An overview of Quantum Cellular Automata
Abstract
Quantum cellular automata consist in arrays of identical finitedimensional quantum systems, evolving in discretetime steps by iterating a unitary operator . Moreover the global evolution is required to be causal (it propagates information at a bounded speed) and translationinvariant (it acts everywhere the same). Quantum cellular automata provide a model/architecture for distributed quantum computation. More generally, they encompass most of discretespace discretetime quantum theory. We give an overview of their theory, with particular focus on structure results; computability and universality results; and quantum simulation results.
This paper is dedicated to my high school Mathematics teacher, Anne Lefèvre.
1 Introduction
Von Neumann provided the modern axiomatisation of quantum theory in terms of the density matrix formalism NeumannQT in 1955. He also invented the cellular automata (CA) model of computation Neumann in 1966, but never brought the two together. Feynman suggested doing so FeynmanQC; FeynmanQCA in 1986, just as he was inventing the very concept of quantum computation (QC).
Indeed, confronted with the inefficiency of classical computers for simulating quantum physics, Feynman realized that one ought to use quantum devices instead FeynmanQC. What better than a quantum system in order to simulate another quantum system? Soon afterwards FeynmanQCA he introduced Quantum Cellular Automata (QCA) for two reasons. First because they constituted a promising architecture for the implementation quantum simulation devices—as demonstrated nowadays with cold atoms on optical lattices, integrated quantum optics or superconducting qubits. Second, because the quantum simulation of a quantum physical phenomena requires that we are able to describe it “in terms of qubits”. Most often, this qubit description is obtained by formulating a discretespace discretetime version of the original continuous description of the phenomena—i.e. a QCA model for it.
Notice that your usual, numerical simulation of the phenomena on a classical computer, would also require that the phenomena be described “in terms of bits”. But these numerical schemes are not usually thought of as being physically legitimate themselves, because they tend to be unaesthetic or worse break fundamental symmetries. For instance, applying a finitedifference methods upon the partial differential equation governing the propagation of a particle, will typically break unitarity, making it unphysical. A QCA model, on the other hand, has to remain physical, and unitary (with some work it may even retain Lorentzcovariance). In this sense, QCA models may be thought of as constituting physically legitimate descriptions of the phenomena, themselves. Moreover some are way simpler, and explanatory than the original continuous description, as we shall see. Thus, the provision of toy models for theoretical physics is another, strong reason to study QCA.
Yet another strong reason lies at the heart of Theoretical Computer Science with the basic question: What is a computer, ultimately? Which key resources are granted to us by nature, for the sake of computing? Tentative answers to these questions have been obtained by abstracting away from particles and forces, to reach formal models of computations—such as the Turing machine. Turing machines used to be our best answer. Nowadays, however, spatial and quantum parallelism need be taken into account, leading us to propose models of distributed quantum computation. Amongst the different models of distributed QC, QCA are of the most established. Just like CA, QCA account for space “as we know it” (i.e. mostly euclidean). Thus, they constitute a framework to model and reason about problems in spatiallysensitive distributed Quantum Computation MazoyerFiring. For instance we may wonder, as we shall see, whether there exists an intrinsically universal QCA, i.e. capable of simulating all others in a spacetimepreserving manner; or whether QCA evolutions are computable altogether.
A QCA is an array of identical finitedimensional quantum systems. The whole array evolves in discretetime steps by iterating a linear operator . Moreover the global evolution is translationinvariant (it acts everywhere in the same way), causal (information cannot be transmitted faster than some fixed number of cells per time step), and unitary (the condition required by the postulate of evolutions in quantum theory, akin to reversibility). See Fig. 1. This axiomatic style definition is the natural ‘quantization’ of the classical definition SchumacherWerner; ArrighiLATA; ArrighiIJUC. But contrary to its classical counterpart the axiomatic definition did not immediately yield a straightforward way of constructing the instances of this model. A great deal of effort has been dedicated towards understanding their structure, in terms of infinitely repeating circuits of local, quantum gates ArrighiUCAUSAL; ArrighiJCSS; ArrighiPQCA.
Roadmap.
We will tackle the above themes in a somewhat reversed order. We will start with the axiomatic definition of QCA, the consequent structure theorems, their origins, in Section 2. This will place us in a position to recall the main universality results and consequences in computability, in Section 3. QCA models of particle physics, whether for the sake of quantum simulation or as toy models for theoretical physics, will be discussed in Section 4. Section 5 summarizes the progress QCA theory has made, and some of the challenges that remain ahead.
Foreword.
I aimed to state the fundamental results of the field in a mathematical manner, and then touch on the many fascinating results around with just a word of explanation—hoping to show how they relate to each other and thereby draw a coherent picture. I favoured the logical order over the chronological, and citations over long and necessarily incomplete lists of surnames. Even with more than citations and in spite of my best efforts, I realize that this remains to some extent a personal account: I needed to select what seemed to be the most significant contributions, and may still be unaware of other great advances. Do get in touch if you have suggestions for the subsequent versions. Whist writing I learnt that a technically more comprehensive review on QCA was being written FarrellyReview. Hopefully the two will complement each other, updating Wiesner.
2 Structure
2.1 State space
Quantum Cellular Automata (QCA) consist in an dimensional array of identical dimensional quantum systems. In other words, each cell is a qudit in , and there are of them, thus the state space should morally be something like “”. Unfortunately this is not a Hilbert space (e.g. the scalar product may diverge). If one is willing to pay the price of abandoning Hilbert spaces and move to algebras BratteliRobinson, then one can make sense of QCA over such a space SchumacherWerner. In ArrighiLATA; ArrighiIJUC we were able to develop a simpler alternative, which is to assume that basic configurations are mostly empty.
Definition 1 (Configurations)
Consider a finite set, called the alphabet, with a distinguished element of , called the empty state. A configuration over is a function , i.e. mapping to , such that the set of the such that , is finite. The set of all configurations will be denoted or just .
Notice that is countable. Thus, we can now consider the Hilbert space of superpositions of configurations.
Definition 2 (State space)
The Hilbert space of configurations is that having canonical orthonormal basis . It will be denoted or just .
2.2 Qca
The global evolution of a QCA is required to be translationinvariant, meaning that it acts everywhere in the same way.
Definition 3
(Translationinvariance) Let denote the translation operator along the dimension, i.e. the linear operator over which maps into , where is such that for all , . A linear operator over is said to be translationinvariant if and only if for every .
Moreover, the global evolution of a QCA is required to be causal, meaning that information propagates at a bounded speed. In order to formulate this property, we need to be able to speak of the state of a cell at time , to say that it should only depends on the state of its neighbours at time . But in order to speak of the state of a subsystem in quantum theory we must switch to the density matrix formalism. Recall NielsenChuang that a density matrix represents a probability distributions over pure states as the corresponding convex sum of projectors . Thus, when pure states evolve according to , density matrices evolve according to . Then, the state of cell at time is obtained by tracing out all the of the other cells, i.e. , with the linear operator such that . Similarly, the state of its neighbours at time is obtained by tracing out the rest, i.e. . We get :
Definition 4 (Causality)
A linear operator over is said to be causal with neighbourhood if and only if for any there exists a function such that for any over , we have where .
To a certain extent, this may be thought of as the equivalent of the local rule of a classical CA. However, unlike for classical CA, this does not straightforwardly yield a local mechanism whereby may be computed from , for two reasons. First, because by itself is not unitary (it maps many cells into one) and thus cannot be considered to be physical. Second, because in quantum theory, knowing the states and of cells and does not entail knowing their joint states —as these may be entangled. Hence, unlike for classical CA, the following axiomatic definition of QCA does not immediately yield a straightforward way of constructing / enumerating all of the instances of the model.
Definition 5 (Qca)
A QCA is a linear operator over which is translationinvariant, causal and unitary.
We needed structure theorems in order to tame this axiomatic definition, into a constructive one.
2.3 Structure theorems
From the above axiomatic definition of QCA, we were able to eventually deduce that every QCA can be directly simulated by a finite depth quantum circuit of local unitary gates, infinitely repeating across space ArrighiUCAUSAL; ArrighiJCSS. In order to do so each cell of the QCA needs be encoded into a doubled up cell.
Theorem 2.1 (Unitary plus causality implies localizability)
ArrighiUCAUSAL; ArrighiJCSS
Let be an dimensional QCA with alphabet . Let be the isometry from to which adds an ancillary empty subcell at , i.e. . This mapping can be trivially extended to whole configurations, yielding the mapping . There exists an dimensional QCA with alphabet , such that , where admits the following mutlilayer quantum circuit representation:
(1) 
with

is the swap between the two subcells at and hence is local to .

is , which turns out to be local to the neighbourhood .
Proof
outline.

Show the equivalence between causality in the Shrödinger picture (Def. 4) and causality in the Heisenberg picture SchumacherWerner, which states that the conjugate of a local operator by a causal operator remains a local operator.

Extend to act only on the right subcells, and show that is a local operator, since it is the conjugate of a local operator.

Remark that plays the role a local update mechanism, as it computes the future state of a cell, and then puts it aside. Show that applying everywhere does implement , up to swaps.
This ‘direct simulation’ of QCA by QCA does show that any QCA can be put into the form of a finite depth quantum circuit, with local unitary gates and , up to a simple encoding. Indeed, and can be done in parallel whenever they do not overlap, i.e. when . In the case, for instance, the construction results in a –layered circuit. This case may seem restrictive, but it is not. Just like for CA IbarraJiang, one can always group the cells into supercells, relative to which the neighbourhood reduces down to .
The notion of ‘intrinsic simulation’ of a QCA by a QCA is obtained by relaxing the notion of direct simulation in two ways. First, by allowing for such groupings of cells into supercells, yielding and . Second, by allowing to be iterated times before it directly simulates . I.e. intrinsically simulates when directly simulates . The notion, made rigourous in ArrighiNUQCA, allowed us to reach an even simpler quantum circuitform for QCA ArrighiPQCA.
Definition 6 (Pqca)
An dimensional partitioned QCA (PQCA) is induced by a scattering unitary taking a hypercube of cells into a hypercube of cells, i.e. acting over , and preserving quiescence, i.e. . Let over and the diagonal translation. The induced global evolution is at even steps, and at odd steps.
In other words, PQCA work by partitioning the grid of cells into supercells, applying a local operation on each supercell, translating the partition along , applying to the new macrocells, etc., as illustrated in Fig. 2. Of course one can now change scale and take the view that supercells are now cells, whose subcells are the former cells. Then the data contained in each cell can be thought of as being subdivided into subcells, about to be sent towards the direction. In this picture, a PQCA is therefore the alternation of a ‘scattering step’ (where gets applied on each cell) followed by an ‘advection step’ (where subcells get synchonously exchanged with their corresponding crossdiagonal neighbour)—a scheme which physicists refer to as Lattice Gas Automata. Either way, PQCA suffer the small inconvenience that homogeneity is now over two steps. For instance, in the Lattice Gas Automata picture, the cells at time are translated by with respect to the cells at time . This can be fixed elegantly by doubling up the lattice like a checkerboard, with black and white lattices ignoring each other. Then PQCA recover fulltranslation invariance and fall back within the class of QCA with neighbourhood , i.e. QCA such that the state of a cell depends only upon that of its diagonal neighbours at the previous time step—but not on its own previous state, see Fig. 5. We were able to prove the following.
Theorem 2.2 (PQCA are intrinsically universal)
ArrighiPQCA Given any dimensional QCA , there exists an dimensional PQCA which intrinsically simulates .
Proof
outline.

Spacegroup the cells of the QCA so that its neighbourhood is down to .

Apply Th. 2.1, yielding successive partitions for applying the in a nonoverlapping way.

Devise a PQCA , with appropriate ancillary systems used to get the timing right, so that the scattering successively performs translated by the vectors in , according to the corresponding partition.
2.4 Advanced structure theorems
The Theorems of Subsec 2.3 show that QCA can be simulated, in a spacetimepreserving manner, by mutlilayer quantum circuits, and that these can in turn be simulated by PQCA. But, does it mean that QCA are exactly mutlilayer quantum circuits, or PQCA? A number of researchers have been pursuing this question. Our current knowledge of these issues varies according to the spatial dimension. It helps to recall the corresponding results on Reversible Cellular Automata (RCA), as these are the classical counterparts of QCA.
In 1D, RCA, are exactly the set of translationinvariant twolayer circuit of reversible gates and partial shifts KariBlock. The analogous result holds true for 1D QCA, as we can show that they are exactly the set of translationinvariant twolayered quantum circuits SchumacherWerner; ArrighiLATA; ArrighiIJUC. This leads to a group theoretical classification of 1D QCA, called the index theory GrossNesmeVogtsWerner—a single number characterizes the flux of information within the 1D QCA. Still, 1D QCA are not exactly PQCA ShakeelMeyer.
In 2D, RCA, are again exactly the set of translationinvariant threelayer circuit of reversible gates and partial shifts KariBlock. We still do not know whether the analogous result holds true for 2D QCA, in spite of the latest thorough efforts to generalize index theory to 2D HastingsClassification. Again we were able to show, thanks to a counterexample coming from RCA KariCircuit, that 2D QCA are not exactly PQCA ArrighiLATA; ArrighiIJUC. This is taken a step further in ShakeelLove, whose characterization of PQCA in terms of inclusion of algebras actually works independent of the spatial dimension.
Apart from this characterization of PQCA, little is known in 3D and beyond. It is open whether RCA coincide with the set translationinvariant circuits of reversible gates and partial shifts KariBlock. Again we do not know whether the analogous result holds true for 3D QCA, but HastingsNonTrivial holds the promises of a counterexample: “either a nontrivial threedimensional qudit QCA exists or a nontrivial twodimensional fermionic QCA exists.” Let us take this opportunity to mention fermionic QCA.
Intuitively, fermions are indistinguishable quantum systems, such that permuting one for another does not change anything but for a global, minus sign. Technically, let denote the linear operator which annihilates the fermion having spin at , let be the corresponding creation, and similarly for the other spins. The fermionic commutation relations are
These indeed entail that , so that , i.e. two identical fermions exclude each other. However, these anticommutation relations between and for any and also entail that is not actually local in the usual, qubit sense. This leads to subtle differences between computing with qubits and computing with fermions PaviaFermions. A fermionic QCA is a quantum evolution which is prescribed in terms of how the fermionic operators evolve, rather than how qubitlocal algebras evolve, see for instance PaviaMolecular; HastingsNonTrivial. However, a direct analog of Theorem 2.1 still holds for fermionic QCA FarrellyCausal. Moreover, FarrellyThesis shows that fermionic QCA and QCA intrinsically simulate each other. In ArrighiQED, the quantum evolution we describe is both a valid QCA and a valid fermionic QCA : whilst the notion of locality differs, that of causality coincides.
Finally, let us mention that Theorem 2.1 cannot be generalized to Noisy QCA, i.e. causal TPCPmaps, as these cannot always be simulated by finitedepth circuits of local TPCPmaps. This is a direct consequence of ArrighiPCA, were we showed that there can be no such structure theorems for classical, probabilistic CA.
2.5 Classical bijective CA, MPU
An RCA is a bijective CA, whose inverse is also causal, and hence a CA. Depending upon the space of configuration which one considers, there may be bijective CA which are not RCA. This is the case over in particular, where causal and bijective does not entail that is causal. It follows that the linear extension of into a unitary operator over may no longer be causal. One way to think about this is that for the update mechanism to be local, the local operation needs be conjugated by causal operators, which may not be the case with . We gave a concrete example of this in ArrighiLATA; ArrighiIJUC, which is defined as follows. Let , and for all and define as the ‘exclusive or’ , , extended so that , . Now let map the configuration into the configuration . One sees that is both bijective over and causal, but that is not causal, because a long subword may either stem from a similar subword or from a subword . It follows that the corresponding QCA is not causal. Indeed, consider the two states
and their two images
where . We can transmit information between arbitrarily distant parties in just one step of as follows.

Prepare the state with the first non quiescent cell in Alice’s lab in Paris and the last non quiescent cell with Bob in New York.

Alice either leaves the state unchanged or performs a local change by applying a phase gate to her cell, changing into .

One step of is performed, leading to either or .

Whether Alice performed or not has now led to a perfectly measurable change from to for Bob—despite him being arbitrarily far remote.
This infinite speedup is intuitively unphysical, and should disallowed: is no QCA.
Notice that the above example would not work over the space of infinite configurations , because is noninjective over that space: the infinite configurations and both map to . Actually, it turns out that over , any bijective CA is an RCA, because causal and bijective does entail that is causal. It follows that the linear extension of such a into a unitary operator is causal, and the local update mechanism is local indeed. Stlll, the QCA may have a much wider radius of causality than it had as an RCA . In ArrighiNEIGH we were able to show that . However, given just one cannot even bound ; in fact there is no computable function such that , as was proven in KariRevUndec. Let us mention that tHooftCA embarked on the program of studying how much physics can be recovered within such quantized RCA .
The early definition DurrWell; DurrUnitary of QCA would allow for these noncausal over ; and had to be abandoned. However, a characterization of 1D QCA as tensor networks of Matrix Product Unitaries (MPU) CiracMPU has recently appeared, which bears strong similarities with the abandoned definition, whilst remaining causal. Indeed, with MPU one looks for a tensor such that the circular tensor network
is unitary matrix for any . Interestingly, this is the quantum analogue of the classical bijectivity over rather than , which made bijective CA an RCA and a QCA.
2.6 Historical notes
Whilst the definition of QCA is now wellestablished, its beginnings were difficult. The first attempt of a definition would require that state vector of a cell at time , should be locallydependent upon the state of its neighbouring cells at time DurrWell; DurrUnitary. Whilst plausible, we realized that this definition was problematic, as it would still allow for information to propagate at an arbitrary speed ArrighiLATA; ArrighiIJUC. The first solid axiomatic definition of QCA was given in SchumacherWerner, in terms of –algebra, together with a broken proof that these were in fact exactly twolayer quatum circuits, infinitely repeating across space. In ArrighiLATA; ArrighiIJUC we rephrased the definition in terms of Hilbert spaces, and clarified the proof as sound for 1D QCA, but gave a counterexample in higherdimensions taken from KariCircuit. In ArrighiUCAUSAL; ArrighiJCSS we were able to obtain Th. 2.1, which we later ArrighiPQCA completed into Th. 2.2. Summarizing, QCA are not exactly PQCA, but they are intrinsically simulated by them.
It took a while, thus, to arrive at the axiomatic definition, and even longer to deduce ways of constructing / enumerating the corresponding instances of this definition. Naturally, this lack of operationality left the gap open for many competing, handson definitions of the same concept. A closer examination shows that these competing operational definitions would fall into three classes: the mutlilayer quantum circuits PerezCheungArrighiUCAUSAL, the twolayer quantum circuits BrennenWilliams; Karafyllidis; NagajWocjan; Raussendorf; SchumacherWerner; VanDam, and PQCA WatrousFOCS; VanDam; InokuchiMizoguchi. In ArrighiPQCA we showed that they all simulate each other in a spacetimepreserving manner, leading us to prefer their simplest, PQCA form.
3 Universality
3.1 Intrinsic universality
In Subsec. 2.3 we recalled the notion of intrinsic simulation between QCA in order to show that PQCA can simulate all other QCA in a spacetimepreserving manner. Now, once the structure of a computational model of computation is wellunderstood, the last step to take in order to try and simplify it even further, is to identify universal instances of the model. Minimal universal instances are particularly useful, as they point towards the threshold physical resources required in order to implement the entire model. Thus, we need to look for a single PQCA which can intrinsically simulate all other PQCA.
In a PQCA, incoming information gets scattered by a fixed ‘scattering unitary’ , before getting redispatched. We need to find a universal scattering unitary , see Fig. 2. 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 simulating any grid network of quantum processing units, in a spacepreserving manner. From a theoretical physics perspective, this is looking for a universal scattering phenomenon, a problem which we could phrase humorous form: “A physicist is taken on a desert Island where he is allowed only one type of elementary particle. Which one would he choose, whose scattering behaviour is rich enough so that it can simulate all that others?”.
We began the search for an intrinsically universal PQCA in dimension , which is feasible ArrighiDCM; ArrighiFI (see also Fig. 2) but difficult, because wires cannot cross over. We then then tackled the problem in the general dimensional case, where we could find a much simpler solution ArrighiSimple; ArrighiNUQCA. Eventually we reached a minimal, dimensional construction ArrighiQGOL. This socalled ‘Quantum Game of Life’ roughly works as follows. Each cell contains just one qubit—in Fig. 3 the cell is represented as little cube, red if the qubit is in state , transparent if it is in state . At even steps, the cube of 8 qubits at cells , as well as all of its –translates, each undergo a scattering unitary , synchronously. At odd steps, the cube of 8 qubits at cells , as well as all of its –translates, each undergo again, synchronously. The scattering unitary is given in Fig. 3, by means of its action over a small number of basis states (the full definition follows by linear extension and assuming rotationinvariance).
Observe that when there is just one red, e.g. at the leftbottomfront corner of the cube, it just moves across the cube. But because of the staggering between even and odd steps, it will find itself at the leftbottomfront corner of the new cube and again move across—this is the mechanism whereby signals are made. We also need to be able to redirect our signals, so we need red walls to form stable patterns, and demand that if a fifth red comes along, it bounces off. The Hadamard is implemented as a special case of this deflection : if the signal bounces on a edge, it skids or bounces, in a quantum superposition. Finally a two qubit interaction happens when two signals cross, and a phase gets added.
The outline of the proof that this PQCA is intrinsically universal is as follows (see Fig. 4 for a 2D illustration of this argument). First make fixedshaped tiles, each implementing one of the universal quantum gates of the quantum circuit model in a fixed number of time steps, out of these walls and signals. Next, combine these tiles into a layout that implements , the scattering unitary of the PQCA to be simulated. Repeat this layout across space. Finally, plug the outputs of each simulated gates into the inputs of the neighbouring ones, so that they feed each other, thereby implementing the staggered structure of the simulated PQCA.
Notice that the hereby constructed scattering unitary is over 8 qubits, which is much more complicated that the 2 qubits gate sets that are universal for quantum circuits. This is because simulation of a QCA has to be done in a parallel, spacetimepreserving manner, and because we must simulate not just one iteration of but several (, …, i.e. after every iteration we must get ready for the next one). Thus intrinsic universality is a much more stringent requirement than quantum circuit universality.
3.2 Other kinds of universality
Quantum Turing machine.
We just constructed a PQCA that is capable of simulating any other PQCA and hence any QCA. But does it mean that this PQCA is capable of running any quantum algorithm? Clearly, the question amounts to whether QCA are universal for QC. In the sense of the quantum circuit model the answer is clearly yes, simply by inspection of the construction of Subsec. 3.1, Fig. 3 and 4 in particular. In the sense of the quantum Turing machine, the answer is clearly yes also, as was proven in WatrousFOCS. In this construction, a QCA with alphabet simulates a Quantum Turing machine with alphabet and internal states . The way this works is that each cell is capable of hosting the head of the Turing machine : it has enough state space to store both the symbol at this location of the tape; the internal state of the head; and whether the head is actually there or not.
Quantum circuit universality.
The socalled “physical universality” is specifickind of quantum circuit universality for QCA, more stringent than the early work of VanDamUniversal. Indeed, in all the abovementioned universality constructions, part of the state space of each cell is used to encode ‘the program’ (i.e. what dynamics is to be simulated), whilst the other part is used to encode ‘the data’ (i.e. the states whose evolution is being simulated). One may demand that this is not the case, and wish to have a convex region in which the data (the input to a quantum circuit) lies untouched, without any preparation, whereas only the surroundings are allowed to code for the program (the quantum circuit to be applied). The requirement is that after a precise number of time steps, the data is to be found at the very same place, evolved according to the specified quantum circuit. Such a construction is achieved in SchaefferPhysUniv. In this construction the data within is left to “explode” into , where it gets treated and redirected towards .
Computability.
The (strong) physical ChurchTuring thesis states that “any function that can be (efficiently) computed by a physical system can be (efficiently) computed by a Turing machine”. Because there are concrete examples of functions that cannot be computed by Turing machines (e.g. the famous halting function ), the physical ChurchTuring thesis makes a strong statement about physics’ (in)ability to compute.
The discovery of QC algorithms has shaken the strong version of the thesis. But what about the original version—could it be that a QC might compute functions that were not computable classically? Quantum theory imposes that physical systems evolve unitarily : according to a unitary matrix when the system is finitedimensional; according to a unitary operator otherwise. It follows that finite quantum circuits can always be simulated (very inefficiently) on a classical computer just via matrix multiplications. Therefore these do not endanger the original version of the thesis. Yet nothing forbids NielsenComputability; Kieu; NielsenMore that Unitary operators, on the other hand, break the thesis, e.g. . That is unless the limitation comes from other physical principles.
That physically motivated limitations lead to the physical ChurchTuring thesis was aready argued in Gandy by Gandy, Turing’s former PhD student. There, the main idea is that causality (i.e. bounded velocity of propagation of information) together with homogeneity (i.e. the rules of physics are the same everywhere and everywhen) and finite density (i.e. bounded number of bits per volume) entail CAlike evolutions, which are computable. The main issue with this proof of the physical ChurchTuring thesis based upon physics principles, is that it complete ignores quantum theory. Quantum theory demands that the bounded density principle be updated, changing the word ‘bits’ to the word ‘qubits’.
When we do so, the updated set of principles entail QCAlike evolutions in the axiomaticstyle of Def. 5. We can then apply Th. 2.1. Armed with a robust notion of computability upon vector spaces ArrighiCIE, we were able show that these are computable ArrighiGANDY. This provided a proof of the ChurchTuring thesis based upon quantum physics principles.
4 Simulation
Let us take a step back to realize how the previous results chain up. In Sec. 2 we showed that discretespace discretetime quantum theory, i.e. QCA, can be intrinsically simulated by PQCA. In Sec. 3 we constructed a minimal, intrinsically universal PQCA. Logically, this entails that any lattice discretespace discretetime quantum physics phenomenon can be expressed within this particular PQCA.
Let us evaluate whether such statements are applicable in practice. Let us pick up one of the simplest and most fundamental physics phenomenon, namely the free propagation of the electron, and see whether it can be reexpressed by means of some simple PQCA.
4.1 The Dirac QCA
The equation governing the free propagation of an electron is called the Dirac equation. Let us describe a PQCA model for it, the socalled Dirac QCA. For this Dirac QCA we adopt the conventions depicted in Fig. 5. Each red (resp. black) wire carries a qubit, which codes for the presence or absence of a leftmoving (resp. rightmoving) electron. Thus there can be at most two electrons per site , (where the red and black wires cross).
The scattering unitary of this QCA is given by
(2)  
(3) 
with and , where stands for the mass of the electron, and the spacetime discretization step. Notice that the components are ordered so that when the mass is zero, the particles do not change direction, i.e. a rightmoving electron is transfered from to , etc.
To see whether this QCA implements the Dirac equation, let us consider the oneparticle sector, i.e. restrict to the QCA to the subspace spanned by states of the form .
Let be the amplitude of the particle being on a red (resp. left) wire at their intersection at point . We have
(4) 
Expanding to the first order in gives
(5)  
(6)  
(7) 
which is the Dirac equation, where .
4.2 Further simulation results
The oneparticle sector : quantum walks.
Subsec. 4.1, illustrated how to restrict QCA to their ‘oneparticle sector’, i.e. configurations of the form with . Each of these represents a single ‘particle’, standing at a some position with internal state . The Hilbert space of superpositions of these configurations is a rather small subspace of , which is better described as —i.e. superpositions of positionstate pairs . The amplitude of is usually written and one needs . This oneparticle sector of QCA has a life of its own. It is the playground for a huge field of research known by the name of Quantum Walks (QW). A QW, therefore, is essentially an operator driving the evolution of a single particle on the lattice, through local unitaries.
One reason for the popularity of QW is that a whole series of novel Quantum Computing algorithms, for the future Quantum Computers, have been discovered via QW, e.g. BooleanEvalQW; ConductivityQW, or are better expressed using QW, e.g the Grover search. In these QWbased algorithms, however, the walker usually explores a graph, which is encoding the instance of the problem, rather than a fixed lattice. No continuum limit is taken.
Our focus here will remain with the other reason, which is the ability of QW to simulate certain quantum physical phenomena, in the continuum limit—thereby providing:

Simple discrete toy models of physical phenomena, that conserve most symmetries (unitarity, homogeneity, causality, sometimes even Lorentzcovariance)—thereby providing playgrounds to discuss foundational questions in Physics.

Quantum simulation schemes, for the nearfuture simulation devices, in the way that was envisioned by Feynman when he invented QC FeynmanQC; FeynmanQCA.

Stable numerical schemes, even for classical computers—thereby guaranteeing convergence as soon as they are consistent.
Subsec. 4.1 is a simplified presentation of the original arguments by BenziSucci; BialynickiBirula; MeyerQLGI suggesting that QW can simulate the Dirac equation. We just gave a rigorous proof of convergence, given regular enough initial conditions, including in –dimensions, in ArrighiDirac—without the need to actually solve the QW evolution as in StrauchPhenomena. An axiomatic derivation of these schemes is given in DAriano; DAriano3D; RaynalDirac. We discussed conservation of symmetries, including Lorentzcovariance in ArrighiLorentzCovariance, and so did PaviaLORENTZ; PaviaLORENTZ2; DebbaschLORENTZ. The KleinGordon equation can also be simulated via this QW once the appropriate decoupling is performed, as explained in IndiansDirac; MolfettaDebbasch and in our subsequent generalization ArrighiKG. The Schrödinger equation can be obtained in a similar fashion, but by scaling space and time differently, i.e. but , see StrauchCTQW; StrauchShrodinger; BoghosianTaylor2.
Once it was realized that QW could simulate free particles, the focus shifted towards simulating particles in some background field cedzich2013propagation; di2014quantum; di2016quantum; arnault2016quantum; marquez2017fermion, by means of nontranslationinvariant QW. The question of the impact, of these inhomogeneous fields, upon the propagation of the walker gave rise to lattice models of Anderson localization WernerLocalization; JoyeLocalization. Surprisingly, they even gave rise to lattice models of particles propagating in curved spacetime MolfettaDebbasch2014Curved; DebbaschWaves, see also our subsequent generalizations ArrighiGRDirac; ArrighiGRDirac3D.
The manyparticle sector.
Recently, the twoparticle sector of QCA was investigated, with the two walkers interacting via a phase (similar to the Thirring model DdV87). This was shown to produce molecular binding between the particles ahlbrecht2012molecular; PaviaMolecular. In the manyparticle sector, the problem of defining a concrete QCA that would simulate a specific interacting QFT had remained out of reach until ArrighiQED. In this paper, we were able to give a first QCA description of QED in –spacetime (a.k.a the Schwinger model).
Trotterization of a nearestneighour hamiltonian.
QCA are in discretespace and discretetime. Let us consider their cousins in discretespace but continuoustime, i.e. lattices of quantum systems interacting according to a nearestneighbour translationinvariant hamiltonian. These have been studied for quite a while in Physics e.g. in condensed matter or statistical quantum mechanics (spin chains, Ising models, Hubbard models…), or towards QC (as candidate architectures FitzsimonsTwamley; Benjamin; Twamley; WeinsteinHellberg, for quantum information transport Bose, for entanglement creation Subrahmanyam; SubrahmanyamLakshminarayan; BrennenWilliams, as universal QC VollbrechtCirac; NagajWocjan…).
Up to groupings and reencodings, focussing here on the 1D case just for simplicity, nearestneighbour translationinvariant hamiltonians work as follows VollbrechtCirac. A global, continuoustime evolution is induced, by giving a hermitian matrix over verifying that , according to
with where stands for as acting over positions and .
From a practical implementation pointofview, nearestneighbour hamiltonians are central, and will be for a long time. Indeed, although there are a number of remarkable exceptions WernerElectricQW; Alberti2QW; Sciarrino, most the leadingedge latticebased quantum simulation devices remain better described as continuoustime evolutions Bloch. From a theoretical pointofview, however, nearestneighbour hamiltonians suffer the same downsides as the rest of nonrelativistic quantum mechanics to which they pertain. Namely, strictly speaking they do allow for superluminalsignalling, hopefully in some negligible, exponentially tailing off manner, relying upon some LiebRobinson type of argument that can sometimes go wrong EisertSupersonic. Intuitively, this is because even though it is the case that in an infinitesimal of time a cell only interacts with its neighbour, this is no longer true after any finite period of time , however small, as includes terms of the form and if information is to propagate at all.
Still, there is a strong connection between nearestneighbours hamiltonians and QCA, which arises from the TrotterKato formula (a.k.a BakerCampbellThomson or operatorsplitting method):
Indeed, let and , and readily get that , where is the PQCA induced by the scattering unitary —back in discretespace discretetime.
Are space and time back on an equal footing, thanks to this ‘trotterization procedure’? Not quite. For this approximation to hold mathematically, one still needs that —for instance by setting and as was done earlier in order to get the nonrelativistic, Schrödinger equation. Thus, QCA arising by trotterizing nearestneighbour hamiltonians are generally nonrelativistic models ArrighiChiral, unless they are carefully engineered otherwise, as we did in ArrighiUnifiedQW. Still, this is not the only use of the TrotterKato formula, which has turned out to be an ubiquitous mathematical tool in this field.
Noise, thermodynamics.
To the best of our knowledge there has not been much studies of noisy QCA, i.e. replacing unitary operators by quantum operators (a.k.a TPCP maps), with a handful of exceptions AvalleNoisyQCA, in the manyparticle sector. The oneparticle sector has been thoroughly studied on the other hand, e.g. studying the transition for QW (ballistic transport) to random walks (diffusion) LoveBoghosian. In ArrighiNoisyQW we studied this transition in parallel with the continuum limit to PDE, where the Dirac equation turns into a Lindblad equation and then a telegraph equation—making the argument that noisy quantum simulation devices can still be useful, to simulate noisy quantum systems.
The connection to toy models of thermodynanics is also a promising one. In RomanelliTemperature the thermalization of a QW is observed. In the manyparticle sector, Clifford QCA SchlingemannWerner; GutschowCLIFFORD (a subcase of QCA which can be classically simulated) were shown to produce fractal pictures NesmeGutschow and then gliders GUWZ, used to show bounds on entanglement propagation and von Neumann entropy creation in QCA. These handson toy models may eventually bring about interesting, complementary pointofview on the blossoming field of quantum thermodynamics, by taking space into account, which is believed to be a key ingredient of the quantumtoclassical transition PazZurek.
5 Conclusion
Summary. Quantum cellular automata (QCA) are quantum evolutions of lattices of quantum systems, as resulting from nearest neighbourinteractions. This sentence, however, could be understood in many ways:

In terms of finitedepth circuits of local quantum quantum gates, infinitely repeating across space—amongst the various shapes of circuits proposed, it is now known that the simplest, namely Partitioned QCA (PQCA), can simulate all the others.

In continuoustime in terms of sum of local hamiltonians. It is now clear that integrating these over a small period of time, yields a discretetime evolution that can again be simulated by a PQCA.

In more abstract terms, as the axiomatic requirement that the evolved state vector; or quasilocal algebra; or density matrix, be locally dependent. Locally dependent state vectors have turned out to make little sense, but the last two were shown to be equivalent and ultimately again simulated by PQCA.
Amongst PQCA, some instances were shown to simulate the quantum Turing machine; the free electron; the electron in an electromagnetic field; the electron in curved spacetime, including in –dimensions. Lately some were shown to model interacting quantum field theories as PQCA, namely the Thirring model, and QED in –dimensions. Ultimately, some particular instances were shown to simulate all other instances.
Perspectives. I would love to see QCA come true, implemented in the labs. This is certainly a fascinating topic in which cold atoms WernerElectricQW; Alberti2QW; Bloch, integrated fiber optics Sciarrino and hopefully superconducting qubits SuperconductingQSim will have a say. As a theoretician it would be adventurous for me to comment much on this perspective. It seems quite likely however that noisy implementations will see the light in the next ten years, with progressive improvements from there. At least nothing, at theoretical level, prevents it. The fact that much physics phenomena can be cast as QCA is an encouragement in this sense; it suggests that physics might naturally implement QCA, at its fundamental level.
But exactly, how much particle physics can be recast as QCA? Will QCA provide us with an alternative mathematical framework for interacting quantum field theories? Hopefully a clearer one, more explanatory, readily providing us with quantum simulation algorithms to draw predictions? These questions are, at the theoretical level, the obvious and most likely continuation of the trend of work which I presented in this overview. I am very optimistic about them : my personal belief is that these will be answered positively, probably within the next ten years. Of course I foresee many technical difficulties along the way, but no good reason why this could not be done. Thus, I wish to take this opportunity to encourage young researchers to engage these noble and realistic aims, hopefully enjoying the same collaborative spirit that has reigned over this research community in the last decade.
I do not believe, however, that QCA can account for General Relativity. Nor do I believe that they constitute the ultimate model of distributed Quantum Computation. In both cases, an ingredient is missing : the ability to depart from the grid and make the topology dynamical. Quantum Causal Graph Dynamics ArrighiQCGD are unitary operators over quantum superpositions of graphs. The graphs constrain the evolution by telling whom can interact with whom ; but at the same time they are the subject of the evolution, as they may vary in time. The possible connections between this further generalization of QCA, and Quantum Gravity, are intriguing.
Acknowledgements
I was lucky to have, as regular coauthors, great researchers such as Pablo Arnault, Cédric Bény, Gilles Dowek, Giuseppe Di Molfetta, Stefano Facchini, Terry Farrelly, Marcelo Forets, Jon Grattage, Iván Márquez, Vincent Nesme, Armando Péres, Zizhu Wang, Reinhard Werner. I would like to thank Jarkko Kari and Grzegorz Rozenberg for inviting me to write this overview, a task which I had been postponing for too long.