Difference between revisions of "Harvard-2017:T3"

From CMB-S4 wiki
Jump to navigationJump to search
 
(3 intermediate revisions by 3 users not shown)
Line 1: Line 1:
 
Back to [[Harvard-2017:_Cosmology_with_CMB-S4#Friday.2C_25_August| Harvard-2017 main page]]
 
Back to [[Harvard-2017:_Cosmology_with_CMB-S4#Friday.2C_25_August| Harvard-2017 main page]]
 
== Parallel Session    T3: Sim WG: sky modeling, component separation and lensing reconstruction(Chair: Blake Sherwin) [Jefferson 256]==
 
== Parallel Session    T3: Sim WG: sky modeling, component separation and lensing reconstruction(Chair: Blake Sherwin) [Jefferson 256]==
 +
Summary:
 +
[[File:s4Review.pdf]]
 +
[[:File:JCH_optimization.pdf| [Session summary] ]]
  
 
See draft schedule below.
 
See draft schedule below.
Line 14: Line 17:
 
- Discussion: open questions and (in particular) what are the next steps? (~10 mins, Raphael)
 
- Discussion: open questions and (in particular) what are the next steps? (~10 mins, Raphael)
  
[[File: Foregrounds.pdf]]
+
[[File:Foregrounds.pdf]]
  
  
Line 52: Line 55:
  
 
== Notes from session ==
 
== Notes from session ==
 +
Note-taker: Jason Henning
  
 +
'''Notes for Foreground section:'''
 +
* Foregrounds dominate everywhere by orders of magnitude
 +
* Model YY=00
 +
** (not physical) model to validate Fisher forecasts in Science book with dust and sync with levels set by BICEP patch.
 +
** Multi-realization.
 +
** Gaussian realizations.
 +
* Model YY=01, 02, 03
 +
** contain instrument noise
 +
** limited resolution
 +
** assumed spectral dependence
 +
** (Different flavors of PySM)
 +
*** differ by freq dependence of dust and sync, and AME
 +
 +
* All models lead to small foreground residuals for S4 strawperson design
 +
** ~ order of magnitude under signal.
 +
*** Partly due to high correlation between freq.
 +
** ILC prescription for foreground removal.
 +
*** BICEP hasn’t achieved same residuals as Raphael.
 +
*** Raphael’s model just requires maps to be properly calibrated to T_cmb.
 +
 +
* Model YY=04
 +
** Ghosh, et al 1611.02418
 +
** based on GASS HI data
 +
** designed for larger decorrelation between freq.
 +
 +
* Model YY=05
 +
** toy model with even more decorrelation
 +
 +
* Models Y=04,05 lea to increase in sigma(r) of ~ 2 and bias for Raphael
 +
*** Bias is ~ size of signal we’re trying to reconstruct.
 +
** Discussion about how much decorrelation we need to account for and how to model it.
 +
*** Victor’s framework still wants large freq separation.
 +
 +
* Model YY=06
 +
** Based on MHD
 +
** assumes constant dust-to-gas ratio
 +
** assume energy spectrum of electrons for synchrotron.
 +
** Reproduces E/B ratios.
 +
** Limited to small patches.
 +
*** CP and RF disagree about how much this model would over-estimate correlation between dust and synchrotron at high Galactic latitude.
 +
** Resolution not good enough for use with lensing as well (sims don’t resolve turbulence).
 +
 +
** Two sources of decorrelation
 +
*** spatial variation of spectral index allowed by Planck.
 +
*** line-of-sight effects
 +
*** decorrelation between freq in dust is small.
 +
 +
** What is the expected level of decorrelation???
 +
** We need sims to understand how we address decorrelation!!
 +
** We need models that address degree scales and needs of delensing survey simultaneously.
 +
** No non-forced sims have been able to show large decorrelation.
 +
 +
'''Next-steps'''
 +
* Add decorrelation into sims, more data, integrate high/low-ell sims.
 +
* Test closely-spaced freqs
 +
 +
'''Alex '''
 +
* NG-Foregrounds
 +
** Models for small-scale pol dust
 +
*** Each gives different amount of BB power (while TT is tight).
 +
** Dust causes NG in lensing map.
 +
*** So far no bias to lensing auto-spectra for the best 5% of sky for Vansyngel+ sims
 +
*** 40% of sky there’s 1% bias in TT lensing, but still nothing in polarization.  Comforting, but only as good as the sims.
 +
 +
** Planck FFP8 sims (Challinor, Allison++2017)
 +
*** Big bias (17% in TT, 62% in EBxEB)
 +
*** Can downweight dusty areas and reduce bias in A_L from EBxEB (TT bias still there).
 +
*** How does NG of dust impact delensing?
 +
*** No average bias, but a lot of patch-patch scatter.
 +
 +
* Takeways:
 +
** TT is biased from dust.
 +
** Polarization has bias that can be removed with downweighting (assuming uniform power over patches).
 +
** No evidence that we do not need dust and synch channels for high-ell survey.
 +
** Can we extend this to multi-freq?
 +
** How much do we trust freq decoherence in sims?
 +
 +
 +
'''Colin Hill (Component Separation)'''
 +
* Simulated T maps
 +
** Missing small-scale Galac synch and AME.
 +
** All components properly correlated.
 +
* Simulated P maos
 +
* Constrained ICL
 +
** specifically project out whatever components you want.
 +
** Variance in final map larger since dof used.
 +
* If foregrounds are lower than noise floor ILC won’t pick them out but they can still bias lensing.
 +
* CP: Why aren’t we using these sims for low-ell forecasts as well?
 +
 +
'''Kyle Story (Lensed B-mode template)'''
 +
* “B-truth”: the best we could ever do
 +
** Leads to 99.9% delensing.
 +
* Quadratic estimator achieves ~ 90% delensing.
 +
* Next steps
 +
** Run these templates through r-estimation.
 +
** Add non-idealities to these maps?
 +
** Incorporate multi-freq, cleaned maps into pipeline.
 +
** LK: Investigating map-based delensing algorithms beyond QE is high-priority.
 +
** Move onto ML estimators.
 +
** '''Marius Millea'''
 +
*** Starting to look at beyond QE delensing (1708.06753)
  
 
== Action items/Next steps ==
 
== Action items/Next steps ==
  
 
Summarize action items here
 
Summarize action items here

Latest revision as of 11:31, 25 August 2017

Back to Harvard-2017 main page

Parallel Session T3: Sim WG: sky modeling, component separation and lensing reconstruction(Chair: Blake Sherwin) [Jefferson 256]

Summary: File:S4Review.pdf [Session summary]

See draft schedule below. (post talks here)

  • [[File: ]]

---

Part 1 (20 mins) - Sky Modeling and Component Separation (moderated by Raphael, Blake may help with discussion)

- Update on recent work in sky modeling / component separation (~10 mins, Raphael)

- Discussion: open questions and (in particular) what are the next steps? (~10 mins, Raphael)

File:Foregrounds.pdf


Part 2 (20 mins) - Lensing (moderated by Blake)

Short updates / talks about recent work with lots of discussion:

- auto / delensing foreground bias estimation (5 mins slides + 3 mins discussion, Alex) File:S4 lensing foregrounds alex aug2014.key.pdf

- small scale frequency cleaning / optimization (2+2 mins, Colin H.) [PDF]

- simulating delensing pipelines / delensing templates (brief update 2+2 mins, Kyle) [PDF]

- Discussion: what should we work on next? (4 mins on future work, Blake)

---



Some suggested discussion topics:

- All subjects: what are next steps for the group in i) sky modeling ii) component separation iii) lensing bias estimation iv) high-ell foreground cleaning and optimization v) delensing simulation


- FG modeling / cleaning: would realistic levels of decorrelation (or other effects) substantially change the optimization? What is enough complexity in foreground modeling? Can we get more data?

- Lensing, biases: what is the current take-home, how can we converge?

- Lensing cleaning, optimization: is current cleaning too conservative? How can we decide on frequency requirements?

- Lensing, delensing sims: issues and challenges, quadratic and max like?


- All topics (discussion at end): For next steps, what are our highest priorities?

Notes from session

Note-taker: Jason Henning

Notes for Foreground section:

  • Foregrounds dominate everywhere by orders of magnitude
  • Model YY=00
    • (not physical) model to validate Fisher forecasts in Science book with dust and sync with levels set by BICEP patch.
    • Multi-realization.
    • Gaussian realizations.
  • Model YY=01, 02, 03
    • contain instrument noise
    • limited resolution
    • assumed spectral dependence
    • (Different flavors of PySM)
      • differ by freq dependence of dust and sync, and AME
  • All models lead to small foreground residuals for S4 strawperson design
    • ~ order of magnitude under signal.
      • Partly due to high correlation between freq.
    • ILC prescription for foreground removal.
      • BICEP hasn’t achieved same residuals as Raphael.
      • Raphael’s model just requires maps to be properly calibrated to T_cmb.
  • Model YY=04
    • Ghosh, et al 1611.02418
    • based on GASS HI data
    • designed for larger decorrelation between freq.
  • Model YY=05
    • toy model with even more decorrelation
  • Models Y=04,05 lea to increase in sigma(r) of ~ 2 and bias for Raphael
      • Bias is ~ size of signal we’re trying to reconstruct.
    • Discussion about how much decorrelation we need to account for and how to model it.
      • Victor’s framework still wants large freq separation.
  • Model YY=06
    • Based on MHD
    • assumes constant dust-to-gas ratio
    • assume energy spectrum of electrons for synchrotron.
    • Reproduces E/B ratios.
    • Limited to small patches.
      • CP and RF disagree about how much this model would over-estimate correlation between dust and synchrotron at high Galactic latitude.
    • Resolution not good enough for use with lensing as well (sims don’t resolve turbulence).
    • Two sources of decorrelation
      • spatial variation of spectral index allowed by Planck.
      • line-of-sight effects
      • decorrelation between freq in dust is small.
    • What is the expected level of decorrelation???
    • We need sims to understand how we address decorrelation!!
    • We need models that address degree scales and needs of delensing survey simultaneously.
    • No non-forced sims have been able to show large decorrelation.

Next-steps

  • Add decorrelation into sims, more data, integrate high/low-ell sims.
  • Test closely-spaced freqs

Alex

  • NG-Foregrounds
    • Models for small-scale pol dust
      • Each gives different amount of BB power (while TT is tight).
    • Dust causes NG in lensing map.
      • So far no bias to lensing auto-spectra for the best 5% of sky for Vansyngel+ sims
      • 40% of sky there’s 1% bias in TT lensing, but still nothing in polarization. Comforting, but only as good as the sims.
    • Planck FFP8 sims (Challinor, Allison++2017)
      • Big bias (17% in TT, 62% in EBxEB)
      • Can downweight dusty areas and reduce bias in A_L from EBxEB (TT bias still there).
      • How does NG of dust impact delensing?
      • No average bias, but a lot of patch-patch scatter.
  • Takeways:
    • TT is biased from dust.
    • Polarization has bias that can be removed with downweighting (assuming uniform power over patches).
    • No evidence that we do not need dust and synch channels for high-ell survey.
    • Can we extend this to multi-freq?
    • How much do we trust freq decoherence in sims?


Colin Hill (Component Separation)

  • Simulated T maps
    • Missing small-scale Galac synch and AME.
    • All components properly correlated.
  • Simulated P maos
  • Constrained ICL
    • specifically project out whatever components you want.
    • Variance in final map larger since dof used.
  • If foregrounds are lower than noise floor ILC won’t pick them out but they can still bias lensing.
  • CP: Why aren’t we using these sims for low-ell forecasts as well?

Kyle Story (Lensed B-mode template)

  • “B-truth”: the best we could ever do
    • Leads to 99.9% delensing.
  • Quadratic estimator achieves ~ 90% delensing.
  • Next steps
    • Run these templates through r-estimation.
    • Add non-idealities to these maps?
    • Incorporate multi-freq, cleaned maps into pipeline.
    • LK: Investigating map-based delensing algorithms beyond QE is high-priority.
    • Move onto ML estimators.
    • Marius Millea
      • Starting to look at beyond QE delensing (1708.06753)

Action items/Next steps

Summarize action items here