UCSD-2019: Analysis/Pipeline Working Group: Low-ell BB

From CMB-S4 wiki
Revision as of 17:30, 18 October 2019 by Cbischoff (talk | contribs) (→‎Notes)
Jump to navigationJump to search

Link back to agenda


  1. Identify key decisions that must be made (and justified) prior to CD-1,
  2. Make progress on (or actually make) those decisions,
  3. Lay out a timeline and process for making each decision, consistent with the post-decision work and internal reviews that will be needed to complete preparations for CD-1,
  4. Ensure that those timelines and processes are understood and supported by the collaboration, and that we (together) believe we can follow them.

Key Issues

  • How do we optimize detector allocation across frequency in the face of uncertainty about foreground properties, for both the SATs and the delensing LAT?
  • What are the advantages of split bands? How does performance compare to non-split bands?
  • Do we need a 20 GHz channel on delensing LAT (for SAT science)
  • How well will de-lensing work in practice?
  • What are the necessary analysis tools to answer these questions?

What is CD-1?

Blatantly copied from Analysis/Pipeline Working Group: Maps to C_ell:

Background/clarifying questions:

  • What does “by CD-1” mean, and what are the implications for when tools need to be in place and working?
    • According to APC white paper (https://arxiv.org/abs/1908.01062), CD-1 is in Q3 of FY2021 (so June 2021?).
      • But according to project office, "Plan [must be] finalized by start of 2020 for delivering...CD-1"
    • Working backward from there, any tool that could reasonably influence a CD-1 decision needs to be in place and working by ... ?
    • Give an example timeline for an example decision?


  1. Introduction by everyone in the room: who? where? what aspects of low-ell BB interest you? 5 minutes
  2. Recap of the plan for this session/CD-1 goals slides[Wu], 3 minutes
  3. Review of Fisher based S4 forecasting thus far leading to DSR appendix A [Ben/Victor/Raphael], 15 minutes
  4. Review of Map based S4 forecasting thus far and ideas for next steps Clem, 10 minutes
  5. Review of forecasting and simulations for Simons Observatory inflation science [Alonso / Errard / Sherwin], slides delensing slides 30 minutes
  6. Plans for working group [all], ~60 minutes
    1. Forecasting
      • Quantitative comparison of SO and CMB-S4 DSR forecasts
      • Optimization of delensing effort: update for DSR sensitivity, motivation for frequency coverage, feedback from real delensing efforts
    2. Data Challenge simulations
      • More / better foreground models
      • Instrumental systematics
      • Other map non-idealities, i.e. filtering / mode loss
      • Coordination with technical groups on instrument configuration, etc
      • What else needs to be demonstrated for CD1?
      • Bottlenecks for sim production
    3. Analysis of Data Challenges
      • Who plans to participate? How do we get more participation?
      • Delensing
    4. Specific Data Challenge plan / timeline
      • Experiment config 06: DSR configuration, in progress
      • Update / reoptimization of delensing survey?
      • Inclusion of instrumental systematics -- which ones? how to include?
      • Data Challenges coming out of Data Management group

Remote attendance

Join Zoom Meeting https://zoom.us/j/340796462?pwd=S1o3Q3NQUlFOWkVVUXVrV1laK0JrQT09

  • Meeting ID: 340 796 462
  • Password: 654876
  • One tap mobile
  • +14086380968,,340796462# US (San Jose)
  • +16465588656,,340796462# US (New York)
  • Dial by your location
  •         +1 408 638 0968 US (San Jose)
  •         +1 646 558 8656 US (New York)
  • Meeting ID: 340 796 462
  • Find your local number: https://zoom.us/u/aeGN6VEA3T


Buza / Racine: CMB-S4 PGW Forecasting

  • Forecasting loop: achieved performance used to forecast CMB-S4 instrument -> generate map-based sims with sky model -> analyze to determine sigma(r) and compare to forecast
  • 9 bands spanning 4 atmospheric windows. Using split-bands to guard against unknown foreground complications. (20 GHz channel is on LAT to avoid huge beams)
  • Calculate detector NETs for South Pole and Chile. NETs are only used to rescale BICEP/Keck achieved performance -- no ab initio sensitivity calculation.
  • Forecast assumes sky model with dust and synchrotron. Foregrounds are allowed to decorrelate -- Fisher calculation assumes 3% dust decorrelation between 217 and 353 GHz.
  • Choose distribution of effort across nine frequencies plus delensing (10 channels) to minimize sigma(r).
  • Data challenge maps use forecasted noise levels but consider many different foreground models. Validates forecast and also determines bias due to different foreground models.
  • Analyze data challenge maps with two pipelines: ILC and parametric likelihood foreground cleaning.
  • For DSR, converted optimized distribution into discrete instrument configurations in consultation with SAT and detectors groups. Settled on "configuration 5" -> reference design.
  • Using detailed sky coverage maps for Chile (deep and shallow) and Pole (deep or wide). Still rescaling from BICEP/Keck noise, but accounting for change in depth and bandpower degrees of freedom. Also applied a foreground mask, which eliminates some of the Chile coverage. Optionally apply foreground penalty -- assume that we can clean down to 1% of foreground, residual acts as bias.
    • End up with variation in sigma(r) as a function of number of SATs in Chile vs Pole. Results are different for r=0 vs r=0.003.
  • Lloyd: We don't currently have optimal amount of delensing throughput. Want to know what delensing throughput is needed to hit measurement requirements. See Marius' figure from DSR but need to know what A_L level is threshold.

Sherwin: SO BB pipeline / delensing

  • SO is very different limit from S4 -- more noise and more sky, so delensing is less important. SO sigma(r) increases by factor of 2 without delensing.
  • Using linearized delensing B-mode template constructed on curved sky. Using Wiener filter to downweight noisy regions, but this isn't much better than just masking. Lensing template treated as virtual frequency band.
  • SO in intermediate regime where delensing is needed, but noise isn't quite good enough for internal delensing. Also using LSS (CIB, galaxies) as lensing tracer. Multi-tracer delensing gives significant improvement over CIB or CMB internal alone. Think that they will get to ~70% delensing once LSST data is available.
  • Running this in a pipeline. Sims include lensing and LSS.
  • Think that LSS calibration will be ok because can do it via cross-spectra with noisy CMB lensing map. Worried about bias due to Galactic dust in CIB maps, but ran some sims and it looks ok.
  • Is multi-tracer delensing useful as cross-check for CMB-S4? Very complicated, but could be helpful for wide survey. Important to marginalize over A_lens because lensing residual might be uncertain.

Errard: SO BB + delensing forecasts

  • Three different analysis pipelines: power-spectrum based, parametric map-based, and blind ILC. Now working to validate full pipelines on realistic sims.
  • Large sky fraction for low-ell survey: 10-20%. Hit regions with higher foregrounds and have to worry about spatial variations of foregrounds.
  • Power-spectrum based pipeline: use auto and cross-spectra, model foregrounds, marginalize with Gaussian likelihood. Hard to account for spatially-varying foregrounds.
  • Map-based cleaning pipeline: xForecast. Approximate full map likelihood by averaging over CMB and noise fluctuations. Validated against full map-based likelihood (BFoRe) and saw good agreement.
  • Analysis of simulated maps produces similar results for all pipelines. Tried some more complicated foreground models.
  • Delensing: for power-spectrum-based pipeline, delensing template is a virtual frequency channel. For map-based pipeline, subtract delensing map from foreground-cleaned CMB map.
  • We should run these pipelines on CMB-S4 sims and vice versa. SO sims are at NERSC and publicly available.

Pryke: PGW map-based sims

  • Most recent set is experiment definition 04 with ten different foreground models. Not up-to-date with DSR experiment design.
  • Hope that foreground models span reality, but no guarantee of this.
    • Three PySM models
    • Tuhin Ghosh model
    • Highly-decorrelated dust model, designed to break things (and it does!)
    • Flauger / Hensley MHD-based model
    • Amplitude-modulated Gaussian, based on Planck
    • MKD multilayer model (Delabrouille): SED deviates from greybody even in small pixels -- produces strong bias in analysis (but also fails goodness-of-fit)
    • Vansyngel model: extends non-Gaussian structure to small scales
  • Map-based sims validate forecast but also allow measurement of bias due to foreground model.
  • Recipe to scale noise from achieved performance to map-based sims:
    1. Use ratios of ideal NETs to scale from achieved performance to S4 forecast
    2. Generate full sky realizations of noise (with tweak needed to recover N_ell)
    3. Divide by sqrt of relative hits map. Raphael notes that this causes small bias in results because of mode coupling.
  • In addition to BK achieved performance, we have new POLARBEAR results with N_ell and hit pattern.
  • Going to make 06 sets of sims using DSR experiment configuration and hits maps.
    • Who takes this over when Clem goes to Pole? Caterina volunteers (maybe Victor or Colin too)
  • What about data challenge maps from Data Management group?
    • We should be ready to analyze those, but keep going on our own sim program.
    • Need to talk to Data Management people to make sure that they use good inputs for the sims, especially noise and sky model.