Photometric Redshifts

download Photometric Redshifts

of 24

  • date post

    03-Jan-2016
  • Category

    Documents

  • view

    25
  • download

    0

Embed Size (px)

description

Photometric Redshifts. PHAT Meeting Pasadena 3-5 Dec 2008. Christian Wolf. data. model. Farb- bibliothek. estimator. Schätzer/ Klassifikator. Frequentist precision statistics: = “Using what IS there: N(z)!”. result. Bayesian frontier exploration: = “What do we (not) know: p(z)=?”. - PowerPoint PPT Presentation

Transcript of Photometric Redshifts

  • Photometric RedshiftsChristian WolfPHAT Meeting Pasadena 3-5 Dec 2008

  • Photo-z Ingredients & Applicationspectral energy distributionPDF: p(z)empirical dataorexternal template2-fittingartificial neural net learning algorithms

  • Model + Estimator Combinations2PDF Ambiguity warningNNNo PDF, no warningTemplate modelCan be extrapolated in z,magCalibration issuesPriors issuesEmpirical modelGood priorsNo calibration issuesCan not be extrapolated Code2NN

    Model

    TemplateEmpirical

  • Class Decision and z-Estimation zCan include morphology etc. in statistics or neural net

  • Examples from COMBO-17 Classification ~98% complete at R
  • Redshift Error RegimesThree regimes in photo-z quality

    SaturationModel-data calibration offsets; intrinsic dimensionality of classTransitionLocally linear colour(z) grid BreakdownGlobally nonlinear colour(z) gridmag

  • Galaxies: Saturation & TransitionR=20R=22R=23.7Galaxies at z~0.45

  • QSOs: Saturation at R
  • Photo-Z TroubleCatastrophic failures & misclassifications

    Large z errors

    Local z bias

    Unrealistic z errorsModel ambiguities in colour space (spotted?)

    PDF too unconstrained

    PDF wrong (calib, prior)

    Mismatch between data and model

  • Common FixesCatastrophic failures & misclassifications

    Large z errors

    Local z bias

    Unrealistic z errorsModel ambiguities in colour space

    PDF too unconstrained

    PDF wrong (calib, prior)

    Mismatch between data and model

  • Add More Data: Wider RangeWider coverage Covers spectral features across wider z rangeAdd NIR dataFor z > 1 galaxiesOnly weak & high-variance features in rest-frame UVRed z>1 galaxies with noisy optical data Add UV data (e.g. GALEX)For QSOs (Ball et al. 2007)Lyman break at z < 2-3Abdalla et al. 2007

  • Add More Data: Narrow FiltersImprove localization and contrast of featuresQSO line detection avoids catastrophic failures at z < 3Galaxies+QSOs: improve zGalaxy, star, QSO, WD, ?Wolf 2001

  • Add More Data: Narrow FiltersImprove localization and contrast of featuresQSO line detection avoids catastrophic failures at z < 3Galaxies+QSOs: improve zGalaxy, star, QSO, WD, ?Wolf 2001

  • Add PriorsImpact: ptot = pprior pcolour Reduce rate of bimodal PDFsReduce larger z (up to 2) Explicit for template modelsLuminosity function, rangeMag / z extrapolation ~ok Implicit for empirical modelsRestricted in mag & zMag extrapolation wrongZ extrapolation impossibleRepresentative sample?

  • Repair (or Make) TemplatesBudavari et al. 2000,2001

  • Repair Models: Uncommon ObjectsCOMBO-17 field Abell 901/2Super-cluster at zspec~ 0.16800 members with R < 21 Photo-z 2002: using SEDs by Kinney et al. (1996)25% of S/N~100 members outliers with zphot~ 0.06 z ~ 0.1 SHOCK!Red spirals! Photo-z 2003: include dust-reddened old SEDs ~1% outliers

  • Photometry & SEDsPhotometry:PSF-matchedCalibration (obs. frame)Artefacts: instrumental, data reductionError distribution, non-Gaussian systematics in Gaussian error floor

    Source variabilityStars: RR Lyr, long-termGalaxies: supernovaePhotometric BlendsTransient blends by moving objectsClose neighboursLine-of-sight projections, strong lensesBinaries, WD+M etc.

    SED compositionAGN componentComposite stellar populations

  • PSF-Matched PhotometryBasic method:Assume Gaussian PSFConvolve to worst PSFPhotometry in aperture AProblems:Local PSF variationsNon-Gaussian PSF (Capak et al. 2007)Special case: Gaussian aperture & PSFStronger weight to brighter object centreAeff = PSF A (space-based aperture)If A and PSF Gaussian, then Aeff Gaussian as wellMinimize computations: Fix Aeff & adjust A to PSF2A = 2eff - 2PSF

    (Rser & Meisenheimer 1991) Non-gaussian aperture & PSFShapelet-based method (Kuijken 2008)

  • What We Understand by NowOrigin of local z biasObserved-frame + rest-frame calibrationNon-flat priors: p(zph|zsp) vs. p(zsp|zph), z>0 Origin of catastrophic outliersUnrecognised z ambiguity in colour spaceWrong data / errors: blends, instrumental issues

    Minimum z variance levelsIntrinsic SED variationsSpectral resolution

  • Local Z Bias: CalibrationN = number of filters, i.e. independent data pointsCalibration offsets in N=3 D1-D normalisation1-D z-bias1-D restframe SED bias1 out of N offset dimensions causes a photo-z bias z More filters smaller z (proj. component ~1/N)Narrow filters small z (larger col/z on feature) Spectroscopy with N~102..3: z without flux calibrationFew-filter photo-zs limited by calibration accuracyMany-filter photo-zs limited by number and resolution of filters

  • Catastrophic OutliersResult from undetected ambiguities: Also: wrong data/errors Example: see shrinking training sample 20% sample in 1:20 ambiguities causes overall 1% unflagged outliers

  • Intrinsic Variety: Z Error Support Example: QSO near g-r~1 or z~3.7Main signal: Ly forest in g, but SEDs >0-D family Training sample in boxRedshift distribution: mean 3.66, rms 0.115RMS/(1+z) = 0.024 Testing sample in boxRMS/(1+z) error 0.023

  • What To Work On: DataDefine most effective & efficient data sets:From simulations (which dont rule out outliers)Describe data correctly:Consistent apertures across bandsTrue photometric scatter by objectMinimise unrecognised error sources in data:Error floor from photometric blends & transients

  • What To Work On: ModelsTemplates etc.:Best templates, rare objects with different SEDsBest priors, best extrapolation in (mag, z)Training samples:Discretization effects, confidence limits on random-nessPropagation into n(z) errors and outlier risksSize matters: What do I need?Combine all approaches? Empirical + extrapolated template model for all kinds of use