Notes and discussions
Anze Slosar
Notes:
Surveys interacting with S4: live on k-plane
Taxonomy of surveys:
projected tracers (CMB lensing, WL, GC): low k_perp, no k_par
Spec surveys: all plane
Foregrounded: no kpar
Cross-correlation: overlap in k modes
Proj: no redshift specificity
Spectro: full 3D, xcorr with CMB only small slice, high SN
Foregrounded: no kpar, 21cm, do not overlap with CMB lensing (projected)
Redshift specificity helps a lot - 3x2pt, CMB lensing
Higher order stats: allow you to combine things that don’t necessarily overlap on k plane
Field reconstructions to recover lost k modes, e.g. 21cm, reconstruct initial modes
CMB: lensing (complete, sample-variance cancelation), tSZ/kSZ, ISW, moving lens, patchy reionization - xcorr with 21cm
Combine kSZ ML: Hotinli et al.
Questions:
Most promising avenue for CMB x 21cm?
→ Probably tidal reconstruction or lensing reconstruction
Utkarsh Giri
Notes:
kSZ: CMB photon scattering by moving electrons
Sec. anisotropy prop to velocity and electron density
Fixed realization of vr produces non-vanishing correlation between deltag and T -> reconstruct velocity mode
Vr probes large scale cosmological modes, on large scales: cosmological fields linearly related, on large scales - vr reconstruction noise smaller than GC shot noise
Application: fNL, sample variance cancelation due to additional tracers
Use vr as additional tracer -> strong fNL constraints
Question: rely on linear noise models, nonlinearities?
Use sims to investigate noise assumptions: 2-3 times larger noise for reconstructions due to nonlinear noises (similar to CMB lensing biases) N1, N3/2
Noises can be captured analytically
Use sims to check SVC -
Questions:
Blake: is it obvious why the N3/2 bias is a lot bigger than the N1? (Sorry if I missed this.)
Manu: How important is accurate modeling of N32?
Colin: what sets the scale at which N3/2 peaks?
David Alonso
Notes:
Projected tracers, e.g. LSST, tracers of matter fluctuations
Focus on CMB lensing and WL, GC
tSZ tomography: cross-correlate tSZ with clustering - constrain gas pressure / mass bias
Cluster mass calibration - high-z clusters
Tomography: reconstruct redshift dependence, Nx2pt
Use tomography to constrain growth - break bias - growth degeneracy with CMB lensing
Growth reconstruction: check probe consistency, consistency with LCDM
Two data sets: DECALS, KiDS / DES / Planck
Decouple growth from background - growth spline with nodes
S8(z): data comb gives evidence of lower growth at 0.2<z<0.6 -> driven by shear
Need CMB lensing for high-z
LCDM good fit with lower S8
Systematics:
Photoz uncertainties: analytic marg, self-calib, CMB x less sensitive
Shear calibration:
Galaxy bias: hybrid EFT
Questions:
Alex Krolewski
Notes:
Improve on CMB lensing auto with GC - growth of structure, brings in galaxy bias -> need all probes
Related to S8 tension
Use CMB lensing as xcheck of optical lensing
Pick galaxy sample with high S/N
High z - overlap with lensing kernel
Angular coverage
WISE: infrared galaxy survey: all-sky, gals with old stellar pops can be easier to detect in WISE
Three galaxy samples: blue, red, green z~0.6 - z~1.5
Need to remove stars using GAIA - contamination left 1%
Redshift distribution: no photoz - > do cross-correlation redshifts with specs, measure n(z) and bias evolution, so just include measured relations in theory
Theory: linear + HO bias (dominated by linear modes), allows fix cosmo & HO bias
Tests on mocks, calibration of scales used, lmax~250-300
Redshift uncertainty: run sep. chains for all nz realizations -> full shape marg.
Results: samples consistent, 2.6sigma tension with Planck in S8
Questions:
David A.: Can you say a bit more about the N(z) marginalization using clustering redshifts? E.g. do you need to worry about the fact that the measurements allow for negative values?
Blake: If I recall right, DES obtained the same S8-only constraint but did not report tension with Planck when considering the full posterior. Could you remind me how you obtained the 2.6sigma tension number and could a similar effect operate here?
Yan-Chuan Cai
Notes:
ISW and CMB lensing around substructures
ISW and CMB lensing: Sourced by same late-time grav potential
ISW: probe of accelerated expansion - time-varying potentials
Photons in voids: cold-spot in CMB, photons in clusters: hot spot
Usually: cross-correlations between primary CMB and LSS tracers (e.g. lensing)
Joint measurements of GC, CMB lensing, ISW
Why superstructures?
Look at generalization of cluster cosmology, peak counts, etc.
Beyond 2pt
Beyond GR - different behaviour in high/low density regions
Granett et al., 2008: amplitude of stacked CMB temperature high compared to LCDM expectation
Question: what is the cause for this?
Repeat analysis using DESI legacy survey
reduce sample variance
No detection for temperature imprints from supervoids
Questions:
Jia: why is there no lack of signal for void (but it is there for clusters..)? I.e.why “lensing is low” is not impacting void? Answer: could be related to photo-z bias, or something physical (e.g. neutrino free-streaming) but no definitive answer.
David A.: there was a recent paper measuring a negative ISW from voids at very high redshifts from QSOs. Do you have any thoughts on that? (this one)
Yan-Chuan Cai: Yes. I think the statistical significance isn’t very high,less than 3sigma I think, similar to many other analyses. It would be great to have more volume to beat down the variance.
Leander Thiele
Notes:
NNs to map N-bodies to baryons: CNN / DeepSet
Predict baryonic fields from DM only - baryons local, i.e. use local ML approach
Usage:
Rapidly generate sims
Interpret trained model and learn about astro
CNN:
DM -> pe, ne
Use CNN on 3D field, redshift zero
Semianalytic models
Main problem: sparsity, only small fraction of volume is interesting, overcome by biasing training sample using zoom-ins
Electron pressure spans large dynamic range (input trafos, semi-analytic model)
Small scales well fit, large scales better than models, projection improves performance
DeepSet:
Concentrate on massive halos -> no transl symmetry
CNNs not best approach: spend lot of resource on empty regions, poor interpretability
Idea: can we use simulations representation to train NN? - DeepSet VAE
Restricted architecture improves on existing models
Questions:
Interpolate between baryonic feedback models? Train on one sim and then do transfer learning on other sim
Jia Liu
Notes:
Why sims? Test pipelines, systematics, astrophysics, covariance, modeling
Why correlated? CMB foregrounds and LSS, not super useful for survey systematics because will be uncorrelated, but extragalactic, astrophysics syst will show up differently, input to train ML models
Typical CMB sims: 2D, gravity only, CMB observables painted -> Sehgal, Stein
Typical LSS sims: 2/3D, more cosmo, hydro/gravity only, curved/flat, smaller boxes
Correlated sims need to accommodate both worlds
current : run by experts in both areas
Yuuki Omori: MultiDark Planck 2
Roadmap:
Basis: gravity-only sims
Then send off to experts and paint on CMB / LSS observables
Coordination!
How to get gravity-only sims: fast full-sky lightcone (e.g. COLA, FastPM)
Challenges: computing, storage, requirements, maintenance, personnel
Questions:
Dongwon ‘DW’ Han
Notes:
Need correlated sims: high-res, multi-frequency, need NG info
Computationally expensive -> use DL
mmDL: CMB kappa -> NG kappa, kSZ, tSZ, CIB, radio + lensed TUQ
Network:
Data augmentation, not enough sims
Conversion / prediction
NG restoration step
Results:
Reproduce source counts, power spectra, cross-correlations, bispectra, trispectra (CMB lensing biases)
Network can recover correlations between large and small scales even though trained on small patches
Variance in sims comparable to Knox
Sims available publicly
Allow mass-production of independent full-sky realizations
Fast: forward-modeling
Future:
Fixed cosmology - CAMELS?
Missing associated catalog - extension to create catalogs
Can we use network to learn optimal summary stats?
Questions: