1 / 45

Learning from spectropolarimetric observations

Learning from spectropolarimetric observations. A. Asensio Ramos Instituto de Astrofísica de Canarias. aasensio.github.io/blog. @ aasensior. github.com/ aasensio. Learning from observations is an ill-posed problem. Follow these four steps. Understand your problem

hiero
Download Presentation

Learning from spectropolarimetric observations

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Learningfrom spectropolarimetricobservations A. Asensio Ramos Instituto de Astrofísica de Canarias aasensio.github.io/blog @aasensior github.com/aasensio

  2. Learningfromobservationsisanill-posedproblem

  3. Followthesefoursteps • Understandyourproblem • Understandthemodelthat ‘generates’ your data • Define a meritfunction • Compute the‘best’ fitbyoptimizingorsamplethismeritfunction Thesolutiontoanymodelfitting has to be probabilistic

  4. Understandyourproblem • Your data has beenobtainedwithaninstrument • Yoursyntheticmodelmightnotexplainwhatyousee • You are surelynotunderstandingyourerrors • Systematics • …

  5. Understandyourgenerativemodel Thisisthemostimportantand complexpart of theinference Example Generativemodel Assumptions • Weassumethat xi are fixed and givenwithzerouncertainty • Uncertainty in themeasurementisGaussianwithzero meanand diagonal covariance

  6. Fromthegenerativemodeltothemeritfunction Likelihood Probabilitythatthemeasured data has been generatedfromthemodel

  7. Why do we do thec2fitting? Thestandardleast-squaresfitting comes fromthe maximization of a Gaussianlikelihood

  8. Somesubtleties • Weights • Do notchangethe position of themaximum • Modifythecurvature at themaximum • Ifnoisestatisticschange, modifythelikelihood

  9. Be aware of theassumptions • Errors are Gaussian • Youknowtheerrors itisdifficulttoestimateuncertaintiesin theerrorsbecauseerrors are already a 2ndorderstatistics • Errors are onlyonthe y axis  x locations are givenwithinfiniteprecision • Themodelincludesthetruth

  10. Whatifwe break theassumptions? Any of ourassumptionsmight be broken • Errors are notGaussian • Wedon’tknowtheerrors • Errors are alsoonthex axis • Themodeldoesnotincludethetruth

  11. Withoutoutliers

  12. Withoutliers Wegetbiasedresults

  13. Modeleverything Ifyoumodelthe data points and theoutliers, youautomatically have a generativemodel and a meritfunctiontooptimize pointsfromthe line badpoint

  14. Fitting He I 10830 Å profiles

  15. Hazel github.com/aasensio/hazel MIT license

  16. Assumptions+ properties • Multi-term atom • Simplified but realistic radiativetransfer effects • One or two components (along LOS or inside pixel) • Magneto-optical effects • MIT license • MPI using master-slavescheme • Scalesalmostlinearlywith N-1 (testedwithup to 500 CPUs) • Pythonwrapperforsynthesis

  17. 3d3D 3p3P 3s3S 2p3P 10830 Å 2s3S

  18. Forward modelling

  19. Problemswithinversion • Robustness • Sensitivitytoparameters • Ambiguities

  20. Robustness: 2-step inversion Global convergence DIRECT Refinement Levenberg-Marquardt Step 1 Step 2 Step 3 DIRECT algorithm (Jones et al 93)

  21. Sensitivitytoparameters: cycles Modifyweights and do cycles Cycle 1 Invertthermodynamicalproperties t, Dvth, vDopp, … Stokes I Cycle 2 Invertmagneticfield vector Stokes Q, U, V

  22. Ambiguities

  23. Ambiguities: off-limbapproach • Do a firstinversionwithHazel • Saturationregime findtheambiguoussolutions (<8) In thesaturationregime(above~40 G for He I 10830)

  24. Ambiguities: off-limbapproach • Do a firstinversionwithHazel • Saturationregime findtheambiguoussolutions (<8) • Foreachsolution, use Hazelto refine theinversion • NowalmostautomaticallywithHazel

  25. Wheretogofromhere? • Do full Bayesianinversion • Modelcomparison • Inversionswithconstraints Modeleverything, includingsystematics, and integrateoutnuisanceparameters

  26. Bayesianinference PyHazel+PyMultinest

  27. Modelcomparison H0 : simple Gaussian H1 : twoGaussians of equalwidthbutunknownamplitude ratio

  28. Modelcomparison H0 : simple Gaussian H1 : twoGaussians of equalwidthbutunknownamplitude ratio

  29. Modelcomparison

  30. Modelcomparison ln R=2.22  weak-moderateevidence in favor of model 1

  31. Constraints

  32. Central stars of planetarynebulae

  33. Bayesianhierarchicalmodel Model FV B1,μ1 Model FV b0 B2,μ2 Model FV B3,μ3

  34. Bayesianhierarchicalmodel

  35. Are solar tornadoes and barbsthesame? Coreof the He I line at 1083.0 nm (~0.8’’) • Full Stokes He I line at 1083.0 nm (VTT+TIP II) • Imaging at thecore of the Hα line (VTT - diffractionlimited MOMFBD) • Imaging at thecore of the Ca II K (VTT - diffractionlimited MOMFBD) • Imagingfrom SDO

  36. Coincidencewithtornadoes in AIA

  37. ``Vertical’’ solutions Field inclination

  38. ``Horizontal’’ solutions Field inclination

  39. Magneticfieldisrobust • Fields are statisticallybelow 20 G • Someregionsreach 50-60 G • Filamentary vertical structures in magneticfieldstrength

  40. Conclusions • Be aware of yourassumptions • Modeleverythingifpossible • Hazelisfreelyavailable • Ambiguities can be problematic • More worktoputchromosphericinversionsat thelevel of photosphericinversions

  41. Announcement IAC Winter SchoolonBayesianAstrophysics La Laguna, November 3-14, 2014

More Related