The Non-Gaussian Universe: Advanced Statistical Inference in the Era of Next-Generation Cosmological Surveys

16 juin 2026, 00:20 → 19 juin 2026, 16:05 Europe/Athens

Dougalis room (Heraklion, Greece)

Jean-Luc Starck (CosmoStat, CEA Paris-Saclay), Marta Spinelli (Observatoire de la Cote d'Azur)

The era of Euclid, LSST, and SKA heralds a profound transformation in observational cosmology. As surveys approach the limits set by cosmic variance and systematic control, extracting maximal information from the cosmic density field demands a fundamental rethinking of how we perform cosmological inference. Traditional two-point statistics, while powerful, capture only a fraction of the information encoded in the nonlinear, non-Gaussian Universe.

This conference explores the emerging paradigm of full-information cosmology, bringing together approaches that move beyond the power spectrum, including high-order statistics, field-level inference, forward modeling, and simulation-based methods. By uniting developments across theory, simulations, and data analysis, the meeting aims to illuminate new pathways toward unbiased, robust, and information-optimal inference, and to define the statistical foundations of cosmology in the era of next-generation surveys.

SOC:  Klea Panayidou, Natalia Porqueres, Marta Spinelli, Jean-Luc Starck, Greg Tsagkatakis, Panos Tsakalides, Stefano Camera

REGISTRATION DEADLINE: May 1st

Confirmed Keynote Speakers

• Carolina Cuesta-Lazaro, Harward, USA
• Zuhui Fan, Pekin University, China 
• Suman Majumdar, Indian Institute of Technology, Indore, India
• Alan Heavens, Imperial College, UK
• Cora Uhlemann,  Bielefeld University, Germany
• Gary Bernstein, Pennsylvania University, USA.
• Nicolas Martinet, Laboratoire d'Astrophysique de Marseille, France
• Giovanni Arico, INFN Bologna, Italy
• Pauline Zarouk , Laboratoire de Physique Nucléaire et de Hautes Énergies, Sorbonne University, Paris, France
• Bhuvnesh Jain, University of Pennsylvania, USA.

mar. 16 juin

14:20 → 14:35 Welcome

Orateur: Jean-Luc Starck (CosmoStat, CEA Paris-Saclay)

14:35 → 17:15 Session 1

14:35 Probing non-gaussian information with hybrid summary statistics

Present-day cosmological cosmic shear fields are non-gaussian, which makes extraction of cosmological information challenging. I report on work done using extreme data compression of physics-based (power spectra) and neural-network derived summary statistics in a fully Bayesian SBI framework. Application to DES Y3 data shows significant improvement in cosmological precision, over 2-point statistics (3 x 2 pt).

Orateur: Alan Heavens (Imperial College London)

15:20 Three-point intrinsic alignments of galaxies and haloes in the FLAMINGO simulations

Weak gravitational lensing is a powerful probe of cosmology, but its interpretation is limited by intrinsic alignments (IA) of galaxies with the large-scale structure. Future surveys will enable precise measurements of higher-order statistics, providing information beyond two-point correlations. Realising this potential requires robust and self-consistent modelling of IA beyond two-point statistics. In this talk, I will present a study of the three-point IA signal measured in the FLAMINGO hydrodynamical simulation. We measure both the three-point correlation function (3PCF) and the third-order aperture mass statistics, using galaxy and halo shapes from the largest simulation, spanning $(2.8,\mathrm{cGpc})^3$. This volume enables high-significance detections of three-point IA over a wide range of scales and triangle configurations. We compare these measurements to predictions from the tree-level effective field theory (EFT) of IA, testing both the full model and several reduced variants. On large scales, the EFT model provides an excellent description of the measured signal, with the inferred alignment amplitude consistent with that obtained from two-point statistics. We find that neglecting higher-order EFT terms can bias constraints and degrades the fit, while enforcing co-evolution relations for these parameters yields performance close to the full model. These relations hold approximately across halo masses and redshifts, making them promising for photometric surveys where simpler models are preferred.

Orateur: Dr Casper Vedder (Leiden University)

15:40 Coffee break

16:10 Adding shape information to the weak lensing peak statistics

Being sensitive to the non-Gaussianity of cosmic structures, weak-lensing (WL) peak statistics have been recognized as an important cosmological probe complementary to 2pt correlation analyses. In this talk, I will present our recent studies on WL peak steepness statistics inspired by machine learning feature explorations. This new analysis is sensitive to the density profile of halos that encodes important information of baryonic feedback and dark matter properties, enabling us to derive additional physical information from WL data besides cosmological parameter constraints. The impacts of galaxy intrinsic alignments on WL peak statistics will also be discussed. An introduction about the Chinese Space Station Survey Telescope (CSST) will be included.

Orateur: Prof. Zuhui Fan (South-Western Institute for Astronomy Research, Yunnan University, China)

16:55 Do Baryons Break Higher-Order Statistics?

To unlock the full potential of Stage IV weak lensing surveys, we must extract non-Gaussian information while mitigating small-scale astrophysical systematics. Positioned strategically between classical 2-point statistics and black-box neural networks, higher-order statistics (HOS) offer a balance of physical interpretability and high constraining power. However, their sensitivity to unmodeled baryonic feedback threatens this theoretical advantage. Leveraging simulation-based inference via neural posterior estimation, we quantify the parameter biases induced by baryons across varying survey footprints. To robustly mitigate these shifts, we investigate the efficacy of survey-dependent scale cuts, restricting our analysis to conservative, quasi-linear regimes. By comparing these mitigated HOS constraints directly against the standard angular power spectrum, we address a defining question for next-generation cosmology: do non-Gaussian statistics actually provide tangible, real-world gains, or is their advantage completely erased by the need to mitigate astrophysical systematics?

Orateur: Andreas Tersenov

mer. 17 juin

10:00 → 19:50 Session 3

10:00 Going Below and Beyond: Extracting non-Gaussian information with one-point statistics

The nonlinear gravitational collapse of matter has shaped the intricate cosmic web of galaxies, creating a diverse range of density environments. Two-point statistics, widely used in galaxy surveys such as Euclid and Rubin LSST, average over these environments, potentially obscuring critical information. To maximise the scientific return of these surveys, we must explore the rich diversity of cosmic structures and non-Gaussian statistics. While higher order N-point correlation functions are often considered, I will explain how one-point statistics capture complementary information to two-point statistics and compress a set of higher-order correlations, providing a powerful tool for probing cosmology and extensions of ΛCDM. I will also explain how these insights can be applied to the one-point statistics of weak lensing and galaxy clustering, as well as their joint distribution, offering new avenues to probe fundamental physics.

Orateur: Prof. Cora Uhlemann (Bielefeld University)

10:45 SBi3PCF: Simulation Based Inference with the Integrated 3 Point Correlation Function

We present SBi3PCF, a simulation-based inference (SBI) framework for analysing a higher-order weak lensing statistic, the integrated 3-point correlation function (i3PCF). Our approach forward-models the cosmic shear field using the CosmoGridV1 suite of N-body simulations, including a comprehensive set of systematic effects such as intrinsic alignment, baryonic feedback, photometric redshift uncertainty, shear calibration bias, and shape noise. Using this, we have produced a set of DES Y3-like synthetic measurements for 2-point shear correlation functions ξ± (2PCFs) and i3PCFs ζ± across 6 cosmological and 11 systematic parameters. Having validated these measurements against theoretical predictions and thoroughly examined for potential systematic biases, we have found that the impact of source galaxy clustering and reduced shear on the i3PCF is negligible for Stage-III surveys. Furthermore, we have tested the Gaussianity assumption for the likelihood of our data vector and found that while the sampling distribution of the 2PCF can be well approximated by a Gaussian function, the likelihood of the combined 2PCF + i3PCF data vector including filter sizes of 90′ and larger can deviate from this assumption. Our SBI pipeline employs masked autoregressive flows to perform neural likelihood estimation and is validated to give statistically accurate posterior estimates. On mock data, we find that including the i3PCF yields a substantial 63.8% median improvement in the figure of merit for Ωm−σ8−w0. These findings are consistent with previous works on the i3PCF and demonstrate that our SBI framework can achieve the accuracy and realism needed to analyse the i3PCF in wide-area weak lensing surveys.

Orateur: Dr David Gebauer (Bielefeld University)

11:05 Coffee break

11:35 Higher Order Weak Lensing Statistics in Euclid

I will review past and ongoing efforts to exploit higher order weak lensing statistics within the Euclid Consortium.

Orateur: Nicolas Martinet (LAM/CNES)

12:20 Confronting the Cosmological Principle: Shaking the foundations of the Cosmos

This talk intends to put the Cosmological Principle (CP) to test by analyzing all-sky surveys of quasars and radio galaxies. CP asserts that the universe is isotropic and homogeneous on large scales. It attributes the Cosmic Microwave Background (CMB) thermal dipole to our local peculiar motion [1], hence giving it the name of kinematic dipole. If this attribution is correct, then all sky surveys of other cosmological probes should show a similar dipole in their distribution throughout the sky. More than forty years ago, researchers postulated the presence of a number count dipole in source distribution (dubbed as the matter dipole) as a test for the CP [2]. However, recent research [3] has found a disagreement between the matter dipole and kinematic dipole, with claims reaching well over 5σ! [4,5] I will put the CP to test by analysis of all sky surveys of quasars [6] and radio galaxies [7,8]. Using Bayesian statistics, I will show that a dipole aligned with the CMB is present. However, the amplitude of the dipole is still in tension with our expectations, bringing the validity of the cosmological principle into question [9,10]. We will also look into the future, discussing some outstanding issues in the domain and possible methods to increase the sophistication of the number count dipole test. [11,12]. [1]. Planck Collaboration et al., A&A, 641, A1 (2020). [2]. Ellis G. F. R., Baldwin J. E., MNRAS, 206, 377 (1984). [3]. Kumar Aluri P., et al., Classical and Quantum Gravity, 40, 094001 (2023). [4]. Secrest et al., ApJ, 937, L31 (2022). [5]. Dam L., Lewis G. F., Brewer B. J., MNRAS, 525, 231 (2023). [6]. Storey-Fischer et. al., ApJ, 964, 69, (2024). [7]. Hale C. L., et al., Publ. Astron. Soc. Australia, 38, e058 (2021). [8]. Condon et al., AJ, 115, 1693 (1998). [9]. Mittal V., Oayda O. T., Lewis G. F., MNRAS, 527, 8497 (2024). [10]. Oyada O. T., Mittal V., Lewis G. F., Murphy, T., MNRAS, 531, 4545 (2024). [11]. Oyada O. T., Mittal V., Lewis G. F., MNRAS 537, 1 (2024). [12]. Mittal V., Oyada O. T., Lewis G. F., OJAp 8, 143 (2025).

Orateur: Dr Vasudev Mittal (School of Physics, The University of Sydney)

12:40 Lunch break

14:30 Field Level Inference for DESI (TBC)

Orateur: Pauline Zarrouk (LPNHE)

15:15 GenSBI: Flow Matching and Diffusion Models for Simulation-Based Inference in JAX

Flow and diffusion generative models are rapidly gaining traction in simulation-based inference (SBI). Beyond neural posterior estimation, these methods naturally extend to neural likelihood estimation and joint density estimation, providing a versatile framework for amortized Bayesian inference with intractable likelihoods. Their strong theoretical foundations, with clear optimization objectives and convergence guarantees, combined with the freedom to use any neural architecture as backbone, have established them as the state of the art on generative modelling benchmarks and driven their rapid adoption in the physical sciences, from gravitational-wave parameter estimation to cosmological field-level inference. In parallel, the JAX ecosystem is becoming the framework of choice for computational cosmology, driven by its support for differentiable simulations, composable program transformations, and seamless hardware acceleration on GPUs and TPUs. Yet the most widely used SBI libraries remain PyTorch-based, creating a disconnect for researchers who develop their forward models and analysis pipelines in JAX. We introduce GenSBI, an open-source, JAX-native library that implements flow matching and score-based diffusion as interchangeable density estimation backends under a unified interface. The library ships with three transformer-based architectures adapted from the recent generative modelling literature and provides an end-to-end workflow entirely native in JAX, from training through posterior calibration (SBC, TARP, LC2ST). Worked examples cover neural posterior estimation on unstructured data, time series, and images, and show how to integrate custom architectures and domain-specific embedding networks for virtually any inference task. GenSBI is designed from the ground up to be extensible at every level, from generative method and neural backbone to solver and inference mode, so that researchers can apply the library out of the box or build new methodologies on top of it as the field advances. The library and all examples are publicly available at gensbi.com.

Orateur: Aurelio Amerio (University of Valencia, IFIC & CSIC)

15:35 Coffee break

16:05 Applications of higher order stats and diffusion models to weak lensing data

Three-point correlations and wavelet based statistics are two quite different varieties of higher order statistics. I will show how they can greatly improve our constraints on systematic uncertainties, including detection of effects not possible with two-point correlations, and thus enable more robust cosmology. I will also introduce analytical and diffusion model approaches to reconstructing weak lensing mass maps from survey data. The trade offs in using diffusion models will be discussed. I will show results of high resolution mass maps and their cross correlations with visible tracers.

Orateur: Bhuvnesh Jain (U Penn)

16:50 Explicit vs. Implicit Likelihood Inference for Weak Lensing Cosmic Shear

Stage-IV weak lensing surveys such as Euclid and the Vera C. Rubin Observatory will deliver unprecedentedly precise cosmological measurements. Fully exploiting their potential requires inference frameworks capable of capturing the rich cosmological information mostly encoded at highly non-linear scales. Traditional likelihood-based approaches often rely on Gaussian assumptions that may break down for higher-order statistics, while modern simulation-based methods bypass explicit likelihood modelling by learning the posterior directly from simulated data — yet whether this added flexibility is truly necessary remains an open question with major implications for the cost and interpretability of future analyses. We present a systematic comparison of Explicit Likelihood Inference (ELI), combining Gaussian Process emulation with a Gaussian likelihood, and Likelihood-Free Inference (LFI) via Neural Density Estimators, applied to a common forward-modelling simulation suite in a Euclid DR3-like configuration constraining Ωm and σ8​. Both pipelines are evaluated across a hierarchy of Gaussian, non-Gaussian, and CNN-based map-level estimators, with posterior calibration validated through a comprehensive diagnostic framework. By directly contrasting these two approaches under controlled conditions, our work provides concrete guidance on the robustness, accuracy, and practical trade-offs of traditional and deep-learning-based inference strategies for next-generation weak lensing cosmology.

Orateur: Dr Simone Vinciguerra (LAM)

jeu. 18 juin

10:00 → 10:45 Generative Solutions for Cosmic Problems

In the era of high-dimensional data and simulation-based science, machine learning is transforming the different stages of the scientific method in astrophysics. I will present a summary of my research connected to generative models in two main areas: First, I will discuss how generative models enable high-dimensional simulation-based inference for cosmology and provide a roadmap to field-level inference. Hydrodynamical simulations are essential for self-consistent predictions of diverse cosmological observables—including the kinetic and thermal Sunyaev-Zel'dovich effects (kSZ, tSZ), fast radio bursts (FRBs), baryonification effects in weak lensing, and galaxy formation. However, their computational expense has traditionally limited their use in large-scale inference. I will show how machine learning emulators of these simulations unlock inference at scale, and how generative models can interpolate across different hydrodynamical simulators that employ varying subgrid assumptions and model different physical processes. Second, I will also present a complementary approach to disentangle representations arising from instrumental systematics versus the underlying astrophysics in a fully data-driven way.

Orateur: Prof. Carolina Cuesta-Lazaro (MIT)

10:45 → 11:05 TBD

11:05 → 11:35 Coffee break

11:35 → 12:20 Full-catalog field-level inference in 2.5 dimensions for imaging surveys

I will describe our progress in conducting field-level cosmological inference from the galaxy catalogs of imaging surveys. We incorporate both the galaxy-density and weak lensing information yielded by every galaxy above some flux threshold, marginalizing over each galaxy's redshift during the inference and without creating redshift bins. The mass field is parameterized in voxels of depth 100-200 Mpc, to match the best line-of-sight resolution available from color information. This reduces the number of latent field parameters >10x, but also precludes use of most analytically inspired galaxy occupation models. We instead develop simple parametric galaxy occupation probabilities which treat line-of-sight structure as stochastic.

Orateur: Prof. Gary Bernstein (University of Pennsylvania)

12:20 → 12:40 Source clustering impact on weak lensing high-order statistics

Orateur: Dr Seung-gyu Hwang (CEA/LCS)

12:40 → 14:30 Lunch break

14:30 → 15:15 Higher-order baryonic modelling for the Euclid mission

Next-generation large-scale structure surveys such as Euclid will deliver percent-level measurements of the cosmic matter distribution, pushing cosmological analyses deep into the nonlinear regime. On these scales, baryonic processes such as AGN feedback and galaxy formation significantly affect matter clustering, posing a major challenge for cosmological inference.

While most existing approaches model baryonic effects at the level of two-point statistics, higher-order statistics such as the matter bispectrum encode additional non-Gaussian information and are essential for fully exploiting forthcoming data. After a brief overview of baryonic effects in higher-order weak lensing statistics, I will present a fast and accurate method for incorporating baryonic effects into higher-order weak lensing analyses. The approach builds on high-resolution simulations combined with post-processing techniques such as cosmology scaling and baryonification to efficiently explore a wide cosmological and astrophysical parameter space. We train a deep neural network to emulate baryonic modifications to the matter bispectrum across scales and redshifts relevant to the Euclid mission and other Stage IV surveys.

We show that our emulator satisfies the precision requirements of the first Euclid data release and enables robust inference of the cosmological information of the small-scale matter bispectrum, further enhancing the constraining power achievable with two-point analyses. More broadly, our work highlights the power of emulation techniques to bridge complex simulations and parameter inference, maximising the scientific return of upcoming surveys.

Orateur: Dr Giovanni Aricò (INFN Bologna)

15:15 → 15:35 Starlet l1 norm for HI intensity mapping

Neutral hydrogen (HI) intensity mapping provides a powerful probe of large-scale structure, containing significant non-Gaussian cosmological information, particularly at late times in the evolution of the Universe. However, traditional two-point statistics, such as the angular power spectrum, fail to capture this information. In this work, we employ the starlet l1​ norm, a multi-scale higher-order statistic that is robust to instrumental effects and well suited for analyzing non-Gaussian features in cosmological fields. Using a simulation-based inference (SBI) framework, we quantify the improvement in cosmological parameter constraints obtained with the starlet l1​ norm compared to those derived from the angular power spectrum. Our analysis demonstrates how incorporating multi-scale higher-order information can enhance the constraining power on cosmological parameters in HI intensity maps.

Orateur: Pauline Gorbatchev (FORTH)

15:35 → 16:05 Coffee break

16:05 → 16:25 Tracing cosmic structure with neutral hydrogen after the Epoch of Reionization

The 21-cm signal produced by neutral hydrogen (HI) is a unique tracer of Large-Scale Structure (LSS) throughout cosmic history. It is the only probe of the matter distribution during the Dark Ages, and enables insight into the astrophysics of the ionization process during the Epoch of Reionization (EoR). After the EoR has concluded, only dense regions toward halo centres (which can self-shield from the ionizing photons) remain neutral and thus detectable in the 21-cm line. Therefore, post-EoR, the HI signal returns to being a strong (if biased) tracer of the underlying large-scale structure. From this late-time signal, we can trace structure formation across the universe for z < 6. Recently, observations have suggested a late end to reionization at z ~ 5.3. Considering this and the effects of inhomogeneous recombination, it is likely that small patches of neutral IGM will remain during the time that halo-based HI emits significantly, resulting in a highly non-Gaussian distribution of neutral hydrogen. In this work, we present a semi-numerical simulation of the neutral hydrogen within halos, which is combined with the outputs from the open-source EoR code 21cmFAST to produce a robust representation of the emission from all HI sources. We make predictions on the detectability of a transition signal within future intensity mapping surveys, and investigate the effect of varying EoR astrophysics on the observed power spectrum. Our simulations predict a drop in power of four orders of magnitude between 4 < z < 7. Assuming an inhomogeneous recombination model, we find a flattening of the power due to lingering neutral islands masking the late-time HI signal for 5 < z < 6.5. Using SKA-Low deep survey parameters, we find HI power spectrum detectability at scales k < 1 h/Mpc for redshifts 3 < z < 7, even when using a horizon limit avoidance scheme to mitigate against foregrounds. Our results suggest a sufficient SNR of the HI power spectrum tracing the underlying halos z < 5, which can be used for late-time cosmology; and that the resulting Ω_HI constraints can trace different reionization scenarios such as a decreased escape fraction. This study implies that SKA-Low intensity mapping observations for 3 < z < 7 will be an important probe to constrain reionization parameters as well as cosmological models. In future work, we will go beyond the power spectrum to look at the evolution and detectability of the HI bispectrum and trispectrum throughout this highly non-Gaussian transition period.

Orateur: Dr Jamie Incley (University of Manchester)

ven. 19 juin

10:00 → 10:45 'Emulator is all you need' for the reionization inference with SKA-Low

The redshifted 21-cm signal from the Epoch of Reionization (EoR) provides a direct probe to the evolution of the state of the intergalactic medium (IGM). Further, its inherent strong and evolving non-Gaussian nature demands a sophisticated signal estimator and a robust inference framework. In this talk, I will discuss our efforts towards the development of machine learning based signal emulators designed to accelerate parameter inference by circumventing computationally expensive simulations. Our early approaches focused on emulating different summary statistics of the signal. We first employed Artificial Neural Networks (ANNs) to predict the 21-cm power spectrum and bispectrum given a set of reionization parameters. This speeds up the inference computation time by several folds. Additionally, we showed that one can achieve 2-4 times tighter constraints on EoR parameters when utilizing all unique k-triangle 21-cm bispectrum with respect to the power spectrum. However, ANN emulators suffered from a critical limitation: they produced only point-value predictions without quantifying associated uncertainties and were highly erroneous in their predictions for the late stages of the EoR. We improved the performance of this emulator by introducing Bayesian Neural Networks (BNNs) architecture. BNN emulators provide posterior distributions of predicted statistics, including prediction uncertainty, demonstrating superior performance, especially for smaller training datasets. These BNN-based emulators proved particularly valuable for emulating the complex signal bispectrum, which exhibits high dynamic range and sign variability across triangular configurations in Fourier space. We then moved to the next frontier i.e. to build field-level emulators. The latest development, CosmoUiT, represents a paradigm shift: a hybrid Vision Transformer-UNet architecture that directly emulates entire 3D 21-cm brightness temperature cubes across the EoR parameter space. This architecture uniquely captures both large-scale morphological features through transformer-based self-attention mechanisms and small-scale variations through convolutional pathways, while remaining conditioned on input reionization parameters. CosmoUiT achieves remarkable computational speedup—orders of magnitude faster than original simulations—while maintaining high fidelity to simulated maps across multiple validation metrics. This progression from compressed summary statistics to full field emulation represents a fundamental shift in EoR analysis methodology, enabling field-level Bayesian inference pipelines possible for the analysis of the SKA-Low observations of the EoR while preserving the physics encoded in the non-Gaussian 21-cm signal structures.

Orateur: Prof. Suman Majumdar (Indian Institute of Technology Indore)

10:45 → 11:05 Impact of non-Gaussianity on EoR 21cm signal detection and inference

The EoR 21cm signal is highly non-Gaussian due to formation and overlap of the ionised regions in the IGM. This non-Gaussianity gives rise to higher-order statistics (bispectra, trispectra, etc.) of the signal. The trispectrum contribution affects the cosmic variance of the observed power spectrum resulting into a saturation in the signal-to-noise ratio (SNR) due to non-zero trispectrum contribution. This effect are not pronounced for the current interferrometers, however due to uprecedented sensitivity of the SKA, it considerably affects the SNR at scales k<0.5 Mpc^-1, particularly during the mid and end stages of reionisation. This also impacts any further inference based on the powerspectrum measurements. Non-Gaussianity of the signal is also treasures information about the sources and the process of reionisation. The bispectrum is the lowest-order non-Gaussian statistics. We show that due to improved sensitivity SKA will be able to measure the bispectrum of the EoR 21cm signal, and using bispectrum in inference will yield more tighter constraints over the astrophysical parameters. We find that 100-1000 hrs of SKA observations are sufficient to obtain better contraints over ionisation parameters.

Orateur: Abinash Kumar Shaw (Max Planck Institute for Astrophysics)

11:05 → 11:35 Coffee break

11:35 → 11:55 The Signature of Strong High-Redshift Radio Backgrounds on the Cosmic Dawn 21-cm Bispectrum

Measurements from the Absolute Radiometer for Cosmology, Astrophysics, and Diffuse Emission 2 (ARCADE-2) reveal a strong radio background in the GHz frequency range. Since the cosmological 21-cm signal is measured relative to the background radiation temperature, the presence of a radio excess can significantly alter its characteristics. Previous studies have explored the impact of an inhomogeneous radio background on the global 21-cm signal and 21-cm power spectrum. This non-uniform radio background is also expected to introduce substantial non-Gaussianity. In this work, using the bispectrum, we analyze the non-Gaussianity in the 21-cm signal in the presence of an excess galactic radio background and investigate how line-of-sight radio fluctuations from early galaxies influence its nature. We find that even a moderate enhancement in radio efficiency in early galaxies significantly affects the small-scale 21-cm bispectrum. Furthermore, the delayed heating transition caused by a galactic radio background shifts the sign change in the squeezed-limit bispectrum to lower redshifts (z ~11), providing a potential observational signature for distinguishing different radio background models. These results demonstrate that the 21-cm bispectrum, particularly in the squeezed limit, is highly sensitive to radio background fluctuations, making it a powerful tool for probing high-redshift radio-loud sources and the physics of the early cosmic epoch.

Orateur: Dr Sudipta Sikder (Tel Aviv University)

11:55 → 12:15 Modelling the Effects of Primary Beam Heterogeneity in SKA-Mid for 21cm Cosmology

Orateur: Dr Tobias Russell (The University of Manchester)

12:15 → 12:35 Capturing non-Gaussian cosmic information with Neutral Hydrogen one-point statistics

Orateur: Dr Bernhard Vos (Bielefeld University)

12:35 → 12:55 Foreground removal in HI 21 cm intensity mapping under frequency-dependent beam distortions

Neutral hydrogen (HI) intensity mapping with single-dish telescopes is a powerful probe of post-reionization cosmology, but extracting the signal is challenging due to bright foregrounds and realistic, frequency-dependent beam effects that degrade most standard methods. I present an evaluation of SDecGMCA, a spherical extension of GMCA that combines sparse component separation with explicit beam deconvolution, using simulations tailored to MeerKAT and SKA-Mid. Compared with model-fitting and other blind source separation techniques, SDecGMCA remains robust under complex beam distortions, suppresses spurious spectral features, and recovers the HI power spectrum to better than 5% accuracy on intermediate angular scales (10<ℓ<200). I will show that an improved version of SDecGMCA, which uses learnlets instead of starlets performs even better removing previous limitations of the method.

Orateur: Athanasia Gkogkou (IA/ICS-FORTH)