutral hydrogen from the epoch of reionization. In this work, we test the performance of removal techniques (FastICA, GMCA, and GPR) on 10 nights of LOFAR data and investigate the possibility of recovering the latest upper limit on the 21 cm signal. Interestingly, we find that GMCA and FastICA reproduce the most recent 2σ upper limit of $\Delta ^2_{21} \lt $ (73)2 mK2 at k = 0.075 hcMpc−1, which resulted from the application of GPR. We also find that FastICA and GMCA begin to deviate from the noise-limit at k-scales larger than ∼0.1 hcMpc−1. We then replicate the data via simulations to see the source of FastICA and GMCA’s limitations, by testing them against various instrumental effects. We find that no single instrumental effect, such as primary beam effects or mode-mixing, can explain the poorer recovery by FastICA and GMCA at larger k-scales. We then test scale-independence of FastICA and GMCA, and find that lower k-scales can be modelled by a smaller number of independent components. For larger scales (k ≳ 0.1 hcMpc−1), more independent components are needed to fit the foregrounds. We conclude that, the current usage of GPR by the LOFAR collaboration is the appropriate removal technique. It is both robust and less prone to overfitting, with future improvements to GPR’s fitting optimization to yield deeper limits.

The degeneracy between primordial non-Gaussianity and foregrounds in 21 cm intensity mapping experiments

ABSTRACT Potential evidence for primordial non-Gaussianity (PNG) is expected to lie in the largest scales mapped by cosmological surveys. Forthcoming 21 cm intensity mapping experiments will aim to probe these scales by surveying neutral hydrogen (H i) within galaxies. However, foreground signals dominate the 21 cm emission, meaning foreground cleaning is required to recover the cosmological signal. The effect this has is to damp the H i power spectrum on the largest scales, especially along the line of sight. Whilst there is agreement that this contamination is potentially problematic for probing PNG, it is yet to be fully explored and quantified. In this work, we carry out the first forecasts on fNL that incorporate simulated foreground maps that are removed using techniques employed in real data. Using an Monte Carlo Markov Chain analysis on an SKA1-MID-like survey, we demonstrate that foreground cleaned data recovers biased values [$f_{\rm NL}= -102.1_{-7.96}^{+8.39}$ (68 per cent CL)] on our fNL = 0 fiducial input. Introducing a model with fixed parameters for the foreground contamination allows us to recover unbiased results ($f_{\rm NL}= -2.94_{-11.9}^{+11.4}$). However, it is not clear that we will have sufficient understanding of foreground contamination to allow for such rigid models. Treating the main parameter $k_\parallel ^\text{FG}$ in our foreground model as a nuisance parameter and marginalizing over it, still recovers unbiased results but at the expense of larger errors ($f_{\rm NL}= 0.75^{+40.2}_{-44.5}$), which can only be reduced by imposing the Planck 2018 prior. Our results show that significant progress on understanding and controlling foreground removal effects is necessary for studying PNG with H i intensity mapping.

Multipole expansion for H i intensity mapping experiments: simulations and modelling

ABSTRACT We present a framework and an open-source python toolkit to analyse the two-point statistics of 3D fluctuations in the context of H i intensity maps using the multipole expansion formalism. We include simulations of the cosmological H i signal using N-body and lognormal methods, foregrounds and their removal, as well as instrumental effects. Using these simulations and analytical modelling, we investigate the impact of foreground cleaning and the instrumental beam on the power spectrum multipoles as well as on the Fourier space clustering wedges. We find that both the instrumental beam and the foreground removal can produce a quadrupole (and a hexadecapole) signal, and demonstrate the importance of controlling and accurately modelling these effects for precision radio cosmology. We conclude that these effects can be modelled with reasonable accuracy using our multipole expansion technique. We also perform a Markov Chain Monte Carlo (MCMC) analysis to showcase the effect of foreground cleaning on the estimation of the H i abundance and bias parameters. The accompanying python toolkit is available at https://github.com/IntensityTools/MultipoleExpansion, and includes an interactive suite of examples to aid new users.

HIR4: cosmology from a simulated neutral hydrogen full sky using Horizon Run 4

ABSTRACT The distribution of cosmological neutral hydrogen will provide a new window into the large-scale structure of the Universe with the next generation of radio telescopes and surveys. The observation of this material, through 21 cm line emission, will be confused by foreground emission in the same frequencies. Even after these foregrounds are removed, the reconstructed map may not exactly match the original cosmological signal, which will introduce systematic errors and offset into the measured correlations. In this paper, we simulate future surveys of neutral hydrogen using the Horizon Run 4 (HR4) cosmological N-body simulation. We generate H i intensity maps from the HR4 halo catalogue, and combine with foreground radio emission maps from the Global Sky Model, to create accurate simulations over the entire sky. We simulate the H i sky for the frequency range 700–800 MHz, matching the sensitivity of the Tianlai pathfinder. We test the accuracy of the fastICA, PCA, and log-polynomial fitting foreground removal methods to recover the input cosmological angular power spectrum and measure the parameters. We show the effect of survey noise levels and beam sizes on the recovered the cosmological constraints. We find that while the reconstruction removes power from the cosmological 21 cm distribution on large scales, we can correct for this and recover the input parameters in the noise-free case. However, the effect of noise and beam size of the Tianlai pathfinder prevents accurate recovery of the cosmological parameters when using only intensity mapping information.

Extracting the 21cm Global Signal using Artificial Neural Networks

The study of the cosmic Dark Ages, Cosmic Dawn, and Epoch of Reionization (EoR) using the all-sky averaged redshifted HI 21cm signal, are some of the key science goals of most of the ongoing or upcoming experiments, for example, EDGES, SARAS, and the SKA. This signal can be detected by averaging over the entire sky, using a single radio telescope, in the form of a Global signal as a function of only redshifted HI 21cm frequencies. One of the major challenges faced while detecting this signal is the dominating, bright foreground. The success of such detection lies in the accuracy of the foreground removal. The presence of instrumental gain fluctuations, chromatic primary beam, radio frequency interference (RFI) and the Earth’s ionosphere corrupts any observation of radio signals from the Earth. Here, we propose the use of Artificial Neural Networks (ANN) to extract the faint redshifted 21cm Global signal buried in a sea of bright Galactic foregrounds and contaminated by different instrumental models. The most striking advantage of using ANN is the fact that, when the corrupted signal is fed into a trained network, we can simultaneously extract the signal as well as foreground parameters very accurately. Our results show that ANN can detect the Global signal with $\gtrsim 92 {{\ \rm per\ cent}}$ accuracy even in cases of mock observations where the instrument has some residual time-varying gain across the spectrum.

Reionisation & Cosmic Dawn Astrophysics from the Square Kilometre Array: Impact of Observing Strategies

Interferometry of the cosmic 21-cm signal is set to revolutionise our understanding of the Epoch of Reionisation (EoR) and the Cosmic Dawn (CD). The culmination of ongoing efforts will be the upcoming Square Kilometre Array (SKA), which will provide tomography of the 21-cm signal from the first billion years of our Universe. Using a galaxy formation model informed by high-z luminosity functions, here we forecast the accuracy with which the first phase of SKA-low (SKA1-low) can constrain the properties of the unseen galaxies driving the astrophysics of the EoR and CD. We consider three observing strategies: (i) deep (1000h on a single field); (ii) medium-deep (100hr on 10 independent fields); and (iii) shallow (10hr on 100 independent fields). Using the 21-cm power spectrum as a summary statistic, and conservatively only using the 21-cm signal above the foreground wedge, we predict that all three observing strategies should recover astrophysical parameters to a fractional precision of ∼0.1 – 10 per cent. The reionisation history is recovered to an uncertainty of $\Delta z \mathrel {\lesssim}0.1$ (1σ) for the bulk of its duration. The medium-deep strategy, balancing thermal noise against cosmic variance, results in the tightest constraints, slightly outperforming the deep strategy. The shallow observational strategy performs the worst, with up to a ∼10 – 60 per cent increase in the recovered uncertainty. We note, however, that non-Gaussian summary statistics, tomography, as well as unbiased foreground removal would likely favour the deep strategy.

Statistical properties of polarized CMB foreground maps

Foreground removal techniques for CMB analyses make specific assumptions about the properties of foregrounds in temperature and in polarization. By investigating the statistics of foreground components more understanding about the degree to which these assumptions are valid can be obtained. In this work we investigate E- and B-mode maps of the two strongest polarized foregrounds, synchrotron and thermal dust emission, with regards to their similarity with Gaussian processes, their spectral variations, and cross-correlations. We perform tests in patches of ∼3.7° size collectively covering the full sky and find most of them to conform to their Gaussian expectation according to the statistics in use. Correlations exhibit distinct differences in E- and B-mode signals, which point towards necessities in foreground removal methods. We discuss potential consequences and possible further directions.