UMICH-2015: Simulations & Data Analysis Break-Out Session 2
- 1 Forecasting for the S4 Science Book
- 1.1 What S&DA do we need to do to support the science sections of the CMB-S4 Science Book?
- 1.2 Some additional "Forecasting for S4 Science Book" Framing Questions
- 1.3 Forecasting using simulations: complexity vs. feasibility?
- 1.4 Sky model for simulations and forecasting
- 1.5 Previous Forecasting plots from S4 Snowmass / Examples of current tools
- 1.6 Who will sign up for what this year?
- 1.7 Notes taken during session (by Tom Crawford)
- 2 Wiki navigation
Forecasting for the S4 Science Book
What S&DA do we need to do to support the science sections of the CMB-S4 Science Book?
What are the right inputs to the simulations?
- Sky model
- scalar, tensor, non-Gaussian, (cosmic strings?)
- Mission model
- beams, band-passes, noise PSDs
- detector electronics, data acquisition
- scanning strategy, flags
- atmosphere, ground pickup
What are the required output(s) of the simulations?
What are the required analysis outputs/metrics/figures of merit?
- Power spectra
What is the right balance between the veracity of the models used in the simulations and the tractability of their production and analysis?
- Spectral domain
- Map domain
- Time domain
What resources are available over the coming year?
- Cycles & storage
Some additional "Forecasting for S4 Science Book" Framing Questions
- How can we make these forecasts truly realistic?
- grounded in experience / achieved performance
- conservative regarding systematics
- conservative regarding foreground complexity
- What drives need for simplicity, transparency, flexibility?
- Forecasting <--> Survey design feedback loop needs to be fast as we explore tradeoffs for different science drivers
- Need flexibility to separately explore large scale / degree scale / arcmin scale survey parameters
- Input assumptions easily changeable to determine dependencies
- Multiple parallel forecasting efforts should coordinate inputs/outputs to
- understand where differences are due to input assumptions vs. methods.
- What are the roles in this forecasting stage for
- timestream level sims (if any)?
- map level sims?
- bandpower level likelihoods / fisher analyses?
- Role of Data Challenges
- not most efficient tool for exploring tradeoffs
- at specific points in survey definition space, can validate Fisher forecasts
- common input sims, what is needed:
- foreground maps
- CMB maps
- noise / systematics maps
- encoding experiment filters: reobs matrices?
- separate challengers / challengees?
- How to parameterize S4 survey specifications
- What assumptions for external datasets? (e.g. BAO from...)
Forecasting using simulations: complexity vs. feasibility?
Large scale simulations with full complexity require long investment and significant resources. How can we go beyond very basic forecasting that neglects foregrounds and systematics and defines an instrument with only a beam size, a sky coverage, and a sensitivity, without implementing the full sims? How many levels of complexity should we consider, e.g.
- basic: sensitivity with Gaussian white stationary noise; gaussian beams; sky fraction
- moderate: inhomogeneous noise with non-stationarity and low-frequency excess
- advanced: model of atmospheric noise based on map differences from existing data on patches or on model; scanning and filtering + map-making for detector sets, scaling-up as a function of number of detectors; simple foreground subtraction with, e.g. decorrelation, NILC, or simple model fitting per pixel
- full: as realistic as possible
A few additional ideas:
- 1/ell noise: usually experiments do not perform as well on large scales as extrapolated from noise rms. Can we work-out an empirical law, possibly based on existing experiments?
- foregrounds: separate polarization issues (for primordial and lensing B-modes mostly) from temperature issues (on small scales, for SZ and extragalactic astrophysics). Should we worry about EE and TE foregrounds for parameter estimations?
- How can we get a "trustable" model of the level of foreground residuals (to treat them as a noise in forecasting)?
Sky model for simulations and forecasting
A significant unknown in any forecasting with S4 is the complexity of foreground emission (the level is approximately known from current models). To go a step beyond first order estimates, we should make simulations of the sky emission with representative foreground complexity, including (plausible / possible) surprises.
An effort is ongoing in the Planck collaboration to produce a final Planck Sky Model (PSM). This tool builds on 10 years of development of modelling sky emission in a parametric way, based on a large collection of data sets (maps observed at various frequencies, catalogues of objects, number counts, etc.).
A description of a early version of the Planck Sky model can be found in Delabrouille, Betoule, Melin, et al. (2013), "The pre-launch Planck Sky Model: a model of sky emission at submillimetre to centimetre wavelengths", A&A; 553, 96 (http://adsabs.harvard.edu/abs/2013A%26A...553A..96D). See also the PSM web page: http://www.apc.univ-paris7.fr/~delabrou/PSM/psm.html
The PSM has been used to generate simulations for Planck described in Planck 2015 results. XII. Full Focal Plane simulations (http://adsabs.harvard.edu/abs/2015arXiv150906348P).
Here is a short overview of the components in the model, their main limitations, and plans to fix those limitations. Suggestions and prioritization would be useful.
- Galactic emission: Low frequency model is still based on WMAP + Haslam 408 MHz maps. Dust based on Planck for the current version used to make Planck FFP8 maps (not public yet). Low frequency foregrounds should be updated on the basis of recent Planck analysis (http://adsabs.harvard.edu/abs/2015arXiv150606660P). All maps are limited by the limited angular resolution of the input observations.
- Synchrotron: based on WMAP + Haslam 408 MHz maps, polarization from a large scale model of the galactic magnetic field and of depolarization. Complemented by random fluctuations on small scales.
- Free-free: Should be de-noised at high galactic latitude. Assumed unpolarized for the moment.
- CO: based on ground based observations (Dame et al.), only part-sky. Assumed unpolarized for the moment.
- Spinning dust: Should be de-noised, frequency dependence of the emission should be pixel-dependent. Assumed unpolarized for the moment.
- Dust: Temperature and Polarization based on Planck HFI (mostly 353 GHz). Several options for scaling with frequency exist.
- Other lines than CO are missing
- Clusters: based on number counts + random distribution on the sky, with relativistic corrections, scaling laws to connect mass to Y. Thermal SZ, kinetic SZ (with random directions), polarized SZ effects are implemented. Missing: contamination by radio and IR sources, correlation with CIB and with lensing.
- CIB (unresolved background of (mostly high redshift) dusty galaxies
- Radio sources
- Local IR galaxies
- How do we go beyond very simple models (e.g. one single power law synchrotron per pixel, one single dust greybody, ...). What is the evidence that we should go beyond such simple models?
- Should we make simulations with (nasty) surprises, e.g. 0.1 to 1% polarization for spinning dust, CO, (free-free?)
- Should we do blind analyses of simulations at this stage ?
- What complementary data do we envisage (should we work on CMB-S4 only, or CMB-S4 + LiteBIRD, COrE+, PIXIE, X-rays, other?)
- should we try to put in the simulated sky signatures of all the effects we will look for in real S4 data (prioritize, set requirements on accuracy of modeling)
- alternatives to the PSM?
Previous Forecasting plots from S4 Snowmass / Examples of current tools
See also http://lanl.arxiv.org/pdf/1402.4108.pdf Wu, Errard, Dvorkin, Lee, McDonald, Slosar, Zahn.
S4 Snowmass Inflation white paper (1309.5381) forecast plot:
How do we want the S4 science book forecast to do better?
- Foreground separation approach (e.g. multifreq/multicomponent tools that can be used, while maintaining simplicity/transparency)
- Delensing treatment (e.g. allow different survey depth assumptions for arcmin vs deg scales)
- Systematics treatment (e.g. include assumption of systematic uncertainty scaling with noise?)
- Alternative B-mode FOM's? (min detectable r, n_t, etc.)
Current "r" forecasting code examples:
- BK / BKP - derived forecasting, Victor Buza File:V buza fisher20150921.pdf
- please volunteer other examples (or effort!)
S4 Snowmass Neutrinos white paper (1309.5383) forecast plot:
How do we want the S4 science book forecast to do better?
Who will sign up for what this year?
Action Item is to set up working groups:
- for FG and cosmology input maps
- science metric forecasters
- coordination of inputs and data challenges
Notes taken during session (by Tom Crawford)
What forecasting is needed for Science Book? - what is deadline? June 2016? - what work needs to happen so we can discuss & iterate at January meeting in Berkeley?
What are right inputs to simulations? - Dunkley: need to have lensed input. - do we absolutely need self-consistent lensed CMB + matter + foregrounds - Crawford: worried about foreground lensing bias if not.
How do we define figures of merit for forecasting? - Kovac: won't get it done here, but we need to assign people to working groups to do this.
What resources are available (people, cpu, codes) over the next year? - come back to this at end of session
How to make forecasts truly realistic? - sensitivity: ground in achieved performance: how? - Dunkley: use realistic N(l) (noise power spectrum) - Kovac: and realistic # of modes (filtering) - Pryke: and # of detectors and observing efficiency - consensus: real N(l) and achieved annual survey depth/efficiency - systematics: include explicitly or use some systematics penalty - Keskitalo: differential pointing & bandpasses very important and should be included at TOD level - Crawford: if include specifically, how? - Borrill, Keskitalo think we should do TOD sims this round (before January); Dunkley, Kovac, Pryke, Crawford think there is not enough time and not well-defined enough - Lawrence: chicken-and-egg problem that you can't characterize systematics before instrument is built, but need systematics in forecasts that define instrument. Needs to be iterative/ongoing. - foregrounds: we all agree we need to be conservative about complexity of foregrounds - how to include? Dunkley: bad old way is as extra noise in Fisher formalism - Flauger: better way is to include foreground parameters in Fisher
Role of data challenges - separate challengers/challengees? - not for first round, according to Pryke & Delabrouille
How to include experiment filters (Kovac: possibility is reobservation matrices)? - Borrill: TOD sims! - Dunkley: at very least, cut off at some l_min.
Multiple forecasts for key science goals? - Kovac: yes, and need to coordinate inputs & outputs to understand differences
How to parameterize S4 survey specifications? - at least 4 dimensions: map depth, f_sky (possibly as a function of depth), observing frequencies, angular scale coverage
Sky model - how to improve PSM at high latitudes where S/N on foregrounds is small? - Dunkley: would like to see results for worst-case polarized spinning dust - Flauger: Colin Hill paper useful for getting polarization angles - PSM also has clusters, sources, etc. - not currently correlated with lensing but that is in the works - Battaglia: why not use a real sim or one of the fast mocks for this? - can generate halo catalog with painted-on SZ, CIB, etc. and associated kappa map in minutes - Crawford: sounds like a possible plan is using a fast mock for extragalactic stuff & PSM for Galactic foregrounds - How can we / should we go beyond simple models? Should we include "nasty surprises"? - Dunkley: Bruce Draine worried/excited about "magnetic dust" - what other data sets to include / produce simulations of? - Pryke: can we steal highly developed code from other projects? - Delabrouille: lots of this is being developed for CORE+ - Dvorkin has forecasting code available - Kovac: what should we assume for a tau constraint? - Pryke: the PSM is a single realization of most stuff with a random component---do we make multiple realizations?
How to make S4 science book better than Snowmass? - Dunkley: would be great to try de-lensing on a realistic map - Dvorkin: treated foregrounds by assuming PSM and a cleaning level - what was assumed about tau? (probably Planck Blue Book) - example of improved Fisher code (realistic N(l), reproduces achieved Bicep/Keck constraints) shown by Kovac/Buza - Bischoff: great but needs better treatment of lensing
Fuller: is analysis code set up to take advantage of higher precision of data? - for instance, how to take advantage of nucleosynthesis data? - Dunkley: Julien Lesgourges is one of the people working on this
Who will sign up for what? What working groups do we need to define/assign people to?