![Page 1: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/1.jpg)
TheNa'onalCenterforAtmosphericResearchissponsoredbytheNa'onalScienceFounda'on.Anyopinions,findingsandconclusionsorrecommenda'onsexpressedinthispublica'onarethoseoftheauthor(s)anddonotnecessarilyreflecttheviewsoftheNa'onalScienceFounda'on.
©UCAR
DARTTutorialPart IV:OtherUpdatesforanObservedVariable
![Page 2: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/2.jpg)
Don’tknowmuchaboutproper'esofthissample.Maynaivelyassumeitisrandomdrawfrom‘truth’.
Ensemblefilters:Priorisavailableasfinitesample.
−4 −2 0 2 40
0.2
0.4
0.6
Prob
abilit
y
Prior Ensemble
ProductofTwoGaussians
p A | BC( ) = p(B | AC)p(A |C)p(B |C)
= p(B | AC)p(A |C)p(B | x)p(x |C)dx∫
![Page 3: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/3.jpg)
Howcanwetakeproductofsamplewithcon'nuouslikelihood?
Fitacon'nuous(Gaussianfornow)distribu'ontosample.−4 −2 0 2 40
0.2
0.4
0.6
Prob
abilit
y
Prior PDF
Prior Ensemble
ProductofTwoGaussians
p A | BC( ) = p(B | AC)p(A |C)p(B |C)
= p(B | AC)p(A |C)p(B | x)p(x |C)dx∫
![Page 4: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/4.jpg)
Observa'onlikelihoodusuallycon'nuous(nearlyalwaysGaussian).
IfObs.likelihoodisn’tGaussian,cangeneralizemethodsbelow.
−4 −2 0 2 40
0.2
0.4
0.6
Prob
abilit
y
Prior PDF
Obs. Likelihood
Prior Ensemble
ProductofTwoGaussians
p A | BC( ) = p(B | AC)p(A |C)p(B |C)
= p(B | AC)p(A |C)p(B | x)p(x |C)dx∫
![Page 5: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/5.jpg)
ProductofpriorGaussianfitandObs.likelihoodisGaussian.
Compu'ngcon'nuousposteriorissimple.BUT,needtohaveaSAMPLEofthisPDF.
−4 −2 0 2 40
0.2
0.4
0.6
Prob
abilit
y
Prior PDF
Obs. Likelihood
Posterior PDF
Prior Ensemble
ProductofTwoGaussians
p A | BC( ) = p(B | AC)p(A |C)p(B |C)
= p(B | AC)p(A |C)p(B | x)p(x |C)dx∫
![Page 6: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/6.jpg)
SamplingPosteriorPDF:
Therearemanywaystodothis.
Exactproper'esofdifferentmethodsmaybeunclear.Trialanderrors'llbestwaytoseehowtheyperform.Willinteractwithproper'esofpredic'onmodels,etc.
−2 −1 0 1 2 30
0.2
0.4
0.6Posterior PDF
Prob
abilit
y
![Page 7: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/7.jpg)
SamplingPosteriorPDF:
−2 −1 0 1 2 30
0.2
0.4
0.6Posterior PDF
Prob
abilit
y
Random Sample
Justdrawarandomsample(filter_kind=5in&assim_tools_nml).
![Page 8: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/8.jpg)
SamplingPosteriorPDF:
−2 −1 0 1 2 30
0.2
0.4
0.6Posterior PDF
Prob
abilit
y
Random Sample; Exact Mean
Justdrawarandomsample(filter_kind=5in&assim_tools_nml).
Can‘playgames’withthissampletoimprove(modify)itsproper'es.
Example: Adjustthemeanofthesampletobeexact. Canalsoadjustthevariancetobeexact.
![Page 9: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/9.jpg)
SamplingPosteriorPDF:
−2 −1 0 1 2 30
0.2
0.4
0.6Posterior PDF
Prob
abilit
y
Random Sample; Exact Mean and Var.
Justdrawarandomsample(filter_kind=5in&assim_tools_nml).
Can‘playgames’withthissampletoimprove(modify)itsproper'es.
Example: Adjustthemeanofthesampletobeexact. Canalsoadjustthevariancetobeexact.
![Page 10: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/10.jpg)
SamplingPosteriorPDF:
−2 −1 0 1 2 30
0.2
0.4
0.6Posterior PDF
Prob
abilit
y
Random Sample; Exact Mean and Var.
Mightalsowanttoeliminaterareextremeoutliers.
NOTE:Proper'esoftheseadjustedsamplescanbequitedifferent.Howtheseproper'esinteractwiththerestoftheassimila'onisanopenques'on.
Justdrawarandomsample(filter_kind=5in&assim_tools_nml).
![Page 11: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/11.jpg)
SamplingPosteriorPDF:
Constructa‘determinis'c’samplewithcertainfeatures.
Forinstance:Samplecouldhaveexactmeanandvariance.
Thisisinsufficienttoconstrainensemble,needotherconstraints.
−3 −2 −1 0 1 2 3 40
0.2
0.4
0.6Posterior PDF
Prob
abilit
y
![Page 12: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/12.jpg)
SamplingPosteriorPDF:
Constructa‘determinis'c’samplewithcertainfeatures(filter_kind=6in&assim_tools_nml;manuallyadjustkurtosis).
Example:Exactsamplemeanandvariance.
Samplekurtosis(related to the sharpness/tailedness of a distribution) is3, which is the expectedvaluefora normal distribution. Startby assuming auniformly-spacedsample andadjus'ngquadra'cally.
−3 −2 −1 0 1 2 3 40
0.2
0.4
0.6Posterior PDF
Prob
abilit
y
Kurtosis 3
![Page 13: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/13.jpg)
SamplingPosteriorPDF:
−3 −2 −1 0 1 2 3 40
0.2
0.4
0.6Posterior PDF
Prob
abilit
y
Kurtosis 3
Kurtosis 2
Example:Exactsamplemeanandvariance.
Samplekurtosis2:lessextremeoutliers,lessdensenearmean.Avoidingoutliersmightbeniceincertainapplica'ons.Samplingheavilynearmeanmightbenice.
Constructa‘determinis'c’samplewithcertainfeatures(filter_kind=6in&assim_tools_nml;manuallyadjustkurtosis).
![Page 14: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/14.jpg)
SamplingPosteriorPDF:
−4 −2 0 2 40
0.2
0.4
0.6
Prior Ensemble
Prob
abilit
yFirsttwomethodsdependonlyonmeanandvarianceofpriorsample.
Example:Supposepriorsampleis(significantly)bimodal?
Mightwanttoretainaddi'onalinforma'onfromprior.RecallthatEnsembleAdjustmentFiltertriedtodothis(Sec'on1).
![Page 15: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/15.jpg)
SamplingPosteriorPDF:
−4 −2 0 2 40
0.2
0.4
Prior Ensemble
Prob
abilit
y
Prior PDFObs. Likelihood
Posterior PDF
Random Posterior Ensemble
Firsttwomethodsdependonlyonmeanandvarianceofpriorsample.
Example:Supposepriorsampleis(significantly)bimodal?
Mightwanttoretainaddi'onalinforma'onfromprior.RecallthatEnsembleAdjustmentFiltertriedtodothis(Sec'on1).
![Page 16: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/16.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
Prob
abilit
y
Prior Ensemble
‘Classical’MonteCarloalgorithmfordataassimila'on
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 17: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/17.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
Prob
abilit
y
Prior Ensemble
Again,fitaGaussiantothesample.
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 18: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/18.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
Prob
abilit
y
Prior Ensemble
Obs. Likelihood
Again,fitaGaussiantothesample.Aretherewaystodothiswithoutcompu'ngpriorsamplestats?
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 19: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/19.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
Prob
abilit
y
Prior Ensemble
Obs. Likelihood
Random Draws from Obs.
Generatearandomdrawfromtheobserva'onlikelihood.Associateitwiththefirstsampleofthepriorensemble.
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 20: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/20.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
Prob
abilit
y
Prior Ensemble
Obs. Likelihood
Random Draws from Obs.
Havesampleofjointpriordistribu'onforobserva'onandpriorMEAN. Adjus'ngthemeanofobs.sampletobeexactimprovesperformance.
Adjus'ngthevariancemayfurtherimproveperformance.Outliersarea poten'alproblem,butcanberemoved.
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 21: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/21.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
Probability
Foreachpriormean/obs.pair,findmeanofposteriorPDF.
DARTTutorialSec'on6:Slide24
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 22: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/22.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
Probability
Priorsamplestandarddevia'ons'llmeasuresuncertaintyofpriormeanes'mate.
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 23: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/23.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
Probability
Priorsamplestandarddevia'ons'llmeasuresuncertaintyofpriormeanes'mate.Obs.likelihoodstandarddevia'onmeasuresuncertaintyofobs.es'mate.
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 24: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/24.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
Prob
abilit
y
Posterior PDF
Takeproductoftheprior/obsdistribu'onsforfirstsample.ThisisthestandardGaussianproduct.
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 25: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/25.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
Probability
Meanofproductisrandomsampleofposterior.Productofrandomsamplesisrandomsampleofproduct.
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 26: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/26.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
Probability
Repeatthisopera'onforeachjointpriorpair.
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 27: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/27.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
Probability
Posteriorsamplemaintainsmuchofpriorsamplestructure.(Thisismoreapparentforlargerensemblesizes.)
Posteriorsamplemeanandvarianceconvergeto‘exact’forlargesamples.
EnsembleKalmanFilter(EnKF)(filter_kind=2in&assim_tools_nml).
![Page 28: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/28.jpg)
EnsembleFilterAlgorithms:EnsembleKernelFilter(EKF)
(filter_kind=3in&assim_tools_nml).
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
Prior Ensemble
Canretainmorecorrectinforma'onaboutnon-Gaussianpriors.Canalsobeusedforobs.likelihoodterminproduct(notshownhere).
![Page 29: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/29.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
Prior Ensemble
Prior PDF
Usually,kernelwidthsareafunc'onofthesamplevariance.
EnsembleKernelFilter(EKF)(filter_kind=3in&assim_tools_nml).
![Page 30: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/30.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
Prior Ensemble
Prior PDFObs. Likelihood
Usually,kernelwidthsareafunc'onofthesamplevariance.
EnsembleKernelFilter(EKF)(filter_kind=3in&assim_tools_nml).
![Page 31: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/31.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
Obs. Likelihood
Example Kernels: Half as Wide as Prior PDF
ApproximatepriorasasumofGaussianscenteredoneachsample.
EnsembleKernelFilter(EKF)(filter_kind=3in&assim_tools_nml).
![Page 32: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/32.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
Obs. Likelihood
Example Kernels: Half as Wide as Prior PDF
Normalized Sum of Kernels
Thees'mateoftheprioristhenormalizedsumofallkernels.
EnsembleKernelFilter(EKF)(filter_kind=3in&assim_tools_nml).
![Page 33: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/33.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
Obs. Likelihood
Applydistribu'velawtotakeproduct: productofthesumisthesumoftheproducts. Otherwise, the
product cannot be analytically determined.
EnsembleKernelFilter(EKF)(filter_kind=3in&assim_tools_nml).
![Page 34: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/34.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
Obs. Likelihood
ComputeproductoffirstkernelwithObs.likelihood.
EnsembleKernelFilter(EKF)(filter_kind=3in&assim_tools_nml).
![Page 35: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/35.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
Obs. Likelihood
But,cannolongerignoretheweighttermforproductofGaussians.Kernelswithmeanfurtherfromobserva'ongetlessweight.
EnsembleKernelFilter(EKF)(filter_kind=3in&assim_tools_nml).
![Page 36: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/36.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
Obs. Likelihood
Con'nuetotakeproductsforeachkernelinturn.Closerkernelsdominateposterior.
EnsembleKernelFilter(EKF)(filter_kind=3in&assim_tools_nml).
![Page 37: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/37.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
Obs. LikelihoodNormalized Sum of Posteriors
Finalposteriorisweight-normalizedsumofkernelproducts.
PosteriorissomewhatdifferentthanforensembleadjustmentorensembleKalmanfilter(muchlessdensityinleolobe.)
EnsembleKernelFilter(EKF)(filter_kind=3in&assim_tools_nml).
![Page 38: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/38.jpg)
EnsembleFilterAlgorithms:
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
Obs. LikelihoodNormalized Sum of Posteriors
Posterior Ensemble
Formingsampleoftheposteriorcanbeproblema'c.Randomsampleissimple.
Determinis'csamplingismuchmoretrickyhere(fewresultsavailable).
EnsembleKernelFilter(EKF)(filter_kind=3in&assim_tools_nml).
![Page 39: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/39.jpg)
EnsembleFilterAlgorithms:
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
PriorProb
abilit
y D
ensi
ty Applyforwardoperatortoeachensemblemember.Getpriorensembleinobserva'onspace.
RankHistogramFilter(filter_kind=8in&assim_tools_nml).
Goal:Wanttohandlenon-Gaussianpriorsorobserva'onlikelihoods.Lowinforma'oncontentobs.mustyieldsmallincrements.
MustperformwellforGaussianpriors.Mustbecomputa'onallyefficient.
![Page 40: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/40.jpg)
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
PriorProb
abilit
y D
ensi
ty
Step1:Getcon'nuouspriordistribu'ondensity.• Place(ens_size+1)-1massbetweenadjacentensemblemembers.• Reminiscentofrankhistogramevalua'onmethod.
EnsembleFilterAlgorithms:
DARTTutorialSec'on6:Slide60
RankHistogramFilter(filter_kind=8in&assim_tools_nml).
![Page 41: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/41.jpg)
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
PriorProb
abilit
y D
ensi
ty
Step1:Getcon'nuouspriordistribu'ondensity.• Par'alGaussiankernelsontails,N(tail_mean,ens_sd).• tail_meanselectedsothat(ens_size+1)-1massisintail.
EnsembleFilterAlgorithms:RankHistogramFilter(filter_kind=8in&assim_tools_nml).
![Page 42: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/42.jpg)
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
PriorProb
abilit
y D
ensi
ty
Step2:Uselikelihoodtocomputeweightforeachensemblemember.• Analogoustoclassicalpar'clefilter.• Canbeextendedtonon-Gaussianobs.likelihoods.
EnsembleFilterAlgorithms:RankHistogramFilter(filter_kind=8in&assim_tools_nml).
![Page 43: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/43.jpg)
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
PriorProb
abilit
y D
ensi
ty
Step2:Uselikelihoodtocomputeweightforeachensemblemember.• Canapproximateinteriorlikelihoodwithlinearfit;forefficiency.
EnsembleFilterAlgorithms:RankHistogramFilter(filter_kind=8in&assim_tools_nml).
![Page 44: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/44.jpg)
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
PriorProb
abilit
y D
ensi
ty
Step3:Computecon'nuousposteriordistribu'on.• Approximatelikelihoodwithtrapezoidalquadrature,takeproduct.
(DisplayedproductnormalizedtomakeposterioraPDF).
EnsembleFilterAlgorithms:RankHistogramFilter(filter_kind=8in&assim_tools_nml).
![Page 45: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/45.jpg)
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
PriorProb
abilit
y D
ensi
ty
Step3:Computecon'nuousposteriordistribu'on.• ProductofpriorGaussiankernelwithlikelihoodfortails.
EnsembleFilterAlgorithms:RankHistogramFilter(filter_kind=8in&assim_tools_nml).
![Page 46: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/46.jpg)
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
PriorProb
abilit
y D
ensi
ty
RHF Posterior
Step4:Computeupdatedensemblemembers:• (ens_size+1)-1ofposteriormassbetweeneachensemblepair.• (ens_size+1)-1ineachtail.• Uninforma'veobserva'on(not shown) would havenoimpact.
EnsembleFilterAlgorithms:RankHistogramFilter(filter_kind=8in&assim_tools_nml).
![Page 47: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/47.jpg)
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
PriorProb
abilit
y D
ensi
ty
RHF Posterior
EAKF Posterior
ComparetostandardensembleadjustmentKalman filter(EAKF).NearlyGaussiancase,differencesaresmall.
EnsembleFilterAlgorithms:RankHistogramFilter(filter_kind=8in&assim_tools_nml).
![Page 48: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/48.jpg)
−3 −2.5 −2 −1.5 −1 −0.5 0 0.5 1
0
1
2
3
PriorProb
abilit
y D
ensi
ty
RHF Posterior
EAKF Posterior
RankHistogramgetsridofprior outlierthatisinconsistentwithobs. EAKFcan’tgetridofthis prior outlier.
LargepriorvariancefromoutliercausesEAKFtoshioallmemberstoomuchtowardsobserva'on (with mean off the page).
EnsembleFilterAlgorithms:RankHistogramFilter(filter_kind=8in&assim_tools_nml).
![Page 49: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/49.jpg)
-3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5
Prob
abilit
y D
ensi
ty
0
0.5
1
1.5
PriorRHF Posterior
EAKF Posterior
Convec've-scalemodels(andlandmodels)haveanalogousbehavior.Convec'onmayfireat‘random’loca'ons.Subsetofensembleswillbeinrightplace,restinwrongplace.Wanttoaggressivelyeliminateconvec'oninwrongplace.
EnsembleFilterAlgorithms:RankHistogramFilter(filter_kind=8in&assim_tools_nml).
![Page 50: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/50.jpg)
DealingwithEnsembleFilterErrorsFix1,2,3independently,HARDbutongoing.O\en,ensemblefilters...1-4:Varianceinfla'on,Increaseprioruncertaintytogiveobsmoreimpact.5.‘Localiza'on’:onlyletobs.impactasetof‘nearby’statevariables.O\ensmoothlydecreaseimpactto0asfunc'onofdistance.
****
1. Model Error
2. h errors; Representativeness
3. ‘Gross’ Obs. Error 4. Sampling Error; Gaussian Assumption
5. Sampling Error; Assuming Linear Statistical Relation
tk
tk+1
tk+2
h h h
DARTTutorialSec'on9:Slide3
![Page 51: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/51.jpg)
Model/FilterError:FilterDivergenceandVarianceInfla'on
−4 −3 −2 −1 00
0.5
1
Prob
abilit
y
"TRUE" Prior PDF
1.Historyofobserva'onsandphysicalsystem=>‘true’distribu'on.
![Page 52: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/52.jpg)
Model/FilterError:FilterDivergenceandVarianceInfla'on
−4 −3 −2 −1 00
0.5
1
Prob
abilit
y
"TRUE" Prior PDF
Variance Deficient PDF
1.Historyofobserva'onsandphysicalsystem=>‘true’distribu'on.2.Samplingerror,somemodelerrorsleadtoinsufficientpriorvariance.3.Canleadto‘filterdivergence’:prioristooconfident,obs.Ignored.
![Page 53: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/53.jpg)
Model/FilterError:FilterDivergenceandVarianceInfla'on
−4 −3 −2 −1 00
0.5
1
Prob
abilit
y
"TRUE" Prior PDF
Variance Deficient PDF
1.Historyofobserva'onsandphysicalsystem=>‘true’distribu'on.2.Samplingerror,somemodelerrorsleadtoinsufficientpriorvariance.3.Canleadto‘filterdivergence’:prioristooconfident,obs.Ignored.
Naïvesolu'onisvarianceinfla'on:justincreasespreadofprior.Forensemblememberi, inflate xi( ) = λ xi − x( )+ x
![Page 54: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/54.jpg)
Model/FilterError:FilterDivergenceandVarianceInfla'on
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
"TRUE" Prior PDF Error in Mean (from model)
1.Historyofobserva'onsandphysicalsystem=>‘true’distribu'on.2.Mostmodelerrorsalsoleadtoerroneousshi\inen'redistribu'on.3.Again,priorcanbeviewedasbeingTOOCERTAIN.
![Page 55: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/55.jpg)
Model/FilterError:FilterDivergenceandVarianceInfla'on
−4 −2 0 2 40
0.2
0.4
0.6
0.8
Prob
abilit
y
"TRUE" Prior PDF Error in Mean (from model)
Variance Inflated
Infla'ngcanamelioratethis.Obviously,ifweknewE(error),we’dcorrectforitdirectly.
1.Historyofobserva'onsandphysicalsystem=>‘true’distribu'on.2.Mostmodelerrorsalsoleadtoerroneousshi\inen'redistribu'on.3.Again,priorcanbeviewedasbeingTOOCERTAIN.
![Page 56: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/56.jpg)
PhysicalSpaceVarianceInfla'on
Inflateallstatevariablesbysameamountbeforeassimila'on.Capabili'es:1. Canbeeffec'veforavarietyofmodels.2. Canmaintainlinearbalances.3. Prior continues to resemble that from the first guess.
4. Simpleandcomputationally cheap.
Liabili'es:
1. Statevariablesnotconstrainedbyobserva'onscan‘blowup’.ForinstanceunobservedregionsnearthetopofAGCMs.
2. Magnitudeofλnormallyselectedbytrialanderror.
![Page 57: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/57.jpg)
PhysicalSpaceVarianceInfla'oninLorenz63
−20
0
20 −20
0
2010
20
30
40
Observa'onoutsideprior:dangeroffilterdivergence.
Observa'oninred.Priorensembleingreen.
![Page 58: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/58.jpg)
PhysicalSpaceVarianceInfla'oninLorenz63
−20
0
20 −20
0
2010
20
30
40
A\erinfla'ng,observa'onisinpriorcloud:filterdivergenceavoided.
Observa'oninred.Priorensembleingreen.
Inflatedensembleinmagenta.
![Page 59: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/59.jpg)
PhysicalSpaceVarianceInfla'oninLorenz63
−20
0
20 −20
0
2010
20
30
40
Priordistribu'onissignificantly‘curved’.
Observa'oninred.Priorensembleingreen.
![Page 60: DART Tutorial Part IV - Mesoscale Research Groupderecho.math.uwm.edu/classes/NWP/21-DARTTutorialSec69.pdf · Don’t know much about properes of this sample. May naively assume it](https://reader031.vdocuments.us/reader031/viewer/2022021603/5cb8563088c99331688b5d40/html5/thumbnails/60.jpg)
PhysicalSpaceVarianceInfla'oninLorenz63
−20
0
20 −20
0
2010
20
30
40
Inflatedprioroutsidealractor.Posteriorwillalsobeoffalractor.
Observa'oninred.Priorensembleingreen.
Inflatedensembleinmagenta.
Canleadtotransientoff-alractorbehavioror…model‘blow-up’.