View metadata, citation and similar papers at core.ac.uk brought to you by CORE provided by Elsevier - Publisher Connector JournalofComputerandSystemSciences78(2012)1883–1898 ContentslistsavailableatSciVerseScienceDirect Journal of Computer and System Sciences www.elsevier.com/locate/jcss Intrinsically universal n-dimensional quantum cellular automata Pablo Arrighia,b, Jonathan Grattagea,∗ aEcoleNormaleSupérieuredeLyon,LaboratoireLIP,46alléed’Italie,69364Lyoncedex07,France bUniversityofGrenoble,LaboratoireLIG,BâtimentIMAGC,220ruedelaChimie,38400Saint-Martin-d’Hères,France a r t i c l e i n f o a b s t r a c t Articlehistory: Several non-axiomatic approaches have been taken to define Quantum Cellular Automata Received20September2010 (QCA); Partitioned QCA (PQCA) are the most canonical. Here we show any QCA can be Receivedinrevisedform8March2011 put into PQCA form. Our construction reconciles the non-axiomatic definitions of QCA, Accepted17November2011 showing that they can all simulate one another, thus they are all equivalent to the Availableonline12January2012 axiomatic definition. A simple n-dimensional QCA capable of simulating all others to arbitrary precision is described, where the initial configuration and the evolution of any Keywords: Quantumcomputation QCA can be encoded within the initial configuration of the intrinsically universal QCA. Cellularautomata SeveralstepsthencorrespondtoonestepofthesimulatedQCA,achievedviaanon-trivial Universality reduction of the problem to universality in quantum circuits. Results are formalised by Quantumcircuits defining generalised n-dimensional intrinsic simulation, preserving topology in that each cell of the simulated QCA is encoded as a group of adjacent cells in the universal QCA. Implicationsarediscussed. ©2011ElsevierInc.Allrightsreserved. 1. Introduction Cellular automata (CA), first introduced by von Neumann [1], consist of an array of identical cells, each of which may take one of a finite number of possible states. The whole array evolves in discrete time steps by iterating a function G. Thisglobalevolution G isshift-invariant(itactsinthesamewayeverywhere)andlocal(informationcannotbetransmitted faster than some fixednumber ofcellspertime step). 1.1. QCA:Importanceandcompetingdefinitions ThemodernaxiomatisationofquantumtheoryintermsofthedensitymatrixformalismwasprovidedbyvonNeumann in1955[2],whoalsodevelopedthecellularautomata(CA)modelofcomputationin1966[1],buthedidnotbringthetwo together. Feynman did so in 1986 [3], just as he was developing the concept of quantum computation (QC). Listed below arethe keymultidisciplinary motivations forstudying QCA,thefirsttwo being those ofFeynman. • Implementationperspective.QCAmayprovideanimportantpathtorealisticimplementationsofQC,mainlybecausethey eliminatetheneedforanexternal,classicalcontroloverthecomputationandhencetheprincipalsourceofdecoherence. Thisisunder investigation [4–8]. • Simulationperspective. QC was first conceived as a way to efficiently simulate other quantum physical systems. Whilst other applications have been invented since, this still remains a likely and important application of QC. However, it * Correspondingauthor. E-mailaddresses:[email protected](P.Arrighi),[email protected](J.Grattage). 0022-0000/$–seefrontmatter ©2011ElsevierInc.Allrightsreserved. doi:10.1016/j.jcss.2011.12.008 1884 P.Arrighi,J.Grattage/JournalofComputerandSystemSciences78(2012)1883–1898 may not be straightforward to encode the theoretical description of a quantum physical system into a QC in a relevant manner, i.e. so that the QC can then provide an accurate and efficient simulation. QCA constitute a natural theoretical setting forthis purpose,inparticular viaQuantum Lattice-Gas Automata [9–13]. • CAperspective. By their definition, CA are shift-invariant and causal. CA are therefore a physics-like model of computa- tion(atermcoinedbyMargolus[14]),astheysharesomefundamentalsymmetriesoftheoreticalphysics:homogeneity (invariance of physical laws in time and space), causality, and (often) reversibility. Thus it is natural, following Margo- lus[15],tostudy theirquantum extensions. • Modelsofcomputationperspective.Therearemany models ofdistributed computation (e.g.CCS, π-calculus), but oftenin such models the idea of space is not directly related to our general understanding, such as our intuitive understanding of relative positions of objects in 3D space. These models are not adequate for reasoning about simple space-sensitive synchronisation problems, such as ‘machine self-reproduction’ [16,1] or the ‘Firing Squad’ problem [17,18]. In contrast, CA were initially used to model spatially distributed computation in space [19]. Moreover, QCA provide a model of QC, andhence constitute a framework tomodel and reasonabout problemsinspatially distributed QC. • Theoreticalphysicsperspective. QCA could provide helpful toy models for theoretical physics [20]. For this purpose it couldbuildbridgesbetweencomputerscienceandtheoreticalphysics,asthepresentpaperattemptstofortheconcept ofuniversality. These motivations demonstrate the importance of studying QCA. Once this is acknowledged researchers are faced with an overabundance of competing definitions. There are four main approaches to defining QCA: the axiomatic style[21–23],themultilayerblockrepresentation[23,24],thetwo-layerblockrepresentation[4,6,21,25–27],andPartitioned QCA(PQCA)[27–29].Anatural first questions toconsider iswhether theyareequivalent, andinwhat sense. 1.2. QCA:Simulation Probably the most well-known CA is Conway’s ‘Game of Life’; a two-dimensional CA which has been shown to be universal for computation in the sense that any Turing Machine (TM) can be encoded within its initial state and then executed by evolution of the CA [30]. As TM are generally considered to be a robust definition of ‘what an algorithm is’ in classical computer science, this result could be perceived as providing a conclusion to the topic of CA simulation. However, this is not the case, as CA do more than just running any algorithm. They run distributed algorithms in a distributed manner, model phenomena together with their spatial structure, and allow the use of the spatial parallelism inherent in the model. These features, modelled by CA and not by TM, are all interesting, and so the concepts of simulation and universality needed be revisited in this context to account for space. This has been done by returning to the original meaning of the word simulation [31–33], namely the ability for one instance of a computational model to simulate other instances of the same computational model. The introduction of a partial order on CA via groupings [34], and subsequent generalisations [35,36], have led to elegant and robust definitions of intrinsic simulation. Intrinsic simulation formalises the ability of a CA to simulate another in a space-preserving manner. Intuitively this is exactly what is needed to show the equivalence between the various competing definitions of QCA, i.e. that they can all simulate each other in a space- preservingmanner. Thedefinition ofintrinsicsimulation hasalreadybeen translated inthequantum context[37],however as it stands this is not sufficient to obtain the desired result. In this paper the definition of intrinsic simulation in the quantum context is discussed and developed, before the equivalence between all the various above-mentioned definitions ofQCAistackled. 1.3. QCA:Simplification Intrinsic universality is the ability to intrinsically simulate any other QCA. Here we show that the axiomatic style QCA, themultilayerblockrepresentationQCA,thetwo-layerblockrepresentationQCA,andthePQCAareequivalent,entailingthat PQCA are intrinsically universal. Here the PQCA is chosen as the prime model as it is the simplest way to describe a QCA. Therefore, the result developed in this work is also a simplifying one for the field of QCA as a whole. From a theoretical physics perspective, showing that ‘Partitioned Quantum Cellular Automata are universal’ is a statement that ‘scattering phenomena areuniversal physical phenomena’. There are several related results in the CA literature. Several influential works by Morita et al. emphasise Reversible PartitionedCAuniversality.Forinstance,theyprovidecomputationuniversalReversiblePartitionedCAconstructions[38,39], and their ability to simulate any CA in the one-dimensional case is also shown [40]. The problem of intrinsically universal Reversible CA (RCA) constructions was tackled by Durand-Lose [41,42]. The difficulty is in having an n-dimensional RCA simulate all other n-dimensional RCA and not, say, the (n−1)-dimensional RCA, otherwise a history-keeping dimension couldbeused, asinToffoli [43].Stronglyrelatedtothisisthe workon block representationsofRCAbyKari[44]. 1.4. QCA:Universality The QCA-related results are focused on universality. Watrous [28] proved that QCA are universal in the sense of QTM. Shepherd, Franz and Werner [45] defined a class of QCA where the scattering unitary Ui changes at each step i (classical P.Arrighi,J.Grattage/JournalofComputerandSystemSciences78(2012)1883–1898 1885 controlQCA).Universalityinthequantumcircuit-sensehasalreadybeenachievedbyVanDam[27],CiracandVollbrecht[8], Nagaj and Wocjan [6], and Raussendorf [26]. In the bounded-size configurations case, circuit universality coincides with intrinsic universality, as noted by Van Dam [27]. Intrinsically universal QCA in the one-dimensional case have also been resolved[37].GiventhecrucialroleofthisinclassicalCAtheory[46],theissueofintrinsicuniversalityinthen-dimensional caseneededtobeaddressed.HavingthenshownthatPQCA,asimplesubclassofQCA,areintrinsicallyuniversal,itremained to show that there existed an n-dimensional PQCA capable of simulating all other n-dimensional PQCA for n>1 to some arbitraryprecision,which weshowinthispaper. PQCA are QCA of a particular form, where incoming information is scattered by a fixed unitary U before being redis- tributed, and this paper shows PQCA to be intrinsically universal. Hence the problem of finding an intrinsically universal PQCAreduces to finding some universal scattering unitary U (as made formal in Section 5.2, Fig. 9). Also, the requirements on U are much more stringent than just quantum circuit universality, as the simulation of a QCA H has to be done in aparallel, space-preserving manner. Moreover,not only asingle iteration of H has to be simulated, but several (H2,...), so that after every simulation the universal PQCA is ready for a further iteration. From a computer architecture point of view, this problem can be recast in terms of finding some fundamental quantum processing unit which is capable of simulat- ing any other network of quantum processing units, in a space-preserving manner. From a theoretical physics perspective, this amounts to specifying a scattering phenomenon that is capable of simulating any other, again in a space-preserving manner. These difficulties can be overcome. A key result shown here is the construction of a simple intrinsically universal n-dimensional QCA. 1.5. Layout The necessary theoretical background for understanding QCA, and hence the problems addressed by this paper, is pro- vided in Section 2. Intrinsic simulation is discussed and generalised in Section 3. In Section 4 the various alternative definitions of QCA are shown to be equivalent to the simplest definition, i.e. PQCA. In Section 5 a simple example of an intrinsically universal PQCA is developed. Section 6 concludes with summary and discussion with ideas for future direc- tions. Thispaperalso integratesthecontributionsoftwo already-published conference papers [47,48]. 2. Definitions 2.1. n-DimensionalQCA This section provides the axiomatic style definitions for n-dimensional QCA. Configurations hold the basic states of an entirearray ofcells, andhence denote thepossible basic statesoftheentireQCA: Definition1(Finiteconfigurations).A(finite)configurationcoverΣ isafunctionc:Zn→Σ,with(i1,...,in)(cid:4)→c(i1,...,in)= cqiu1i..e.isnc,esnutcshtattheaotfthΣer.eThexeisstestaof(paollssfiibnliyteecmopntfiyg)ufirantiitoenssetovIesraΣtisfwyiinllgb(ei1d,e.n..o,tiend)∈C/ΣI .⇒ci1...in=q,whereq isadistinguished fin SincethisworkrelatestoQCAratherthanCA,theglobalstateofaQCAcanbeasuperpositionoftheseconfigurations.To construct the separable Hilbert space of superpositions of configurations the set of configurations must be countable. Thus finite, unbounded, configurations are considered. The quiescent state of a CA is analogous to the blank symbol of a Turing machine tape;it symbolises theemptyspace, andlocally quiescent cellsremainquiescent. Definition2 (Superpositionsofconfigurations). Let HCΣ be the Hilbert space of configurations. Each finite configuration c fin is associated with a unit vector |c(cid:7), such that the family (|c(cid:7))c∈CΣ is an orthonormal basis of HCΣ. A superpositionof fin fin configurationsisthenaunitvectorin HCΣ. fin Definition3 (Unitarity). A linear operator G:HCΣ →HCΣ is unitary if and only if {G|c(cid:7)|c∈CΣfin} is an orthonormal basis fin fin of HCΣ. fin Definition4(Shift-invariance).Considertheshiftoperation,fork∈{1,...,n},whichtakesconfiguration c toc(cid:8) whereforall (i1,...,in) we have c(cid:8)i1...ik...in=ci1...ik+1...in. Let σk:HCΣfin→HCΣfin denote its linear extension to a superpositions of configu- rations. Alinear operator G:HCΣ →HCΣ issaidtobe shift-invariantifandonly if Gσk=σkG foreachk. fin fin Thefollowingdefinitioncapturesthecausalityofthedynamics.Imposingtheconditionthatthestateassociatedtoacell (its reduced density matrix) is a function of the neighbouring cells is equivalent to stating that information propagates at abounded speed. 1886 P.Arrighi,J.Grattage/JournalofComputerandSystemSciences78(2012)1883–1898 Definition5 (Causality). A linear operator G:HCΣ →HCΣ is said to be causal if and only if for any (i1,...,in)∈Zn, there existsafunction f suchthat ρ(cid:8)|N= f(ρ|N) forfianll ρ ovefirnHCΣ,where: fin N={i1,i1+1}×···×{in,in+1}, ρ|N means the restriction of ρ to the neighbourhood N in the sense of the partial trace,and ρ(cid:8)=GρG†. In the classical case, the definition is that the letter to be read in some given cell i at time t+1 depends only on the state of the cells i to i+1 at time t. Transposed to a quantum setting, the above definition is obtained. To know the state ofcell number i,onlythe states of cells i and i+1 beforethe evolutionneedbe known. More precisely, this restrictive definition of causality is known in the classical case as a 1-neighbourhood cellular au- 2 tomaton, because the most natural way to represent such an automaton is to shift the cells by 1 at each step, so that 2 visuallythestate ofacelldepends onthestateofthetwocellsunderit.Thisdefinitionofcausality isnot restrictive,as by grouping cells into “supercells” any CA with an arbitrary finite neighbourhood N can be made into a 1-neighbourhood CA. 2 The same method can be applied to QCA, so this definition of causality holds without loss of generality. However, the f in the above definition does not directly lead to a constructive definition of a cellular automaton, unlike the local transition function inthe classical case [23]. Thisapproachleadstothefollowingdefinitionofann-dimensionalQCA.Ithasbeengivenpreviously[22,23],butclearly ∗ stemsfromanequivalent definition inthe literature,phrased intermsofhomomorphism ofa C -algebra [21]. Definition6 (QCA). An n-dimensional quantum cellular automaton (QCA) is an operator G:HCΣ →HCΣ which is unitary, fin fin shift-invariant andcausal. Whilst this is clearly the natural, axiomatic definition QCA, it remains a non-constructive one. In this sense it can be compared to the Curtis–Hedlund [49] definition of CA as the set of continuous, shift-invariant functions. These definitions characterise (Q)CA via the global, composable properties that they must have; but they do not provide an operational, hands-on descriptionoftheirdynamics. 2.2. Multilayerblockrepresentation What is meant by an operational description of a QCA? A central tool and concept in this paper is that of a (multilayer) block representation of QCA. Intuitively, we say that a QCA G admits a block representation when it can be expressed as blocks,i.e.localunitaries,composedinspace(viathetensorproduct)andtime(viaoperatorcomposition),therebyforming a finite-depth quantum circuit infinitely repeating across space. The structure theorem given in previous work [23] states that anyQCAcaninfact berepresentedinsuchaway: Theorem 1 (n-Dimensional QCA multilayer block representation). Let G be ann-dimensional QCA with alphabet Σ. Let E be an isometryfrom HΣ →HΣ ⊗HΣ suchthat E|ψx(cid:7)=|q(cid:7)⊗|ψx(cid:7).Thismappingcanbetriviallyextendedtowholeconfigurations, yieldingamappingE:HCfiΣn→HCfiΣn2.Therethenexistsann-dimensionalQCAHonalphabetΣ2,suchthatHE=EG,andHadmits a2n-layerblockrepresentation.MoreoverHisoftheform (cid:2)(cid:3) (cid:4)(cid:2)(cid:5) (cid:4) H= S Kx (1) where: • (Kx)isacollectionofcommutingunitaryoperatorsallidenticaluptoshift,eachlocaliseduponeachneighbourhoodNx; • SistheswapgateoverHΣ⊗HΣ,hencelocaliseduponeachnodex. This theorem therefore bridges the gap between the axiomatic style definition of QCA and the operational descriptions of QCA. Again, it should be compared with the Curtis–Hedlund [49] theorem, which shows the equivalence between the axiomaticdefinitionofCAandthemoreoperational,standarddefinition,withalocalfunctionappliedsynchronouslyacross space.OnecanarguethattheformgiveninEq.(1)isnotthatsimple.Acontributionofthispaperistosimplifyitdownto PQCA. Amongst the operational definitions of QCA listed in Section 1, only that of Pérez-Delgado and Cheung [24] is not two- layer.Theydirectlystate, aftersome interestinginformalarguments, that QCAareofaformsimilar tothatgiveninEq.(1). In other words, this theorem demonstrates that starting from an axiomatic definition of QCA, such as Schumacher and Werner’s[21],onecanderiveacircuit-likestructureforn-dimensionalQCA,therebyextendingtheirresultton dimensions. It also demonstrates that operational definitions [24] can be given a rigorous axiomatics. These factors demonstrate that the definitions of Pérez-Delgado and Cheung [24] and Schumacher and Werner [21] are actually equivalent, up to ancillary cells. P.Arrighi,J.Grattage/JournalofComputerandSystemSciences78(2012)1883–1898 1887 This shows that the axiomatic definition of QCA given in Section 2.1 is equivalent to a multilayer block representation. Thereare,however,severalotherdefinitionsofQCA,i.e.two-layerblockrepresentationsandPQCA.Theaimistonowshow that all definitions of QCA can be reconciled via intrinsic simulation. A quantum version of intrinsic simulation has already beendeveloped[37],butonlyforone-dimensionalQCA,anditisnotgeneralenoughtostatetherequiredequivalence.This difficulty is addressed in the next section, where a new concept of intrinsic simulation forn-dimensional QCA is developed withthe requiredproperties. 3. Intrinsicsimulationofn-dimensionalQCA IntrinsicsimulationofoneCAbyanotherwasdiscussedinformallyinSection1.2.Apedagogicaldiscussionintheclassical case was given by Ollinger [35], and quantised intrinsic simulation has been formalised in the one-dimensional case [37]. This definition is extended to n dimensions (and relaxed, see details below) here. The potential use of this concept in theoreticalphysicsisalso discussed. Intuitively, ‘G simulates H’ is shown by translating the contents of each cell of H into cells of G, running G, and then reversing the translation; this three step process amounts to running H. This translation should be simple (it should not provide a “hidden” way to compute G), should preserve the topology (each cell of H is encoded into cells of G in a way which preserves neighbours), and should be faithful (no information should be lost in translation). This latter requirement relates to the isometry property of quantum theory, i.e. an inner product preserving evolution with Enc†Enc=I. This same requirementagreeswiththe translation being a physical process.Thefollowing definitionsarethusderived. Definition7 (Isometriccoding). Consider ΣG and ΣH, two alphabets with distinguished quiescent states qG and qH, and such that |ΣH|(cid:2)|ΣG|. Consider HΣG and HΣH the Hilbert spaces having these alphabets as their basis, and HCfiGn, HCfiHn theHilbert spaces offiniteconfigurations overthese alphabets. exteLnetdsEinbteoaanniissoommeettrriicclliinneeaarrmmaappfErnomc=H((cid:6)ΣHZntoE)HfΣroGmwHhiCchH pinretsoeHrveCsGq,uwiehsiccehnwcee, ci.ael.lsaunchistohmatetEri|cqeHn(cid:7)c=od|qinGg(cid:7).. It trivially fin fin |qHL(cid:7)e⊗t |DqGb(cid:7)e.IatntriisvoimalleytreicxtleinnedasrinmtoapanfroismomHeΣtrGictloinHeaΣrHm⊗apHDΣeGcw=h(i(cid:6)chZanlsDo)pfrreosmerHveCsGquiinetsoceHncCeH,i⊗nHthCeGse,nwsheicthhawt eDc|qaGll(cid:7)a=n fin fin fin isometricdecoding. Theisometries E and D define an isometriccoding ifthe following condition issatisfied: (cid:7) (cid:8) ∀|ψ(cid:7)∈HCH, ∃|φ(cid:7)∈HCG / |ψ(cid:7)⊗|φ(cid:7)=Dec Enc|ψ(cid:7) . fin fin (Here Dec isunderstood tomorally bean inversefunction of Enc,butsome garbage |φ(cid:7) may beomitted.) Definition8 (Directsimulation). Consider ΣG and ΣH, two alphabets with distinguished quiescent states qG and qH, and two QCA G and H over these alphabets. We say that G directly simulates H, if and only if there exists an isometric coding suchthat (cid:7) (cid:8) (cid:7) (cid:7) (cid:8)(cid:8) ∀i∈N, ∀|ψ(cid:7)∈HCH, ∃|φ(cid:7)∈HCG / Gi|ψ(cid:7) ⊗|φ(cid:7)=Dec Hi Enc|ψ(cid:7) . fin fin Unfortunately this is not enough for intrinsic simulation, as it implies that |ΣH|=|ΣG|. It is often desirable that G simulates H eventhough thetranslation: – takes severalcells of H intoseveralcellsof G; – demands severalstepsof G inordertosimulate severalstepsof H. Hencethe groupingofcellsisrequired. Definition9(Grouping).LetG beann-dimensionalQCAoveralphabetΣ.Letsandt betwointegers,q(cid:8) awordinΣ(cid:8)=Σsn. Consider the iterate global evolution Gt up to a grouping of each hypercube of sn adjacent cells into one supercell. If this operator can be considered to be a QCA G(cid:8) over Σ(cid:8) with quiescent symbol q(cid:8), then we say that G(cid:8) is an (s,t,q(cid:8))-grouping of G. A natural way to continue would be to define an intrinsically universal QCA. However, due to the continuity of H, this approximationcanonly be upto (cid:8). InSection5we provideauniversalQCAwithaboundonthefiniteerror. Definition10(Intrinsicsimulation).Consider ΣG and ΣH,twoalphabetswithdistinguished quiescent statesqG andqH,and (cid:8) twoQCAG and H overthesealphabets.WesaythatG intrinsicallysimulates H ifandonlyifthereexistsG ,somegrouping (cid:8) (cid:8) (cid:8) of G,and H ,somegroupingof H,suchthat G directlysimulates H . 1888 P.Arrighi,J.Grattage/JournalofComputerandSystemSciences78(2012)1883–1898 Fig.1.Theconceptofintrinsicsimulationmadeformal. In other words, G intrinsically simulates H if and only if there exists some isometry E which translates supercells of H intosupercellsof G,suchthatif G isiteratedandthentranslatedback,thewholeprocessisequivalenttoaniterationof H. Thisunderstanding isshown schematically inFig. 1. Comparedwithpreviouswork[37],theconceptofintrinsicsimulationhasbeenmodifiedtoallowthegroupinginFig.1 onthesimulated QCAside, and thisvariationisimportant toTheorem3.Thisisanalogous totheclassical case [36]. A natural way to follow would be to define the notion of an intrinsically universal QCA. However due to the continuous nature of the underlying Hilbert spaces, no QCA can be intrinsically universal in an exact sense. We can only hope to have a ‘dense’ QCA, i.e. one which can simulate any other up to some precision (cid:8), which can then be made arbitrarily small. In Section5we providesuch a construction, togetherwithbounds on (cid:8). The study of QC aims to address the issues related to the physical nature of computing, and over the last twenty years there have been a number of quantisations of the classical models of computation, and novel results on the complexity of the tasks that can be encoded in these models. It could be said that theoretical physics has aided theoretical computer science via this path. Within this context, it is likely that the reverse path could also be productive. This would be part of a bigger trend where theoretical physics departs from looking at ‘matter’ (particles interacting, scattering, forces, etc.) and seeks to look at ‘information’ (entropy, observation, information exchanges between systems, etc.), in an attempt to clarify its own concepts. An example of this is the huge impact that quantum information theory has had on the understanding of foundational concepts such as entanglement [50] and decoherence [51]. A computer science based approach can help to understand physical principles, not only in terms of ‘information’, but also in terms of the ‘dynamics of information’, i.e. information processing. Looking at computer science, a fundamental concept in computation theory is universality. An instance of a model of computation is universal if it can simulate any other; this would also be a useful concept in physics. For example, if trying to reconcile two rather different mechanics (quantum theory and general relativity, say), finding such a minimal,universal physicalphenomenon would provide something simple to frame, so that the focus can be on reconciling the mechanics, while rich enough to guarantee that some arbitrarily complex phenomenon can be incorporated into this reconciled me- chanics. However,thefollowing mustbe considered: • Firstly, a universal TM should be able to simulate each object independently in its own space. The universal physical phenomenon should be some elementaryunit of computation that can be combined to form a 3D network, accounting forspace and interactionsacross space satisfactorily. • Secondly, the universal TM is slow at simulating quantum physical phenomena, which suggests that it is not rich enough. The universal physical phenomenon should therefore be a universal model of quantum computation, which accounts forthecost ofsimulation. The work that has been presented in this section formalises an idea of universality which fits both these criteria, namely intrinsicuniversality overQCA. 4. Constructions Now that an appropriate notion of intrinsic simulation has been developed, the problem of showing an equivalence between thedifferent operational definitions ofQCAisaddressed here. 4.1. Downtotwolayers:BlockQCA Quantisations of block representations of CA are generally presented as two-layer; cf. [4,25,6,26,21,27]. This is captured bythe definitionofaBlock QCA(BQCA),where H⊗2n is H⊗···⊗H,repeated 2n times: P.Arrighi,J.Grattage/JournalofComputerandSystemSciences78(2012)1883–1898 1889 Fig.2.BQCA.TheelementaryunitaryevolutionsU0andU1arealternatedrepeatedlyasshown,in1D. Fig.3.SketchofaBQCAsimulatingaQCA.Theoriginalcellxiscodedintofourcells,atthecentre(E).ItstartsbyconsideringtheNorth–Westasattime0 itwillcomputeitsNorth–Westsuccessor,andthenmoveclockwise.Attime1itwillcomputeitsNorth–Eastsuccessor,etc. Definition11 (BQCA). A block n-dimensional quantum cellular automaton (BQCA) is defined by two unitary operators U0 and U1 such that Ui:HΣ⊗2(cid:6)n →HΣ⊗2n, and Ui|qq...qq(cid:7)=|qq...qq(cid:7), i.e. each takes 2n cells into 2n cells and preserves quiescence. Consider Gi=( 2ZnUi) the operator over H. The induced global evolution is G0 at odd time steps, and σG1 ateventime steps,where σ isa translation byonein all directions(Fig. 2). ShowingtheequivalenceoftheQCAandBQCAaxiomaticsisnottrivial.Inonedirectionthisissimple,asBQCAareuni- tary,causal,andshift-invariant,andhencefallundertheaxiomaticsandTheorem1(strictlyspeakingweneedtogroupeach hypercube of 2n adjacent cells into a supercell, see Definition 9). However, there are several factors to consider regarding theability ofBQCA tosimulate any QCA,which arenowaddressed. In the form given by Theorem 1, each cell x at time t is successively involved in 2n computations governed by a local unitary K,whoseaimistocomputethenextstateofacellwithinaradius 1 fromxattimet+1.Intwodimensions,acellx 2 usesthecellsWest,North–WestandNorthtoworkoutitsNorth–Westsuccessor,andthenthecellsNorth,North–East,East ofittocomputetheNorth–Eastsuccessor(similarlyfortheSouth–EastandtheSouth–Westsuccessors).Tomimicthiswith a BQCA, each original cell can be encoded into four cells, arranged so that the original cell x starts in the North–West quadrantofthefourcells.ThefirstlayeroftheBQCAappliesthelocalunitary K tocomputetheNorth–Westsuccessorof x. The second layer of the BQCA moves the orig(cid:6)inal cell x in the North–West quadrant. Each full application of the evolution of the BQCA corresponds only to one layer ( K), hence it will take four steps for this BQCA to simulate one step of the QCA.Fig. 3showsasketchofthemethodused. There are some considerations to be discussed. When cell x is turning clockwise in the example, the cell to its North is turning anticlockwise. Hence we need some ancillary data coding for the path to be taken by the original cell x within the four coding cells. Also, Theorem 1 finishes with a Swap between the ‘computed tape’, where the results have been stored, and the ‘uncomputed tape’ (i.e. what remains of the original cell after having computed all of its successors) which is not shown in the sketch. Hence the number of layers of K computed so far has to be tracked, so that the Swap occurs at the appropriate step. The Swap also needs to know where the results have been stored in order to move them correctly. All of this has tobe arranged spatially and efficiently,and onesuch method isshown inFigs.4and 5. BQCA can therefore simulate QCA up to a relatively simple encoding, using blocks of four cells. This explains the need for grouping on the simulated QCA side in the revised quantised intrinsic simulation, as in Fig. 1. Encoding groups of cells rather than individual cells is also required for the PQCA discussion (videinfra). This encoding is given for two dimensions, but the construct clearly generalises to n dimensions. Hence QCA (Definition 6) provide a rigorous axiomatics for BQCA (Definition 11), andBQCAprovideaconvenient operationaldescriptionofQCA.We haveshownthat: Theorem2(BQCAareuniversal).Givenanyn-dimensionalQCAH,thereexistsann-dimensionalBQCAGwhichsimulatesH. 1890 P.Arrighi,J.Grattage/JournalofComputerandSystemSciences78(2012)1883–1898 Fig.4.BQCAsimulatingaQCA.Thegreyareasdenotetheneighbourhoodwheretheactionofkx,thefirstlayeroftheBQCA,willbesignificant–i.e.agroup offourcellswhereitwillperformaKxoperationtoworkoutasuccessor.Wherethissuccessorwillbestoredisindicatedby(Rx).AtthenextstepRxhas appeared,andtheregistershavebeenreshuffledduetothesecondlayeroftheBQCA,whichactsaccordingtotherotation-directionmark.Thesecondlayer alsoincreasestheclockcountandincludesthefinalswappingstep,whichonlyhappensattime3.Thereitensuresthat R0 becomes A, R1 becomes B, etc.Whichregistersaretobeswappedwithoneanothercanbecalculatedfromtherotationandarrowmarks.EachstepismadeformalbyFig.5. Fig.5.OperationsusedinFig.4.kappliesa K operationwheneversomedataispresent(datacarriesanextrabittodistinguishitfrom|q(cid:7),say).TheU operationreshufflesthedatabyrotatingitinthedirectiongivenbytheindicatorinthetopleft(clockwiseoranticlockwise),andincrementstheindex counter.Finally,cswapactsastheidentityinallcasesexceptwhentheindexis3,whenitswapstheresultofthecomputationswiththedata,readyfor thenextround. 4.2. Downtoonescatteringunitary:PQCA Quantisations of partitioned representations of CA are given in several works [29,27,28]. These constitute the simplest approach to defining QCA. It is therefore interesting to consider whether QCA (as in Definition 6) provide a rigorous ax- iomatics for PQCA, and if PQCA provide a convenient operational description of QCA. A PQCA is essentially a BQCA where thetwo layersapplythesame unitaryoperation,shifted appropriately. Definition12 (PQCA). A partitioned n-dimensional quantum cellular automaton (PQCA) is defined by a scattering unitary operator U suchthat U:HΣ⊗2n→HΣ⊗2n,and U|qq.(cid:6)..qq(cid:7)=|qq...qq(cid:7),i.e.thattakesahypercubeof2n cellsintoahypercube of 2n cellsandpreservequiescence. Consider G=( 2ZnU),theoperatorover H.Theinduced global evolution is G atodd timesteps, and σG at eventimesteps, where σ isa translation byoneinall directions(Fig. 6). Following previous results (Section 4.1), it is only necessary to show that PQCA can simulate BQCA. Both PQCA and BQCA are two-layer; the only difference is that for BQCA those two layers may be different (e.g. compare Figs. 6 and 2), whereasforPQCAthereisonlyasinglescatteringunitary.SoaU-definedPQCA,withaU capableofperformingU0 and U1 P.Arrighi,J.Grattage/JournalofComputerandSystemSciences78(2012)1883–1898 1891 Fig.6.Partitionedone-dimensionalQCAwithscatteringunitaryU.Eachlinerepresentsaquantumsystem,inthiscaseawholecell.Eachsquarerepresents ascatteringunitaryU whichisappliedtotwocells.Timeflowsupwards. Fig.7.PQCAsimulatingaBQCA.TheQCAisdecoratedwithcontrolqubitsfollowingasimpleencodingprocedure(left),whichallowthescatteringunitaryU (centre)toactaseitherU0orU1,accordingtothelayer(right).Theblackboxcanbeanyunitary. alternativelyascontrolledbysomeancillarysuffices.Thishasbeenshownforonedimension[37]andisgivenherefortwo dimensions inFig. 7.It isclearthat theconstruct givenheregeneraliseston dimensions. Theorem3(PQCAareuniversal).Givenanyn-dimensionalQCAH,thereexistsann-dimensionalPQCAGwhichsimulatesH. ThereforeitcanbeconcludedthatPQCAarethemostcanonicalandgeneraloperationaldescriptionofQCA.Moregener- ally, by showing here that the various definitions of QCA available [52,27,4,6,26,25,29,24] are equivalent, this demonstrates that awell-axiomatised, concrete,andoperationaln-dimensional QCAisnowavailable. 5. AnintrinsicallyuniversalQCA In Section 2.1 the formal definition of n-dimensional PQCA was discussed (Fig. 6), and the formal definition of intrinsic simulation was recalled (Fig. 15). The aim now is to find a particular U-defined PQCA which is capable of intrinsically simulatingany V-definedPQCA,forany V.Inordertodescribesucha U-definedPQCAindetail,twothingsmustbegiven: the dimensionality of the cells (including the meaning attached to each of the states they may take), and the way the scattering unitary U acts upon these cells. First we discuss the general scheme used to solve this problem, and then we describe thePQCAimplementing it. 5.1. Circuituniversalityversusintrinsicuniversalityinhigherdimensions As already discussed, intrinsic universality refers to the ability for one CA to simulate any other CA in a way which preserves the spatial structure of the simulated CA. Conversely, computation universality refers to the ability of a CA to simulate any TM, and hence run any algorithm. Additionally, circuit universality is the ability of one CA to simulate any circuit. These are Nand gate circuits for classical circuits and classical CA, and Toffoli gate circuits for reversible circuits and reversible CA. Informally, in a quantum setting, circuit universality is the ability of a PQCA to simulate any unitary evolution expressed as a combination of a universal set of quantum gates, such as the standard gate set: Cnot,R(π) (also knownasthe π gate),andtheHadamardgate.TherelationshipsbetweenthesethreeconceptsofCAuniversalityha4vebeen 8 notedpreviously[33].AcomputationuniversalCAisalsoacircuituniversalCA,becausecircuitsarefinitarycomputations.In addition,anintrinsicuniversalCAisalsoacomputationuniversalCA,becauseitcansimulateanyCA,includingcomputation universalCA.Hence intrinsicuniversality impliescomputation universality,which impliescircuituniversality. In onedimension thisisnot an equivalence. Intuitively, computation universality requiresmorethan circuituniversality, namely the ability to loop the computation, which is not trivial for CA. Similarly, intrinsic universality requires more than computation universality, such as the ability to simulate multiple communicating TM. In the classical setting there are formalresultsthat distinguish these ideas [35]. In n dimensions, it is often assumed in the classical CA literature that circuit universality implies intrinsic universal- ity, and hence that these are equivalent [35]. Strictly speaking this is not true. Consider a two-dimensional CA which runs one-dimensionalCAinparallel.Iftheone-dimensionalCAiscircuit/computationuniversal,butnotcomputation/intrinsically 1892 P.Arrighi,J.Grattage/JournalofComputerandSystemSciences78(2012)1883–1898 Fig.8.FlatteningaPQCAintoasimulatingPQCA.(a):Considerfourcells(white,lightgrey,darkgrey,black)ofaPQCAhavingscatteringunitaryV.Thefirst layerPQCAapplies V tothesefourcells,thenthesecondlayerapplies V atthefourcorners.(b):Weneedtoflattenthissothatthetwolayersbecome non-overlapping.Thefirstlayercorrespondstothecentresquare,andthesecondlayertothefourcornersquares.Atthebeginningthesignals(white,light grey,darkgrey,black)codingforthesimulatedcellsareinthecentresquare.TheyundergoV,andaredirectedtowardsthebottomleft,topleft,bottom right,andtoprightsquaresrespectively,wheretheyundergoV butpairedupwithsomeothersignals,etc. Fig.9.FlatteningaPQCAintoasimulatingPQCA(cont’d).(c):Withinthecentralsquaretheincomingsignalsarebunchedtogethersoastoundergoa quantumcircuitwhichimplements V,andarethendispatchedtowardsthefourcorners.Thisdiagramdoesnotmakeexplicitanumberofsignaldelays, whichmaybeneededtoensurethattheyarrivesynchronouslyatthebeginningofthecircuitimplementingV.(d):Withinthecentralrectangle,thecircuit whichimplementsV isitselfacombinationofsmallercircuitsforimplementingauniversalsetofquantumgatessuchasCnot,HadamardandtheR(π), 4 togetherwithdelays.TheseareimplementedasexplainedinSections5.3and5.4. universal, then this is also true for the two-dimensional CA. Similarly, in the PQCA setting, the two-dimensional construc- tionsin [24]and [26] arecircuituniversal but not intrinsically universal. However, this remains a useful intuition: Indeed, CA admit a block representation, where these blocks are permutations forreversibleCA,whileforPQCAtheblocksareunitarymatrices.Thustheevolutionofany(reversible/quantum)CAcanbe expressed as an infinite (reversible/quantum) circuit of (reversible/quantum) gates repeating across space. If a CA is circuit universal, and if it is possible to wire together different circuit components in different regions of space, then the CA can simulate the block representation of any CA, and hence can simulate any CA in a way which preserves its spatial structure. It is intrinsically universal. This is the route followed next in constructing the intrinsically universal n-dimensional PQCA. Firsttheconstructionofthe‘wires’,whichcancarryinformationacrossdifferentregionsofspace,isconsidered.Herethese are signals which can be redirected or delayed using barriers, with each signal holding a qubit of information. Secondly, the ‘circuit-pieces’ are constructed, by implementing quantum gates which can be combined. One and two qubit gates are implemented asobstacles to,andinteractionsof,these signals. 5.2. FlatteningaPQCAintospace In the classical CA literature it is considered enough to show that the CA implements some wires carrying signals, and some universal gates acting upon them, to prove that an n-dimensional CA is in fact intrinsically universal. Any CA can be encoded into a ‘wire and gates’ arrangement following the above argument, but this has never been made explicit in the literature. This section makes more precise how to flatten any PQCA in space, so that it is simulated by a PQCA which implements quantum wires and universal quantum gates. Flattening a PQCA means that the infinitely repeating, two-layered circuit is arranged in space so that at the beginning all the signals carrying qubits find themselves in circuit- pieces which implement a scattering unitary of the first layer, and then all synchronously exit and travel to circuit-pieces implementing the scattering unitary of the second layer, etc. An algorithm for performing this flattening can be provided, however the process will not be described in detail, for clarity and to follow the classical literature, which largely ignores thisprocess. Theflattening processcanbe expressedinthreesteps: Firstly, the V-definedPQCAisexpandedinspace bycoding each cell into a hypercube of 2n cells. This allows enough space for the scattering unitary V to be applied on non-overlapping hypercubes of cells, illustrated in the two-dimensional case in Fig. 8(a). Secondly, the hypercubes where V is applied must be connected with wires, as shown in Fig. 8(b). Within these hypercubes wiring is required so that incoming signals are bunched together to undergo a circuit implementation of V, and are then dispatched appropriately, as shown in Fig. 9(c). This requires both time and space expansions, with factors that depend non-trivially (but uninterestingly) upon the size of the circuit implementation of V and the way the wiring and gates work in the simulating PQCA. Next, an encoding of the