Computer Physics Communications 146 (2002) 295–301

www.elsevier.com/locate/cpc

Physical quantum algorithms

David A. Meyer
Project in Geometry and Physics, Department of Mathematics, University of California/San Diego, La Jolla, CA 92093-0112, USA
Received 7 September 2001

Abstract
I review the differences between classical and quantum systems, emphasizing the connection between no-hidden variable
theorems and superior computational power of quantum computers. Using quantum lattice gas automata as examples, I describe
possibilities for efficient simulation of quantum and classical systems with a quantum computer. I conclude with a list of research
directions.  2002 Published by Elsevier Science B.V.
PACS: 03.67.Lx; 05.10.-a

Keywords: Quantum simulation; Quantum lattice gas automata

1. Introduction ficiently on a quantum computer [2]. More recently,

Deutsch [3], Jozsa [4], Simon [5], Shor [6], Grover [7]
There are two paths towards quantum computing: and others have noted that a quantum computer could
one is teleological and the other is practical. The tele- solve classical problems as well. In this primarily ped-
ological path—described 35 years ago in the proph- agogical paper I describe some of the steps which
esy known as Moore’s Law [1]—leads down through have been taken along this practical path, and spec-
smaller and smaller device sizes where quantum ef- ulate about some steps further along it.
fects become wilder and wilder. Eventually, rather
than domesticating them for classical computation, ex-
2. Simulating quantum systems classically
perimental physicists and engineers believe they will
be able to preserve them for quantum computation.
Let me begin by reviewing the reasons quantum
The practical path, on the other hand, is paved with
systems are believed to be hard to simulate on clas-
the desire to solve specific problems efficiently. In an
sical computers. Traditionally these are known as ‘no
amusing role-reversal, it is theoretical physicists, com-
hidden variable theorems’. Each is a statement that
puter scientists, and mathematicians who follow this
no classical model with specified constraints can re-
path. The first steps along it were taken 20 years ago produce quantum mechanical results. Consideration
by Feynman, who suggested that since quantum sys- of two of them, the Gleason/Kochen–Specker theo-
tems seem to be very hard to simulate on a classical rem [8,9] and Bell’s theorem [10], reveals both their
computer, perhaps they could be simulated more ef- heuristic power and their weaknesses.
In 1957 Gleason proved that for Hilbert spaces of
E-mail address: [email protected] (D.A. Meyer). dimension at least 3, any non-negative measure on
0010-4655/02/$ – see front matter  2002 Published by Elsevier Science B.V.
PII: S 0 0 1 0 - 4 6 5 5 ( 0 2 ) 0 0 4 2 0 - 4
296 D.A. Meyer / Computer Physics Communications 146 (2002) 295–301

states which is quantum mechanical (i.e. for any or- The states for which Bell’s theorem rules out classi-
thogonal basis {êi } the measure sums to 1) must de- cal hidden variables are entangled, i.e. ones for which
rive from a density matrix [8]. In 1967 Kochen and the state of multiple particles cannot be described as
Specker made the contradiction with a classical hidden the product of states for each particle individually.
variable model more explicit [9]: They constructed a Since this is true for all but a set of measure 0 in the
finite set of unit vectors in R3 with the property that space of all pure states, Bell’s theorem and its gener-
every attempt to assign values 0 or 1 to each vec- alizations (see [15] for a recent survey) indicate that
tor satisfying the condition that in each orthogonal most quantum states cannot even be described by rea-
triple two vectors get 1 and the third gets 0 must fail. sonable (in the sense of local) classical models. This
That is, no classical ‘hidden variable’ can be assigned is a more subtle problem than simply the large size of
to pre-determine which outcome of each of some fi- the state space, which we consider next.
nite set of complete measurements of the spin-squared
of a spin-1 particle will be observed (since the spin
eigenvalues are {−1, 0, +1}, two of the spin-squareds 3. Dynamics
are 1 and one is 0 for any complete measurement).
Such a hidden variable would be non-contextual, in The dimension of the Hilbert space describing the
the sense that its value on each vector would specify state of a system of multiple particles grows exponen-
the spin-squared observed for that measurement, inde- tially in the number of particles: 2n for n spin- 21 par-
ticles, for example. This exponential explosion, how-
pendently of which complete measurement including
ever, is not enough to preclude classical simulation.
it is performed. One can argue, however, that noncon-
Consider a classical, probabilistic lattice gas. On a ho-
textuality is too strict a condition to place on hidden
mogeneous one-dimensional lattice of size n there are
variables—perhaps the results of measurement should
4n basis states si , since each lattice site can be occu-
depend on hidden variables inherent in the measur-
pied by no more than 1 particle with each of the two
ing device, which might differ for each complete mea-
possible velocities. A general state s is a convex com-
surement [11]. Furthermore, the measurements must
bination:
be exactly along the vectors Kochen and Specker con- n −1 n −1
4
4

structed, but from a computational complexity per-
spective, infinite precision is suspect [12]—and in fact s= pi si with pi = 1, pi
0.
i=0 i=0
one can show that without additional assumptions one
cannot prove a Kochen–Specker theorem using only Evolving the whole state, i.e. the probability distrib-
finite precision measurements [13]. ution, is therefore an exponentially difficult problem
Both of these weaknesses are absent from Bell’s in the size of the lattice. Nevertheless, such lattice
theorem [10]. He proved that the results of local gas models are used regularly (see, e.g., [16]). But
measurements on specific states of pairs of spin- 21 one does not evolve the whole probability distribution.
particles, i.e. vectors in C2 ⊗C2 , cannot be reproduced Rather, one samples it, by evolving a single si to a sin-
by any local, classical hidden variable model. Here gle si at the next timestep, using some random num-
ber generator. Multiple runs sample the final distribu-
‘local’ means restricted to individual particles. This
tion. A quantum lattice gas automaton (QLGA, which
result is robust against measurement imprecision, and
I will describe in more detail in §4) is also described at
locality of the hidden variables seems justified on
each timestep by a vector in a space with basis {|si
}—
physical grounds—the finite speed of light and the
where |·
is the standard Dirac notation for vectors in
locality of physical interactions. In fact, these are the
Hilbert space [17]:
same grounds upon which we base our models of
n −1
4
n −1
4

computation: At each timestep a classical or quantum
Turing machine changes only the state of the head and |ψ
= ai |si
with |ai |2 = 1, ai ∈ C.
the symbol written on the tape cell where the head is i=0 i=0
located [3,14]; it does not make non-local changes of Evolving the QLGA state has classical computational
all the cells of the tape simultaneously, for example. complexity comparable to evolving the whole state of
 D.A. Meyer / Computer Physics Communications 146 (2002) 295–301 297

the probabilistic LGA. But in the quantum case, this 4. Quantum simulations
cannot be reduced by sampling individual histories:
each has a complex amplitude so the histories with In §§1–3 I have tried to explain the heuristic that
each given final state interfere. classical simulation of quantum systems is difficult,
Thus interference seems to be the phenomenon while noting what remains to be proved to make
which makes quantum dynamics hard to simulate clas- such a claim rigorous. Now let us consider Feynman’s
sically. In fact, although the multi-particle structure of proposed solution: simulation with quantum comput-
a system is important, entanglement per se seems to ers [2]. The standard model of quantum computa-
be less relevant: In liquid state NMR quantum com- tion allows polynomially many local (i.e. acting non-
puting experiments [18], for example, the state is not trivially on only 1 or 2 qubits) gate operations [25].
entangled at any timestep (more precisely, since the This is a reasonable model since in principle it can
system is in a mixed state—a convex combination of be realized by a quantum system with a local Hamil-
pure states—the state is separable) [19]. Nevertheless, tonian. Feynman’s proposal has been verified in this
it seems to be difficult to construct a reasonable lo- model for quantum systems defined by local Hamil-
cal hidden variable model for the dynamics [20], i.e. tonians [26–31]. More exotic quantum systems can
the dynamics seems difficult to simulate classically. also be simulated efficiently with a standard quantum
To make this more than heuristic, however, we would computer: Fractional quantum Hall systems, for ex-
need a dynamical Gleason/Kochen–Specker/Bell-type ample, have Hamiltonians which vanish on the phys-
theorem which applies even for evolution through a se- ical states; the only nontrivial unitary transformations
have global (topological) origin. Nevertheless, Freed-
quence of unentangled states. Perhaps some hint about
man, Kitaev and Wang have shown that such topologi-
a way to do this may be found in Laflamme’s response
cal quantum field theories can be simulated efficiently
to the separability criticism of NMR quantum compu-
with a standard quantum computer [32].
tation [21].
A particularly simple architecture for a quantum
Of course, some demonstrations of the absence of
computer is a QLGA [33]. Although I’m not aware
classical models for quantum dynamics already exist.
of a demonstration that classical LGA are capable of
These are more commonly known as quantum algo-
universal computation, their similarity to the reversible
rithms for oracle problems; since each consists of a
billiard ball model of Margolus [34] suggests that
sequence of unitary operations, they are dynamical re- they may be; since QLGA specialize to deterministic
sults. Grover’s quantum search algorithm, for exam- LGA, they would be also. Whether they can efficiently
ple, solves the problem of identifying a ∈ {0, 1}n given (i.e. with polynomial overhead) simulate quantum gate
an oracle which responds to√a query x ∈ {0, 1}n by re- arrays is, I believe, also an open question. In the other
turning δxa , using only O( 2n ) queries [7]. Classi- direction, QLGA can be simulated efficiently on a
cally, any algorithm would require O(2n ) queries. For standard quantum computer, but have theoretical and
n > 2, the state is entangled at every timestep after possibly practical advantages: They directly simulate
the first [22]. Possibly more to the point is Bernstein quantum systems and are possibly more easily realized
and Vazirani’s algorithm which solves the problem of experimentally than arbitrary quantum gate arrays.
identifying a ∈ {0, 1}n given an oracle which responds The possible configurations for each particle on
to a query x ∈ {0, 1}n by returning x · a mod 2, us- a one-dimensional lattice L are labeled by pairs
ing only 1 quantum query [23]. Classically, any al- (x, α) ∈ L × {±1}, where x is the position and α
gorithm would require O(n) queries. And this quan- the velocity. A classical lattice gas evolution rule
tum algorithm works without creating entanglement at consists of an advection stage (x, α) → (x + α, α),
any timestep [24]. These results suggest that while a followed by a scattering stage. Each particle in a
theorem on the impossibility of efficient classical sim- QLGA [33] exists in states which  are superpositions
ulation of quantum dynamics may exist, it will have of the classical
 states: |ψ
= ψx,α |x, α
, where
to count all the elementary operations, not just the 1 = ψ|ψ
= ψ̄x,α ψx,α . The evolution rule must be
queries, which will presumably make it more difficult unitary; the most general with the same form as the
to prove. classical rule is:
298 D.A. Meyer / Computer Physics Communications 146 (2002) 295–301

advect

ψx,α |x, α
−→ ψx,α |x + α, α
5. Simulating classical systems
scatter

−→ ψx,α Sαα  |x + α, α 
, In the previous sections we have seen that there are
where the scattering matrix is quantum algorithms to efficiently simulate multipar-
  ticle quantum systems which seem to be difficult to
cos m i sin m simulate classically. Since, as I noted in the Introduc-
S= .
i sin m cos m tion, there are efficient quantum algorithms to solve
Fig. 1 illustrates this quantum evolution: at m = 0 classical problems, a natural question is whether quan-
it specializes to the classical deterministic lattice gas tum computers can simulate classical systems [41,42].
rule. The x = t → 0 limit of this discrete time Assuming quantum mechanics is a correct description
evolution is the Dirac equation [33]; the x 2 = t → of the world, the existence of a classical description
0 limit is the Schrödinger equation [35]. for macroscopic physics means that quantum com-
This QLGA model can be extended to include puters can simulate classical physics with constant
multiple particles with a unitary two particle scattering overhead—although the constant factor may be some-
rule thing like 1023 , i.e. a number of quantum degrees of
freedom sufficiently large that subsystems decohere
|x, α, x, −α
→ eis |x, α, x, −α
and can be identified as the classical objects to be sim-
shown in Fig. 1. With these rules the 1-dimensional ulated.
QLGA discretizes the quantum field theory described Can we do better? That is, could there be quantum
by the (1 + 1)-dimensional massive Thirring model speedups for classical physics? Yepez has proposed
[33,36]. These rules also preserve the symmetry (i.e. that the answer is ‘yes’. Using a “Type II” quantum
bosonic or fermionic) of the wave function under par- computer in which the state is measured, locally, af-
ticle exchange [37]. The QLGA rules can be general- ter each timestep and then reset using a lattice Boltz-
ized to discretize the multi-particle Schrödinger equa- mann rule [43]. A model like this can achieve at most a
tion in arbitrary dimensions [35]; it seems more diffi- constant speedup, corresponding to reduced computa-
cult, however, to create QLGA which discretize rela- tional cost for local evolution. In practice, of course, a
tivistic evolution in higher dimensions [38]. large constant improvement can be tremendously use-
The fact that the QLGA rules are homogeneous, ful, but perhaps it is possible to do better. More pre-
i.e. the same at each lattice site and at each timestep, cisely, using a standard quantum computer, can classi-
suggests that they might be easier to implement than cal systems be simulated more efficiently than is pos-
general quantum gate arrays which are not. Possi- sible classically? Lidar and Biham have shown that
ble physical systems in which they might be imple- the answer to this question is also ‘yes’, for the non-
mented include crystals—as originally proposed by dynamical problem of sampling the ground state dis-
Feynman [2] and more recently in the context of solid tribution of a spin glass [41].
state NMR [39]—or optical lattices [40]. A detailed There are also QLGA results which suggest that
proposal for physical implementation in such systems certain aspects of classical dynamics can be simulated
could motivate experimental work towards realization more efficiently quantum mechanically. Consider clas-
of QLGA. sical diffusion of a particle in a linear potential, as
shown in Fig. 2. A discrete model for the evolution
is a biased random walk, with prob(x) ∝ e−∇V x .
The results of simulating the evolution with a classical
(probabilistic) LGA are shown in Fig. 3. The average
position of the particle satisfies
   
x(t) − x(0) ∝ −∇V t.
In QLGA the evolution rules can be modified to
include a potential by incorporating an x dependent
Fig. 1. The general evolution rules for the one-dimensional QLGA. phase multiplication, i.e. e−iV (x) at each timestep [44],
 D.A. Meyer / Computer Physics Communications 146 (2002) 295–301 299

Fig. 2. The probability distribution for the position of a classical Fig. 4. The evolution of expected position for the QLGA model of a
particle after diffusing in a linear potential. The particle was initially quantum particle subject to a linear potential.
at the origin.
lation of more complicated classical dynamics may be
possible.

6. Conclusion

In conclusion, let me reiterate the open questions

discussed in this paper:

Is there a proof that (some) quantum dynamics is

difficult to simulate classically? Can it be difficult
even when the state is unentangled (separable) at
each timestep?
Fig. 3. The evolution of expected position for the random walk
model of a classical particle diffusing in a linear potential.
In case QLGA become a practical architecture for
quantum computers, can they simulate the standard
which one might imagine implementing in a physical model of quantum computation with no more than
system with an applied, spatially varying magnetic polynomial overhead?
field, for example. Fig. 4 shows the result of a QLGA
simulation with a linear potential. Now the average What are possible physical implementations of
position of the particle approximately satisfies QLGA?
    What are the correct QLGA models for relativistic
x(t) − x(0) ∝ −∇V t 2 .
quantum systems in more than 1 spatial dimension?
That is, this quantum system simulates the evolution
of the average position of a classical particle diffus- and most importantly,
ing in a linear potential quadratically faster than does
the classical simulation shown in Fig. 3. I must em- Are there quantum algorithms which speed up the
phasize that it is only the average position which is simulation of classical physics?
being simulated accurately, not the whole probabil-
ity distribution. Furthermore, the quadratic speedup Positive answers to this last question will broaden the
only holds on timescales t  2π/∇V . On longer possible uses for a quantum computer and help justify
timescales the evolution is periodic [45]. Nevertheless, the immense commitment of resources which seems
this very simple example suggests that efficient simu- likely to be necessary to develop a scalable one.
300 D.A. Meyer / Computer Physics Communications 146 (2002) 295–301

Acknowledgements [13] D.A. Meyer, Finite precision measurement nullifies the

Kochen–Specker theorem, Phys. Rev. Lett. 83 (1999) 3751–
3754.
This work was supported in part by the National
[14] A.M. Turing, On computable numbers, with an application
Security Agency (NSA) and Advanced Research and to the Entscheidungsproblem, Proc. London Math. Soc. 442
Development Activity (ARDA) under Army Research (1936) 230–265.
Office (ARO) contract number DAAG55-98-1-0376, [15] R.F. Werner, M.M. Wolf, Quantum Information and Computa-
and by the Air Force Office of Scientific Research tion 1 (2001) 1–25.
(AFOSR) under grant number F49620-01-1-0494. [16] G.D. Doolen, U. Frisch, B. Hasslacher, S. Orszag, S. Wolfram
(Eds.), Lattice Gas Methods for Partial Differential Equations,
Proceedings, Santa Fe Institute Studies in the Science of
Complexity, Vol. IV, Addison-Wesley, New York, 1990.
References [17] P.A.M. Dirac, The Principles of Quantum Mechanics, 4th edn.,
Oxford University Press, Oxford, 1958.
[1] G.E. Moore, Cramming more components onto integrated [18] I.L. Chuang, L.M.K. Vandersypen, X. Zhou, D.W. Leung,
circuits, Electronics 38 (1965) 114–117. S. Lloyd, Experimental realization of a quantum algorithm,
[2] R.P. Feynman, Simulating physics with computers, Int. J. Nature 393 (1998) 143–146;
Theor. Phys. 21 (1982) 467–488; Quantum mechanical com- D.G. Gory, W. Mass, M. Price, E. Knill, R. Laflamme, W.H.
puters, Found. Phys. 16 (1986) 507–531. Zurek, T.F. Havel, S.S. Somaroo, Experimental quantum error
[3] D. Deutsch, Quantum theory, the Church–Turing principle correction, Phys. Rev. Lett. 81 (1998) 2152–2155;
and the universal quantum computer, Proc. Roy. Soc. London J.A. Jones, M. Mosca, R.H. Hansen, Implementation of a
A 400 (1985) 97–117. quantum search algorithm on a nuclear magnetic resonance
[4] D. Deutsch, R. Jozsa, Rapid solution of problems by quantum computer, Nature 393 (1998) 344–346.
computation, Proc. Roy. Soc. London A 439 (1992) 553–558. [19] S.L. Braunstein, C.M. Caves, R. Jozsa, N. Linden, S. Popescu,
[5] D.R. Simon, On the power of quantum computation, in: R. Schack, Separability of very noisy mixed states and implica-
S. Goldwasser (Ed.), Proceedings of the 35th Symposium on tions for NMR quantum computing, Phys. Rev. Lett. 83 (1999)
Foundations of Computer Science, Santa Fe, NM, 20–22 No- 1054–1057.
vember 1994, IEEE Computer Society Press, Los Alamitos, [20] R. Schack, C.M. Caves, Classical model for bulk-ensemble
CA, 1994, pp. 116–123; On the power of quantum computa- NMR quantum computation, Phys. Rev. A 60 (1999) 4354–
tion, SIAM J. Comput. 26 (1997) 1474–1483. 4362; subsequent work indicates that there is a local classi-
[6] P.W. Shor, Algorithms for quantum computation: discrete log- cal model with exponentially much local information (N.C.
arithms and factoring, in: S. Goldwasser (Ed.), Proceedings Menicucci, C.M. Caves, Local realistic model for the dynam-
of the 35th Symposium on Foundations of Computer Science, ics of bulk-ensemble NMR information processing, quant-
Santa Fe, NM, 20–22 November 1994, IEEE Computer Soci- ph/0111152).
ety Press, Los Alamitos, CA, 1994, pp. 124–134; Polynomial- [21] R. Laflamme, Review of ‘Separability of very noisy mixed
time algorithms for prime factorization and discrete logarithms states and implications for NMR quantum computing’,
on a quantum computer, SIAM J. Comput. 26 (1997) 1484– Quick Reviews in Quantum Computation and Information,
1509. http://quickreviews.org/qc/.
[7] L.K. Grover, A fast quantum mechanical algorithm for data- [22] D.A. Meyer, N.R. Wallach, Global entanglement in multiparti-
base search, in: Proceedings of the 28th Annual ACM Sym- cle systems, quant-ph/0108104.
posium on the Theory of Computing, Philadelphia, PA, 22– [23] E. Bernstein, U. Vazirani, Quantum complexity theory, in:
24 May 1996, ACM, New York, 1996, pp. 212–219; Quantum Proceedings of the 25th ACM Symposium on Theory of
mechanics helps in searching for a needle in a haystack, Phys. Computing, San Diego, CA, 16–18 May 1993, ACM Press,
Rev. Lett. 79 (1997) 325–328. New York, 1993, pp. 11–20; Quantum complexity theory,
[8] A.M. Gleason, Measures on the closed subspaces of a Hilbert SIAM J. Comput. 26 (1997) 1411–1473.
space, J. Math. Mech. 6 (1957) 885–893. [24] D.A. Meyer, Sophisticated quantum search without entangle-
[9] S. Kochen, E.P. Specker, The problem of hidden variables in ment, Phys. Rev. Lett. 85 (2000) 2014–2017.
quantum mechanics, J. Math. Mech. 17 (1967) 59–87. [25] M.H. Freedman, Foundations of Computational Mathematics 2
[10] J.S. Bell, On the Einstein–Podolsky–Rosen paradox, Physics 1 (2002) 145–154.
(1964) 195–200. [26] C. Zalka, Efficient simulation of quantum systems by quantum
[11] J.S. Bell, On the problem of hidden variables in quantum computers, Proc. Roy. Soc. London A 454 (1998) 313–322.
mechanics, Rev. Mod. Phys. 38 (1966) 447–452. [27] S. Wiesner, Simulations of many-body quantum systems by a
[12] A. Schönhage, On the power of random access machines, in: quantum computer, quant-ph/9603028.
H.A. Maurer (Ed.), Automata, Languages and Programming, [28] S. Lloyd, Universal quantum simulators, Science 273 (1996)
Proceedings of the 6th Colloquium, Graz, Austria, 16–20 July 1073–1078.
1979, Lecture Notes in Computer Science, Vol. 71, Springer, [29] B.M. Boghosian, W. Taylor IV, Simulating quantum mechan-
New York, 1979, pp. 520–529. ics on a quantum computer, Phys. Rev. D 120 (1998) 30–42.
 D.A. Meyer / Computer Physics Communications 146 (2002) 295–301 301

[30] D.S. Abrams, S. Lloyd, Simulation of many-body Fermi [38] G.V. Riazanov, The Feynman path integral for the Dirac
systems on a universal quantum computer, Phys. Rev. Lett. 79 equation, Sov. Phys. JETP 6 (1958) 1107–1113.
(1997) 2586–2589. [39] F. Yamaguchi, Y. Yamamoto, Crystal lattice quantum com-
[31] G. Ortiz, J.E. Gubernatis, E. Knill, R. Laflamme, Phys. Rev. puter, Appl. Phys. A 68 (1999) 1–8.
A 64 (2001) 022319/1-14. [40] I.H. Deutsch, G.K. Brennen, P.S. Jessen, Quantum computing
[32] M.H. Freedman, A. Kitaev, Z. Wang, Simulation of topological with neutral atoms in an optical lattice, Fortsch. Phys. 48
field theories by quantum computers, quant-ph/0001071. (2000) 925–943.
[33] D.A. Meyer, From quantum cellular automata to quantum [41] D.A. Lidar, O. Biham, Simulating Ising spin glasses on a
lattice gases, J. Statist. Phys. 85 (1996) 551–574. quantum computer, Phys. Rev. E 56 (1997) 3661–3681.
[34] N. Margolus, Physics-like models of computation, Physica [42] J. Yepez, A quantum lattice-gas model for computation of fluid
D 10 (1984) 81–95. dynamics, Phys. Rev. E 63 (2001) 046702.
[35] B.M. Boghosian, W. Taylor IV, A quantum lattice-gas model [43] J. Yepez, Lattice-gas quantum computation, Int. J. Mod. Phys.
for the many-particle Schrödinger equation in d dimensions, C 9 (1998) 1587–1596.
Phys. Rev. E 8 (1997) 705–716. [44] D.A. Meyer, Quantum mechanics of lattice gas automata: One
[36] C. Destri, H.J. de Vega, Light-cone lattice approach to fermi- particle plane waves and potentials, Phys. Rev. E 55 (1997)
onic theories in 2D, Nucl. Phys. B 290 (1987) 363–391. 5261–5269.
[37] D.A. Meyer, Quantum lattice gases and their invariants, Int. J. [45] D.A. Meyer, H. Blumer, J. Statist. Phys. 107 (2002) 225–239.
Mod. Phys. C 8 (1997) 717–735.