Upload
others
View
7
Download
0
Embed Size (px)
Citation preview
1
ObjectiveValueIsAlwaysNewcombizable
ArifAhmedandJackSpencer
Words:5885
Words,includingfootnotesandappendices:9060
Thisessayarguesthatevidentialdecisiontheory(EDT)isincompatiblewithoptionshaving
objectivevalues.
After somescene-setting (§§1-2),weconsider threearguments forour thesis: the
argument fromNewcomb’sProblem(§§3-4), theargument fromExpectationism(§§5-6),
andtheargumentfromNewcombizability(§7).Thefirsttwoargumentsfailforinstructive
reasons.But the third succeeds.EDT is inconsistentwithoptionshavingobjectivevalues
becauseobjectivevalueisalwaysNewcombizable.
§1/ObjectiveandSubjective
Webeginbysupposing:
TwoOughts.Thereisbothanobjectiveoughtandasubjectiveought.
2
OnecandenyTwoOughts—Mackie(1977),forinstance,rejectstheobjectiveought,and
Moore(1903)andThomson(2008)rejectthesubjectiveought.1Butcaseslikethefollowing
convincemanyphilosopherstoacceptit:2
Miners.Tenminersaretrapped inshaftAorshaftB,but theagentdoesnotknow
which.Floodwatersthreatentofloodtheshafts.Theagenthasenoughsandbagsto
blockoneshaft,butnotboth.Iftheagentblocksoneshaft,allthewaterwillgointo
theother,killinganyminersinside.Iftheagentblocksneithershaft,bothshaftswill
fillhalfwaywithwater,andjustoneminer,thelowestintheshaft,willbekilled.(In
fact,theminersaretrappedinshaftA.)
With regard to Miners, one might ask: what ought the agent to do? And is the agent’s
uncertaintyrelevant towhatsheought todo?IfTwoOughtsis true, thesequestions lack
univocalanswers.
Intheobjectivesense,theagent’suncertaintyisirrelevant.Howeverconfidentsheis
thattheminersareinshaftB,sheobjectivelyoughttoblockshaftA.Inthesubjectivesense,
theuncertaintyisparamount.Iftheagentis5%confidentthattheminersareinshaftAand
95%confidentthattheyareinshaftB,thenshesubjectivelyoughttoblockshaftB.If,aswe
supposehenceforward,theagentis50%confidentthattheminersareinshaftAand50%
1 Also seeKolodnyandMacFarlane2010,which argues for onlyoneought that is neitherpurely
objectivenorpurelysubjective. 2Parfit(unpublished);cf.Jackson1991andRegan1980.ForadifferentmotivationofTwoOughts,
seeOddieandMenzies1992.
3
confidentthattheyareinshaftB,shesubjectivelyoughttoblockneithershaft—despite
knowingforcertainthatblockingneithershaftisnotwhatsheobjectivelyoughttodo.
Given the distinction between objective and subjective ought, we can saywhat it
wouldbeforoptionstohaveobjectiveandsubjectivevalues.
Saying that options have objective values is equivalent to saying that some
consequentializabletheoryoftheobjectiveoughtistrue.Inotherwords,it’stosaythatthere
is a numerically representable propertyof options such that,whenever an agent faces a
decision,theoptionsthatareobjectivelypermissibleforhertochooseareexactlytheoptions
thatmaximizetheproperty.
Somephilosophersconjecturethatevery(plausible)theoryoftheobjectiveoughtis
consequentializable.3Althoughwewillnotassumethat thisconjecture istrue, if it is,our
argumentsestablishanevenmoresurprisingmetaethicalthesis—thatEDTisincompatible
withevery(plausible)theoryoftheobjectiveought.
Saying that options have subjective value is equivalent to saying that some
consequentializabletheoryofthesubjectiveoughtistrue.Likeobjectivevalues,subjective
values are numerically representable properties of options. But unlike objective values,
subjectivevaluesarehadonlyrelativetoacredencefunction.Sotoclaimthatoptionshave
subjective values is to claim that there is some numerically representable property of
optionsrelativetoacredence functionsuchthat,wheneveranagent facesadecision, the
optionsthataresubjectivelypermissiblefortheagenttochooseareexactlytheoptionsthat
maximizethepropertyrelativetotheagent’scredencefunction.
3Seee.g.Dreier1993;2011,Louise2004,Portmore2007.
4
TounderstandwhyEDTisinconsistentwithoptionshavingobjectivevalues,itwill
behelpfultocontrastitwithitschiefrival,causaldecisiontheory(CDT).
§2/EDTandCDT
EDTandCDTarebothconsequentializabletheoriesofthesubjectiveought.Buttheydisagree
overwhatsubjectivevalueis.
According to EDT, subjective value is evidential expected value. To characterize
evidentialexpectedvalue,weneedsomefamiliarformalism.
Let𝑊 = {𝑤%,𝑤', … ,𝑤)}bethesetofpossibleworlds.Weassumeforsimplicitythat
𝑊isfinite.Andtoeasetheformalism,weignorethedistinctionbetweenapossibleworld
anditssingleton.
Let𝐴 = {𝑎%, 𝑎', … , 𝑎-}bethesetoftheagent’soptions,eachofwhichisaproposition
(i.e. a setofpossibleworlds) that theagent canmake truebydeciding.4Weassume that
optionsarealwayspair-wiseexclusiveandjointlyexhaustive.
Let 𝐶 be the agent’s credence function, a probability function mapping each
proposition to a real number in the unit interval and thereby representing the agent’s
confidencethatthepropositionistrue.Let𝐶(𝑝)betheagent’scredencein𝑝.Andintheusual
way,let𝐶(𝑝|𝑞) = 𝐶(𝑝𝑞)/𝐶(𝑞)betheagent’sconditionalcredencein𝑝given𝑞.
Finally,let𝑢betheworld-valuationfunctionmappingeachpossibleworldtoareal
numberandtherebyrepresentingthevalueofthatworld.Valuecomesinmanyflavors,and
4FollowingJeffrey1983:83-4.
5
forourpurposesanyflavorwilldo.Ifweareinterestedinmorality,wemaytake𝑢(𝑤)to
representhowmorallygoodwis.Ifweareinterestedinprudentialvalue,wemaytake𝑢(𝑤)
torepresenthowprudentiallygood(i.e.desirable)w is;andsoon.OurclaimthatEDTis
incompatiblewithoptionshavingobjectivevaluesholdsforanyflavorofvaluethatmight
belongtoindividualworlds.
With this formalism in place we can characterize evidential expected value. The
evidentialexpectedvalueof𝑎,written𝑉(𝑎),is:
EvidentialExpectedValue:𝑉(𝑎) = ∑ 𝐶(𝑤|𝑎)𝑢(𝑤)8 .5
According to CDT, subjective value is causal expected value. The simplest way to
characterizecausalexpectedvalue,which,forconvenience,weadopthere,involvesahighly
discerning causal similaritymetric.6 Thatmeans: ameasure of distance betweenpossible
worlds(asimilaritymetric)suchthat: (i) it ishighlydiscerning: foranyoption𝑎andany
world𝑤9 ,some𝑎-world𝑤:is,bythelightsofthismeasure,uniquelyclosestto𝑤97;(ii)itis
5Forsimplicity,thischaracterizationofevidentialexpectedvalueinvolvescredencesinindividual
worlds, thus departing from the technical assumption in standard axiomatizations of EDT that
preferenceisdefinedoveranatomlessfield(Jeffrey1983:146).Thisisharmless,sincenothingthat
wesaydependsontherepresentationtheoremforEDTthatrequiresatomlessnessforitsproof;in
anycase,wecoulddowithoutatomsaslongasvalueisnot‘gunky’(asHájek(2013:438)putsit).6Fordiscussionofothercharacterizations,seee.g.Ahmed2014,ch.2,Joyce1999ch.5,Lewis1981.7Nonaturalsimilaritymetricisquitethisdiscerning:fordiscussionseeLewis1973andJoyce1999
sect.5.4.Amorerealisticproposalisthat,foranyoption𝑎andworld𝑤9 ,thereissomesetof𝑎-worlds,
{𝑤:, 𝑤;, … ,𝑤)}thatare,bythelightsofthecausalsimilaritymetric,moresimilarto𝑤9 thanareany
other𝑎-worlds.Nothingexcepteaseofexpositionisgoingtoturnonthis.
6
causal:itreckonstheclosest𝑎-worldto𝑤9 tomatch𝑤9 onallparticularmattersoffactto
which𝑎iscausallyirrelevant.
Giventhishighlydiscerningcausalsimilaritymetric,wecangivetruth-conditionsfor
nonbacktracking counterfactual conditionals. If ⟨𝑎 ⇒ 𝑤:⟩ is the nonbacktracking
counterfactual—thatworld𝑤:wouldhavebeenactualhadoption𝑎beenchosen—then
⟨𝑎 ⇒ 𝑤:⟩istrueatsomepossibleworld𝑤9 ifandonlyif𝑤:isthe𝑎-worldthatis,bythelights
ofthehighlydiscerningcausalsimilaritymetric,uniquelymostsimilarto𝑤9 .
The causal expected value of option𝑎, written𝑈(𝑎), is a function of the agent’s
credencesinthesenonbacktrackingcounterfactuals:
CausalExpectedValue:𝑈(𝑎) = ∑ 𝐶(⟨𝑎 ⇒ 𝑤⟩)𝑢(𝑤)8 .
§3/TheArgumentfromNewcomb’sProblem
ThefirstpurportedargumentforourthesisistheargumentfromNewcomb’sProblem.EDT
and CDT often agree on which option maximizes subjective value, but not always. One
famouscruxis:
Newcomb’sProblem.Thereisatransparentboxandanopaquebox.Theagenthastwo
options:she can takeonly theopaquebox (she can ‘one-box’), or takebothboxes
(‘two-box’).The transparentbox contains$1,000.Theopaquebox containseither
$1,000,000or$0,dependingonapredictionmadeyesterdaybyareliablepredictor.
Ifthepredictorpredictedthattheagentwouldtwo-box,thentheopaqueboxcontains
7
$0. If the predictor predicted that the agentwould one-box, then the opaque box
contains$1,000,000.Theagentknowsallofthis.
The agent reasonably takes her choice to be strong evidence about whether the
opaqueboxcontains$1,000,000or$0.Equatingunitsofvalueanddollars,the𝑉-scoreof
one-boxingisthereforeslightlyunder1,000,000,andthatoftwo-boxingslightlyover1,000.8
Buttwo-boxinguniquelymaximizes𝑈.Interpretingthefollowingcounterfactualsin
termsofacausalsimilaritymetric,theagent’scredencethatshewouldhavereceived$1,000
hadshetwo-boxedexactlyequalshercredencethatshewouldhavereceived$0hadsheone-
boxed, and her credence that she would have received $1,001,000 had she two-boxed
exactlyequalshercredencethatshewouldhavereceived$1,000,000hadsheone-boxed.
The𝑈-scoreof two-boxing is thereforeexactly1,000greater than thatofone-boxing,no
matterwhatthe𝑈-scoreofone-boxingis.
TheargumentfromNewcomb’sProblemreferstoNewcomb’sProblemandhastwo
premises:
Newcomb Knowledge: If options have objective values then: an agent facing
Newcomb’sProblemknowsforcertainthattwo-boxinguniquelymaximizesobjective
value.
8E.g.iftheagentexpectsthepredictortobeaccurateon99%oftrialswhicheveroptionischosen,
then the V-score of one-boxing is (0)(0.01) + (0.99)(1,000,000) = 990,000; and that of two-
boxingis(1,000)(0.99) +(0.01)(1,001,000) = 11,000.
8
Bridge:Ifoptionshaveobjectivevaluesthen:ifanagent’soptionsare𝑎%, 𝑎',… , 𝑎- ,
andtheagentknowsforcertainthatoption𝑎9 uniquelymaximizesobjectivevalue,
thenoption𝑎9 uniquelymaximizessubjectivevalue.
NewcombKnowledgeismotivatedbytheclaimthatobjectivevalueiscausalvalue.
Thecausalvalueofanoptionisthevalueoftheworldthatwouldhavebeenactualhadthe
agentchosentheoption.Thatis:if⟨𝑎 ⇒ 𝑤:⟩istrueatworld𝑤9 ,thenthecausalvalueof𝑎at
𝑤9 is𝑢(𝑤:).Causalvalueisafamiliarconceptionofobjectivevalue.Itfeatures,forexample,
inMoore’sdefenseofconsequentialism,wherehewrites:
Inordertoshowthatanyaction[maximizesobjectivevalue],itisnecessarytoknow
both what are the other conditions, which will, conjointly with it, determine its
effects;toknowexactlywhattheeffectsoftheseconditionswillbe;andtoknowall
theeventswhichwillbe inanywayaffectedbyouractionsthroughoutan infinite
future.Wemusthaveall thiscausalknowledge….Andnotonlythis:wemustalso
possessallthisknowledgewithregardtotheeffectsofeverypossiblealternative;and
mustthenbeabletoseebycomparisonthatthetotalvalueduetotheexistenceofthe
actioninquestionwillbegreaterthanthatwhichwouldbeproducedbyanyofthese
alternatives.(1903:149)
ItisuncontroversialthatanagentfacingNewcomb’sProblemknowsforcertainthat
two-boxinguniquelymaximizescausalvalue.So,ifobjectivevalueiscausalvalue,Newcomb
Knowledgeistrue.
9
Bridgeismotivatedbymetaethicalconsiderations.Thesubjectiveoughtarisesfrom
theagent’s subjectiveuncertainty about theobjectiveought.When anagent isuncertain
what she objectively ought to do, there can be a discrepancy between what the agent
subjectively and objectively ought to do. Indeed, asMiners illustrates, when an agent is
uncertainaboutsheobjectivelyoughttodo,anagentcanknowforcertainthattheydiffer.
Butifanagentknowsforcertainwhatsheobjectivelyoughttodo,thentheobjectiveand
subjectiveoughtmust coincide.9TwoOughts thusentails the following conditional: if an
agentknowsforcertainthatsheobjectivelyoughttochooseoption𝑎9 ,shesubjectivelyought
tochooseoption𝑎9 .10Andgiventhatoptionshavebothobjectiveandsubjectivevalues,this
conditionalisequivalenttotheconsequentofBridge.11
9Zimmermandefendsthestrongerclaimthat‘[a]nagent[subjectively]oughttoperformanactifand
only ifhebelieves that it is [objectively] thebest option thathehas’ (2008:5).WethinkMiners
refutestheleft-to-rightdirectionofthisclaim;buttheright-to-leftdirectionisplausibleandentails
Bridge.
10OneapparentchallengetoBridgeinvolvesupwardmonotonicity.TheagentinMinersobjectively
oughttoblockshaftA.Iftheobjectiveoughtisupwardmonotonic,thentheagentinMinerswill(if
rational)knowforcertainthatsheobjectivelyoughttoblocksomeshaftorother.Butitisnotthe
casethattheagentsubjectivelyoughttoblocksomeshaftorother.Inresponseonemightreasonably
denyupwardmonotonicity;butinanycase,theexampledoesnotthreatenBridgeasstated,because
Bridgeonlycoversanagent’soptions,andsinceoptionsarepair-wiseexclusive,blockingsomeshaft
orothercannotbeanoptionifblockingshaftAisanoption.11Bridgemustbedistinguishedfromnearbyprinciples.
Onenearbyprincipleconcernspermissibility.Itsays:ifanagentknowsforcertainthatan
option is objectively permissible, then the option is subjectively permissible. This principle is
plausible, but less plausible than Bridge. Opaque sweetening cases, discussed inHare 2010, put
pressureonthispermissibilityprinciple,butputnopressureonBridge.
10
ItmightappearthattheargumentfromNewcomb’sProblemestablishesourthesis
— that Nozick, way back in 1969, showed that EDT is inconsistentwith options having
objectivevalues.ButtheargumentfromNewcomb’sProblemcanberesisted.
§4/ManyConceptionsofObjectiveValue
Bridgeisundeniable,butNewcombKnowledgeisnot.AnagentfacingNewcomb’sProblem
knowsforcertainthattwo-boxinguniquelymaximizescausalvalue.ButwhyshouldEDT’ists
grantthatobjectivevalueiscausalvalue?Afterall,therearemanyconceptionsofobjective
Another nearby principle concerns inevitable knowledge. It says: if an agent knows for
certainthatshewouldbeobjectivelyrequiredtochooseaparticularoptionifsheknewthatpwas
trueandalsoknowsforcertainthatshewouldbeobjectivelyrequiredtochoosethatsameoptionif
sheknewthatpwasfalse,thentheagentissubjectivelyrequiredtochoosetheoption.Thisinevitable
knowledge principle is clearly false. It might be the case that an agent, who is uncertain about
whetherp,objectivelyoughttopayasmallsumofmoneytocometolearnthetruth-valueofp,even
thoughtheagentknowsforcertainthatshewouldbeobjectivelyrequirednottopaythesmallsum
ifsheknewthatpwastrueorknewthatpwasfalse.
ThedifferencebetweenBridgeandtheprincipleconcerninginevitableknowledgeisrelevant
tothediscussioninHare2016.Haremakestwoclaims:(1)thatkillingonetosavefiveissubjectively
permissiblewhenitisunknownwhichofthesixisbeingsacrificedtosavetheotherfive,and(2)that
killingonetosavefiveissubjectivelyimpermissibleiftheidentityoftheonebeingsacrificedtosave
thefivewasknown.Itisnotobvioustousthatbothclaimsaretrue.Buteveniftheyare,theypose
nothreattoBridge.Weclaimthatif(1)istrue,thenitisalsoobjectivelypermissibletokilloneto
savefivewhentheidentityoftheonebeingsacrificedtosavethefiveisunknown.Andweclaimthat
if(2)istrue,thenitisalsoobjectivelyimpermissibletokillonetosavefivewhentheidentityofthe
onebeingsacrificed tosave the five isknown.Hare’sexamplemightbeacounterexample to the
inevitableknowledgeprinciple,butitisnotacounterexampletoBridge.
11
value(whereby‘conceptionofobjectivevalue’wemeanaproposalaboutwhatobjective
valueis).
Anaturalconceptionofobjectivevaluehastwocomponents.Thefirstisasimilarity
metric.Theinputtoasimilaritymetricissomeoption𝑎andsomeworld𝑤9 ,andtheoutput
issomesetF𝑤:,𝑤;,… , 𝑤)Gof𝑎-worldsthatare,onthatmetric,themostsimilara-worldsto
𝑤9 .Thesecondcomponentisamethodofaveraging.Theobjectivevalueofoption𝑎atworld
𝑤9 issomeweightedaverageof𝑢H𝑤:I, 𝑢(𝑤;), … , 𝑢(𝑤)),thevaluesofthe𝑎-worldsthatare,
bythelightsofthesimilaritymetric,mostsimilarto𝑤9 .Ifasimilaritymetricalwaysoutputs
asetcontainingonlyonepossibleworld,wecandisregardthemethodofaveraging,forthe
objectivevalueofanoption𝑎atworld𝑤9 isthenjustthevalueofthe𝑎-worldthatis,bythe
lightsofthesimilaritymetric,uniquelymostsimilarto𝑤9 .Causalvalueisonememberofthis
familyofnaturalconceptionsofobjectivevalue,buttherearemanyothers.Andmanyare
uniquelymaximizedbyone-boxinginNewcomb’sProblem.12
Forinstance,considerwhatHorgansays,inhisdefenseofone-boxing:
I shall assume that there is indeed a standard resolution of the vagueness of the
similarityrelationamongworlds,andthatLewis’saccountofitisessentiallycorrect.
Returning to Newcomb’s problem, it is clear that [the] premises…of the one-box
12Thefullclassofobjectivevalueconceptsextendswellbeyondthisnaturalfamily.Initsbroadest
sense, an objective value concept is any function taking each pair of a possible world and a
propositiontoarealnumber.Althoughwefinditmorenaturalandintuitivetofocusontherestricted
family that themain textdescribes,ourresultextends toanyobjectivevalueconceptwithin this
broaderclass.WeclaimthatEDTrulesoutidentifyingobjectivevaluewithanyfunctionwithinthat
broaderclass.
12
argument cannot…be true under the standard resolution. For the beingmade his
predictionaboutmychoice,andhaseitherputthe$1millioninthe[opaquebox]or
not,wellbeforeIchoose.Thus,hisactual-worldpredictionandtheactual-worldstate
of[theopaquebox]remainintactintheclosestworldinwhichItakebothboxes,and
alsointheclosestworldinwhichItake[theopaquebox]only.
[… The] intuitive plausibility of the one-box argument rests upon a
nonstandardresolution,onethatseemsquiteappropriate in thiscontext. Itdiffers
fromthestandardresolutiontotheextentthatitgivestopprioritytomaintainingthe
being’spredictivecorrectnessinthenearestpossibleworldwhereItakebothboxes,
and also in the nearest world where I take [only the opaque box]. Under this
backtrackingresolution…theclosestworldinwhichItakebothboxesisoneinwhich
thebeingcorrectlypredictedthisandputnothingin[theopaquebox],andtheclosest
worldinwhichItakeonly[theopaquebox]isoneinwhichhecorrectlypredictedthis
andput$1millionin[it].(1981:336)
AlthoughHorgandoesnotgivedetails,wecanspelloutasimilaritymetriconpossible
worldsthatsatisfieshisdesiderata.OnewaytodoitwouldbetograftHorgan’s‘toppriority’
onto something like Lewis’s (1979) lexicographic criteria for measuring the relative
similarityofworlds,insistingthatwhatcountsmostforclosenessofapossibleworld𝑤in
this context is whether the predictor’s accuracy at𝑤 with regard to the agent’s choice
13
matchesthepredictor’sactualaccuracyonthatquestion.13WethenusetheHorganmetric
todefineaconceptionofobjectivevalue,whichwemightcall:
Horganvalue:if𝑤:isthe𝑎-worldthatismostsimilarto𝑤9 bytheHorganmetric,
thentheHorganvalueofaat𝑤9 is𝑢(𝑤9).
13Moreformally,defineameasureoftherelativesimilarityofworlds𝑤and𝑤′toafixedworld𝑤9
byappealtofivepartialordersonworlds:
1) 𝑤 >% 𝑤′ iff: the Newcomb predictor’s correctness on this occasion at𝑤 matches the
predictor’saccuracyat𝑤9 ,butthepredictor’scorrectnessat𝑤′doesnot.
2) 𝑤 >' 𝑤′iff:therearebig,widespreadanddiverseviolationsofthelawsof𝑤9 at𝑤′and
notat𝑤.
3) 𝑤 >L 𝑤′iff:thespatio-temporalregionthroughoutwhichperfectmatchoverparticular
factswith𝑤9 prevailsislargerat𝑤thanat𝑤′.
4) 𝑤 >M 𝑤′iff:therearesmall,localized,simpleviolationsofthelawsof𝑤9 at𝑤′andnotat
𝑤.
5) 𝑤 >N 𝑤′iff:𝑤achievesapproximatesimilarityto𝑤9 overmattersofparticularfactand
𝑤′doesnot.
Say that 𝑤: is Horgan-closer to 𝑤9 than is 𝑤; if and only if: either (i) F𝑛P𝑤; >- 𝑤:G = ∅ and
F𝑛P𝑤: >- 𝑤;G ≠ ∅;or(ii)minF𝑛P𝑤; >- 𝑤:G > minF𝑛P𝑤: >- 𝑤;G.
Criteria2)-5)arefromLewis(1979:47-8),minusLewis’shedgingover5).Takenjointlyas
ananalysisofthe‘standardresolution’ofclosenessinnaturallanguage,2)-5)areimplausible(see
e.g.McDermott1999forcriticism).Ofcoursewearenotmakinganyclaimsaboutnaturallanguage
counterfactualsbutratherusingLewis’scriteriatoconstructakindofobjectivevaluethatone-boxing
maximizes.
14
Often,theHorganvalueofanoptionisequaltoitscausalvalue.Forinstance,equating
livesandunitsof value, inMiners, the causalvalueofblocking shaftA (B) is10 (0), and
likewisetheHorganvalueofblockingshaftA(B)is10(0).
But in Newcomb’s Problem, the two diverge. The causal metric holds fixed the
contentsoftheopaqueboxbutnotthepredictor’scorrectness.Whatevermoneyisinthe
opaqueboxatthetwo-boxing-worldthatiscausallyclosesttotheactualworldisalsointhe
opaqueboxattheone-boxing-worldthatiscausallyclosest,sothecausalvalueoftwo-boxing
exceedsthatofone-boxingby1,000.TheHorgansimilaritymetricholdsfixedthepredictor’s
correctnessbutnotthecontentsoftheopaquebox.So,ifthepredictorisactuallycorrect,the
two-boxing-worldthatisHorgan-closestisonewheretheopaqueboxcontains$0,andthe
one-boxing-worldthatisHorgan-closestisonewheretheopaqueboxcontains$1,000,000.
Ifthepredictorisactuallyincorrect,thesituationisreversed—thetwo-boxing-worldthat
isHorgan-closest isonewhere theopaquebox contains$1,000,000, and theone-boxing-
worldthat isHorgan-closest isonewheretheopaqueboxcontains$0. Ineithercase, the
Horganvalueofsomeoptiondivergesfromitscausalvalue.
If objective value is Horgan value, Newcomb Knowledge is false. An agent facing
Newcomb’sProblemdoesnotknowforcertainthattwo-boxinguniquelymaximizesHorgan
value. After all, she is confident that the predictor is accurate on thisoccasion, so she is
confident thatone-boxinguniquelymaximizesHorganvalue. Indeed, if objectivevalue is
Horganvalueandthepredictorisknowntobesufficientlyreliable,theagentmightknowfor
certainthatone-boxinguniquelymaximizesobjectivevalue.AndHorganvalueisnotunique
15
in this respect. Uncountably many conceptions of objective value can be known to be
uniquelymaximizedbyone-boxinginNewcomb’sProblem.14
The argument from Newcomb’s Problem is therefore too quick. The argument
establishessomething.SinceBridgeistrue,theargumentestablishesthatEDTisinconsistent
with any conception of objective value that validates Newcomb Knowledge. But most
conceptionsofobjectivevalueinvalidateNewcombKnowledge.Withoutsomeindependent
argument that the true conceptionofobjectivevaluevalidatesNewcombKnowledge, the
argumentfromNewcomb’sProblemfailstoestablishourthesis.
§5/TheArgumentfromExpectationism
ThesecondpurportedargumentforourthesisistheargumentfromExpectationism.
Expectationismisathesisabouthowobjectivevalueandsubjectivevaluerelate.It
saysthatsubjectivevalueisexpectedobjectivevalue.Let⟨𝑂(𝑎) = 𝑣⟩bethepropositionthat
(onsomearbitrarilychosenscale)theobjectivevalueofanoption𝑎 is𝑣 ∈ ℝ.Then,more
formally,wehave:
14Thusconsider theviewonwhichtheobjectivevalueofanoption𝑎 isgivenbytheconditional
chancesatsometimebeforethedecision.Supposethepredictionisat𝑡% ,andconsidertheconditional
chances taggedtosomeearlier𝑡X .Wecan formulateasimilaritymetric thatmakestheclosesta-
worldsto𝑤9 bethoseinthesetofworlds{𝑤:, 𝑤;, … ,𝑤)}thathavepositivechanceattime𝑡Xat𝑤9 ,
conditionalon𝑎.Tocalculatetheobjectivevalue,averagethevaluesoftheoutputtedworldsbytheir
respectivechancesconditionalon𝑎.Thus,accordingtotheproposal,theobjectivevalueof𝑎atworld
𝑤9 equals∑ 𝐶ℎZ[,\](𝑤|𝑎)𝑢(𝑤)8 .Ifthepredictorhasa99%chanceofcorrectnessbackat𝑡X ,thenthe
objectivevalueofone-boxingis(0)(0.01) +(1,000,000)(0.99) = 990,000,andtheobjectivevalue
oftwo-boxingis(1,000)(0.99) +(0.01)(1,001,000) = 11,000.
16
Expectationism:Foranycredencefunction𝐶,thesubjectivevalueofarelativeto𝐶
is∑ 𝑣𝐶(⟨𝑂(𝑎) = 𝑣⟩)^ .
Expectationism is widely accepted. Often it’s just assumed,15 but there are some
explicitdefensesofit.Forexample,onemightdefendExpectationismbyclaimingthatthe
subjectivevalueofanoptionshouldbetheagent’sbestestimateofitsobjectivevalue,and
thenarguingthattheagent’sbestestimateofaquantityistheirexpectationofit.16
Inthedisputebetweenone-boxersandtwo-boxers,Expectationismisneutral.Ifwe
combineExpectationismwiththeclaimthatobjectivevalueiscausalvalue,wegetCDT,since
thecausalexpectedvalueofanoptionistheagent’sexpectationofthecausalvalueofthe
option.17 But we can combine Expectationism with conceptions of objective value that
invalidateNewcombKnowledge.Forexample,ifwecombineExpectationismwiththeclaim
thatobjectivevalueisHorganvalue,thenwegetHorgandecisiontheory(HDT)—theview
thatthesubjectivevalueofanoptionistheagent’sexpectationoftheHorganvalueofthe
option.Whereastwo-boxinguniquelymaximizesexpectedcausalvalue(i.e.causalexpected
value),one-boxinguniquelymaximizesexpectedHorganvalue.
15Seee.g.Parfit1984:25.16ForadefenseofExpectationismalongtheselines,seee.g.OddieandMenzies1994andPettigrew
2015.17Proof:Let𝑂Z(𝑎)bethecausalvalueofoption𝑎atworld𝑤.Theexpectationofthecausalvalueof
𝑎relativetocredencefunctionCis∑ 𝐶(𝑤)𝑂Z(𝑎)8 .LetWibetheworldsatwhich⟨𝑎 ⇒ 𝑤9⟩istrue.
Then ∑ 𝐶(𝑤)𝑂Z(𝑎)8 = ∑ 𝐶(𝑤)𝑢(𝑤%) +⋯+ ∑ 𝐶(𝑤)𝑢(𝑤-)8̀8a = 𝐶(⟨𝑎 ⇒ 𝑤%⟩)𝑢(𝑤%) + ⋯+
𝐶(⟨𝑎 ⇒ 𝑤-⟩)𝑢(𝑤-) =∑ 𝐶(⟨𝑎 ⇒ 𝑤⟩)𝑢(𝑤) = 𝑈(𝑎)8 .
17
ButExpectationismishostiletoEDT,aswenowargue.
Everyremotelyplausibleconceptionofobjectivevaluemustallowanagenttoregard
aasevidenceaboutwhattheobjectivevalueofais.Takethesimplestcase,inwhichanagent
isuncertainwhethertheobjectivevalueofoption𝑎 is𝑣%or𝑣'.Everyremotelyplausible
conceptionofobjectivevaluemustvalidate:
Relevance:Itispossiblethatanagent’scredencesbesuchthat,forsome𝑣% ≠ 𝑣':
i. 𝐶(⟨𝑂(𝑎) = 𝑣%⟩ ∨ ⟨𝑂(𝑎) = 𝑣'⟩) = 1,
ii. 𝐶(⟨𝑂(𝑎) = 𝑣%⟩) = 𝑥 < 1,and
iii. 𝐶(⟨𝑂(𝑎) = 𝑣%⟩|𝑎) = 𝑦 ≠ 𝑥.
Relevancemustholdbecausethefactthatanagentregards𝑎asevidenceastowhether𝑝
cannotprecludetheobjectivevalueof𝑎fromdependingonwhether𝑝.
Tosee thismore concretely, consideravariationonMiners.On thisvariation, the
agentrememberstheminers’tellingherwhichshafttheywouldbeworkingin,butcannot
consciouslyrecallwhich.Asbefore,sheis50%confidentthattheminersareinshaftAand
50%confidentthattheyareinshaftB.Butshe(reasonably)thinksthatthereisanonzero
chancethatherunconsciousmemorywillinfluenceherchoiceifshechoosestoblockoneof
the shafts. Therefore, her confidence that the miners are in shaft A, conditional on her
blockingshaftA,is(say)52%,upfrom50%.
Sinceclauses(i)and(ii)ofRelevanceclearlyhold in thiscase,anyonewhodenies
RelevancewouldhavetoholdthattheobjectivevalueofblockingshaftAdoesnotdepend
onwheretheminersare.Theywouldhavetoholdthattheobjectivevalueofblockingshaft
18
AataworldatwhichtheminersareinshaftAisequaltotheobjectivevalueofblockingshaft
AataworldatwhichtheminersareinshaftB.Butthat’sabsurd.Onanyremotelyplausible
conceptionofobjectivevalue,theobjectivevalueofblockingshaftAdependsonwherethe
minersactuallyare,eveniftheagentregardsblockingshaftAasevidenceaboutwherethe
minersare.Thus,Relevancemusthold.18
Nowwecanprove:
Result#1.Relevance,Expectationism,andEDTarejointlyinconsistent.
The(verysimple)proofisinAppendixA.Informally,theideaisthatExpectationismmakes
thesubjectivevalueofanoptiondependonlyontheagent’sunconditionalcredencesinits
havingthisorthatobjectivevalue,whereasEDTmakesitturnonitsexpectedobjectivevalue
conditionalonitsperformance.RelevancethereforeimpliesthatEDTandExpectationismcan
disagree about the subjective value of an option. For instance, in the variant onMiners,
18Insayingthis,weareeffectivelysettingaside‘indexical’valueconceptsofthesortthatHájekand
Pettit(2004)discussinconnectionwithLewis’s(1988,1996)argumentsagainst‘DesireasBelief’
(DAB).Indexicalvaluesdependnotonlyonthestateofthemind-independentworldbutalsoonthe
beliefs of the agent herself.We agreewithHájek and Pettit that indexical value concepts evade
Lewis’sarguments.TheyalsoviolateRelevance.(Forinstance,iftheindexicalvalueofanoption𝑎is
just𝑉(𝑎),then,solongasconditionalizationcanchangetheevidentialexpectedvalueofthetautology
(seeBradleyandStefánsson2016:699-702),Relevanceasappliedtoindexicalvaluefails.)Butwe
denythatindexicalvalueisobjective.Thenotionofobjectivevaluethatinterestsusissuchthat,if
optionshaveobjectivevalues, theobjectivevalueof (say)blockingshaftA inMinersdependson
wheretheminersactuallyare,regardlessofwhattheagentthinks.(FormoreonDABseen.23.)
19
Expectationism implies that the subjective value of blocking shaft A is 5, whereas EDT
reckonsitat5.2.
TheargumentfromExpectationismexploitsResult#1.ItsaysthatEDTisinconsistent
withoptionshavingobjectivevaluebecause,ifoptionshaveobjectivevalues,Relevanceand
Expectationismarebothtrue.
The argument from Newcomb’s Problemwas too narrow. It showed that EDT is
inconsistentwithanyconceptionofobjectivevaluethatvalidatesNewcombKnowledge,but
was silentabout conceptionsofobjective thatdonotvalidateNewcombKnowledge.The
argumentfromExpectationism,bycontrast,purportstoestablishthatEDTisinconsistent
withevery(remotelyplausible)conceptionofobjectivevalue,eventhose,likeHorganvalue,
that invalidate Newcomb Knowledge. It might seem, then, that the argument from
Expectationismestablishesourthesis—thataproperunderstandingofthemathematical
relationshipbetweensubjectivevalueandobjectivevaluerevealsthatEDTisinconsistent
withoptionshavingobjectivevalues.ButtheargumentfromExpectationismcanberesisted.
§6/ResistingExpectationism
EveryremotelyplausibleconceptionofobjectivevaluevalidatesRelevance,sotheargument
fromExpectationismestablishesthatEDTandExpectationismareinconsistent.Evidential
expectedvalueisnotexpectedobjectivevalue.CDT’istsandHDT’istsagreethatsubjective
valueisexpectedobjectivevalueanddisagreeaboutwhatobjectivevalueis,butEDT’istsdo
notshareinthisagreement.NoconceptionofobjectivevaluestandstoEDTascausalvalue
standstoCDT.
20
ButExpectationismcanbequestioned.Ifoptionshavebothobjectiveandsubjective
values,thentheremustbesomewell-behaved,intimaterelationshipbetweenthesubjective
valueofanoptionrelativetoanagent’scredencefunctionandtheagent’shypothesesabout
itsobjectivevalue.Butthisrelationshipneedn’tbeexpectation.
Aminimalnecessaryconditionfortherelationshipbetweenobjectiveandsubjective
valuebeingwell-behavedandintimateisthetruthofthefollowingprinciple:
Certain Reflection: For any credence function 𝐶, if 𝐶(⟨𝑂(𝑎) = 𝑣⟩) = 1, then the
subjectivevalueofoption𝑎relativeto𝐶equals𝑣.
ButCertainReflectionisstrictlyweakerthanExpectationism,aswecanseebyconsidering
someviewsthatverifyCertainReflectionwhilefalsifyingExpectationism.
FirstconsiderMaximin—theviewthatthesubjectivevalueof𝑎relativetoanagent’s
credencefunctionistheleast𝑣suchthattheagentassignsnonzerocredenceto⟨𝑂(𝑎) = 𝑣⟩.
MaximinsatisfiesCertainReflection,butitfalsifiesExpectationism.
Next,considerRisk-adjustedExpectationism—theviewthatthesubjectivevalueof𝑎
isnotthestraightexpectationofobjectivevaluebutratheraweightedsumofitspossible
objectivevaluesthatattachesmoreimportancetosomeofthesepossibilitiesthantoothers
dependingontheirrankandnotonlyontheirprobability: for instance, itweightsworse
possibilities more heavily than better ones, other things being equal.19 On most such
weightings, the view that subjective value is risk-adjusted expectation ofobjective value
19Buchak2013ch.2.
21
falsifiesExpectationism,butagainitsatisfiesCertainReflection,becauseifyouarecertainof
whattheobjectivevalueofanoptionis,thentherearenoalternativehypothesesaboutthis
towhichyoucangivemoreorlessweightdependingontheirrank.20
Athirdalternative,whichisamenabletoEDT,takessubjectivevaluetobeconditional
expectedobjectivevalue.Wecallthis:
ConditionalExpectationism:Foranycredencefunction𝐶,thesubjectivevalueofa
relativeto𝐶is∑ 𝑣𝐶(⟨𝑂(𝑎) = 𝑣⟩|𝑎)^ .21
ConditionalExpectationismentailsCertainReflection. AndgivenCertainReflection,EDT
entailsConditionalExpectationism.22
20Wecanrealizethisideaformallybymeansofadistortioni.e.anon-decreasingfunction𝑟: [0,1] →
[0,1] such that 𝑟(0) = 0, 𝑟(1) = 1. For any option𝑎, arrange its epistemically possible objective
valuesinincreasingorder𝑥%, 𝑥', … , 𝑥-.ThecorrespondingversionofRisk-adjustedExpectationism
(RE) says that relative to a credence function 𝐶 the subjective value of 𝑎 is 𝜎(𝑎) = 𝑥% +
∑ (𝑥9l% − 𝑥9)𝑟H𝐶(⟨𝑂(𝑎) ≥ 𝑥9l%⟩)I-o%9p% .If𝑟istheidentityfunctioni.e.𝑟(𝑥) = 𝑥thensubjectivevalue
coincides with expected objective value; but if 𝑟 is convex to the 𝑥-axis, e.g. if 𝑟(𝑥) = 𝑥', then
subjective value weights worse possibilities more heavily than expectationism demands. But
trivially,itsatisfiesCertainReflection.21Oddie1994:460.AlsoseeBroome’s(1991)discussionofConditionalExpectationism(whichhe
somewhatmisleadinglycallsDesire-as-Expectation).22Proof:Let𝑤9 beanyworldatwhich𝑎and⟨𝑂(𝑎) = 𝑣⟩arebothtrue,andlet𝐶concentrateallofits
credenceon𝑤9 .Then,byCertainReflection,thesubjectivevalueof𝑎relativeto𝐶is𝑣.And,byEDT,
thesubjectivevalueof𝑎relativeto𝐶is∑ 𝐶(𝑤|𝑎)𝑢(𝑤8 ) = 𝑢(𝑤9).So𝑢(𝑤9) = 𝑣.ThismeansthatEDT
entailsConditional Expectationism, since,∑ 𝐶(𝑤|𝑎)𝑢(𝑤) = ∑ 𝐶(𝑤|𝑎)𝑢(𝑤) +⋯ =Z∈⟨q(r)p^a⟩∩rZ∈8
∑ 𝐶(𝑤|𝑎)𝑣% + ⋯ = ∑ 𝑣𝐶(⟨𝑂(𝑎) = 𝑣⟩|𝑎)^Z∈⟨q(r)p^a⟩ .
22
There is something intuitiveaboutExpectationism,which says that the subjective
valueofanoptionshouldbetheagent’sestimateofitsobjectivevalue,andsomeonewho
acceptsConditionalExpectationismmustrejectExpectationism,since,givenRelevance,the
two claims cannot both be true. But there is also something intuitive about Conditional
Expectationism,which says that the subjective value of an option should be the agent’s
estimateofitsobjectivevalueinworldswhereitisrealized.Thetheoreticalcostofrejecting
ExpectationisminfavorofConditionalExpectationismthusseemstouslow.Andthereisno
argumentfromConditionalExpectationism.WhereasRelevance,Expectationism,andEDT
arejointlyinconsistent,Relevance,ConditionalExpectationism,andEDTareconsistent.
TheassumptionthatEDT’istsmustacceptExpectationismisunjustified.(Evenamong
opponentsofEDT,Expectationismisnotcommonground.)So,withoutsomeindependent
argumentthatEDT’istsmustacceptExpectationism,theargumentfromExpectationismfails
toestablishourthesis.23
23WeshouldbrieflyrelatethepresentdiscussiontoLewis’s(1986,1988)argumentagainsttheanti-
Humean‘DesireasBelief’(DAB)thesis,towhichtheargumentfromExpectationismbearsanobvious
resemblance.ThebasicideabehindLewis’sargumentisthatifevidentialexpectedvalue𝑉measures
theagent’sdesire for the truthofaproposition, thenDABsays that𝑉(𝐴) = 𝐶H�̇�I,where �̇� isthe
propositionthatitisgoodthat𝐴.(Lewis’sproofinvolvesanungradednotionofgoodnessbutcould
easily be extended to cover a graded notion, as the argument fromExpectationism does.) Given
Lewis’sassumptionthat𝑉(𝐴|𝐴) = 𝑉(𝐴) (whichhasbeenquestioned:seeBradleyandStefánsson
2016:699-702)itfollowsfromDABthat𝐶H�̇�P𝐴I = 𝐶H�̇�I;butthisisinconsistentwiththeanalogue
of Relevance that Lewis implicitly assumes (1996: 309). Oneway for the anti-Humean to resist
Lewis’sargument(Price1989:122)wouldbetoreformulatetheanti-Humeanthesisas‘Desireas
ConditionalBelief’(DACB),whichsaysthat𝑉(𝐴) = 𝐶H�̇�P𝐴I.ThisthesisgivesnotractiontoLewis’s
argument,forjustthesamereasonthatConditionalExpectationismgivesnonetotheargumentfrom
Expectationism. In response, Lewis argues (1996: 310-11) that DACB is a version of desire by
23
§7/TheArgumentfromNewcombizability
ThethirdpurportedargumentforourthesisistheargumentfromNewcombizability.
ItstartsfromaweakeningofBridge.Aswesaid,Bridgeseemsundeniable.Ifoptions
haveobjectivevaluesthen:ifanagent’soptionsare𝑎%, 𝑎', … , 𝑎-,andtheagentknowsfor
certain that option 𝑎9 uniquely maximizes objective value, then 𝑎9 uniquely maximizes
subjective value. But the subjective values of options relative to a credence function
supervene on the agent’s credences and theworld valuations.We therefore canweaken
Bridge, appealing to any of the agent’s certainties, and not just those that constitute
knowledge.Theresultis:
Dominance. If options have objective values then: if an agent’s options are
𝑎%, 𝑎', … , 𝑎- , and the agent is certain that option𝑎9 uniquelymaximizes objective
value,then𝑎9 uniquelymaximizessubjectivevalue.
necessity: it is committed to the existence of aproposition𝐺 which the agent desires to be true
whateverhercredences.WearenotclearwhythispointhasanyforceagainstDACB.Maybetheidea
isthattheanti-Humeanthesisisadescriptivepsychologicalthesisaboutaperson’sdesires,andthat
itissimplyfalseasamatteroffactthatthereisanythingthateveryonedesires(cf.Lewis1996:304-
5). However this may be, we are clear enough that no analogous point threatens Conditional
Expectationism. After all, the latter is a normative thesis, because of the connection between
subjectivevalueandwhatonesubjectivelyoughttodo.Evenifthereisnothingthateveryonedoes
value,theremightbesomethingthateveryoneshouldvalue.Inshort,wethink:(a)thatthesame
objectionarisesagainstbothLewis’sargumentandtheargumentfromExpectationism;and(b)that
eveniftheformersurvivesit,thelatterdoesnot.
24
DominanceislikeCertainReflection.CertainReflectionensuresthatsubjectivevalue
conformstoobjectivevaluenumerically—itsaysthat,ifanagentiscertainthattheobjective
valueofanoptionequals𝑣, then,relativeto theagent’scredence function, thesubjective
value of the option also equals𝑣. Dominance ensures that subjective value conforms to
objective valueordinally— it says that, if an agent is certain that some optionuniquely
maximizesobjectivevalue, then,relativeto theagent’scredence function, theoptionalso
uniquelymaximizessubjectivevalue.
Dominance is entailed bymany views about how objective and subjective values
relate, including Minimax, Risk-adjusted Expectationism, CDT, HDT, and any form of
Expectationism.24ButDominanceishostiletoEDT.ThereisageneralizationofRelevance—
24ProofthatExpectationismimpliesdominance:let𝑂Z(𝑎)betheobjectivevalueofoption𝑎atworld
w.If𝐶(⟨𝑂(𝑎) > 𝑂(𝑏)⟩) = 1andExpectationismistrue,then,relativeto𝐶,thedifferencebetweenthe
subjectivevalueofoption𝑎andthesubjectivevalueofoption𝑏equals∑ 𝐶(𝑤)(𝑂Z(𝑎) − 𝑂Z(𝑏))8 .
Sincetheagentiscertainthattheobjectivevalueof𝑎exceedsthatof𝑏,atanyworld𝑤towhich𝐶
assignsnonzerocredence,𝑂Z(𝑎) − 𝑂Z(𝑏)ispositive.Hence,∑ 𝐶(𝑤)(𝑂Z(𝑎) − 𝑂Z(𝑏))8 ispositive,
whichentailsthat,relativeto𝐶,thesubjectivevalueofoption𝑎exceedsthesubjectivevalueof𝑏.
Proofthatmaximinimpliesdominance:let𝑎 = min{𝑥|𝐶(⟨𝑂(𝑎) = 𝑥⟩) > 0}.If𝑎 ≤ 𝑏then𝐶H⟨𝑂(𝑎) ≤
𝑏⟩I > 0, so 𝐶(⟨𝑂(𝑎) ≤ 𝑂(𝑏)⟩ > 0. Contrapositively, 𝐶(⟨𝑂(𝑎) > 𝑂(𝑏)⟩) = 1 implies 𝑎 > 𝑏, so the
maximinsubjectivevalueof𝑎exceeds thatof𝑏.For theproof thatRisk-adjustedExpectationism
implies dominance, see Buchak 2013: 245-6. Note that this proof assumes that the distortion
function𝑟isstrictlyincreasing.Butevenifweassumeonlythat𝑟isnon-decreasingwecanstillprove
thefollowingweakeneddominanceprinciple:ifanagent’soptionsare𝑎%, 𝑎',… , 𝑎-,andtheagentis
certainthatoption𝑎9uniquelymaximizesobjectivevalue,thennootheroptionuniquelymaximizes
subjective value. The argument from Newcombizability still goes through on this weakening of
Dominance.
25
a principle that is strictly stronger than Relevance, but no less obviously true — that
contradictstheconjunctionofEDTandDominance.
Wewillbuilduptotherelevantprinciple in twostages.Tostart,supposethat the
agentiscertainthattheobjectivevalueofoption𝑎%exceedstheobjectivevalueofoption𝑎'
bysomedefinitepositivemargin𝑧,butisuncertainwhethertheobjectivevaluesof𝑎%and
𝑎'equal𝑣%and𝑣' = 𝑣% − 𝑧,respectively,orinsteadequal𝑣Land𝑣M = 𝑣L − 𝑧,respectively,
where𝑣% − 𝑣L > 𝑧.Anyremotelyplausibleconceptionofobjectivevaluemustvalidate:
BaselineRelevance.Itispossiblefortheagent’scredencestobesuchthat:
i. 𝐶(⟨𝑂(𝑎%) = 𝑣%⟩ ∨ ⟨𝑂(𝑎%) = 𝑣L⟩) = 1,
ii. 𝐶(⟨𝑂(𝑎%) = 𝑣%⟩|𝑎%) = 𝑥 < 1,
iii. 𝐶(⟨𝑂(𝑎%) = 𝑂(𝑎') + 𝑧⟩) = 1,and
iv. 𝐶(⟨𝑂(𝑎%) = 𝑣%⟩|𝑎') = 𝑦 ≠ 𝑥.
The rationale behind Baseline Relevance is the same as that behind Relevance. Baseline
Relevanceholdsbecauseanyremotelyplausibleconceptionofobjectivevaluemustallowan
agenttoregard𝑎%asevidenceaboutwhattheobjectivevalueof𝑎% is,eveniftheagentis
certainthattheobjectivevalueof𝑎%exceedstheobjectivevalueofsomeotheroption𝑎'by
somemargin𝑧.
IfBaselineRelevanceholdsofeveryremotelyplausibleconceptionofobjectivevalue,
thensotoodoesavariantinwhichclause(iv)isstrongerthanabareinequality.Usingthe
samenotation,theprincipleis:
26
Newcombizability.Itispossiblefortheagent’scredencestosatisfy:
i. 𝐶(⟨𝑂(𝑎%) = 𝑣%⟩ ∨ ⟨𝑂(𝑎%) = 𝑣L⟩) = 1,
ii. 𝐶(⟨𝑂(𝑎%) = 𝑣%⟩|𝑎%) = 𝑥 < 1,
iii. 𝐶(⟨𝑂(𝑎%) = 𝑂(𝑎') + 𝑧⟩) = 1,and
iv. 𝐶(⟨𝑂(𝑎%) = 𝑣%⟩|𝑎') = 𝑦 > 𝑥 + y^ao^z
.
Asfaraswecansee,almostanyconceptionofobjectivevalueisNewcombizable.25
Indeed,thereseemstobeageneralrecipeforNewcombizing.Let𝑂beanyconceptionof
objectivevalue.Ifthereareoptions𝑎%and𝑎',thentherearepropositions𝑆% =def. ⟨𝑂(𝑎%) =
𝑣% ∧ 𝑂(𝑎') = 𝑣% − 𝑧⟩ and 𝑆' =def. ⟨𝑂(𝑎%) = 𝑣L ∧ 𝑂(𝑎') = 𝑣L − 𝑧⟩. We can therefore
constructadecisionprobleminwhichthepayoffsareasfollows:
𝑺𝟏 𝑺𝟐
𝒂𝟏 𝑣% 𝑣L
𝒂𝟐 𝑣' = 𝑣% − 𝑧 𝑣M = 𝑣L − 𝑧
Table1
25Thereasonforthequalification‘almost’isjustthis.(i)Weareassumingthattheobjectivevalue
conceptimpliesthatobjectivevaluetakesontheappropriatevaluesforsomeoptions𝑎%and𝑎'in
somepossibleworlds.Thus,forexample,wedisregardanyconceptionofobjectivevalueonwhich
every option has the same objective value at every world. (ii)We are assuming that any given
distributionofobjectivevaluesofoptionsacrosspossibleworldsisconsistentwithanydistribution
ofcredencesacrossthoseworlds.Thisdoesnotseemcontentioustous,sincetheobjectivevalueof
options are at least sometimes totally insensitive to changes in the agent’s credences. See the
correspondingdiscussionofRelevanceatn.18.
27
Weconstructacredencefunction𝐶ontheatomsF𝑎9𝑆:G9,:p%,'asfollows.Let𝑥 =%'�1 − y
^ao^z�.
Let 𝑦 = %M�3 + y
^ao^z�. Choose an arbitrary 𝑘, 0 < 𝑘 < 1. Let 𝐶(𝑎%𝑠%) = 𝑥𝑘, 𝐶(𝑎%𝑠') =
(1 − 𝑥)𝑘,𝐶(𝑎'𝑠%) = 𝑦(1 − 𝑘)and𝐶(𝑎'𝑠') = (1 − 𝑦)(1 − 𝑘).Thenitiseasytocheckthat𝐶
satisfiesallofclauses(i)-(iv)intheNewcombizabilitycondition.Ifoneneedsaback-story,
imaginethatthemechanismthattypicallycausesonetochoose𝑎%alsoandindependently
tendstopromoteastate𝑆'inwhichbothoptionspossesslessofwhicheverkindofobjective
valueisatissue.26
26Asaconcreteillustrationofthisgeneralresult,notethatHorganvalueisNewcombizable.Suppose
that youhave a choicebetween takingbox1 andbox2, bothboxesbeingopaque.Takingbox1
releases2unitsofwelfare;takingbox2doesnothing.ButaNewcombianpredictorhaswrittendown
whatheexpectsyourchoicetobe;andifyoumanagetooutwitthepredictor,yougetabonusof8
unitsofwelfare.Moreover,youknowthatthepredictorissomewhatbetteratpredictingpeoplewho
choosebox1(successrateof0.625)thanatpredictingpeoplewhochoosebox2(successrateof
0.1875).Thisneednotbebecauseyouareantecedentlyandrobustlyconfidentthatthepredictionis
thatyouwillchoosebox1;itcouldbethatactivationofthepartofyourbrainthatinclinesyouto
choosebox2interfereswiththepredictor’sbrain-scanningdevice.(Forasimilarexampleseethe
‘Semi-Frustrater’probleminSpencerandWellsforthcoming.)
ThiscasesatisfiesthefourclausesofNewcombizability.(i)youarecertainthattheHorgan
valueof(𝑎%)takingbox1iseither𝑣% = 10(ifthepredictorisactuallyinaccurate)or𝑣L = 2(ifthe
predictorisactuallyaccurate).Tospelloutwhy:on theHorganresolutionofcounterfactuals, the
closest𝑎%-worldtoactuality, call it𝑤%, isonewhere thepredictor isaccurate ifandonly ifhe is
actuallyaccurate.Soifthepredictorisactuallyinaccuratethen𝑢(𝑤%) = 10;otherwise𝑢(𝑤%) = 2.So
certainly𝑂(𝑎%) = 10 ∨ 𝑂(𝑎%) = 2.(ii)YourconfidencethattheHorganvalueoftakingbox1is10,
giventhatyoutakebox1, is𝑥 = 0.375.(iii)YouarecertainthattheHorganvalueoftakingbox1
exceedstheHorganvalueof(𝑎')takingbox2by𝑧 = 2 units,becauseifthepredictorisactually
accuratethentheformeris2andthelatteriszero,andifthepredictorisinaccuratethentheformer
is10andthelatteris8.(iv)YourconfidencethattheHorganvalueoftakingbox1is10,giventhat
28
If every remotely plausible conception of objective value is Newcombizable, then
DominanceandEDTcannotbothbetrue,because,asweproveinAppendixB:
Result#2.Newcombizability,Dominance,andEDTarejointlyinconsistent.
Theintuitiveideabehindthisisasfollows.Whateverobjectivevalueis,wecanconstructa
casewhereanoption𝑎%hasmoreofitthananotheroption𝑎'ateverypossibleworld;but
thedominatedoption,𝑎',isverygoodevidencethatbothoptionshavehighobjectivevalue.
Dominancethereforedemandsthatthesubjectivevalueof𝑎%exceedsthatof𝑎';butEDT
impliestheopposite.
The argument fromNewcombizability follows fromResult #2. It says that EDT is
inconsistentwithoptionshavingobjectivevaluesbecause,ifoptionshaveobjectivevalues,
NewcombizabilityandDominancearetrue.
The argument from Expectationism can be resisted by retreating from
Expectationism to Conditional Expectationism, but no analogous ‘conditionalizing’
maneuverhelpsagainsttheargumentfromNewcombizability.Toseethis,considerafew
proposals.
TheobviousfirstproposalistorejectDominanceinfavorof:
you take box 2, is 𝑦 = 0.8125 > 𝑥 + y
^ao^z= 0.625. So the case represents a Newcombization of
Horganvalue.
29
ConditionalDominance1:Ifoptionshaveobjectivevaluesthen:ifanagent’soptions
are𝑎%, 𝑎',… , 𝑎-,andtheagentiscertainthatoption𝑎9 uniquelymaximizesobjective
valuegiventhat𝑎9 isrealized,thenoption𝑎9 uniquelymaximizessubjectivevalue.27
Butthisgetsusnofurther,becauseConditionalDominance1entailsDominance.Ifanagent
iscertainthat𝑎9 uniquelymaximizesobjectivevalue,thentheagentisalsocertainthat𝑎9
uniquely maximizes objective value given that 𝑎9 is realized. So, Newcombizability,
ConditionalDominance1,andEDTarejointlyinconsistent.
AsecondproposalmightbetorejectDominanceinfavorof:
ConditionalDominance2:Ifoptionshaveobjectivevaluesthen:ifanagent’soptions
are 𝑎%, 𝑎',… , 𝑎-, and the agent is certain that: the objective value of option 𝑎9
conditional on its realization exceeds the objective value of any other option 𝑎;
conditionalonitsrealization,thenoption𝑎9 uniquelymaximizessubjectivevalue.28
Conditional Dominance 2 is not inconsistent with the conjunction of EDT and
Newcombizability,butthat’sbecauseitsaysnothingmeaningfulatall.Theobjectivevalueof
anoptionisnothadrelativetoacredencefunction,sothereisnosuchthingastheobjective
valueofanoptionconditionalonitsrealization(orconditionalonanything).
27Formally,wecanwritetheconsequentofConditionalDominance1:𝐶(⋀ ⟨𝑂(𝑎9) > 𝑂(𝑎;)⟩;�9 |𝑎9) =
1 → ⋀ 𝜎(𝑎9) > 𝜎(𝑎;);�9 .28 Formally, we might try writing the consequent of Conditional Dominance 2 as follows:
𝐶(⋀ ⟨𝑂(𝑎9|𝑎9) > 𝑂(𝑎;|𝑎;)⟩;�9 ) = 1 → ⋀ 𝜎(𝑎9) > 𝜎(𝑎;);�9 .
30
Onecouldtrytogive‘conditionalobjectivevalue’anobjectivemeaning,perhapsby
explicatingitasacounterfactualconditionalrelativetosomesimilaritymetric.Theobjective
valueofanoptionconditionalonitsrealizationthenwouldbetheobjectivevaluethatthe
optionwouldhavewereittoberealized.Thiswouldgiveus:
ConditionalDominance3:Ifoptionshaveobjectivevaluesthen:ifanagent’soptions
are𝑎%, 𝑎',… , 𝑎-,andtheagentiscertainthat:theobjectivevalueofoption𝑎9 wereit
toberealizedexceedstheobjectivevalueofanyotheroption𝑎;wereittoberealized,
thenoption𝑎9 uniquelymaximizessubjectivevalue.29
Butthisjustrearrangesthedeck-chairs.If𝑂isaconceptionofobjectivevalue,thenthe𝑂-
valueofanoptionwere it toberealized is justanother conceptionofobjectivevalue,and
ConditionalDominance3isjustDominanceassertedaboutit.Thisalternativeconceptionof
objectivevalueis,likeanyremotelyplausibleconception,Newcombizable,sonoprogressis
made.Newcombizability,ConditionalDominance3,andEDTarejointlyinconsistent.
ThereisnowayaroundDominance,justasthereisnowayaroundCertainReflection.
To reject either is to reject thewhole idea of objective values—objective properties of
options to which subjective values conform. Anyone who rejects Certain Reflection or
Dominance rejects the idea that differences between objective and subjective values are
alwaysmereartifactsoftheagent’suncertaintyaboutobjectivevalues.
29 Formally, we might write the consequent of Conditional Dominance 3 as: 𝐶H∀𝑛⋀ : (𝑎9 ⇒;�9
⟨𝑂(𝑎9) = 𝑛⟩) → (𝑎; ⇒ ⟨𝑂(𝑎;) < 𝑛⟩)I = 1 → ⋀ 𝜎(𝑎9) > 𝜎(𝑎;);�9 , where ⇒ is the selected
counterfactualoperator.
31
WhattheargumentfromNewcombizabilityrevealsisthereal,metaethicallessonof
Newcomb’sProblem.InthefiftyyearssinceNozickintroducedtheproblem,therehasbeen
much ado about causation. Over and again, opponents of EDT make the same causal
observations:theagentfacingNewcomb’sProblemhasnocontroloverthecontentsofthe
opaquebox;theamountofmoneycontainedintheopaqueboxatthecausallyclosestone-
boxingworld is likewise contained in the opaque box in the causally closest two-boxing
world; the agent is in a position to know for certain that the causal valueof two-boxing
exceedsthecausalvalueofone-boxing,andsoon.Theseobservationssuggestthatthelesson
ofNewcomb’sProblemhassomethingessentiallytodowithcausation.
Butitdoesn’t.Causallanguageisusedbecauseitispresupposedthatobjectivevalue
iscausalvalue,butthemetaethicallessonofNewcomb’sProblemconcernstherelationship
betweenobjectivevalueandsubjectivevalue,onanyplausibleconceptionofobjectivevalue.
Takeanyremotelyplausibleconceptionofobjectivevalue,beitcausalorwhollynoncausal.
ThatconceptionofobjectivevaluewillbeNewcombizable.TherewillbeacasewhereEDT
recommends an option that the agent knows for certain to be objectivelyworse on that
conceptionthantheonlyalternative.TheEDT’ist’sclaimthatsubjectivevalueisevidential
expected value thuswill be inconsistentwith the claim that that conception of objective
valueistrue.
Maximin, Risk-adjusted Expectationism, and the various form of Expectationism,
including CDT and HDT, are consistent with options having objective values. Indeed,
arguably these theories are defensible only if options have objective values. But EDT is
metaethicallyverydifferent.EDTisnotconsistentwithoptionshavingobjectivevalues.
32
§8/Conclusion
Pastthispoint,theauthorsofthispaperpartways.WeagreethatEDTisinconsistentwith
optionshavingobjectivevalue,butwedon’tagreeaboutwhattomakeofthatfact.Oneusof
hasantecedentcommitmentstooptionshavingobjectivevalues,soisinclinedtotakethe
inconsistency to amount to a metaethical refutation of EDT. The other has antecedent
commitments toEDT,so is inclined totake the inconsistency toamount toametaethical
refutationoftheclaimthatoptionshaveobjectivevalues.Obviously,wecannotsettlehere
whethertorejectEDTortheclaimthatoptionshaveobjectivevalue.Butoneofthemhasgot
togo.
33
AppendixA
HereweproveResult#1:thatRelevance,Expectationism,andEDTarejointlyinconsistent.
Start from a possible case thatwitnesses the truth of Relevance. Then, according to the
agent’s credences, 𝐶(⟨𝑂(𝑎) = 𝑣%⟩ ∨ ⟨𝑂(𝑎) = 𝑣'⟩) = 1, 𝐶(⟨𝑂(𝑎) = 𝑣%⟩) = 𝑥 < 1, and
𝐶(⟨𝑂(𝑎) = 𝑣%⟩|𝑎) = 𝑦 ≠ 𝑥.Theagent’sexpectationoftheobjectivevalueofoption𝑎is:
∑ 𝐶(⟨𝑂(𝑎) = 𝑣⟩)𝑣� = 𝑥𝑣% + (1 − 𝑥)𝑣' = 𝑥(𝑣% − 𝑣') +𝑣'.
TheV-valueof𝑎is:
∑ 𝐶(⟨𝑂(𝑎) = 𝑣⟩|𝑎)𝑣� = 𝑦𝑣% + (1 − 𝑦)𝑣' = 𝑦(𝑣% −𝑣') + 𝑣'.
IfExpectationismistrue,thesubjectivevalueof𝑎relativetoCisequaltotheexpectationof
theobjectivevalueof𝑎relativeto𝐶.Hence:
𝑥(𝑣% − 𝑣') +𝑣' = 𝑦(𝑣% −𝑣') + 𝑣'.
But 𝑣% ≠ 𝑣' implies that 𝑥(𝑣% − 𝑣') +𝑣' = 𝑦(𝑣% −𝑣') + 𝑣' only if 𝑥 = 𝑦, and 𝑥 ≠ 𝑦.
Therefore,EDT,Relevance,andExpectationismcannotallbetrue.QED.
Note thatExpectationism,aswehave characterized it, is a thesis aboutnumerical
values.Itsaysthatthenumericalrepresentationofthesubjectivevalueofanoptionrelative
toa credence function (on somearbitrarily chosen scale)mustbeara certain arithmetic
34
relationtothenumericalrepresentationoftheobjectivevalueoftheoption(onthatsame
scale).Butonemightbeconcernedwithamoregeneral,purelynormativethesis:namely,
NormativeExpectationism.Agentsalwayssubjectivelyoughttomaximizeexpected
objectivevalue.
TheproofabovedoesnotestablishthatEDT,Relevance,andNormativeExpectationismare
inconsistent, but the inconsistency between these three claims can now be proven very
simply. Just imagine a case which, relative to some arbitrary determination of a scale,
satisfiesRelevance, and supposewithout lossof generality that𝑣% > 𝑣' and𝑥 > 𝑦. Then
imagineachoicebetweenoptions𝑎and𝑏,where𝑎isasdescribedaboveand𝑏isalottery
withchance �l�'ofrealizinganoutcomewithobjectivevalue𝑣%andchance1 − �
�l�'�of
realizinganoutcomewithobjectivevalue𝑣'.NormativeExpectationism implies that the
agentsubjectivelyoughttorealize𝑎,butEDTimpliesthattheagentsubjectivelyoughtto
realize𝑏.
AppendixB
Here we prove Result #2: that Newcombizability, Dominance, and EDT are jointly
inconsistent.StartfromacasethatwitnessesthetruthofNewcombizability.Then,according
totheagent’scredences:𝐶(⟨𝑂(𝑎%) = 𝑣%⟩ ∨ ⟨𝑂(𝑎%) = 𝑣L⟩) = 1,𝐶(⟨𝑂(𝑎%) = 𝑣%⟩|𝑎%) = 𝑥 < 1,
𝐶(⟨𝑂(𝑎%) = 𝑂(𝑎') + 𝑧⟩) = 1, and 𝐶(⟨𝑂(𝑎%) = 𝑣%⟩|𝑎') = 𝑦 > 𝑥 + y^ao^z
. Then since EDT
entailsConditionalExpectationism(seen.22),theV-valueofoption𝑎%equals:
35
∑ 𝐶(⟨𝑂(𝑎%) = 𝑣⟩|𝑎%)𝑣� = 𝑥𝑣% + (1 − 𝑥)𝑣L = 𝑥(𝑣% − 𝑣L) + 𝑣L.
TheV-valueofoption𝑎'equals:
∑ 𝐶(⟨𝑂(𝑎') = 𝑣⟩|𝑎')𝑣� = 𝑦𝑣' + (1 − 𝑦)𝑣M = 𝑦(𝑣% − 𝑧) + (1 − 𝑦)(𝑣L − 𝑧) =
𝑦(𝑣% − 𝑣L) + 𝑣L − 𝑧.
Andsince𝑦 > 𝑥 + y^ao^z
,
𝑦(𝑣% − 𝑣L) + 𝑣L − 𝑧 > �𝑥 + y^ao^z
� (𝑣% − 𝑣L) + 𝑣L − 𝑧 = 𝑥(𝑣% − 𝑣L) + 𝑣L.
Thus,ifEDTistrue,thesubjectivevalueof𝑎'relativetoCexceedsthesubjectivevalueof𝑎%
relativetoC.
Buttheagentiscertainthattheobjectivevalueof𝑎%exceedstheobjectivevalueof
𝑎'.So,ifDominanceistrue,thesubjectivevalueof𝑎%relativetoCexceedsthatof𝑎'relative
toC.Therefore,Newcombizability,EDT,andDominancecannotallbetrue.QED.
36
References
Ahmed,A.2014.Evidence,DecisionandCausality.Cambridge:CambridgeUniversityPress.
Buchak,L.2013.RiskandRationality.Oxford:OxfordUniversityPress.
Bradley,R.andH.O.Stefánsson.2016.Desire,ExpectationandInvariance.Mind125:691-
725.
Broome,J.1991.Desire,BeliefandExpectation.Mind100:265-7.
Dreier,J.1993.StructuresofNormativeTheories.Monist76:22-40.
———. 2011. In Defense of Consequentializing. In Timmons, M. (ed.),Oxford Studies in
NormativeEthics,Vol.1.Oxford:OxfordUniversityPress:97-119.
Hájek,A.2015.OnthePluralityofLewis’sTrivialityResults. InLoewer,B.andJ.Schaffer
(ed.),ACompaniontoDavidLewis.Oxford:Blackwell:425-45.
———andP.Pettit.2004.DesireBeyondBelief.AustralasianJournalofPhilosophy82:77-
92.
Hare,C.2010.TaketheSugar.Analysis70:237-47.
———.2016.ShouldWeWishWelltoAll?PhilosophicalReview125:451-72.
Horgan,T.1981.CounterfactualsandNewcomb’sProblem.JournalofPhilosophy78:331-56.
Jackson, F. 1991. Decision-Theoretic Consequentialism and the Nearest and Dearest
Objection.Ethics101:461-88.
Jeffrey,R.C.1983.TheLogicofDecision.Seconded.Chicago:UniversityofChicagoPress.
Joyce,J.1999.TheFoundationsofCausalDecisionTheory.Cambridge:CambridgeUniversity
Press.
Kolodny,N.andJ.MacFarlane.2010.IfsandOughts.JournalofPhilosophy107:115-43.
Lewis,D.K.1973.Counterfactuals.Oxford:Blackwell.
37
———.1979.CounterfactualDependenceandTime’sArrow.Noȗs13:455-76.Reprintedin
hisPhilosophicalPapersVolumeII(1986).Oxford:OxfordUniversityPress.
———.1981.CausalDecisionTheory.AustralasianJournalofPhilosophy59:5-30.
———.1988.DesireasBelief.Mind97:323-32.
———.1996.DesireasBeliefII.Mind105:303-13.
Louise,J.2004.RelativityofValueandtheConsequentialistUmbrella.PhilosophicalQuarterly
54:518-36.
McDermott,M.1999.CounterfactualsandAccessPoints.Mind108:291-334.
Mackie,J.L.1977.Ethics:InventingRightandWrong.London:PenguinBooks.
Moore,G.E.1903.PrincipiaEthica.Cambridge:CambridgeUniversityPress.
Oddie,G.1994.Harmony,Purity,Truth.Mind103:451-72.
———andP.Menzies.1992.AnObjectivist’sGuidetoSubjectiveValueEthics102:512-33.
Parfit,D.1984.ReasonsandPersons.Oxford:OxfordUniversityPress.
———.Unpublished.WhatWeTogetherDo.
http://individual.utoronto.ca/stafforini/parfit/parfit_-_what_we_together_do.pdf
Pettigrew, R. 2015. Risk, Rationality, and Expected Utility Theory Canadian Journal of
Philosophy45:796-826.
Portmore,D.W.2007.ConsequentializingMoralTheories.PacificPhilosophicalQuarterly88:
39-73.
Price,H.1989.DefendingDesire-as-Belief.Mind98:119-27.
Regan,D.1980.UtilitarianismandCooperation.Oxford:OxfordUniversityPress.
Spencer, J. and I. Wells. Forthcoming. Why Take Both Boxes? Philosophy and
PhenomenologicalResearch.
38
Thomson,J.J.2008.Normativity.Peru,Illinois:OpenCourt.
Zimmerman,M.2008.LivingWithUncertainty.Cambridge:CambridgeUniversityPress.