Share
Embed
GT ICR - Ian Hothi - Exploring the Epoch of Reionization with Scattering Transforms
François Orieux
September 30, 2026
Transcript
88 slides
-
Exploring the Epoch of Reionization with Scattering Transforms Ian Hothi
Collaborators: Erwan Allys (LPENS) , Benoit Semelin (LUX), Romain Meriot (Imperial College), Francois Boulanger (LPENS) 1 -
The Epoch of Reionization Why is it so important? •
Sets the stage for the Universe we see today • The rst structures formed during this Epoch • The IGM, today, is ionized and it is during this period where the IGM went from neutral to ionized • ItÕs a 1 billion year epoch in the history of the Universe that we know little fi about 4 -
Indirect Constraints Constraints from Planck & The G-P Trough ¥Planck:
Z_start≈14 ¥G-P Trough: Z_end≈6 ¥ 7 -
Indirect Constraints Constraints from Planck & The G-P Trough ¥Planck:
Z_start≈14 ¥G-P Trough: Z_end≈6 ¥ As we have accumulate more Quasar spectra, Bosman+18, the end of reionization appears to be more extended ending at z = 5.3 (Qin+21) ¥ 8 -
The 21cm Signal fi Hyper ne Transition Line of Hydrogen
10 Oyama (2015) -
11 The 21cm Signal A Probe of the IGM dIν
= − Iν + Bν(ν) dτν -
12 The 21cm Signal A Probe of the IGM dIν
= − Iν + Bν(ν) dτν 2 Iνc Tb(ν) ≈ 2 2kBν -
13 The 21cm Signal A Probe of the IGM dTb
= − Tb + TCMB dτν -
14 The 21cm Signal A Probe of the IGM Tb(τν)
= TS(1 − e −τν ) + TCMBe −τν TS = Tb(τν = 0) -
The 21cm Signal A Probe of the IGM 2 TCMB
Ωbh 0.24 H(z) 1+z δTb = 28(1 + δ)xHI 1 − ( TS )( 0.0223 ) ( Ωm )( 10 )[ δrvr + H(z) ] 15 -
The 21cm Signal A Probe of the IGM 2 TCMB
Ωbh 0.24 H(z) 1+z δTb = 28(1 + δ)xHI 1 − ( TS )( 0.0223 ) ( Ωm )( 10 )[ δrvr + H(z) ] Matter Density 16 -
The 21cm Signal A Probe of the IGM 2 TCMB
Ωbh 0.24 H(z) 1+z δTb = 28(1 + δ)xHI 1 − ( TS )( 0.0223 ) ( Ωm )( 10 )[ δrvr + H(z) ] Matter Density Neutral Fraction 17 -
The 21cm Signal A Probe of the IGM 2 TCMB
Ωbh 0.24 H(z) 1+z δTb = 28(1 + δ)xHI 1 − ( TS )( 0.0223 ) ( Ωm )( 10 )[ δrvr + H(z) ] Matter Density Neutral Fraction Cosmology 18 -
How to Observe the EoR Towards a rst detection Current
interferometers do not have the sensitivity to detect tomographic maps of the EoR fi They instead aim for a statistical detection of the EoR 28 -
How to Observe the EoR Power Spectrum 2D Sky Redsh
ift/Fre 29 quenc y -
How to Observe the EoR Power Spectrum P(k) k k=
33 2 2 2 kx + ky + kz -
Detecting the EoR How to mitigate these contaminants? Foreground Avoidance
38 Foreground Removal -
Detecting the EoR How to mitigate these contaminants? Foreground Avoidance
k∥ = | kz | k⊥ = 40 2 2 kx + ky -
Detecting the EoR How to mitigate these contaminants? Foreground Removal
Observed Signal 21cm Signal 41 Jelić et al. (2008) -
Detecting the EoR How to mitigate these contaminants? Foreground Avoidance
Foreground Removal k∥ = | kz | 42 k⊥ = 2 2 kx + ky Jelić et al. (2008) -
Detecting the EoR Power Spectrum Upper Limits 43 Courtesy of
Florent Mertens -
Detecting the EoR What are the current limitations? • Current
foreground removal techniques work well, but not well enough • They are highly-in uenced by the instrument • They are focused on detecting the power spectrum, and so miss out on key information from the non-Gaussian nature of the 21cm signal. • As the 21cm signal is weak, any residual 45 ff fl systematic or foreground can a ect the Ôobserved signalÕ -
Moving Beyond The Power Spectrum 21cm Map Emission 46 The
21cm Signal is highly non-Gaussian -
Moving Beyond The Power Spectrum The Bispectrum • The instrument
highly impacts the nonGaussianity of the signal • Changing itÕs amplitude and sign • Current foreground removal techniques are unable to recover the bispectrum • Further showing their reliance on power spectrum recovery • There is a need for a low-variance probe of the non-Gaussianity 48 -
Moving Beyond The Power Spectrum Scattering Transforms ∫ S +
⊛ ⊛ + Secondary Convolution - Secondary wavelet probing larger scales 54 -
Scattering Transforms Current Applications Building a Family of Statistics •
Wavelet Scattering Transforms [Allys et al. (2019)] • Wavelet Phase Harmonics [Allys et al. (2020)] • Scattering Covariances [Cheng et al. (2023)] Parameter Inference • Interstellar Medium [Allys et al. (2019), Regaldo et al. (2020), Lei et al. (2022)] • Large Scale Structures [Eickenberg et al. (2022), Valogiannis et al. (2022a, 2022b)] • EoR [Grieg et al. (2022), Hothi et. al (2024)] Component Separation • Dust Emission [Auclair et al. (2024), Regaldo et al. (2024)] 55 -
Scattering Transforms Component Separation • Scattering Transforms have been successfully
used for component separation on 2D Maps ¥ Right: Dust separated from CIB • However, the EoR is a more complicated task • The signal to foreground ratio << 1 • The statistics of the foregrounds are not well known • The observed data are 3D Lightcones not 2D maps 57 Auclair et. al (2024) -
Scattering Transforms Component Separation • Scattering Transforms have been successfully
used for component separation on 2D Maps ¥ Right: Dust separated from CIB • However, the EoR is a more complicated task • The signal to foreground ratio << 1 • The statistics of the foregrounds are not well known • The observed data are 3D Lightcones not 2D maps 58 -
Developing 3D Scattering Transform Development of a 3D Wavelet Set
• This is non-trivial, but to get a good sampling of the cylindrical plane: ⃗ • Consider a 1D Gaussian wavelet de ned along the k -axis, ϕ ( k ) z z z ⃗ • Consider a 2D Gaussian wavelet de ned in the k − k plane, ϕ ( k ) x ϕ ( k ⃗) z z fi fi ⨂ 61 y ⃗ ) ϕxy( k xy xy xy -
Generative Models of EoR Lightcones How well does the 3D
Scattering Transform Summarise? White Targe t Noise Light cone ϕstarting ϕt Gradient Descent Loss function to be minimised 2 ℒ(ϕt, ϕi) = | ϕt − ϕi | Synth esise ST coe cient at iteration i 63 ffi ϕend ≈ ϕt d Lig htcon e -
Generative Models of EoR Lightcones Can we reproduce other independent
statistics? We take a single sample (Lightcone) and apply the 3D scattering transform statistic From this single sample, we generate 30 realisations 64 -
Conclusions I Generative Model • We have developed a 3D
Scattering Transform Tool: • It is able to fully characterise an EoR Lightcone • We are able to produce generative models of Lightcones, that can reproduce the statistics of the target eld • Future Goals: • Develop a component separation tool fi • Apply to real interferometric data 67 -
Parameter Inference Goals: • We want to produce a statistic
that is able to take advantage of the spectroscopic nature of the data - 2+1 Statistics • We want this statistic to be compressed, keeping the essential information with the fewest components • Use sher analysis to see how informative our statistic versus traditionally used statistics fi 69 -
Summary of the Statistics Compared 74 Statistics Wavelet + Scaling
3D Spherically averaged PS 3D Gaussian + Log10 binning 2D PS 2D Gaussian + Log10 binning WST_w 2D Gaussian + Log10 binning -
How will we compare them Fisher Analysis ∂S ∂S −1
θ Σ Fij = ∂θi ∂θj S represents a vector encompassing the expected values of the statistics θi corresponds to a given parameter of the simulation, that we try to constrain Σ corresponds to the covariance of these statistics, which arises from cosmic variance or the thermal noise in high-noise cases Fisher Matrix relates to the minimum variance of any unbiased estimator 76 -
How will we compare them Fisher Analysis + θi +
Sθi fid θi − θi − Sθi ∂S ∂S −1 θ Σ Fij = ∂θi ∂θj 77 -
How will we compare them Fisher Analysis + θi δθi
+ Sθi fid θi − θi − Sθi ∂S ∂S −1 θ Σ Fij = ∂θi ∂θj 78 -
How will we compare them Fisher Analysis + θi δθi
+ Sθi δS fid θi − θi − Sθi ∂S ∂S −1 θ Σ Fij = ∂θi ∂θj 79 -
How will we compare them Fisher Analysis + θi δθi
+ Sθi δS fid θi − θi − Sθi ∂S ∂S −1 θ Σ Fij = ∂θi ∂θj 80 -
81 Simulation information ¥ We use 21cmFast for the simulation:
−1 ¥ 200h Mpc (128x256x256) ¥ 128 freq./redshift channels at SKA resolution ¥ Simulated between z = 8.82 (144.60 MHz) and z = 9.33 (137.46 MHz). ¥ We vary the following parameters: ± T : 50000 5000 K ¥ vir ¥ Rmax= 15 ± 5 Mpc ¥ ζ= 30 ± 5 -
Compression of Redshift Evolution For however nk coefficients of a
statistic, we have 128 redshift points, leading to 128nk coefficients ϕk(z) k Z 82 -
Compression of Redshift Evolution We then compress this evolution, by
applying continuous wavelets to the redshift evolution ϕk (J) compressed k J 83 -
Results s The statistics will be denoted by ϕ̄l:j s
is the statistic l is the summary used on the evolution along the z direction j are the scales that are summarised along the z direction. WM For example, ϕ̄ℓ 1,ℓ 2:1,2 represents the evolution-compressed wavelet moments, 1 2 84 computed at j = 1 and j = 2 scales for both ℓ and ℓ norms. -
Conclusions II Parameter Inference • We have developed a 2+1
Scattering Transform Formalism that can summarise 3D EoR data • We have shown that in this 2+1 regime, the Scattering Transform provides tighter parameter constraints that the traditionally used statistic (the power spectrum) • Even in high noise cases, the 2+1 outperforms the spherically-averaged power spectrum 87 -
General Conclusion Exploring the EoR with Scattering Transforms • We
have developed 3D scattering transform statistics tailored to the EoR • Component Separation: Developed a 3D wavelet set that can appropriately probe the cylindrical space of the signal, without mixing di erent regions • Showed that this 3D scattering transform statistic can capture the salient features of the signal, paving the way for future component separation methods • Parameter Inference: Developed a summarised 2+1 scattering transform ff that can take advantage of the spectral nature of the 21cm signal to enhance parameter constraints 88