hierarchical probabilistic inference of cosmic sheardlaw/wfirst2020s/slides/schneider.pdf ·...
TRANSCRIPT
-
LLNL-PRES-733055This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under contract DE-AC52-07NA27344. Lawrence Livermore National Security, LLC
HierarchicalProbabilisticInferenceofCosmicShearAstronomyinthe2020s:SynergieswithWFIRST
MichaelD.SchneiderwithJoshMeyersandWillDawson
June 27, 2017
Collaborators:D. Bard, D. Hogg, D. Lang, P. Marshall, K. Ng
-
LLNL-PRES-7330552
§ WFIRST&LSSTareideallysuitedforajointcosmicshearmeasurementtoconstraincosmologicalparametersanddarkenergy
§ NewshearinferencemethodsarerequiredtofullyexploitthesensitivitiesofWFIRST&LSST
§ Ahierarchicalprobabilisticforwardmodelapproachshowsthemostpromiseformeetingshearbiasrequirements
§ Theseprobabilisticalgorithmsenableboth:— Exploitationofinformationinnewcosmologicalstatistics— Moreflexiblecomputingpipelinestoingestandinterpretnewdata
Summary
-
LLNL-PRES-7330553
Introduction
-
LLNL-PRES-7330554
Cosmicshear measurement
§ Thelensingbylargescalestructure§ Lookingforverysmallsignalunderverylarge
amountofnoise§ Wedon’tknow“unsheared”shapes,butcan
(roughly)assumetheyareisotropicallydistributed
§ Cosmicsheardistortsstatisticalisotropy;galaxyellipticitiesbecomecorrelated
§ ExquisiteprobeofDE,ifsystematicscanbecontrolled
§ LSST:willmeasurefewbilliongalaxyellipticities.ExcellentsensitivitytobothDEandsystematics!
4
Cosmic shearsignal iscomparableto ellipticity of the Earth, ~0.3%
- D. Wittman
-
LLNL-PRES-7330555
WhatwilltheWFIRSTHLSaddtocosmicshear?Improvedtomography,reducedbiasmeanstighterdarkenergyconstraints
§ Deblending— 30– 50%ofLSSTgalaxydetectionswillbemultipleresolvedobjects
asseenbyWFIRST— Shearbias:Mostblendsarechancealignmentsofgalaxiesat
differentredshifts— Photo-zbias:mixedcolors/biasedphotometry— Assertnumber&propertiesofblendcomponentsgivenHLS
overlapwithLSST— TrainstatisticalcalibrationforentireLSSTfootprint
§ Improvedphoto-z’sfromLSSToptical+WFIRSTNIRbands
§ HigherfidelityLSScross-correlations(fromgrism survey)— Breakmanysystematicsandcosmologicalparameterdegeneracies
§ Reducedshearbias?
5
-
LLNL-PRES-7330556
Weaklensingofgalaxies:theforwardmodel
Unknown & dominates signal
Want this
Marginalize
Constrained by
Image credit: GREAT08, Bridle et al.
-
LLNL-PRES-7330557
Wewon’tjusthavemoredatawith‘StageIV’surveys.- We’reinanerawithqualitativelynewcomputingcapabilities
Dijkstra’s Law
A quantitative difference is also a qualitative difference if the quantitative difference is greater than an order of magnitude.
Figure credit: Wikipedia
> 2 orders of magnitude since SDSS era
-
LLNL-PRES-7330558
Qualitativechangesincomputingenablenewscientificmethods
- Mark Seager, CTO for the HPC Ecosystem at Intel(interview in Inside HPC on June 6, 2016)
“…predictive simulation has brought together theory and experiment in such a compelling way that it’s fundamentally extended the scientific method for the first time since Galileo Galilei invented the telescope in 1609…”
-
LLNL-PRES-7330559
§ We’refacingsystematics-limitedmeasurements— End-to-endsimulationsoftheexperimentarethebestapproachtoimproveaccuracy&
precision
— Tiesdataandsimulationmoreintricatelythaninpastcosmologypipelines
§ Imageandcatalogsummarystatisticsarenolongergoodenoughtomeetnextgenerationsciencerequirements— Probabilistichierarchicalmodelsandrelatedmachine-learningapproachesshowpromise
butaremuchmorecomputationallyintensive
— Potentialchangestothetraditional‘facility’/‘user’separateanalysisstages
Data+Computeconvergenceincosmology– DOEASCRinitiative,April2016
Removing the line between ‘analysis’ and ‘simulation’.
-
LLNL-PRES-73305510
Catalog cross-matchingbetweenspaceandgroundisconfusedbysignificantobjectblendingasseenbyLSST
Ground:(Subaru(Suprime0Cam(
Space:(Hubble(ACS(
LSST blend fractions estimated from Subaru & HST overlapping imaging
Dawson+2015
-
LLNL-PRES-73305511
Shear bias
-
LLNL-PRES-73305512
ShapetoShear:NoiseBias
§ Ellipticity:
§ Ensembleaverageellipticity isanunbiasedestimatorofshear.
§ However,maximumlikelihoodellipticity inamodelfitisnotunbiased.
§ Ellipticity isanon-linearfunctionofpixelvalues.
e =a� ba+ b
exp(2i✓)
-
LLNL-PRES-73305513
1. Calibrateusingsimulations.(im3shape,sfit)
— Butcorrectionsareupto50xlargerthanexpectedsensitivity!
2. Propagateentireellipticity distributionfunctionP(ellip |data)
— UseBayes’theorem:P(ellip |data)∝ P(data|ellip)P(ellip)
— MeasureP(ellip)indeepfields.(lensfit,ngmix,FDNT).
— Infersimultaneouslywithshearinahierarchicalmodel.(MBI).
MitigatingNoiseBias– atleast2strategies
-
LLNL-PRES-73305514
Ahierarchicalmodelforthegalaxydistribution
§ σe =intrinsicellipticity dispersion
§ eint =galaxyintrinsicellipticity
§ g=shear
§ esh =galaxyshearedellipticity
§ PSF=pointspreadfunction
§ D=modelimage
§ σn =pixelnoise
§ D=data:observedimage
arXiv:1411.2608
-
LLNL-PRES-73305515
Ourgraphicalmodeltellsushowtofactorthejointlikelihood
§ Useaprobabilisticgraphicalmodeltoencodethefactorizationofthejointprobabilitydistributionofvariablesinthemodel.
§ Wedon’tcareaboutesh forcosmology,sointegrateitout.
Huge complicated integral to compute for every posterior evaluation.
/Z Y
ij
P⇣D̂ij|PSFj, �n,j, eshi
⌘Y
i
P⇣eshi |g, �e
⌘P (g) P (�e) d{eshi }
Pr⇣g,�e| {PSF}j , {�n,j , {Dij}}
⌘
/Z
dngal�eshi
2
4Y
ij
Pr�Dij |PSFj ,�n,j , eshi
�3
5"Y
i
Pr�eshi |g,�e
�Pr(g)Pr(�e)
#
arXiv:1411.2608
-
LLNL-PRES-73305516
ImportanceSampling allowstractabledivide&computeWethusestimatethepseudo-marginallikelihood forshear
§ Don’tgobacktopixelsforeverytimewesampleanewgorσe.
§ Foreachgalaxy,drawimagemodelparametersamplesunderafixed“interim”prior.Thisisembarrassinglyparallelizable.
§ Usereweightedsamplestoapproximatetheintegral viaMonteCarlo.
1
ConditionalPrior
InterimPosterior
InterimPrior
Likelihood
Ongoing research question:How many interim samples are needed?
-
LLNL-PRES-73305517
Sourcecharacterizationviaprobabilisticimagemodeling
Infer image model parameters via MCMC under an interim prior distribution for the galaxy and PSF parameters.
MBI GREAT3 analysis with:The Tractor (Lang & Hogg)
Now use GalSim + MCMC
GalSim models inside an MCMC chain – Can it be made fast enough?
-
LLNL-PRES-73305518
Exampleinterimposteriorinferencesforgalaxystampimages
-
LLNL-PRES-73305519
ProbabilisticforwardmodelingcanmeetLSSTshearbiasrequirements… atleastwhentestedonsimulatedimages
§ GREAT3CGC-likesetup— 200’fields’withconstantshearperfield— 10kgalaxiesperfield
§ Marginalize7parameterspergalaxy:— e1,e2,HLR,flux,dx,dy,n— Notable:Sersic indexmarginalized
§ HaveNOTmarginalizedPSF(yet!)
Immediate takeaway: Hierarchical inference performs significantly better than ensemble average maximum likelihood ellipticity.
-
LLNL-PRES-73305520
Multi-epoch & multi-telescope data sets
-
LLNL-PRES-73305521
Howdowecombinemultipleobservationsofthesamegalaxy?Naïvelywemustjointfitallepochssimultaneously
arXiv:1511.03095Generalized Multiple Importance SamplingElvira, Martino, Luengo, & Bugallo
Problem: Imagine we have fit pixel data from LSST year 1. How do we incorporate year 2 observations without redoing (expensive) calculations?
Solution: Consider single-epoch samples as draws from a multi-modal importance sampling distribution:
-
LLNL-PRES-73305522
‘cross-pollination’ needed:
Evaluate the likelihood of epoch i given model parameter samples
from epoch j, for all combinations of i, j.
A standard scatter / gather operation
Multipleimportancesampling(MIS)viaweightedpseudo-marginals
1. Sample from the conditional posterior for each epoch individually
2. Evaluate the ratio of the conditional posterior for each epoch i to that of the MIS sampling distribution
-
LLNL-PRES-73305523
MultipleimportancesamplingenablesstreamingdataanalysisEfficiencyissignificantlyenhancedbyusingolddataasasampling‘prior’
§ Drawparametersamplesfromfirstepochunderanominalinterimprior
§ Drawsamplesfromsubsequentepochswithapriorinformedbypreviousepochsamples
§ Simulationstudiesshow:— ~10%ofsampleshavesignificantweight
whencombining200epochsinstreamingfashion
-
LLNL-PRES-73305524
PSF marginalization
-
LLNL-PRES-73305525
MarginalizingPSFs:MISmakesthistractable
§ LSSTwillhave~200epochsperobjectperfilter— WeaimtomarginalizethePSF∏n,i inevery
epoch— Themarginalizationisconstrainedby:
• ConsistencyofPSFrealizationsoverthefocalplaneforeachepoch
• Consistencyoftheunderlyingsourcemodelacrossepochs
§ Simplestapproach(statistically,notcomputationally):Infergalaxymodelsgivenallepochimagingsimultaneously— “Interim”samplesareofsize:~10galaxy
params +200*~4PSFparams =~1kparameters!
-
LLNL-PRES-73305526
1. Fitstarfootprintsinallepochsviaprobabilisticforwardmodels
2. MarginalizestarimageparameterstoconstraintheglobalfieldPSFmodelforeachepoch— Stateoftheopticsaberrations,and— Distributionofatmosphereturbulencestatistics
3. Fitallgalaxyfootprintsineachepochviaforwardmodels— UsePSFmodelsdrawnfromthemarginalposteriorgiventhestarimages
4. RunThresherontheinterimgalaxysamplesforallepochs(via‘cross-pollinator’)
ThepipelineforPSFmarginalization
One approximation needed: Marginalize PSF model independently for each field location
-
LLNL-PRES-73305527
Probabilistic cosmological one-point statistics
-
LLNL-PRES-73305528
Probabilisticcosmologicalmassmapping
Zero E/B mode mixing by construction
Objective: infer the 3D gravitational potential of the initial conditions
arXiv:1610.06673
-
LLNL-PRES-73305529
Hierarchicalinferenceofcosmologicallensingmassdistributions
Validation with simulations A real merging galaxy cluster New:• Linear and
nonlinear scales reconstructed in one framework
• No E/B mode mixing by construction
Schneider+2017, ApJ
-
LLNL-PRES-73305530
Applicationtodata:WeaklensingmassmapsforAbel781merginggalaxyclustersasseenbytheDeepLensSurvey
Our method Previously published method
-
LLNL-PRES-73305531
Latent features of the galaxy distribution
-
LLNL-PRES-73305532
Pr(eint)isnotGaussian!
§ WouldrathernotassertaparticularparametricformforP(eint).
§ Usea“non-parametric”distribution:aDirichletProcessMixtureModel
3
Ellipticities from COSMOS
-
LLNL-PRES-73305533
Hierarchicalinferenceofintrinsic galaxyproperties
Specify a Dirichlet Process (DP) for the distribution of intrinsic galaxy property hyper-parameters
The DP is a ‘non-parametric’ distribution with discrete support
The DP distribution allows clustering of data points (e.g., galaxies) to infer latent structure in the data.
-
LLNL-PRES-73305534
GibbsupdatesintheDirichlet Processmodel
Pr(cn = c`|c�n,!n,↵,X ) = bN�n,c Pr(dn|↵c` ,X ), 8` 6= n
Pr(cn 6= c`8` 6= n|c�n,!n,X ) = bZ
Pr(dn|↵,X )G0(↵) d↵,
↵cn ⇠ G0 (↵cn)NcnY
`=1
Pr(d`|↵cn ,X ))
Latent class assignments are updated with different conditional distributions depending on whether any other observations are assigned to the current class.
The DP mixture parameters are simply updated with the posterior given all observations currently associated with the given latent class.
ZPr(dn|↵,X )G0(↵) d↵ =
ZnN
NX
k=1
Prmarg(!nk|a)Pr(!nk|I0)
Prmarg(!nk|a) ⌘Z
d↵cn G0(↵cn |a)Pr(!nk|↵cn)
Pr(cn 6= c`8` 6= n|c�n,!n,X ) = bZ
Pr(dn|↵,X )G0(↵) d↵
Highlighted integral is expensive to compute in general.
With importance sampling we only require the DP base distribution to be conjugate to the distribution of galaxy properties – NOT the likelihood.
Neal (2000)
-
LLNL-PRES-73305535
Asimulationstudywith100galaxiesvalidatestheDPmodel
100 galaxies drawn from 1 of 2 Gaussian ellipticity distributions
-
LLNL-PRES-73305536
Simulationstudy:Wecanbeatthetraditional‘shapenoise’statisticalerrorboundbyinferringlatentstructureinthedata
100 galaxies drawn from 1 of 2 Gaussian ellipticity distributions
3x improvement in cosmic shear precision
-
LLNL-PRES-73305537
GREAT3results
§ TestedhierarchicalapproachusingsimulationsfromthethirdGRavitationallEnsingAccuracyTest(GREAT3).
§ Hierarchicalinferenceperformssignificantlybetterthanensembleaveragemaximumlikelihoodellipticity.
§ TheDPMMellipticitypriorperformsbetterthanthesingleGaussianellipticityprior.
Mean ML ellipticityHierarchical Inference
: 13% shear calibration errorsH.I. : 4% shear calibration errors
Dirichlet Process Inference
DP : 1-2% shear calibration errors
Input shearSh
ear r
esid
uals
-
LLNL-PRES-73305538
Multi-variateDPmixturemodel (inprogress):“standardizable”ellipticities.
§ Ellipticalgalaxieshaveanarrowerintrinsicellipticitydistributionthanlate-type.Highersensitivitytoshear!
§ Ellipticals/spiralsalsodistinguishablebycolorandmorphology(e.g.,Sersicindex,Ginicoefficient,asymmetry),potentiallyprovidingadditionalvariableswithwhichtocluster.
§ Othercorrelationstoexploit?
-
LLNL-PRES-73305539
ApplicationtotheDeepLensSurvey:realgalaxiesrequireatleast2latentclasses(ignoringlensing)
We infer 2 latent classes given only an ellipticity catalog
Preliminary: The marginal posterior distribution of ellipticity variance from the Deep Lens Survey
photo-z: [0.65, 0.7]
-
LLNL-PRES-73305540
TheprobabilisticweaklensingworkflowplanforLSST
Probabilistic pipeline
-
LLNL-PRES-73305541
§ Cosmic shear is systematics limited & signal is dominated by PSF and astrophysics
— A probabilistic approach is warranted to infer a small signal and mitigate biases
§ A hierarchical probabilistic model for cosmic shear can trade bias for variance, but also can increase precision by learning latent structure in the galaxy distribution.
§ Importance sampling methods allow tractable approaches to a probabilistic forward model of LSST & WFIRST imaging
— With billions of galaxies and hundreds of epochs per galaxy modeling LSST or WFIRST imaging requires an approach to separating analyses of data subsets, even though statistically correlated
§ We are able to sample from a probabilistic model with multiple hierarchies to marginalize both correlated image systematics and astrophysical properties of galaxies.
Summary
Probabilistic