1: \documentclass[aps,twoside,twocolumn]{revtex4}
2:
3: \usepackage{amsmath}
4: \usepackage{amsfonts}
5: \usepackage{amssymb}
6: \usepackage{epsfig}
7:
8:
9: \def\jtitle#1{#1} % use this line to include titles of journal articles
10: %\def\jtitle#1{} % and this in order not to
11:
12: \newtheorem{theorem}{Theorem}
13: \newtheorem{lemma}[theorem]{Lemma}
14: \newtheorem{definition}[theorem]{Definition}
15: \newtheorem{proposition}[theorem]{Proposition}
16: \newtheorem{corr}[theorem]{Corollary}
17: \def\proof{\par\vskip12pt\noindent{\it Proof}}
18: \def\qed{\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill
19: \quad\hbox{\leavevmode \hbox to.77778em{%
20: \hfil\vrule \vbox to.675em%
21: {\hrule width.6em\vfil\hrule}\vrule\hfil}}}
22:
23: \def\AA{{\cal A}}\def\BB{{\cal B}}\def\CC{{\cal C}}\def\DD{{\cal D}}\def\EE{{\cal E}}
24: \def\HH{{\cal H}}
25: \def\MM{{\cal M}}\def\ZZ{{\cal Z}}
26: \def\idty{\hbox{\rm\openone}}
27: \def\tr{{\rm tr}}
28: \def\sym{{\rm sym}}
29: \def\norm#1{\left\Vert{#1}\right\Vert}
30: \def\stsp#1{{\frak S}(#1)}% state space
31: \def\bigtimes{\mathop{\times}}
32:
33: \def\Prob{{\mathbb P}}
34: \def\MFalg{{\cal M\mkern-3mu\cal F}(A,B)}
35: \def\Cx{{\mathbb{C}}}\def\Ir{\hbox{$\mathbb{Z}$}}
36: \def\ket#1{\vert#1\rangle}
37: \def\bra#1{\langle#1\vert}
38: \def\ketbra#1#2{\vert#1\rangle\langle#2\vert}
39: \def\braket#1#2{\right\langle#1\mid\vert#2\left\rangle}
40: \def\Nei{{\cal N}}% Neighbourhood scheme
41: \def\loclat{{\cal L}} % quotient lattice for local rules
42: \def\qone{{\bf1}}
43: \def\dtwos{d^{\,2^{\scriptstyle s}}}
44: \def\grule{T}
45: \def\lrule{\grule_0}
46: \def\brule{\grule_\Box}
47:
48: \def\spp{{\bf s}} %linear support of subspace of tensor product
49: \def\Spp{{\bf S}} % support algebra
50:
51: \def\OO{{\mathcal O}}
52: \def\AO#1{{\mathfrak A}({\OO}_{#1})}
53: \def\spacelike{\hbox{$/\mkern-9mu\backslash$}}
54: \def\Field{{\mathbb F}}
55:
56: \begin{document}
57: %* Top Matter
58: \title{Reversible Quantum Cellular Automata}
59: \author{B. Schumacher}
60: \email{schumacherb@kenyon.edu} \affiliation{Kenyon College, Dept.
61: of Physics, Gambier, Ohio 43022-9623, USA}
62: \author{R.~F. Werner}
63: \email{R.Werner@TU-BS.DE} \affiliation{Inst. Math. Phys.,
64: TU-Braunschweig, Mendelssohnstra{\ss}e 3,
65: D-38106 Braunschweig, Germany}
66: \begin{abstract}
67: We define quantum cellular automata as infinite quantum lattice
68: systems with discrete time dynamics, such that the time step
69: commutes with lattice translations and has strictly finite
70: propagation speed. In contrast to earlier definitions this allows
71: us to give an explicit characterization of all local rules
72: generating such automata. The same local rules also generate the
73: global time step for automata with periodic boundary conditions.
74: Our main structure theorem asserts that any quantum cellular
75: automaton is structurally reversible, i.e., that it can be
76: obtained by applying two blockwise unitary operations in a
77: generalized Margolus partitioning scheme. This implies that, in
78: contrast to the classical case, the inverse of a nearest neighbor
79: quantum cellular automaton is again a nearest neighbor automaton.
80:
81: We present several construction methods for quantum cellular
82: automata, based on unitaries commuting with their translates, on
83: the quantization of (arbitrary) reversible classical cellular
84: automata, on quantum circuits, and on Clifford transformations
85: with respect to a description of the single cells by finite Weyl
86: systems. Moreover, we indicate how quantum random walks can be
87: considered as special cases of cellular automata, namely by
88: restricting a quantum lattice gas automaton with local particle
89: number conservation to the single particle sector.
90:
91: \end{abstract}
92: \maketitle
93:
94: \section{Introduction}
95: The idea of generalizing the classical notion of cellular automata
96: to the quantum regime is certainly not new. Indeed, it is already
97: present in Feynman's famous paper \cite{Feynman} from 1982, in
98: which he argues that quantum computation might more powerful than
99: classical. However, although there have been several formal
100: definitions of quantum cellular automata over the years, the
101: theory is not in good shape at the moment, and a systematic
102: exploration of the general properties of such systems on the one
103: hand and of the potential for computational applications on the
104: other has hardly begun. We believe that this is partly due to
105: deficiencies of the existing approaches, and therefore propose a
106: new one, which is very natural, and only requires a few basic
107: assumptions: a discrete cell structure with a finite quantum
108: system for every cell and translation symmetry, a discrete time
109: step for the global system, reversibility, and finite propagation
110: speed.
111:
112: Quantum cellular automata (``QCAs'') are of interest to several
113: fields. There are obvious connections to the statistical mechanics
114: of lattice systems, and potential applications to ultraviolet
115: regularization of quantum field theories. In Quantum Computer
116: Science they appear as one natural model of computation extending
117: the well-developed theory of classical cellular automata into the
118: quantum domain. But also the experimental side is rapidly
119: developing: quantum computing in optical lattices \cite{optlat}
120: and arrays of microtraps \cite{mictrap} are among the most
121: promising candidates for the first quantum computer that does
122: useful computations \cite{QdCA}. It is typical for such systems
123: that the addressing of individual cells is much harder than a
124: change of external parameters affecting all cells in the same way
125: \cite{Benjamin}. But this is just the theoretical description of a
126: cellular automaton. Possible tasks, which have the same built-in
127: translation invariance are simulations of solid state models.
128: Typically classical simulations run into problems already for
129: moderate systems sizes, precisely because of the dimension and
130: complexity explosion which Feynman noted, and for which he
131: proposed quantum computation as a cure. The theory presented in
132: this paper can be considered as providing the first elements of an
133: assembly language for such simulations.
134:
135: The problems which have plagued previous attempts to define QCAs
136: begin with the definition of a system of infinitely many cells.
137: Consider the simplest operation such an automaton should be able
138: to perform: applying the same unitary transformation separately to
139: each cell. This would involve multiplying infinitely many phases,
140: so there is really no well-defined unitary operator describing the
141: global state change. Therefore, the quantization approach ``just
142: make the transition function unitary'' does not work very well.
143: Similarly, the notion of state vectors as amplitude assignments to
144: uncountably many classical configurations is ill-defined. But this
145: would be a candidate for the ``configurations'' of a QCA, which
146: causes problems for a definition of QCAs in terms of
147: configurations and their transformations. Various approaches
148: \cite{Watrous,vanDam,Gruska,santa,Feynman} will be described and
149: commented in Section~\ref{sec:others}. A constructive method to
150: obtain QCAs, which is common to most of these approaches
151: (including ours) is partitioning the system into blocks of cells,
152: applying blockwise unitary transformations, and possibly iterating
153: such operations. Model studies based on such constructions (e.g.,
154: \cite{Brennen}) therefore produce results independently of the
155: definition problems.
156:
157: In order to arrive at a satisfactory notion of QCAs, it is helpful
158: to draw on ideas from a discipline, which has been dealing with
159: infinite arrays of simple quantum systems for a long time, i.e.,
160: the statistical mechanics of quantum spin systems. Infinite
161: systems have been considered particularly in the algebraic
162: approach to such systems \cite{BraRo}. The basic idea is to focus
163: on the observables rather than the states, i.e., to work in the
164: Heisenberg picture rather than the Schr\"odinger picture
165: \cite{Paschen}. The main advantage is that in contrast to
166: localized states, it does make sense speak of local observables
167: \cite{Haag}, i.e., observables requiring a measurement only of a
168: finite collection of cells. The global transition rule of a
169: cellular automaton is then a transformation $T$ on the observable
170: algebra of the infinite system. As always in the Heisenberg
171: picture, the interpretation of such a transformation is that
172: `preparing a state, running the automaton for one step, and then
173: measuring the observable $A$' gives exactly the same expectations
174: as preparing the same state and measuring $T(A)$. As always, $T$
175: must be completely positive and satisfy $T(\idty)=\idty$. But more
176: importantly we can state the crucial localization property of
177: QCAs: When $A$ is localized on a region $\Lambda$ of the lattice,
178: then $T(A)$ should be localized in
179: $\Lambda+\Nei=\{x+n|\;x\in\Lambda,\ n\in\Nei\}$, where $\Nei$ is
180: the {\it neighborhood scheme} of the QCA.
181:
182: To our knowledge, this view of QCAs was first used in \cite{RiWe},
183: where the approach to equilibrium in a QCA with irreversible local
184: rules (based on a partitioning scheme) was investigated. The
185: general picture of QCAs remained unsatisfactory, however, because
186: the partitioning scheme seemed a rather special way of
187: constructing a QCA. For a satisfactory theory of QCAs we demand
188: that there should be a direct connection between the {\it global
189: transition rule} $T$ and the {\it local transition rule}: If we
190: know the global transition rule, we should be able to extract
191: immediately the local rule in a unique way, and conversely, from
192: the local rule we should be able to synthesize the global rule.
193: The class of global rules should have an axiomatic specification,
194: the most important of which would be the existence of a finite
195: neighborhood scheme in the above sense. On the other hand, for the
196: local transition rules we would prefer a {\it constructive
197: characterization}. That is, there should be a procedure for
198: obtaining all local rules leading to global rules with the
199: specified properties, in which all choices are clearly
200: parameterized.
201:
202: The partitioning QCAs of \cite{RiWe} failed to meet these
203: requirements, because they provided a construction, but no
204: axiomatic characterization of the global rules obtained in this
205: way. In particular, it remained unclear whether two steps of such
206: an automaton could be considered as a single step of an automaton
207: with enlarged neighborhood scheme. The idea enabling the present
208: paper was that all these difficulties vanish if we restrict to the
209: class of {\it reversible} QCAs. The axiomatic characterization is
210: extremely simple: In addition to the above locality condition we
211: assume that the global rule must have an inverse, which is again
212: an admissible quantum channel. This is equivalent to saying that
213: $T$ must be an automorphism of the observable algebra. Then the
214: local rule is simply the restriction of this automorphism to the
215: algebra of a single cell. Conversely, since every observable can
216: be obtained as a linear combination of products of single-cell
217: observables, the local rule determines the global automorphism.
218: This allows us to subsume all the known constructions of QCAs, but
219: also to prove a general structure theorem: every reversible QCA is
220: structurally reversible, i.e., we can write the local rule in a
221: partitioning scheme involving two unitary matrices, which makes it
222: apparent how the global rule can be unitarily implemented on
223: arbitrarily large regions, and how to obtain the local rule of the
224: inverse.
225:
226: A further bonus from our proof of the structure theorem is that it
227: does not actually require the global rule to be an automorphism:
228: it works under the prima facie much weaker assumption that the
229: global rule is a homomorphism (and not necessarily onto). Then
230: invertibility follows (see Corollary~\ref{cor:invert} below).
231: Therefore, invertibility was not included in Definition~1, which
232: makes it much easier to verify whether a proposed rule is indeed a
233: QCA.
234:
235: The structural invertibility was an open problem in the theory of
236: classical reversible cellular automata in higher dimensional
237: lattices until recently \cite{Kari1}. Hence, since our proof of
238: the quantum result is rather simple, it appears that some proofs
239: in the classical domain can be simplified by going quantum.
240:
241: Our paper is organized as follows: In Section~\ref{sec:def}, we
242: begin with the axiomatic definition of QCAs in the sense described
243: above. Its counterpart, the constructive description is given in
244: the form of a collection of basic constructions and examples in
245: Section~\ref{sec:basicon}. It turns out that one of these
246: constructions, based on partitioning, is already sufficient to
247: obtain all QCAs in the sense of our definition. This rather
248: surprising result is stated and proved in Section~\ref{sec:struc}.
249: The ideas of the proof also allows us to give an explicit
250: parameterization of the simplest class of QCAs: nearest neighbor
251: automata in one dimension with one qubit per cell. As mentioned in
252: the introduction, the current literature on the subject is mostly
253: based on a definition we do not find satisfactory. We discuss
254: these, and some further related definitions, in more detail in
255: Section~\ref{sec:others}. Finally, in an appendix we provide some
256: mathematical background on finite dimensional C*-algebras, which
257: play a key role in the proof of Theorem~\ref{mainthm}.
258:
259:
260:
261:
262: \section{Definition of QCAs}\label{sec:def}
263: We consider an infinite cubic array of cells, labelled by integer
264: vectors $x\in\Ir^s$, where $s\geq1$ is the spatial dimension of
265: the lattice \cite{otherlat}. Each cell contains a $d$-level
266: quantum system with the same finite $d\geq2$. That is to say, with
267: each cell $x\in\Ir^s$, we associate the observable algebra $\AA_x$
268: of the cell, and each of these algebras is an isomorphic copy of
269: the algebra of complex $d\times d$-matrices. When
270: $\Lambda\subset\Ir^s$ is a finite subset, we denote by
271: $\AA(\Lambda)$ the algebra of observables belonging to all cells
272: in $\Lambda$, i.e., the tensor product
273: $\bigotimes_{x\in\Lambda}\AA_x$. By tensoring with unit operators
274: on $\Lambda_2\setminus\Lambda_1$ we consider $\AA(\Lambda_1)$ as a
275: subalgebra of $\AA(\Lambda_2)$, whenever
276: $\Lambda_1\subset\Lambda_2$. In this way the product $A_1A_2$ of
277: $A_i\in\AA(\Lambda_i)$ becomes a well-defined element of
278: $\AA(\Lambda_1\cup\Lambda_2)$. Moreover, tensoring with the
279: identity does not change the norm, so we get a normed algebra of
280: {\it local observables}, whose completion is called the {\it
281: quasi-local algebra}\cite{BraRo}, and will be denoted by
282: $\AA(\Ir^s)$. Similarly, for other infinite subsets
283: $\Lambda\subset\Ir^s$ we define $\AA(\Lambda)$ as the closure of
284: the union of all $\AA(\Lambda')$ with $\Lambda'\subset \Lambda$
285: finite.
286:
287: When $x\in\Ir^s$ is a {\it lattice translation}, we denote by
288: $\tau_x$ the isomorphism from each $\AA_y$ to $\AA_{x+y}$, and its
289: extensions from $\AA(\Lambda)\to\AA(\Lambda+x)$, by shifting every
290: site. Here we have used the notation
291: $\Lambda+x=\{y+x|y\in\Lambda\}$ for shifted lattice subsets, which
292: we also extend to
293: $\Lambda_1+\Lambda_2=\{x_1+x_2|x_i\in\Lambda_i\}$.
294:
295: A {\it state} $\omega$ of the spin system is a linear functional
296: on $\AA(\Ir^s)$, which is positive in the sense that
297: $\omega(X^*X)\geq0$ and normalized as $\omega(\idty)=1$.
298: Equivalently, a state is given by a family $\omega_\Lambda$ of
299: density operators on $(\Cx^d)^{\otimes\Lambda}$ (for each finite
300: $\Lambda$), such that
301: $\omega(X)=\tr(\omega_\Lambda X)$ for $X\in\AA(\Lambda)$. The
302: local density matrices have to satisfy the consistency condition
303: that, for $\Lambda_1\subset\Lambda_2$, $\omega_{\Lambda_1}$ is
304: obtained from $\omega_{\Lambda_2}$ by tracing out all tensor
305: factors in $\Lambda_2\setminus\Lambda_1$. Note that a state does
306: not correspond to a configuration of a classical automaton, but
307: rather to a probability distribution over global configurations.
308:
309: \begin{definition}
310: A {\bf Quantum Cellular Automaton}
311: with neighborhood scheme $\Nei\subset\Ir^s$
312: is an homomorphism $\grule:\AA(\Ir^s)\to\AA(\Ir^s)$ of the quasi-local
313: algebra, which commutes with lattice translations, and satisfies
314: the locality condition
315: $\grule(\AA(\Lambda))\subset\AA(\Lambda+\Nei)$ for every finite
316: set $\Lambda\subset\Ir^s$.
317: The local {\bf transition rule} of a cellular automaton is the
318: homomorphism $\lrule:\AA_0\to\AA(\Nei)$. \end{definition}
319:
320: Note that a unitary operator for the time evolution is not necessary in
321: this formulation. Instead we have replaced it by its action on
322: observables. Of course, the time step has to be read in the Heisenberg
323: picture. That is, measuring some local observable $A\in\AA(\Lambda)$ at
324: time $t+1$ is equivalent to measuring the observable $\grule(A)$ at time
325: $t$. The transition rule thus describes this backwards calculation for a
326: single cell. $\lrule(\AA_0)$ is some isomorphic copy of the one-cell
327: algebra embedded in a possibly quite complicated way into the algebra of
328: neighboring cells. The relationship between the global evolution and the
329: one-site transition rule is as simple as it should be:
330:
331: \begin{lemma}\label{lem:localrule}{\ }\newline
332: (1) The global homomorphism $\grule$ is uniquely determined by the
333: local transition rule $\lrule$. \newline
334: (2) A homomorphism $\lrule:\AA_0\to\AA(\Nei)$ is the transition
335: rule of a cellular automaton if and only if for all $x\in\Ir^s$ such that
336: $\Nei\cap(\Nei+x)\neq\emptyset$ the algebras $\lrule(\AA_0)$ and
337: $\tau_x\bigl(\lrule(\AA_0)\bigr)$ commute elementwise. \end{lemma}
338:
339: \proof:\quad By translation invariance the action of
340: $\grule_x:\AA_x\to\AA(\Nei+x)$ is determined as
341: $\grule_x(A_x)=\tau_x\lrule\tau_{-x}(A_x)$ on all other cells.
342: Moreover, because $\grule$ is a homomorphism, the extension to any
343: local algebra is also fixed. Explicitly, consider a product
344: $\bigotimes_{x\in\Lambda}A_x=\prod_{x\in\Lambda}A_x$ of one-site
345: operators. This equation just expresses our identification of the
346: one site algebras $\AA_x$ with subalgebras of $\AA(\Lambda)$ by
347: tensoring with unit operators. Then the homomorphism property of
348: the global evolution requires that
349: \begin{equation}\label{alfaglobal}
350: \grule\left(\bigotimes_{x\in\Lambda} A_x\right)
351: =\prod_{x\in\Lambda} \grule_x(A_x)\;.
352: \end{equation}
353: Note that the product on the right hand side cannot be replaced
354: by a tensor product, because the factors have overlapping
355: localization regions $x+\Nei$. Moreover, the argument of $\grule$
356: is a product of commuting factors, hence so is the right hand
357: side. Hence the the commutativity condition (2) is necessary.
358:
359: Conversely, if the factors $\grule_x(A_x)$ commute, their product
360: is unambiguously defined. Since every local observable can be
361: expressed as a linear combination of tensor products,
362: Eq.~(\ref{alfaglobal}) defines a homomorphism on the local
363: algebra, as required. This shows the converse of (2), and since
364: we have given an explicit formula of $\grule$ in terms of $\lrule$
365: it also shows (1). \qed
366:
367:
368:
369: It is clear that the commutation condition of the Lemma can be
370: expressed as a finite set of equations, and can therefore be
371: verified effectively. Since only a small portion of the lattice is
372: needed in this verification, the same steps are needed to check
373: local transition rules fore QCAs on graphs which locally look like
374: $\Ir^s$. Such graphs can be seen as integer lattices with {\it
375: periodic boundary conditions}. The only condition we will need to
376: impose is that the periods for the boundary condition are not too
377: small compared with the size of neighborhood scheme $\Nei$ (a
378: condition called ``regularity'' below). No algebraic conditions on
379: the homomorphism $\lrule$ are needed. Therefore we can turn the
380: construction around, and immediately get a QCA on the infinite
381: lattice from a QCA with finitely many cells. Since for finitely
382: many cells a homomorphism is always just implemented by a unitary
383: matrix, this allows us to define QCAs even for the infinite
384: system, just by specifying unitary matrices with suitable
385: properties.
386:
387: Let us describe the periodic boundary QCAs more precisely and see
388: what conditions are needed for the neighborhoods. The cells of a
389: system with periodic boundary conditions arise from the cells in
390: $\Ir^s$ by identifying certain cells, namely all those differing
391: by a vector $\gamma$ in some subgroup $\Gamma\subset\Ir^s$. The
392: set of cells is thus identified with the quotient
393: $\loclat=\Ir^s/\Gamma$. For each point in $x\in\loclat$, i.e.,
394: each equivalence class $x=x_0+\Gamma$ of identified cells, only
395: one observable algebra is given. The sites $\loclat$ can be
396: represented in a so called fundamental domain of $\Gamma$ like the
397: parallelogram in Fig.~\ref{figlattis}. But there seems to be no
398: way to draw this nicely as a set of square cells.
399:
400: For points $x=(x_0+\Gamma)\in\loclat$, i.e., we can define the
401: translation $x+n=x_0+n+\Gamma$. Consider now a neighborhood scheme
402: $\Nei\subset\Ir^s$. The neighborhood of $x\in\loclat$ is then
403: $x+\Nei\subset\loclat$. Note that as far as the model with
404: periodic boundary conditions is concerned we could change each
405: $n\in\Nei$ by a lattice vector in $\Gamma$ without changing the
406: neighborhood of any point. But since we are interested in the
407: connection with the infinite model, we do not admit this
408: ambiguity. The neighborhood scheme is called {\it regular} for the
409: given periodic structure given by $\loclat$, if the equations we
410: have to check for the commutation rule of a cellular automaton are
411: the same in both cases. Since these equations depend only on the
412: intersections between translates of $\Nei$ we only need to make
413: sure that the geometry of intersections is the same. So suppose
414: that neighborhoods on $\loclat$ intersect, say
415: $(x+\Nei)\cap(y+\Nei)\neq\emptyset$. This means that there is a
416: translation $m\in\Ir^s$ such that $x+m=y$ and also
417: $\Nei\cap(m+\Nei)\neq\emptyset$. Clearly, the first condition
418: determines $m$ up to a lattice translation $\Gamma$, and the
419: second is an intersection condition on the infinite lattice.
420: Obviously, $x+(\Nei\cap(\Nei+m))\subset(x+\Nei)\cap(y+\Nei)$. But
421: the inclusion could be strict if there is more than one $m$ with
422: the required properties. This is precisely the case regularity
423: must exclude. To summarize, we call a neighborhood scheme
424: $\Nei\subset\Ir^s$ {\it regular} for a subgroup
425: $\Gamma\subset\Ir^s$, if $\Nei\cap(m+\Nei)\neq\emptyset$ and
426: $n\in\Gamma$, $n\neq0$, imply that $\Nei\cap(m+n+\Nei)=\emptyset$.
427: A more compact equivalent form is
428: $(\Nei+\Nei-\Nei-\Nei)\cap\Gamma=\{0\}$. An example of a regular
429: neighborhood is given in Figure~\ref{figlattis}. That the same
430: neighborhood becomes non-regular for a smaller lattice is shown in
431: Fig.~\ref{figirreg}.
432:
433:
434:
435: \begin{figure}[htb]
436: \epsfxsize=5cm \epsffile{peribound.eps}
437: \caption{\label{figlattis}{\it Periodic boundary conditions.
438: The parallelogram represents the given finite lattice $\loclat$,
439: where cells cut by opposite boundaries must be suitably joined.
440: The marks on the corners join up to a full circle.
441: Two translates of the same neighbourhood are shown.
442: It is regular, because the geometry of intersections is
443: the same as on the infinite plane. }}
444: \end{figure}
445:
446:
447: \begin{figure}[htb]
448: \epsfxsize=2.5cm \epsffile{irregular.eps}
449: \caption{\label{figirreg}{\it The same neighborhood scheme
450: is not regular on a 5$\times$4 torus: Again we have two translates,
451: but their intersection (cross-hatched) cannot be realized on the
452: infinite square lattice as intersection of two such neighborhoods. }}
453: \end{figure}
454:
455:
456:
457:
458:
459: Then checking condition (2) of the Lemma for the QCA
460: on $\loclat$ and for the QCA on $\Ir^s$ are exactly equivalent. We
461: call this useful principle the {\it Wrapping Lemma:}
462:
463: \begin{lemma} The QCA transition rules on a finite lattice
464: $\loclat$ with respect to a regular neighborhood scheme $\Nei$ are
465: in one-to-one correspondence with the transition rules for QCAs on
466: $\Ir^s$ with the same neighborhood scheme. \end{lemma}
467:
468:
469: \par
470:
471:
472:
473: \section{Basic Construction Methods}\label{sec:basicon}
474: \subsection{Commuting Unitaries and Phases }\label{sec:communitary}
475: Consider a unitary operator $U_0$ in some local algebra
476: $\AA(\widetilde\Nei)$, which commutes with all its translates
477: $U_x=\tau_x(U_0)$, up to a phase. That is, we require that there
478: are complex numbers $\zeta_x$ with $|\zeta_x|=1$ such that
479: $U_0\tau_x(U_0)=\zeta_x\tau_x(U_0)U_0$ for
480: $(\widetilde\Nei+x)\cap\widetilde\Nei\neq\emptyset$. This is
481: equivalent to
482: \begin{equation}\label{localprojective}
483: U_xU_y=\zeta_{y-x}\;U_yU_x
484: \end{equation}
485: We can then formally define a unitary operator
486: \begin{equation}\label{infUnitary}
487: \mbox{``\ }U=\prod_{{x\in\Ir^s}} U_x
488: \mbox{\ ''.}
489: \end{equation}
490: Here the the scare quotes indicate that there is no way this
491: infinite product can be made sense of. However, we can define
492: instead the action of this ``operator'' on local observables. To
493: this end, note that the actions $A\mapsto U_x^*AU_x$ commute for
494: different $x$. Moreover, if $x+\Nei$ does not intersect the
495: localization region of $A$, this action is the identity.
496: Therefore, in the infinite product of these operations only a
497: finite set is {\it not} the identity, and their product defines an
498: automorphism $\grule$. More formally, we have
499: \begin{equation}\label{commutealpha}
500: \grule(A)=\lim_{\Lambda\nearrow\Ir^s} U_\Lambda^*AU_\Lambda\;,
501: \end{equation}
502: where $A\in\AA(\Ir^s)$, and
503: $U_\Lambda=\prod_{{x\in\Lambda}}U_x$.
504: The limit is over any sequence of finite sets, eventually
505: absorbing all lattice points, and if $A$ is localized in a finite
506: region, the limit is actually constant for sufficiently large
507: $\Lambda$.
508:
509: The local transition rule is found by
510: applying $\grule$ to a single cell. This gives the neighborhood scheme
511: \begin{equation}\label{neiCommute}
512: \Nei=\bigcup_x\{\widetilde\Nei+x| 0\in\widetilde\Nei+x\}
513: =\widetilde\Nei-\widetilde\Nei\;.
514: \end{equation}
515:
516: We mention three special cases of this construction:
517: \begin{itemize}
518: \item When $\widetilde\Nei=\Nei=\{0\}$, we have to choose a unitary operator
519: $U_0\in\AA_0$ acting on a single cell. Thus the cellular automaton
520: acts by applying the same unitary rotation separately to every
521: cell.
522: \item Fix a basis in $\Cx^d$ and
523: consider any unitary $U_0$ which is diagonal in the corresponding
524: product basis. Clearly, this guarantees that $U_0$ commutes with
525: its translates. That is, we can generate a QCA by an arbitrary
526: choice of {\it local phases} in some computational basis. A
527: prominent example is an Ising interaction
528: \begin{equation}\label{ising}
529: H=(\idty-\sigma_3)\otimes(\idty-\sigma_3)
530: \end{equation}
531: turned on for a suitable finite time (e.g., $t=\pi/4$), which
532: generates the initial entangled state of a one-way quantum
533: computer \cite{onewaycomp} together with a cellwise Hadamard
534: rotation. \item Choose at every site a family of unitaries $V_p$,
535: $p=1,\ldots,d^2$ commuting up to a phase. Examples are the Pauli
536: matrices for $d=2$, or a discrete Weyl system. Then $U_0$ can be
537: any finite product of operators from this family, localized on
538: neighboring cells. The automata generated in this way form a
539: group.
540: \end{itemize}
541:
542: \noindent For all QCAs constructed from commuting unitaries we
543: have the strange property that they show {\it no propagation}, in
544: the sense that the localization region does not increase when we
545: iterate the automaton. The reason is that $n$ steps are
546: implemented by a product of commuting unitaries, each of which
547: appears $n$ times. Hence it is exactly equivalent to take instead
548: a single step with the basic unitary operator $U_0$ replaced by
549: $U_0^n$.
550:
551: Of course, most QCAs do have propagation. For example, we could
552: take a single step constructed from commuting unitaries, followed
553: by a site-wise rotation along a skew axis. Another rich class of
554: propagating QCAs is given in the next section.
555:
556: \subsection{Quantization of Classical Reversible CAs}\label{sec:classical}
557: A typical feature of the local phase automata is that they leave
558: invariant the algebra of operators diagonal in the chosen basis.
559: This algebra $\DD$ is the quasi-local algebra of a classical CA
560: embedded into the quantum system, which has $d$ states per cell,
561: when $\AA_x$ is the algebra of $d\times d$-matrices. It is
562: therefore natural to look for cellular automata, which leave this
563: classical subalgebra $\DD$ invariant as a set, but not
564: elementwise. Clearly, such QCAs induce a classical CA on the
565: classical subsystem. In fact, {\it every} reversible classical CA
566: can be obtained in this way.
567:
568: Intuitively, this is seen as follows: the classical CA can be run
569: with periodic boundary conditions, for simplicity, so we have only
570: a finite set $\loclat$ of cells. It then defines a permutation of
571: the $d^{|\loclat|}$ classical configurations, which we can
572: interpret immediately as a unitary permutation operator $U$. This
573: unitary operator is now used to implement the local evolution of
574: the QCA. All we need to verify is that for $A\in\AA_0$ the
575: evolution $U^*AU$ is indeed contained in some local algebra
576: $\AA(\Nei)$ for some regular neighborhood $\Nei$. Then the
577: wrapping Lemma asserts that the QCA is also well-defined on the
578: infinite lattice.
579:
580: However, this argument must be more subtle than it looks: it is
581: well known, that the injectivity of a classical CA on the infinite
582: lattice implies the existence of an inverse CA \cite{Rich}, but
583: the inverse is hard to compute, because there is no a priori upper
584: bound on its neighborhood size. Superficially, the inverse
585: neighborhoods do not seem to enter the above argument. However,
586: the argument for the locality $U^*AU\in\AA(\Nei)$ requires more
587: than the locality of the classical rule and the unitarity of $U$.
588: Consider, for example, the rule
589: \begin{equation}\label{xorautomat}
590: c^{t+1}_{x}=c^{t}_{x+1}+c^{t}_{x}+c^{t}_{x-1}\;,
591: \end{equation}
592: where $c^{t}_{x}\in\{0,1\}$, and addition is mod$2$. With
593: periodic boundary conditions of length $L$ this is an invertible
594: transformation unless $L$ is divisible by 3. This proviso is not
595: of the form ``for sufficiently large $L$...'' , which means that
596: the classical automaton does not allow a local inversion, i.e.,
597: there is no inverse cellular automaton. By the wrapping Lemma, it
598: is clear that for this rule we cannot find a quantum version
599: either\cite{classicalW}. This shows that the {\it local}
600: invertibility of the classical CA must enter the argument. We
601: therefore assume now that the classical CA is locally invertible,
602: and the lattice $\loclat$ is chosen sufficiently large, so that
603: the neighborhood schemes for both the classical automaton and its
604: inverse are regular for $\loclat$.
605:
606: The classical configurations are functions
607: $\underline{a}:\loclat\to A$, where $A=\{1,\ldots,d\}$ denotes the
608: set of classical states for each cell, and at the same time labels
609: the computational basis of $\Cx^d$. We will write
610: $\underline{a}\in A^\loclat$, and denote by
611: $\underline{a}_x=\underline{a}(x)$ the value of the cell $x$ in
612: configuration $\underline{a}$.
613:
614: Extending this to a product basis, each configuration
615: $\underline{a}\in A^\loclat$ determines a basis vector
616: $\ket{\underline{a}}$. Then the global unitary transition operator
617: is defined by $U\ket{\underline{a}}=\ket{F({\underline{a}})}$,
618: where $F$ denotes the global classical transition function. The
619: transition rule of the QCA is determined by computing all matrix
620: elements of the operator $\lrule(\ketbra{c_0}{e_0})$, i.e.,
621: \begin{eqnarray}\label{classmat}
622: \bigl\langle \underline{a}\bigr|\lrule(\ketbra{c_0}{e_0}) \bigl|\underline{b}\bigr\rangle
623: &=&\bigl\langle \underline{a}\bigr|U^*(\ketbra{c_0}{e_0}\otimes
624: \idty^{\loclat\setminus\{0\}})U
625: \bigl|\underline{b}\bigr\rangle
626: \nonumber\\
627: &=&\bigl\langle F(\underline{a})\bigr|(\ketbra{c_0}{e_0}\otimes
628: \idty^{\loclat\setminus\{0\}})
629: \bigl|F(\underline{b})\bigr\rangle
630: \nonumber
631: \end{eqnarray}
632: This expression is $=1$, if
633: \begin{eqnarray}\label{transCCA}
634: \mbox{if}\ F(\underline{a})_0&=&c_0,\quad
635: F(\underline{b})_0=e_0,\ \nonumber\\
636: \mbox{and}\ F(\underline{a})_x&=&F(\underline{b})_x
637: \quad\mbox{for}\ x\neq0
638: \end{eqnarray}
639: and $=0$ otherwise. We have to show that this is of the form
640: $X\otimes\idty^{\otimes\loclat\setminus\Nei}$.
641:
642: \begin{lemma}\label{classlem}Let $F$ be a classical cellular automaton
643: with neighborhood scheme $\Nei_C$, which has an inverse automaton
644: with neighborhood scheme $\Nei_I$. Then $\grule$ as defined above
645: is a QCA with neighborhood scheme $\Nei=\Nei_C-\Nei_C-\Nei_I$.
646: \end{lemma}
647:
648: Of course, this Lemma only gives an upper bound on the size of the
649: neighborhood scheme. Depending on the particular automaton, $\Nei$
650: may be much smaller.
651:
652: \proof\/: We have to show that for all $e_0,c_0$, the operator
653: $\lrule(\ketbra{c_0}{e_0})$ is of the form
654: $X\otimes\idty^{\otimes\loclat\setminus\Nei}$ with
655: $X\in\AA(\Nei)$. This is equivalent to two conditions: on the one
656: hand the matrix elements $\bigl\langle
657: \underline{a}\bigr|\lrule(\ketbra{c_0}{e_0})
658: \bigl|\underline{b}\bigr\rangle$ must vanish, whenever $a_x\neq
659: b_x$ for some point $x\notin\Nei$, and, moreover, the value of the
660: matrix elements must be independent of $a_x$ for such $x$.
661:
662: Suppose that the matrix element is non-zero, i.e.,
663: condition~(\ref{transCCA}) holds. Then for all $y$ such that
664: $y+\Nei_I\neq0$ the computation of the the values of $a_y$ and
665: $b_y$ from $F(\underline{a})$ and $F(\underline{b})$ will give the
666: same values of $a_y$ and $b_y$. In other words, the diagonality
667: condition holds for $y\notin(-\Nei_I)$.
668:
669: The dependence of the matrix element on $a_x$ is also governed by
670: condition~(\ref{transCCA}): since the matrix elements can only be
671: $0$ or $1$, we have to show that the validity of the condition
672: does not depend on $a_x$ for $x\notin\Nei$, given the range of
673: equality of $a'$s and $b'$s established in the previous paragraph.
674: Indeed, once that diagonality is established, one can see that
675: most of the conditions~(\ref{transCCA}) become redundant. If $x$
676: is such that $(x+\Nei_C)\cap(-\Nei_I)=\emptyset$, then
677: $F(\underline{a})_x$ and $F(\underline{b})_x$ are computed by the
678: local rule of $F$ from identical data, so they must be equal. It
679: therefore suffices to consider the condition for those finitely
680: many $x$ for which a dependence remains possible, i.e.,
681: $x\in(-\Nei_C-\Nei_I)$. But then only those $a_y$ enter, which
682: contribute to $F(\underline{a})_x$ via the local rule (and
683: similarly for $\underline{b}$). This restricts $y$ to
684: $(\Nei_C-\Nei_C-\Nei_I)$, and this is what the Lemma claims as the
685: localization region. \qed
686:
687:
688: \subsection{Partitioning}
689: The easiest way to build a cellular automaton with readily
690: verified locality properties is based on the cellwise unitary
691: rotations, with the modification of changing the partitioning of
692: the system into cells \cite{Lloyd}. The typical construction would
693: thus be:
694: \begin{enumerate}
695: \item possibly divide the given cells into suitable subcells, by
696: writing the one-cell Hilbert space $\Cx^d$ as a tensor product of
697: other spaces.
698: \item partition the set of cells of the previous step into blocks in some periodic
699: way: every cell now belongs to exactly one block, and any two blocks
700: are connected by a lattice translation.
701: \item apply the same unitary operator to each block algebra.
702: \item possibly repeat this procedure with different block
703: partitions
704: \item possibly split and regroup once again to come back to the original pattern of
705: $d$-dimensional cells.
706: \end{enumerate}
707:
708: Every step in this construction is well-defined for the global
709: system for the same reason that cell-wise rotations are valid QCA
710: operations. Moreover, the unitary operators doing each of the
711: steps are essentially arbitrary, and by just inverting the steps
712: we can immediately construct the inverse QCA. This is why
713: partitioned cellular automata are sometimes called {\it
714: structurally reversible}. Moreover, the partitioning idea allows
715: one to construct QCAs from {\it irreversible} local rules just as
716: easily as reversible ones \cite{Brennen,RiWe}.
717:
718: Our main theorem (Theorem~\ref{mainthm} below) will tell us that
719: {\it every} QCA can be written in partitioned form. For nearest
720: neighbor automata it is sufficient to take two steps in which
721: cells are grouped in cubes of side 2, possibly with a choice of
722: unequal cell sizes in the intermediate step (see
723: Section~\ref{sec:margolus}). This is known as the Margolus
724: partitioning scheme (see Figure~\ref{margolus}).
725:
726: \begin{figure}[htb]
727: {\epsfxsize=5cm \epsffile{margolus0.eps}
728: \caption{\label{margolus}{\it The Margolus partitioning scheme in $s=2$ dimensions.
729: Operations are alternatingly applied to the solid and to the dashed
730: partitioning into $2\times2$ squares. The square shape of the cells is irrelevant,
731: as they only serve to label localized quantum systems.}}}
732: \end{figure}
733:
734:
735:
736: \subsection{Circuits}
737:
738: Of course, it is natural to think of a cellular automaton as a
739: physical device, which just happens to be infinitely extended (or
740: periodically closed). This suggests building QCAs from some basic
741: supply of circuit elements, whose properties will then ensure that
742: the overall operation makes sense. The mathematical details of the
743: description must then work out automatically, because ``Hardware
744: cannot lie''. Consider the following example
745:
746: \begin{figure}[htb]
747: \epsfxsize=5cm \epsffile{circuit1.eps}
748: \caption{\label{circ1}{\it A proposed QCA circuit.}}
749: \end{figure}
750:
751: Here the flow of information is from top to bottom. The symbol
752: stands for a CNOT gate, by which the bit on the line with circle
753: and cross (the ``target bit'') is flipped if and only if the value
754: of the bit on the line with the fat dot (the ``control bit'') is
755: ``$1$''. This is a standard gate also for quantum computation. We
756: can readily compute the action of this device on classical
757: information: each output bit depends only on the input on the same
758: line and the line one step to the right. Presumably, this would
759: also be the description of the action on computational basis
760: states in the quantum case. So can we not just build this device,
761: and construct its proper mathematical description along with the
762: hardware?
763:
764: The problem with this automaton becomes apparent already when we
765: try to compute its classical inverse: this requires at each switch
766: to know the value of a control bit, which is not yet determined.
767: In fact, the inverse does not exist, because the initial states
768: ``all {\tt1}'' and ``all {\tt0}'' are both mapped to ``all
769: {\tt0}''. So this device is not reversible, even classically.
770: (Incidentally, this also holds for periodic boundary conditions,
771: so it is not a problem of infinite size). What went wrong? The
772: problem is the {\it timing} of the gates. In fact, in the usual
773: gate model of quantum computation it is assumed that each gate is
774: executed at a certain time. Here the times overlap, and if we
775: insist on the gates being executed, say from the left to the
776: right, we either need infinitely many operation times per step, or
777: we run into time ordering problems at the boundary condition.
778:
779: One way to avoid this problem is to insist on some finite number
780: of clock cycles per QCA step, that each gate is executed in one of
781: these cycles, and that the gates running in the same cycle do not
782: access the same registers. An ``unscrambled version'' of the above
783: impossible QCA is drawn in Fig.~\ref{circ2}
784:
785: \begin{figure}[htb]
786: \epsfxsize=5cm \epsffile{circuit2.eps}
787: \caption{\label{circ2}{\it An operational QCA circuit
788: taking two clock cycles}}
789: \end{figure}
790:
791: Clearly, this is a partitioning QCA and, in fact, the description
792: we just gave of a circuit with timing constraints is nothing but
793: the definition of a partitioning QCA.
794:
795: \subsection{Clifford automata}
796:
797: In the implementations of quantum computation there is often a
798: separation between ``easily implemented'' operations and others,
799: which may be more costly. For example, ''linear'' transformations
800: on the quantum light field can be performed with mirrors, beam
801: splitters and phase plates, whereas squeezing or photon number
802: counting are more costly. A similar choice of a subgroup of
803: ``easy'' operations for qubit quantum computation is the group of
804: transformations, which take tensor products of Pauli matrices into
805: tensor products of Pauli matrices, the so called {\it Clifford
806: group}\cite{cliff}. Indeed in some implementations these play a
807: special role, and can be executed in parallel \cite{onewaycomp}.
808: Clearly, it is important to understand this subgroup completely,
809: although it is also clear that such transformations alone will not
810: allow quantum computational speedup.
811:
812: The natural mathematical setting for the investigation of Clifford
813: QCAs are discrete Weyl systems, with a finite Weyl system acting
814: at each site, and the tensor products of one-site Weyl operators
815: generating a Weyl system with infinitely many degrees of freedom.
816: If the local Weyl system has prime dimension $d$, one can give a
817: very explicit description of the group of Clifford QCAs, which
818: will be presented elsewhere \cite{OurCliff}. For illustration let
819: us just take qubit automata $d=2$ in one dimension.
820:
821: What is needed to define such an automaton? Since the Pauli
822: matrices $\sigma_x$ and $\sigma_z$ generate the one site algebra,
823: we only need to specify the two operators $\lrule(\sigma_x)$ and
824: $\lrule(\sigma_y)$. The Clifford property means that these two
825: operators must be tensor products of Pauli matrices, so we have
826: \begin{eqnarray}\label{CQCAstring}
827: \lrule(\sigma_x)&=&\sigma_{\xi_{-N}}\otimes\cdots\sigma_{\xi_{0}}
828: \otimes\cdots\sigma_{\xi_{N}}
829: \equiv\sigma(\xi)\\
830: \lrule(\sigma_x)&=&\sigma_{\eta_{-N}}\otimes\cdots\sigma_{\eta_{0}}
831: \otimes\cdots\sigma_{\eta_{N}}
832: \equiv\sigma(\eta)\;,\nonumber
833: \end{eqnarray}
834: where each $\xi_i,\eta_i$ can take the values $0,x,y,z$, with
835: $\sigma_0=\idty$. Now the two strings
836: $\xi=(\xi_{-N},\ldots,\xi_{N})$ and
837: $\eta=(\eta_{-N},\ldots,\eta_{N})$ completely characterize the
838: automaton. But which strings are allowed? From the general theory
839: we immediately get the necessary and sufficient conditions:
840: $\sigma(\xi)$ must commute with all its translates, the same holds
841: for $\sigma(\eta)$. Moreover, $\sigma(\xi)$ commutes with
842: $\tau_i(\sigma(\eta))$ for $i\neq0$ and anti-commutes for $i=0$.
843: Since tensor products of Pauli matrices always either commute or
844: anti-commute, one can run a computer search for all examples with
845: low neighborhood size $N$. This turns up the surprising result
846: that (up to a common translation) the strings $\xi$ and $\eta$
847: must be {\it palindromes}, i.e., $\xi_{-k}=\xi_k$.
848:
849: The general theory \cite{OurCliff} confirms this, and moreover
850: etablishes an isomorphism of the group of Clifford QCAs with the
851: group of $2\times2$-matrices,
852: \begin{equation}\label{CQCAmat}
853: \left(\begin{array}{cc}
854: \xi_+(z)&\eta_+(z)\\\xi_-(z)&\eta_-(z)
855: \end{array}\right)\;,
856: \end{equation}
857: whose entries are polynomials over the two-element field
858: $\Field_2=\{0,1\}$ in one indeterminate $z$, such that the
859: determinate is the constant polynomial $1$. Here the coefficients
860: of $\xi_\pm$ are bit strings which together determine the string
861: $\xi_0,\xi_1,\ldots\xi_N$ used in Eq.~(\ref{CQCAstring}).
862:
863: One can also find a simple set of generators: Apart from one-site
864: transformations and the shift only one QCA is needed:
865: \begin{eqnarray}\label{CQCAxzx}
866: \lrule(\sigma_x)&=&\idty\otimes\sigma_{z}
867: \otimes\idty\\
868: \lrule(\sigma_x)&=&\sigma_{z}\otimes\sigma_{x}\otimes\sigma_{z}\;,\nonumber
869: \end{eqnarray}
870: possibly, however, acting not between neighboring sites as written
871: here, but between sites at a fixed distance $L$, so that the chain
872: breaks up into $L$ non-interacting chains.
873:
874: The prototype (\ref{CQCAxzx}) has a number of interesting
875: properties. For example, the iterates $T^t(\sigma(\zeta))$ of an
876: initial Pauli product $\sigma(\zeta)$ have a specific form: they
877: consist of a Pauli product moving to the left, and another one
878: moving to the right, each at maximal speed, and the expanding
879: space between these patterns is filled by one of four
880: possibilities: all $\idty$, all $\sigma_y$ or an alternating
881: patterns of $\sigma_x\otimes\sigma_y$, or the same shifted by one
882: cell.
883:
884:
885:
886: \section{Structure}\label{sec:struc}
887: In this section we will employ the commutativity property of
888: transition rules to get some information about the structure of
889: possible rules, aiming at the proof that all QCA transition rules
890: can be understood in a partitioning scheme.
891:
892: Without loss of generality, we consider only nearest neighbor
893: rules on a cubic lattice. For a non-cubic lattice we can choose a
894: family of basic cells (a ``fundamental domain'') such that all
895: cells are generated from these basic ones by translation
896: symmetries. By considering the fundamental domain and its
897: translates as new cells, we effectively get a family of lattice
898: cells labelled by $\Ir^s$. If the neighborhood scheme involves
899: more than nearest neighbors, we can again enlarge cells. Of
900: course, these operations partly destroy the underlying lattice
901: symmetry, so that operations on regrouped cells may fail to have
902: the translation (or other) symmetry of the original lattice.
903: However, for the proof of structural invertibility a regrouped
904: lattice of ``supercells'' works just as well.
905:
906: We begin by describing the geometry of the generalized Margolus
907: partitioning scheme. In the following subsection we state the main
908: theorem: this scheme indeed suffices for all QCAs. The proof
909: relies on the concept of ``support algebras'', and is described in
910: Subsection~\ref{sec:supp}. When the support algebras are abelian,
911: one can characterize the possible QCAs at the single cell level,
912: without partitioning (see Subsection~\ref{sec:abelianDD}). This
913: allows us to determine explicitly all nearest neighbor qubit
914: automata in one dimension (Section~\ref{sec:qubits}).
915:
916:
917: \subsection{Generalized Margolus Partitioning}
918: \label{sec:margolus}
919:
920: Consider a cellular automaton with one-site algebra $\AA_0=\MM_d$,
921: lattice $\Ir^s$, and {\it nearest neighborhood}
922: scheme:
923: \begin{equation}\label{NN}
924: \Nei=\{x\in\Ir^s|\ \forall_i|x_i|\leq1\}\;.
925: \end{equation}
926: We will use a cell grouping introduced by
927: Margolus \cite{MarTo}. In this scheme one ``supercell'' is the
928: unit cube
929: \begin{equation}\label{cube}
930: \Box=\{x\in\Ir^s|\ \forall_i x_i=0,1\}\;,
931: \end{equation}
932: which consists of $2^s$ cells. The even translates $\Box+2x$ with
933: $x\in\Ir^s$ cover the whole lattice. We denote by $Q$ the set of
934: $2^s$ {\it quadrant vectors} $q\in\Ir^s$ for which each component
935: is $q_i\in\{-1,+1\}$ (see Fig~\ref{margolusT}). In particular the
936: vector into the positive quadrant (with all $q_i=+1$) will be
937: denoted by $\qone$. Note that the sum of two quadrant vectors is
938: an even lattice translation in $2\Ir^s$. Moreover, the $2^s$ cubes
939: $\Box+q$ for $q\in Q$ are disjoint and their union contains all
940: neighborhoods of cells in $\Box$. Just like the even translates of
941: $\Box$, the cubes $\Box+\qone+2x$ with $x\in\Ir^s$ form a
942: partition of the lattice (compare Fig.~\ref{margolus}).
943:
944: \begin{figure}[htb]
945: \epsfxsize=5cm \epsffile{margolusT.eps}
946: \caption{\label{margolusT}{\it The Margolus scheme in $s=2$ dimensions.
947: The basic cube `` $\Box$'' is at the center, with the origin marked ``\,$0$''.
948: The shaded squares are reached from the origin by quadrant vectors.
949: Dashed outlines mark the copies $\Box+q$ shifted by quadrant vectors.}}
950: \end{figure}
951:
952: A partitioned automaton based on the Margolus scheme would
953: alternatingly apply blockwise unitary transformations to the cubes
954: in the two partitions. But not every QCA can be written in this
955: way. A good counterexample is the shift in the quadrant direction
956: $\qone$. Here the entire quantum information in the $\Ir^s$ cells
957: $\Box$ will have to be moved to $\Box+\qone$ although these blocks
958: have only a single cell as overlap. It turns out, however, that a
959: slight generalization of the Margolus scheme suffices to represent
960: every QCA: all we have to do is to allow different cell sizes in
961: the intermediate step. For the rest of this subsection we will
962: explain the resulting scheme.
963:
964: With each quadrant vector $q\in Q$ we associate an observable
965: algebra $\BB_q\subset\AA(\Box+q)$, which is isomorphic to the
966: algebra of $n(q)\times n(q)$-matrices for some integer $n(q)$.
967: Since $\BB_q$ is contained in a local algebra, it makes sense to
968: consider its (even) translates $\tau_x(\BB_q)\subset\AA(\Box+q+x)$
969: with $x\in2\Ir^s$. In particular, $\AA(\Box+\qone)$ contains all
970: the algebras $\tau_{\qone-q}(\BB_q)$. A crucial assumption of our
971: construction is that these subalgebras of $\AA(\Box+\qone)$
972: commute, and together span $\AA(\Box+q)$. This is possible if and
973: only if the equation
974: \begin{equation}\label{blockdims}
975: \prod_{q\in Q}n(q)=\dtwos
976: \end{equation}
977: holds for the matrix dimensions. Note that each cell $\AA(\Box+x)$
978: has this dimension, whether or not $x$ is even or not.
979:
980: The local rule of an automaton now defines (and is defined by) a
981: homomorphism
982: \begin{equation}\label{locMarg}
983: \brule:\AA(\Box)\to\prod_{q\in Q}\BB_q\;.
984: \end{equation}
985: Since the dimensions of domain and range are the same, such a
986: homomorphism is necessarily an isomorphism, i.e., we can find
987: suitable bases in each cell and for each of the matrix algebras
988: $\BB_q$ such that $\brule(A)=UAU^*$ for a unitary operator
989: \begin{equation}\label{bruleU}
990: U:\bigotimes_{x\in\Box}\Cx^d\longrightarrow\bigotimes_q
991: \Cx^{n(q)}\;.
992: \end{equation}
993: Such a unitary operator by itself does not fix a QCA, because if
994: we only take $\BB_q$ as an abstract matrix algebra, we still need
995: to specify how $\tau_{\qone-q}\BB_q$ is contained in
996: $\AA(\Box+\qone)$ or, equivalently, to specify the isomorphism of
997: $\bigotimes_q\tau_{\qone-q}\BB_q$ with $\AA(\Box+\qone)$. This
998: will be affected by another unitary operator
999: \begin{equation}\label{bruleV}
1000: V:\bigotimes_q \Cx^{n(q)}\longrightarrow\bigotimes_{x\in\Box+\qone}\Cx^d\;.
1001: \end{equation}
1002:
1003: Any pair of unitaries $(U,V)$ according to
1004: Eqs.~(\ref{bruleU},\ref{bruleV}) specifies a transformation $T$ on
1005: local algebras, which satisfies all requirements for a cellular
1006: automaton, except translation invariance: $T$ only commutes with
1007: even translations. The scheme in one and two lattice dimensions is
1008: visualized in Figures~\ref{margolus1}-\ref{figchops}.
1009:
1010: \begin{figure}[htb]
1011: \epsfxsize=5cm \epsffile{test.eps}
1012: \caption{\label{margolus1}{\it Generalized Margolus Scheme in $s=1$ dimension.
1013: The algebras $\BB_{+1}$ and $\BB_{-1}$ are symbolized by the different size cells
1014: in the intermediate step.}}
1015: \end{figure}
1016:
1017: \begin{figure}[htb]
1018: \epsfxsize=7cm \epsffile{pegacyc.eps}
1019: \caption{\label{pegasus}{\it Generalized Margolus Scheme in $s=2$ dimensions.
1020: The four subalgebras $\BB_q$ are symbolized by the shapes in Fig.~\ref{figchops}.
1021: The tesselation is due to M.C. Escher (1956). }}
1022: \end{figure}
1023: %
1024: %\vskip40pt
1025: \begin{figure}[htb]
1026: \epsfxsize=3cm \epsffile{pegpiece.eps}
1027: \caption{\label{figchops}{\it The four shapes representing the algebras
1028: $\BB_q$ in Fig.~\ref{pegasus}. The arrows indicate the appropriate quadrant vectors $q$.
1029: }}
1030: \end{figure}
1031:
1032:
1033: If we want $\grule$, as constructed from unitaries $U$ and $V$, to
1034: be a proper cellular automaton with full translation invariance,
1035: there will be additional conditions on these unitaries.
1036: Unfortunately, these conditions are not easily written down and
1037: solved in the general case. We also note that $U$ and $V$ are not
1038: uniquely determined by the automaton: we have the freedom to
1039: choose a basis in every $\Cx^{n(q)}$. Changing this basis amounts
1040: to a cellwise rotation included in $U$, which is immediately
1041: undone by the $V$-step. The key feature of automata in a
1042: partitioned scheme is {\it structural reversibility}, as described
1043: in the following Lemma.
1044:
1045: \begin{lemma}\label{invMarge}
1046: Let $\grule$ be a homomorphism constructed from unitaries $(U,V)$
1047: in the generalized Margolus scheme. Then $\grule$ is invertible,
1048: and $\grule^{-1}$ is also a generalized Margolus automaton.
1049: Moreover, if $\grule$ commutes with all (not just even)
1050: translations, then both $\grule$ and $\grule^{-1}$ are nearest
1051: neighbor QCAs.
1052: \end{lemma}
1053:
1054: \proof: The Margolus unitaries defining $\grule^{-1}$ are
1055: $(V^*,U^*)$. To check the localization properties, it is helpful
1056: not to think of these transformations in the apparently
1057: time-asymmetric scheme of Fig.~\ref{pegasus}, but to take $\BB_q$
1058: as an algebra localized in the intersection of the cubes $\Box$
1059: and $(\Box+q)$, i.e., as localized at the cell $(\qone+q)/2$. Note
1060: that $\grule^{-1}$ is a two-sided inverse, and hence uniquely
1061: determined by $\grule$.
1062:
1063: Therefore, if $\grule$ commutes with all translations, so does
1064: $\grule^{-1}$. It remains to check that if $\grule$ is a Margolus
1065: automaton commuting with translations, it is actually a nearest
1066: neighbor automaton, i.e., a QCA with neighborhood scheme $\Nei$
1067: from (\ref{NN}). Since, for every $x\in\Box$, we have
1068: $0\in(\Box-x)$, we have
1069: \begin{equation}
1070: \grule(\AA_0)
1071: \subset\grule(\AA(\Box-x))
1072: \subset\bigotimes_{q\in Q} \AA(\Box+q-x)\;.
1073: \end{equation}
1074: Since this is valid for {\it all} $x\in\Box$, we have
1075: $\grule(\AA_0)\subset\AA(\widetilde\Nei)$, with
1076: \begin{equation}\label{neimarg}
1077: \widetilde\Nei=\bigcap_{x\in\,\Box}\bigcup_{q\in Q}(\Box+q-x)=\Nei
1078: \end{equation}
1079: \qed
1080:
1081:
1082:
1083:
1084: \subsection{Main Theorem}
1085:
1086: \vbox{
1087: \begin{theorem}\label{mainthm}
1088: Let $\grule$ be the global transition homomorphism of a nearest
1089: neighbor quantum cellular automaton on the lattice $\Ir^s$ with
1090: single-cell algebra $\AA_0=\MM_d$. Then $\grule$ can be
1091: represented in the generalized Margolus partitioning scheme, i.e.,
1092: $\grule$ restricts to an isomorphism
1093: \begin{equation}
1094: \grule:\AA(\Box)\longrightarrow \bigotimes_{q\in Q}\BB_q \;,
1095: \end{equation}
1096: where for each quadrant vector $q\in Q$, the subalgebra
1097: $\BB_q\subset\AA(\Box+q)$ is a full matrix algebra,
1098: $\BB_q\cong\MM_{n(q)}$. These algebras and the matrix dimensions
1099: $n(q)$, which satisfy Eq.~(\ref{blockdims}), are uniquely
1100: determined by $\grule$.
1101: \end{theorem}
1102: }
1103:
1104:
1105: Combining this with Lemma~\ref{invMarge}, we get
1106:
1107: \begin{corr}\label{cor:invert}
1108: The inverse of a nearest neighbor QCA exists, and is a nearest
1109: neighbor QCA.
1110: \end{corr}
1111:
1112: Note that this result is in stark contrast to the classical
1113: situation. In the classical case the inverse of an injective CA is
1114: a CA, i.e., locally invertible, but it is a highly non-trivial
1115: matter to determine the neighborhood scheme of the inverse, which
1116: can be much larger than the neighborhood of the automaton itself.
1117: This is not a contradiction with the observation that every
1118: classical CA can be quantized (see Section~\ref{sec:classical}):
1119: In order to construct a QCA from a classical CA, we needed the
1120: neighborhood size of the inverse.
1121:
1122: The proof of the Theorem will be given in
1123: subsection~\ref{sec:proof}. The key idea is to construct
1124: $\BB_q\subset\AA(\Box+q)$ explicitly from the inclusion
1125: $\grule(\AA(\Box))\subset\bigotimes_q\AA(\Box+q)$. This
1126: construction, which will also be useful independently, will be
1127: described in the next section.
1128:
1129:
1130: \subsection{Support Algebras of Local Rules}\label{sec:supp}
1131:
1132: By definition, the transition rule $\lrule$ maps one cell algebra
1133: into a tensor product of neighboring ones. Therefore we need to
1134: investigate just how one subalgebra can sit inside a tensor
1135: product of others.
1136:
1137: Consider a subalgebra $\AA\subset\BB_1\otimes\BB_2$ of tensor
1138: product. For the moment let us forget about the multiplication
1139: laws, and just consider these as vector spaces, with the tensor
1140: product known from the (multi-)linear algebra of finite
1141: dimensional vector spaces. Then we can expand each $a\in\AA$ into
1142: a sum $a=\sum_\mu b_\mu^{(1)}\otimes b_\mu^{(2)}$. But we might
1143: get by just using a small subset of operators $b_\mu^{(i)}$. The
1144: smallest subspace of $\BB_1$ sufficient for these expansions will
1145: be called the {\it support} of $\AA$ on the first factor, and will
1146: be denoted by $\spp(\AA,\BB_1)$. For a more formal definition note
1147: that each $a\in\AA$ can be expanded uniquely in the form
1148: $a=\sum_\mu a_\mu\otimes e_\mu$, where $\{e_\mu\}$ is fixed a
1149: basis of $\BB_2$. Then $\spp(\AA,\BB_1)$ is the linear span of all
1150: $a_\mu$ in this expansion, and is clearly independent of the basis
1151: $\{e_\mu\}$ chosen for $\BB_2$. Then it is clear that
1152: $b_\mu^{(i)}\in \spp(\AA,\BB_i)$ indeed suffice to expand every
1153: $a\in\AA$, i.e.,
1154: \begin{equation}\label{a2supp}
1155: \AA\subset \spp(\AA,\BB_1)\otimes \spp(\AA,\BB_2)
1156: \subset\BB_1\otimes\BB_2\;.
1157: \end{equation}
1158: The analogous relation for $\AA$ contained in a tensor product of
1159: more factors is seen in the same way.
1160:
1161: Now we remember the algebraic structure: $\spp(\AA,\BB_i)$ is a
1162: linear subspace of a C*-algebra, so we can define the {\it support
1163: algebra} $\Spp(\AA,\BB_i)$ of $\AA\subset \bigotimes_i\BB_i$ on
1164: one factor $\BB_i$ as the subalgebra of $\BB_i$ generated by the
1165: elements of $\spp(\AA,\BB_i)$ \cite{Zanardi}. Then we also have
1166: \begin{equation}\label{a2Supp}
1167: \AA\subset \Spp(\AA,\BB_1)\otimes \Spp(\AA,\BB_2)
1168: \subset\BB_1\otimes\BB_2\;.
1169: \end{equation}
1170: Note that since any observable algebra $\AA$ is closed under adjoints, so is
1171: $\Spp(\AA,\BB_1)$. The crucial fact we need about such inclusions
1172: is the following:
1173:
1174: \begin{lemma} \label{sppcomm}
1175: Let $\AA_1\subset\BB_1\otimes\BB_2$ and
1176: $\AA_2\subset\BB_2\otimes\BB_3$ be subalgebras such that
1177: $\AA_1\otimes\idty_3$ and $\idty_1\otimes\AA_2$ commute in
1178: $\BB_1\otimes\BB_2\otimes\BB_3$. Then $\Spp(\AA_1,\BB_2)$ and
1179: $\Spp(\AA_2,\BB_2)$ commute in $\BB_2$.
1180: \end{lemma}
1181:
1182: \proof:\quad Pick bases $\{e_\mu\}\subset\BB_1$ and
1183: $\{e'_\nu\}\subset\BB_2$, and let $a\in\AA_1$ and $a'\in\AA_2$.
1184: Then we may expand uniquely: $a=\sum_\mu e_\mu\otimes a_\mu$ and
1185: $a'=\sum_\nu a'_\nu\otimes e'_\nu$. Then by assumption
1186: \begin{displaymath}
1187: 0=[a\otimes\idty_3,\idty_1\otimes a']
1188: =\sum_{\mu\nu}e_\mu\otimes [a_\mu,a'_\nu]\otimes e'_\nu \;.
1189: \end{displaymath}
1190: Now since the elements $e_\mu\otimes e'_\nu$ are a basis of
1191: $\BB_1\otimes\BB_3$, this expansion is unique, so we must have
1192: $[a_\mu,a'_\nu]=0$ for all $\mu,\nu$. Clearly, this property also
1193: transfers to the algebras generated by the $a_\mu$ and $a'_\nu$,
1194: i.e., to the support algebras noted in the Lemma. \qed
1195:
1196:
1197: \subsection{Proof of the Main Theorem}\label{sec:proof}
1198:
1199: We apply the construction of support algebras to the inclusion
1200: \begin{equation}\label{inclusion}
1201: \grule(\AA(\Box))\subset\bigotimes_q\AA(\Box+q)
1202: \end{equation}
1203: and define
1204: \begin{equation}\label{defBBq}
1205: \BB_q=\Spp(\grule(\AA(\Box)),\AA(\Box+q))\;.
1206: \end{equation}
1207:
1208:
1209: As a finite dimensional C*-algebra, each $\BB_q$ is isomorphic to
1210: $\BB_q=\bigoplus_\mu\MM_{n(q,\mu)}$ (See
1211: Proposition~\ref{Csform}). Now $\grule(\AA(\Box))$ is
1212: homomorphically embedded into $\bigotimes_q\BB_q$, with
1213: $\grule(\idty)=\idty$. Hence by Proposition~\ref{Cshom} we know
1214: that for any choice of summands $\mu_q$ we must have that
1215: $\dtwos$, the matrix dimension of $\AA(\Box)$, divides
1216: $\prod_qn(q,\mu_q)$. This gives a lower bound on the block sizes.
1217:
1218: In order to get an upper bound, consider the support algebras
1219: contained in some shifted cube, such as $\Box+\qone$. These are
1220: \begin{eqnarray}
1221: \Spp\Bigl(\grule(\AA(\Box+\qone-q))&,&\AA(\Box+\qone)\Bigr)
1222: \nonumber\\
1223: &=&\tau_{\qone-q}\Spp\Bigl(\grule(\AA(\Box)),\AA(\Box+q)\Bigr)
1224: \nonumber\\
1225: &=&\tau_{\qone-q}(\BB_q)\;
1226: \end{eqnarray}
1227: Since the $\AA(\Box+\qone-q)$ commute, so do their images under
1228: $\grule$ and, by Lemma~\ref{sppcomm}, so do the algebras
1229: $\tau_{\qone-q}(\BB_q)\subset\AA(\Box+\qone)$. However, we do not
1230: know a priori that these algebras are contained in
1231: $\AA(\Box+\qone)$ like a tensor product: When
1232: $z_{q,\mu_q}\in\BB_q$ is a central projection onto one of the
1233: matrix blocks of $\BB_q$, it is clear that the product
1234: $\prod_qz_{q,\mu_q}$ is a central element of the algebra generated
1235: by the $\BB_q$, but it might be zero. On the other hand, there
1236: must be {\it some} choice of blocks $\mu_q$, for which this is
1237: non-zero, and for this combination $\prod_q\MM_{n(q,\mu_q)}$,
1238: which is isomorphic to $\MM_n$ with $n=\prod_qn(q,\mu_q)$, is a
1239: direct summand of $\prod_q\BB_q\subset\AA(\Box+\qone)$. Hence we
1240: get the inequality
1241: \begin{equation}\label{dimineq}
1242: \dtwos\geq\prod_qn(q,\mu_q)\;.
1243: \end{equation}
1244:
1245:
1246: On the other hand, by the first step, the left hand side divides
1247: the right hand side of this inequality, so the we must have
1248: equality. This also implies that only one summand can be present
1249: in $\BB_q$, so we get $\BB_q\cong\MM_{n(q)}$ with
1250: $n(q)=n(q,\mu_q)$. That $\grule$ is an isomorphism from
1251: $\AA(\Box)$ onto $\bigotimes_q\BB_q$ follows by a direct dimension
1252: count, since a *-homomorphism between full matrix algebras of
1253: equal dimension can only be zero or an isomorphism. \qed
1254:
1255:
1256: \subsection{QCAs with abelian neighborhood}\label{sec:abelianDD}
1257: In a sense, Theorem~\ref{mainthm} gives a complete constructive
1258: procedure for QCAs in terms of the two unitary operators $U$, $V$
1259: with a constraint. Unfortunately, however, it does not seem to be
1260: easy to give a general solution of the constraint equations
1261: expressing the translation invariance (rather than the invariance
1262: by even translations). Therefore it is suggestive to repeat the
1263: analysis of support algebras also on the single cell level.
1264: Setting
1265: \begin{equation}\label{Dlocal}
1266: \DD_x=\Spp\bigl(\lrule(\AA_0), \AA_x\bigr)
1267: \end{equation}
1268: we have
1269: \begin{equation}\label{Dlocalsub}
1270: \lrule(\AA_0) \subset\bigotimes_{x\in\Nei}\DD_x
1271: \end{equation}
1272: It is clear that since $\lrule(\AA_0)$ is non-abelian, at least
1273: one of the algebras $\DD_x$ must also be non-abelian. The simplest
1274: case in this regard will be when all $\DD_x$ are abelian and
1275: commute with each other, except one, say $\DD_0$, which then has
1276: to isomorphic to the full cell algebra $\AA_0$ by
1277: Prop.\ref{Cshom}. Since all $\DD_x$ commute, we can jointly
1278: diagonalize them and this fixes a canonical basis for every cell.
1279: When $\ket\mu$ denotes the basis vectors, we can write the local
1280: transition rule as
1281: \begin{equation}\label{conditionalrule}
1282: \lrule(A)
1283: =\sum_{\mu_\Nei} U(\mu_\Nei)^*AU(\mu_\Nei)\otimes
1284: \bigotimes_{0\neq x\in\Nei}\ketbra{\mu_x}{\mu_x}\;,
1285: \end{equation}
1286: where the sum runs over all tuples $\mu_\Nei$ of basis labels
1287: $\mu_x$ for $x\in\Nei, x\neq0$ and, for each such tuple, $U(\mu_\Nei)$ is
1288: a unitary operator. Thus $\lrule$ describes a {\it conditional unitary
1289: operation} on cell $0$, where the conditioning is in some fixed
1290: ``computational basis''.
1291:
1292: We now need to analyze the constraints on these unitaries needed to make
1293: this homomorphism $\lrule$ a local transition rule. As a first step we
1294: look at the simplest case:
1295:
1296: \begin{lemma} Let $U_\mu,V_\nu\in\MM_d$ be unitary operators ($\mu,\nu=1,\ldots,d$)
1297: such that, for all $A,B\in\MM_d$,
1298: \begin{equation}\label{condscommute}
1299: \sum_{\mu\nu}\Bigl[U_\mu^*AU_\mu\otimes\ketbra\mu\mu,\;
1300: \ketbra\nu\nu\otimes V_\nu^*BV_\nu\Bigr]=0
1301: \end{equation}
1302: Then there are unitary $U,V\in\MM_d$ such that, for all $\mu$,
1303: $U^*U_\mu$ and $V^*V_\mu$ are diagonal.
1304: \end{lemma}
1305:
1306: \proof :\quad Let us abbreviate $A_\mu=U_\mu^*AU_\mu$,
1307: $B_\nu=V_\nu^*BV_\nu$, and take the matrix element of equation
1308: (\ref{condscommute}) in the product basis vectors
1309: $\bra{\alpha\beta}\cdots\ket{\gamma\delta}$. This gives
1310: \begin{equation}
1311: \bra\alpha A_\beta\ket\gamma\; \bra\beta B_\gamma\ket\delta
1312: =\bra\alpha A_\delta\ket\gamma\; \bra\beta B_\alpha\ket\delta
1313: \nonumber
1314: \end{equation}
1315: Now set $B=V_\alpha\ketbra{\beta'}{\delta'}V_\alpha^*$, with
1316: $\beta'\neq\beta$. Then $B_\alpha=\ketbra{\beta'}{\delta'}$, and
1317: $\bra\beta B_\alpha\ket\delta=0$, and the right hand side vanishes.
1318: Hence, for every $A$ and every $\delta'$
1319: \begin{equation}
1320: \bra\alpha A_\beta\ket\gamma\;
1321: \bra\beta V_\gamma^* V_\alpha\ket{\beta'}
1322: \bra{\delta'}V_\alpha^*V_\gamma \ket\delta
1323: =0.
1324: \nonumber
1325: \end{equation}
1326: Now the first factor can be made non-zero by an appropriate choice
1327: of $A$, and the third factor can be made non-zero by choosing
1328: $\delta'$, because the unitary operator $V_\alpha^*V_\gamma$
1329: cannot annihilate $\ket\delta$. It follows that
1330: $\bra\beta V_\gamma^* V_\alpha\ket{\beta'}=0$ vanishes for all
1331: indices, or $V_\gamma^* V_\alpha$ is diagonal for all
1332: $\alpha,\gamma$. Hence the Lemma follows with $V=V_1$, and the
1333: statement for $U$ follows by symmetry.\qed
1334:
1335: Let us apply this Lemma to the one-dimensional nearest neighbor
1336: case. Then a local rule of the form (\ref{conditionalrule}) can be
1337: written as
1338: \begin{equation}\label{conditionalrule1D}
1339: \lrule(A)
1340: =\sum_{\mu\nu} \ketbra\mu\mu\otimes U_{\mu\nu}^*AU_{\mu\nu}
1341: \otimes\ketbra{\nu}{\nu};.
1342: \end{equation}
1343: The remaining commutation condition is, for arbitrary one-site
1344: observables $A,B$,
1345: \begin{eqnarray}\label{commute1D}
1346: 0&=&[\lrule(A)\otimes\idty, \idty\otimes\lrule(B)]
1347: \nonumber\\
1348: &=&\sum_{\mu\nu\mu'\nu'} \ketbra\mu\mu\otimes
1349: \Bigr[U_{\mu\nu}^*AU_{\mu\nu}\otimes\ketbra{\nu}{\nu},
1350: \nonumber\\ &&\qquad
1351: \ketbra{\mu'}{\mu'}\otimes U_{\mu'\nu'}^*BU_{\mu'\nu'}
1352: \Bigr]\otimes\ketbra{\nu'}{\nu'}
1353: \end{eqnarray}
1354: Hence we can apply the Lemma to the commutator separately for
1355: every pair $\mu,\nu'$, and we find that up to a common unitary all
1356: $U_{\mu\nu}$ have to commute. Again, up to a cell-wise unitary
1357: rotation, we can choose the common eigenbasis of the $U_{\mu\nu}$
1358: as the same basis in which the conditions are written, i.e.,
1359: \begin{equation}\label{phase1D}
1360: U_{\mu\nu}=\sum_\kappa u(\mu\kappa\nu) \ketbra\kappa\kappa \;,
1361: \end{equation}
1362: with some phase function $u$ depending on three neighboring basis
1363: labels. However, this phase function $u$ is not arbitrary:
1364: unitaries of the form (\ref{phase1D}) in this way may still fail
1365: to satisfy (\ref{commute1D}). If we insert $A=\ketbra ab$ and
1366: $B=\ketbra{a'}{b'}$ we get the functional equation
1367: \begin{equation}\label{uuuu}
1368: \frac{u(\mu b a')}{u(\mu a a')}\cdot \frac{u(bb'\nu')}{u(ba'\nu')}
1369: =\frac{u(\mu b b')}{u(\mu a b')}\cdot \frac{u(ab'\nu')}{u(aa'\nu')}\;.
1370: \end{equation}
1371: Since we want to classify solutions up to a cell-wise rotation, we
1372: can take one of the unitaries $U_{\mu\nu}$ to be the identity, say
1373: $U_{11}=\idty$, or $u(1x1)=1$. Moreover, an overall phase of
1374: $U_{\mu\nu}$ is irrelevant, and we can choose this so
1375: $u(\mu1\nu)=1$. Then in (\ref{uuuu}) we take $a=b'=\nu'=1$, which
1376: gives
1377: \begin{equation}\label{uuuusolve}
1378: u(\mu ba')=u(\mu b1)u(ba'1)
1379: \end{equation}
1380: Thus $u$ is already determined by the two-variable function
1381: $(a,b)\mapsto u(a,b,1)$. It is easy to check that any choice of
1382: this function yields a solution of (\ref{uuuu}) via
1383: (\ref{uuuusolve}).
1384:
1385: We can summarize the result as follows:
1386:
1387: \begin{proposition}\label{prop:abelianDD}
1388: Let $\lrule$ be the local transition rule of
1389: a QCA such that $\DD_1$ and $\DD_{-1}$ are both abelian. Then,
1390: with respect to a basis in which these algebras are diagonal,
1391: there is a phase gate on $\Cx^d\otimes \Cx^d$:
1392: \begin{equation}\label{phasegate}
1393: U=\sum_{ab}u(a,b)\ketbra{ab}{ab} \;,
1394: \end{equation}
1395: normalized such that $u(1,b)=u(b,1)=1$, and a one-site unitary
1396: $V$ such that
1397: \begin{eqnarray}\label{phasealfa}
1398: \lrule(A)&=&X^*(\idty\otimes V^*AV\otimes\idty)X \qquad \mbox{with}\nonumber\\
1399: X&=& (U\otimes\idty_3)(\idty_1\otimes U)\;. \nonumber
1400: \end{eqnarray}
1401: \end{proposition}
1402:
1403:
1404: \subsection{Unilateral Automata}
1405:
1406: Another case of automata in one dimension, which can be characterized
1407: completely, are automata with neighborhood scheme $\Nei=\{0,1\}$ (or,
1408: symmetrically, $\Nei=\{-1,0\}$). The analysis is almost identical to that
1409: of Theorem~\ref{mainthm}, and gives full matrix algebras
1410: $\DD_i=\Spp(\lrule(\AA_0),\AA_i)=\MM_{n_i}$ and such that $n_0n_1=d$. As
1411: in the case of the Theorem, the local rules combines an arbitrary unitary
1412: $U:\Cx^d\to\Cx^{n_0}\otimes\Cx^{n_1}$ with a unitary
1413: $V:\Cx^{n_1}\otimes\Cx^{n_0}\to\Cx^d$. We only mention these to state
1414: the following classification of the simplest case:
1415:
1416: \subsection{Nearest neighbor qubit automata in one
1417: dimension}\label{sec:qubits}
1418:
1419: Consider the right and left support algebras
1420: $\DD_{-1},\DD_{0},\DD_{+1}$ as in (\ref{Dlocal}). These are
1421: subalgebras of the $2\times2$-matrices, which leaves three
1422: possibilities: each of these algebras can either be trivial
1423: ($\DD_i=\Cx\idty$), an abelian two-state algebra (isomorphic to
1424: the diagonal matrices, or the full algebra $\MM_2$.
1425:
1426: Suppose that at least one of the algebras $\DD_{\pm1}$, say
1427: $\DD_{-1}$, is trivial. Then we have a unilateral automaton. Since
1428: $n_0n_1=2$ we must have either $n_0=2, n_1=1$, a cell-wise unitary
1429: rotation or $n_0=1, n_1=2$ a right shift, possibly combined with a
1430: cell-wise rotation.
1431:
1432: Suppose that none of the algebras $\DD_{\pm1}$ is trivial. Then
1433: because $\DD_{-1}$ commutes with $\DD_{+1}$ neither algebra can be
1434: the full matrix algebra, since that would force the other to be
1435: trivial. It follows that both are abelian, and commute. Hence
1436: after a basis change (by another cell-wise rotation) we can take
1437: $\DD_{-1}=\DD_{+1}$ as the algebra of diagonal
1438: $2\times2$-matrices. This brings us into the situation of
1439: Section~\ref{sec:abelianDD}, and we find a phase rotation.
1440: Choosing the normalization in Proposition~\ref{prop:abelianDD}, we
1441: have only one free parameter left, i.e., we have an automaton
1442: built from commuting unitaries $U_x$ (cf.
1443: Section~\ref{sec:basicon}), which are {\it phase gates}
1444: \begin{equation}\label{phigate}
1445: U_x=\left(\begin{matrix}1&0&0&0\\0&1&0&0\\0&0&1&0\\
1446: 0&0&0&e^{i\phi}
1447: \end{matrix}\right)\;.
1448: \end{equation}
1449:
1450: The classification is hence complete. It is represented in
1451: Fig.\ref{qbit}.
1452: \begin{figure}[htb]
1453: \epsfxsize=6cm \epsffile{qbit.eps}
1454: \caption{\label{qbit}{\it All nearest neighbor qubit automata arise by combining
1455: cell-wise unitary rotations $(C)$ with right shifts $(R)$, left shifts $(L)$,
1456: or phase gates $(P)$.}}
1457: \end{figure}
1458:
1459: It is interesting to compare this with the classical case, which
1460: can be stated very similarly: then the only cell-wise operations
1461: are identity and global flip. Of course, the possibility of phase
1462: gates does not arise, leaving 6 classical possibilities.
1463:
1464: Beyond qubits, a good classification exists for Clifford automata
1465: in prime dimension \cite{OurCliff}. However, for general
1466: three-level systems the classification is likely to be
1467: complicated, since there is already a host of reversible classical
1468: nearest neighbor automata.
1469:
1470: \section{Other Approaches}\label{sec:others}
1471: In this section we take a look at the various proposals for
1472: defining QCAs, and show how they relate to the approach taken in
1473: this paper.
1474:
1475: \subsection{Feynman, and Transition Quasi-Probabilities}\label{sec:Feynman}
1476: The idea of a quantum cellular automaton is clearly present in
1477: Feynmans's famous 1981 lecture \cite{Feynman}. Although he
1478: suggests not only that a theory might and should be developed, and
1479: that it might even be taken seriously as fundamental physical
1480: theory, he does not actually develop a notion of QCAs in this
1481: article. The context in which he does write down a transition rule
1482: for a QCA, is where he discusses the possibility of simulating a
1483: QCA with a probabilistic cellular automaton. He emphasizes that
1484: this would involve something like negative transition
1485: probabilities and closes this section saying ``... I wanted to
1486: explain that if I try my best to make the equations look as near
1487: as possible to what would be imitable by a classical probabilistic
1488: computer, I get into trouble''.
1489:
1490: Feynman's idea for making a QCA look as classical as possible is
1491: to replace ``transition probabilities'' by Wigner function-like
1492: ``transition quasi-probabilities''. This approach is an
1493: interesting contribution to the definition of {\it irreversible}
1494: QCAs, which is still plagued with problems. However, as Feynman is
1495: clearly aware, it fails, and it is instructive to analyze this
1496: failure in the case of reversible automata.
1497:
1498: The transition function of a classical probabilistic cellular
1499: automaton (for a set $S$ of single-cell states) is a set of
1500: probabilities $M(s'|s_{\Nei})$ specifying the probability of
1501: finding a state $s'\in S$, when the configuration of the neighbors
1502: at the previous time step is $s_\Nei$. Using the ``simultaneous
1503: and independent update rule'' the probability for finding a
1504: configuration $s_\Lambda$ in a finite region $\Lambda\subset\Ir^s$
1505: becomes
1506: \begin{equation}\label{PCA}
1507: M(s'_\Lambda|s)=\prod_{x\in\Lambda}M(s'_x|s_{\Nei+x})\;.
1508: \end{equation}
1509: Note that this is readily read as a statement in the Heisenberg
1510: picture: the probability for $s'_\Lambda$ is expressed as the
1511: expectation of a random variable in the previous time step, namely
1512: the right hand side of (\ref{PCA}), considered as a function of
1513: the variables $s_x$. In the quantum case, $M(s'|s_\Nei)$ would
1514: then become an observable in $\AA_\Nei$, depending in a linear
1515: (and completely positive) way on an observable $s'\in\AA_0$. This
1516: is precisely a description of the local transition rule $\lrule$.
1517: Note that the commutation condition for local rules
1518: (Lemma~\ref{lem:localrule}) implies that the product is
1519: well-defined, independently of the ordering of the factors. In a
1520: more general quantum context, such as irreversible QCA evolutions
1521: for which the global rule will not respect the product, we cannot
1522: be sure of this property: additional information about the
1523: ordering of factors in (\ref{PCA}) would have to be supplied, but
1524: even an ordering fixed by some convention would not prevent the
1525: evolution from sometimes taking hermitian elements to
1526: non-hermitian elements.
1527:
1528: This latter problem: the ordering of factors and the hermiticity
1529: is neatly solved by Feynman's approach of quasi-probabilities. He
1530: expands all qubit operators in a special basis of four hermitian
1531: operators $F(\xi)\in\MM_2$, $\xi\in\{{++},{+-},{-+},{--}\}$:
1532: \begin{equation}\label{Feyman_Wigner}
1533: F_{uv}=\left(\begin{matrix}
1534: (1+u)/2 &v(1-iu)/4\\
1535: v(1+iu)/4& (1-u)/2
1536: \end{matrix}\right) \;, (u,v=\pm1)\;.
1537: \end{equation}
1538: The expectations $f_\rho(\xi)=\tr(\rho F(\xi))$ of these operators
1539: are the analogs of the Wigner function (see \cite{Wootters,Paz}
1540: for later elaborations on Wigner functions in finite dimension).
1541: Of course, by taking tensor products of these operators we get
1542: Wigner functions for multi-qubit systems. Now we can expand the
1543: transition rule for a single cell in Wigner operators:
1544: \begin{equation}\label{alfa0Wig}
1545: \lrule(F(\eta_0))
1546: =\sum_{\xi_\Nei}M(\eta_0|\xi_\Nei)
1547: \bigotimes_{x\in\Nei} F(\xi_x) \;,
1548: \end{equation}
1549: with the transition quasi-probabilities $M(\eta_0|\xi_\Nei)$.
1550: These are usually not positive, but this is a minor inconvenience
1551: as long as the physical transition operator $\grule$ is positive.
1552: Equation~(\ref{alfa0Wig}) is just an expansion of the local rule
1553: in a basis of Wigner operators, so a local rule is completely
1554: equivalent to a set of transition quasi-probabilities. The
1555: difference between the quasi-probability approach and ours is how
1556: the global rule is constructed. Let us consider, for simplicity,
1557: the image of a two-site observable $F(\eta_1)\otimes F(\eta_2)$
1558: under a one-dimensional nearest neighbor automaton. According to
1559: (\ref{PCA}), transition probabilities must be combined as
1560: \begin{equation}\label{alfa1Wig}
1561: \begin{array}{l}\displaystyle
1562: \grule_{\rm quasi}(F(\eta_1)\otimes F(\eta_2))\\ \displaystyle
1563: {}\quad=\sum_{\xi_0,\xi_1,\xi_2,\xi_3}M(\eta_1|\xi_0,\xi_1,\xi_2)
1564: M(\eta_2|\xi_1,\xi_2,\xi_3)
1565: \rule[-25pt]{0pt}{45pt}\\
1566: \displaystyle{}\ \qquad\qquad\times
1567: F(\xi_0)\otimes F(\xi_1)\otimes F(\xi_2)\otimes F(\xi_3)\;.
1568: \end{array}
1569: \end{equation}
1570: We can extend this to arbitrarily large configurations to get the
1571: time evolution of an automaton in the Heisenberg picture. Since
1572: the transition quasi-probabilities are real, this evolution will
1573: automatically preserve hermiticity, and since $\sum_\xi
1574: F(\xi)=\idty$ all normalization and locality properties will
1575: automatically come out correctly. Of course, there will be no
1576: operator ordering problem, since we multiply at the level of
1577: functions, and all this will work for probabilistic irreversible
1578: and reversible transition rules alike.
1579:
1580: On the other hand, under the homomorphism $\grule$ the tensor
1581: product $F(\eta_1)\otimes F(\eta_2)$ is evolved to
1582: \begin{equation}\label{alfa2Wig}
1583: \begin{array}{l}\displaystyle
1584: \grule(F(\eta_1)\otimes F(\eta_2))\\ \displaystyle
1585: {}\ = \sum_{{\xi_0,\xi_1,\xi_2,\xi_3}\atop{\xi'_0,\xi'_1,\xi'_2,\xi'_3}}
1586: M(\eta_1|\xi_0,\xi_1,\xi_2)
1587: M(\eta_2|\xi'_1,\xi'_2,\xi'_3)
1588: \rule[-25pt]{0pt}{45pt}\\\displaystyle
1589: {}\quad \times
1590: F(\xi_0)\otimes F(\xi_1)F(\xi'_1)
1591: \otimes F(\xi_2)F(\xi'_2)\otimes F(\xi'_3)
1592: \end{array}
1593: \end{equation}
1594: Then (\ref{alfa1Wig})=(\ref{alfa2Wig}), iff
1595: $F(\xi)F(\eta)=\delta_{\xi\eta}F(\xi)$, which is obviously true
1596: for the minimal projections in a classical observable algebra, but
1597: obviously false for the Wigner operators. So the automata studied
1598: in this paper are not (or not in general) representable as
1599: quasi-probabilistic CAs.
1600:
1601: Feynman uses the approach based on (\ref{alfa1Wig}) only as a
1602: caricature of a quantum process. He points out that negative
1603: transition quasi-probabilities exclude a simulation of the global
1604: time step (\ref{PCA}) by successive independent random trials. For
1605: him this indicates the increased complexity of quantum
1606: computation.
1607:
1608: Whether or not this approach works as a definition of (possibly
1609: also irreversible) QCAs hinges on the question of positivity.
1610: Non-positive transition probabilities would not be serious if they
1611: lead to the construction of a legitimate (i.e., completely
1612: positive) global transition rule. Unfortunately, however, there is
1613: no indication that quantum positivity improves in the passage from
1614: local to global rule. This can be seen already for a simple phase
1615: gate array, i.e., (P) in Fig.~\ref{qbit}. For generic phase angle
1616: $\varphi$ neither the local transition nor the global transition
1617: e.g., the expression (\ref{alfa2Wig}), have positive transition
1618: quasi-probabilities, although, of course, they correspond to
1619: completely positive operations. On the other hand, the two-site
1620: transition rule (\ref{alfa1Wig}) takes some positive operators
1621: into non-positive ones.
1622:
1623: An interesting special case occurs for phase angle $\phi=\pi$. In
1624: that case, the local rule does give rise to positive, and even
1625: {\it deterministic} transition quasi-probabilities. They belong
1626: to a classical deterministic CA, which like the phase gate QCA is
1627: its own inverse. But if it is applied as in (\ref{alfa1Wig}), it
1628: violates positivity.
1629:
1630:
1631:
1632: \subsection{Watrous et al.}
1633: One of the first serious attempts at the definition of QCAs was by
1634: Watrous \cite{Watrous}. It is based on another ``quantization'' of
1635: the transition probability formula (\ref{PCA}), inspired by
1636: Feynman's notion that in quantum theory one must replace
1637: probabilities by amplitudes. Thus one tries a product formula like
1638: (\ref{PCA}) for the transition amplitudes, presumably defining in
1639: this way the unitary transition operator for the whole process:
1640: \begin{equation}\label{watrousCA}
1641: U(a|b)=\prod_x u(a_x|b_{\Nei+x})\;,
1642: \end{equation}
1643: where $a,b$ are classical configurations, labelling the basis
1644: states of the QCA, and $b_{\Nei+x}$ is the configuration $b$
1645: restricted to the neighborhood of $x$. The function $u$ plays the
1646: role of the local transition rule. Basically the same definition
1647: is also used in van Dam \cite{vanDam}, where it is phrased as an
1648: assignment of a a product vector to every basis state in the
1649: computational basis. Further work in this approach is to be found
1650: in \cite{santa}, and in the textbook \cite{Gruska}, or a recent
1651: introduction \cite{Aoun}. For the sake of discussion let us call
1652: an automaton defined by (\ref{watrousCA}) a WQCA. There are
1653: several problems with this formula:
1654:
1655: \begin{enumerate}
1656: \item
1657: The infinite product may not be defined. This may be resolved by
1658: either introducing a ``quiescent state'' which is invariant under
1659: the evolution, and considering only superpositions of
1660: configurations which are quiescent outside a finite region
1661: \cite{Watrous,santa}, or to look at periodic boundary conditions
1662: only \cite{vanDam}. Either approach works, but none of these
1663: workarounds is necessary in our definition.
1664:
1665: \item $U$ from (\ref{watrousCA}) has no reason to be unitary, and
1666: most of the time it isn't. In other words, there is no
1667: straightforward way of characterizing those local transition
1668: amplitudes $u$ for which the formula does indeed define an
1669: isometric (or stronger: a unitary) operator, in which case the
1670: rule is called ``well-formed'' (or unitary). Whereas positivity
1671: and normalization of the local rule $M$ on the right hand side of
1672: (\ref{PCA}) guarantee the corresponding properties for the global
1673: evolution, no equally simple criterion exists for well-formedness
1674: (but see \cite{santa} for algorithms in the one-dimensional case).
1675:
1676: \item Many unitary operators are not of the form
1677: (\ref{watrousCA}). To begin with, the definition depends on the
1678: choice of a preferred basis. In general, the product of two WQCA
1679: unitaries need not be a WQCA (with larger neighborhood), and the
1680: inverse of a WQCA may fail to be a WQCA. This
1681: makes it hard to build a general theory on this definition.
1682: \newline
1683: To get an example of these phenomena note that a necessary
1684: condition for a unitary of Watrous form is that the computational
1685: basis states are mapped to product states. Moreover, if we follow
1686: the unitary operator by a site-wise unitary rotation, or precede
1687: it by a product of phase gates (i.e., introducing additional phase
1688: factors independent of the output labels $a_x$), we stay in this
1689: class. But now consider a sitewise Hadamard rotation, followed by
1690: a product of phase gates. It is easy to verify that such a map
1691: does not take the computational basis states to product states (or
1692: Briegel's one-way computers would not work.) So this is not a
1693: WQCA, although it is the product of two WQCAs, and its inverse is
1694: also a WQCA .
1695:
1696: \item Even in the cases where the formula does work, such as,
1697: e.g., for the multiplication by local phases, the size of the
1698: neighborhood in (\ref{watrousCA}) is not the neighborhood scheme
1699: describing the propagation of observable effects (the $\Nei$ of
1700: our definition), but rather the $\widetilde\Nei$ from
1701: Section~\ref{sec:communitary}. In fact, it is not even clear
1702: whether a WQCA is necessarily local in our sense: We did not
1703: manage to exclude the possibility that a local measurement after
1704: one time step might allow inferences about arbitrarily distant
1705: modifications of the state.
1706:
1707: \item in the original paper \cite{Watrous}, Watrous also looks at
1708: partitioned automata. Since we prove that all QCAs can be obtained
1709: by a partitioning scheme, it might seem that WQCA$\supset$QCA, in
1710: seeming contradiction to the Example in 3 above. However, Watrous
1711: uses only a special case of partitioning, namely a unitary
1712: followed by a permutation of subcells. Only if we extend the class
1713: of WQCAs and include all their products, we get all QCAs.
1714:
1715:
1716: \end{enumerate}
1717: To summarize: Judged by the criteria in the introduction, the
1718: notion of Watrous is not a satisfactory formalization of the idea
1719: of quantum cellular automata. What is lacking is the easy passage
1720: from local to global rules, and from global axiomatic to local
1721: constructive description. For all problems involving the iteration
1722: of automata, the fact that the class is not closed under
1723: composition and inverse, puts a premature end to studies based on
1724: WQCAs.
1725:
1726:
1727: \subsection{Richter and Werner}
1728:
1729: As mentioned earlier, the observable-based approach underlying our
1730: definition was first used in \cite{RiWe} by one of us, but with a
1731: focus on the irreversible case. In that paper partitioning
1732: (dissipative cell evolution combined with permutation of subcells)
1733: was used to allow a free construction satisfying a global locality
1734: condition.
1735:
1736: It is not so clear what should replace the axiomatic definition in
1737: the irreversible case. A possible condition is to just replace the
1738: local rule $\lrule$ by a completely positive map, and insist on
1739: commutativity as before. This does define a global evolution step
1740: \cite{RiWe,Tak}. In view of the reversible case this would seem
1741: like a good definition. However, it is again not clear whether the
1742: composition of two such transformations will again be of the same
1743: kind. Nor is it clear how to describe the class of transformations
1744: obtained by several such ``simultaneous independent update''
1745: steps.
1746:
1747: It turns out that the analysis of Theorem~\ref{mainthm} can partly
1748: be repeated in the irreversible case, thereby reducing the
1749: possibilities somewhat. This line will be pursued elsewhere.
1750:
1751:
1752: \subsection{Wolfram}
1753: In a recent thick book \cite{Wolfram} S. Wolfram has argued that
1754: the that the universe might be a big cellular automaton following
1755: simple rules (see also \cite{Zuse}). Wolfram uses only classical
1756: structures, expressing the belief, however, that quantum
1757: structures might emerge from classical rules generating sufficient
1758: complexity. Since Einstein failed with a program like that (he
1759: thought of overdetermined non-linear field equations, rather than
1760: CAs) it would be nice to see the details worked out.
1761:
1762: \subsection{Quantum random walks}
1763: The term ``Quantum cellular automaton'' has sometimes been used
1764: \cite{Zeilinger,Iwo,Meyer,konno} for a unitary evolution of a
1765: particle on a discretized space. The total Hilbert space of such a
1766: system is $\ell^2(\Ir^s)$, the space of square summable complex
1767: functions on the lattice, i.e., the direct sum rather than the
1768: tensor product of the one-cell spaces. The classical analogue of
1769: such a system is a single classical particle moving on a lattice,
1770: e.g., in a random walk. Therefore much better terminology to call
1771: these quantum systems {\it quantum random walks}, rather than
1772: cellular automata. Such systems have been proposed for purposes of
1773: quantum computation, in particular for search problems on graphs
1774: (see \cite{Kempe} for a review).
1775:
1776: Quantum random walks require localization properties not unlike
1777: those of QCAs. To see the connection it is interesting to consider
1778: the connection between classical random walks and CAs: A random
1779: walk can be seen as a special CA, with each cell either empty or
1780: occupied, started in a configuration with exactly one occupied
1781: cell. Of course, the overall CA dynamics should respect this
1782: constraint. Because the CA rule is local we can then also put
1783: arbitrarily many particles on the lattice, and the dynamics is
1784: well-defined by the random walk, as long as the particles do not
1785: collide. What happens on collision is a piece of information which
1786: must be supplied in order to make a random walk into a CA, i.e.,
1787: in order to pass from a random walk to a (possibly
1788: ``interacting'') {\it diffusion}.
1789:
1790: Consider a QCA with a special ``empty'' state specified for the
1791: single cell algebra. This means that we can define a global
1792: quantity {\it particle number}, which ought to be conserved by the
1793: QCA. Technically this is the infinite sum of 0's and 1's, and not
1794: a well defined observable. However, we can consider this formal
1795: sum as a lattice interaction generating the time evolution of
1796: ``gauge transformations'', acting as an infinite product of
1797: unitaries, as in Section~\ref{sec:communitary}. For a QCA
1798: commuting with such transformations, the ``one-particle'' Hilbert
1799: space can be defined, and the QCA dynamics restricted to this
1800: subspace is a unitary evolution of the random walk type. Note that
1801: we have not assumed that the dimension $d$ of the one-site algebra
1802: is $2$, i.e., occupied cells (``particles'') may have an internal
1803: structure. This turns out to be necessary: a standard classical
1804: random walk, with only the empty/occupied distinction and nearest
1805: neighbor interaction cannot be reversible. Similarly, a unitary on
1806: $\ell^2(\Ir^s)$ cannot be strictly local \cite{Zeilinger}, and we
1807: do need internal states \cite{Meyer2}. If there is only one
1808: particle, this is the same as saying that we have only the
1809: empty/occupied distinction for the cells, but we have also a
1810: ``quantum coin'' which helps determining the steps.
1811:
1812: The standard model of a quantum random walk \cite{Kempe} uses a
1813: qubit coin, i.e., three states for the QCA. All random walks with
1814: such a coin are parameterized by a single unitary
1815: $2\times2$-matrix $U$: we can describe the internal states
1816: (``chirality'' \cite{konno}) as ``go right'' and ``go left''. A
1817: trivial, but globally well-defined evolution step is defined by
1818: following these instructions. The general case arises by following
1819: this with a sitewise unitary rotation by $U$, representing the
1820: quantum coin flip.
1821:
1822: Now the question arises: can we consider such random walks as the
1823: one-particle component of a QCA allowing arbitrarily many
1824: particles? Indeed this is possible, even in many ways, and it
1825: would be very interesting to classify all possibilities, hence all
1826: ``interactions''. One possibility is to second quantize the random
1827: walk, which leads to a Boson system allowing, at each site, an
1828: arbitrary number of particles. It is clear from the formalism of
1829: second quantization that all the locality properties required of a
1830: QCA are then satisfied. Clearly, this is the non-interacting
1831: option. If we insist on finitely many states per cell we introduce
1832: some kind of hard core interaction. A trivial way of doing it with
1833: four states is this: Consider two infinite qubit chains, called
1834: the left moving and the right moving chain. The ``free'' time
1835: evolution $S$ is indeed shifting the two chains separately
1836: according to this description. Now at each lattice site we take
1837: the right moving and the left moving one-site algebras together to
1838: define a single cell of the QCA, with the four states labelled
1839: `empty', `R', `L', and `RL'. We then select a unitary $U$ leaving
1840: the empty and the doubly occupied state invariant, but shuffling
1841: the `R' and `L' states as before. Clearly, the combination of the
1842: free evolution $S$ with the sitewise application of $U$ defines a
1843: QCA whose one-particle sector is the given random walk.
1844:
1845:
1846: \subsection{Spacetime localized algebras}
1847:
1848: A quantum field theory can be considered from two equivalent
1849: points of view. On the one hand one can consider the fields at
1850: some time fixed time as the basic dynamical variables, e.g., the
1851: Cauchy data at time zero. On the other hand, for the discussion of
1852: relativistic causality, it is convenient to consider as
1853: fundamental the family of algebras $\AO{}$ associated with the
1854: measurements in some space-time region $\mathcal O$ \cite{Haag}.
1855: Similarly, one can look at a QCA from these two points of view,
1856: the ``Cauchy data'' point of view being what we described so far.
1857: For the space-time view consider the extended lattice $\Ir^{s+1}$
1858: of space-time points $(x,t)$, with $x\in\Ir^s$, the same spatial
1859: lattice as before, and $t\in\Ir$ a discrete time point.
1860:
1861: The global C*-algebra will be the same as before, but we introduce
1862: additional subalgebras, namely
1863: \begin{equation}\label{Axt}
1864: \AA_{x,t}=\grule^{-t}(\AA_x)\;.
1865: \end{equation}
1866: Since $\grule$ commutes with lattice translations we can define a
1867: joint set of space-time translations $\grule_{x,t}$, combining a
1868: lattice translation by $x$ with $\grule^{-t}$. For any subset
1869: $\OO\subset\Ir^{s+1}$ we define $\AO{}$ as the C*-algebra
1870: generated by all the $\AA_{x,t}$ with $(x,t)\in\OO$. In order to
1871: study the localization properties of of these algebras, let us say
1872: that a sequence of lattice points
1873: \begin{equation}\label{slpath}
1874: (x_0,t_0),\ (x_1,t_0+1),\ (x_2,t_0+2),\ldots,(x_N,t_N)
1875: \end{equation}
1876: is (forward) {\it timelike}, if $x_{k+1}-x_k\in\Nei$ for all $k$.
1877: Then we define two regions $\OO_1,\OO_2\subset\Ir^{s+1}$ to be
1878: {\it spacelike separated} (notation: $\OO_1\spacelike\OO_2$), if
1879: no timelike curve passes through some point of $\OO_1$ and also
1880: through some point of $\OO_2$. Moreover, we say that $\OO_1$ is
1881: {\it causally dependent} on $\OO_2$ (notation:
1882: $\OO_1\vartriangleleft\OO_2$), if every timelike curve which
1883: passes through (any point of) $\OO_1$ also passes through $\OO_2$.
1884:
1885: Then it is clear that $\OO_1\spacelike\OO_2$ implies that $\AO1$
1886: and $\AO2$ commute elementwise. Moreover, if $\OO_2$ is a finite
1887: set of lattice points $(x,t)$, all with the same $t$, and if
1888: $\OO_1\vartriangleleft\OO_2$, then $\AO1\subset\AO2$.
1889:
1890: Conversely, if we have a net of local algebras $\AO{}$, defined
1891: for arbitrary finite subsets $\OO$ of the spacetime-lattice, if
1892: the spacetime translations act by automorphism of the total
1893: algebra such that $T_{(x,t)}(\AO{})={\mathfrak A}({\OO}+(x,t))$,
1894: and that the above locality properties hold, then the time zero
1895: algebras form a QCA in the sense of Section~\ref{sec:def}.
1896:
1897: This way of viewing QCAs may indeed be useful for making the
1898: connection to relativistic field theories, or for a detailed study
1899: of the growth of localization regions.
1900:
1901:
1902:
1903: \subsection{Non-commuting cells}
1904: As is standard in quantum theory we described the cells of the
1905: automaton as subsystems in the usual tensor product sense. In a
1906: slightly more axiomatic style we could have postulated, that
1907: arbitrary measurements in separate cells can be carried out
1908: jointly, implying the commutation between the different cells.
1909: While this is certainly very natural, more general commutation
1910: rules between cells may be considered. An obvious example are
1911: anti-commutation rules, i.e., we could think of a Fermi gas on a
1912: lattice. This would, of course, change the requirements on local
1913: rules, but it is quite clear how to adapt our definition to that
1914: case.
1915:
1916: Another very interesting deviation from commuting cells is studied
1917: in \cite{Seiler}, in connection with non-commutative analogs of
1918: 2-surfaces with constant negative curvature, the sine-Gordon
1919: equation, and 2-dimensional lattice systems in magnetic field
1920: (Hofstadter butterfly). Roughly speaking the structure
1921: investigated has has a single variable, say a unitary $Q_x$, at
1922: each cell ``$x$''. Non-commutativity comes in, because any pair of
1923: neighboring unitaries forms a discrete Weyl system. At distances
1924: $\geq2$ the variables commute. It is clear that (reversible)
1925: dynamical rules can be described exactly along the lines of our
1926: definition, as automorphisms respecting this algebraic structure,
1927: and also the localization (up to a finite enlargement of
1928: localization regions). Of course, this is not the place to explore
1929: such structures, and we point again to the book \cite{Seiler} for
1930: more material.
1931:
1932: \appendix
1933: \section{Finite Dimensional C*-algebras}\label{APPCstar}
1934:
1935: Almost all the hard work in any textbook on C*-algebras goes into
1936: aspects of the theory, which become entirely trivial for finite
1937: dimensional C*-algebras. Of course, some algebras in this paper
1938: are infinite dimensional, notably the quasi-local C*-algebra
1939: describing the infinite system. But the key arguments use only the
1940: finite dimensional structure. Therefore we will give in this
1941: Appendix a quick summary of C*-algebra theory as it applies to the
1942: finite dimensional case.
1943:
1944: In order to make the paper more accessible to communities in which
1945: algebraic terminology is less current (e.g., most theoretical
1946: physicists, and classical computer scientists) we start out with
1947: an extended glossary, in which the basic notions are defined and
1948: some basic facts are noted. This will be followed by some
1949: structure theorems which we need in the body of the paper. Of
1950: course, all this cannot replace a serious textbook. We recommend
1951: \cite{BraRo,Tak,Dix}.
1952:
1953:
1954: \subsection{C*-Glossary}
1955:
1956: The operations making up the abstract structure of C*-algebras are
1957: inspired by those known from algebras of operators on a Hilbert
1958: space. In fact, every algebra of operators on a finite dimensional
1959: Hilbert space, which is also closed under taking adjoints,
1960: satisfies the definition, and every abstract C*-algebra is
1961: isomorphic to an operator algebra. The following list of relevant
1962: operations and concepts may serve as a glossary. The algebra under
1963: consideration is usually denoted by $\AA$.
1964:
1965: \begin{itemize}
1966: \item {\it Addition} and multiplication by {\it complex scalars}. This makes
1967: $\AA$ a vector space over $\Cx$, which we assume to be finite
1968: dimensional from now on.
1969:
1970: \item {\it Multiplication}. We denote the product by $AB\in\AA$,
1971: when $A,B\in\AA$. The product is distributive and associative (but
1972: not necessarily commutative). If the algebra is commutative or
1973: ``abelian'' the system under consideration is classical.
1974:
1975: \item The {\it adjoint} or ``star operation'' denoted by $A^*\in\AA$,
1976: when $A\in\AA$. This is conjugate linear (or ``antilinear''),
1977: which means that $(\lambda A)^*=\overline{\lambda}A^*$, and
1978: $(A+B)^*=A^*+B^*$. The adjoint satisfies $(AB)^*=B^*A^*$.
1979: Physicists often write $A^\dagger$ for the adjoint.
1980:
1981: \item The {\it norm}, which is a positive number $\norm{A}$
1982: associated with each $A\in\AA$. With respect to the algebraic
1983: structures, the norm satisfies $\norm{A+B}\leq\norm A+\norm B$,
1984: $\norm{\lambda A}=|\lambda|\;\norm A$, $\norm{AB}\leq\norm A\norm
1985: B$ and $\norm{A^*A}=\norm A^2$. $\norm A=0$ implies $A=0$. In
1986: contrast to the general case, the norm is uniquely determined by
1987: the algebraic structures.
1988:
1989: \item An {\it identity}. In contrast to the general case, in finite
1990: dimensional C*-algebras there is always a unique element $\idty$
1991: satisfying $\idty A=A\idty=A$.
1992: \item A {\it homomorphism} between C*-algebras is a map $\Phi:\AA\to\BB$
1993: between C*-algebras, preserving the algebraic structures. That is,
1994: $\Phi$ is linear, $\Phi(AB)=\Phi(A)\Phi(B)$, and
1995: $\Phi(A^*)=\Phi(A)^*$. The latter property is sometimes emphasized
1996: by speaking of *-homomorphims. Homomorphisms $\Phi:\AA\to\AA$ are
1997: called {\it endomorphisms} of $\AA$, and if an inverse
1998: homomorphism exists, $\Phi$ is called an {\it isomorphism} or an
1999: {\it automorphism} (if $\AA=\BB$). For general homomorphisms,
2000: $\Phi(\idty)$ is a projection in $\BB$. If $\Phi(\idty)=\idty$, we
2001: call it {\it unital}.
2002: \item An {\it ordering}. We write $A\geq0$, if there is a
2003: $B\in\AA$ such that $A=B^*B$. The set of positive elements is a
2004: convex cone in the set of hermitian ($A^*=A$) elements. In an
2005: operator algebra, the positive elements are precisely those
2006: hermitian ones with all eigenvalues non-negative.
2007: \item The {\it center} of $\AA$ is the subalgebra $\ZZ(\AA)\subset\AA$ of
2008: elements $Z$ such that $ZA=AZ$ for all $A\in\AA$.
2009: \item A {\it state} on $\AA$ is a linear functional
2010: $\omega:\AA\to\Cx$ which is positive (i.e., $A\geq0$ implies
2011: $\omega(A)\geq0$) and normalized (i.e., $\omega(\idty)=1$).
2012: \item A {\it trace} on $\AA$ is a positive linear functional $\tau$ such
2013: that $\tau(AB)=\tau(BA)$.
2014:
2015: \item a positive linear functional $\omega$ is called {\it
2016: faithful} if $A\geq0$ and $\omega(A)=0$ imply $A=0$. Every finite
2017: dimensional C*-algebra has a faithful trace. On an operator
2018: algebra, the usual trace of operators (which we will denote by
2019: $\tr$) is an example.
2020: \item Given a faithful trace $\tau$, every positive linear
2021: functional can be written as $\omega(A)=\tau(\rho A)$, for a
2022: unique $\rho\in\AA$, $\rho\geq0$, which is called the {\it
2023: density operator} of $\omega$ with respect to $\tau$. $\omega$ is
2024: also a trace iff $\rho\in\ZZ(\AA)$.
2025:
2026: \item An element $P\in\AA$ is a called a {\it projection} if $P^*=P=P^2$.
2027: It is called a {\it minimal projection} if for any projection $Q$,
2028: $Q\leq P$ implies $Q=0$ or $Q=P$.
2029:
2030: \item The {\it direct sum} $\AA=\bigoplus_\mu\AA_\mu$ of a finite collection of
2031: finite dimensional C*-algebras $\AA_\mu$ is the vector space
2032: direct sum, i.e., elements are tuples of components
2033: $A_\mu\in\AA_\mu$, with componentwise algebraic operations. In
2034: operator algebras each term in the sum corresponds to a diagonal
2035: block in a block matrix decomposition.
2036: \item The {\it tensor product} $\AA=\bigotimes_\mu\AA_\mu$ is the
2037: vector space tensor product, with product and adjoint defined as
2038: the unique linear (resp. conjugate linear) extensions of
2039: $(\otimes_\mu A_\mu)(\otimes_\mu B_\mu)=\otimes_\mu (A_\mu B_\mu)$
2040: and $(\otimes_\mu A_\mu)^*=\otimes_\mu A_\mu^*$. In operator
2041: algebras one forms this product by taking first the tensor
2042: products of the underlying Hilbert spaces, and taking the algebra
2043: generated by all tensor product operators.
2044: \end{itemize}
2045:
2046: \noindent The first four items on this list are the definition of
2047: C*-algebras. Note that order and unit are explicitly defined
2048: in terms of the algebraic structure (linear operations,
2049: multiplication and adjoint), and the norm is also defined
2050: explicitly as
2051: \begin{equation}\label{defnorm}
2052: \norm A=\inf\{\lambda>0| \exists_B\;A^*A+B^*B=\lambda^2\idty\}
2053: \end{equation}
2054: It might thus seem superfluous to list the norm among the defining
2055: elements. In the infinite dimensional case it is needed, of
2056: course, to formulate the topological completeness requirement (and
2057: completeness is needed in turn to construct $B$ in
2058: (\ref{defnorm})). However, even in the finite dimensional case the
2059: implication $(\norm A=0)\Rightarrow (A=0)$ carries non-trivial
2060: information by excluding the existence of nonzero elements $A_i$
2061: such that $\sum_iA_i^*A_i=0$.
2062:
2063: \subsection{C*-structure}
2064:
2065: Consider a single Hermitian element $A\in\AA$. Then since $\AA$ is
2066: finite dimensional, the powers $A^n$ must be linearly dependent,
2067: i.e., there is a characteristic polynomial $p(A)=\sum_kc_kA^k=0$.
2068: From this one readily constructs polynomials $p_\ell$ such that
2069: $p_\ell(A)$ is a projection, and
2070: \begin{equation}\label{babyspectral}
2071: A=\sum_\ell a_\ell p_\ell(A)
2072: \end{equation}
2073: where $a_\ell$ are the distinct roots of $p(a)=0$. This is called
2074: the {\it spectral theorem}. It implies, in particular, that any
2075: finite dimensional C*-algebra has many projections (which may fail
2076: in infinite dimension).
2077:
2078: This fact will be used in the following fundamental structure
2079: theorem. Recall that by $\MM_n$ we denote the algebra of complex
2080: $n\times n$ matrices.
2081:
2082: \begin{proposition}\label{Csform}
2083: Every finite dimensional C*-algebra $\AA$ is
2084: characterized uniquely up to isomorphism by a finite sequence
2085: $n_1\geq n_2\geq\cdots\geq1$ of numbers such that
2086: \begin{equation}\label{isoCstar}
2087: \AA \cong\bigoplus_\mu \MM_{n_\mu}\;.
2088: \end{equation}
2089: \end{proposition}
2090:
2091: The basic idea of the proof is to consider the center of $\AA$,
2092: which is a finite dimensional abelian C*-algebra. The minimal
2093: projections $z_\mu$ of the center decompose the algebra into a
2094: direct sum $\AA=\bigoplus_\mu z_\mu\AA$, in which each of the
2095: summands has trivial center. The building blocks $z_\mu\AA$ are
2096: then seen to be isomorphic to full matrix algebras $\MM_{n_\mu}$,
2097: where $n_\mu$ is the maximal number of mutually orthogonal
2098: projections in $z_\mu\AA$.
2099:
2100:
2101: \begin{proposition}\label{Cshom}
2102: If $\Phi:\MM_d\to \bigoplus_\mu\MM_{n(\mu)}$ is a *-homomorphism
2103: such that $\Phi(\idty)=\idty$, each $n(\mu)$ has to be divisible
2104: by $d$.
2105: \end{proposition}
2106:
2107: By considering the composition of the given homomorphism with the
2108: projection onto one summand, which is also a homomorphism, we can
2109: consider the case of a single summand. Thus $\Phi$ becomes a
2110: representation of $\MM_d$ on a Hilbert space of dimension
2111: $n(\mu)$, which can be decomposed into irreducible
2112: representations. It is a basic property of $\MM_d$, however, that
2113: all its irreducible representations are unitarily equivalent to
2114: the defining representation on $\Cx^d$, so $n(\mu)$ must be $d$
2115: times the multiplicity (number of isomorphic irreducible summands)
2116: of this representation.
2117:
2118:
2119: \begin{proposition}\label{Cscom}
2120: If $\AA\cong\bigoplus_\mu \MM_{n(\mu)}$ and $\BB\cong\bigoplus_\nu
2121: \MM_{m(\nu)}$ are commuting subalgebras of $\BB(\HH)$, the algebra
2122: $\AA\BB$ is also decomposed into direct summands, each of which
2123: arises by multiplying a summand form each of the algebras. These
2124: occur with integer multiplicities $r_{\mu\nu}\geq0$ such that
2125: \begin{equation}\label{multip}
2126: \sum_{\mu\nu}r_{\mu\nu}n(\mu)m(\nu)
2127: =\dim\HH\;.
2128: \end{equation}
2129: \end{proposition}
2130:
2131: Let $A_\mu\in\MM_{n(\mu)}\subset\AA$ and
2132: $B_\nu\in\MM_{m(\nu)}\subset\BB$ be elements from the respective
2133: blocks. Then $A_\mu\otimes B_\nu\mapsto A_\mu B_\nu$ is a
2134: representation of $\MM_{n(\mu)\cdot
2135: m(\nu)}\cong\MM_{n(\mu)}\otimes\MM_{m(\nu)}$ on $\HH$, which may
2136: however be zero since we cannot guarantee that it preserves the
2137: identity. The rest of the argument is as for the previous
2138: proposition.
2139:
2140:
2141:
2142:
2143: \section{Acknowledgements}
2144: The main lines of this paper were conceived in March 2003, in
2145: discussions at the Centro Ettore Majorana in Erice , whose
2146: hospitality we acknowledge. We also benefited from discussions
2147: with I. Cirac, M. Wilkens, and C.H. Bennett.
2148:
2149:
2150: \begin{thebibliography}{99}
2151:
2152: \bibitem{Feynman}R. Feynman, Simulating physics with computers,
2153: {\it Int. J. Theor. Phys. \bf21} (1982) 467-488, Reprinted in
2154: A.J.G. Hey (ed.), {\it Feynman and Computation}, Perseus Books
2155: 1999.
2156:
2157: \bibitem{optlat}O. Mandel, M. Greiner, A. Widera, T. Rom, T.W. H\"ansch,
2158: and I. Bloch,
2159: \jtitle{Coherent transport of neutral atoms in spin-dependent optical
2160: lattice potentials,}
2161: {\it Phys. Rev. Lett. \bf91}, 010407 (2003)
2162:
2163: \bibitem{mictrap}R. Dumke, M. Volk, T. Muether, F.B.J.
2164: Buchkremer, G. Birkl, and W. Ertmer,
2165: \jtitle{ Microoptical Realization of Arrays of Selectively Addressable
2166: Dipole Traps: A Scalable Configuration for Quantum Computation
2167: with Atomic Qubits,}
2168: {\it Phys. Rev. Lett. \bf89}, 097903 (2002) and quant-ph/0110140
2169:
2170: \bibitem{QdCA} Further away are proposals based on arrays of
2171: quantum dots, which are also published under the heading ``quantum
2172: cellular automata''. These are ideas for new hardware for {\it
2173: classical} computing, possibly replacing CMOS technology. In order
2174: to avoid confusion with the ideas in which quantum coherence plays
2175: a key role (as it does in our paper), many authors from that
2176: community are now using the more precise term ``quantum {\it dot}
2177: cellular automata''. For an overview see the home page of the
2178: Notre Dame group (www.nd.edu/\~{}qcahome), or: P.D. Tougaw, C.S.
2179: Lent, \jtitle{Logical devices implemented using quantum cellular
2180: automata,} {\it J.Appl.Phys. \bf75} (1994) 1818.
2181:
2182: \bibitem{Benjamin} S. C. Benjamin,
2183: \jtitle{Schemes for parallel quantum computation without local
2184: control of qubits,}
2185: {\it Phys. Rev. A \bf61} 020301 (2000)
2186:
2187: \bibitem{Watrous}J. Watrous: On one-dimensional quantum cellular automata.
2188: In Proceedings of the 36th Annual Symposium on Foundations of
2189: Computer Science, 1995, pp.~528--537.
2190:
2191: \bibitem{vanDam}W. van Dam: Quantum cellular automata, Master
2192: Thesis, Computer Science Nijmegen, Summer 1996
2193:
2194: \bibitem{Gruska}J. Gruska: {\it Quantum Computing}, (McGraw-Hill,
2195: Cambridge 1999). QCAs are treated in Section 4.3.
2196:
2197: \bibitem{santa}C. D\"urr and M. Santha: A decision procedure for
2198: unitary linear quantum cellular automata,
2199: quant-ph/9604007;\newline
2200: C. D\"urr, H. L\^eTanh and M. Santha,
2201: \jtitle{A decision procedure for
2202: well-formed linear quantum cellular automata,}
2203: {\it Rand. Struct. Algorithms \bf11}, 381-394 (1997)
2204: and cs.DS/9906024
2205:
2206: \bibitem{Brennen}G. K. Brennen and J. E. Williams,
2207: \jtitle{Entanglement dynamics in 1D quantum cellular automata, }
2208: quant-ph/0306056
2209:
2210: \bibitem{BraRo}O. Bratteli and D. Robinson:
2211: {\it Operator algebras and quantum statistical mechanics}, vol. I
2212: (Springer 1979)
2213:
2214: \bibitem{Paschen}A C*-algebraic framework was also applied to QCAs
2215: in a recent thesis. However, it was applied to the state side,
2216: rather than the observables, making locality properties harder to
2217: see. K. Paschen: \"Uber Reversibilit\"at, Nicht-Determiniertheit
2218: und Quantenrechnen in Zellularautomaten, Dissertation in
2219: Informatik (PhD Thesis in Computer Science), Karlsruhe 2002
2220:
2221: \bibitem{Haag}R. Haag: {\it Local quantum physics}, Springer 1996
2222:
2223: \bibitem{RiWe}S. Richter and R.F. Werner,
2224: \jtitle{Ergodicity of quantum cellular automata,}
2225: {\it J. Stat. Phys. \bf82} (1996) 963-998 and cond-mat/9504001
2226:
2227: \bibitem{Kari1}J. Kari,
2228: \jtitle{On the circuit depth of structurally reversible cellular automata,}
2229: {\it Fund.Inform. \bf34} (2003) 1–-15
2230:
2231: \bibitem{otherlat}The basic concepts in this paragraph work in any lattice
2232: structure. In fact, they do not even require translation
2233: invariance and could be formulated for possibly different spins
2234: (given by possibly infinite dimensional C*-algebras) localized on
2235: the nodes of a finite or infinite graph.
2236:
2237: \bibitem{onewaycomp}R. Raussendorf, D. E. Browne and H.-J. Briegel,
2238: \jtitle{The one-way quantum computer - a non-network model of quantum
2239: computation,}
2240: {\it J. Mod. Opt \bf 49}, 1299 (2002).
2241:
2242: \bibitem{Rich}D. Richardson,
2243: \jtitle{Tesselation with local transformations,}
2244: {\it J.Comp.Syst.Sci. \bf6} (1972) 373--388
2245:
2246: \bibitem{classicalW}What appears here as a pathology is allowed in
2247: the (periodic boundary version) of Watrous QCAs, i.e., the
2248: Wrapping Lemma fails for that structure. A more systematic
2249: study of the possibility shown by our example was carried out in\\
2250: S. Inokuchi, Y. Mizoguchi,
2251: \jtitle{Generalized partitioned quantum
2252: cellular automata and quantumization of classical CA,}
2253: quant-ph/0312102.
2254:
2255: \bibitem{Lloyd}S. Lloyd,
2256: \jtitle{A potentially realizable quantum computer,}
2257: {\it Science \bf261}, 1569--1571 (1993)
2258:
2259: \bibitem{cliff} We follow this terminology, although it is not clear what
2260: Clifford had to with this. In field theory these transformations
2261: would be called ``quasi-free'', or ``Bogolyubov automorphisms'',
2262: in phase space quantum mechanics ``metaplectic transformations''.
2263:
2264: \bibitem{OurCliff}D. Schlingemann, R.F. Werner, The structure of
2265: Clifford quantum cellular automata, in preparation.
2266:
2267: \bibitem{MarTo}T. Toffoli and M. Margolus,
2268: \jtitle{Invertible Cellular automata: a review,}
2269: {\it Physica D\bf 45}(1990) 229-253
2270:
2271: \bibitem{Zanardi}The support algebra of a single element, an
2272: interaction Hamiltonian was also introduced under the name
2273: ``interaction algebra'' by P.Zanardi: Stabilization of quantum
2274: information: a unified dynamical-algebraic approach,
2275: quant-ph/0203008
2276:
2277: \bibitem{Wootters} K.S. Gibbons, M.J. Hoffman, W.K. Wootters,
2278: \jtitle{Discrete phase space based on finite fields,}
2279: quant-ph/0401155
2280:
2281: \bibitem{Paz}J. P. Paz: Discrete Wigner functions and the phase space representation of
2282: quantum teleportation, quant-ph/0204150
2283:
2284: \bibitem{Aoun}B. Aoun, M. Tarifi, Quantum cellular automata,
2285: quant-ph/0401123.
2286:
2287: \bibitem{Tak}M. Takesaki, {\it Theory of operator algebras, I},
2288: Springer 1979
2289:
2290: \bibitem{Wolfram} S. Wolfram, {\it A new kind of science},
2291: (Self-published, Wolfram Media Inc. 2002)
2292:
2293: \bibitem{Zuse} K. Zuse, {\it Rechnender Raum},
2294: Schriften zur Datenverarbeitung, Band 1, Vieweg,
2295: Braunschweig 1969.
2296:
2297: \bibitem{Zeilinger}G. Gr\"ossing and A. Zeilinger,
2298: \jtitle{Quantum cellular automata,}
2299: {\it Complex Systems \bf2}, 197--208 (1988)
2300:
2301: \bibitem{Iwo}I. Bialynicki-Birula,
2302: \jtitle{Weyl, Dirac, and Maxwell equations on a lattice as unitary cellular automata, }
2303: {\it Phys. Rev. D \bf49}, 6920--6927 (1994)
2304:
2305: \bibitem{Meyer}D. A. Meyer,
2306: \jtitle{From quantum cellular automata to quantum lattice gases, }
2307: {\it J. Stat. Phys. \bf85},551--574 (1996)
2308:
2309: \bibitem{konno}N. Konno, K. Mitsuda, T. Soshi, H.J. Yoo,
2310: \jtitle{Quantum walks and reversible cellular automata},
2311: quant-ph/0403107
2312:
2313: \bibitem{Kempe}J. Kempe,
2314: \jtitle{Quantum random walks: an introductory overview,}
2315: {\it Contemp.Phys. \bf44} (2003) 307 -– 327
2316:
2317: \bibitem{Meyer2}D. A. Meyer: Unitarity in one dimensional nonlinear
2318: quantum cellular automata, quant-ph/9605023;
2319: From quantum cellular automata to quantum lattice gases, {\it
2320: J.Stat.Phys. \bf85} (1996) 551--574
2321:
2322: \bibitem{Seiler}A.I. Bobenko, R. Seiler (eds.),
2323: {it Discrete integrable geometry and physics},
2324: Clarendon Press, Oxford 1999. The book has several articles
2325: connecting to the strucutre mentionend in the text. It is best to
2326: pick up the pointers in the introduction by the editors.
2327:
2328: \bibitem{Dix}J. Dixmier, {\it C*-algebras}, North Holland 1977
2329:
2330:
2331:
2332:
2333:
2334:
2335: \end{thebibliography}
2336: \end{document}
2337: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2338:
2339: \bibitem{Durand}J. Durand-Lose: Representing reversible cellular
2340: automata with reversible block cellular automata. In R. Cori, J.
2341: Mazoyer, M. Morvan, and R. Mosery, (eds.): Discrete models,
2342: combinatorics, computation and geometry (DM-CCG'01), volume AA.
2343: Dicrete Mathematics and Theoretical Computer Science, 2001.
2344:
2345:
2346: \bibitem{Boykett}T. Boykett: Efficient exhasutive enumeration of reversible one
2347: dimensional cellular automata, manuscript 40 pages.
2348: http://verdi.algebra.uni-linz.ac.at/~tim/enumeration.ps.gz
2349:
2350: \bibitem{Zhmud}E.M. Zhmud, {\it Math. USSR Sbornik}, {\bf15}
2351: (1971)7-29, and {\bf 16} (1972)1-16.
2352:
2353: \
2354:
2355: [25] G. ’t Hooft, K.Isler and S. Kalitzin, “Quantum field
2356: theoretic behavior of a deterministic cellular automaton”, Nucl.
2357: Phys.B386,495 (1992).
2358:
2359: [26] K. Svozil, “Are quantum fields cellular automata?”, Physics
2360: Letters A, 153, (1986).
2361:
2362: [27] M.D. Kostin, “Cellular automata for quantum systems”, J.
2363: Phys. A 26, L209 (1993).
2364: