ebook img

A categorical semantics for causal structure PDF

0.29 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview A categorical semantics for causal structure

A categorical semantics for causal structure AleksKissinger and SanderUijlen InstituteforComputingandInformationSciences,RadboudUniversity 7 1 {aleks,suijlen}@cs.ru.nl 0 2 January 26,2017 n a J Abstract 5 2 We presenta categorical construction for modellingboth definite and indefinite causal struc- tureswithinageneralclassofprocesstheoriesthatincludeclassicalprobabilitytheoryandquan- ] tumtheory. Unlikepriorconstructionswithincategoricalquantummechanics, theobjectsofthis h p theoryencode fine-grainedcausal relationshipsbetween subsystemsand givea newmethod for - expressingandderivingconsequencesforabroadclassofcausalstructures.Toillustratethispoint, t n we show that this framework admits processes with definite causal structures, namely one-way a signallingprocesses,non-signallingprocesses,andquantumn-combs,aswellasprocesseswithin- u definitecausalstructure,suchasthequantumswitchandtheprocessmatricesofOreshkov,Costa, q and Brukner. We furthermoregive derivationsof their operational behaviour using simple, dia- [ grammaticaxioms. 2 v 2 1 Introduction 3 7 Broadly,causalstructuresidentifywhicheventsorprocessestakingplaceacrossspaceandtimecan, 4 0 inprinciple,serveascausesoreffectsofoneanother. Forinstance,thecausalstructureofrelativistic . spacetimeisgivenbydemandingeventscanonlyhaveacausalinfluenceonothereventswhichcan 1 0 bereachedwithoutexceedingthespeedoflight. 7 Inthe context of quantum theory, causalrelationships between inputs and outputs to quantum 1 processes have been expressed in a variety of ways. Perhaps the simplest are in the form of non- : v signalling constraints, which guarantee that distant agents are not capable of sending information i X fasterthatthespeedoflight,e.g. toaffecteachother’smeasurementoutcomes[3]. Quantumstrate- gies [15] and more recently quantum combs [4] offer a means of expressing more intricate causal r a relationships,intheformofchainsofcausallyorderedinputsandoutputs. Furthermore,ithasbeen shown recently that one can formulate a theory that is locally consistent with quantum theory yet assumes no fixed background causal structure [22]. Interestingly, such a theory admits indefinite causalstructures.Namely,itallowsonetoexpressprocesseswhichinhabitaquantumsuperposition of causal orders. If physically realisable, such processes can outperform any theory that demands a fixed causalordering in certainnon-local games such as ‘guess your neighbour’s input’ [22] and quantum computational tasks such as single-shot channel discrimination [6]. Perhaps even more surprisingly,ithasbeenshownrecentlythat,inthepresenceofthreeormoreparties,causalbounds canbeviolatedevenwithinatheorythatbehaveslocallylikeclassicalprobabilitytheory[2]. The keyingredientin studying (andvarying) causalstructuresof processesis the development ofacoherenttheoryofhigherordercausalprocesses. Thatis,ifwetreatlocalagents(orevents,labora- tories,etc.) asfirstordercausalprocesses,thentheactofcomposingtheseprocessesinaparticular causal order should be treated as a second-order process. In this paper, we develop a categorical 1 framework for expressing and reasoning about such higher order processes. This starts from the generalcontextofaprecausalcategory,whichisacompactclosed categorythatsatisfiesafourextra axiomsthatprovideenough additionalstructuretoreasonaboutcausalrelationshipsbetweensys- temsandtoprovesimpleno-goresultssuchasnotime-travel,i.e.theimpossibilityofprocesssending informationintoitsowncausalpast. Most importantly, precausal categories have a special discarding process defined on each object, whichenablesonetostatethecausalitypostulateforaprocessΦ: A B: → B Φ = A A Thishasaclearoperationalintuition: iftheoutputofaprocessisdiscarded,itdoesn’tmatterwhichprocess occurred.Whileseeminglyobvious,thiscondition,originallygivenby[5]inthecontextofoperational probabilistic theories, is surprisingly powerful. For instance, it is equivalent to the non-signalling propertyforjointprocessesarisingfromsharedcorrelations[13]. Startingfromaprecausalcategory ,wegiveaconstructionofthe -autonomouscategoryCaus[ ] C ∗ C ofhigher-ordercausalprocesses,intowhichthecategoryof(first-order)causalprocessessatisfying the equation above embeds fully and faithfully. Our main examples start from the precausalcate- goriesofmatricesofpositiverealnumbersandcompletelypositivemaps,whichwillyieldcategories ofhigher-orderclassicalstochasticprocessesandhigher-orderquantumchannels,respectively. Whilecategoricalquantummechanics[1]hastypicallyfocussedoncompactclosedcategoriesof quantumprocesses,weshowthatthis -autonomousstructureyieldsamuchrichertypesystemfor ∗ describingcausalrelationships betweensystems. A simple, yetstriking example is in the use of ⊗ vs. `toformthetypesofprocessesonajointsystem: (A⊸A′) (B⊸B′) non-signallingprocesses ⊗ ← (A⊸A′)`(B⊸B′) allprocesses ← Usingthistypesystem,wegivelogicalcharacterisationsofnon-signallingandone-waynon-signalling processes and then prove that these are equivalent to the operational characterisations in terms of their interactions with the discardingprocess. We thengo on tocharacterisehigher-ordersystems, notablybipartite,andn-partitesecond-ordercausalprocessesandgiveseveralexampleswhichare knowntoexhibitindefinitecausalstructure,namelytheOCBprocessfrom[22],theclassicaltripar- titeprocessfrom[2],andanabstractversionofthequantumswitchdefinedin[6]. Finally,weprove using just the structure of Caus[ ] thatthe switchdoes notadmita causalorderingbyreducingto C notime-travel. Relatedwork. Thisworkwasinspiredby[23],whichaimsforauniformdescriptionofhigher- orderquantum operations in termsof generalised Choi operators. However, rather thanrelyingon the linearstructure of spacesof operators, we workpurelyintermsof the -autonomous structure ∗ andtheprecausalaxioms,whichconcernthecompositionalbehaviourofdiscardingprocesses. The constructionofCaus[ ]isavariantofthedoublegluingconstructionusedin[17]toconstructmodels C oflinearlogic. Inthelanguageofthatpaper,ourconstructionconsistsofbuildingthe‘tightorthog- onality category’induced bya focussed orthogonality on 1 (I,I), then restrictingtoobjects I { } ⊆ C satisfying the flatness condition in Definition 4.2. Since it is -autonomous, Caus[ ] indeed gives ∗ C a model of multiplicative linear logic, enabling us to enlist the aid of linear-logic based tools for provingtheoremsaboutcausaltypes. Wecommentbrieflyonthisintheconclusion. Acknowledgements.BothauthorsaregratefultoPauloPerinottiforhisinputandsharingadraft versionof[23].WewouldalsoliketothankCˇaslavBrukner,MattyHoban,andSamStatonforuseful discussions. This work is supported by the ERC under the EuropeanUnion’s Seventh Framework Programme(FP7/2007-2013)/ERCgrantno320571. 2 2 Preliminaries We work in the context of symmetric monoidal categories (SMCs). An SMC consists of a collection of objects ob( ), for every pair of objects, A,B ob( ) a set (A,B) of morphisms, associative C ∈ C C sequentialcomposition‘ ’withunits1 forall A ob( ),associative(uptoisomorphism)parallel A ◦ ∈ C composition ‘ ’ for objects and morphisms with unit I ob( ), and swap maps σ : A B A,B ⊗ ∈ C ⊗ → B A, satisfying the usual equations one would expect for composition and tensor product. For ⊗ simplicity,wefurthermoreassume isstrict,i.e. C (A B) C = A (B C) A I = A = I A ⊗ ⊗ ⊗ ⊗ ⊗ ⊗ ThisisnolossofgeneralitysinceeverySMCisequivalenttoastrictone.Fordetails,[21]isastandard reference. WewishtotreatSMCsastheoriesofphysicalprocesses,henceweoftenrefertoobjectsassystems and morphisms a processes. We will also extensively use string diagram notation for SMCs, where systemsaredepictedaswires,processesandboxes,and: g g f := f g:= g f ◦ ⊗ f B A 1A := A 1I := σA,B := A B Note that diagrams should be read from bottom-to-top. A process ρ : I A is called a state, a → processπ : A I iscalledaneffect,andλ : I I iscalledanumber. Depictedasstringdiagrams: → → state:= effect:= π number:=λ ρ NumbersinanSMCalwaysformacommutativemonoidwith‘multiplication’ andunittheidentity ◦ morphism1 . Wetypicallywrite1 simplyas1. I I We will begin with a category and construct a new category Caus[ ] of higher-order causal C C processes. In order to make this construction, we first need a mechanism for expressing higher- orderprocesses. Compactclosureprovidessuchamechanismthatisconvenientwithinthegraphical language andalreadyfamiliarwithin the literatureonquantum channels, in the guise of the Choi- Jamiołkowskiisomorphism. Definition2.1. AnSMC iscalledcompactclosedifeveryobject A hasadualobject A . Thatis,for ∗ C everyAthereexistsmorphismsη : I A Aandǫ : A A I,satisfying: A ∗ A ∗ → ⊗ ⊗ → (ǫ 1 ) (1 η ) =1 (1 ǫ ) (η 1 )=1 A A A A A A A A A A ⊗ ◦ ⊗ ∗⊗ ◦ ⊗ ∗ ∗ Werefertoη andǫ asacupandacap,denotedgraphicallyas and ,respectively. A A Inthisnotation,theequationsinDefinition2.1become: A A∗ = = A∗ A A A∗ A A∗ 3 Itisalwayspossibletochoosecupsandcapsinsuchawaythatthecanonicalisomorphisms I = I, ∗ ∼ (A B) = A B ,andA = A areallinfactequalities. Wewillassumethisisthecasethroughout ⊗ ∗ ∼ ∗⊗ ∗ ∼ ∗∗ thispaper. Crucially, two morphisms in a compact closed category are equal if and only if their string di- agrams are the same. That is, if one diagram can be continuously deformed into the other while mainingtheconnectionsbetweenboxes. Hence,whenwedrawastringdiagram,wemeananycom- positionofboxesviacups,caps,andswapswhichyieldsthegivendiagram,uptodeformation. See [24]foranoverviewofstringdiagramlanguagesformonoidalcategories. Compactclosed categoriesexhibitprocess-stateduality,thatis, processes f : A B arein1-to-1 → correspondencewithstatesρ : I A B: f → ∗⊗ f f (1) 7→ ρf Hence, we treateverythingasa‘state’in and write f : X asshorthand for f : I X. Inthis C → notation,statesareoftheformρ : A,effectsπ : A ,andgeneralprocesses f : A B. Furthermore, ∗ ∗ ⊗ wewon’trequire‘output’wirestoexitupward,andweallowirregularly-shapedboxes.Forexample, wecanwriteaprocessw : A B C Dusing‘comb’notation: ∗ ∗ ⊗ ⊗ ⊗ D w C := A∗ B C∗ D (2) B w A Note that we adopt the convention that an A-labelled ‘input’ wire is of the same type as an A - ∗ labelled‘output’. While both the LHS and the RHS in equation (2) are notation for the same process w, the LHS is strongly suggestive of a second-order mapping from processes B C to processes A D. → → Composition in this notation simply meansapplying the appropriate‘cap’processesto plug wires together: D A∗ D C w Φ := B C∗ B∗ C w Φ B A Remark2.2. Sinceoddly-shapedboxesdon’tuniquelyfixanyorderingofsystemswithrespectto , ⊗ wewilloften‘name’eachsystembygivingitaunique typeandassumesystemsarepermutedvia σ-mapswhenevernecessary. Thisisacommonpracticee.g. inthequantuminformationliterature. Our key exampleswill be Mat(R+) and CPM, which contain stochastic matricesand quantum channels,respectively. Example 2.3. The category Mat(R+) has as objects natural numbers. Morphisms f : m n are → n mmatrices.Compositionisgivenbymatrixmultiplication: (g f)j :=∑ fkgj,m n :=mnand × ◦ i k i k ⊗ f gistheKroneckerproductofmatrices: ⊗ 4 (f g)kl := fkgl ⊗ ij i j Consequently, thetensor unit I = 1,sostatesarecolumnvectorsρ : 1 n, effectsarerowvectors → π : n 1,andnumbersareλ R+. Mat(R+)iscompactclosedwithn = n∗,wherecupsandcaps → ∈ aregivenbytheKroneckerdeltaδ : ij ηij :=δ =: ǫ ij ij Example 2.4. The category CPM has as objects (H), (K),... for Hilbert spaces H,K,... and as B B morphisms completely positive maps Φ : (H) (K). (H) (K) = (H K), hence I = B → B B ⊗B B ⊗ (C) = C. Statesarethereforepositive operatorsandeffectsare(un-normalised)quantum effects, B ∼ i.e. completely positive maps of the form: π(ρ) := Tr(ρP), for some positive operator P. As with Mat(R+), numbers are R+. CPM is also compact closed, with cups and caps given by the (un- normalised)Bellstate: η = Φ Φ ǫ(ρ)=Tr(ρ Φ Φ ) 0 0 0 0 | ih | | ih | where Φ =∑ i i .Consequentially, (H) = (H)andequation(1)givesthebasis-dependent | 0i i| i⊗| i B ∗ B version,i.e. ‘Choi-style’,oftheChoi-Jamiołkowskiisomorphism[19]. The biggest convenience of a compact closed structure is also its biggest drawback: all higher- order structure collapses to first-order structure! As we will soon see, there is a pronounced dif- ferencebetweenfirst-ordercausalprocesseswhichwe introduceinthe nextsection, andgenuinely higher-order causal processes. Thus, while it is natural to take to be compact closed, we expect C Caus[ ]tobeadifferentkindofcategory,whichallowsthisgenuinehigher-orderstructure. C Definition2.5. A -autonomouscategoryisasymmetricmonoidalcategoryequippedwithafulland ∗ faithfulfunctor( ) : op suchthat,byletting: ∗ − C → C A⊸B :=(A B∗)∗ (3) ⊗ thereexistsanaturalisomorphism: (A B,C)= (A,B⊸C) (4) C ⊗ ∼ C ⊸ Asthe notationandtheisomorphism (4)suggest, A B isthe systemwhose statescorrespond ⊸ toprocessesfromAtoB.Weadopttheprogrammers’conventionthat hasprecedenceover and ⊸ ⊗ associatestotheright: A B⊸C :=(A B)⊸C A⊸B⊸C := A⊸(B⊸C) ⊗ ⊗ Either expressionabove representsthe system whose states areprocesseswith two inputs. Indeed ⊸ ⊸ ⊸ (4)impliesthatA B C = A B C. Since issymmetric,wecanre-arrangetheinputsatwill, ⊗ ∼ C i.e. ⊸ ⊸ ⊸ ⊸ A B C = B A C (5) ∼ ⊸ Unlikein acompactclosed category, the object A B canbe differentfromsimply A B. In- ∗ ⊗ deedanycompactclosedcategoryis -autonomous,whereitadditionallyholdsthat: ∗ A⊗B ∼=(A∗⊗B∗)∗ (6) inwhichcase: A⊸B :=(A⊗B∗)∗ ∼= A∗⊗B∗∗ ∼= A∗⊗B However,ina -autonomouscategory,theRHSof(6)isnotA B,butsomethingnew,calledthe ∗ ⊗ ‘par’of AandB: 5 A`B:=(A∗ B∗)∗ (7) ⊗ Thisnewoperationinheritsitsgoodbehaviourfrom : ⊗ A`(B`C) =(A`B)`C A`B = B`A ∼ ∼ Soacompactclosedcategoryisjusta -autonomouscategorywhere = `. However,thislittle ∗ ⊗ tweakyieldsamuchricherstructureofhigher-ordermaps. WethinkofA Basthejointstatespace ⊗ ofAandB,whereasA`BisliketakingthespaceofmapsfromA toB.For(firstorder)statespaces, ∗ thesearebasicallythesame,butaswegotohigherorderspaces,A`Btendstobemuchbiggerthan A B. ⊗ 3 Precausal categories Precausalcategoriesgiveauniverseofallprocesses,andprovideenoughstructureforustoidentify whichofthoseprocessessatisfyfirst-orderandhigher-ordercausalityconstraints. As noted in [10, 11, 12, 5], the crucial ingredient for defining causality is a preferred discarding process fromeverysystem Ainto I. Usingthiseffect,wecandefinecausalityasfollows: A Definition3.1. For systems A and B with discardingprocesses, a process Φ : A B issaid to be → causalif: B Φ = A (8) A Intuitively,causalitymeansthatifwedisregardtheoutputofaprocess,itdoesnotmatterwhich processoccurred.Wedefinediscardingon A Band I intheobviousway: ⊗ := := 1 (9) A B A B I ⊗ Hencecausalstatesproduce1whendiscarded: causal = 1 ρ ⇐⇒ ρ Sincediscardingthe‘output’ofaneffectπ : A I istheidentity,thereisauniquecausaleffectfor → anysystem,namelydiscardingitself: π causal π = ⇐⇒ Example3.2. ForMat(R+),thediscardingprocessisarowvectorconsistingentirelyof1’s;itsends astatetothesumoveritsvectorentries: = 1 1 1 = ∑ ρi ρ ··· i (cid:0) (cid:1) So, causality is precisely the statement that a vector of positive numbers sums to 1, i.e. forms a probabilitydistribution. Consequentially, thecausalityequation(8)foraprocessΦstatesthateach columnofΦmustsumto1. Thatis,Φisastochasticmap. Example 3.3. For CPM, discard is the trace. Hence causal states are density operators and causal processesaretrace-preservingcompletelypositivemaps,a.k.a.quantumchannels. 6 Discardingnotonlyallowsustoexpresswhenaprocessiscausal,italsoallowsustorepresent causalrelationshipsbetweenthesystemsinvolved. Forexample Definition3.4. AcausalprocessΦisone-waynon-signallingwith AbeforeB(written A B)ifthere (cid:22) existsaprocessΦ : A A suchthat ′ ′ → A′ B′ A′ Φ = Φ (10) ′ A B A B Itisone-waynonsignallingwithB AifthereexistsΦ suchthat ′′ (cid:22) A′ B′ B′ Φ = Φ (11) ′′ A B B Suchaprocessiscallednon-signallingifitisbothnon-signallingwith A BandB A. (cid:22) (cid:22) If A hasdiscarding,wecanalsoproduceastatefor A: ∗ A := A A∗ (12) Theseingredientsallowustomakethefollowingdefinition: Definition3.5. Aprecausalcategoryisacompactclosedcategory suchthat: C (C1) hasdiscardingprocessesforeverysystem,compatiblewiththemonoidalstructureasin(9). C (C2) Forevery(non-zero)system A,thedimensionof A: dA := A isaninvertiblescalar. (C3) hasenoughcausalstates: C  ρcausal. f = g  = f = g ∀ ⇒ ρ ρ       (C4) Second-ordercausalprocessesfactorise: ∀Φcausal. ∃Φ1,Φ2causal.    Φ  2 =  w Φ =  ⇒  w = Φ1      7 (C1)enablesonetotalkaboutcausalprocesseswithin . (C2)enablesustorenormalisecertain C processestoproducecausalones. Forexample,everynon-zerosystemhasatleastonecausalstate, calledtheuniformstate. Itisobtainedbynormalising(12): A := d−A1 A Then: A = d−A1 A = d−A1dA = 1 Notethatweallow tohaveazeroobject,inwhichcased =0isnotrequiredtobeinvertible. 0 C (C3)saysthatprocessesarecharacterisedbytheirbehaviouroncausalstates. Since iscompact C closed,(C3)impliesthatitsufficestolookonlyatproductstatestoidentifyaprocess. Proposition3.6. Foranycompactclosedcategory ,(C3)isequivalentto: C ∀ρ1,ρ2causal.  Φ = Φ  = Φ = Φ (13) ′ ⇒ ′    ρ1 ρ2 ρ1 ρ2      Proof. (C3) follows from (13) by taking one of the two systems involved to be trivial. Conversely, assumethepremiseof(13). Applying(C3)onetimeyields: Φ = Φ ′ ρ2 ρ2 forallcausalstatesρ . Bendingthewireyields: 2 Φ = Φ′ ρ2 ρ2 Hencewecanapply(C3)asecondtime. Bendingthewirebackdowngivestheresult. (C4) is perhaps the least transparent. It says that the only mappings from causal processes to causal processes are ‘circuits with holes’, i.e. those mappings which arise from plugging a causal processintoalargercircuitofcausalprocesses.Thiscanequivalentlybesplitintotwosmallerpieces, whichmaylookmorefamiliartosomereaders. Proposition3.7. Foracompactclosedcategory satisfying(C1),(C2),and(C3),condition(C4)is C equivalenttothefollowingtwoconditions: (C4′) Causalone-waysignallingprocessesfactorise: ∃Φ′causal.  ∃Φ1,Φ2causal.    = Φ2 Φ = Φ ⇒  Φ =   ′   Φ1        8 (C5′) Forallw : A B∗: ⊗ ∀Φcausal. ∃ρcausal.     = w Φ = 1 ⇒ w =            ρ          Proof. Assume(C4)holds,andletΦsatisfythepremiseof(C4′). Then,forcausalΨ,wehave: Φ Φ Ψ = ′ = = Φ Ψ Ψ ′ Hence,by(C4)thereexistcausalΦ ,Φ suchthat: 1′ 2 Φ Φ = 2 Φ 1 Deformingthengivesthefactorisationin(C4′). Togetto(C5′)from(C4),wetaketheinputandoutputsystemsofwtrivial,whichgives: Φ 2 w = = =: ρ Φ Φ 1 1 where the second ‘=’ follows from the factthat discardingis the unique causal effect. Conversely, suppose(C4′)and(C5′)hold. Letwbeasecondordercausalprocess. Then,foranycausalρandΦ: w Φ = 1 ρ So,by(C5′): Lemma3.8 w = = w = w ⇒ ρ′ ρ ρ ρ Then,by(C3): 9 w = w =: w′ Hence w satisfies the premise of (C4′), up to diagram deformation, where Φ := w and Φ′ := w′, whichimpliesthatitfactorsasin(C4). Lemma3.8. Foranyw : A B : ∗ ⊗  ρ. w =   w = w  ∃ ⇐⇒  ρ                Proof. ( )immediatelyfollowsfromdiagramdeformation. For( ),assumetheleftmostequation ⇐ ⇒ above. Then we canobtainanexpressionfor ρ byplugging the uniform state into w and applying causalityoftheuniformstate: w = = ρ ρ Substitutingthisexpressionbackinforρyields: B B B w = = = w B A A w ρ A A Example3.9. Mat(R+) and CPMareboth precausalcategories. Seeappendix for proofsof condi- tions(C1)-(C4). Notethatwecanreadilygeneraliseone-waynon-signallingfromDefinition(3.4)tonsystems. Definition3.10. AprocessΦisone-waynon-signalling(A ... A A )if 1 n 1 n (cid:22) (cid:22) − (cid:22) ... ... Φ = Φ ′ ... ... withΦ one-waynon-signallingwith A ... A . ′ 1 n 1 (cid:22) (cid:22) − Then(C4′)impliesangeneraln-foldversionofitself: Proposition3.11. LetΦbeone-waynon-signallingwith A ... A ,thenthereexistsΦ ,...,Φ 1 n 1 n (cid:22) (cid:22) suchthat 10

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.