Classical and quantum causal inference

54
Classical and quantum causal inference (An introduction to techniques and open questions) Sally Shrapnel

Transcript of Classical and quantum causal inference

Classical and quantum causal inference

(An introduction to techniques and open questions)

SallyShrapnel

Classicalcausalinference

...theLaplacianconceptionismoreintunewithhumanintuitions.Thefewesoteric

quantumexperimentsthatconflictwiththepredictionsoftheLaplacianconception

evokesurpriseanddisbelief,andtheydemandscientistsgiveupdeeplyentrenched

intuitionsaboutlocalityandcausality.Ourobjectiveistopreserve,explicateand

satisfy- notdestroy- thoseintuitions.(Pearl,2009)[26]

Classicalcausalinference

Glymour onBellexperiments:

...realexperiments...createassociationsthathavenocausalexplanation

consistentwiththeMarkovassumption,andtheMarkovassumptionmustbe

appliedtoobtainthatconclusion.Youcansaythereisnocausalexplanationof

thephenomenon,orthatthereisacausalexplanationbutitdoesn’tsatisfythe

Markovassumption.(Glymour,2006)[124]

Classicalcausalinference

Currentquantumtheoryandexperimentsshowthatthisassumptiondoesnotholdat

thequantumlevel,wheretherearenonlocalvariablesthatappeartohaveadirect

causaleffectonothers.WhilethesecasesdonotimplythatthecausalMarkov

assumptiondoesnothold,theydosuggestthatwemayseemoreviolationsofthis

assumptionatthequantumlevel.However,inpractice,theCausalMarkovassumption

assumptionappearstobeareasonableworkingassumptioninmostmacroscopic

systems.

Outline

1. Classicalcausalinference

2. Quantumcausalinference

3. Greatintheory,butwhataboutinpractice?

4. Machinelearningtotherescue?

Causalinference

Petersetal,2018

p(Ph,A)

p(Ph,B)

Causalstructurematters

Petersetal,2018

interventionistcausation

CausationisNOTsimplycorrelation

Correlationsdonotenableustodistinguishbetweeneffectiveandineffectivestrategiesthatbringaboutspecificends(Cartwright,1979)

“interventioninvariantmodel”

interventionistcausation

CausationisNOTsimplycorrelation

• correlationsdonotenableustodistinguishbetweeneffectiveandineffectivestrategiesthatbringaboutspecificends(Cartwright,1979)

Goldstandard

interventionistcausation

Interveningandrandomisation oftennotpossible

• Lookforfootprintsinthedatatogiveuscluestothetopologyofthecausalstructure

• Needprinciplesorassumptionstogofromjointdistributionoverobservedvariablestocausalstructure.

Causalinference

Reichenbach’sprinciple:ifXandYarestatisticallydependent,then

1. thereexistseitheracommoncauseZoradirectcausalrelationshipbetweenXandY,and

2. ZscreensoffXandYfromeachother(X⊥Y⏐ Z)

X

Z

Y

X Y

X Y

Causalinference

Jointdistributionisthesameinallthreecases:causalstructure capturedbydifferentpossiblefactorisations

X

Z

Y

X Y

X Y

! 𝑝 𝑋 𝑍 𝑝 𝑌 𝑍 𝑝(𝑍)�

)

𝑝(𝑌)𝑝(𝑋|𝑌)

𝑝(𝑋)𝑝(𝑌|𝑋)

𝑝(𝑋, 𝑌)

CausalgraphicalmodelCasualmodel=graph+causalparametersCausalparameters=distributionofeachvariableconditionedonitsparents

X4

X3

X5

X1

X2

Structuralcausalmodel

𝑋1 = 𝑓1(𝑁1)𝑋2 = 𝑓2(𝑁2)

𝑋3 = 𝑓3(𝑋1, 𝑁3)𝑋4 = 𝑓4(𝑋3, 𝑁4)

𝑋5 = 𝑓5 𝑋3, 𝑋2, 𝑁5 X4

X3

X5

X1

X2

N4

N5

N2N3

N1

MarkovCondition

X4

X3

X5

X1

X2

N4

N5

N2N3

N1

parentsNon-descendants

Graphically,children areconditionallyindependentoftheirnon-descendants,giventheirparents.

(𝑋 ⊥ 𝑌 𝑍 𝐺→ (𝑋 ⊥ 𝑌 𝑍 𝑃

FaithfulnessCondition

X4

X3

X5

X1

X2

N4

N5

N2N3

N1

parentsNon-descendants

Graphically,childrenareconditionallyindependentoftheirnon-descendants,giventheirparents.

Independenciesfoundinthedistributionareonly thoseimpliedbytheMarkovcondition

(𝑋 ⊥ 𝑌 𝑍 𝐺 → (𝑋 ⊥ 𝑌 𝑍 𝑃

(𝑋 ⊥ 𝑌 𝑍 𝐺 ← (𝑋 ⊥ 𝑌 𝑍 𝑃

FaithfulnessCondition

X4

X3

X5

X1

X2

N4

N5

N2N3

N1

parentsNon-descendants

Graphically,childrenareconditionallyindependentoftheirnon-descendants,giventheirparents.

Independenciesfoundinthedistributionareonly thoseimpliedbytheMarkovcondition

(𝑋 ⊥ 𝑌 𝑍 𝐺 → (𝑋 ⊥ 𝑌 𝑍 𝑃

(𝑋 ⊥ 𝑌 𝑍 𝐺 ← (𝑋 ⊥ 𝑌 𝑍 𝑃

Interventions

X4

X3

X5

X1

X2

N4

N5

N2

N1

Interventionseparatesvariablefromantecedentcauses

𝑋1 = 𝑓1(𝑁1)𝑋2 = 𝑓2(𝑁2)

𝑋3 = 𝑓3(X1,N3)𝑋4 = 𝑓4(𝑋3, 𝑁4)

Causalinference

X4

X3=2

X5

X1

X2

N4

N5

N2

N1

Interventionseparatesvariablefromantecedentcauses

𝑋1 = 𝑓1(𝑁1)𝑋2 = 𝑓2(𝑁2)

𝑋3 = 2𝑋4 = 𝑓4(𝑋3, 𝑁4)

Whyisitcausal?

LocalCPDaregeneratedbyautonomouscausalmechanisms thatexistbetweenparentsandchildren.

Causalstructureofmodelisisomorphictonetworkofautonomouscausalmechanisms.

Interventions bringlocaldistributionundercontrolofexperimenteranddon’tdisruptothermechanismsinthemodel

Thisstructurehaspragmaticvalue– ittellsushowtoacttobringaboutcertainends

Latentvariables?

GraphnotMarkovian andFaithful?

Implicitassumptionthatthisisalwaysaconsequenceofunmeasuredcommoncauses(latentvariables)

Inprinciple,locatingandmeasuringsuchvariablesoughttorestoreMarkovianity tothemodelforsomecausalstructure.

Constraintbasedmethods

Algorithms:PC

– Identifiesadjacenciesviadependencies

– IdentifiesV-structuresX YZXZXYZXYZY

– Propagationrulestoorientremainingedgestoavoidcycles

– IdentifiesMarkovequivalentsetofDAGsunderCMA,CFA,acyclicity andCSA

Constraintbasedmethods

PROBLEMS

Givesyougraphbutnotthestructuralequations(can’tinfercounterfactuals).

StatisticaltestsacceptorrejectCIatagivenconfidenceintervalandgraphstructureisverysensitivetohowyousetthisconfidencelevel.

CItestingisunjustifiedforfinitedata(forarbitrarilycomplexfunctions)

Propagationrulestendstopropagateerrors.

Whatabouttwovariables?

AlternativestoCItesting?

Generalprinciples?

Keyideatoretainisindependenceofcausalmechanisms:

“Statisticalcorrelationsbetweenvariablesinasystemaretheresultofacausalgenerativeprocess thatiscomposedofautonomousmodulesthatdonotinformorinfluenceeachother.”

Forbivariatecasethisreducestoindependenceofcause andmechanismrelatingcausetoeffect.

𝑝 𝑐, 𝑒 = 𝑝 𝑐)𝑝 𝑒 𝑐 (𝑐 → 𝑒

Asymmetryofcauseandeffect

𝑝 𝑥, 𝑦 = 𝑝 𝑥)𝑝 𝑦 𝑥 (𝑥 → 𝑦

𝑝 𝑥, 𝑦 = 𝑝 𝑦)𝑝 𝑥 𝑦 (𝑦 → 𝑥

𝑦 = 𝑥@ + 𝑥 + 𝑁

Wantlowstructuralvariabilityofmechanismfordifferentinputvaluesofcause

Alternatives?

Makeassumptionsaboutdatageneratingcausalmechanisms:

Bivariate

1. Additivenoisemodel(ANM)• assumenonlinearfunction,additivenoise,noiseindependentof

cause• Admitsonlyasingleunidentifiablecase:linearfunctionwithgaussian

causeandnoise

2. Post-nonlinearmodel(PNL)• Addsanextranonlinearfunctionwhichisinvertible• Morenon-identifiablecases

3. GaussianProcessInferencemodel(GPI)

4. Algorithmiccomplexity– exploitasymmetriesinfactorisation accordingtoKolmogorovcomplexitymeasures(selectssimplestexplanation)

Alternatives?

Multivariate:

1. ConstraintbasedmethodsPC,FCI(relaxesCS),RFCI.Allrequiredata++

2. Scorebasedalgorithmssearchmodelspaceandminimise aglobalscore• GESexploresgraphspaceusingoperators“addedge”,“removeedge”,

“reverseedge”andoptimises accordingtoBIC.• FGESmorecomputationallyefficient

3. Hybridalgorithms=constraint+scorebased• Max-minHillclimbing(MMHC)buildsskeletonusingCIteststhen

performsagreedyhill-climbingsearchtoorientedges• GFCIusesFGEStosketchgraphandthenCItoorientedges

4.Exploitasymmetrybetweencauseandeffect(LinGAM)+MORE• linearfunctions,non-gaussian sourcenodes,additivenoise

Crowdsourcedcausaldiscovery

Guyon,(2013)

Principledmethods(withvoting)– 0.6accuracyonclassificationtask

Generalised – 0.8(lowlevelfeaturesofjointdistribution- 9000!)

Notlearningthefunctionalrelationships,justdirectionality

machinelearningmethods

Data=x Model=f(x)p(y|x)

10

Optimise andregularise!

Label=y

machinelearningmethods

Supervisedlearning

Largedatabasesofsampledvariablepairswithknowncause-effectrelation.

Castasclassificationproblem.

Taskistogeneralise solutiontounseendatasets(borrowexistingregularisationmethodsfromML).

BUTneedtofeedit“groundtruth”models~motivatedbycommonsense,domainknowledge.

machinelearningmethods

Generativemodels((CiGAN,CausalGAN,CGNN)

SAM“StructuralAgnosticModel”withpenalised adversariallearning.

Recoverscasualgraphfromdata(learnsbothjointandinterventionaldistributions).

RelaxesCMC,CFC,CSA,acyclicity.

Scalestohundredsofvariables.

Estimatesbothstructureofgraphandfunctionalcausalmechanisms.

90%accurate.

Needtofeedit“groundtruth”models– sensitiveto“Bayeserror”.

Needsverification!

Classicalcausalinference

Causalinferencefromjointdistributionsisverydifficult.

Nicetheoreticalandconceptualapproaches,butmanypracticaldifficulties.

Machinelearningseemsapromising(bruteforce)approachbutatthecostofinterpretability(fornow).

Causalinferencetotheexclusionofotherdomainknowledgeisanoccupationalhazard…

Parachuteusetopreventdeathandmajortraumarelatedtogravitationalchallenge:systematicreviewofrandomised controlledtrialsBMJ 2003; 327

Objectives Todeterminewhetherparachutesareeffectiveinpreventingmajortraumarelatedtogravitationalchallenge.Design Systematicreviewofrandomised controlledtrials.Datasources:Medline,WebofScience,Embase,andtheCochraneLibrarydatabases;appropriateinternetsitesandcitationlists.Studyselection: Studiesshowingtheeffectsofusingaparachuteduringfreefall.Mainoutcomemeasure Deathormajortrauma,definedasaninjuryseverityscore>15.ResultsWewereunabletoidentifyanyrandomised controlledtrialsofparachuteintervention.

Conclusions Aswithmanyinterventionsintendedtopreventillhealth,theeffectivenessofparachuteshasnotbeensubjectedtorigorousevaluationbyusingrandomised controlledtrials.Advocatesofevidencebasedmedicinehavecriticisedtheadoptionofinterventionsevaluatedbyusingonlyobservationaldata.Wethinkthateveryonemightbenefitifthemostradicalprotagonistsofevidencebasedmedicineorganised andparticipatedinadoubleblind,randomised,placebocontrolled,crossovertrialoftheparachute.

Bellinequalities

2006Glymour:

Belltheoremisaparticularexampleofamoregeneraltheoryofcausalinference:

Causalgraphicalmodels

Glymour,Clark "MarkovPropertiesandQuantumExperiments,"inW.DemopoulosandI.Pitowsky,eds. PhysicalTheoryandItsInterpretation:EssaysinHonorofJeffreyBub,Springer2006.

Quantumcausalinference

“AnycausalmodelwhichcanreproduceBell-inequalityviolationswhilerespectingtheobservedindependences…willnecessarilyviolateaprinciplethatisatthecoreofallthebestcausaldiscoveryalgorithms [Faithfulness].”

Quantumcausalinference?

r

Non-localityviolatesfaithfulness.

• Settingsaremarginallyindependent𝐴 ⊥ 𝐵

• Nosignalling𝑋 ⊥ 𝐵 𝐴; 𝑌 ⊥ 𝐴 𝐵

Quantumcausalinference

r

Retrocausality violatesfaithfulness.

• Settingsaremarginallyindependent𝐴 ⊥ 𝐵

• Nosignalling𝑋 ⊥ 𝐵 𝐴; 𝑌 ⊥ 𝐴 𝐵

Quantumcausalinference

r

Superdeterminism violatesfaithfulness.

• Settingsaremarginallyindependent𝐴 ⊥ 𝐵

• Nosignalling𝑋 ⊥ 𝐵 𝐴; 𝑌 ⊥ 𝐴 𝐵

Quantumcausalinference

Conclusion:CItestingnotrichenoughmethodologytocapturequantumcausalrelations

Threeapproaches

1. ClassicalapproachPlugquantumdataintoclassicalalgorithms.Usefulascertificationforquantumness.Doesitreallyhelpwithcausalexplanation?

2. Quantumdomainisacausal.

3. Tryanddevelopamoregeneralversionofcausalinference.

Acausal?

Alternativeapproach

Assumequantumcausalstructureisprimitive,buildacausaltheoryfromthegroundup(usingmathematicalobjectsfromQM),thatrecoversclassicalcausalstructureinasuitablelimit.

Writedefinitionsaccordingtothewayphysicistsuse quantumtheorytomakeinterventionistinferencesthatdistinguishbetweeneffectiveandineffectivestrategies(inmostgeneralform).

Obeysindependenceassumptionsthatunderpinclassicalcausalinference.

CostaandShrapnel(2016)“Quantumcausalmodelling”,NJP,18,063062

Shrapnel(2016)“Usinginterventionstodiscoverquantumcausalstructure”PhDthesis,http://espace.library.uq.edu.au/view/UQ:411093

Shrapnel(2017)“DiscoveringQuantumCausalModels,”TheBritishJournalforthePhilosophyofScience(advancearticle)

Desiderata

1. Empirical =Theformalismshouldallowforthediscoveryofcausalstructurefromempiricaldata(causalstructure- canactasanoracleforinterventions).

2. Explanation =All correlationsbetweenempiricallyderiveddatashouldbeaccountedforvianotionsofdirect,indirectorcommoncauserelations,i.e.thereshouldbeno“unexplained”correlations.

3. Classicality=Classicalcausalmodelsshouldberecoveredasalimitingcaseofquantumones.

Quantumcausalmodels

A

B

C

D

G

F

E

Variables=regions

Values=CP maps

Intervention=instruments

CausalMechanisms=CPTP

Causalstructure=process

Generalised circuit

Processisgeneratedbyautonomouscausalmechanisms(channels).

Mechanisms=deterministicunitaries withunmodelled noise

=

Contextuality?

=

Shrapnel,CostaandMilburn,NJP (2018)ShrapnelandCosta,Quantum (2018)

𝑝 𝑗 = 𝑇𝑟(𝐸J𝜌)

Unique!

Unique!

Quantumcausalmodels

A

B

C

D

G

F

E

Variables=regions

Values=CPmaps

Intervention=instruments

CausalMechanisms=CPTP

Causalstructure=process

=

Assumptions:independenceofinterventionsandmechanisms.Do-calculusisignoreincomingstateandre-prepareoutgoingstate.Markovianity – setoflinearconstraintsontheprocess.Faithfulness– nofine-tunedmechanisms(measuretheoreticsense).

Quantumcausalmodels

A

B

C

D

G

F

E

Causalmodelis“interventioninvariant”

Non-Markovian?(unmodelled (latent)commoncause)- extendingthemodeltoincludesuchnodesrestorestheMarkovpropertytothecausalgraph

Canprovethatallclassicalcausalmodelscanbegivenaquantumrepresentation(allinstrumentsfixedinsomebasis)

Inprinciple:

MeasurementDataà graphthat

(i) Accountsforallcorrelations(classicalorquantum)viaautonomouscausalmechanisms

(ii) Actsasanoracleforfutureinterventions

Needcausaldiscoveryalgorithms….

Discoveryalgorithm:

1. Determinesifprocessiscausallyordered

2. Checksifalllatentvariablesareincluded

3. IfMarkoviangivesuniqueDAG

Giarmatzi andCosta,NQI,2018

Stillworktodo

Quantumcausalinferenceiscomputationallyand practicallyveryhard.

Needtoinputprocesswhichrequiresinformationallycompletetomography–exponentialinnumberofvariables.

Needtoknowdimensionofinputsystemsandnumberofsubsystems.

Switch?Indefinitecausalorder?LargerclassesofWthataretheoreticallypossible.

Supervisedlearningofprocess

Task=>classifyprocessasMarkovianvsnon-Markovian,=>estimatedimensionofnon-Markovianenvironment.

SimulatedataandusesupervisedMLtechnique– labelledprocesses.

TrainedRandomForestRegressor,testonunseendata(frominsideandoutsideoutsidetrainingrange)

99%accurateonMarkovianvsnon-Markovian.

95%accurateondimensionalityofenvironment.

Nolossofaccuracyorgeneralityonless-than-informationallycompletedata(only20%offeaturesincluded).

Supervisedlearningofprocess

Caveats:

Simulateddata– tryexperimentaldatanext.

Scaling?

Transferabletodifferentsizeprocesses?

Interpretability?

Conclusions

Causalinferenceisimpossiblewithoutassumptions.

Independenceofcausalmechanisms(includinginterventions)iskey.

Machinelearningmethodsmayprovideapowerfultoolforidentificationofcausalstructure(butatsomecosttointerpretability).

Implicationsof“theoryblind”classical-quantumcausalinference?

References

Acknowledgements:

Causalparents?Hardy,Brukner,Costa,Oreshkov,Spekkens,Liefer,Chiribella,Tucci,Ried,Cavalcanti,Lal,Henson,Pusey,Chaves,Pienaar,Giarmatzi…..+manymore(Lloyd)Causalsisters?Reidetal.,Allenetal.,Causalchildren?Schmid,Pienaar....+more

Books:“Elementsofcausaldiscovery”,Petersetal.,(2018);“DeepLearning”Goodfellowetal.,(2016)